wangzixiang
/
lerobot
mirror of https://github.com/huggingface/lerobot.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into bugfix/opencv-script-types

This commit is contained in:
Steven Palma 2025-04-17 15:18:50 +02:00 committed by GitHub
parent 8af3316a87 a8db91c40e
commit 18d969a6ec
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
706 changed files with 17244 additions and 24412 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
.dockerignore
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Misc
.git
tmp
@ -59,7 +73,7 @@ pip-log.txt
pip-delete-this-directory.txt
# Unit test / coverage reports
!tests/data
!tests/artifacts
htmlcov/
.tox/
.nox/

14
.gitattributes vendored
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
*.memmap filter=lfs diff=lfs merge=lfs -text
*.stl filter=lfs diff=lfs merge=lfs -text
*.safetensors filter=lfs diff=lfs merge=lfs -text

14
.github/ISSUE_TEMPLATE/bug-report.yml vendored
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
name: "\U0001F41B Bug Report"
description: Submit a bug report to help us improve LeRobot
body:

26
.github/workflows/build-docker-images.yml vendored
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Inspired by
# https://github.com/huggingface/peft/blob/main/.github/workflows/build_docker_images.yml
name: Builds
@ -8,6 +22,8 @@ on:
schedule:
- cron: "0 1 * * *"
permissions: {}
env:
PYTHON_VERSION: "3.10"
@ -25,11 +41,14 @@ jobs:
- name: Set up Docker Buildx
uses: docker/setup-buildx-action@v3
with:
cache-binary: false
- name: Check out code
uses: actions/checkout@v4
with:
lfs: true
persist-credentials: false
- name: Login to DockerHub
uses: docker/login-action@v3
@ -60,11 +79,14 @@ jobs:
- name: Set up Docker Buildx
uses: docker/setup-buildx-action@v3
with:
cache-binary: false
- name: Check out code
uses: actions/checkout@v4
with:
lfs: true
persist-credentials: false
- name: Login to DockerHub
uses: docker/login-action@v3
@ -89,9 +111,13 @@ jobs:
steps:
- name: Set up Docker Buildx
uses: docker/setup-buildx-action@v3
with:
cache-binary: false
- name: Check out code
uses: actions/checkout@v4
with:
persist-credentials: false
- name: Login to DockerHub
uses: docker/login-action@v3

16
.github/workflows/nightly-tests.yml vendored
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Inspired by
# https://github.com/huggingface/peft/blob/main/.github/workflows/nightly.yml
name: Nightly
@ -7,6 +21,8 @@ on:
schedule:
- cron: "0 2 * * *"
permissions: {}
# env:
# SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
jobs:

74
.github/workflows/quality.yml vendored
View File

@ -1,15 +1,29 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
name: Quality
on:
workflow_dispatch:
workflow_call:
pull_request:
branches:
- main
push:
branches:
- main
permissions: {}
env:
PYTHON_VERSION: "3.10"
@ -19,7 +33,9 @@ jobs:
runs-on: ubuntu-latest
steps:
- name: Checkout Repository
uses: actions/checkout@v3
uses: actions/checkout@v4
with:
persist-credentials: false
- name: Set up Python
uses: actions/setup-python@v4
@ -30,55 +46,27 @@ jobs:
id: get-ruff-version
run: |
RUFF_VERSION=$(awk '/repo: https:\/\/github.com\/astral-sh\/ruff-pre-commit/{flag=1;next}/rev:/{if(flag){print $2;exit}}' .pre-commit-config.yaml)
echo "RUFF_VERSION=${RUFF_VERSION}" >> $GITHUB_ENV
echo "ruff_version=${RUFF_VERSION}" >> $GITHUB_OUTPUT
- name: Install Ruff
run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
env:
RUFF_VERSION: ${{ steps.get-ruff-version.outputs.ruff_version }}
run: python -m pip install "ruff==${RUFF_VERSION}"
- name: Ruff check
run: ruff check
run: ruff check --output-format=github
- name: Ruff format
run: ruff format --diff
poetry_check:
name: Poetry check
typos:
name: Typos
runs-on: ubuntu-latest
steps:
- name: Checkout Repository
uses: actions/checkout@v3
uses: actions/checkout@v4
with:
persist-credentials: false
- name: Install poetry
run: pipx install poetry
- name: Poetry check
run: poetry check
poetry_relax:
name: Poetry relax
runs-on: ubuntu-latest
steps:
- name: Checkout Repository
uses: actions/checkout@v3
- name: Install poetry
run: pipx install poetry
- name: Install poetry-relax
run: poetry self add poetry-relax
- name: Poetry relax
id: poetry_relax
run: |
output=$(poetry relax --check 2>&1)
if echo "$output" | grep -q "Proposing updates"; then
echo "$output"
echo ""
echo "Some dependencies have caret '^' version requirement added by poetry by default."
echo "Please replace them with '>='. You can do this by hand or use poetry-relax to do this."
exit 1
else
echo "$output"
fi
- name: typos-action
uses: crate-ci/typos@v1.29.10

33
.github/workflows/test-docker-build.yml vendored
View File

@ -1,15 +1,29 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Inspired by
# https://github.com/huggingface/peft/blob/main/.github/workflows/test-docker-build.yml
name: Test Dockerfiles
on:
pull_request:
branches:
- main
paths:
# Run only when DockerFile files are modified
- "docker/**"
permissions: {}
env:
PYTHON_VERSION: "3.10"
@ -22,29 +36,28 @@ jobs:
steps:
- name: Check out code
uses: actions/checkout@v4
with:
persist-credentials: false
- name: Get changed files
id: changed-files
uses: tj-actions/changed-files@v44
uses: tj-actions/changed-files@3f54ebb830831fc121d3263c1857cfbdc310cdb9 #v42
with:
files: docker/**
json: "true"
- name: Run step if only the files listed above change
- name: Run step if only the files listed above change # zizmor: ignore[template-injection]
if: steps.changed-files.outputs.any_changed == 'true'
id: set-matrix
env:
ALL_CHANGED_FILES: ${{ steps.changed-files.outputs.all_changed_files }}
run: |
echo "matrix=${{ steps.changed-files.outputs.all_changed_files}}" >> $GITHUB_OUTPUT
build_modified_dockerfiles:
name: Build modified Docker images
needs: get_changed_files
runs-on:
group: aws-general-8-plus
if: ${{ needs.get_changed_files.outputs.matrix }} != ''
if: needs.get_changed_files.outputs.matrix != ''
strategy:
fail-fast: false
matrix:
@ -52,9 +65,13 @@ jobs:
steps:
- name: Set up Docker Buildx
uses: docker/setup-buildx-action@v3
with:
cache-binary: false
- name: Check out code
uses: actions/checkout@v4
with:
persist-credentials: false
- name: Build Docker image
uses: docker/build-push-action@v5

131
.github/workflows/test.yml vendored
View File

@ -1,15 +1,28 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
name: Tests
on:
pull_request:
branches:
- main
paths:
- "lerobot/**"
- "tests/**"
- "examples/**"
- ".github/**"
- "poetry.lock"
- "pyproject.toml"
- ".pre-commit-config.yaml"
- "Makefile"
- ".cache/**"
push:
@ -20,10 +33,16 @@ on:
- "tests/**"
- "examples/**"
- ".github/**"
- "poetry.lock"
- "pyproject.toml"
- ".pre-commit-config.yaml"
- "Makefile"
- ".cache/**"
permissions: {}
env:
UV_VERSION: "0.6.0"
jobs:
pytest:
name: Pytest
@ -34,6 +53,7 @@ jobs:
- uses: actions/checkout@v4
with:
lfs: true # Ensure LFS files are pulled
persist-credentials: false
- name: Install apt dependencies
# portaudio19-dev is needed to install pyaudio
@ -41,25 +61,19 @@ jobs:
sudo apt-get update && \
sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
- name: Install poetry
run: |
pipx install poetry && poetry config virtualenvs.in-project true
echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
# TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
- name: Set up Python 3.10
uses: actions/setup-python@v5
- name: Install uv and python
uses: astral-sh/setup-uv@v5
with:
enable-cache: true
version: ${{ env.UV_VERSION }}
python-version: "3.10"
cache: "poetry"
- name: Install poetry dependencies
run: |
poetry install --all-extras
- name: Install lerobot (all extras)
run: uv sync --all-extras
- name: Test with pytest
run: |
pytest tests -v --cov=./lerobot --durations=0 \
uv run pytest tests -v --cov=./lerobot --durations=0 \
-W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
-W ignore::UserWarning:torch.utils.data.dataloader:558 \
-W ignore::UserWarning:gymnasium.utils.env_checker:247 \
@ -74,66 +88,63 @@ jobs:
- uses: actions/checkout@v4
with:
lfs: true # Ensure LFS files are pulled
persist-credentials: false
- name: Install apt dependencies
run: sudo apt-get update && sudo apt-get install -y ffmpeg
- name: Install poetry
run: |
pipx install poetry && poetry config virtualenvs.in-project true
echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
# TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
- name: Set up Python 3.10
uses: actions/setup-python@v5
- name: Install uv and python
uses: astral-sh/setup-uv@v5
with:
enable-cache: true
version: ${{ env.UV_VERSION }}
python-version: "3.10"
- name: Install poetry dependencies
run: |
poetry install --extras "test"
- name: Install lerobot
run: uv sync --extra "test"
- name: Test with pytest
run: |
pytest tests -v --cov=./lerobot --durations=0 \
uv run pytest tests -v --cov=./lerobot --durations=0 \
-W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
-W ignore::UserWarning:torch.utils.data.dataloader:558 \
-W ignore::UserWarning:gymnasium.utils.env_checker:247 \
&& rm -rf tests/outputs outputs
# TODO(aliberts, rcadene): redesign after v2 migration / removing hydra
# end-to-end:
# name: End-to-end
# runs-on: ubuntu-latest
# env:
# MUJOCO_GL: egl
# steps:
# - uses: actions/checkout@v4
# with:
# lfs: true # Ensure LFS files are pulled
end-to-end:
name: End-to-end
runs-on: ubuntu-latest
env:
MUJOCO_GL: egl
steps:
- uses: actions/checkout@v4
with:
lfs: true # Ensure LFS files are pulled
persist-credentials: false
# - name: Install apt dependencies
# # portaudio19-dev is needed to install pyaudio
# run: |
# sudo apt-get update && \
# sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
- name: Install apt dependencies
# portaudio19-dev is needed to install pyaudio
run: |
sudo apt-get update && \
sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
# - name: Install poetry
# run: |
# pipx install poetry && poetry config virtualenvs.in-project true
# echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
- name: Install uv and python
uses: astral-sh/setup-uv@v5
with:
enable-cache: true
version: ${{ env.UV_VERSION }}
python-version: "3.10"
# - name: Set up Python 3.10
# uses: actions/setup-python@v5
# with:
# python-version: "3.10"
# cache: "poetry"
- name: Install lerobot (all extras)
run: |
uv venv
uv sync --all-extras
# - name: Install poetry dependencies
# run: |
# poetry install --all-extras
- name: venv
run: |
echo "PYTHON_PATH=${{ github.workspace }}/.venv/bin/python" >> $GITHUB_ENV
# - name: Test end-to-end
# run: |
# make test-end-to-end \
# && rm -rf outputs
- name: Test end-to-end
run: |
make test-end-to-end \
&& rm -rf outputs

19
.github/workflows/trufflehog.yml vendored
View File

@ -1,10 +1,23 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
on:
push:
name: Secret Leaks
permissions:
contents: read
permissions: {}
jobs:
trufflehog:
@ -14,6 +27,8 @@ jobs:
uses: actions/checkout@v4
with:
fetch-depth: 0
persist-credentials: false
- name: Secret Scanning
uses: trufflesecurity/trufflehog@main
with:

20
.gitignore vendored
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Logging
logs
tmp
@ -49,6 +63,10 @@ share/python-wheels/
*.egg
MANIFEST
# uv/poetry lock files
poetry.lock
uv.lock
# PyInstaller
# Usually these files are written by a python script from a template
# before PyInstaller builds the exe, so as to inject date/other infos into it.
@ -60,7 +78,7 @@ pip-log.txt
pip-delete-this-directory.txt
# Unit test / coverage reports
!tests/data
!tests/artifacts
htmlcov/
.tox/
.nox/

61
.pre-commit-config.yaml
View File

@ -1,7 +1,29 @@
exclude: ^(tests/data)
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
exclude: "tests/artifacts/.*\\.safetensors$"
default_language_version:
python: python3.10
repos:
##### Meta #####
- repo: meta
hooks:
- id: check-useless-excludes
- id: check-hooks-apply
##### Style / Misc. #####
- repo: https://github.com/pre-commit/pre-commit-hooks
rev: v5.0.0
hooks:
@ -13,25 +35,40 @@ repos:
- id: check-toml
- id: end-of-file-fixer
- id: trailing-whitespace
- repo: https://github.com/adhtruong/mirrors-typos
rev: v1.31.1
hooks:
- id: typos
args: [--force-exclude]
- repo: https://github.com/asottile/pyupgrade
rev: v3.19.0
rev: v3.19.1
hooks:
- id: pyupgrade
- repo: https://github.com/astral-sh/ruff-pre-commit
rev: v0.8.2
rev: v0.11.5
hooks:
- id: ruff
args: [--fix]
- id: ruff-format
- repo: https://github.com/python-poetry/poetry
rev: 1.8.0
hooks:
- id: poetry-check
- id: poetry-lock
args:
- "--check"
- "--no-update"
##### Security #####
- repo: https://github.com/gitleaks/gitleaks
rev: v8.21.2
rev: v8.24.3
hooks:
- id: gitleaks
- repo: https://github.com/woodruffw/zizmor-pre-commit
rev: v1.5.2
hooks:
- id: zizmor
- repo: https://github.com/PyCQA/bandit
rev: 1.8.3
hooks:
- id: bandit
args: ["-c", "pyproject.toml"]
additional_dependencies: ["bandit[toml]"]

53
CONTRIBUTING.md
View File

@ -129,38 +129,71 @@ Follow these steps to start contributing:
🚨 **Do not** work on the `main` branch.
4. for development, we use `poetry` instead of just `pip` to easily track our dependencies.
If you don't have it already, follow the [instructions](https://python-poetry.org/docs/#installation) to install it.
4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
Set up a development environment with conda or miniconda:
```bash
conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev
```
To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
If you're using `uv`, it can manage python versions so you can instead do:
```bash
poetry install --sync --extras "dev test"
uv venv --python 3.10 && source .venv/bin/activate
```
To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
using `poetry`
```bash
poetry sync --extras "dev test"
```
using `uv`
```bash
uv sync --extra dev --extra test
```
You can also install the project with all its dependencies (including environments):
using `poetry`
```bash
poetry install --sync --all-extras
poetry sync --all-extras
```
using `uv`
```bash
uv sync --all-extras
```
> **Note:** If you don't install simulation environments with `--all-extras`, the tests that require them will be skipped when running the pytest suite locally. However, they *will* be tested in the CI. In general, we advise you to install everything and test locally before pushing.
Whichever command you chose to install the project (e.g. `poetry install --sync --all-extras`), you should run it again when pulling code with an updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the new dependencies.
Whichever command you chose to install the project (e.g. `poetry sync --all-extras`), you should run it again when pulling code with an updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the new dependencies.
The equivalent of `pip install some-package`, would just be:
using `poetry`
```bash
poetry add some-package
```
When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
using `uv`
```bash
poetry lock --no-update
uv add some-package
```
When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
using `poetry`
```bash
poetry lock
```
using `uv`
```bash
uv lock
```
5. Develop the features on your branch.
As you work on the features, you should make sure that the test suite
@ -195,7 +228,7 @@ Follow these steps to start contributing:
git commit
```
Note, if you already commited some changes that have a wrong formatting, you can use:
Note, if you already committed some changes that have a wrong formatting, you can use:
```bash
pre-commit run --all-files
```
@ -258,7 +291,7 @@ sudo apt-get install git-lfs
git lfs install
```
Pull artifacts if they're not in [tests/data](tests/data)
Pull artifacts if they're not in [tests/artifacts](tests/artifacts)
```bash
git lfs pull
```

250
Makefile
View File

@ -1,11 +1,25 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
.PHONY: tests
PYTHON_PATH := $(shell which python)
# If Poetry is installed, redefine PYTHON_PATH to use the Poetry-managed Python
POETRY_CHECK := $(shell command -v poetry)
ifneq ($(POETRY_CHECK),)
PYTHON_PATH := $(shell poetry run which python)
# If uv is installed and a virtual environment exists, use it
UV_CHECK := $(shell command -v uv)
ifneq ($(UV_CHECK),)
PYTHON_PATH := $(shell .venv/bin/python)
endif
export PATH := $(dir $(PYTHON_PATH)):$(PATH)
@ -20,171 +34,109 @@ build-gpu:
test-end-to-end:
${MAKE} DEVICE=$(DEVICE) test-act-ete-train
${MAKE} DEVICE=$(DEVICE) test-act-ete-train-resume
${MAKE} DEVICE=$(DEVICE) test-act-ete-eval
${MAKE} DEVICE=$(DEVICE) test-act-ete-train-amp
${MAKE} DEVICE=$(DEVICE) test-act-ete-eval-amp
${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train-with-online
${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
${MAKE} DEVICE=$(DEVICE) test-default-ete-eval
${MAKE} DEVICE=$(DEVICE) test-act-pusht-tutorial
test-act-ete-train:
python lerobot/scripts/train.py \
policy=act \
policy.dim_model=64 \
env=aloha \
wandb.enable=False \
training.offline_steps=2 \
training.online_steps=0 \
eval.n_episodes=1 \
eval.batch_size=1 \
device=$(DEVICE) \
training.save_checkpoint=true \
training.save_freq=2 \
policy.n_action_steps=20 \
policy.chunk_size=20 \
training.batch_size=2 \
training.image_transforms.enable=true \
hydra.run.dir=tests/outputs/act/
--policy.type=act \
--policy.dim_model=64 \
--policy.n_action_steps=20 \
--policy.chunk_size=20 \
--policy.device=$(DEVICE) \
--env.type=aloha \
--env.episode_length=5 \
--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
--dataset.image_transforms.enable=true \
--dataset.episodes="[0]" \
--batch_size=2 \
--steps=4 \
--eval_freq=2 \
--eval.n_episodes=1 \
--eval.batch_size=1 \
--save_freq=2 \
--save_checkpoint=true \
--log_freq=1 \
--wandb.enable=false \
--output_dir=tests/outputs/act/
test-act-ete-train-resume:
python lerobot/scripts/train.py \
--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
--resume=true
test-act-ete-eval:
python lerobot/scripts/eval.py \
-p tests/outputs/act/checkpoints/000002/pretrained_model \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=8 \
device=$(DEVICE) \
test-act-ete-train-amp:
python lerobot/scripts/train.py \
policy=act \
policy.dim_model=64 \
env=aloha \
wandb.enable=False \
training.offline_steps=2 \
training.online_steps=0 \
eval.n_episodes=1 \
eval.batch_size=1 \
device=$(DEVICE) \
training.save_checkpoint=true \
training.save_freq=2 \
policy.n_action_steps=20 \
policy.chunk_size=20 \
training.batch_size=2 \
hydra.run.dir=tests/outputs/act_amp/ \
training.image_transforms.enable=true \
use_amp=true
test-act-ete-eval-amp:
python lerobot/scripts/eval.py \
-p tests/outputs/act_amp/checkpoints/000002/pretrained_model \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=8 \
device=$(DEVICE) \
use_amp=true
--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
--policy.device=$(DEVICE) \
--env.type=aloha \
--env.episode_length=5 \
--eval.n_episodes=1 \
--eval.batch_size=1
test-diffusion-ete-train:
python lerobot/scripts/train.py \
policy=diffusion \
policy.down_dims=\[64,128,256\] \
policy.diffusion_step_embed_dim=32 \
policy.num_inference_steps=10 \
env=pusht \
wandb.enable=False \
training.offline_steps=2 \
training.online_steps=0 \
eval.n_episodes=1 \
eval.batch_size=1 \
device=$(DEVICE) \
training.save_checkpoint=true \
training.save_freq=2 \
training.batch_size=2 \
training.image_transforms.enable=true \
hydra.run.dir=tests/outputs/diffusion/
--policy.type=diffusion \
--policy.down_dims='[64,128,256]' \
--policy.diffusion_step_embed_dim=32 \
--policy.num_inference_steps=10 \
--policy.device=$(DEVICE) \
--env.type=pusht \
--env.episode_length=5 \
--dataset.repo_id=lerobot/pusht \
--dataset.image_transforms.enable=true \
--dataset.episodes="[0]" \
--batch_size=2 \
--steps=2 \
--eval_freq=2 \
--eval.n_episodes=1 \
--eval.batch_size=1 \
--save_checkpoint=true \
--save_freq=2 \
--log_freq=1 \
--wandb.enable=false \
--output_dir=tests/outputs/diffusion/
test-diffusion-ete-eval:
python lerobot/scripts/eval.py \
-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=8 \
device=$(DEVICE) \
--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
--policy.device=$(DEVICE) \
--env.type=pusht \
--env.episode_length=5 \
--eval.n_episodes=1 \
--eval.batch_size=1
test-tdmpc-ete-train:
python lerobot/scripts/train.py \
policy=tdmpc \
env=xarm \
env.task=XarmLift-v0 \
dataset_repo_id=lerobot/xarm_lift_medium \
wandb.enable=False \
training.offline_steps=2 \
training.online_steps=0 \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=2 \
device=$(DEVICE) \
training.save_checkpoint=true \
training.save_freq=2 \
training.batch_size=2 \
training.image_transforms.enable=true \
hydra.run.dir=tests/outputs/tdmpc/
test-tdmpc-ete-train-with-online:
python lerobot/scripts/train.py \
env=pusht \
env.gym.obs_type=environment_state_agent_pos \
policy=tdmpc_pusht_keypoints \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=10 \
device=$(DEVICE) \
training.offline_steps=2 \
training.online_steps=20 \
training.save_checkpoint=false \
training.save_freq=10 \
training.batch_size=2 \
training.online_rollout_n_episodes=2 \
training.online_rollout_batch_size=2 \
training.online_steps_between_rollouts=10 \
training.online_buffer_capacity=15 \
eval.use_async_envs=true \
hydra.run.dir=tests/outputs/tdmpc_online/
--policy.type=tdmpc \
--policy.device=$(DEVICE) \
--env.type=xarm \
--env.task=XarmLift-v0 \
--env.episode_length=5 \
--dataset.repo_id=lerobot/xarm_lift_medium \
--dataset.image_transforms.enable=true \
--dataset.episodes="[0]" \
--batch_size=2 \
--steps=2 \
--eval_freq=2 \
--eval.n_episodes=1 \
--eval.batch_size=1 \
--save_checkpoint=true \
--save_freq=2 \
--log_freq=1 \
--wandb.enable=false \
--output_dir=tests/outputs/tdmpc/
test-tdmpc-ete-eval:
python lerobot/scripts/eval.py \
-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=8 \
device=$(DEVICE) \
test-default-ete-eval:
python lerobot/scripts/eval.py \
--config lerobot/configs/default.yaml \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=8 \
device=$(DEVICE) \
test-act-pusht-tutorial:
cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/created_by_Makefile.yaml
python lerobot/scripts/train.py \
policy=created_by_Makefile.yaml \
env=pusht \
wandb.enable=False \
training.offline_steps=2 \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=2 \
device=$(DEVICE) \
training.save_model=true \
training.save_freq=2 \
training.batch_size=2 \
training.image_transforms.enable=true \
hydra.run.dir=tests/outputs/act_pusht/
rm lerobot/configs/policy/created_by_Makefile.yaml
--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
--policy.device=$(DEVICE) \
--env.type=xarm \
--env.episode_length=5 \
--env.task=XarmLift-v0 \
--eval.n_episodes=1 \
--eval.batch_size=1

119
README.md
View File

@ -23,15 +23,24 @@
</div>
<h2 align="center">
<p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">New robot in town: SO-100</a></p>
<p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
Build Your Own SO-100 Robot!</a></p>
</h2>
<div align="center">
<img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
<p>We just added a new tutorial on how to build a more affordable robot, at the price of $110 per arm!</p>
<p>Teach it new skills by showing it a few moves with just a laptop.</p>
<p>Then watch your homemade robot act autonomously 🤯</p>
<p>Follow the link to the <a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">full tutorial for SO-100</a>.</p>
<p><strong>Meet the SO-100 – Just $110 per arm!</strong></p>
<p>Train it in minutes with a few simple moves on your laptop.</p>
<p>Then sit back and watch your creation act autonomously! 🤯</p>
<p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
Get the full SO-100 tutorial here.</a></p>
<p>Want to take it to the next level? Make your SO-100 mobile by building LeKiwi!</p>
<p>Check out the <a href="https://github.com/huggingface/lerobot/blob/main/examples/11_use_lekiwi.md">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>
<img src="media/lekiwi/kiwi.webp?raw=true" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
</div>
<br/>
@ -89,14 +98,25 @@ conda create -y -n lerobot python=3.10
conda activate lerobot
```
When using `miniconda`, install `ffmpeg` in your environment:
```bash
conda install ffmpeg -c conda-forge
```
> **NOTE:** This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
> - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
> ```bash
> conda install ffmpeg=7.1.1 -c conda-forge
> ```
> - _[On Linux only]_ Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
Install 🤗 LeRobot:
```bash
pip install -e .
```
> **NOTE:** Depending on your platform, If you encounter any build errors during this step
you may need to install `cmake` and `build-essential` for building some of our dependencies.
On linux: `sudo apt-get install cmake build-essential`
> **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
`sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
- [aloha](https://github.com/huggingface/gym-aloha)
@ -122,10 +142,7 @@ wandb login
├── examples # contains demonstration examples, start here to learn about LeRobot
| └── advanced # contains even more examples for those who have mastered the basics
├── lerobot
| ├── configs # contains hydra yaml files with all options that you can override in the command line
| | ├── default.yaml # selected by default, it loads pusht environment and diffusion policy
| | ├── env # various sim environments and their datasets: aloha.yaml, pusht.yaml, xarm.yaml
| | └── policy # various policies: act.yaml, diffusion.yaml, tdmpc.yaml
| ├── configs # contains config classes with all options that you can override in the command line
| ├── common # contains classes and utilities
| | ├── datasets # various datasets of human demonstrations: aloha, pusht, xarm
| | ├── envs # various sim environments: aloha, pusht, xarm
@ -213,7 +230,7 @@ A `LeRobotDataset` is serialised using several widespread file formats for each
- videos are stored in mp4 format to save space
- metadata are stored in plain json/jsonl files
Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can use the `local_files_only` argument and specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
### Evaluate a pretrained policy
@ -222,87 +239,48 @@ Check out [example 2](./examples/2_evaluate_pretrained_policy.py) that illustrat
We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
```bash
python lerobot/scripts/eval.py \
-p lerobot/diffusion_pusht \
eval.n_episodes=10 \
eval.batch_size=10
--policy.path=lerobot/diffusion_pusht \
--env.type=pusht \
--eval.batch_size=10 \
--eval.n_episodes=10 \
--policy.use_amp=false \
--policy.device=cuda
```
Note: After training your own policy, you can re-evaluate the checkpoints with:
```bash
python lerobot/scripts/eval.py -p {OUTPUT_DIR}/checkpoints/last/pretrained_model
python lerobot/scripts/eval.py --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
```
See `python lerobot/scripts/eval.py --help` for more instructions.
### Train your own policy
Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
Check out [example 3](./examples/3_train_policy.py) that illustrate how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
In general, you can use our training script to easily train any policy. Here is an example of training the ACT policy on trajectories collected by humans on the Aloha simulation environment for the insertion task:
To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.
```bash
python lerobot/scripts/train.py \
policy=act \
env=aloha \
env.task=AlohaInsertion-v0 \
dataset_repo_id=lerobot/aloha_sim_insertion_human \
```
The experiment directory is automatically generated and will show up in yellow in your terminal. It looks like `outputs/train/2024-05-05/20-21-12_aloha_act_default`. You can manually specify an experiment directory by adding this argument to the `train.py` python command:
```bash
hydra.run.dir=your/new/experiment/dir
```
In the experiment directory there will be a folder called `checkpoints` which will have the following structure:
```bash
checkpoints
├── 000250 # checkpoint_dir for training step 250
│ ├── pretrained_model # Hugging Face pretrained model dir
│ │ ├── config.json # Hugging Face pretrained model config
│ │ ├── config.yaml # consolidated Hydra config
│ │ ├── model.safetensors # model weights
│ │ └── README.md # Hugging Face model card
│ └── training_state.pth # optimizer/scheduler/rng state and training step
```
To resume training from a checkpoint, you can add these to the `train.py` python command:
```bash
hydra.run.dir=your/original/experiment/dir resume=true
```
It will load the pretrained model, optimizer and scheduler states for training. For more information please see our tutorial on training resumption [here](https://github.com/huggingface/lerobot/blob/main/examples/5_resume_training.md).
To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding:
```bash
wandb.enable=true
```
A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](./examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
![](media/wandb.png)
Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
#### Reproduce state-of-the-art (SOTA)
We have organized our configuration files (found under [`lerobot/configs`](./lerobot/configs)) such that they reproduce SOTA results from a given model variant in their respective original works. Simply running:
We provide some pretrained policies on our [hub page](https://huggingface.co/lerobot) that can achieve state-of-the-art performances.
You can reproduce their training by loading the config from their run. Simply running:
```bash
python lerobot/scripts/train.py policy=diffusion env=pusht
python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
```
reproduces SOTA results for Diffusion Policy on the PushT task.
Pretrained policies, along with reproduction details, can be found under the "Models" section of https://huggingface.co/lerobot.
## Contribute
If you would like to contribute to 🤗 LeRobot, please check out our [contribution guide](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md).
### Add a new dataset
<!-- ### Add a new dataset
To add a dataset to the hub, you need to login using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
```bash
@ -320,7 +298,7 @@ python lerobot/scripts/push_dataset_to_hub.py \
See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py).
If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py). -->
### Add a pretrained policy
@ -330,7 +308,7 @@ Once you have trained a policy you may upload it to the Hugging Face hub using a
You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
- `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
- `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
- `config.yaml`: A consolidated Hydra training configuration containing the policy, environment, and dataset configs. The policy configuration should match `config.json` exactly. The environment config is useful for anyone who wants to evaluate your policy. The dataset config just serves as a paper trail for reproducibility.
- `train_config.json`: A consolidated configuration containing all parameter userd for training. The policy configuration should match `config.json` exactly. Thisis useful for anyone who wants to evaluate your policy or for reproducibility.
To upload these to the hub, run the following:
```bash
@ -417,3 +395,6 @@ Additionally, if you are using any of the particular policy architecture, pretra
year={2024}
}
```
## Star History
[![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)

14
benchmarks/video/README.md
View File

@ -21,7 +21,7 @@ How to decode videos?
## Variables
**Image content & size**
We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an appartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
For these reasons, we run this benchmark on four representative datasets:
- `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
@ -51,7 +51,7 @@ For a comprehensive list and documentation of these parameters, see the ffmpeg d
### Decoding parameters
**Decoder**
We tested two video decoding backends from torchvision:
- `pyav` (default)
- `pyav`
- `video_reader` (requires to build torchvision from source)
**Requested timestamps**
@ -63,7 +63,7 @@ This of course is affected by the `-g` parameter during encoding, which specifie
Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
Additionally, because some policies might request single timestamps that are a few frames appart, we also have the following scenario:
Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
- `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
@ -85,8 +85,8 @@ However, due to how video decoding is implemented with `pyav`, we don't have acc
**Average Structural Similarity Index Measure (higher is better)**
`avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
One aspect that can't be measured here with those metrics is the compatibility of the encoding accross platforms, in particular on web browser, for visualization purposes.
h264, h265 and AV1 are all commonly used codecs and should not be pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
- `yuv420p` is more widely supported across various platforms, including web browsers.
- `yuv444p` offers higher color fidelity but might not be supported as broadly.
@ -114,9 +114,9 @@ We tried to measure the most impactful parameters for both encoding and decoding
Additional encoding parameters exist that are not included in this benchmark. In particular:
- `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
- `-tune` which allows to optimize the encoding for certains aspects (e.g. film quality, fast decoding, etc.).
- `-tune` which allows to optimize the encoding for certain aspects (e.g. film quality, fast decoding, etc.).
See the documentation mentioned above for more detailled info on these settings and for a more comprehensive list of other parameters.
See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
- `torchaudio`

32
benchmarks/video/capture_camera_feed.py
View File

@ -17,12 +17,21 @@
import argparse
import datetime as dt
import os
import time
from pathlib import Path
import cv2
import rerun as rr
# see https://rerun.io/docs/howto/visualization/limit-ram
RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
rr.init("lerobot_capture_camera_feed")
rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
now = dt.datetime.now()
capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
if not capture_dir.exists():
@ -39,24 +48,21 @@ def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
frame_index = 0
while True:
start_time = time.time()
while time.time() - start_time < duration:
ret, frame = cap.read()
if not ret:
print("Error: Could not read frame.")
break
cv2.imshow("Video Stream", frame)
rr.log("video/stream", rr.Image(frame.numpy()), static=True)
cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
frame_index += 1
# Break the loop on 'q' key press
if cv2.waitKey(1) & 0xFF == ord("q"):
break
# Release the capture and destroy all windows
# Release the capture
cap.release()
cv2.destroyAllWindows()
# TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.
if __name__ == "__main__":
@ -86,5 +92,11 @@ if __name__ == "__main__":
default=720,
help="Height of the captured images.",
)
parser.add_argument(
"--duration",
type=int,
default=20,
help="Duration in seconds for which the video stream should be captured.",
)
args = parser.parse_args()
display_and_save_video_stream(**vars(args))

2
benchmarks/video/run_video_benchmark.py
View File

@ -67,7 +67,7 @@ def parse_int_or_none(value) -> int | None:
def check_datasets_formats(repo_ids: list) -> None:
for repo_id in repo_ids:
dataset = LeRobotDataset(repo_id)
if dataset.video:
if len(dataset.meta.video_keys) > 0:
raise ValueError(
f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
)

37
docker/lerobot-cpu/Dockerfile
View File

@ -1,32 +1,29 @@
# Configure image
ARG PYTHON_VERSION=3.10
FROM python:${PYTHON_VERSION}-slim
# Configure environment variables
ARG PYTHON_VERSION
ARG DEBIAN_FRONTEND=noninteractive
# Install apt dependencies
RUN apt-get update && apt-get install -y --no-install-recommends \
build-essential cmake \
libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
speech-dispatcher \
&& apt-get clean && rm -rf /var/lib/apt/lists/*
# Create virtual environment
RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
RUN python -m venv /opt/venv
ENV DEBIAN_FRONTEND=noninteractive
ENV MUJOCO_GL="egl"
ENV PATH="/opt/venv/bin:$PATH"
RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
# Install LeRobot
# Install dependencies and set up Python in a single layer
RUN apt-get update && apt-get install -y --no-install-recommends \
build-essential cmake git \
libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
speech-dispatcher libgeos-dev \
&& ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python \
&& python -m venv /opt/venv \
&& apt-get clean && rm -rf /var/lib/apt/lists/* \
&& echo "source /opt/venv/bin/activate" >> /root/.bashrc
# Clone repository and install LeRobot in a single layer
COPY . /lerobot
WORKDIR /lerobot
RUN pip install --upgrade --no-cache-dir pip
RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
RUN /opt/venv/bin/pip install --upgrade --no-cache-dir pip \
&& /opt/venv/bin/pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
--extra-index-url https://download.pytorch.org/whl/cpu
# Set EGL as the rendering backend for MuJoCo
ENV MUJOCO_GL="egl"
# Execute in bash shell rather than python
CMD ["/bin/bash"]

2
docker/lerobot-gpu-dev/Dockerfile
View File

@ -13,7 +13,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
sed gawk grep curl wget zip unzip \
tcpdump sysstat screen tmux \
libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
speech-dispatcher \
speech-dispatcher portaudio19-dev libgeos-dev \
python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
&& apt-get clean && rm -rf /var/lib/apt/lists/*

40
docker/lerobot-gpu/Dockerfile
View File

@ -1,30 +1,24 @@
FROM nvidia/cuda:12.4.1-base-ubuntu22.04
# Configure image
# Configure environment variables
ARG PYTHON_VERSION=3.10
ARG DEBIAN_FRONTEND=noninteractive
# Install apt dependencies
RUN apt-get update && apt-get install -y --no-install-recommends \
build-essential cmake \
libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
speech-dispatcher \
python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
&& apt-get clean && rm -rf /var/lib/apt/lists/*
# Create virtual environment
RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
RUN python -m venv /opt/venv
ENV DEBIAN_FRONTEND=noninteractive
ENV MUJOCO_GL="egl"
ENV PATH="/opt/venv/bin:$PATH"
RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
# Install LeRobot
# Install dependencies and set up Python in a single layer
RUN apt-get update && apt-get install -y --no-install-recommends \
build-essential cmake git \
libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
speech-dispatcher libgeos-dev \
python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
&& ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python \
&& python -m venv /opt/venv \
&& apt-get clean && rm -rf /var/lib/apt/lists/* \
&& echo "source /opt/venv/bin/activate" >> /root/.bashrc
# Clone repository and install LeRobot in a single layer
COPY . /lerobot
WORKDIR /lerobot
RUN pip install --upgrade --no-cache-dir pip
RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"
# Set EGL as the rendering backend for MuJoCo
ENV MUJOCO_GL="egl"
RUN /opt/venv/bin/pip install --upgrade --no-cache-dir pip \
&& /opt/venv/bin/pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"

573
examples/10_use_so100.md
View File

@ -1,90 +1,190 @@
This tutorial explains how to use [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot.
# Using the [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot
## Source the parts
## Table of Contents
Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with link to source the parts, as well as the instructions to 3D print the parts, and advices if it's your first time printing or if you don't own a 3D printer already.
- [A. Source the parts](#a-source-the-parts)
- [B. Install LeRobot](#b-install-lerobot)
- [C. Configure the Motors](#c-configure-the-motors)
- [D. Step-by-Step Assembly Instructions](#d-step-by-step-assembly-instructions)
- [E. Calibrate](#e-calibrate)
- [F. Teleoperate](#f-teleoperate)
- [G. Record a dataset](#g-record-a-dataset)
- [H. Visualize a dataset](#h-visualize-a-dataset)
- [I. Replay an episode](#i-replay-an-episode)
- [J. Train a policy](#j-train-a-policy)
- [K. Evaluate your policy](#k-evaluate-your-policy)
- [L. More Information](#l-more-information)
**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
## A. Source the parts
## Install LeRobot
Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
and advice if it's your first time printing or if you don't own a 3D printer.
Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
## B. Install LeRobot
> [!TIP]
> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
On your computer:
1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
#### 2. Restart shell
Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
#### 3. Create and activate a fresh conda environment for lerobot
<details>
<summary><strong>Video install instructions</strong></summary>
<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
</details>
```bash
mkdir -p ~/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
rm ~/miniconda3/miniconda.sh
~/miniconda3/bin/conda init bash
conda create -y -n lerobot python=3.10
```
2. Restart shell or `source ~/.bashrc`
3. Create and activate a fresh conda environment for lerobot
Then activate your conda environment (do this each time you open a shell to use lerobot!):
```bash
conda create -y -n lerobot python=3.10 && conda activate lerobot
conda activate lerobot
```
4. Clone LeRobot:
#### 4. Clone LeRobot:
```bash
git clone https://github.com/huggingface/lerobot.git ~/lerobot
```
5. Install LeRobot with dependencies for the feetech motors:
#### 5. Install ffmpeg in your environment:
When using `miniconda`, install `ffmpeg` in your environment:
```bash
conda install ffmpeg -c conda-forge
```
#### 6. Install LeRobot with dependencies for the feetech motors:
```bash
cd ~/lerobot && pip install -e ".[feetech]"
```
For Linux only (not Mac), install extra dependencies for recording datasets:
```bash
conda install -y -c conda-forge ffmpeg
pip uninstall -y opencv-python
conda install -y -c conda-forge "opencv>=4.10.0"
```
Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms :robot:.
Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
## Configure the motors
## C. Configure the motors
Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the use of our scripts below.
> [!NOTE]
> Throughout this tutorial you will find videos on how to do the steps, the full video tutorial can be found here: [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I).
**Find USB ports associated to your arms**
To find the correct ports for each arm, run the utility script twice:
### 1. Find the USB ports associated to each arm
Designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6 (F1...F6 and L1...L6).
#### a. Run the script to find port
<details>
<summary><strong>Video finding port</strong></summary>
<video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
<video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
</details>
To find the port for each bus servo adapter, run the utility script:
```bash
python lerobot/scripts/find_motors_bus_port.py
```
#### b. Example outputs
Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
Remove the usb cable from your MotorsBus and press Enter when done.
[...Disconnect leader arm and press Enter...]
The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
The port of this MotorsBus is /dev/tty.usbmodem575E0031751
Reconnect the usb cable.
```
Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
Remove the usb cable from your MotorsBus and press Enter when done.
[...Disconnect follower arm and press Enter...]
The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
The port of this MotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the usb cable.
```
Troubleshooting: On Linux, you might need to give access to the USB ports by running:
#### c. Troubleshooting
On Linux, you might need to give access to the USB ports by running:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
**Configure your motors**
Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
#### d. Update config file
IMPORTANTLY: Now that you have your ports, update the **port** default values of [`SO100RobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
```python
@RobotConfig.register_subclass("so100")
@dataclass
class So100RobotConfig(ManipulatorRobotConfig):
calibration_dir: str = ".cache/calibration/so100"
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: int | None = None
leader_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem58760431091", <-- UPDATE HERE
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
},
),
}
)
follower_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem585A0076891", <-- UPDATE HERE
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
},
),
}
)
```
### 2. Assembling the Base
Let's begin with assembling the follower arm base
#### a. Set IDs for all 12 motors
<details>
<summary><strong>Video configuring motor</strong></summary>
<video src="https://github.com/user-attachments/assets/ef9b3317-2e11-4858-b9d3-f0a02fb48ecf"></video>
<video src="https://github.com/user-attachments/assets/f36b5ed5-c803-4ebe-8947-b39278776a0d"></video>
</details>
Plug your first motor F1 and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate. Replace the text after --port to the corresponding follower control board port and run this command in cmd:
```bash
python lerobot/scripts/configure_motor.py \
--port /dev/tty.usbmodem58760432961 \
@ -94,7 +194,8 @@ python lerobot/scripts/configure_motor.py \
--ID 1
```
Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
> [!NOTE]
> These motors are currently limited. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
Then unplug your motor and plug the second motor and set its ID to 2.
```bash
@ -108,71 +209,304 @@ python lerobot/scripts/configure_motor.py \
Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
**Remove the gears of the 6 leader motors**
Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
**Add motor horn to the motors**
Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). For SO-100, you need to align the holes on the motor horn to the motor spline to be approximately 1:30, 4:30, 7:30 and 10:30.
Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
#### b. Remove the gears of the 6 leader motors
## Assemble the arms
<details>
<summary><strong>Video removing gears</strong></summary>
Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
<video src="https://github.com/user-attachments/assets/0c95b88c-5b85-413d-ba19-aee2f864f2a7"></video>
## Calibrate
</details>
Follow the video for removing gears. You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
## D. Step-by-Step Assembly Instructions
**Step 1: Clean Parts**
- Remove all support material from the 3D-printed parts.
---
### Additional Guidance
<details>
<summary><strong>Video assembling arms</strong></summary>
<video src="https://github.com/user-attachments/assets/488a39de-0189-4461-9de3-05b015f90cca"></video>
</details>
**Note:**
This video provides visual guidance for assembling the arms, but it doesn't specify when or how to do the wiring. Inserting the cables beforehand is much easier than doing it afterward. The first arm may take a bit more than 1 hour to assemble, but once you get used to it, you can assemble the second arm in under 1 hour.
---
### First Motor
**Step 2: Insert Wires**
- Insert two wires into the first motor.
<img src="../media/tutorial/img1.jpg" style="height:300px;">
**Step 3: Install in Base**
- Place the first motor into the base.
<img src="../media/tutorial/img2.jpg" style="height:300px;">
**Step 4: Secure Motor**
- Fasten the motor with 4 screws. Two from the bottom and two from top.
**Step 5: Attach Motor Holder**
- Slide over the first motor holder and fasten it using two screws (one on each side).
<img src="../media/tutorial/img4.jpg" style="height:300px;">
**Step 6: Attach Motor Horns**
- Install both motor horns, securing the top horn with a screw. Try not to move the motor position when attaching the motor horn, especially for the leader arms, where we removed the gears.
<img src="../media/tutorial/img5.jpg" style="height:300px;">
<details>
<summary><strong>Video adding motor horn</strong></summary>
<video src="https://github.com/user-attachments/assets/ef3391a4-ad05-4100-b2bd-1699bf86c969"></video>
</details>
**Step 7: Attach Shoulder Part**
- Route one wire to the back of the robot and the other to the left or in photo towards you (see photo).
- Attach the shoulder part.
<img src="../media/tutorial/img6.jpg" style="height:300px;">
**Step 8: Secure Shoulder**
- Tighten the shoulder part with 4 screws on top and 4 on the bottom
*(access bottom holes by turning the shoulder).*
---
### Second Motor Assembly
**Step 9: Install Motor 2**
- Slide the second motor in from the top and link the wire from motor 1 to motor 2.
<img src="../media/tutorial/img8.jpg" style="height:300px;">
**Step 10: Attach Shoulder Holder**
- Add the shoulder motor holder.
- Ensure the wire from motor 1 to motor 2 goes behind the holder while the other wire is routed upward (see photo).
- This part can be tight to assemble, you can use a workbench like the image or a similar setup to push the part around the motor.
<div style="display: flex;">
<img src="../media/tutorial/img9.jpg" style="height:250px;">
<img src="../media/tutorial/img10.jpg" style="height:250px;">
<img src="../media/tutorial/img12.jpg" style="height:250px;">
</div>
**Step 11: Secure Motor 2**
- Fasten the second motor with 4 screws.
**Step 12: Attach Motor Horn**
- Attach both motor horns to motor 2, again use the horn screw.
**Step 13: Attach Base**
- Install the base attachment using 2 screws.
<img src="../media/tutorial/img11.jpg" style="height:300px;">
**Step 14: Attach Upper Arm**
- Attach the upper arm with 4 screws on each side.
<img src="../media/tutorial/img13.jpg" style="height:300px;">
---
### Third Motor Assembly
**Step 15: Install Motor 3**
- Route the motor cable from motor 2 through the cable holder to motor 3, then secure motor 3 with 4 screws.
**Step 16: Attach Motor Horn**
- Attach both motor horns to motor 3 and secure one again with a horn screw.
<img src="../media/tutorial/img14.jpg" style="height:300px;">
**Step 17: Attach Forearm**
- Connect the forearm to motor 3 using 4 screws on each side.
<img src="../media/tutorial/img15.jpg" style="height:300px;">
---
### Fourth Motor Assembly
**Step 18: Install Motor 4**
- Slide in motor 4, attach the cable from motor 3, and secure the cable in its holder with a screw.
<div style="display: flex;">
<img src="../media/tutorial/img16.jpg" style="height:300px;">
<img src="../media/tutorial/img19.jpg" style="height:300px;">
</div>
**Step 19: Attach Motor Holder 4**
- Install the fourth motor holder (a tight fit). Ensure one wire is routed upward and the wire from motor 3 is routed downward (see photo).
<img src="../media/tutorial/img17.jpg" style="height:300px;">
**Step 20: Secure Motor 4 & Attach Horn**
- Fasten motor 4 with 4 screws and attach its motor horns, use for one a horn screw.
<img src="../media/tutorial/img18.jpg" style="height:300px;">
---
### Wrist Assembly
**Step 21: Install Motor 5**
- Insert motor 5 into the wrist holder and secure it with 2 front screws.
<img src="../media/tutorial/img20.jpg" style="height:300px;">
**Step 22: Attach Wrist**
- Connect the wire from motor 4 to motor 5. And already insert the other wire for the gripper.
- Secure the wrist to motor 4 using 4 screws on both sides.
<img src="../media/tutorial/img22.jpg" style="height:300px;">
**Step 23: Attach Wrist Horn**
- Install only one motor horn on the wrist motor and secure it with a horn screw.
<img src="../media/tutorial/img23.jpg" style="height:300px;">
---
### Follower Configuration
**Step 24: Attach Gripper**
- Attach the gripper to motor 5.
<img src="../media/tutorial/img24.jpg" style="height:300px;">
**Step 25: Install Gripper Motor**
- Insert the gripper motor, connect the motor wire from motor 5 to motor 6, and secure it with 3 screws on each side.
<img src="../media/tutorial/img25.jpg" style="height:300px;">
**Step 26: Attach Gripper Horn & Claw**
- Attach the motor horns and again use a horn screw.
- Install the gripper claw and secure it with 4 screws on both sides.
<img src="../media/tutorial/img26.jpg" style="height:300px;">
**Step 27: Mount Controller**
- Attach the motor controller on the back.
<div style="display: flex;">
<img src="../media/tutorial/img27.jpg" style="height:300px;">
<img src="../media/tutorial/img28.jpg" style="height:300px;">
</div>
*Assembly complete – proceed to Leader arm assembly.*
---
### Leader Configuration
For the leader configuration, perform **Steps 1–23**. Make sure that you removed the motor gears from the motors.
**Step 24: Attach Leader Holder**
- Mount the leader holder onto the wrist and secure it with a screw.
<img src="../media/tutorial/img29.jpg" style="height:300px;">
**Step 25: Attach Handle**
- Attach the handle to motor 5 using 4 screws.
<img src="../media/tutorial/img30.jpg" style="height:300px;">
**Step 26: Install Gripper Motor**
- Insert the gripper motor, secure it with 3 screws on each side, attach a motor horn using a horn screw, and connect the motor wire.
<img src="../media/tutorial/img31.jpg" style="height:300px;">
**Step 27: Attach Trigger**
- Attach the follower trigger with 4 screws.
<img src="../media/tutorial/img32.jpg" style="height:300px;">
**Step 28: Mount Controller**
- Attach the motor controller on the back.
<div style="display: flex;">
<img src="../media/tutorial/img27.jpg" style="height:300px;">
<img src="../media/tutorial/img28.jpg" style="height:300px;">
</div>
*Assembly complete – proceed to calibration.*
## E. Calibrate
Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
**Manual calibration of follower arm**
/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
#### a. Manual calibration of follower arm
> [!IMPORTANT]
> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
You will need to move the follower arm to these positions sequentially:
| 1. Zero position | 2. Rotated position | 3. Rest position |
|---|---|---|
| ------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| <img src="../media/so100/follower_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/so100/follower_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/so100/follower_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
Make sure both arms are connected and run this script to launch manual calibration:
```bash
python lerobot/scripts/control_robot.py calibrate \
--robot-path lerobot/configs/robot/so100.yaml \
--robot-overrides '~cameras' --arms main_follower
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--robot.cameras='{}' \
--control.type=calibrate \
--control.arms='["main_follower"]'
```
**Manual calibration of leader arm**
Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
#### b. Manual calibration of leader arm
Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
| 1. Zero position | 2. Rotated position | 3. Rest position |
|---|---|---|
| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
Run this script to launch manual calibration:
```bash
python lerobot/scripts/control_robot.py calibrate \
--robot-path lerobot/configs/robot/so100.yaml \
--robot-overrides '~cameras' --arms main_leader
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--robot.cameras='{}' \
--control.type=calibrate \
--control.arms='["main_leader"]'
```
## Teleoperate
## F. Teleoperate
**Simple teleop**
Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/so100.yaml \
--robot-overrides '~cameras' \
--display-cameras 0
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--robot.cameras='{}' \
--control.type=teleoperate
```
**Teleop with displaying cameras**
#### a. Teleop with displaying cameras
Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/so100.yaml
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--control.type=teleoperate
```
## Record a dataset
## G. Record a dataset
Once you're familiar with teleoperation, you can record your first dataset with SO-100.
@ -189,87 +523,102 @@ echo $HF_USER
Record 2 episodes and upload your dataset to the hub:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/so100.yaml \
--fps 30 \
--repo-id ${HF_USER}/so100_test \
--tags so100 tutorial \
--warmup-time-s 5 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 2 \
--push-to-hub 1
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/so100_test \
--control.tags='["so100","tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=2 \
--control.push_to_hub=true
```
## Visualize a dataset
Note: You can resume recording by adding `--control.resume=true`.
If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
## H. Visualize a dataset
If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
```bash
echo ${HF_USER}/so100_test
```
If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
```bash
python lerobot/scripts/visualize_dataset_html.py \
--repo-id ${HF_USER}/so100_test
--repo-id ${HF_USER}/so100_test \
--local-files-only 1
```
## Replay an episode
## I. Replay an episode
Now try to replay the first episode on your robot:
```bash
python lerobot/scripts/control_robot.py replay \
--robot-path lerobot/configs/robot/so100.yaml \
--fps 30 \
--repo-id ${HF_USER}/so100_test \
--episode 0
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--control.type=replay \
--control.fps=30 \
--control.repo_id=${HF_USER}/so100_test \
--control.episode=0
```
## Train a policy
## J. Train a policy
To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
```bash
python lerobot/scripts/train.py \
dataset_repo_id=${HF_USER}/so100_test \
policy=act_so100_real \
env=so100_real \
hydra.run.dir=outputs/train/act_so100_test \
hydra.job.name=act_so100_test \
device=cuda \
wandb.enable=true
--dataset.repo_id=${HF_USER}/so100_test \
--policy.type=act \
--output_dir=outputs/train/act_so100_test \
--job_name=act_so100_test \
--policy.device=cuda \
--wandb.enable=true
```
Let's explain it:
1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/so100_test`.
2. We provided the policy with `policy=act_so100_real`. This loads configurations from [`lerobot/configs/policy/act_so100_real.yaml`](../lerobot/configs/policy/act_so100_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
3. We provided an environment as argument with `env=so100_real`. This loads configurations from [`lerobot/configs/env/so100_real.yaml`](../lerobot/configs/env/so100_real.yaml).
4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so100_test`.
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
Training should take several hours. You will find checkpoints in `outputs/train/act_so100_test/checkpoints`.
## Evaluate your policy
To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so100_test` policy:
```bash
python lerobot/scripts/train.py \
--config_path=outputs/train/act_so100_test/checkpoints/last/pretrained_model/train_config.json \
--resume=true
```
## K. Evaluate your policy
You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/so100.yaml \
--fps 30 \
--repo-id ${HF_USER}/eval_act_so100_test \
--tags so100 tutorial eval \
--warmup-time-s 5 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 10 \
-p outputs/train/act_so100_test/checkpoints/last/pretrained_model
python lerobot/scripts/control_robot.py \
--robot.type=so100 \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/eval_act_so100_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=10 \
--control.push_to_hub=true \
--control.policy.path=outputs/train/act_so100_test/checkpoints/last/pretrained_model
```
As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
1. There is an additional `-p` argument which indicates the path to your policy checkpoint with (e.g. `-p outputs/train/eval_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_so100_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_so100_test`).
1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_act_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so100_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so100_test`).
## More
## L. More Information
Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
If you have any question or need help, please reach out on Discord in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).

597
examples/11_use_lekiwi.md Normal file
View File

@ -0,0 +1,597 @@
# Using the [LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi) Robot with LeRobot
## Table of Contents
- [A. Source the parts](#a-source-the-parts)
- [B. Install software Pi](#b-install-software-on-pi)
- [C. Setup LeRobot laptop/pc](#c-install-lerobot-on-laptop)
- [D. Assemble the arms](#d-assembly)
- [E. Calibrate](#e-calibration)
- [F. Teleoperate](#f-teleoperate)
- [G. Record a dataset](#g-record-a-dataset)
- [H. Visualize a dataset](#h-visualize-a-dataset)
- [I. Replay an episode](#i-replay-an-episode)
- [J. Train a policy](#j-train-a-policy)
- [K. Evaluate your policy](#k-evaluate-your-policy)
> [!TIP]
> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#mobile-so-100-arm`](https://discord.com/channels/1216765309076115607/1318390825528332371).
## A. Source the parts
Follow this [README](https://github.com/SIGRobotics-UIUC/LeKiwi). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts, and advice if it's your first time printing or if you don't own a 3D printer.
Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
### Wired version
If you have the **wired** LeKiwi version you can skip the installation of the Raspberry Pi and setting up SSH. You can also run all commands directly on your PC for both the LeKiwi scripts and the leader arm scripts for teleoperating.
## B. Install software on Pi
Now we have to setup the remote PC that will run on the LeKiwi Robot. This is normally a Raspberry Pi, but can be any PC that can run on 5V and has enough usb ports (2 or more) for the cameras and motor control board.
### Install OS
For setting up the Raspberry Pi and its SD-card see: [Setup PI](https://www.raspberrypi.com/documentation/computers/getting-started.html). Here is explained how to download the [Imager](https://www.raspberrypi.com/software/) to install Raspberry Pi OS or Ubuntu.
### Setup SSH
After setting up your Pi, you should enable and setup [SSH](https://www.raspberrypi.com/news/coding-on-raspberry-pi-remotely-with-visual-studio-code/) (Secure Shell Protocol) so you can login into the Pi from your laptop without requiring a screen, keyboard and mouse in the Pi. A great tutorial on how to do this can be found [here](https://www.raspberrypi.com/documentation/computers/remote-access.html#ssh). Logging into your Pi can be done in your Command Prompt (cmd) or if you use VSCode you can use [this](https://marketplace.visualstudio.com/items?itemName=ms-vscode-remote.remote-ssh) extension.
### Install LeRobot
On your Raspberry Pi:
#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
#### 2. Restart shell
Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
#### 3. Create and activate a fresh conda environment for lerobot
<details>
<summary><strong>Video install instructions</strong></summary>
<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
</details>
```bash
conda create -y -n lerobot python=3.10
```
Then activate your conda environment (do this each time you open a shell to use lerobot!):
```bash
conda activate lerobot
```
#### 4. Clone LeRobot:
```bash
git clone https://github.com/huggingface/lerobot.git ~/lerobot
```
#### 5. Install ffmpeg in your environment:
When using `miniconda`, install `ffmpeg` in your environment:
```bash
conda install ffmpeg -c conda-forge
```
#### 6. Install LeRobot with dependencies for the feetech motors:
```bash
cd ~/lerobot && pip install -e ".[feetech]"
```
## C. Install LeRobot on laptop
If you already have install LeRobot on your laptop you can skip this step, otherwise please follow along as we do the same steps we did on the Pi.
> [!TIP]
> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
On your computer:
#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
#### 2. Restart shell
Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
#### 3. Create and activate a fresh conda environment for lerobot
<details>
<summary><strong>Video install instructions</strong></summary>
<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
</details>
```bash
conda create -y -n lerobot python=3.10
```
Then activate your conda environment (do this each time you open a shell to use lerobot!):
```bash
conda activate lerobot
```
#### 4. Clone LeRobot:
```bash
git clone https://github.com/huggingface/lerobot.git ~/lerobot
```
#### 5. Install ffmpeg in your environment:
When using `miniconda`, install `ffmpeg` in your environment:
```bash
conda install ffmpeg -c conda-forge
```
#### 6. Install LeRobot with dependencies for the feetech motors:
```bash
cd ~/lerobot && pip install -e ".[feetech]"
```
Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms and Mobile base :robot:.
Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
# D. Assembly
First we will assemble the two SO100 arms. One to attach to the mobile base and one for teleoperation. Then we will assemble the mobile base.
## SO100 Arms
### Configure motors
The instructions for configuring the motors can be found [Here](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#c-configure-the-motors) in step C of the SO100 tutorial. Besides the ID's for the arm motors we also need to set the motor ID's for the mobile base. These needs to be in a specific order to work. Below an image of the motor ID's and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ID's for the wheels are 7, 8 and 9.
<img src="../media/lekiwi/motor_ids.webp?raw=true" alt="Motor ID's for mobile robot" title="Motor ID's for mobile robot" width="60%">
### Assemble arms
[Assemble arms instruction](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#d-assemble-the-arms)
## Mobile base (LeKiwi)
[Assemble LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi)
### Update config
Both config files on the LeKiwi LeRobot and on the laptop should be the same. First we should find the Ip address of the Raspberry Pi of the mobile manipulator. This is the same Ip address used in SSH. We also need the usb port of the control board of the leader arm on the laptop and the port of the control board on LeKiwi. We can find these ports with the following script.
#### a. Run the script to find port
<details>
<summary><strong>Video finding port</strong></summary>
<video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
<video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
</details>
To find the port for each bus servo adapter, run the utility script:
```bash
python lerobot/scripts/find_motors_bus_port.py
```
#### b. Example outputs
Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
[...Disconnect leader arm and press Enter...]
The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
Reconnect the usb cable.
```
Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
```
Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
[...Disconnect follower arm and press Enter...]
The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the usb cable.
```
#### c. Troubleshooting
On Linux, you might need to give access to the USB ports by running:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
#### d. Update config file
IMPORTANTLY: Now that you have your ports of leader and follower arm and ip address of the mobile-so100, update the **ip** in Network configuration, **port** in leader_arms and **port** in lekiwi. In the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py) file. Where you will find something like:
```python
@RobotConfig.register_subclass("lekiwi")
@dataclass
class LeKiwiRobotConfig(RobotConfig):
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: int | None = None
# Network Configuration
ip: str = "172.17.133.91"
port: int = 5555
video_port: int = 5556
cameras: dict[str, CameraConfig] = field(
default_factory=lambda: {
"mobile": OpenCVCameraConfig(camera_index="/dev/video0", fps=30, width=640, height=480),
"mobile2": OpenCVCameraConfig(camera_index="/dev/video2", fps=30, width=640, height=480),
}
)
calibration_dir: str = ".cache/calibration/lekiwi"
leader_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem585A0077581",
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
},
),
}
)
follower_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/ttyACM0",
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
"left_wheel": (7, "sts3215"),
"back_wheel": (8, "sts3215"),
"right_wheel": (9, "sts3215"),
},
),
}
)
teleop_keys: dict[str, str] = field(
default_factory=lambda: {
# Movement
"forward": "w",
"backward": "s",
"left": "a",
"right": "d",
"rotate_left": "z",
"rotate_right": "x",
# Speed control
"speed_up": "r",
"speed_down": "f",
# quit teleop
"quit": "q",
}
)
mock: bool = False
```
## Wired version
For the wired LeKiwi version your configured IP address should refer to your own laptop (127.0.0.1), because leader arm and LeKiwi are in this case connected to own laptop. Below and example configuration for this wired setup:
```python
@RobotConfig.register_subclass("lekiwi")
@dataclass
class LeKiwiRobotConfig(RobotConfig):
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: int | None = None
# Network Configuration
ip: str = "127.0.0.1"
port: int = 5555
video_port: int = 5556
cameras: dict[str, CameraConfig] = field(
default_factory=lambda: {
"front": OpenCVCameraConfig(
camera_index=0, fps=30, width=640, height=480, rotation=90
),
"wrist": OpenCVCameraConfig(
camera_index=1, fps=30, width=640, height=480, rotation=180
),
}
)
calibration_dir: str = ".cache/calibration/lekiwi"
leader_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem585A0077581",
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
},
),
}
)
follower_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem58760431061",
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
"left_wheel": (7, "sts3215"),
"back_wheel": (8, "sts3215"),
"right_wheel": (9, "sts3215"),
},
),
}
)
teleop_keys: dict[str, str] = field(
default_factory=lambda: {
# Movement
"forward": "w",
"backward": "s",
"left": "a",
"right": "d",
"rotate_left": "z",
"rotate_right": "x",
# Speed control
"speed_up": "r",
"speed_down": "f",
# quit teleop
"quit": "q",
}
)
mock: bool = False
```
# E. Calibration
Now we have to calibrate the leader arm and the follower arm. The wheel motors don't have to be calibrated.
### Calibrate follower arm (on mobile base)
> [!IMPORTANT]
> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
You will need to move the follower arm to these positions sequentially:
| 1. Zero position | 2. Rotated position | 3. Rest position |
| ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| <img src="../media/lekiwi/mobile_calib_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
Make sure the arm is connected to the Raspberry Pi and run this script (on the Raspberry Pi) to launch manual calibration:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--robot.cameras='{}' \
--control.type=calibrate \
--control.arms='["main_follower"]'
```
### Wired version
If you have the **wired** LeKiwi version please run all commands including this calibration command on your laptop.
### Calibrate leader arm
Then to calibrate the leader arm (which is attached to the laptop/pc). You will need to move the leader arm to these positions sequentially:
| 1. Zero position | 2. Rotated position | 3. Rest position |
| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
Run this script (on your laptop/pc) to launch manual calibration:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--robot.cameras='{}' \
--control.type=calibrate \
--control.arms='["main_leader"]'
```
# F. Teleoperate
> [!TIP]
> If you're using a Mac, you might need to give Terminal permission to access your keyboard. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
To teleoperate SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--control.type=remote_robot
```
Then on your laptop, also run `conda activate lerobot` and this script:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--control.type=teleoperate \
--control.fps=30
```
> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`. For the `--control.type=remote_robot` you will also need to set `--control.viewer_ip` and `--control.viewer_port`
You should see on your laptop something like this: ```[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.``` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
| Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
| ---------- | ------------------ | ---------------------- |
| Fast | 0.4 | 90 |
| Medium | 0.25 | 60 |
| Slow | 0.1 | 30 |
| Key | Action |
| --- | -------------- |
| W | Move forward |
| A | Move left |
| S | Move backward |
| D | Move right |
| Z | Turn left |
| X | Turn right |
| R | Increase speed |
| F | Decrease speed |
> [!TIP]
> If you use a different keyboard you can change the keys for each command in the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py).
### Wired version
If you have the **wired** LeKiwi version please run all commands including both these teleoperation commands on your laptop.
## Troubleshoot communication
If you are having trouble connecting to the Mobile SO100, follow these steps to diagnose and resolve the issue.
### 1. Verify IP Address Configuration
Make sure that the correct ip for the Pi is set in the configuration file. To check the Raspberry Pi's IP address, run (on the Pi command line):
```bash
hostname -I
```
### 2. Check if Pi is reachable from laptop/pc
Try pinging the Raspberry Pi from your laptop:
```bach
ping <your_pi_ip_address>
```
If the ping fails:
- Ensure the Pi is powered on and connected to the same network.
- Check if SSH is enabled on the Pi.
### 3. Try SSH connection
If you can't SSH into the Pi, it might not be properly connected. Use:
```bash
ssh <your_pi_user_name>@<your_pi_ip_address>
```
If you get a connection error:
- Ensure SSH is enabled on the Pi by running:
```bash
sudo raspi-config
```
Then navigate to: **Interfacing Options -> SSH** and enable it.
### 4. Same config file
Make sure the configuration file on both your laptop/pc and the Raspberry Pi is the same.
# G. Record a dataset
Once you're familiar with teleoperation, you can record your first dataset with LeKiwi.
To start the program on LeKiwi, SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--control.type=remote_robot
```
If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
```bash
huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
```
Store your Hugging Face repository name in a variable to run these commands:
```bash
HF_USER=$(huggingface-cli whoami | head -n 1)
echo $HF_USER
```
On your laptop then run this command to record 2 episodes and upload your dataset to the hub:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/lekiwi_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=2 \
--control.push_to_hub=true
```
Note: You can resume recording by adding `--control.resume=true`.
### Wired version
If you have the **wired** LeKiwi version please run all commands including both these record dataset commands on your laptop.
# H. Visualize a dataset
If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
```bash
echo ${HF_USER}/lekiwi_test
```
If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
```bash
python lerobot/scripts/visualize_dataset_html.py \
--repo-id ${HF_USER}/lekiwi_test \
--local-files-only 1
```
# I. Replay an episode
Now try to replay the first episode on your robot:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--control.type=replay \
--control.fps=30 \
--control.repo_id=${HF_USER}/lekiwi_test \
--control.episode=0
```
## J. Train a policy
To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
```bash
python lerobot/scripts/train.py \
--dataset.repo_id=${HF_USER}/lekiwi_test \
--policy.type=act \
--output_dir=outputs/train/act_lekiwi_test \
--job_name=act_lekiwi_test \
--policy.device=cuda \
--wandb.enable=true
```
Let's explain it:
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/lekiwi_test`.
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
Training should take several hours. You will find checkpoints in `outputs/train/act_lekiwi_test/checkpoints`.
## K. Evaluate your policy
You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
```bash
python lerobot/scripts/control_robot.py \
--robot.type=lekiwi \
--control.type=record \
--control.fps=30 \
--control.single_task="Drive to the red block and pick it up" \
--control.repo_id=${HF_USER}/eval_act_lekiwi_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=10 \
--control.push_to_hub=true \
--control.policy.path=outputs/train/act_lekiwi_test/checkpoints/last/pretrained_model
```
As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_act_lekiwi_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_lekiwi_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_lekiwi_test`).

186
examples/11_use_moss.md
View File

@ -2,7 +2,7 @@ This tutorial explains how to use [Moss v1](https://github.com/jess-moss/moss-ro
## Source the parts
Follow this [README](https://github.com/jess-moss/moss-robot-arms). It contains the bill of materials, with link to source the parts, as well as the instructions to 3D print the parts, and advices if it's your first time printing or if you don't own a 3D printer already.
Follow this [README](https://github.com/jess-moss/moss-robot-arms). It contains the bill of materials with link to source the parts, as well as the instructions to 3D print the parts and advice if it's your first time printing or if you don't own a 3D printer already.
**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
@ -31,16 +31,15 @@ conda create -y -n lerobot python=3.10 && conda activate lerobot
git clone https://github.com/huggingface/lerobot.git ~/lerobot
```
5. Install LeRobot with dependencies for the feetech motors:
5. Install ffmpeg in your environment:
When using `miniconda`, install `ffmpeg` in your environment:
```bash
cd ~/lerobot && pip install -e ".[feetech]"
conda install ffmpeg -c conda-forge
```
For Linux only (not Mac), install extra dependencies for recording datasets:
6. Install LeRobot with dependencies for the feetech motors:
```bash
conda install -y -c conda-forge ffmpeg
pip uninstall -y opencv-python
conda install -y -c conda-forge "opencv>=4.10.0"
cd ~/lerobot && pip install -e ".[feetech]"
```
## Configure the motors
@ -83,6 +82,54 @@ sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
```
#### Update config file
IMPORTANTLY: Now that you have your ports, update the **port** default values of [`MossRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
```python
@RobotConfig.register_subclass("moss")
@dataclass
class MossRobotConfig(ManipulatorRobotConfig):
calibration_dir: str = ".cache/calibration/moss"
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: int | None = None
leader_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem58760431091", <-- UPDATE HERE
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
},
),
}
)
follower_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": FeetechMotorsBusConfig(
port="/dev/tty.usbmodem585A0076891", <-- UPDATE HERE
motors={
# name: (index, model)
"shoulder_pan": [1, "sts3215"],
"shoulder_lift": [2, "sts3215"],
"elbow_flex": [3, "sts3215"],
"wrist_flex": [4, "sts3215"],
"wrist_roll": [5, "sts3215"],
"gripper": [6, "sts3215"],
},
),
}
)
```
**Configure your motors**
Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
```bash
@ -129,28 +176,32 @@ Next, you'll need to calibrate your Moss v1 robot to ensure that the leader and
You will need to move the follower arm to these positions sequentially:
| 1. Zero position | 2. Rotated position | 3. Rest position |
|---|---|---|
| ------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| <img src="../media/moss/follower_zero.webp?raw=true" alt="Moss v1 follower arm zero position" title="Moss v1 follower arm zero position" style="width:100%;"> | <img src="../media/moss/follower_rotated.webp?raw=true" alt="Moss v1 follower arm rotated position" title="Moss v1 follower arm rotated position" style="width:100%;"> | <img src="../media/moss/follower_rest.webp?raw=true" alt="Moss v1 follower arm rest position" title="Moss v1 follower arm rest position" style="width:100%;"> |
Make sure both arms are connected and run this script to launch manual calibration:
```bash
python lerobot/scripts/control_robot.py calibrate \
--robot-path lerobot/configs/robot/moss.yaml \
--robot-overrides '~cameras' --arms main_follower
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--robot.cameras='{}' \
--control.type=calibrate \
--control.arms='["main_follower"]'
```
**Manual calibration of leader arm**
Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
| 1. Zero position | 2. Rotated position | 3. Rest position |
|---|---|---|
| ------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
| <img src="../media/moss/leader_zero.webp?raw=true" alt="Moss v1 leader arm zero position" title="Moss v1 leader arm zero position" style="width:100%;"> | <img src="../media/moss/leader_rotated.webp?raw=true" alt="Moss v1 leader arm rotated position" title="Moss v1 leader arm rotated position" style="width:100%;"> | <img src="../media/moss/leader_rest.webp?raw=true" alt="Moss v1 leader arm rest position" title="Moss v1 leader arm rest position" style="width:100%;"> |
Run this script to launch manual calibration:
```bash
python lerobot/scripts/control_robot.py calibrate \
--robot-path lerobot/configs/robot/moss.yaml \
--robot-overrides '~cameras' --arms main_leader
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--robot.cameras='{}' \
--control.type=calibrate \
--control.arms='["main_leader"]'
```
## Teleoperate
@ -158,18 +209,22 @@ python lerobot/scripts/control_robot.py calibrate \
**Simple teleop**
Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/moss.yaml \
--robot-overrides '~cameras' \
--display-cameras 0
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--robot.cameras='{}' \
--control.type=teleoperate
```
**Teleop with displaying cameras**
Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/moss.yaml
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--control.type=teleoperate
```
## Record a dataset
@ -189,40 +244,46 @@ echo $HF_USER
Record 2 episodes and upload your dataset to the hub:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/moss.yaml \
--fps 30 \
--repo-id ${HF_USER}/moss_test \
--tags moss tutorial \
--warmup-time-s 5 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 2 \
--push-to-hub 1
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/moss_test \
--control.tags='["moss","tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=2 \
--control.push_to_hub=true
```
Note: You can resume recording by adding `--control.resume=true`.
## Visualize a dataset
If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
```bash
echo ${HF_USER}/moss_test
```
If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
```bash
python lerobot/scripts/visualize_dataset_html.py \
--repo-id ${HF_USER}/moss_test
--repo-id ${HF_USER}/moss_test \
--local-files-only 1
```
## Replay an episode
Now try to replay the first episode on your robot:
```bash
python lerobot/scripts/control_robot.py replay \
--robot-path lerobot/configs/robot/moss.yaml \
--fps 30 \
--repo-id ${HF_USER}/moss_test \
--episode 0
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--control.type=replay \
--control.fps=30 \
--control.repo_id=${HF_USER}/moss_test \
--control.episode=0
```
## Train a policy
@ -230,20 +291,18 @@ python lerobot/scripts/control_robot.py replay \
To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
```bash
python lerobot/scripts/train.py \
dataset_repo_id=${HF_USER}/moss_test \
policy=act_moss_real \
env=moss_real \
hydra.run.dir=outputs/train/act_moss_test \
hydra.job.name=act_moss_test \
device=cuda \
wandb.enable=true
--dataset.repo_id=${HF_USER}/moss_test \
--policy.type=act \
--output_dir=outputs/train/act_moss_test \
--job_name=act_moss_test \
--policy.device=cuda \
--wandb.enable=true
```
Let's explain it:
1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/moss_test`.
2. We provided the policy with `policy=act_moss_real`. This loads configurations from [`lerobot/configs/policy/act_moss_real.yaml`](../lerobot/configs/policy/act_moss_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
3. We provided an environment as argument with `env=moss_real`. This loads configurations from [`lerobot/configs/env/moss_real.yaml`](../lerobot/configs/env/moss_real.yaml).
4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/moss_test`.
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
Training should take several hours. You will find checkpoints in `outputs/train/act_moss_test/checkpoints`.
@ -252,21 +311,24 @@ Training should take several hours. You will find checkpoints in `outputs/train/
You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/moss.yaml \
--fps 30 \
--repo-id ${HF_USER}/eval_act_moss_test \
--tags moss tutorial eval \
--warmup-time-s 5 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 10 \
-p outputs/train/act_moss_test/checkpoints/last/pretrained_model
python lerobot/scripts/control_robot.py \
--robot.type=moss \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/eval_act_moss_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=10 \
--control.push_to_hub=true \
--control.policy.path=outputs/train/act_moss_test/checkpoints/last/pretrained_model
```
As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
1. There is an additional `-p` argument which indicates the path to your policy checkpoint with (e.g. `-p outputs/train/eval_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_moss_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_moss_test`).
1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_act_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_moss_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_moss_test`).
## More

18
examples/1_load_lerobot_dataset.py
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script demonstrates the use of `LeRobotDataset` class for handling and processing robotic datasets from Hugging Face.
It illustrates how to load datasets, manipulate them, and apply transformations suitable for machine learning tasks in PyTorch.
@ -105,7 +119,7 @@ print(dataset.features[camera_key]["shape"])
delta_timestamps = {
# loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
camera_key: [-1, -0.5, -0.20, 0],
# loads 8 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
# loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
"observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
# loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
"action": [t / dataset.fps for t in range(64)],
@ -129,6 +143,6 @@ dataloader = torch.utils.data.DataLoader(
for batch in dataloader:
print(f"{batch[camera_key].shape=}") # (32, 4, c, h, w)
print(f"{batch['observation.state'].shape=}") # (32, 5, c)
print(f"{batch['observation.state'].shape=}") # (32, 6, c)
print(f"{batch['action'].shape=}") # (32, 64, c)
break

54
examples/2_evaluate_pretrained_policy.py
View File

@ -1,6 +1,25 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This scripts demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
```bash
pip install -e ".[pusht]"
```
"""
from pathlib import Path
@ -10,7 +29,6 @@ import gymnasium as gym
import imageio
import numpy
import torch
from huggingface_hub import snapshot_download
from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
@ -18,25 +36,15 @@ from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
output_directory = Path("outputs/eval/example_pusht_diffusion")
output_directory.mkdir(parents=True, exist_ok=True)
# Download the diffusion policy for pusht environment
pretrained_policy_path = Path(snapshot_download("lerobot/diffusion_pusht"))
# OR uncomment the following to evaluate a policy from the local outputs/train folder.
# Select your device
device = "cuda"
# Provide the [hugging face repo id](https://huggingface.co/lerobot/diffusion_pusht):
pretrained_policy_path = "lerobot/diffusion_pusht"
# OR a path to a local outputs/train folder.
# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
policy.eval()
# Check if GPU is available
if torch.cuda.is_available():
device = torch.device("cuda")
print("GPU is available. Device set to:", device)
else:
device = torch.device("cpu")
print(f"GPU is not available. Device set to: {device}. Inference will be slower than on GPU.")
# Decrease the number of reverse-diffusion steps (trades off a bit of quality for 10x speed)
policy.diffusion.num_inference_steps = 10
policy.to(device)
# Initialize evaluation environment to render two observation types:
# an image of the scene and state/position of the agent. The environment
@ -47,7 +55,17 @@ env = gym.make(
max_episode_steps=300,
)
# Reset the policy and environmens to prepare for rollout
# We can verify that the shapes of the features expected by the policy match the ones from the observations
# produced by the environment
print(policy.config.input_features)
print(env.observation_space)
# Similarly, we can check that the actions produced by the policy will match the actions expected by the
# environment
print(policy.config.output_features)
print(env.action_space)
# Reset the policy and environments to prepare for rollout
policy.reset()
numpy_observation, info = env.reset(seed=42)

113
examples/3_train_policy.py
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""This scripts demonstrates how to train Diffusion Policy on the PushT environment.
Once you have trained a model with this script, you can try to evaluate it on
@ -8,22 +22,54 @@ from pathlib import Path
import torch
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
from lerobot.common.datasets.utils import dataset_to_policy_features
from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
from lerobot.configs.types import FeatureType
# Create a directory to store the training checkpoint.
output_directory = Path("outputs/train/example_pusht_diffusion")
output_directory.mkdir(parents=True, exist_ok=True)
# Number of offline training steps (we'll only do offline training for this example.)
# Adjust as you prefer. 5000 steps are needed to get something worth evaluating.
training_steps = 5000
device = torch.device("cuda")
log_freq = 250
def main():
# Create a directory to store the training checkpoint.
output_directory = Path("outputs/train/example_pusht_diffusion")
output_directory.mkdir(parents=True, exist_ok=True)
# Set up the dataset.
delta_timestamps = {
# # Select your device
device = torch.device("cuda")
# Number of offline training steps (we'll only do offline training for this example.)
# Adjust as you prefer. 5000 steps are needed to get something worth evaluating.
training_steps = 5000
log_freq = 1
# When starting from scratch (i.e. not from a pretrained policy), we need to specify 2 things before
# creating the policy:
# - input/output shapes: to properly size the policy
# - dataset stats: for normalization and denormalization of input/outputs
dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
features = dataset_to_policy_features(dataset_metadata.features)
output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
input_features = {key: ft for key, ft in features.items() if key not in output_features}
# Policies are initialized with a configuration class, in this case `DiffusionConfig`. For this example,
# we'll just use the defaults and so no arguments other than input/output features need to be passed.
cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
# We can now instantiate our policy with this config and the dataset stats.
policy = DiffusionPolicy(cfg, dataset_stats=dataset_metadata.stats)
policy.train()
policy.to(device)
# Another policy-dataset interaction is with the delta_timestamps. Each policy expects a given number frames
# which can differ for inputs, outputs and rewards (if there are some).
delta_timestamps = {
"observation.image": [i / dataset_metadata.fps for i in cfg.observation_delta_indices],
"observation.state": [i / dataset_metadata.fps for i in cfg.observation_delta_indices],
"action": [i / dataset_metadata.fps for i in cfg.action_delta_indices],
}
# In this case with the standard configuration for Diffusion Policy, it is equivalent to this:
delta_timestamps = {
# Load the previous image and state at -0.1 seconds before current frame,
# then load current image and state corresponding to 0.0 second.
"observation.image": [-0.1, 0.0],
@ -32,38 +78,29 @@ delta_timestamps = {
# and 14 future actions with a 0.1 seconds spacing. All these actions will be
# used to supervise the policy.
"action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
}
dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
}
# Set up the the policy.
# Policies are initialized with a configuration class, in this case `DiffusionConfig`.
# For this example, no arguments need to be passed because the defaults are set up for PushT.
# If you're doing something different, you will likely need to change at least some of the defaults.
cfg = DiffusionConfig()
policy = DiffusionPolicy(cfg, dataset_stats=dataset.meta.stats)
policy.train()
policy.to(device)
# We can then instantiate the dataset with these delta_timestamps configuration.
dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
# Create dataloader for offline training.
dataloader = torch.utils.data.DataLoader(
# Then we create our optimizer and dataloader for offline training.
optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
dataloader = torch.utils.data.DataLoader(
dataset,
num_workers=4,
batch_size=64,
shuffle=True,
pin_memory=device != torch.device("cpu"),
pin_memory=device.type != "cpu",
drop_last=True,
)
)
# Run training loop.
step = 0
done = False
while not done:
# Run training loop.
step = 0
done = False
while not done:
for batch in dataloader:
batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
output_dict = policy.forward(batch)
loss = output_dict["loss"]
batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
loss, _ = policy.forward(batch)
loss.backward()
optimizer.step()
optimizer.zero_grad()
@ -75,5 +112,9 @@ while not done:
done = True
break
# Save a policy checkpoint.
policy.save_pretrained(output_directory)
# Save a policy checkpoint.
policy.save_pretrained(output_directory)
if __name__ == "__main__":
main()

357
examples/4_train_policy_with_script.md
View File

@ -1,193 +1,223 @@
This tutorial will explain the training script, how to use it, and particularly the use of Hydra to configure everything needed for the training run.
This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
> **Note:** The following assume you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
## The training script
LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
LeRobot offers a training script at [`lerobot/scripts/train.py`](../lerobot/scripts/train.py). At a high level it does the following:
- Loads a Hydra configuration file for the following steps (more on Hydra in a moment).
- Makes a simulation environment.
- Makes a dataset corresponding to that simulation environment.
- Makes a policy.
- Initialize/load a configuration for the following steps using.
- Instantiates a dataset.
- (Optional) Instantiates a simulation environment corresponding to that dataset.
- Instantiates a policy.
- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
## Basics of how we use Hydra
Explaining the ins and outs of [Hydra](https://hydra.cc/docs/intro/) is beyond the scope of this document, but here we'll share the main points you need to know.
First, `lerobot/configs` has a directory structure like this:
```
.
├── default.yaml
├── env
│ ├── aloha.yaml
│ ├── pusht.yaml
│ └── xarm.yaml
└── policy
├── act.yaml
├── diffusion.yaml
└── tdmpc.yaml
```
**_For brevity, in the rest of this document we'll drop the leading `lerobot/configs` path. So `default.yaml` really refers to `lerobot/configs/default.yaml`._**
When you run the training script with
## Overview of the configuration system
In the training script, the main function `train` expects a `TrainPipelineConfig` object:
```python
python lerobot/scripts/train.py
# train.py
@parser.wrap()
def train(cfg: TrainPipelineConfig):
```
Hydra is set up to read `default.yaml` (via the `@hydra.main` decorator). If you take a look at the `@hydra.main`'s arguments you will see `config_path="../configs", config_name="default"`. At the top of `default.yaml`, is a `defaults` section which looks likes this:
You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
```yaml
defaults:
- _self_
- env: pusht
- policy: diffusion
When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
```python
@dataclass
class TrainPipelineConfig:
dataset: DatasetConfig
env: envs.EnvConfig | None = None
policy: PreTrainedConfig | None = None
```
in which `DatasetConfig` for example is defined as such:
```python
@dataclass
class DatasetConfig:
repo_id: str
episodes: list[int] | None = None
video_backend: str = "pyav"
```
This logic tells Hydra to incorporate configuration parameters from `env/pusht.yaml` and `policy/diffusion.yaml`. _Note: Be aware of the order as any configuration parameters with the same name will be overidden. Thus, `default.yaml` is overridden by `env/pusht.yaml` which is overidden by `policy/diffusion.yaml`_.
This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
From the command line, we can specify this value with using a very similar syntax `--dataset.repo_id=repo/id`.
Then, `default.yaml` also contains common configuration parameters such as `device: cuda` or `use_amp: false` (for enabling fp16 training). Some other parameters are set to `???` which indicates that they are expected to be set in additional yaml files. For instance, `training.offline_steps: ???` in `default.yaml` is set to `200000` in `diffusion.yaml`.
By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
Thanks to this `defaults` section in `default.yaml`, if you want to train Diffusion Policy with PushT, you really only need to run:
```bash
python lerobot/scripts/train.py
```
However, you can be more explicit and launch the exact same Diffusion Policy training on PushT with:
```bash
python lerobot/scripts/train.py policy=diffusion env=pusht
```
This way of overriding defaults via the CLI is especially useful when you want to change the policy and/or environment. For instance, you can train ACT on the default Aloha environment with:
```bash
python lerobot/scripts/train.py policy=act env=aloha
```
There are two things to note here:
- Config overrides are passed as `param_name=param_value`.
- Here we have overridden the defaults section. `policy=act` tells Hydra to use `policy/act.yaml`, and `env=aloha` tells Hydra to use `env/aloha.yaml`.
_As an aside: we've set up all of our configurations so that they reproduce state-of-the-art results from papers in the literature._
## Overriding configuration parameters in the CLI
Now let's say that we want to train on a different task in the Aloha environment. If you look in `env/aloha.yaml` you will see something like:
```yaml
# lerobot/configs/env/aloha.yaml
env:
task: AlohaInsertion-v0
```
And if you look in `policy/act.yaml` you will see something like:
```yaml
# lerobot/configs/policy/act.yaml
dataset_repo_id: lerobot/aloha_sim_insertion_human
```
But our Aloha environment actually supports a cube transfer task as well. To train for this task, you could manually modify the two yaml configuration files respectively.
First, we'd need to switch to using the cube transfer task for the ALOHA environment.
```diff
# lerobot/configs/env/aloha.yaml
env:
- task: AlohaInsertion-v0
+ task: AlohaTransferCube-v0
```
Then, we'd also need to switch to using the cube transfer dataset.
```diff
# lerobot/configs/policy/act.yaml
-dataset_repo_id: lerobot/aloha_sim_insertion_human
+dataset_repo_id: lerobot/aloha_sim_transfer_cube_human
```
Then, you'd be able to run:
```bash
python lerobot/scripts/train.py policy=act env=aloha
```
and you'd be training and evaluating on the cube transfer task.
An alternative approach to editing the yaml configuration files, would be to override the defaults via the command line:
## Specifying values from the CLI
Let's say that we want to train [Diffusion Policy](../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
```bash
python lerobot/scripts/train.py \
policy=act \
dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
env=aloha \
env.task=AlohaTransferCube-v0
--dataset.repo_id=lerobot/pusht \
--policy.type=diffusion \
--env.type=pusht
```
There's something new here. Notice the `.` delimiter used to traverse the configuration hierarchy. _But be aware that the `defaults` section is an exception. As you saw above, we didn't need to write `defaults.policy=act` in the CLI. `policy=act` was enough._
Putting all that knowledge together, here's the command that was used to train https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human.
Let's break this down:
- To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../lerobot/common/policies)
- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../lerobot/common/envs/configs.py)
Let's see another example. Let's say you've been training [ACT](../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
```bash
python lerobot/scripts/train.py \
hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human \
device=cuda
env=aloha \
env.task=AlohaTransferCube-v0 \
dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
policy=act \
training.eval_freq=10000 \
training.log_freq=250 \
training.offline_steps=100000 \
training.save_model=true \
training.save_freq=25000 \
eval.n_episodes=50 \
eval.batch_size=50 \
wandb.enable=false \
--policy.type=act \
--dataset.repo_id=lerobot/aloha_sim_insertion_human \
--env.type=aloha \
--output_dir=outputs/train/act_aloha_insertion
```
> Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
There's one new thing here: `hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human`, which specifies where to save the training output.
## Using a configuration file not in `lerobot/configs`
Above we discusses the our training script is set up such that Hydra looks for `default.yaml` in `lerobot/configs`. But, if you have a configuration file elsewhere in your filesystem you may use:
We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
Looking at the [`AlohaEnv`](../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
```bash
python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
python lerobot/scripts/train.py \
--policy.type=act \
--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
--env.type=aloha \
--env.task=AlohaTransferCube-v0 \
--output_dir=outputs/train/act_aloha_transfer
```
Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
## Loading from a config file
As a concrete example, this becomes particularly handy when you have a folder with training outputs, and would like to re-run the training. For example, say you previously ran the training script with one of the earlier commands and have `outputs/train/my_experiment/checkpoints/pretrained_model/config.yaml`. This `config.yaml` file will have the full set of configuration parameters within it. To run the training with the same configuration again, do:
Now, let's assume that we want to reproduce the run just above. That run has produced a `train_config.json` file in its checkpoints, which serializes the `TrainPipelineConfig` instance it used:
```json
{
"dataset": {
"repo_id": "lerobot/aloha_sim_transfer_cube_human",
"episodes": null,
...
},
"env": {
"type": "aloha",
"task": "AlohaTransferCube-v0",
"fps": 50,
...
},
"policy": {
"type": "act",
"n_obs_steps": 1,
...
},
...
}
```
We can then simply load the config values from this file using:
```bash
python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
python lerobot/scripts/train.py \
--config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
--output_dir=outputs/train/act_aloha_transfer_2
```
`--config_path` is also a special argument which allows to initialize the config from a local config file. It can point to a directory that contains `train_config.json` or to the config file itself directly.
Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
```bash
python lerobot/scripts/train.py \
--config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
--output_dir=outputs/train/act_aloha_transfer_2
--policy.n_action_steps=80
```
> Note: While `--output_dir` is not required in general, in this case we need to specify it since it will otherwise take the value from the `train_config.json` (which is `outputs/train/act_aloha_transfer`). In order to prevent accidental deletion of previous run checkpoints, we raise an error if you're trying to write in an existing directory. This is not the case when resuming a run, which is what you'll learn next.
`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
```bash
python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
```
will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
## Resume training
Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
Let's reuse the command from the previous run and add a few more options:
```bash
python lerobot/scripts/train.py \
--policy.type=act \
--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
--env.type=aloha \
--env.task=AlohaTransferCube-v0 \
--log_freq=25 \
--save_freq=100 \
--output_dir=outputs/train/run_resumption
```
Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
Here we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can showcase resumption. You should be able to see some logging and have a first checkpoint within 1 minute (depending on hardware). Wait for the first checkpoint to happen, you should see a line that looks like this in your terminal:
```
INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
```
Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
```bash
python lerobot/scripts/train.py \
--config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
--resume=true
```
You should see from the logging that your training picks up from where it left off.
Another reason for which you might want to resume a run is simply to extend training and add more training steps. The number of training steps is set by the option `--steps`, which is 100 000 by default.
You could double the number of steps of the previous run with:
```bash
python lerobot/scripts/train.py \
--config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
--resume=true \
--steps=200000
```
## Outputs of a run
In the output directory, there will be a folder called `checkpoints` with the following structure:
```bash
outputs/train/run_resumption/checkpoints
├── 000100 # checkpoint_dir for training step 100
│ ├── pretrained_model/
│ │ ├── config.json # policy config
│ │ ├── model.safetensors # policy weights
│ │ └── train_config.json # train config
│ └── training_state/
│ ├── optimizer_param_groups.json # optimizer param groups
│ ├── optimizer_state.safetensors # optimizer state
│ ├── rng_state.safetensors # rng states
│ ├── scheduler_state.json # scheduler state
│ └── training_step.json # training step
├── 000200
└── last -> 000200 # symlink to the last available checkpoint
```
## Fine-tuning a pre-trained policy
In addition to the features currently in Draccus, we've added a special `.path` argument for the policy, which allows to load a policy as you would with `PreTrainedPolicy.from_pretrained()`. In that case, `path` can be a local directory that contains a checkpoint or a repo_id pointing to a pretrained policy on the hub.
For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
```bash
python lerobot/scripts/train.py \
--policy.path=lerobot/act_aloha_sim_transfer_cube_human \
--dataset.repo_id=lerobot/aloha_sim_insertion_human \
--env.type=aloha \
--env.task=AlohaInsertion-v0
```
When doing so, keep in mind that the features of the fine-tuning dataset would have to match the input/output features of the pretrained policy.
## Typical logs and metrics
When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you configured your run correctly. The final configuration will also be saved with the checkpoint.
After that, you will see training log like this one:
```
INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
```
or evaluation log like:
or evaluation log:
```
INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
```
These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
- `smpl`: number of samples seen during training.
- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
- `epch`: number of time all unique samples are seen (epoch).
@ -200,14 +230,45 @@ These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here
Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
---
## In short
So far we've seen how to train Diffusion Policy for PushT and ACT for ALOHA. Now, what if we want to train ACT for PushT? Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
We'll summarize here the main use cases to remember from this tutorial.
#### Train a policy from scratch – CLI
```bash
python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
python lerobot/scripts/train.py \
--policy.type=act \ # <- select 'act' policy
--env.type=pusht \ # <- select 'pusht' environment
--dataset.repo_id=lerobot/pusht # <- train on this dataset
```
Please, head on over to our [advanced tutorial on adapting policy configuration to various environments](./advanced/train_act_pusht/train_act_pusht.md) to learn more.
#### Train a policy from scratch - config file + CLI
```bash
python lerobot/scripts/train.py \
--config_path=path/to/pretrained_model \ # <- can also be a repo_id
--policy.n_action_steps=80 # <- you may still override values
```
Or in the meantime, happy coding! 🤗
#### Resume/continue a training run
```bash
python lerobot/scripts/train.py \
--config_path=checkpoint/pretrained_model/ \
--resume=true \
--steps=200000 # <- you can change some training parameters
```
#### Fine-tuning
```bash
python lerobot/scripts/train.py \
--policy.path=lerobot/act_aloha_sim_transfer_cube_human \ # <- can also be a local path to a checkpoint
--dataset.repo_id=lerobot/aloha_sim_insertion_human \
--env.type=aloha \
--env.task=AlohaInsertion-v0
```
---
Now that you know the basics of how to train a policy, you might want to know how to apply this knowledge to actual robots, or how to record your own datasets and train policies on your specific task?
If that's the case, head over to the next tutorial [`7_get_started_with_real_robot.md`](./7_get_started_with_real_robot.md).
Or in the meantime, happy training! 🤗

37
examples/5_resume_training.md
View File

@ -1,37 +0,0 @@
This tutorial explains how to resume a training run that you've started with the training script. If you don't know how our training script and configuration system works, please read [4_train_policy_with_script.md](./4_train_policy_with_script.md) first.
## Basic training resumption
Let's consider the example of training ACT for one of the ALOHA tasks. Here's a command that can achieve that:
```bash
python lerobot/scripts/train.py \
hydra.run.dir=outputs/train/run_resumption \
policy=act \
dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
env=aloha \
env.task=AlohaTransferCube-v0 \
training.log_freq=25 \
training.save_checkpoint=true \
training.save_freq=100
```
Here we're using the default dataset and environment for ACT, and we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can test resumption. You should be able to see some logging and have a first checkpoint within 1 minute. Please interrupt the training after the first checkpoint.
To resume, all that we have to do is run the training script, providing the run directory, and the resume option:
```bash
python lerobot/scripts/train.py \
hydra.run.dir=outputs/train/run_resumption \
resume=true
```
You should see from the logging that your training picks up from where it left off.
Note that with `resume=true`, the configuration file from the last checkpoint in the training output directory is loaded. So it doesn't matter that we haven't provided all the other configuration parameters from our previous command (although there may be warnings to notify you that your command has a different configuration than than the checkpoint).
---
Now you should know how to resume your training run in case it gets interrupted or you want to extend a finished training run.
Happy coding! 🤗

401
examples/7_get_started_with_real_robot.md
View File

@ -36,16 +36,14 @@ Using `pip`:
pip install -e ".[dynamixel]"
```
Or using `poetry`:
Using `poetry`:
```bash
poetry install --sync --extras "dynamixel"
poetry sync --extras "dynamixel"
```
/!\ For Linux only, ffmpeg and opencv requires conda install for now. Run this exact sequence of commands:
Using `uv`:
```bash
conda install -c conda-forge ffmpeg
pip uninstall opencv-python
conda install -c conda-forge "opencv>=4.10.0"
uv sync --extra "dynamixel"
```
You are now ready to plug the 5V power supply to the motor bus of the leader arm (the smaller one) since all its motors only require 5V.
@ -54,24 +52,56 @@ Then plug the 12V power supply to the motor bus of the follower arm. It has two
Finally, connect both arms to your computer via USB. Note that the USB doesn't provide any power, and both arms need to be plugged in with their associated power supply to be detected by your computer.
*Copy pasting python code*
Now you are ready to configure your motors for the first time, as detailed in the sections below. In the upcoming sections, you'll learn about our classes and functions by running some python code in an interactive session, or by copy-pasting it in a python file.
If you have already configured your motors the first time, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
In the upcoming sections, you'll learn about our classes and functions by running some python code, in an interactive session, or by copy-pasting it in a python file. If this is your first time using the tutorial., we highly recommend going through these steps to get deeper intuition about how things work. Once you're more familiar, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/koch.yaml \
--robot-overrides '~cameras' # do not instantiate the cameras
python lerobot/scripts/control_robot.py \
--robot.type=koch \
--control.type=teleoperate
```
It will automatically:
1. Detect and help you correct any motor configuration issues.
2. Identify any missing calibrations and initiate the calibration procedure.
3. Connect the robot and start teleoperation.
1. Identify any missing calibrations and initiate the calibration procedure.
2. Connect the robot and start teleoperation.
### a. Control your motors with DynamixelMotorsBus
You can use the [`DynamixelMotorsBus`](../lerobot/common/robot_devices/motors/dynamixel.py) to communicate with the motors connected as a chain to the corresponding USB bus. This class leverages the Python [Dynamixel SDK](https://emanual.robotis.com/docs/en/software/dynamixel/dynamixel_sdk/sample_code/python_read_write_protocol_2_0/#python-read-write-protocol-20) to facilitate reading from and writing to the motors.
**First Configuration of your motors**
You will need to unplug each motor in turn and run a command the identify the motor. The motor will save its own identification, so you only need to do this once. Start by unplugging all of the motors.
Do the Leader arm first, as all of its motors are of the same type. Plug in your first motor on your leader arm and run this script to set its ID to 1.
```bash
python lerobot/scripts/configure_motor.py \
--port /dev/tty.usbmodem58760432961 \
--brand dynamixel \
--model xl330-m288 \
--baudrate 1000000 \
--ID 1
```
Then unplug your first motor and plug the second motor and set its ID to 2.
```bash
python lerobot/scripts/configure_motor.py \
--port /dev/tty.usbmodem58760432961 \
--brand dynamixel \
--model xl330-m288 \
--baudrate 1000000 \
--ID 2
```
Redo the process for all your motors until ID 6.
The process for the follower arm is almost the same, but the follower arm has two types of motors. For the first two motors, make sure you set the model to `xl430-w250`. _Important: configuring follower motors requires plugging and unplugging power. Make sure you use the 5V power for the XL330s and the 12V power for the XL430s!_
After all of your motors are configured properly, you're ready to plug them all together in a daisy-chain as shown in the original video.
**Instantiate the DynamixelMotorsBus**
To begin, create two instances of the [`DynamixelMotorsBus`](../lerobot/common/robot_devices/motors/dynamixel.py), one for each arm, using their corresponding USB ports (e.g. `DynamixelMotorsBus(port="/dev/tty.usbmodem575E0031751"`).
@ -105,10 +135,10 @@ The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
Reconnect the usb cable.
```
Troubleshooting: On Linux, you might need to give access to the USB ports by running:
Troubleshooting: On Linux, you might need to give access to the USB ports by running this command with your ports:
```bash
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1
sudo chmod 666 /dev/tty.usbmodem575E0032081
sudo chmod 666 /dev/tty.usbmodem575E0031751
```
*Listing and Configuring Motors*
@ -117,13 +147,11 @@ Next, you'll need to list the motors for each arm, including their name, index,
To assign indices to the motors, run this code in an interactive Python session. Replace the `port` values with the ones you identified earlier:
```python
from lerobot.common.robot_devices.motors.configs import DynamixelMotorsBusConfig
from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
leader_port = "/dev/tty.usbmodem575E0031751"
follower_port = "/dev/tty.usbmodem575E0032081"
leader_arm = DynamixelMotorsBus(
port=leader_port,
leader_config = DynamixelMotorsBusConfig(
port="/dev/tty.usbmodem575E0031751",
motors={
# name: (index, model)
"shoulder_pan": (1, "xl330-m077"),
@ -135,8 +163,8 @@ leader_arm = DynamixelMotorsBus(
},
)
follower_arm = DynamixelMotorsBus(
port=follower_port,
follower_config = DynamixelMotorsBusConfig(
port="/dev/tty.usbmodem575E0032081",
motors={
# name: (index, model)
"shoulder_pan": (1, "xl430-w250"),
@ -147,45 +175,57 @@ follower_arm = DynamixelMotorsBus(
"gripper": (6, "xl330-m288"),
},
)
leader_arm = DynamixelMotorsBus(leader_config)
follower_arm = DynamixelMotorsBus(follower_config)
```
*Updating the YAML Configuration File*
IMPORTANTLY: Now that you have your ports, update [`KochRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
```python
@RobotConfig.register_subclass("koch")
@dataclass
class KochRobotConfig(ManipulatorRobotConfig):
calibration_dir: str = ".cache/calibration/koch"
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: int | None = None
Next, update the port values in the YAML configuration file for the Koch robot at [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml) with the ports you've identified:
```yaml
[...]
robot_type: koch
leader_arms:
main:
_target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
port: /dev/tty.usbmodem575E0031751 # <- Update
motors:
leader_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": DynamixelMotorsBusConfig(
port="/dev/tty.usbmodem585A0085511", <-- UPDATE HERE
motors={
# name: (index, model)
shoulder_pan: [1, "xl330-m077"]
shoulder_lift: [2, "xl330-m077"]
elbow_flex: [3, "xl330-m077"]
wrist_flex: [4, "xl330-m077"]
wrist_roll: [5, "xl330-m077"]
gripper: [6, "xl330-m077"]
follower_arms:
main:
_target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
port: /dev/tty.usbmodem575E0032081 # <- Update
motors:
"shoulder_pan": [1, "xl330-m077"],
"shoulder_lift": [2, "xl330-m077"],
"elbow_flex": [3, "xl330-m077"],
"wrist_flex": [4, "xl330-m077"],
"wrist_roll": [5, "xl330-m077"],
"gripper": [6, "xl330-m077"],
},
),
}
)
follower_arms: dict[str, MotorsBusConfig] = field(
default_factory=lambda: {
"main": DynamixelMotorsBusConfig(
port="/dev/tty.usbmodem585A0076891", <-- UPDATE HERE
motors={
# name: (index, model)
shoulder_pan: [1, "xl430-w250"]
shoulder_lift: [2, "xl430-w250"]
elbow_flex: [3, "xl330-m288"]
wrist_flex: [4, "xl330-m288"]
wrist_roll: [5, "xl330-m288"]
gripper: [6, "xl330-m288"]
[...]
"shoulder_pan": [1, "xl430-w250"],
"shoulder_lift": [2, "xl430-w250"],
"elbow_flex": [3, "xl330-m288"],
"wrist_flex": [4, "xl330-m288"],
"wrist_roll": [5, "xl330-m288"],
"gripper": [6, "xl330-m288"],
},
),
}
)
```
Don't forget to set `robot_type: aloha` if you follow this tutorial with [Aloha bimanual robot](aloha-2.github.io) instead of Koch v1.1
This configuration file is used to instantiate your robot across all scripts. We'll cover how this works later on.
**Connect and Configure your Motors**
Before you can start using your motors, you'll need to configure them to ensure proper communication. When you first connect the motors, the [`DynamixelMotorsBus`](../lerobot/common/robot_devices/motors/dynamixel.py) automatically detects any mismatch between the current motor indices (factory set to `1`) and the specified indices (e.g., `1, 2, 3, 4, 5, 6`). This triggers a configuration procedure that requires you to unplug the power cord and motors, then reconnect each motor sequentially, starting from the one closest to the bus.
@ -248,6 +288,11 @@ Steps:
- Scan for devices. All 12 motors should appear.
- Select the motors one by one and move the arm. Check that the graphical indicator near the top right shows the movement.
** There is a common issue with the Dynamixel XL430-W250 motors where the motors become undiscoverable after upgrading their firmware from Mac and Windows Dynamixel Wizard2 applications. When this occurs, it is required to do a firmware recovery (Select `DYNAMIXEL Firmware Recovery` and follow the prompts). There are two known workarounds to conduct this firmware reset:
1) Install the Dynamixel Wizard on a linux machine and complete the firmware recovery
2) Use the Dynamixel U2D2 in order to perform the reset with Windows or Mac. This U2D2 can be purchased [here](https://www.robotis.us/u2d2/).
For either solution, open DYNAMIXEL Wizard 2.0 and select the appropriate port. You will likely be unable to see the motor in the GUI at this time. Select `Firmware Recovery`, carefully choose the correct model, and wait for the process to complete. Finally, re-scan to confirm the firmware recovery was successful.
**Read and Write with DynamixelMotorsBus**
To get familiar with how `DynamixelMotorsBus` communicates with the motors, you can start by reading data from them. Copy past this code in the same interactive python session:
@ -312,27 +357,27 @@ Alternatively, you can unplug the power cord, which will automatically disable t
**Instantiate the ManipulatorRobot**
Before you can teleoperate your robot, you need to instantiate the [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) using the previously defined `leader_arm` and `follower_arm`.
Before you can teleoperate your robot, you need to instantiate the [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) using the previously defined `leader_config` and `follower_config`.
For the Koch v1.1 robot, we only have one leader, so we refer to it as `"main"` and define it as `leader_arms={"main": leader_arm}`. We do the same for the follower arm. For other robots (like the Aloha), which may have two pairs of leader and follower arms, you would define them like this: `leader_arms={"left": left_leader_arm, "right": right_leader_arm},`. Same thing for the follower arms.
For the Koch v1.1 robot, we only have one leader, so we refer to it as `"main"` and define it as `leader_arms={"main": leader_config}`. We do the same for the follower arm. For other robots (like the Aloha), which may have two pairs of leader and follower arms, you would define them like this: `leader_arms={"left": left_leader_config, "right": right_leader_config},`. Same thing for the follower arms.
You also need to provide a path to a calibration directory, such as `calibration_dir=".cache/calibration/koch"`. More on this in the next section.
Run the following code to instantiate your manipulator robot:
```python
from lerobot.common.robot_devices.robots.configs import KochRobotConfig
from lerobot.common.robot_devices.robots.manipulator import ManipulatorRobot
robot = ManipulatorRobot(
robot_type="koch",
leader_arms={"main": leader_arm},
follower_arms={"main": follower_arm},
calibration_dir=".cache/calibration/koch",
robot_config = KochRobotConfig(
leader_arms={"main": leader_config},
follower_arms={"main": follower_config},
cameras={}, # We don't use any camera for now
)
robot = ManipulatorRobot(robot_config)
```
The `robot_type="koch"` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
The `KochRobotConfig` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `robot_type="aloha"` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected. If you need to run manual calibration, simply update `calibration_dir` to `.cache/calibration/aloha`.
For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
**Calibrate and Connect the ManipulatorRobot**
@ -343,18 +388,18 @@ When you connect your robot for the first time, the [`ManipulatorRobot`](../lero
Here are the positions you'll move the follower arm to:
| 1. Zero position | 2. Rotated position | 3. Rest position |
|---|---|---|
| ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| <img src="../media/koch/follower_zero.webp?raw=true" alt="Koch v1.1 follower arm zero position" title="Koch v1.1 follower arm zero position" style="width:100%;"> | <img src="../media/koch/follower_rotated.webp?raw=true" alt="Koch v1.1 follower arm rotated position" title="Koch v1.1 follower arm rotated position" style="width:100%;"> | <img src="../media/koch/follower_rest.webp?raw=true" alt="Koch v1.1 follower arm rest position" title="Koch v1.1 follower arm rest position" style="width:100%;"> |
And here are the corresponding positions for the leader arm:
| 1. Zero position | 2. Rotated position | 3. Rest position |
|---|---|---|
| ----------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- |
| <img src="../media/koch/leader_zero.webp?raw=true" alt="Koch v1.1 leader arm zero position" title="Koch v1.1 leader arm zero position" style="width:100%;"> | <img src="../media/koch/leader_rotated.webp?raw=true" alt="Koch v1.1 leader arm rotated position" title="Koch v1.1 leader arm rotated position" style="width:100%;"> | <img src="../media/koch/leader_rest.webp?raw=true" alt="Koch v1.1 leader arm rest position" title="Koch v1.1 leader arm rest position" style="width:100%;"> |
You can watch a [video tutorial of the calibration procedure](https://youtu.be/8drnU9uRY24) for more details.
During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to mesure if the values changed negatively or positively.
During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
Finally, the rest position ensures that the follower and leader arms are roughly aligned after calibration, preventing sudden movements that could damage the motors when starting teleoperation.
@ -579,9 +624,11 @@ Note: Some cameras may take a few seconds to warm up, and the first frame might
Finally, run this code to instantiate and connectyour camera:
```python
from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
camera = OpenCVCamera(camera_index=0)
config = OpenCVCameraConfig(camera_index=0)
camera = OpenCVCamera(config)
camera.connect()
color_image = camera.read()
@ -603,7 +650,7 @@ uint8
With certain camera, you can also specify additional parameters like frame rate, resolution, and color mode during instantiation. For instance:
```python
camera = OpenCVCamera(camera_index=0, fps=30, width=640, height=480)
config = OpenCVCameraConfig(camera_index=0, fps=30, width=640, height=480)
```
If the provided arguments are not compatible with the camera, an exception will be raised.
@ -617,18 +664,20 @@ camera.disconnect()
**Instantiate your robot with cameras**
Additionaly, you can set up your robot to work with your cameras.
Additionally, you can set up your robot to work with your cameras.
Modify the following Python code with the appropriate camera names and configurations:
```python
robot = ManipulatorRobot(
KochRobotConfig(
leader_arms={"main": leader_arm},
follower_arms={"main": follower_arm},
calibration_dir=".cache/calibration/koch",
cameras={
"laptop": OpenCVCamera(0, fps=30, width=640, height=480),
"phone": OpenCVCamera(1, fps=30, width=640, height=480),
"laptop": OpenCVCameraConfig(0, fps=30, width=640, height=480),
"phone": OpenCVCameraConfig(1, fps=30, width=640, height=480),
},
)
)
robot.connect()
```
@ -652,39 +701,20 @@ torch.Size([3, 480, 640])
255
```
Also, update the following lines of the yaml file for Koch robot [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml) with the names and configurations of your cameras:
```yaml
[...]
cameras:
laptop:
_target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
camera_index: 0
fps: 30
width: 640
height: 480
phone:
_target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
camera_index: 1
fps: 30
width: 640
height: 480
```
### d. Use `control_robot.py` and our `teleoperate` function
This file is used to instantiate your robot in all our scripts. We will explain how this works in the next section.
### d. Use `koch.yaml` and our `teleoperate` function
Instead of manually running the python code in a terminal window, you can use [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) to instantiate your robot by providing the path to the robot yaml file (e.g. [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml)) and control your robot with various modes as explained next.
Instead of manually running the python code in a terminal window, you can use [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) to instantiate your robot by providing the robot configurations via command line and control your robot with various modes as explained next.
Try running this code to teleoperate your robot (if you dont have a camera, keep reading):
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/koch.yaml
python lerobot/scripts/control_robot.py \
--robot.type=koch \
--control.type=teleoperate
```
You will see a lot of lines appearing like this one:
```
INFO 2024-08-10 11:15:03 ol_robot.py:209 dt: 5.12 (195.1hz) dtRlead: 4.93 (203.0hz) dtRfoll: 0.19 (5239.0hz)
INFO 2024-08-10 11:15:03 ol_robot.py:209 dt: 5.12 (195.1hz) dtRlead: 4.93 (203.0hz) dtWfoll: 0.19 (5239.0hz)
```
It contains
@ -694,21 +724,12 @@ It contains
- `dtRlead: 4.93 (203.0hz)` which is the number of milliseconds it took to read the position of the leader arm using `leader_arm.read("Present_Position")`.
- `dtWfoll: 0.22 (4446.9hz)` which is the number of milliseconds it took to set a new goal position for the follower arm using `follower_arm.write("Goal_position", leader_pos)` ; note that writing is done asynchronously so it takes less time than reading.
Note: you can override any entry in the yaml file using `--robot-overrides` and the [hydra.cc](https://hydra.cc/docs/advanced/override_grammar/basic) syntax. If needed, you can override the ports like this:
Importantly: If you don't have any camera, you can remove them dynamically with this [draccus](https://github.com/dlwh/draccus) syntax `--robot.cameras='{}'`:
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/koch.yaml \
--robot-overrides \
leader_arms.main.port=/dev/tty.usbmodem575E0031751 \
follower_arms.main.port=/dev/tty.usbmodem575E0032081
```
Importantly: If you don't have any camera, you can remove them dynamically with this [hydra.cc](https://hydra.cc/docs/advanced/override_grammar/basic) syntax `'~cameras'`:
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/koch.yaml \
--robot-overrides \
'~cameras'
python lerobot/scripts/control_robot.py \
--robot.type=koch \
--robot.cameras='{}' \
--control.type=teleoperate
```
We advise to create a new yaml file when the command becomes too long.
@ -744,23 +765,23 @@ for _ in range(record_time_s * fps):
Importantly, many utilities are still missing. For instance, if you have cameras, you will need to save the images on disk to not go out of RAM, and to do so in threads to not slow down communication with your robot. Also, you will need to store your data in a format optimized for training and web sharing like [`LeRobotDataset`](../lerobot/common/datasets/lerobot_dataset.py). More on this in the next section.
### a. Use `koch.yaml` and the `record` function
### a. Use the `record` function
You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) to achieve efficient data recording. It encompasses many recording utilities:
1. Frames from cameras are saved on disk in threads, and encoded into videos at the end of recording.
1. Frames from cameras are saved on disk in threads, and encoded into videos at the end of each episode recording.
2. Video streams from cameras are displayed in window so that you can verify them.
3. Data is stored with [`LeRobotDataset`](../lerobot/common/datasets/lerobot_dataset.py) format which is pushed to your Hugging Face page (unless `--push-to-hub 0` is provided).
4. Checkpoints are done during recording, so if any issue occurs, you can resume recording by re-running the same command again. You can also use `--force-override 1` to start recording from scratch.
3. Data is stored with [`LeRobotDataset`](../lerobot/common/datasets/lerobot_dataset.py) format which is pushed to your Hugging Face page (unless `--control.push_to_hub=false` is provided).
4. Checkpoints are done during recording, so if any issue occurs, you can resume recording by re-running the same command again with `--control.resume=true`. You will need to manually delete the dataset directory if you want to start recording from scratch.
5. Set the flow of data recording using command line arguments:
- `--warmup-time-s` defines the number of seconds before starting data collection. It allows the robot devices to warmup and synchronize (10 seconds by default).
- `--episode-time-s` defines the number of seconds for data recording for each episode (60 seconds by default).
- `--reset-time-s` defines the number of seconds for resetting the environment after each episode (60 seconds by default).
- `--num-episodes` defines the number of episodes to record (50 by default).
- `--control.warmup_time_s=10` defines the number of seconds before starting data collection. It allows the robot devices to warmup and synchronize (10 seconds by default).
- `--control.episode_time_s=60` defines the number of seconds for data recording for each episode (60 seconds by default).
- `--control.reset_time_s=60` defines the number of seconds for resetting the environment after each episode (60 seconds by default).
- `--control.num_episodes=50` defines the number of episodes to record (50 by default).
6. Control the flow during data recording using keyboard keys:
- Press right arrow `->` at any time during episode recording to early stop and go to resetting. Same during resetting, to early stop and to go to the next episode recording.
- Press left arrow `<-` at any time during episode recording or resetting to early stop, cancel the current episode, and re-record it.
- Press escape `ESC` at any time during episode recording to end the session early and go straight to video encoding and dataset uploading.
7. Similarly to `teleoperate`, you can also use `--robot-path` and `--robot-overrides` to specify your robots.
7. Similarly to `teleoperate`, you can also use the command line to override anything.
Before trying `record`, if you want to push your dataset to the hub, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
```bash
@ -771,27 +792,29 @@ Also, store your Hugging Face repository name in a variable (e.g. `cadene` or `l
HF_USER=$(huggingface-cli whoami | head -n 1)
echo $HF_USER
```
If you don't want to push to hub, use `--push-to-hub 0`.
If you don't want to push to hub, use `--control.push_to_hub=false`.
Now run this to record 2 episodes:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/koch.yaml \
--fps 30 \
--repo-id ${HF_USER}/koch_test \
--tags tutorial \
--warmup-time-s 5 \
--episode-time-s 30 \
--reset-time-s 30 \
--num-episodes 2
python lerobot/scripts/control_robot.py \
--robot.type=koch \
--control.type=record \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.fps=30 \
--control.repo_id=${HF_USER}/koch_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=2 \
--control.push_to_hub=true
```
This will write your dataset locally to `~/.cache/huggingface/lerobot/{repo-id}` (e.g. `data/cadene/koch_test`) and push it on the hub at `https://huggingface.co/datasets/{HF_USER}/{repo-id}`. Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
You can look for other LeRobot datasets on the hub by searching for `LeRobot` tags: https://huggingface.co/datasets?other=LeRobot
Remember to add `--robot-overrides '~cameras'` if you don't have any cameras and you still use the default `koch.yaml` configuration.
You will see a lot of lines appearing like this one:
```
INFO 2024-08-10 15:02:58 ol_robot.py:219 dt:33.34 (30.0hz) dtRlead: 5.06 (197.5hz) dtWfoll: 0.25 (3963.7hz) dtRfoll: 6.22 (160.7hz) dtRlaptop: 32.57 (30.7hz) dtRphone: 33.84 (29.5hz)
@ -803,20 +826,10 @@ It contains:
- `dtRlead: 5.06 (197.5hz)` which is the delta time of reading the present position of the leader arm.
- `dtWfoll: 0.25 (3963.7hz)` which is the delta time of writing the goal position on the follower arm ; writing is asynchronous so it takes less time than reading.
- `dtRfoll: 6.22 (160.7hz)` which is the delta time of reading the present position on the follower arm.
- `dtRlaptop:32.57 (30.7hz) ` which is the delta time of capturing an image from the laptop camera in the thread running asynchrously.
- `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchrously.
- `dtRlaptop:32.57 (30.7hz) ` which is the delta time of capturing an image from the laptop camera in the thread running asynchronously.
- `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchronously.
Troubleshooting:
- On Linux, if you encounter a hanging issue when using cameras, uninstall opencv and re-install it with conda:
```bash
pip uninstall opencv-python
conda install -c conda-forge opencv=4.10.0
```
- On Linux, if you encounter any issue during video encoding with `ffmpeg: unknown encoder libsvtav1`, you can:
- install with conda-forge by running `conda install -c conda-forge ffmpeg` (it should be compiled with `libsvtav1`),
- or, install [Homebrew](https://brew.sh) and run `brew install ffmpeg` (it should be compiled with `libsvtav1`),
- or, install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1),
- and, make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/koch_test) that you can obtain by running:
@ -824,7 +837,7 @@ At the end of data recording, your dataset will be uploaded on your Hugging Face
echo https://huggingface.co/datasets/${HF_USER}/koch_test
```
### b. Advices for recording dataset
### b. Advice for recording dataset
Once you're comfortable with data recording, it's time to create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings.
@ -842,6 +855,8 @@ python lerobot/scripts/visualize_dataset_html.py \
--repo-id ${HF_USER}/koch_test
```
Note: You might need to add `--local-files-only 1` if your dataset was not uploaded to hugging face hub.
This will launch a local web server that looks like this:
<div style="text-align:center;">
<img src="../media/tutorial/visualize_dataset_html.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="100%">
@ -853,11 +868,12 @@ A useful feature of [`lerobot/scripts/control_robot.py`](../lerobot/scripts/cont
To replay the first episode of the dataset you just recorded, run the following command:
```bash
python lerobot/scripts/control_robot.py replay \
--robot-path lerobot/configs/robot/koch.yaml \
--fps 30 \
--repo-id ${HF_USER}/koch_test \
--episode 0
python lerobot/scripts/control_robot.py \
--robot.type=koch \
--control.type=replay \
--control.fps=30 \
--control.repo_id=${HF_USER}/koch_test \
--control.episode=0
```
Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
@ -869,51 +885,18 @@ Your robot should replicate movements similar to those you recorded. For example
To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
```bash
python lerobot/scripts/train.py \
dataset_repo_id=${HF_USER}/koch_test \
policy=act_koch_real \
env=koch_real \
hydra.run.dir=outputs/train/act_koch_test \
hydra.job.name=act_koch_test \
device=cuda \
wandb.enable=true
--dataset.repo_id=${HF_USER}/koch_test \
--policy.type=act \
--output_dir=outputs/train/act_koch_test \
--job_name=act_koch_test \
--policy.device=cuda \
--wandb.enable=true
```
Let's explain it:
1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/koch_test`.
2. We provided the policy with `policy=act_koch_real`. This loads configurations from [`lerobot/configs/policy/act_koch_real.yaml`](../lerobot/configs/policy/act_koch_real.yaml). Importantly, this policy uses 2 cameras as input `laptop` and `phone`. If your dataset has different cameras, update the yaml file to account for it in the following parts:
```yaml
...
override_dataset_stats:
observation.images.laptop:
# stats from imagenet, since we use a pretrained vision model
mean: [[[0.485]], [[0.456]], [[0.406]]] # (c,1,1)
std: [[[0.229]], [[0.224]], [[0.225]]] # (c,1,1)
observation.images.phone:
# stats from imagenet, since we use a pretrained vision model
mean: [[[0.485]], [[0.456]], [[0.406]]] # (c,1,1)
std: [[[0.229]], [[0.224]], [[0.225]]] # (c,1,1)
...
input_shapes:
observation.images.laptop: [3, 480, 640]
observation.images.phone: [3, 480, 640]
...
input_normalization_modes:
observation.images.laptop: mean_std
observation.images.phone: mean_std
...
```
3. We provided an environment as argument with `env=koch_real`. This loads configurations from [`lerobot/configs/env/koch_real.yaml`](../lerobot/configs/env/koch_real.yaml). It looks like
```yaml
fps: 30
env:
name: real_world
task: null
state_dim: 6
action_dim: 6
fps: ${fps}
```
It should match your dataset (e.g. `fps: 30`) and your robot (e.g. `state_dim: 6` and `action_dim: 6`). We are still working on simplifying this in future versions of `lerobot`.
4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/koch_test`.
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
For more information on the `train` script see the previous tutorial: [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md)
@ -978,34 +961,36 @@ for _ in range(inference_time_s * fps):
busy_wait(1 / fps - dt_s)
```
### a. Use `koch.yaml` and our `record` function
### a. Use our `record` function
Ideally, when controlling your robot with your neural network, you would want to record evaluation episodes and to be able to visualize them later on, or even train on them like in Reinforcement Learning. This pretty much corresponds to recording a new dataset but with a neural network providing the actions instead of teleoperation.
To this end, you can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/koch.yaml \
--fps 30 \
--repo-id ${HF_USER}/eval_koch_test \
--tags tutorial eval \
--warmup-time-s 5 \
--episode-time-s 30 \
--reset-time-s 30 \
--num-episodes 10 \
-p outputs/train/act_koch_test/checkpoints/last/pretrained_model
python lerobot/scripts/control_robot.py \
--robot.type=koch \
--control.type=record \
--control.fps=30 \
--control.repo_id=${HF_USER}/eval_act_koch_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=10 \
--control.push_to_hub=true \
--control.policy.path=outputs/train/act_koch_test/checkpoints/last/pretrained_model
```
As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
1. There is an additional `-p` argument which indicates the path to your policy checkpoint with (e.g. `-p outputs/train/eval_koch_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_koch_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_koch_test`).
1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_koch_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_koch_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_koch_test`).
### b. Visualize evaluation afterwards
You can then visualize your evaluation dataset by running the same command as before but with the new inference dataset as argument:
```bash
python lerobot/scripts/visualize_dataset.py \
--repo-id ${HF_USER}/eval_koch_test
--repo-id ${HF_USER}/eval_act_koch_test
```
## 6. Next step

63
examples/8_use_stretch.md
View File

@ -43,21 +43,19 @@ conda create -y -n lerobot python=3.10 && conda activate lerobot
git clone https://github.com/huggingface/lerobot.git ~/lerobot
```
6. Install LeRobot with stretch dependencies:
6. When using `miniconda`, install `ffmpeg` in your environment:
```bash
conda install ffmpeg -c conda-forge
```
7. Install LeRobot with stretch dependencies:
```bash
cd ~/lerobot && pip install -e ".[stretch]"
```
> **Note:** If you get this message, you can ignore it: `ERROR: pip's dependency resolver does not currently take into account all the packages that are installed.`
For Linux only (not Mac), install extra dependencies for recording datasets:
```bash
conda install -y -c conda-forge ffmpeg
pip uninstall -y opencv-python
conda install -y -c conda-forge "opencv>=4.10.0"
```
7. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
8. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
```bash
stretch_system_check.py
```
@ -92,20 +90,24 @@ Serial Number = stretch-se3-3054
**Calibrate (Optional)**
Before operating Stretch, you need to [home](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#homing) it first. Be mindful about giving Stretch some space as this procedure will move the robot's arm and gripper. Now run this command:
```bash
python lerobot/scripts/control_robot.py calibrate \
--robot-path lerobot/configs/robot/stretch.yaml
python lerobot/scripts/control_robot.py \
--robot.type=stretch \
--control.type=calibrate
```
This is equivalent to running `stretch_robot_home.py`
> **Note:** If you run any of the LeRobot scripts below and Stretch is not poperly homed, it will automatically home/calibrate first.
> **Note:** If you run any of the LeRobot scripts below and Stretch is not properly homed, it will automatically home/calibrate first.
**Teleoperate**
Before trying teleoperation, you need activate the gamepad controller by pressing the middle button. For more info, see Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/hello_robot/#gamepad-teleoperation).
Now try out teleoperation (see above documentation to learn about the gamepad controls):
> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/stretch.yaml
python lerobot/scripts/control_robot.py \
--robot.type=stretch \
--control.type=teleoperate
```
This is essentially the same as running `stretch_gamepad_teleop.py`
@ -125,16 +127,18 @@ echo $HF_USER
Record one episode:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/stretch.yaml \
--fps 20 \
--repo-id ${HF_USER}/stretch_test \
--tags stretch tutorial \
--warmup-time-s 3 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 1 \
--push-to-hub 0
python lerobot/scripts/control_robot.py \
--robot.type=stretch \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/stretch_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=2 \
--control.push_to_hub=true
```
> **Note:** If you're using ssh to connect to Stretch and run this script, you won't be able to visualize its cameras feed (though they will still be recording). To see the cameras stream, use [tethered](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup) or [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup).
@ -142,11 +146,12 @@ python lerobot/scripts/control_robot.py record \
**Replay an episode**
Now try to replay this episode (make sure the robot's initial position is the same):
```bash
python lerobot/scripts/control_robot.py replay \
--robot-path lerobot/configs/robot/stretch.yaml \
--fps 20 \
--repo-id ${HF_USER}/stretch_test \
--episode 0
python lerobot/scripts/control_robot.py \
--robot.type=stretch \
--control.type=replay \
--control.fps=30 \
--control.repo_id=${HF_USER}/stretch_test \
--control.episode=0
```
Follow [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) to train a policy on your data and run inference on your robot. You will need to adapt the code for Stretch.

130
examples/9_use_aloha.md
View File

@ -2,7 +2,7 @@ This tutorial explains how to use [Aloha and Aloha 2 stationary](https://www.tro
## Setup
Follow the [documentation from Trossen Robotics](https://docs.trossenrobotics.com/aloha_docs/getting_started/stationary/hardware_setup.html) for setting up the hardware and plugging the 4 arms and 4 cameras to your computer.
Follow the [documentation from Trossen Robotics](https://docs.trossenrobotics.com/aloha_docs/2.0/getting_started/stationary/hardware_setup.html) for setting up the hardware and plugging the 4 arms and 4 cameras to your computer.
## Install LeRobot
@ -30,16 +30,14 @@ conda create -y -n lerobot python=3.10 && conda activate lerobot
git clone https://github.com/huggingface/lerobot.git ~/lerobot
```
5. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
5. When using `miniconda`, install `ffmpeg` in your environment:
```bash
cd ~/lerobot && pip install -e ".[dynamixel, intelrealsense]"
conda install ffmpeg -c conda-forge
```
For Linux only (not Mac), install extra dependencies for recording datasets:
6. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
```bash
conda install -y -c conda-forge ffmpeg
pip uninstall -y opencv-python
conda install -y -c conda-forge "opencv>=4.10.0"
cd ~/lerobot && pip install -e ".[dynamixel, intelrealsense]"
```
## Teleoperate
@ -50,17 +48,22 @@ Teleoperation consists in manually operating the leader arms to move the followe
2. Our code assumes that your robot has been assembled following Trossen Robotics instructions. This allows us to skip calibration, as we use the pre-defined calibration files in `.cache/calibration/aloha_default`. If you replace a motor, make sure you follow the exact instructions from Trossen Robotics.
By running the following code, you can start your first **SAFE** teleoperation:
> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/aloha.yaml \
--robot-overrides max_relative_target=5
python lerobot/scripts/control_robot.py \
--robot.type=aloha \
--robot.max_relative_target=5 \
--control.type=teleoperate
```
By adding `--robot-overrides max_relative_target=5`, we override the default value for `max_relative_target` defined in `lerobot/configs/robot/aloha.yaml`. It is expected to be `5` to limit the magnitude of the movement for more safety, but the teleoperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot-overrides max_relative_target=null` to the command line:
By adding `--robot.max_relative_target=5`, we override the default value for `max_relative_target` defined in [`AlohaRobotConfig`](lerobot/common/robot_devices/robots/configs.py). It is expected to be `5` to limit the magnitude of the movement for more safety, but the teleoperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot.max_relative_target=null` to the command line:
```bash
python lerobot/scripts/control_robot.py teleoperate \
--robot-path lerobot/configs/robot/aloha.yaml \
--robot-overrides max_relative_target=null
python lerobot/scripts/control_robot.py \
--robot.type=aloha \
--robot.max_relative_target=null \
--control.type=teleoperate
```
## Record a dataset
@ -80,27 +83,29 @@ echo $HF_USER
Record 2 episodes and upload your dataset to the hub:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/aloha.yaml \
--robot-overrides max_relative_target=null \
--fps 30 \
--repo-id ${HF_USER}/aloha_test \
--tags aloha tutorial \
--warmup-time-s 5 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 2 \
--push-to-hub 1
python lerobot/scripts/control_robot.py \
--robot.type=aloha \
--robot.max_relative_target=null \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/aloha_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=2 \
--control.push_to_hub=true
```
## Visualize a dataset
If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
```bash
echo ${HF_USER}/aloha_test
```
If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
```bash
python lerobot/scripts/visualize_dataset_html.py \
--repo-id ${HF_USER}/aloha_test
@ -109,16 +114,17 @@ python lerobot/scripts/visualize_dataset_html.py \
## Replay an episode
**/!\ FOR SAFETY, READ THIS /!\**
Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot-overrides max_relative_target=5` to your command line as explained above.
Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot.max_relative_target=5` to your command line as explained above.
Now try to replay the first episode on your robot:
```bash
python lerobot/scripts/control_robot.py replay \
--robot-path lerobot/configs/robot/aloha.yaml \
--robot-overrides max_relative_target=null \
--fps 30 \
--repo-id ${HF_USER}/aloha_test \
--episode 0
python lerobot/scripts/control_robot.py \
--robot.type=aloha \
--robot.max_relative_target=null \
--control.type=replay \
--control.fps=30 \
--control.repo_id=${HF_USER}/aloha_test \
--control.episode=0
```
## Train a policy
@ -126,49 +132,51 @@ python lerobot/scripts/control_robot.py replay \
To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
```bash
python lerobot/scripts/train.py \
dataset_repo_id=${HF_USER}/aloha_test \
policy=act_aloha_real \
env=aloha_real \
hydra.run.dir=outputs/train/act_aloha_test \
hydra.job.name=act_aloha_test \
device=cuda \
wandb.enable=true
--dataset.repo_id=${HF_USER}/aloha_test \
--policy.type=act \
--output_dir=outputs/train/act_aloha_test \
--job_name=act_aloha_test \
--policy.device=cuda \
--wandb.enable=true
```
Let's explain it:
1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/aloha_test`.
2. We provided the policy with `policy=act_aloha_real`. This loads configurations from [`lerobot/configs/policy/act_aloha_real.yaml`](../lerobot/configs/policy/act_aloha_real.yaml). Importantly, this policy uses 4 cameras as input `cam_right_wrist`, `cam_left_wrist`, `cam_high`, and `cam_low`.
3. We provided an environment as argument with `env=aloha_real`. This loads configurations from [`lerobot/configs/env/aloha_real.yaml`](../lerobot/configs/env/aloha_real.yaml). Note: this yaml defines 18 dimensions for the `state_dim` and `action_dim`, corresponding to 18 motors, not 14 motors as used in previous Aloha work. This is because, we include the `shoulder_shadow` and `elbow_shadow` motors for simplicity.
4. We provided `device=cuda` since we are training on a Nvidia GPU.
1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/aloha_test`.
2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
For more information on the `train` script see the previous tutorial: [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md)
Training should take several hours. You will find checkpoints in `outputs/train/act_aloha_test/checkpoints`.
## Evaluate your policy
You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
```bash
python lerobot/scripts/control_robot.py record \
--robot-path lerobot/configs/robot/aloha.yaml \
--robot-overrides max_relative_target=null \
--fps 30 \
--repo-id ${HF_USER}/eval_act_aloha_test \
--tags aloha tutorial eval \
--warmup-time-s 5 \
--episode-time-s 40 \
--reset-time-s 10 \
--num-episodes 10 \
--num-image-writer-processes 1 \
-p outputs/train/act_aloha_test/checkpoints/last/pretrained_model
python lerobot/scripts/control_robot.py \
--robot.type=aloha \
--control.type=record \
--control.fps=30 \
--control.single_task="Grasp a lego block and put it in the bin." \
--control.repo_id=${HF_USER}/eval_act_aloha_test \
--control.tags='["tutorial"]' \
--control.warmup_time_s=5 \
--control.episode_time_s=30 \
--control.reset_time_s=30 \
--control.num_episodes=10 \
--control.push_to_hub=true \
--control.policy.path=outputs/train/act_aloha_test/checkpoints/last/pretrained_model \
--control.num_image_writer_processes=1
```
As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
1. There is an additional `-p` argument which indicates the path to your policy checkpoint with (e.g. `-p outputs/train/eval_aloha_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_aloha_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_aloha_test`).
3. We use `--num-image-writer-processes 1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constent 30 fps during inference. Feel free to explore different values for `--num-image-writer-processes`.
1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with (e.g. `outputs/train/eval_act_aloha_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_aloha_test`).
2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_aloha_test`).
3. We use `--control.num_image_writer_processes=1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constant 30 fps during inference. Feel free to explore different values for `--control.num_image_writer_processes`.
## More
Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth explaination.
Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth explanation.
If you have any question or need help, please reach out on Discord in the channel `#aloha-arm`.

14
examples/6_add_image_transforms.py → examples/advanced/1_add_image_transforms.py
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and

87
examples/advanced/1_train_act_pusht/act_pusht.yaml
View File

@ -1,87 +0,0 @@
# @package _global_
# Change the seed to match what PushT eval uses
# (to avoid evaluating on seeds used for generating the training data).
seed: 100000
# Change the dataset repository to the PushT one.
dataset_repo_id: lerobot/pusht
override_dataset_stats:
observation.image:
# stats from imagenet, since we use a pretrained vision model
mean: [[[0.485]], [[0.456]], [[0.406]]] # (c,1,1)
std: [[[0.229]], [[0.224]], [[0.225]]] # (c,1,1)
training:
offline_steps: 80000
online_steps: 0
eval_freq: 10000
save_freq: 100000
log_freq: 250
save_model: true
batch_size: 8
lr: 1e-5
lr_backbone: 1e-5
weight_decay: 1e-4
grad_clip_norm: 10
online_steps_between_rollouts: 1
delta_timestamps:
action: "[i / ${fps} for i in range(${policy.chunk_size})]"
eval:
n_episodes: 50
batch_size: 50
# See `configuration_act.py` for more details.
policy:
name: act
# Input / output structure.
n_obs_steps: 1
chunk_size: 100 # chunk_size
n_action_steps: 100
input_shapes:
observation.image: [3, 96, 96]
observation.state: ["${env.state_dim}"]
output_shapes:
action: ["${env.action_dim}"]
# Normalization / Unnormalization
input_normalization_modes:
observation.image: mean_std
# Use min_max normalization just because it's more standard.
observation.state: min_max
output_normalization_modes:
# Use min_max normalization just because it's more standard.
action: min_max
# Architecture.
# Vision backbone.
vision_backbone: resnet18
pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
replace_final_stride_with_dilation: false
# Transformer layers.
pre_norm: false
dim_model: 512
n_heads: 8
dim_feedforward: 3200
feedforward_activation: relu
n_encoder_layers: 4
# Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
# that means only the first layer is used. Here we match the original implementation by setting this to 1.
# See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
n_decoder_layers: 1
# VAE.
use_vae: true
latent_dim: 32
n_vae_encoder_layers: 4
# Inference.
temporal_ensemble_coeff: null
# Training and loss computation.
dropout: 0.1
kl_weight: 10.0

70
examples/advanced/1_train_act_pusht/train_act_pusht.md
View File

@ -1,70 +0,0 @@
In this tutorial we will learn how to adapt a policy configuration to be compatible with a new environment and dataset. As a concrete example, we will adapt the default configuration for ACT to be compatible with the PushT environment and dataset.
If you haven't already read our tutorial on the [training script and configuration tooling](../4_train_policy_with_script.md) please do so prior to tackling this tutorial.
Let's get started!
Suppose we want to train ACT for PushT. Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
```bash
python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
```
We need to adapt the parameters of the ACT policy configuration to the PushT environment. The most important ones are the image keys.
ALOHA's datasets and environments typically use a variable number of cameras. In `lerobot/configs/policy/act.yaml` you may notice two relevant sections. Here we show you the minimal diff needed to adjust to PushT:
```diff
override_dataset_stats:
- observation.images.top:
+ observation.image:
# stats from imagenet, since we use a pretrained vision model
mean: [[[0.485]], [[0.456]], [[0.406]]] # (c,1,1)
std: [[[0.229]], [[0.224]], [[0.225]]] # (c,1,1)
policy:
input_shapes:
- observation.images.top: [3, 480, 640]
+ observation.image: [3, 96, 96]
observation.state: ["${env.state_dim}"]
output_shapes:
action: ["${env.action_dim}"]
input_normalization_modes:
- observation.images.top: mean_std
+ observation.image: mean_std
observation.state: min_max
output_normalization_modes:
action: min_max
```
Here we've accounted for the following:
- PushT uses "observation.image" for its image key.
- PushT provides smaller images.
_Side note: technically we could override these via the CLI, but with many changes it gets a bit messy, and we also have a bit of a challenge in that we're using `.` in our observation keys which is treated by Hydra as a hierarchical separator_.
For your convenience, we provide [`act_pusht.yaml`](./act_pusht.yaml) in this directory. It contains the diff above, plus some other (optional) ones that are explained within. Please copy it into `lerobot/configs/policy` with:
```bash
cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/act_pusht.yaml
```
(remember from a [previous tutorial](../4_train_policy_with_script.md) that Hydra will look in the `lerobot/configs` directory). Now try running the following.
<!-- Note to contributor: are you changing this command? Note that it's tested in `Makefile`, so change it there too! -->
```bash
python lerobot/scripts/train.py policy=act_pusht env=pusht
```
Notice that this is much the same as the command that failed at the start of the tutorial, only:
- Now we are using `policy=act_pusht` to point to our new configuration file.
- We can drop `dataset_repo_id=lerobot/pusht` as the change is incorporated in our new configuration file.
Hurrah! You're now training ACT for the PushT environment.
---
The bottom line of this tutorial is that when training policies for different environments and datasets you will need to understand what parts of the policy configuration are specific to those and make changes accordingly.
Happy coding! 🤗

104
examples/advanced/2_calculate_validation_loss.py
View File

@ -1,3 +1,17 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""This script demonstrates how to slice a dataset and calculate the loss on a subset of the data.
This technique can be useful for debugging and testing purposes, as well as identifying whether a policy
@ -9,27 +23,27 @@ on the target environment, whether that be in simulation or the real world.
"""
import math
from pathlib import Path
import torch
from huggingface_hub import snapshot_download
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
device = torch.device("cuda")
# Download the diffusion policy for pusht environment
pretrained_policy_path = Path(snapshot_download("lerobot/diffusion_pusht"))
# OR uncomment the following to evaluate a policy from the local outputs/train folder.
# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
def main():
device = torch.device("cuda")
policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
policy.eval()
policy.to(device)
# Download the diffusion policy for pusht environment
pretrained_policy_path = "lerobot/diffusion_pusht"
# OR uncomment the following to evaluate a policy from the local outputs/train folder.
# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
# Set up the dataset.
delta_timestamps = {
policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
policy.eval()
policy.to(device)
# Set up the dataset.
delta_timestamps = {
# Load the previous image and state at -0.1 seconds before current frame,
# then load current image and state corresponding to 0.0 second.
"observation.image": [-0.1, 0.0],
@ -38,47 +52,53 @@ delta_timestamps = {
# and 14 future actions with a 0.1 seconds spacing. All these actions will be
# used to calculate the loss.
"action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
}
}
# Load the last 10% of episodes of the dataset as a validation set.
# - Load dataset metadata
dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
# - Calculate train and val episodes
total_episodes = dataset_metadata.total_episodes
episodes = list(range(dataset_metadata.total_episodes))
num_train_episodes = math.floor(total_episodes * 90 / 100)
train_episodes = episodes[:num_train_episodes]
val_episodes = episodes[num_train_episodes:]
print(f"Number of episodes in full dataset: {total_episodes}")
print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
# - Load train an val datasets
train_dataset = LeRobotDataset("lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps)
val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
# Load the last 10% of episodes of the dataset as a validation set.
# - Load dataset metadata
dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
# - Calculate train and val episodes
total_episodes = dataset_metadata.total_episodes
episodes = list(range(dataset_metadata.total_episodes))
num_train_episodes = math.floor(total_episodes * 90 / 100)
train_episodes = episodes[:num_train_episodes]
val_episodes = episodes[num_train_episodes:]
print(f"Number of episodes in full dataset: {total_episodes}")
print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
# - Load train an val datasets
train_dataset = LeRobotDataset(
"lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
)
val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
# Create dataloader for evaluation.
val_dataloader = torch.utils.data.DataLoader(
# Create dataloader for evaluation.
val_dataloader = torch.utils.data.DataLoader(
val_dataset,
num_workers=4,
batch_size=64,
shuffle=False,
pin_memory=device != torch.device("cpu"),
drop_last=False,
)
)
# Run validation loop.
loss_cumsum = 0
n_examples_evaluated = 0
for batch in val_dataloader:
# Run validation loop.
loss_cumsum = 0
n_examples_evaluated = 0
for batch in val_dataloader:
batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
output_dict = policy.forward(batch)
loss, _ = policy.forward(batch)
loss_cumsum += output_dict["loss"].item()
loss_cumsum += loss.item()
n_examples_evaluated += batch["index"].shape[0]
# Calculate the average loss over the validation set.
average_loss = loss_cumsum / n_examples_evaluated
# Calculate the average loss over the validation set.
average_loss = loss_cumsum / n_examples_evaluated
print(f"Average loss on validation set: {average_loss:.4f}")
print(f"Average loss on validation set: {average_loss:.4f}")
if __name__ == "__main__":
main()

222
examples/port_datasets/pusht_zarr.py
View File

@ -1,222 +0,0 @@
import shutil
from pathlib import Path
import numpy as np
import torch
from lerobot.common.datasets.lerobot_dataset import LEROBOT_HOME, LeRobotDataset
from lerobot.common.datasets.push_dataset_to_hub._download_raw import download_raw
PUSHT_TASK = "Push the T-shaped blue block onto the T-shaped green target surface."
PUSHT_FEATURES = {
"observation.state": {
"dtype": "float32",
"shape": (2,),
"names": {
"axes": ["x", "y"],
},
},
"action": {
"dtype": "float32",
"shape": (2,),
"names": {
"axes": ["x", "y"],
},
},
"next.reward": {
"dtype": "float32",
"shape": (1,),
"names": None,
},
"next.success": {
"dtype": "bool",
"shape": (1,),
"names": None,
},
"observation.environment_state": {
"dtype": "float32",
"shape": (16,),
"names": [
"keypoints",
],
},
"observation.image": {
"dtype": None,
"shape": (3, 96, 96),
"names": [
"channel",
"height",
"width",
],
},
}
def build_features(mode: str) -> dict:
features = PUSHT_FEATURES
if mode == "keypoints":
features.pop("observation.image")
else:
features.pop("observation.environment_state")
features["observation.image"]["dtype"] = mode
return features
def load_raw_dataset(zarr_path: Path):
try:
from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
ReplayBuffer as DiffusionPolicyReplayBuffer,
)
except ModuleNotFoundError as e:
print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
raise e
zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
return zarr_data
def calculate_coverage(zarr_data):
try:
import pymunk
from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
except ModuleNotFoundError as e:
print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
raise e
block_pos = zarr_data["state"][:, 2:4]
block_angle = zarr_data["state"][:, 4]
num_frames = len(block_pos)
coverage = np.zeros((num_frames,))
# 8 keypoints with 2 coords each
keypoints = np.zeros((num_frames, 16))
# Set x, y, theta (in radians)
goal_pos_angle = np.array([256, 256, np.pi / 4])
goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
for i in range(num_frames):
space = pymunk.Space()
space.gravity = 0, 0
space.damping = 0
# Add walls.
walls = [
PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
]
space.add(*walls)
block_body, block_shapes = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
block_geom = pymunk_to_shapely(block_body, block_body.shapes)
intersection_area = goal_geom.intersection(block_geom).area
goal_area = goal_geom.area
coverage[i] = intersection_area / goal_area
keypoints[i] = torch.from_numpy(PushTEnv.get_keypoints(block_shapes).flatten())
return coverage, keypoints
def calculate_success(coverage: float, success_threshold: float):
return coverage > success_threshold
def calculate_reward(coverage: float, success_threshold: float):
return np.clip(coverage / success_threshold, 0, 1)
def main(raw_dir: Path, repo_id: str, mode: str = "video", push_to_hub: bool = True):
if mode not in ["video", "image", "keypoints"]:
raise ValueError(mode)
if (LEROBOT_HOME / repo_id).exists():
shutil.rmtree(LEROBOT_HOME / repo_id)
if not raw_dir.exists():
download_raw(raw_dir, repo_id="lerobot-raw/pusht_raw")
zarr_data = load_raw_dataset(zarr_path=raw_dir / "pusht_cchi_v7_replay.zarr")
env_state = zarr_data["state"][:]
agent_pos = env_state[:, :2]
action = zarr_data["action"][:]
image = zarr_data["img"] # (b, h, w, c)
episode_data_index = {
"from": np.concatenate(([0], zarr_data.meta["episode_ends"][:-1])),
"to": zarr_data.meta["episode_ends"],
}
# Calculate success and reward based on the overlapping area
# of the T-object and the T-area.
coverage, keypoints = calculate_coverage(zarr_data)
success = calculate_success(coverage, success_threshold=0.95)
reward = calculate_reward(coverage, success_threshold=0.95)
features = build_features(mode)
dataset = LeRobotDataset.create(
repo_id=repo_id,
fps=10,
robot_type="2d pointer",
features=features,
image_writer_threads=4,
)
episodes = range(len(episode_data_index["from"]))
for ep_idx in episodes:
from_idx = episode_data_index["from"][ep_idx]
to_idx = episode_data_index["to"][ep_idx]
num_frames = to_idx - from_idx
for frame_idx in range(num_frames):
i = from_idx + frame_idx
frame = {
"action": torch.from_numpy(action[i]),
# Shift reward and success by +1 until the last item of the episode
"next.reward": reward[i + (frame_idx < num_frames - 1)],
"next.success": success[i + (frame_idx < num_frames - 1)],
}
frame["observation.state"] = torch.from_numpy(agent_pos[i])
if mode == "keypoints":
frame["observation.environment_state"] = torch.from_numpy(keypoints[i])
else:
frame["observation.image"] = torch.from_numpy(image[i])
dataset.add_frame(frame)
dataset.save_episode(task=PUSHT_TASK)
dataset.consolidate()
if push_to_hub:
dataset.push_to_hub()
if __name__ == "__main__":
# To try this script, modify the repo id with your own HuggingFace user (e.g cadene/pusht)
repo_id = "lerobot/pusht"
modes = ["video", "image", "keypoints"]
# Uncomment if you want to try with a specific mode
# modes = ["video"]
# modes = ["image"]
# modes = ["keypoints"]
raw_dir = Path("data/lerobot-raw/pusht_raw")
for mode in modes:
if mode in ["image", "keypoints"]:
repo_id += f"_{mode}"
# download and load raw dataset, create LeRobotDataset, populate it, push to hub
main(raw_dir, repo_id=repo_id, mode=mode)
# Uncomment if you want to load the local dataset and explore it
# dataset = LeRobotDataset(repo_id=repo_id, local_files_only=True)
# breakpoint()

19
lerobot/__init__.py
View File

@ -58,7 +58,6 @@ available_tasks_per_env = {
],
"pusht": ["PushT-v0"],
"xarm": ["XarmLift-v0"],
"dora_aloha_real": ["DoraAloha-v0", "DoraKoch-v0", "DoraReachy2-v0"],
}
available_envs = list(available_tasks_per_env.keys())
@ -86,23 +85,6 @@ available_datasets_per_env = {
"lerobot/xarm_push_medium_image",
"lerobot/xarm_push_medium_replay_image",
],
"dora_aloha_real": [
"lerobot/aloha_static_battery",
"lerobot/aloha_static_candy",
"lerobot/aloha_static_coffee",
"lerobot/aloha_static_coffee_new",
"lerobot/aloha_static_cups_open",
"lerobot/aloha_static_fork_pick_up",
"lerobot/aloha_static_pingpong_test",
"lerobot/aloha_static_pro_pencil",
"lerobot/aloha_static_screw_driver",
"lerobot/aloha_static_tape",
"lerobot/aloha_static_thread_velcro",
"lerobot/aloha_static_towel",
"lerobot/aloha_static_vinh_cup",
"lerobot/aloha_static_vinh_cup_left",
"lerobot/aloha_static_ziploc_slide",
],
}
available_real_world_datasets = [
@ -221,7 +203,6 @@ available_policies_per_env = {
"xarm": ["tdmpc"],
"koch_real": ["act_koch_real"],
"aloha_real": ["act_aloha_real"],
"dora_aloha_real": ["act_aloha_real"],
}
env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

45
lerobot/common/constants.py Normal file
View File

@ -0,0 +1,45 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# keys
import os
from pathlib import Path
from huggingface_hub.constants import HF_HOME
OBS_ENV = "observation.environment_state"
OBS_ROBOT = "observation.state"
OBS_IMAGE = "observation.image"
OBS_IMAGES = "observation.images"
ACTION = "action"
# files & directories
CHECKPOINTS_DIR = "checkpoints"
LAST_CHECKPOINT_LINK = "last"
PRETRAINED_MODEL_DIR = "pretrained_model"
TRAINING_STATE_DIR = "training_state"
RNG_STATE = "rng_state.safetensors"
TRAINING_STEP = "training_step.json"
OPTIMIZER_STATE = "optimizer_state.safetensors"
OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
SCHEDULER_STATE = "scheduler_state.json"
# cache dir
default_cache_path = Path(HF_HOME) / "lerobot"
HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
if "LEROBOT_HOME" in os.environ:
raise ValueError(
f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"
"'LEROBOT_HOME' is deprecated, please use 'HF_LEROBOT_HOME' instead."
)

68
lerobot/common/datasets/backward_compatibility.py Normal file
View File

@ -0,0 +1,68 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import packaging.version
V2_MESSAGE = """
The dataset you requested ({repo_id}) is in {version} format.
We introduced a new format since v2.0 which is not backward compatible with v1.x.
Please, use our conversion script. Modify the following command with your own task description:
```
python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \\
--repo-id {repo_id} \\
--single-task "TASK DESCRIPTION." # <---- /!\\ Replace TASK DESCRIPTION /!\\
```
A few examples to replace TASK DESCRIPTION: "Pick up the blue cube and place it into the bin.", "Insert the
peg into the socket.", "Slide open the ziploc bag.", "Take the elevator to the 1st floor.", "Open the top
cabinet, store the pot inside it then close the cabinet.", "Push the T-shaped block onto the T-shaped
target.", "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.", "Fold the
sweatshirt.", ...
If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
"""
V21_MESSAGE = """
The dataset you requested ({repo_id}) is in {version} format.
While current version of LeRobot is backward-compatible with it, the version of your dataset still uses global
stats instead of per-episode stats. Update your dataset stats to the new format using this command:
```
python lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py --repo-id={repo_id}
```
If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
"""
FUTURE_MESSAGE = """
The dataset you requested ({repo_id}) is only available in {version} format.
As we cannot ensure forward compatibility with it, please update your current version of lerobot.
"""
class CompatibilityError(Exception): ...
class BackwardCompatibilityError(CompatibilityError):
def __init__(self, repo_id: str, version: packaging.version.Version):
message = V2_MESSAGE.format(repo_id=repo_id, version=version)
super().__init__(message)
class ForwardCompatibilityError(CompatibilityError):
def __init__(self, repo_id: str, version: packaging.version.Version):
message = FUTURE_MESSAGE.format(repo_id=repo_id, version=version)
super().__init__(message)

312
lerobot/common/datasets/compute_stats.py
View File

@ -13,202 +13,164 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from copy import deepcopy
from math import ceil
import numpy as np
import einops
import torch
import tqdm
from lerobot.common.datasets.utils import load_image_as_numpy
def get_stats_einops_patterns(dataset, num_workers=0):
"""These einops patterns will be used to aggregate batches and compute statistics.
def estimate_num_samples(
dataset_len: int, min_num_samples: int = 100, max_num_samples: int = 10_000, power: float = 0.75
) -> int:
"""Heuristic to estimate the number of samples based on dataset size.
The power controls the sample growth relative to dataset size.
Lower the power for less number of samples.
Note: We assume the images are in channel first format
For default arguments, we have:
- from 1 to ~500, num_samples=100
- at 1000, num_samples=177
- at 2000, num_samples=299
- at 5000, num_samples=594
- at 10000, num_samples=1000
- at 20000, num_samples=1681
"""
dataloader = torch.utils.data.DataLoader(
dataset,
num_workers=num_workers,
batch_size=2,
shuffle=False,
)
batch = next(iter(dataloader))
stats_patterns = {}
for key in dataset.features:
# sanity check that tensors are not float64
assert batch[key].dtype != torch.float64
# if isinstance(feats_type, (VideoFrame, Image)):
if key in dataset.meta.camera_keys:
# sanity check that images are channel first
_, c, h, w = batch[key].shape
assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
# sanity check that images are float32 in range [0,1]
assert batch[key].dtype == torch.float32, f"expect torch.float32, but instead {batch[key].dtype=}"
assert batch[key].max() <= 1, f"expect pixels lower than 1, but instead {batch[key].max()=}"
assert batch[key].min() >= 0, f"expect pixels greater than 1, but instead {batch[key].min()=}"
stats_patterns[key] = "b c h w -> c 1 1"
elif batch[key].ndim == 2:
stats_patterns[key] = "b c -> c "
elif batch[key].ndim == 1:
stats_patterns[key] = "b -> 1"
else:
raise ValueError(f"{key}, {batch[key].shape}")
return stats_patterns
if dataset_len < min_num_samples:
min_num_samples = dataset_len
return max(min_num_samples, min(int(dataset_len**power), max_num_samples))
def compute_stats(dataset, batch_size=8, num_workers=8, max_num_samples=None):
"""Compute mean/std and min/max statistics of all data keys in a LeRobotDataset."""
if max_num_samples is None:
max_num_samples = len(dataset)
def sample_indices(data_len: int) -> list[int]:
num_samples = estimate_num_samples(data_len)
return np.round(np.linspace(0, data_len - 1, num_samples)).astype(int).tolist()
# for more info on why we need to set the same number of workers, see `load_from_videos`
stats_patterns = get_stats_einops_patterns(dataset, num_workers)
# mean and std will be computed incrementally while max and min will track the running value.
mean, std, max, min = {}, {}, {}, {}
for key in stats_patterns:
mean[key] = torch.tensor(0.0).float()
std[key] = torch.tensor(0.0).float()
max[key] = torch.tensor(-float("inf")).float()
min[key] = torch.tensor(float("inf")).float()
def auto_downsample_height_width(img: np.ndarray, target_size: int = 150, max_size_threshold: int = 300):
_, height, width = img.shape
def create_seeded_dataloader(dataset, batch_size, seed):
generator = torch.Generator()
generator.manual_seed(seed)
dataloader = torch.utils.data.DataLoader(
dataset,
num_workers=num_workers,
batch_size=batch_size,
shuffle=True,
drop_last=False,
generator=generator,
)
return dataloader
if max(width, height) < max_size_threshold:
# no downsampling needed
return img
# Note: Due to be refactored soon. The point of storing `first_batch` is to make sure we don't get
# surprises when rerunning the sampler.
first_batch = None
running_item_count = 0 # for online mean computation
dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
for i, batch in enumerate(
tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute mean, min, max")
):
this_batch_size = len(batch["index"])
running_item_count += this_batch_size
if first_batch is None:
first_batch = deepcopy(batch)
for key, pattern in stats_patterns.items():
batch[key] = batch[key].float()
# Numerically stable update step for mean computation.
batch_mean = einops.reduce(batch[key], pattern, "mean")
# Hint: to update the mean we need x̄ₙ = (Nₙ₋₁x̄ₙ₋₁ + Bₙxₙ) / Nₙ, where the subscript represents
# the update step, N is the running item count, B is this batch size, x̄ is the running mean,
# and x is the current batch mean. Some rearrangement is then required to avoid risking
# numerical overflow. Another hint: Nₙ₋₁ = Nₙ - Bₙ. Rearrangement yields
# x̄ₙ = x̄ₙ₋₁ + Bₙ * (xₙ - x̄ₙ₋₁) / Nₙ
mean[key] = mean[key] + this_batch_size * (batch_mean - mean[key]) / running_item_count
max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
downsample_factor = int(width / target_size) if width > height else int(height / target_size)
return img[:, ::downsample_factor, ::downsample_factor]
if i == ceil(max_num_samples / batch_size) - 1:
break
first_batch_ = None
running_item_count = 0 # for online std computation
dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
for i, batch in enumerate(
tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute std")
):
this_batch_size = len(batch["index"])
running_item_count += this_batch_size
# Sanity check to make sure the batches are still in the same order as before.
if first_batch_ is None:
first_batch_ = deepcopy(batch)
for key in stats_patterns:
assert torch.equal(first_batch_[key], first_batch[key])
for key, pattern in stats_patterns.items():
batch[key] = batch[key].float()
# Numerically stable update step for mean computation (where the mean is over squared
# residuals).See notes in the mean computation loop above.
batch_std = einops.reduce((batch[key] - mean[key]) ** 2, pattern, "mean")
std[key] = std[key] + this_batch_size * (batch_std - std[key]) / running_item_count
def sample_images(image_paths: list[str]) -> np.ndarray:
sampled_indices = sample_indices(len(image_paths))
if i == ceil(max_num_samples / batch_size) - 1:
break
images = None
for i, idx in enumerate(sampled_indices):
path = image_paths[idx]
# we load as uint8 to reduce memory usage
img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True)
img = auto_downsample_height_width(img)
for key in stats_patterns:
std[key] = torch.sqrt(std[key])
if images is None:
images = np.empty((len(sampled_indices), *img.shape), dtype=np.uint8)
stats = {}
for key in stats_patterns:
stats[key] = {
"mean": mean[key],
"std": std[key],
"max": max[key],
"min": min[key],
images[i] = img
return images
def get_feature_stats(array: np.ndarray, axis: tuple, keepdims: bool) -> dict[str, np.ndarray]:
return {
"min": np.min(array, axis=axis, keepdims=keepdims),
"max": np.max(array, axis=axis, keepdims=keepdims),
"mean": np.mean(array, axis=axis, keepdims=keepdims),
"std": np.std(array, axis=axis, keepdims=keepdims),
"count": np.array([len(array)]),
}
return stats
def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
"""Aggregate stats of multiple LeRobot datasets into one set of stats without recomputing from scratch.
def compute_episode_stats(episode_data: dict[str, list[str] | np.ndarray], features: dict) -> dict:
ep_stats = {}
for key, data in episode_data.items():
if features[key]["dtype"] == "string":
continue # HACK: we should receive np.arrays of strings
elif features[key]["dtype"] in ["image", "video"]:
ep_ft_array = sample_images(data) # data is a list of image paths
axes_to_reduce = (0, 2, 3) # keep channel dim
keepdims = True
else:
ep_ft_array = data # data is already a np.ndarray
axes_to_reduce = 0 # compute stats over the first axis
keepdims = data.ndim == 1 # keep as np.array
The final stats will have the union of all data keys from each of the datasets.
ep_stats[key] = get_feature_stats(ep_ft_array, axis=axes_to_reduce, keepdims=keepdims)
The final stats will have the union of all data keys from each of the datasets. For instance:
- new_max = max(max_dataset_0, max_dataset_1, ...)
# finally, we normalize and remove batch dim for images
if features[key]["dtype"] in ["image", "video"]:
ep_stats[key] = {
k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items()
}
return ep_stats
def _assert_type_and_shape(stats_list: list[dict[str, dict]]):
for i in range(len(stats_list)):
for fkey in stats_list[i]:
for k, v in stats_list[i][fkey].items():
if not isinstance(v, np.ndarray):
raise ValueError(
f"Stats must be composed of numpy array, but key '{k}' of feature '{fkey}' is of type '{type(v)}' instead."
)
if v.ndim == 0:
raise ValueError("Number of dimensions must be at least 1, and is 0 instead.")
if k == "count" and v.shape != (1,):
raise ValueError(f"Shape of 'count' must be (1), but is {v.shape} instead.")
if "image" in fkey and k != "count" and v.shape != (3, 1, 1):
raise ValueError(f"Shape of '{k}' must be (3,1,1), but is {v.shape} instead.")
def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
"""Aggregates stats for a single feature."""
means = np.stack([s["mean"] for s in stats_ft_list])
variances = np.stack([s["std"] ** 2 for s in stats_ft_list])
counts = np.stack([s["count"] for s in stats_ft_list])
total_count = counts.sum(axis=0)
# Prepare weighted mean by matching number of dimensions
while counts.ndim < means.ndim:
counts = np.expand_dims(counts, axis=-1)
# Compute the weighted mean
weighted_means = means * counts
total_mean = weighted_means.sum(axis=0) / total_count
# Compute the variance using the parallel algorithm
delta_means = means - total_mean
weighted_variances = (variances + delta_means**2) * counts
total_variance = weighted_variances.sum(axis=0) / total_count
return {
"min": np.min(np.stack([s["min"] for s in stats_ft_list]), axis=0),
"max": np.max(np.stack([s["max"] for s in stats_ft_list]), axis=0),
"mean": total_mean,
"std": np.sqrt(total_variance),
"count": total_count,
}
def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
"""Aggregate stats from multiple compute_stats outputs into a single set of stats.
The final stats will have the union of all data keys from each of the stats dicts.
For instance:
- new_min = min(min_dataset_0, min_dataset_1, ...)
- new_mean = (mean of all data)
- new_max = max(max_dataset_0, max_dataset_1, ...)
- new_mean = (mean of all data, weighted by counts)
- new_std = (std of all data)
"""
data_keys = set()
for dataset in ls_datasets:
data_keys.update(dataset.meta.stats.keys())
stats = {k: {} for k in data_keys}
for data_key in data_keys:
for stat_key in ["min", "max"]:
# compute `max(dataset_0["max"], dataset_1["max"], ...)`
stats[data_key][stat_key] = einops.reduce(
torch.stack(
[ds.meta.stats[data_key][stat_key] for ds in ls_datasets if data_key in ds.meta.stats],
dim=0,
),
"n ... -> ...",
stat_key,
)
total_samples = sum(d.num_frames for d in ls_datasets if data_key in d.meta.stats)
# Compute the "sum" statistic by multiplying each mean by the number of samples in the respective
# dataset, then divide by total_samples to get the overall "mean".
# NOTE: the brackets around (d.num_frames / total_samples) are needed tor minimize the risk of
# numerical overflow!
stats[data_key]["mean"] = sum(
d.meta.stats[data_key]["mean"] * (d.num_frames / total_samples)
for d in ls_datasets
if data_key in d.meta.stats
)
# The derivation for standard deviation is a little more involved but is much in the same spirit as
# the computation of the mean.
# Given two sets of data where the statistics are known:
# σ_combined = sqrt[ (n1 * (σ1^2 + d1^2) + n2 * (σ2^2 + d2^2)) / (n1 + n2) ]
# where d1 = μ1 - μ_combined, d2 = μ2 - μ_combined
# NOTE: the brackets around (d.num_frames / total_samples) are needed tor minimize the risk of
# numerical overflow!
stats[data_key]["std"] = torch.sqrt(
sum(
(
d.meta.stats[data_key]["std"] ** 2
+ (d.meta.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2
)
* (d.num_frames / total_samples)
for d in ls_datasets
if data_key in d.meta.stats
)
)
return stats
_assert_type_and_shape(stats_list)
data_keys = {key for stats in stats_list for key in stats}
aggregated_stats = {key: {} for key in data_keys}
for key in data_keys:
stats_with_key = [stats[key] for stats in stats_list if key in stats]
aggregated_stats[key] = aggregate_feature_stats(stats_with_key)
return aggregated_stats

156
lerobot/common/datasets/factory.py
View File

@ -14,103 +14,105 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from pprint import pformat
import torch
from omegaconf import ListConfig, OmegaConf
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
from lerobot.common.datasets.transforms import get_image_transforms
from lerobot.common.datasets.lerobot_dataset import (
LeRobotDataset,
LeRobotDatasetMetadata,
MultiLeRobotDataset,
)
from lerobot.common.datasets.transforms import ImageTransforms
from lerobot.configs.policies import PreTrainedConfig
from lerobot.configs.train import TrainPipelineConfig
IMAGENET_STATS = {
"mean": [[[0.485]], [[0.456]], [[0.406]]], # (c,1,1)
"std": [[[0.229]], [[0.224]], [[0.225]]], # (c,1,1)
}
def resolve_delta_timestamps(cfg):
"""Resolves delta_timestamps config key (in-place) by using `eval`.
def resolve_delta_timestamps(
cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
) -> dict[str, list] | None:
"""Resolves delta_timestamps by reading from the 'delta_indices' properties of the PreTrainedConfig.
Doesn't do anything if delta_timestamps is not specified or has already been resolve (as evidenced by
the data type of its values).
"""
delta_timestamps = cfg.training.get("delta_timestamps")
if delta_timestamps is not None:
for key in delta_timestamps:
if isinstance(delta_timestamps[key], str):
# TODO(rcadene, alexander-soare): remove `eval` to avoid exploit
cfg.training.delta_timestamps[key] = eval(delta_timestamps[key])
def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotDataset:
"""
Args:
cfg: A Hydra config as per the LeRobot config scheme.
split: Select the data subset used to create an instance of LeRobotDataset.
All datasets hosted on [lerobot](https://huggingface.co/lerobot) contain only one subset: "train".
Thus, by default, `split="train"` selects all the available data. `split` aims to work like the
slicer in the hugging face datasets:
https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
As of now, it only supports `split="train[:n]"` to load the first n frames of the dataset or
`split="train[n:]"` to load the last n frames. For instance `split="train[:1000]"`.
cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
delta_timestamps against.
Returns:
The LeRobotDataset.
dict[str, list] | None: A dictionary of delta_timestamps, e.g.:
{
"observation.state": [-0.04, -0.02, 0]
"observation.action": [-0.02, 0, 0.02]
}
returns `None` if the the resulting dict is empty.
"""
if not isinstance(cfg.dataset_repo_id, (str, ListConfig)):
raise ValueError(
"Expected cfg.dataset_repo_id to be either a single string to load one dataset or a list of "
"strings to load multiple datasets."
delta_timestamps = {}
for key in ds_meta.features:
if key == "next.reward" and cfg.reward_delta_indices is not None:
delta_timestamps[key] = [i / ds_meta.fps for i in cfg.reward_delta_indices]
if key == "action" and cfg.action_delta_indices is not None:
delta_timestamps[key] = [i / ds_meta.fps for i in cfg.action_delta_indices]
if key.startswith("observation.") and cfg.observation_delta_indices is not None:
delta_timestamps[key] = [i / ds_meta.fps for i in cfg.observation_delta_indices]
if len(delta_timestamps) == 0:
delta_timestamps = None
return delta_timestamps
def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
"""Handles the logic of setting up delta timestamps and image transforms before creating a dataset.
Args:
cfg (TrainPipelineConfig): A TrainPipelineConfig config which contains a DatasetConfig and a PreTrainedConfig.
Raises:
NotImplementedError: The MultiLeRobotDataset is currently deactivated.
Returns:
LeRobotDataset | MultiLeRobotDataset
"""
image_transforms = (
ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
)
# A soft check to warn if the environment matches the dataset. Don't check if we are using a real world env (dora).
if cfg.env.name != "dora":
if isinstance(cfg.dataset_repo_id, str):
dataset_repo_ids = [cfg.dataset_repo_id] # single dataset
else:
dataset_repo_ids = cfg.dataset_repo_id # multiple datasets
for dataset_repo_id in dataset_repo_ids:
if cfg.env.name not in dataset_repo_id:
logging.warning(
f"There might be a mismatch between your training dataset ({dataset_repo_id=}) and your "
f"environment ({cfg.env.name=})."
if isinstance(cfg.dataset.repo_id, str):
ds_meta = LeRobotDatasetMetadata(
cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
)
resolve_delta_timestamps(cfg)
image_transforms = None
if cfg.training.image_transforms.enable:
cfg_tf = cfg.training.image_transforms
image_transforms = get_image_transforms(
brightness_weight=cfg_tf.brightness.weight,
brightness_min_max=cfg_tf.brightness.min_max,
contrast_weight=cfg_tf.contrast.weight,
contrast_min_max=cfg_tf.contrast.min_max,
saturation_weight=cfg_tf.saturation.weight,
saturation_min_max=cfg_tf.saturation.min_max,
hue_weight=cfg_tf.hue.weight,
hue_min_max=cfg_tf.hue.min_max,
sharpness_weight=cfg_tf.sharpness.weight,
sharpness_min_max=cfg_tf.sharpness.min_max,
max_num_transforms=cfg_tf.max_num_transforms,
random_order=cfg_tf.random_order,
)
if isinstance(cfg.dataset_repo_id, str):
# TODO (aliberts): add 'episodes' arg from config after removing hydra
delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
dataset = LeRobotDataset(
cfg.dataset_repo_id,
delta_timestamps=cfg.training.get("delta_timestamps"),
cfg.dataset.repo_id,
root=cfg.dataset.root,
episodes=cfg.dataset.episodes,
delta_timestamps=delta_timestamps,
image_transforms=image_transforms,
video_backend=cfg.video_backend,
revision=cfg.dataset.revision,
video_backend=cfg.dataset.video_backend,
)
else:
raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
dataset = MultiLeRobotDataset(
cfg.dataset_repo_id,
delta_timestamps=cfg.training.get("delta_timestamps"),
cfg.dataset.repo_id,
# TODO(aliberts): add proper support for multi dataset
# delta_timestamps=delta_timestamps,
image_transforms=image_transforms,
video_backend=cfg.video_backend,
video_backend=cfg.dataset.video_backend,
)
logging.info(
"Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
f"{pformat(dataset.repo_id_to_index, indent=2)}"
)
if cfg.get("override_dataset_stats"):
for key, stats_dict in cfg.override_dataset_stats.items():
for stats_type, listconfig in stats_dict.items():
# example of stats_type: min, max, mean, std
stats = OmegaConf.to_container(listconfig, resolve=True)
if cfg.dataset.use_imagenet_stats:
for key in dataset.meta.camera_keys:
for stats_type, stats in IMAGENET_STATS.items():
dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
return dataset

28
lerobot/common/datasets/image_writer.py
View File

@ -38,22 +38,40 @@ def safe_stop_image_writer(func):
return wrapper
def image_array_to_image(image_array: np.ndarray) -> PIL.Image.Image:
def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image:
# TODO(aliberts): handle 1 channel and 4 for depth images
if image_array.ndim == 3 and image_array.shape[0] in [1, 3]:
if image_array.ndim != 3:
raise ValueError(f"The array has {image_array.ndim} dimensions, but 3 is expected for an image.")
if image_array.shape[0] == 3:
# Transpose from pytorch convention (C, H, W) to (H, W, C)
image_array = image_array.transpose(1, 2, 0)
elif image_array.shape[-1] != 3:
raise NotImplementedError(
f"The image has {image_array.shape[-1]} channels, but 3 is required for now."
)
if image_array.dtype != np.uint8:
# Assume the image is in [0, 1] range for floating-point data
image_array = np.clip(image_array, 0, 1)
if range_check:
max_ = image_array.max().item()
min_ = image_array.min().item()
if max_ > 1.0 or min_ < 0.0:
raise ValueError(
"The image data type is float, which requires values in the range [0.0, 1.0]. "
f"However, the provided range is [{min_}, {max_}]. Please adjust the range or "
"provide a uint8 image with values in the range [0, 255]."
)
image_array = (image_array * 255).astype(np.uint8)
return PIL.Image.fromarray(image_array)
def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path):
try:
if isinstance(image, np.ndarray):
img = image_array_to_image(image)
img = image_array_to_pil_image(image)
elif isinstance(image, PIL.Image.Image):
img = image
else:

474
lerobot/common/datasets/lerobot_dataset.py
View File

@ -13,62 +13,68 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import contextlib
import logging
import os
import shutil
from functools import cached_property
from pathlib import Path
from typing import Callable
import datasets
import numpy as np
import packaging.version
import PIL.Image
import torch
import torch.utils
from datasets import load_dataset
from huggingface_hub import create_repo, snapshot_download, upload_folder
from datasets import concatenate_datasets, load_dataset
from huggingface_hub import HfApi, snapshot_download
from huggingface_hub.constants import REPOCARD_NAME
from huggingface_hub.errors import RevisionNotFoundError
from lerobot.common.datasets.compute_stats import aggregate_stats, compute_stats
from lerobot.common.constants import HF_LEROBOT_HOME
from lerobot.common.datasets.compute_stats import aggregate_stats, compute_episode_stats
from lerobot.common.datasets.image_writer import AsyncImageWriter, write_image
from lerobot.common.datasets.utils import (
DEFAULT_FEATURES,
DEFAULT_IMAGE_PATH,
EPISODES_PATH,
INFO_PATH,
STATS_PATH,
TASKS_PATH,
append_jsonlines,
backward_compatible_episodes_stats,
check_delta_timestamps,
check_timestamps_sync,
check_version_compatibility,
create_branch,
create_empty_dataset_info,
create_lerobot_dataset_card,
embed_images,
get_delta_indices,
get_episode_data_index,
get_features_from_robot,
get_hf_features_from_features,
get_hub_safe_version,
get_safe_version,
hf_transform_to_torch,
is_valid_version,
load_episodes,
load_episodes_stats,
load_info,
load_stats,
load_tasks,
serialize_dict,
validate_episode_buffer,
validate_frame,
write_episode,
write_episode_stats,
write_info,
write_json,
write_parquet,
)
from lerobot.common.datasets.video_utils import (
VideoFrame,
decode_video_frames_torchvision,
decode_video_frames,
encode_video_frames,
get_safe_default_codec,
get_video_info,
)
from lerobot.common.robot_devices.robots.utils import Robot
# For maintainers, see lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md
CODEBASE_VERSION = "v2.0"
LEROBOT_HOME = Path(os.getenv("LEROBOT_HOME", "~/.cache/huggingface/lerobot")).expanduser()
CODEBASE_VERSION = "v2.1"
class LeRobotDatasetMetadata:
@ -76,19 +82,36 @@ class LeRobotDatasetMetadata:
self,
repo_id: str,
root: str | Path | None = None,
local_files_only: bool = False,
revision: str | None = None,
force_cache_sync: bool = False,
):
self.repo_id = repo_id
self.root = Path(root) if root is not None else LEROBOT_HOME / repo_id
self.local_files_only = local_files_only
self.revision = revision if revision else CODEBASE_VERSION
self.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
try:
if force_cache_sync:
raise FileNotFoundError
self.load_metadata()
except (FileNotFoundError, NotADirectoryError):
if is_valid_version(self.revision):
self.revision = get_safe_version(self.repo_id, self.revision)
# Load metadata
(self.root / "meta").mkdir(exist_ok=True, parents=True)
self.pull_from_repo(allow_patterns="meta/")
self.load_metadata()
def load_metadata(self):
self.info = load_info(self.root)
self.stats = load_stats(self.root)
self.tasks = load_tasks(self.root)
check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
self.tasks, self.task_to_task_index = load_tasks(self.root)
self.episodes = load_episodes(self.root)
if self._version < packaging.version.parse("v2.1"):
self.stats = load_stats(self.root)
self.episodes_stats = backward_compatible_episodes_stats(self.stats, self.episodes)
else:
self.episodes_stats = load_episodes_stats(self.root)
self.stats = aggregate_stats(list(self.episodes_stats.values()))
def pull_from_repo(
self,
@ -98,21 +121,16 @@ class LeRobotDatasetMetadata:
snapshot_download(
self.repo_id,
repo_type="dataset",
revision=self._hub_version,
revision=self.revision,
local_dir=self.root,
allow_patterns=allow_patterns,
ignore_patterns=ignore_patterns,
local_files_only=self.local_files_only,
)
@cached_property
def _hub_version(self) -> str | None:
return None if self.local_files_only else get_hub_safe_version(self.repo_id, CODEBASE_VERSION)
@property
def _version(self) -> str:
def _version(self) -> packaging.version.Version:
"""Codebase version used to create this dataset."""
return self.info["codebase_version"]
return packaging.version.parse(self.info["codebase_version"])
def get_data_file_path(self, ep_index: int) -> Path:
ep_chunk = self.get_episode_chunk(ep_index)
@ -202,54 +220,65 @@ class LeRobotDatasetMetadata:
"""Max number of episodes per chunk."""
return self.info["chunks_size"]
@property
def task_to_task_index(self) -> dict:
return {task: task_idx for task_idx, task in self.tasks.items()}
def get_task_index(self, task: str) -> int:
def get_task_index(self, task: str) -> int | None:
"""
Given a task in natural language, returns its task_index if the task already exists in the dataset,
otherwise creates a new task_index.
otherwise return None.
"""
task_index = self.task_to_task_index.get(task, None)
return task_index if task_index is not None else self.total_tasks
return self.task_to_task_index.get(task, None)
def save_episode(self, episode_index: int, episode_length: int, task: str, task_index: int) -> None:
self.info["total_episodes"] += 1
self.info["total_frames"] += episode_length
def add_task(self, task: str):
"""
Given a task in natural language, add it to the dictionary of tasks.
"""
if task in self.task_to_task_index:
raise ValueError(f"The task '{task}' already exists and can't be added twice.")
if task_index not in self.tasks:
self.info["total_tasks"] += 1
task_index = self.info["total_tasks"]
self.task_to_task_index[task] = task_index
self.tasks[task_index] = task
self.info["total_tasks"] += 1
task_dict = {
"task_index": task_index,
"task": task,
}
append_jsonlines(task_dict, self.root / TASKS_PATH)
def save_episode(
self,
episode_index: int,
episode_length: int,
episode_tasks: list[str],
episode_stats: dict[str, dict],
) -> None:
self.info["total_episodes"] += 1
self.info["total_frames"] += episode_length
chunk = self.get_episode_chunk(episode_index)
if chunk >= self.total_chunks:
self.info["total_chunks"] += 1
self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
self.info["total_videos"] += len(self.video_keys)
write_json(self.info, self.root / INFO_PATH)
if len(self.video_keys) > 0:
self.update_video_info()
write_info(self.info, self.root)
episode_dict = {
"episode_index": episode_index,
"tasks": [task],
"tasks": episode_tasks,
"length": episode_length,
}
self.episodes.append(episode_dict)
append_jsonlines(episode_dict, self.root / EPISODES_PATH)
self.episodes[episode_index] = episode_dict
write_episode(episode_dict, self.root)
# TODO(aliberts): refactor stats in save_episodes
# image_sampling = int(self.fps / 2) # sample 2 img/s for the stats
# ep_stats = compute_episode_stats(episode_buffer, self.features, episode_length, image_sampling=image_sampling)
# ep_stats = serialize_dict(ep_stats)
# append_jsonlines(ep_stats, self.root / STATS_PATH)
self.episodes_stats[episode_index] = episode_stats
self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats else episode_stats
write_episode_stats(episode_index, episode_stats, self.root)
def write_video_info(self) -> None:
def update_video_info(self) -> None:
"""
Warning: this function writes info from first episode videos, implicitly assuming that all videos have
been encoded the same way. Also, this means it assumes the first episode exists.
@ -259,8 +288,6 @@ class LeRobotDatasetMetadata:
video_path = self.root / self.get_video_file_path(ep_index=0, vid_key=key)
self.info["features"][key]["info"] = get_video_info(video_path)
write_json(self.info, self.root / INFO_PATH)
def __repr__(self):
feature_keys = list(self.features)
return (
@ -286,7 +313,7 @@ class LeRobotDatasetMetadata:
"""Creates metadata for a LeRobotDataset."""
obj = cls.__new__(cls)
obj.repo_id = repo_id
obj.root = Path(root) if root is not None else LEROBOT_HOME / repo_id
obj.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
obj.root.mkdir(parents=True, exist_ok=False)
@ -306,12 +333,21 @@ class LeRobotDatasetMetadata:
# TODO(aliberts, rcadene): implement sanity check for features
features = {**features, **DEFAULT_FEATURES}
obj.tasks, obj.stats, obj.episodes = {}, {}, []
# check if none of the features contains a "/" in their names,
# as this would break the dict flattening in the stats computation, which uses '/' as separator
for key in features:
if "/" in key:
raise ValueError(f"Feature names should not contain '/'. Found '/' in feature '{key}'.")
features = {**features, **DEFAULT_FEATURES}
obj.tasks, obj.task_to_task_index = {}, {}
obj.episodes_stats, obj.stats, obj.episodes = {}, {}, {}
obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, robot_type, features, use_videos)
if len(obj.video_keys) > 0 and not use_videos:
raise ValueError()
write_json(obj.info, obj.root / INFO_PATH)
obj.local_files_only = True
obj.revision = None
return obj
@ -324,8 +360,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
image_transforms: Callable | None = None,
delta_timestamps: dict[list[float]] | None = None,
tolerance_s: float = 1e-4,
revision: str | None = None,
force_cache_sync: bool = False,
download_videos: bool = True,
local_files_only: bool = False,
video_backend: str | None = None,
):
"""
@ -335,7 +372,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
- On your local disk in the 'root' folder. This is typically the case when you recorded your
dataset locally and you may or may not have pushed it to the hub yet. Instantiating this class
with 'root' will load your dataset directly from disk. This can happen while you're offline (no
internet connection), in that case, use local_files_only=True.
internet connection).
- On the Hugging Face Hub at the address https://huggingface.co/datasets/{repo_id} and not on
your local disk in the 'root' folder. Instantiating this class with this 'repo_id' will download
@ -355,7 +392,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
- info contains various information about the dataset like shapes, keys, fps etc.
- stats stores the dataset statistics of the different modalities for normalization
- tasks contains the prompts for each task of the dataset, which can be used for
task-conditionned training.
task-conditioned training.
- hf_dataset (from datasets.Dataset), which will read any values from parquet files.
- videos (optional) from which frames are loaded to be synchronous with data from parquet files.
@ -417,24 +454,28 @@ class LeRobotDataset(torch.utils.data.Dataset):
timestamps is separated to the next by 1/fps +/- tolerance_s. This also applies to frames
decoded from video files. It is also used to check that `delta_timestamps` (when provided) are
multiples of 1/fps. Defaults to 1e-4.
revision (str, optional): An optional Git revision id which can be a branch name, a tag, or a
commit hash. Defaults to current codebase version tag.
sync_cache_first (bool, optional): Flag to sync and refresh local files first. If True and files
are already present in the local cache, this will be faster. However, files loaded might not
be in sync with the version on the hub, especially if you specified 'revision'. Defaults to
False.
download_videos (bool, optional): Flag to download the videos. Note that when set to True but the
video files are already present on local disk, they won't be downloaded again. Defaults to
True.
local_files_only (bool, optional): Flag to use local files only. If True, no requests to the hub
will be made. Defaults to False.
video_backend (str | None, optional): Video backend to use for decoding videos. There is currently
a single option which is the pyav decoder used by Torchvision. Defaults to pyav.
video_backend (str | None, optional): Video backend to use for decoding videos. Defaults to torchcodec when available int the platform; otherwise, defaults to 'pyav'.
You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
"""
super().__init__()
self.repo_id = repo_id
self.root = Path(root) if root else LEROBOT_HOME / repo_id
self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
self.image_transforms = image_transforms
self.delta_timestamps = delta_timestamps
self.episodes = episodes
self.tolerance_s = tolerance_s
self.video_backend = video_backend if video_backend else "pyav"
self.revision = revision if revision else CODEBASE_VERSION
self.video_backend = video_backend if video_backend else get_safe_default_codec()
self.delta_indices = None
self.local_files_only = local_files_only
# Unused attributes
self.image_writer = None
@ -443,64 +484,92 @@ class LeRobotDataset(torch.utils.data.Dataset):
self.root.mkdir(exist_ok=True, parents=True)
# Load metadata
self.meta = LeRobotDatasetMetadata(self.repo_id, self.root, self.local_files_only)
# Check version
check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION)
self.meta = LeRobotDatasetMetadata(
self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
)
if self.episodes is not None and self.meta._version >= packaging.version.parse("v2.1"):
episodes_stats = [self.meta.episodes_stats[ep_idx] for ep_idx in self.episodes]
self.stats = aggregate_stats(episodes_stats)
# Load actual data
try:
if force_cache_sync:
raise FileNotFoundError
assert all((self.root / fpath).is_file() for fpath in self.get_episodes_file_paths())
self.hf_dataset = self.load_hf_dataset()
except (AssertionError, FileNotFoundError, NotADirectoryError):
self.revision = get_safe_version(self.repo_id, self.revision)
self.download_episodes(download_videos)
self.hf_dataset = self.load_hf_dataset()
self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
# Check timestamps
check_timestamps_sync(self.hf_dataset, self.episode_data_index, self.fps, self.tolerance_s)
timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy()
episode_indices = torch.stack(self.hf_dataset["episode_index"]).numpy()
ep_data_index_np = {k: t.numpy() for k, t in self.episode_data_index.items()}
check_timestamps_sync(timestamps, episode_indices, ep_data_index_np, self.fps, self.tolerance_s)
# Setup delta_indices
if self.delta_timestamps is not None:
check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
# Available stats implies all videos have been encoded and dataset is iterable
self.consolidated = self.meta.stats is not None
def push_to_hub(
self,
branch: str | None = None,
tags: list | None = None,
license: str | None = "apache-2.0",
tag_version: bool = True,
push_videos: bool = True,
private: bool = False,
allow_patterns: list[str] | str | None = None,
upload_large_folder: bool = False,
**card_kwargs,
) -> None:
if not self.consolidated:
logging.warning(
"You are trying to upload to the hub a LeRobotDataset that has not been consolidated yet. "
"Consolidating first."
)
self.consolidate()
ignore_patterns = ["images/"]
if not push_videos:
ignore_patterns.append("videos/")
create_repo(
hub_api = HfApi()
hub_api.create_repo(
repo_id=self.repo_id,
private=private,
repo_type="dataset",
exist_ok=True,
)
upload_folder(
if branch:
hub_api.create_branch(
repo_id=self.repo_id,
folder_path=self.root,
branch=branch,
revision=self.revision,
repo_type="dataset",
ignore_patterns=ignore_patterns,
exist_ok=True,
)
upload_kwargs = {
"repo_id": self.repo_id,
"folder_path": self.root,
"repo_type": "dataset",
"revision": branch,
"allow_patterns": allow_patterns,
"ignore_patterns": ignore_patterns,
}
if upload_large_folder:
hub_api.upload_large_folder(**upload_kwargs)
else:
hub_api.upload_folder(**upload_kwargs)
if not hub_api.file_exists(self.repo_id, REPOCARD_NAME, repo_type="dataset", revision=branch):
card = create_lerobot_dataset_card(
tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
)
card.push_to_hub(repo_id=self.repo_id, repo_type="dataset")
create_branch(repo_id=self.repo_id, branch=CODEBASE_VERSION, repo_type="dataset")
card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
if tag_version:
with contextlib.suppress(RevisionNotFoundError):
hub_api.delete_tag(self.repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
hub_api.create_tag(self.repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
def pull_from_repo(
self,
@ -510,11 +579,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
snapshot_download(
self.repo_id,
repo_type="dataset",
revision=self.meta._hub_version,
revision=self.revision,
local_dir=self.root,
allow_patterns=allow_patterns,
ignore_patterns=ignore_patterns,
local_files_only=self.local_files_only,
)
def download_episodes(self, download_videos: bool = True) -> None:
@ -528,16 +596,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
files = None
ignore_patterns = None if download_videos else "videos/"
if self.episodes is not None:
files = [str(self.meta.get_data_file_path(ep_idx)) for ep_idx in self.episodes]
if len(self.meta.video_keys) > 0 and download_videos:
files = self.get_episodes_file_paths()
self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
def get_episodes_file_paths(self) -> list[Path]:
episodes = self.episodes if self.episodes is not None else list(range(self.meta.total_episodes))
fpaths = [str(self.meta.get_data_file_path(ep_idx)) for ep_idx in episodes]
if len(self.meta.video_keys) > 0:
video_files = [
str(self.meta.get_video_file_path(ep_idx, vid_key))
for vid_key in self.meta.video_keys
for ep_idx in self.episodes
for ep_idx in episodes
]
files += video_files
fpaths += video_files
self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
return fpaths
def load_hf_dataset(self) -> datasets.Dataset:
"""hf_dataset contains all the observations, states, actions, rewards, etc."""
@ -550,7 +624,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
# TODO(aliberts): hf_dataset.set_format("torch")
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def create_hf_dataset(self) -> datasets.Dataset:
features = get_hf_features_from_features(self.features)
ft_dict = {col: [] for col in features}
hf_dataset = datasets.Dataset.from_dict(ft_dict, features=features, split="train")
# TODO(aliberts): hf_dataset.set_format("torch")
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
@property
@ -617,7 +699,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
if key not in self.meta.video_keys
}
def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict:
def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict[str, torch.Tensor]:
"""Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a
Segmentation Fault. This probably happens because a memory reference to the video loader is created in
@ -626,9 +708,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
item = {}
for vid_key, query_ts in query_timestamps.items():
video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
frames = decode_video_frames_torchvision(
video_path, query_ts, self.tolerance_s, self.video_backend
)
frames = decode_video_frames(video_path, query_ts, self.tolerance_s, self.video_backend)
item[vid_key] = frames.squeeze(0)
return item
@ -647,8 +727,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
query_indices = None
if self.delta_indices is not None:
current_ep_idx = self.episodes.index(ep_idx) if self.episodes is not None else ep_idx
query_indices, padding = self._get_query_indices(idx, current_ep_idx)
query_indices, padding = self._get_query_indices(idx, ep_idx)
query_result = self._query_hf_dataset(query_indices)
item = {**item, **padding}
for key, val in query_result.items():
@ -665,6 +744,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
for cam in image_keys:
item[cam] = self.image_transforms(item[cam])
# Add task as a string
task_idx = item["task_index"].item()
item["task"] = self.meta.tasks[task_idx]
return item
def __repr__(self):
@ -680,10 +763,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
def create_episode_buffer(self, episode_index: int | None = None) -> dict:
current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
return {
"size": 0,
**{key: current_ep_idx if key == "episode_index" else [] for key in self.features},
}
ep_buffer = {}
# size and task are special cases that are not in self.features
ep_buffer["size"] = 0
ep_buffer["task"] = []
for key in self.features:
ep_buffer[key] = current_ep_idx if key == "episode_index" else []
return ep_buffer
def _get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path:
fpath = DEFAULT_IMAGE_PATH.format(
@ -705,25 +791,35 @@ class LeRobotDataset(torch.utils.data.Dataset):
temporary directory — nothing is written to disk. To save those frames, the 'save_episode()' method
then needs to be called.
"""
# TODO(aliberts, rcadene): Add sanity check for the input, check it's numpy or torch,
# check the dtype and shape matches, etc.
# Convert torch to numpy if needed
for name in frame:
if isinstance(frame[name], torch.Tensor):
frame[name] = frame[name].numpy()
validate_frame(frame, self.features)
if self.episode_buffer is None:
self.episode_buffer = self.create_episode_buffer()
# Automatically add frame_index and timestamp to episode buffer
frame_index = self.episode_buffer["size"]
timestamp = frame.pop("timestamp") if "timestamp" in frame else frame_index / self.fps
self.episode_buffer["frame_index"].append(frame_index)
self.episode_buffer["timestamp"].append(timestamp)
# Add frame features to episode_buffer
for key in frame:
if key not in self.features:
raise ValueError(key)
if key == "task":
# Note: we associate the task in natural language to its task index during `save_episode`
self.episode_buffer["task"].append(frame["task"])
continue
if self.features[key]["dtype"] not in ["image", "video"]:
item = frame[key].numpy() if isinstance(frame[key], torch.Tensor) else frame[key]
self.episode_buffer[key].append(item)
elif self.features[key]["dtype"] in ["image", "video"]:
if key not in self.features:
raise ValueError(
f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
)
if self.features[key]["dtype"] in ["image", "video"]:
img_path = self._get_image_file_path(
episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
)
@ -731,80 +827,95 @@ class LeRobotDataset(torch.utils.data.Dataset):
img_path.parent.mkdir(parents=True, exist_ok=True)
self._save_image(frame[key], img_path)
self.episode_buffer[key].append(str(img_path))
else:
self.episode_buffer[key].append(frame[key])
self.episode_buffer["size"] += 1
def save_episode(self, task: str, encode_videos: bool = True, episode_data: dict | None = None) -> None:
def save_episode(self, episode_data: dict | None = None) -> None:
"""
This will save to disk the current episode in self.episode_buffer. Note that since it affects files on
disk, it sets self.consolidated to False to ensure proper consolidation later on before uploading to
the hub.
This will save to disk the current episode in self.episode_buffer.
Use 'encode_videos' if you want to encode videos during the saving of this episode. Otherwise,
you can do it later with dataset.consolidate(). This is to give more flexibility on when to spend
time for video encoding.
Args:
episode_data (dict | None, optional): Dict containing the episode data to save. If None, this will
save the current episode in self.episode_buffer, which is filled with 'add_frame'. Defaults to
None.
"""
if not episode_data:
episode_buffer = self.episode_buffer
validate_episode_buffer(episode_buffer, self.meta.total_episodes, self.features)
# size and task are special cases that won't be added to hf_dataset
episode_length = episode_buffer.pop("size")
tasks = episode_buffer.pop("task")
episode_tasks = list(set(tasks))
episode_index = episode_buffer["episode_index"]
if episode_index != self.meta.total_episodes:
# TODO(aliberts): Add option to use existing episode_index
raise NotImplementedError(
"You might have manually provided the episode_buffer with an episode_index that doesn't "
"match the total number of episodes in the dataset. This is not supported for now."
)
if episode_length == 0:
raise ValueError(
"You must add one or several frames with `add_frame` before calling `add_episode`."
)
episode_buffer["index"] = np.arange(self.meta.total_frames, self.meta.total_frames + episode_length)
episode_buffer["episode_index"] = np.full((episode_length,), episode_index)
# Add new tasks to the tasks dictionary
for task in episode_tasks:
task_index = self.meta.get_task_index(task)
if task_index is None:
self.meta.add_task(task)
if not set(episode_buffer.keys()) == set(self.features):
raise ValueError()
# Given tasks in natural language, find their corresponding task indices
episode_buffer["task_index"] = np.array([self.meta.get_task_index(task) for task in tasks])
for key, ft in self.features.items():
if key == "index":
episode_buffer[key] = np.arange(
self.meta.total_frames, self.meta.total_frames + episode_length
)
elif key == "episode_index":
episode_buffer[key] = np.full((episode_length,), episode_index)
elif key == "task_index":
episode_buffer[key] = np.full((episode_length,), task_index)
elif ft["dtype"] in ["image", "video"]:
# index, episode_index, task_index are already processed above, and image and video
# are processed separately by storing image path and frame info as meta data
if key in ["index", "episode_index", "task_index"] or ft["dtype"] in ["image", "video"]:
continue
elif len(ft["shape"]) == 1 and ft["shape"][0] == 1:
episode_buffer[key] = np.array(episode_buffer[key], dtype=ft["dtype"])
elif len(ft["shape"]) == 1 and ft["shape"][0] > 1:
episode_buffer[key] = np.stack(episode_buffer[key])
else:
raise ValueError(key)
self._wait_image_writer()
self._save_episode_table(episode_buffer, episode_index)
ep_stats = compute_episode_stats(episode_buffer, self.features)
self.meta.save_episode(episode_index, episode_length, task, task_index)
if encode_videos and len(self.meta.video_keys) > 0:
if len(self.meta.video_keys) > 0:
video_paths = self.encode_episode_videos(episode_index)
for key in self.meta.video_keys:
episode_buffer[key] = video_paths[key]
# `meta.save_episode` be executed after encoding the videos
self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats)
ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
check_timestamps_sync(
episode_buffer["timestamp"],
episode_buffer["episode_index"],
ep_data_index_np,
self.fps,
self.tolerance_s,
)
video_files = list(self.root.rglob("*.mp4"))
assert len(video_files) == self.num_episodes * len(self.meta.video_keys)
parquet_files = list(self.root.rglob("*.parquet"))
assert len(parquet_files) == self.num_episodes
# delete images
img_dir = self.root / "images"
if img_dir.is_dir():
shutil.rmtree(self.root / "images")
if not episode_data: # Reset the buffer
self.episode_buffer = self.create_episode_buffer()
self.consolidated = False
def _save_episode_table(self, episode_buffer: dict, episode_index: int) -> None:
episode_dict = {key: episode_buffer[key] for key in self.hf_features}
ep_dataset = datasets.Dataset.from_dict(episode_dict, features=self.hf_features, split="train")
ep_dataset = embed_images(ep_dataset)
self.hf_dataset = concatenate_datasets([self.hf_dataset, ep_dataset])
self.hf_dataset.set_transform(hf_transform_to_torch)
ep_data_path = self.root / self.meta.get_data_file_path(ep_index=episode_index)
ep_data_path.parent.mkdir(parents=True, exist_ok=True)
write_parquet(ep_dataset, ep_data_path)
ep_dataset.to_parquet(ep_data_path)
def clear_episode_buffer(self) -> None:
episode_index = self.episode_buffer["episode_index"]
@ -833,7 +944,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
def stop_image_writer(self) -> None:
"""
Whenever wrapping this dataset inside a parallelized DataLoader, this needs to be called first to
remove the image_write in order for the LeRobotDataset object to be pickleable and parallelized.
remove the image_writer in order for the LeRobotDataset object to be pickleable and parallelized.
"""
if self.image_writer is not None:
self.image_writer.stop()
@ -873,38 +984,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
return video_paths
def consolidate(self, run_compute_stats: bool = True, keep_image_files: bool = False) -> None:
self.hf_dataset = self.load_hf_dataset()
self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
check_timestamps_sync(self.hf_dataset, self.episode_data_index, self.fps, self.tolerance_s)
if len(self.meta.video_keys) > 0:
self.encode_videos()
self.meta.write_video_info()
if not keep_image_files:
img_dir = self.root / "images"
if img_dir.is_dir():
shutil.rmtree(self.root / "images")
video_files = list(self.root.rglob("*.mp4"))
assert len(video_files) == self.num_episodes * len(self.meta.video_keys)
parquet_files = list(self.root.rglob("*.parquet"))
assert len(parquet_files) == self.num_episodes
if run_compute_stats:
self.stop_image_writer()
# TODO(aliberts): refactor stats in save_episodes
self.meta.stats = compute_stats(self)
serialized_stats = serialize_dict(self.meta.stats)
write_json(serialized_stats, self.root / STATS_PATH)
self.consolidated = True
else:
logging.warning(
"Skipping computation of the dataset statistics, dataset is not fully consolidated."
)
@classmethod
def create(
cls,
@ -933,7 +1012,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
)
obj.repo_id = obj.meta.repo_id
obj.root = obj.meta.root
obj.local_files_only = obj.meta.local_files_only
obj.revision = None
obj.tolerance_s = tolerance_s
obj.image_writer = None
@ -943,19 +1022,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
# TODO(aliberts, rcadene, alexander-soare): Merge this with OnlineBuffer/DataBuffer
obj.episode_buffer = obj.create_episode_buffer()
# This bool indicates that the current LeRobotDataset instance is in sync with the files on disk. It
# is used to know when certain operations are need (for instance, computing dataset statistics). In
# order to be able to push the dataset to the hub, it needs to be consolidated first by calling
# self.consolidate().
obj.consolidated = True
obj.episodes = None
obj.hf_dataset = None
obj.hf_dataset = obj.create_hf_dataset()
obj.image_transforms = None
obj.delta_timestamps = None
obj.delta_indices = None
obj.episode_data_index = None
obj.video_backend = video_backend if video_backend is not None else "pyav"
obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
return obj
@ -975,13 +1048,12 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
delta_timestamps: dict[list[float]] | None = None,
tolerances_s: dict | None = None,
download_videos: bool = True,
local_files_only: bool = False,
video_backend: str | None = None,
):
super().__init__()
self.repo_ids = repo_ids
self.root = Path(root) if root else LEROBOT_HOME
self.tolerances_s = tolerances_s if tolerances_s else {repo_id: 1e-4 for repo_id in repo_ids}
self.root = Path(root) if root else HF_LEROBOT_HOME
self.tolerances_s = tolerances_s if tolerances_s else dict.fromkeys(repo_ids, 0.0001)
# Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
# are handled by this class.
self._datasets = [
@ -993,7 +1065,6 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
delta_timestamps=delta_timestamps,
tolerance_s=self.tolerances_s[repo_id],
download_videos=download_videos,
local_files_only=local_files_only,
video_backend=video_backend,
)
for repo_id in repo_ids
@ -1021,7 +1092,10 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
self.image_transforms = image_transforms
self.delta_timestamps = delta_timestamps
self.stats = aggregate_stats(self._datasets)
# TODO(rcadene, aliberts): We should not perform this aggregation for datasets
# with multiple robots of different ranges. Instead we should have one normalization
# per robot.
self.stats = aggregate_stats([dataset.meta.stats for dataset in self._datasets])
@property
def repo_id_to_index(self):

56
lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md
View File

@ -1,56 +0,0 @@
## Using / Updating `CODEBASE_VERSION` (for maintainers)
Since our dataset pushed to the hub are decoupled with the evolution of this repo, we ensure compatibility of
the datasets with our code, we use a `CODEBASE_VERSION` (defined in
lerobot/common/datasets/lerobot_dataset.py) variable.
For instance, [`lerobot/pusht`](https://huggingface.co/datasets/lerobot/pusht) has many versions to maintain backward compatibility between LeRobot codebase versions:
- [v1.0](https://huggingface.co/datasets/lerobot/pusht/tree/v1.0)
- [v1.1](https://huggingface.co/datasets/lerobot/pusht/tree/v1.1)
- [v1.2](https://huggingface.co/datasets/lerobot/pusht/tree/v1.2)
- [v1.3](https://huggingface.co/datasets/lerobot/pusht/tree/v1.3)
- [v1.4](https://huggingface.co/datasets/lerobot/pusht/tree/v1.4)
- [v1.5](https://huggingface.co/datasets/lerobot/pusht/tree/v1.5)
- [v1.6](https://huggingface.co/datasets/lerobot/pusht/tree/v1.6) <-- last version
- [main](https://huggingface.co/datasets/lerobot/pusht/tree/main) <-- points to the last version
Starting with v1.6, every dataset pushed to the hub or saved locally also have this version number in their
`info.json` metadata.
### Uploading a new dataset
If you are pushing a new dataset, you don't need to worry about any of the instructions below, nor to be
compatible with previous codebase versions. The `push_dataset_to_hub.py` script will automatically tag your
dataset with the current `CODEBASE_VERSION`.
### Updating an existing dataset
If you want to update an existing dataset, you need to change the `CODEBASE_VERSION` from `lerobot_dataset.py`
before running `push_dataset_to_hub.py`. This is especially useful if you introduce a breaking change
intentionally or not (i.e. something not backward compatible such as modifying the reward functions used,
deleting some frames at the end of an episode, etc.). That way, people running a previous version of the
codebase won't be affected by your change and backward compatibility is maintained.
However, you will need to update the version of ALL the other datasets so that they have the new
`CODEBASE_VERSION` as a branch in their hugging face dataset repository. Don't worry, there is an easy way
that doesn't require to run `push_dataset_to_hub.py`. You can just "branch-out" from the `main` branch on HF
dataset repo by running this script which corresponds to a `git checkout -b` (so no copy or upload needed):
```python
from huggingface_hub import HfApi
from lerobot import available_datasets
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
api = HfApi()
for repo_id in available_datasets:
dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
branches = [b.name for b in dataset_info.branches]
if CODEBASE_VERSION in branches:
print(f"{repo_id} already @{CODEBASE_VERSION}, skipping.")
continue
else:
# Now create a branch named after the new version by branching out from "main"
# which is expected to be the preceding version
api.create_branch(repo_id, repo_type="dataset", branch=CODEBASE_VERSION, revision="main")
print(f"{repo_id} successfully updated @{CODEBASE_VERSION}")
```

85
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_cabinet.txt
View File

@ -1,85 +0,0 @@
https://drive.google.com/file/d/1_SOJkgfP5yZyVjMhTt3nwhvyUjcnlI51/view?usp=drive_link
https://drive.google.com/file/d/1rmgN8UUzph1qwJnzG1d-uOafodn-gLvb/view?usp=drive_link
https://drive.google.com/file/d/1NYQ-XxsBVinB6dUoZmVWweT83367P3i2/view?usp=drive_link
https://drive.google.com/file/d/1oAv_j74zxxCJieMG7r5Vl2BeHK1__3s3/view?usp=drive_link
https://drive.google.com/file/d/1wFUJQROsrTJt64YRuIeExhFjr2wnK5uu/view?usp=drive_link
https://drive.google.com/file/d/1KzL3Tt0Le7jVl58XVRUcmigmXjyiuhbK/view?usp=drive_link
https://drive.google.com/file/d/1qy_YBladeHtianSSGtgAPSHtMin7msvf/view?usp=drive_link
https://drive.google.com/file/d/1rA_F0V_qL_nyuC_0aBKCisF4-0TIkF2Y/view?usp=drive_link
https://drive.google.com/file/d/1hw-8qMpz9VgSt62XoASqNRuPECpCwJQP/view?usp=drive_link
https://drive.google.com/file/d/1BpHOl9rKMzdvNGka6js7C0s40hH6vnDA/view?usp=drive_link
https://drive.google.com/file/d/1PazhkhiDnJ-OUMyDVDFxEZNKQQqHiNWS/view?usp=drive_link
https://drive.google.com/file/d/1lZ665R6ATl57dypxH4dGJ2NSt6XYnbuz/view?usp=drive_link
https://drive.google.com/file/d/1V9HzLaf-tlG15wUzT7KrTDCS_z1vi5NV/view?usp=drive_link
https://drive.google.com/file/d/1aKauWiXoKqbNwn_2xs4MrmLlaNYlVNmO/view?usp=drive_link
https://drive.google.com/file/d/1WVD5DFhriO1YmmOgiVHhacR6HWoTPxav/view?usp=drive_link
https://drive.google.com/file/d/1_X43WgeBAsfkhH9EmpyPki8U9joMeAGC/view?usp=drive_link
https://drive.google.com/file/d/1t8x0GqWoNKWtnBsB7_D40Z34nL9ak4kf/view?usp=drive_link
https://drive.google.com/file/d/15V_f26WaKOXjKnq2T3HRWAmtQUi4lbu2/view?usp=drive_link
https://drive.google.com/file/d/11VFIAsiSDsMOBANgrOcZBpKB9AFWnLy7/view?usp=drive_link
https://drive.google.com/file/d/1M0NS7vVaxJv3FHnuRYtdwTFYF7We4LxP/view?usp=drive_link
https://drive.google.com/file/d/1mR0OItTNqFnVLoczcyKYlm6drAy778lO/view?usp=drive_link
https://drive.google.com/file/d/1NbVFWDQAh-z4JJ4D-Zw6Lps9kdvpqh2j/view?usp=drive_link
https://drive.google.com/file/d/1JQoZGBzl4W3QG26-n39tefcGN0fDRMbB/view?usp=drive_link
https://drive.google.com/file/d/1VBjHl-TvZpncopvasIP5G9gecbB2a5f6/view?usp=drive_link
https://drive.google.com/file/d/1VzSf6zaB21nahm7MsPwroXbJ84NIwq0b/view?usp=drive_link
https://drive.google.com/file/d/1OtNnfMEydNtZOcivs4k6E_uJSpf8PkGy/view?usp=drive_link
https://drive.google.com/file/d/14nVvpvsrFr_03Pa_N7MKzwnRwibOUYM6/view?usp=drive_link
https://drive.google.com/file/d/1M8li6duiO2r3lv_9HhF_XJn0oZUIEK5F/view?usp=drive_link
https://drive.google.com/file/d/1Cpzea6fO14lxAaNfSBifqoa4ekhCiLD1/view?usp=drive_link
https://drive.google.com/file/d/1mbxRTm5vlbsY9UJ0jfjM6j9D7kPJjBpG/view?usp=drive_link
https://drive.google.com/file/d/1RXD1i6IfWsHRlCxVmG04h2h5Ycm_WwZN/view?usp=drive_link
https://drive.google.com/file/d/1QFqFSwDGOk1BkgGmqgCcc2BRWnJ6R3MA/view?usp=drive_link
https://drive.google.com/file/d/1bFqWR8DQM0ZUxxtS2bl-RANQvukeFLzp/view?usp=drive_link
https://drive.google.com/file/d/1pR-rH3yNGoyPdD4hJ6-3lXQ-PstBx9du/view?usp=drive_link
https://drive.google.com/file/d/107OAwLY-hva9HeQLIK7VCh-ytdDabVjr/view?usp=drive_link
https://drive.google.com/file/d/1Tpl08QOaSZ37GTO4awFWSdD8wBR9xdlT/view?usp=drive_link
https://drive.google.com/file/d/1MR164AOM-0S1T6RX8xKTV2IHyaCvpqAW/view?usp=drive_link
https://drive.google.com/file/d/1_wknJfVnStIhJ82lU_QtcrwahsqYIsr8/view?usp=drive_link
https://drive.google.com/file/d/1ZuEktWrbYkTx0l5pj3WiZ2CJrfbDOHNo/view?usp=drive_link
https://drive.google.com/file/d/15G_10hkkkq6yxvyI5NGZirlF-RzduR2F/view?usp=drive_link
https://drive.google.com/file/d/1DBKxg3ONqh7dhLuX6oh1Yyo2x383V1Hp/view?usp=drive_link
https://drive.google.com/file/d/1B5iDBkTUr5vopDddV_fHud18SqAHhauS/view?usp=drive_link
https://drive.google.com/file/d/1acwFV0eenRkki1QcjSKH5xqOtys-P3Pr/view?usp=drive_link
https://drive.google.com/file/d/1S47BI83xyrh-FKXsvAQqer98Biu_p8XK/view?usp=drive_link
https://drive.google.com/file/d/1JL6DmBZl3uyq9dyLfgSqtGF06e7E9JwM/view?usp=drive_link
https://drive.google.com/file/d/16WvRS4Kjog8Pxgr0E3sGGnI01YwL9Uql/view?usp=drive_link
https://drive.google.com/file/d/12ttGqL33IPWg0-s1SD44rr22M6LiSQBr/view?usp=drive_link
https://drive.google.com/file/d/1OyZqqnldTU_DliRbr6x0C4a_iWPwIN7j/view?usp=drive_link
https://drive.google.com/file/d/1oYk00IpLnR9fesLfD15Ebe7nVBffEbcS/view?usp=drive_link
https://drive.google.com/file/d/1eyE2-MQduCEqCd-5_kl5zsoOEERAzpZD/view?usp=drive_link
https://drive.google.com/file/d/1ir1Ya-vO0d97pfvbePlUeuKTTRc0qIMU/view?usp=drive_link
https://drive.google.com/file/d/1hOi-JnqlMt47gVnLZHMTqeojyYVErohl/view?usp=drive_link
https://drive.google.com/file/d/1NFFw5_PqigQ7xGqsL-MNq2B1r5yAscCf/view?usp=drive_link
https://drive.google.com/file/d/1uftq1-Zlh8d2sNLWrlVcKYQUwZTD7o24/view?usp=drive_link
https://drive.google.com/file/d/1-ax19dSLPacVgk000T-m3l4flPcg07pM/view?usp=drive_link
https://drive.google.com/file/d/126y-lgn86-ZmCz8hooF1THKJGGObw3OB/view?usp=drive_link
https://drive.google.com/file/d/1JiDniK0VmDIkk92AbBILb8J2Ba59PWML/view?usp=drive_link
https://drive.google.com/file/d/1kr8nPIRljiU0R4J9SMgj80o1FPQxzu9z/view?usp=drive_link
https://drive.google.com/file/d/1bbThWRij1pKBh_kFgV8FwK0sXtTHBoLX/view?usp=drive_link
https://drive.google.com/file/d/1WenzDW6lxk1xkOFm-OiGFfc0ROskAuKU/view?usp=drive_link
https://drive.google.com/file/d/1MiKRzuzUn1yN-k_6kPJJzIGy7dT-nnsD/view?usp=drive_link
https://drive.google.com/file/d/17rRg2tcmB-gNhQ0KoZJQmNfyFeoij1jH/view?usp=drive_link
https://drive.google.com/file/d/11mokBpvrY3ld6sY5WztREtJ1jgqfQV70/view?usp=drive_link
https://drive.google.com/file/d/1Il_6IOx9NDp1bX_KHizJfBwzTufTmn86/view?usp=drive_link
https://drive.google.com/file/d/1KswtJGsxJ7eeBDAmNA_aeLjOxcH6MIxa/view?usp=drive_link
https://drive.google.com/file/d/1gzMhi5uWu4C3Y6WbQ3L-08V96GxTZrRR/view?usp=drive_link
https://drive.google.com/file/d/1nRQFtaBxfUCYc2W90Qibh0kHCt6YQCfc/view?usp=drive_link
https://drive.google.com/file/d/1vs-gyW-KheqHbUATwAhA2mmR9GOGw7f_/view?usp=drive_link
https://drive.google.com/file/d/1MuxzGOA2fgLaHryq82KkQumtuRJGcUOC/view?usp=drive_link
https://drive.google.com/file/d/1IIwxZnGlqrXLUXqG6yMO0r7uhCvhpk9e/view?usp=drive_link
https://drive.google.com/file/d/1vE7XPyaFcXP4DtTY5Y9WKIt7zWgmX-Cr/view?usp=drive_link
https://drive.google.com/file/d/1j-bIV09gr21RC3-x1N_pK4RPLV3fmWKz/view?usp=drive_link
https://drive.google.com/file/d/1t3nW1rD3S-EL0Oymb5U7ZAj5UMkydkln/view?usp=drive_link
https://drive.google.com/file/d/14hbfHCdMKtJZ41F9CQReMec2jeRFTOqR/view?usp=drive_link
https://drive.google.com/file/d/1x-hUyOSne5BW0AzQ3W6_Pf4g5yXQWi9M/view?usp=drive_link
https://drive.google.com/file/d/1sw9JqRg6E-3P84I3ZhzTrJMu0vuiaMmP/view?usp=drive_link
https://drive.google.com/file/d/1LuqhQlL4MGZhB_6THmkovRxrlP26BbdC/view?usp=drive_link
https://drive.google.com/file/d/15C5K6v_lkjnMSmUvVyqHQKwh2N166e7K/view?usp=drive_link
https://drive.google.com/file/d/1ns_9eSsQeeoZ10nlbkLy8tu0GmJFSnkt/view?usp=drive_link
https://drive.google.com/file/d/1NpzWJeK6CqjxzjIMYe6aYdX8xGsQwD4o/view?usp=drive_link
https://drive.google.com/file/d/1NMLezwufKJ9_8xTc9KQThSzVVD71B9Ui/view?usp=drive_link
https://drive.google.com/file/d/1aa71DCUqs6oXlIxX35jgsmsgm-NlDxPV/view?usp=drive_link
https://drive.google.com/file/d/1UJzkIZzAL0j-D5YQBnoq7mHvttASy12O/view?usp=drive_link
https://drive.google.com/file/d/1nPgx36HIJFb7oI94VbRzWjpPP2GANxzG/view?usp=drive_link
https://drive.google.com/file/d/1NovAP-KVJjqcuvWy3d6G4ptGGAIDqcCx/view?usp=drive_link

55
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_chair.txt
View File

@ -1,55 +0,0 @@
https://drive.google.com/file/d/11M3Ye0r5agMaaicPbVGD0q2Hb3rGklbb/view?usp=drive_link
https://drive.google.com/file/d/1-tx7SvYYgSvXCvnf_EI2OVdwK-CkFY6S/view?usp=drive_link
https://drive.google.com/file/d/1EWJunmOpMHaU1hE106wwpbkGYcjQXYAF/view?usp=drive_link
https://drive.google.com/file/d/1IDn95Z7FSiCckrSENtGV4u3RyFHNQSDY/view?usp=drive_link
https://drive.google.com/file/d/1CwzvWj1i7QOtqrZvsCZ6BdZaKNDfpN32/view?usp=drive_link
https://drive.google.com/file/d/1HvAvlhm77nAD3Td24QPSeq8lw-Rl_aOh/view?usp=drive_link
https://drive.google.com/file/d/1t-suKYOPhXH666RpAYNRp2QU_DOy3AeM/view?usp=drive_link
https://drive.google.com/file/d/18xpKgWh7RWyjMN5PkLTOo-AxsAadAuRw/view?usp=drive_link
https://drive.google.com/file/d/1oci5Eto-ztv-AQNz8EnwZveBIhxvk-xJ/view?usp=drive_link
https://drive.google.com/file/d/1Y-t_4vxdE6NpHO0DLJR8f3mD0Q-Wj5-c/view?usp=drive_link
https://drive.google.com/file/d/1lylRqbbbB8bgtpsBWMPACmHJreuKmllv/view?usp=drive_link
https://drive.google.com/file/d/1yliSyMig_NXShWfQx6qyW7Ijf2Y5lFK6/view?usp=drive_link
https://drive.google.com/file/d/1XXhwJsJbeb7KXAooGvJapnm9bjnGUmxS/view?usp=drive_link
https://drive.google.com/file/d/1_xs1f3hW2JArKyvfF7UWubWjyROGTLs6/view?usp=drive_link
https://drive.google.com/file/d/1WVEHpr6EqKCZbkHapQSTXJq4xE4SWFT-/view?usp=drive_link
https://drive.google.com/file/d/1RqOHv9pEQGvW8NUA7ynffFmG999TL_Az/view?usp=drive_link
https://drive.google.com/file/d/1cu5AgD2gh-uA3PFJmzxxzNaF3qOSlYY1/view?usp=drive_link
https://drive.google.com/file/d/1SsrXqiPclNrnYToPZ9Uq-k3y0C4qdHT1/view?usp=drive_link
https://drive.google.com/file/d/1-J7EXf0vjkLIfSqT8ICEsP6CTjzSLBop/view?usp=drive_link
https://drive.google.com/file/d/11O7ewUmoZXfyyKjy_6B5RW4DpjICxqBT/view?usp=drive_link
https://drive.google.com/file/d/1iic44kZoCsjNsfAz2cMstZ9-WQvAhblF/view?usp=drive_link
https://drive.google.com/file/d/1yLV1lVX-2WnWQldGlnQZ0x7QBuDiVkL3/view?usp=drive_link
https://drive.google.com/file/d/1Tybp9ru98TTbGn4eyROpUQwDFuALWXmk/view?usp=drive_link
https://drive.google.com/file/d/13E9OTMiipVJByDs5-J19oWwAz7l94LTN/view?usp=drive_link
https://drive.google.com/file/d/1EeTpJQdMSliw4JzSMtJ6CyTvVdexjM4M/view?usp=drive_link
https://drive.google.com/file/d/1NHyNwoFqzeAu-1_PSpq5JfxaiD_xbpn9/view?usp=drive_link
https://drive.google.com/file/d/1fJcS0phDp4xm_FyGaJ5wr9Pe4KqtHaxD/view?usp=drive_link
https://drive.google.com/file/d/12AqrLUaewDPEcFRqPZeZFb_TQ0Lfi3At/view?usp=drive_link
https://drive.google.com/file/d/1x_hd4Qsq1oJS-aj2t3qM7WbbV7KZj05b/view?usp=drive_link
https://drive.google.com/file/d/14OUSUArmsB068hs6BuEIXQhI1Cyz8Sf0/view?usp=drive_link
https://drive.google.com/file/d/16zlzh1T5zeUJQnFf382NXkFEKEnDub4O/view?usp=drive_link
https://drive.google.com/file/d/1IbDltmN-NEFCNtr1TO4ILxEgQ94rtjWv/view?usp=drive_link
https://drive.google.com/file/d/15gmlf8Gx9455pZ1AlqcCSwh3nDPxMzSr/view?usp=drive_link
https://drive.google.com/file/d/1qHpRL1oZfIMo_vxnm8qfwQ-7l0BZIVva/view?usp=drive_link
https://drive.google.com/file/d/1H1xskIgiFZivkYn23rMzH3xePGOh3VTC/view?usp=drive_link
https://drive.google.com/file/d/1avls6Pv0kYiCMNVknbc1zQsgy64MUDMM/view?usp=drive_link
https://drive.google.com/file/d/1MmWVgCj5khc8KMIifmt3EzF1o-CtPyyn/view?usp=drive_link
https://drive.google.com/file/d/1U0kCc_xqW0WNppf4sbnK14euWKdPZtzB/view?usp=drive_link
https://drive.google.com/file/d/16CaEyQscOuhLj23PEGDTL9DeyNkohkMn/view?usp=drive_link
https://drive.google.com/file/d/1Iu8uM6UUJ0zW8tvN-9UiOe_4oSNzEutg/view?usp=drive_link
https://drive.google.com/file/d/1UImqiBaIxCR-1DNJaZhHqeHhaySOtVIr/view?usp=drive_link
https://drive.google.com/file/d/1VpU2V_leIoRIyv_lAvE7eLHBG8DxCTnp/view?usp=drive_link
https://drive.google.com/file/d/1_Q8J27OT3Xby7QY6yHvIJauFRWEMxkRm/view?usp=drive_link
https://drive.google.com/file/d/1bantmVo1L9Xz4tbiNw_a1UC2Z_HPO1wT/view?usp=drive_link
https://drive.google.com/file/d/1IRIXMJMCBDkBjbaHvAlEiBogSvZ1jK_3/view?usp=drive_link
https://drive.google.com/file/d/1mAHXKjiFbjwydypW2t5Lv8_H5x6nHegl/view?usp=drive_link
https://drive.google.com/file/d/1SfyY796fLrBCMY39OcyuxZafqSCRZPZk/view?usp=drive_link
https://drive.google.com/file/d/1X-44sZ8CcfzIskc0dvSx882o1yFhHaZB/view?usp=drive_link
https://drive.google.com/file/d/1BOIWCCCk6DLD4Bmvc75ZbbLi9AQm-1ao/view?usp=drive_link
https://drive.google.com/file/d/1RuyDtRE1kk76sw-wP8vx5SgLoPF3PA_H/view?usp=drive_link
https://drive.google.com/file/d/1c4eoQiBbGuy3CTAQDUSkd84Ponh1roAQ/view?usp=drive_link
https://drive.google.com/file/d/19PXB9z4Ljq6dsbf9TqcOrrP5SRbw2Tc_/view?usp=drive_link
https://drive.google.com/file/d/1nn1VVZVoIXWdYDozR7XHXE4mPLQG80PQ/view?usp=drive_link
https://drive.google.com/file/d/1MBdFGOKPV8GUhwoSsJ_Ky3qAMLM2Bv3K/view?usp=drive_link
https://drive.google.com/file/d/1of3k_M-7Nh3I1TndcWedxK4ca9dn8Sc5/view?usp=drive_link

20
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_elevator.txt
View File

@ -1,20 +0,0 @@
https://drive.google.com/file/d/12ctkOAdkCNGN1JLbZb5ww3XTBn2LFpGI/view?usp=drive_link
https://drive.google.com/file/d/1G_Vd46_4fq6O64gHHjUbJX5Ld44ZZx0y/view?usp=drive_link
https://drive.google.com/file/d/1uKgUy73B3xBogQAOUhfZjO0X5qZGsi2c/view?usp=drive_link
https://drive.google.com/file/d/1fu9cIrfI-fE2LhdGUxbx7-8Ci_PF8Ypm/view?usp=drive_link
https://drive.google.com/file/d/1Ygk9ZPJzx8xw2A9JF3NHbJ44TqnvSTQR/view?usp=drive_link
https://drive.google.com/file/d/18m5xPuccNsEB20WPshm3zhxmXc6k63ED/view?usp=drive_link
https://drive.google.com/file/d/1DiqqxC44rriviRQpqogcv0-EB-Y6nr9g/view?usp=drive_link
https://drive.google.com/file/d/1qPdaoTVDizJXkfXLioWU7iJ8hqCXSyOQ/view?usp=drive_link
https://drive.google.com/file/d/1Fj9kIA_mG7f67WFfACJEaZ7izcHG7vUm/view?usp=drive_link
https://drive.google.com/file/d/1WpYehZnI2P7dUdJPfkE-ij1rqCnjZEbB/view?usp=drive_link
https://drive.google.com/file/d/1_zwWkT4jPyzB38STWb6whlzsPzXmfA9r/view?usp=drive_link
https://drive.google.com/file/d/1U6-J4I_fPlSFFGfhZPxS5_YzKXwXIZYp/view?usp=drive_link
https://drive.google.com/file/d/1pRhxxcTfZp5tQo_EScvJUwfc3amiS6Vk/view?usp=drive_link
https://drive.google.com/file/d/1lWLntqra83RlYU_gN7Vostnfydf6gutd/view?usp=drive_link
https://drive.google.com/file/d/1vIBKo0x-NYEHV1FvRpco1lQMpRdAWAIL/view?usp=drive_link
https://drive.google.com/file/d/1pdrLV3JTQou_XH0Aap61Ssf60iVKm1jJ/view?usp=drive_link
https://drive.google.com/file/d/1QTsLoQ7SwmKdQHjBGVDaR2uTwfFwtrOf/view?usp=drive_link
https://drive.google.com/file/d/1Gytai8M_12J36GY6L_TulEcOC-035jwS/view?usp=drive_link
https://drive.google.com/file/d/14LJudNc629NT-i8xreXtzl27ce_DxOFJ/view?usp=drive_link
https://drive.google.com/file/d/1sBvPCODbzxGAI0S3lgN5cSG9Go3lRi00/view?usp=drive_link

18
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_shrimp.txt
View File

@ -1,18 +0,0 @@
https://drive.google.com/file/d/1MJn9GbC8p9lN4gC9KDMLEkTkP_gGpXj0/view?usp=drive_link
https://drive.google.com/file/d/1-4LXgjl7ZCOgp-8GCJmFRD8OeqN5Jf7-/view?usp=drive_link
https://drive.google.com/file/d/1Ho06Ce0SPbqU3juaMxNUwAt3zCRLGC8W/view?usp=drive_link
https://drive.google.com/file/d/1ivHoj7_7olBSxH-Y8kqXEW7ttITK-45j/view?usp=drive_link
https://drive.google.com/file/d/1qjY4hM_IvZ8cq2II_n9MeJbvyeuN4oBP/view?usp=drive_link
https://drive.google.com/file/d/1rKVhO_f92-7sw13T8hTVrza3B9oAVgoy/view?usp=drive_link
https://drive.google.com/file/d/1pcLPHO8fBkc1-CRa88tyQtEueE4xiXNi/view?usp=drive_link
https://drive.google.com/file/d/1Vev_chCsIeEdvQ8poEYNsOJFGy_QU8kZ/view?usp=drive_link
https://drive.google.com/file/d/1l5G4zpRkxSLCQjvGPYSN4zfCvVRQuzMz/view?usp=drive_link
https://drive.google.com/file/d/14vgthE1eoakXkr2-DRw50E6lAqYOiUuE/view?usp=drive_link
https://drive.google.com/file/d/17nPSmKKmgQ2B7zkzWrZYiLM3RBuFod82/view?usp=drive_link
https://drive.google.com/file/d/1QcDsxplVvb_ID9BVrihl5FvlC-j7waXi/view?usp=drive_link
https://drive.google.com/file/d/18pEejBpI-eEVaWAAjBCyC0vgbX3T1Esj/view?usp=drive_link
https://drive.google.com/file/d/1H8eH6_IRODtEFT6WoM77ltR5OoOrqXmI/view?usp=drive_link
https://drive.google.com/file/d/1IWlpFRZhoxyG4nS13CWK4leZVk5wbNx4/view?usp=drive_link
https://drive.google.com/file/d/1PbZA8_OCGmMLxNP9xbkLRSChniL4uGxl/view?usp=drive_link
https://drive.google.com/file/d/1p9XAdmG2f_WeflNO4DIJ_tr1rK6M9B4B/view?usp=drive_link
https://drive.google.com/file/d/1nS59Et1cNAvKo3Y4SeSGRuZD5TvBbCF3/view?usp=drive_link

1
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wash_pan.txt
View File

@ -1 +0,0 @@
https://drive.google.com/drive/folders/1S8eFg98IaGAIKVZ8QFWG1bx4mHa-O204

4
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wipe_wine.txt
View File

@ -1,4 +0,0 @@
https://drive.google.com/drive/folders/1tC_g1AJ8lglBLY-fjsQrG6DMBa3Ucp-0
https://drive.google.com/file/d/1fG_Yi2MJrFjiUVN3XoiWXLtTxHlwwaDv/view?usp=drive_link
https://drive.google.com/file/d/1WX32VWfzzX3Blmd06DRxLwFbMJfVe7P4/view?usp=drive_link
https://drive.google.com/file/d/18onsX3vXg3xkFwP5bVUCjdV4n9TRn0C9/view?usp=drive_link

3
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_human.txt
View File

@ -1,3 +0,0 @@
https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF
https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link
https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link

3
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_scripted.txt
View File

@ -1,3 +0,0 @@
https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N
https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link
https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link

3
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_human.txt
View File

@ -1,3 +0,0 @@
https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo
https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link
https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link

3
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_scripted.txt
View File

@ -1,3 +0,0 @@
https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj
https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link
https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_battery.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/19qS_n7vKgDcPeTMnvDHQ5-n73xEbJz5D
https://drive.google.com/file/d/1oC31By0A2bsBeHyUwBdQw1z4ng6yi9Za/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_candy.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1m5rQ6UVH8Q9RQp_6c0CxkQ88-L-ScO7q
https://drive.google.com/file/d/1wHz2qcmwcVG0C0CZ9MjQDQcmj4OY9_a3/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1seQGay470nGQ-knBI5TjsTr8iL9Qws5q
https://drive.google.com/file/d/1T89hSX5U99wLGvGTE7yUBaQPOpyj6Sai/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee_new.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1t3eDc5Rg0DveyRe8oTm6Dia_FYU5mXyf
https://drive.google.com/file/d/1TXFaduTakvS0ZWJqKCX-HIvYglum_5CY/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_cups_open.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1Z9X3DNzd6LS0FFjQemNUMoMA5yk5VQOh
https://drive.google.com/file/d/1Wlyc0vTkjXuWB6zbaVOWhEfD7BmPgUV_/view?usp=drive_link

53
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_fork_pick_up.txt
View File

@ -1,53 +0,0 @@
https://drive.google.com/drive/folders/1DYgB4ifX4uIid9m9jnC0Zdz8Nf7ZC0fc
https://drive.google.com/file/d/1Eb-NRNk_FmVleCbU_Ng5Y4dfcjTKN7Rv/view?usp=drive_link
https://drive.google.com/file/d/1dkhjEADakT-44l9jf-nK4x89kr4yG_qb/view?usp=drive_link
https://drive.google.com/file/d/14hDhgcZkVqNExGb4tIXpSjMshhqZETch/view?usp=drive_link
https://drive.google.com/file/d/1zVMEHpHbuNyP5A_lYU7RPSLB-4V0yfZw/view?usp=drive_link
https://drive.google.com/file/d/1JtgDjBvy7FnRpFzrx_foC3quorYQFAR-/view?usp=drive_link
https://drive.google.com/file/d/1EHdneB6F-PP0dQlX8qPaXbxmKoBy_YwO/view?usp=drive_link
https://drive.google.com/file/d/17Z0jjVBy1OPKREPu77_n_rQzorDiapji/view?usp=drive_link
https://drive.google.com/file/d/1F4i23qPJ_qTf5jWjfLo4ARGJChznYWt3/view?usp=drive_link
https://drive.google.com/file/d/1kZtXWM3uS0-rLblydBfJ0mMcVnMMXw9w/view?usp=drive_link
https://drive.google.com/file/d/1mNODox87xFfY5Z_o5mcLsr8SHb39jDik/view?usp=drive_link
https://drive.google.com/file/d/1Ob44VdmEUA93FKDECiRb5Ogz2xQg5IWp/view?usp=drive_link
https://drive.google.com/file/d/1fdQLdjj3Cwv33R1wZhfrLz9Del8mqgHb/view?usp=drive_link
https://drive.google.com/file/d/1Yu3L3ft21zP__XL8pCfhb788ZleuW1n5/view?usp=drive_link
https://drive.google.com/file/d/1ozBBWXVZ9hXDh9ooHUNroHdYm8UDqnhJ/view?usp=drive_link
https://drive.google.com/file/d/1o0TGqvfWw_Lunxb5ubKDS21Lr_WC0h75/view?usp=drive_link
https://drive.google.com/file/d/1jZnd5eP5L6BH5l98BPN6OnoQx3fu8e9n/view?usp=drive_link
https://drive.google.com/file/d/1S5sYbz8wcLYp0V67v13i4PRcBxodn4Hg/view?usp=drive_link
https://drive.google.com/file/d/1rFeg_x6ftJYwPtBv34D3h2L2cpDLeR4G/view?usp=drive_link
https://drive.google.com/file/d/1GvS3lcm4o6nm_scUk0XxKeVFNmzjucDZ/view?usp=drive_link
https://drive.google.com/file/d/1-9i0riphC7NhhDahcQfD1QoBXP5gF90A/view?usp=drive_link
https://drive.google.com/file/d/15p_IqGsMbKuvzMS872THAZr-3SBtb1Fr/view?usp=drive_link
https://drive.google.com/file/d/1ToyYcBfJL8gbQn0q_59zPLsFmm7dmMJo/view?usp=drive_link
https://drive.google.com/file/d/1e_7PNH7CYafE4pAebP7ZdI7XFbmEcy_i/view?usp=drive_link
https://drive.google.com/file/d/1JoabvGVsIQdug2xOhUIhetEIyDM91y_Y/view?usp=drive_link
https://drive.google.com/file/d/1kOMw1y0lmnVaCjwZICfzCsx6e0Z8MNGR/view?usp=drive_link
https://drive.google.com/file/d/16it_wd1JOevUQTK2_CvF_pBACTgpIPgM/view?usp=drive_link
https://drive.google.com/file/d/1IRcCj9HnJSfbyMgr5XEERGlEnWeZQwOc/view?usp=drive_link
https://drive.google.com/file/d/1Z2dIJfq_S3liGmPN9Rphvkmucnmw7tlb/view?usp=drive_link
https://drive.google.com/file/d/1J3NoAjzndGx9yNyaBOJHdNny1epzUoBt/view?usp=drive_link
https://drive.google.com/file/d/18nOvxV1k8FSmBrhT4TPo2sKKSZXougyx/view?usp=drive_link
https://drive.google.com/file/d/1CT8FxclafFMjSd7gCWVw3VSeryeiF04i/view?usp=drive_link
https://drive.google.com/file/d/16M9KVqQMFfSsXfypK0bocFft8Nz3j2Rt/view?usp=drive_link
https://drive.google.com/file/d/18QPVkw6bj6HW8LTPrQLWrrUX4R6RcF42/view?usp=drive_link
https://drive.google.com/file/d/1hQTVtA5hBTE_StXpJafTZJ3tgt2VQQ_t/view?usp=drive_link
https://drive.google.com/file/d/1Dn-d5g69H6EgAWgsFdrcbJKtz7ySsCQ8/view?usp=drive_link
https://drive.google.com/file/d/13hMr16483P7ALYv73yMRUN37fJdVQM62/view?usp=drive_link
https://drive.google.com/file/d/1848yN3XMN5zJMEgApt6KzrWgfRPfimtv/view?usp=drive_link
https://drive.google.com/file/d/1oAD9kSnS0fTgj-CjD4u9VdZ5X67IOIMa/view?usp=drive_link
https://drive.google.com/file/d/1ilzIWLCCG5b_KgF5s0wdN2I5-lFNpwC1/view?usp=drive_link
https://drive.google.com/file/d/1rjsT2YBjnidxod1s9s-myAYz8boHr-WB/view?usp=drive_link
https://drive.google.com/file/d/18Gg48HTub15bd8qzbhiCUufbVy0fbN5G/view?usp=drive_link
https://drive.google.com/file/d/1WsSnQSqmMTVSRwrhT1Y-v782My2zcjLm/view?usp=drive_link
https://drive.google.com/file/d/1ea9ZCvoyc-xqiFXgeDcA_mOWsw7VUuoi/view?usp=drive_link
https://drive.google.com/file/d/1wv1v3-XhPgbNzp62BXbJTDzMPu2tlDUc/view?usp=drive_link
https://drive.google.com/file/d/18-ikzt8LoZ83Gi3goKCELs4U4z8hrRoF/view?usp=drive_link
https://drive.google.com/file/d/16Bjhp7JNCXkGuLvyNcZowAx3W-Y-15DV/view?usp=drive_link
https://drive.google.com/file/d/1Gc-KRI-xwcp1fMR55ugbrLg_5y3SPde-/view?usp=drive_link
https://drive.google.com/file/d/1oP72Q386Z4Sy5MMm-t5yNogIe5Van_9k/view?usp=drive_link
https://drive.google.com/file/d/112T90eDUDVH-SyOV7UnZl5bscAH2hcfq/view?usp=drive_link
https://drive.google.com/file/d/1y-uKOesRRhjgDtFbG_j65f4SGg0v8XDg/view?usp=drive_link
https://drive.google.com/file/d/1LOP05OagoI3km-ZKQBrS204A85UVk7Ok/view?usp=drive_link
https://drive.google.com/file/d/1QkHQKgasVzWsmdPvkXgGhWyQ84d93_Az/view?usp=drive_link

1
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pingpong_test.txt
View File

@ -1 +0,0 @@
https://drive.google.com/drive/folders/1Ut2cv6o6Pkfgg46DgwVUM7Z5PkNG8eJ-

1
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pro_pencil.txt
View File

@ -1 +0,0 @@
https://drive.google.com/drive/folders/1FqxPV0PgvgIu8XFjtvZSPSExuNcxVVAY

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_screw_driver.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1SKtG0ct9q0nVdYssJNMWSOjikcXliT58
https://drive.google.com/file/d/1nchD21O30B3i3LDoqramo1zgW5YvpJIN/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_tape.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_thread_velcro.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

2
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_towel.txt
View File

@ -1,2 +0,0 @@
https://drive.google.com/drive/folders/1fAD7vkyTGTFB_nGXIKofCU1U05oE3MFv
https://drive.google.com/file/d/1XzyQ2B6LLvcurIonOpEu4nij2qwNWshH/view?usp=drive_link

53
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup.txt
View File

@ -1,53 +0,0 @@
https://drive.google.com/drive/folders/13EQsVsnxT86K20QAoyE_YpsFbQ7fZQdu
https://drive.google.com/file/d/1-W_JHghZG65FNTVhw1SXhtQrazdLL3Ue/view?usp=drive_link
https://drive.google.com/file/d/1VwRJgdWUo-2nQaNM7Bs77-fsm8iwUxEo/view?usp=drive_link
https://drive.google.com/file/d/1wFzGRo5iYA13WLi6IV1ry64RyahQBFio/view?usp=drive_link
https://drive.google.com/file/d/1IKtQzQ-n-UTv64hYpReu2R4cqUvmNQqD/view?usp=drive_link
https://drive.google.com/file/d/1GicVci9OiuuZZH79i5Mg7AtWod94MzwT/view?usp=drive_link
https://drive.google.com/file/d/1JVnIoR7EIQp70T4eAf9RX65JcTrzsjQc/view?usp=drive_link
https://drive.google.com/file/d/1W2xr4h23ucjPrc-mBEeqnACsfaImpc0p/view?usp=drive_link
https://drive.google.com/file/d/10xj_0V7A07o3uCa7v5omUrTC0YlPW8H3/view?usp=drive_link
https://drive.google.com/file/d/1FOc3EMaCy8Mb0_a7PuXLAwKwvxkbKmwU/view?usp=drive_link
https://drive.google.com/file/d/143PgDXBcf2GQ0Q07ZPMVMfBgZDd5sLJG/view?usp=drive_link
https://drive.google.com/file/d/1pE5Tyj0LlGbGWvUzuhixp86Ibu55Ez3I/view?usp=drive_link
https://drive.google.com/file/d/141668b1VzX80ncrVJPzhkoAeIFB4MEK9/view?usp=drive_link
https://drive.google.com/file/d/1bw12lo37p1ZvRvErHsll7cEYi2OxscvZ/view?usp=drive_link
https://drive.google.com/file/d/1zfnMFvbgBjl6SzYhksbaOzfbwLrCN6tb/view?usp=drive_link
https://drive.google.com/file/d/1-GIszA6mUJMaNB-tdh9r9skc77SWA0VX/view?usp=drive_link
https://drive.google.com/file/d/1fTB0zWFYU6zh4IIUFT2zX_OkwYqmElwY/view?usp=drive_link
https://drive.google.com/file/d/1gPIPNKGmrO9c7gKF7SP0SuUYbIBBq8z1/view?usp=drive_link
https://drive.google.com/file/d/12JeJ-dQd5lYyn6PlDOGdE-ChVeiZ-Uv0/view?usp=drive_link
https://drive.google.com/file/d/100_20cgCqerU6qoh3TfTbwLy9mlDAFEG/view?usp=drive_link
https://drive.google.com/file/d/111oAGJ76ku_pYgbBoIdZAC1_XEQcPI__/view?usp=drive_link
https://drive.google.com/file/d/1UhC8L-354ZQ2gblPFGI35EMsVwfpuKa0/view?usp=drive_link
https://drive.google.com/file/d/1sIXQSgUR_xdrNtGrL6QGBnkLMKErsIp1/view?usp=drive_link
https://drive.google.com/file/d/16Ax77bDSIXnsn4GFL8XYKKT1P6bPpfMd/view?usp=drive_link
https://drive.google.com/file/d/1pgRVYwwVIsWq_qsWqZpe1UBzZfF5Fa9D/view?usp=drive_link
https://drive.google.com/file/d/1jtimaZkWsY1P5gC2bbS64H_WCUU7HXN2/view?usp=drive_link
https://drive.google.com/file/d/1N6Bh02P-RiTEgtx1YH1Db_X3TGpP-X_r/view?usp=drive_link
https://drive.google.com/file/d/14Fy8EwJ8d9Vh97Yt1VOvUChSCrfIjBij/view?usp=drive_link
https://drive.google.com/file/d/1IRuv42dvIMPuKhcMZmuXaBjJ-lPFOmQd/view?usp=drive_link
https://drive.google.com/file/d/16XWzNY2D8ucVVn5geBgsVdhm3ppO4que/view?usp=drive_link
https://drive.google.com/file/d/1xsVOoQgthK_L_SDrmq_JvQgUpAvPEAY8/view?usp=drive_link
https://drive.google.com/file/d/1bZbw66DyEMvnJnzkdUUNbKjvNKg8KFYM/view?usp=drive_link
https://drive.google.com/file/d/1CyTVkdrNGGpouCXr4CfhKbMzE6Ah3oo3/view?usp=drive_link
https://drive.google.com/file/d/1hDRyeM-XEDpHXpptbT8LvNnlQUR3PWOh/view?usp=drive_link
https://drive.google.com/file/d/1XhHWxbra8Iy5irQZ83IvxwaJqHq9x4s1/view?usp=drive_link
https://drive.google.com/file/d/1haZcn6aM1o4JlmP9tJj3x2enrxiPaDSD/view?usp=drive_link
https://drive.google.com/file/d/1ypDyuUTbljaBZ34f-t7lj3O_0bRmyX2n/view?usp=drive_link
https://drive.google.com/file/d/1ILEEZo_tA9_ChIAprr2mPaNVKZi5vXsO/view?usp=drive_link
https://drive.google.com/file/d/1U7nVYFaGE8vVTfLCW33D74xOjDcqfgyJ/view?usp=drive_link
https://drive.google.com/file/d/1rZ93_rmCov5SMDxPkfM3qthcRELZrQX6/view?usp=drive_link
https://drive.google.com/file/d/1mYO1b_csddtyE3qT6cwLiw-m2w2_1Lxh/view?usp=drive_link
https://drive.google.com/file/d/1xz7Q5x2jikY8wJQjMRQpRws6AnfWlHm5/view?usp=drive_link
https://drive.google.com/file/d/1OO8GaO-0FrSZRd1kxMYwBmubyiLOWnbl/view?usp=drive_link
https://drive.google.com/file/d/1EXn4NVDmf-4_HCy34mYwT-vwK2CFI9ev/view?usp=drive_link
https://drive.google.com/file/d/10hH70XhXRL9C5SnAG4toHtfHqfJUJo4H/view?usp=drive_link
https://drive.google.com/file/d/18tiBcxea0guUai4lwsXQvt0q2LZ8ZnnJ/view?usp=drive_link
https://drive.google.com/file/d/1Q8R8qv37vk5PQ5kQ2ibx6BFLOySD0VpX/view?usp=drive_link
https://drive.google.com/file/d/17aNriHzjhdibCyuUjQoMFZqjybJZtggG/view?usp=drive_link
https://drive.google.com/file/d/1LVjEYHSdeKm6CotU1QguIeNEPaIaFl_1/view?usp=drive_link
https://drive.google.com/file/d/1ufAhE_EkgJ85slg2EW8aW_grOzE_Lmxd/view?usp=drive_link
https://drive.google.com/file/d/1wtzLtXrkw9eXRGESTPIOlpl1tInu-b2m/view?usp=drive_link
https://drive.google.com/file/d/1Mk5qvVtD_QHwGOUApRq76TUw2T5THu6f/view?usp=drive_link
https://drive.google.com/file/d/1y1WQ3hboWVJ68KEYQQ3OhreGuaUpSgwc/view?usp=drive_link

52
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup_left.txt
View File

@ -1,52 +0,0 @@
https://drive.google.com/drive/folders/1dxWh6YFZUDt6qXIoxgD9bla3CiFjZ11C
https://drive.google.com/file/d/1hNBJN00SCAlOl0ZEgm7RRGbAGDjyBs0p/view?usp=drive_link
https://drive.google.com/file/d/17He0CVwXGeoMmXg4SHKo-osNn7YPKVL7/view?usp=drive_link
https://drive.google.com/file/d/1laNKUVID1x2CV6a2O2WQjwFewKu4lidL/view?usp=drive_link
https://drive.google.com/file/d/1pNf36xbZJGRArYLmNAvRj5y6CoqdC6kB/view?usp=drive_link
https://drive.google.com/file/d/1_4E1-y3JXk5I0ebycLYM70YDPK9g52gZ/view?usp=drive_link
https://drive.google.com/file/d/1PHfzhGPdbolKyOpS3FnR2w7Q8zUlJXSk/view?usp=drive_link
https://drive.google.com/file/d/17ls2PPN-Pi3tEuK059cwV2_iDT8aGhOO/view?usp=drive_link
https://drive.google.com/file/d/1LWsg6PmCT00Kv_N_slrmcwKmQPGoBT3k/view?usp=drive_link
https://drive.google.com/file/d/12LckrchoHTUVH7rxi8J7zD9dA19GXvoW/view?usp=drive_link
https://drive.google.com/file/d/1VqrJKjAIkj5gtFXL69grdSeu9CyaqnSw/view?usp=drive_link
https://drive.google.com/file/d/1g5rQYDBZvW-kUtYPeyF3qmd53v6k7kXu/view?usp=drive_link
https://drive.google.com/file/d/10kUgaSJ0TS7teaG83G3Rf_DG4XGrBt6A/view?usp=drive_link
https://drive.google.com/file/d/1je9XmneZQZvTma5adMJICUPDovW3ppei/view?usp=drive_link
https://drive.google.com/file/d/1v28r6bedwZGbUPVVTVImXhK-42XdtGfj/view?usp=drive_link
https://drive.google.com/file/d/1-TEEx9sGVvzMMaNXYfQMtY2JJ6cvl0dT/view?usp=drive_link
https://drive.google.com/file/d/1YdBKdJFP9rJWBUX7qrOYL_gfUA8o6J9M/view?usp=drive_link
https://drive.google.com/file/d/1X9vffwQHNUSKLXr2RlYNtbWDIFCIDfdF/view?usp=drive_link
https://drive.google.com/file/d/11hqesqa5kvEe5FABUnZRcvmOhR373cYM/view?usp=drive_link
https://drive.google.com/file/d/1ltTTECjEcbQPgS3UPRgMzaE2x9n6H7dC/view?usp=drive_link
https://drive.google.com/file/d/1Zxqfa29JdwT-bfMpivi6IG2vz34d21dD/view?usp=drive_link
https://drive.google.com/file/d/11LQlVxS5hz494dYUJ_PNRPx2NHIJbQns/view?usp=drive_link
https://drive.google.com/file/d/1i1JhNtnZpO_E8rAv8gxBP3ZTZRvcvsZi/view?usp=drive_link
https://drive.google.com/file/d/11jOXAr2EULUO4Qkm748634lg4UUFho5U/view?usp=drive_link
https://drive.google.com/file/d/1rj67wur8DdB_Pipwx24bY43xu4X1eQ5e/view?usp=drive_link
https://drive.google.com/file/d/15ZTm6lO6f_JQy_4SNfrOu3iPYn1Ro8mh/view?usp=drive_link
https://drive.google.com/file/d/1q4gBtqWPJtCwXEvknGgN0WHGp7Vfn1b9/view?usp=drive_link
https://drive.google.com/file/d/1t17keyre47AYqm8GgXiQ7EcvcUkeSiDQ/view?usp=drive_link
https://drive.google.com/file/d/1OYUPGxtZgOF86Ng_BEOTXm_XOYpuQPsO/view?usp=drive_link
https://drive.google.com/file/d/1cBjbGHi3dwWHtx6r9EQJi0JT_CE3LuHt/view?usp=drive_link
https://drive.google.com/file/d/14qaMyF0mcbCB-fCYKNyo5_2NahSC6D5u/view?usp=drive_link
https://drive.google.com/file/d/12FgX86eA7Y5co9ULBVK80XMsiKQSs-Ri/view?usp=drive_link
https://drive.google.com/file/d/1yvoHWidf-jdBVw6qCCXOFfkVwKj_2hPk/view?usp=drive_link
https://drive.google.com/file/d/1a2SugsSDlC8UtUrFzp-_KAwyZckQOvdQ/view?usp=drive_link
https://drive.google.com/file/d/1l8pILBFSAosypWJMza2K09Vm7rug9axm/view?usp=drive_link
https://drive.google.com/file/d/1hfPQ8dBCk97PnOhq6_MIISm3IEzcOxJG/view?usp=drive_link
https://drive.google.com/file/d/1PPAUwlJCFKpms8cqF_k1v2_fCgDBOc3S/view?usp=drive_link
https://drive.google.com/file/d/1lVKQZeqFfK3amEmLuFhYLUFQ2eyE8rOW/view?usp=drive_link
https://drive.google.com/file/d/1K9iPMLfDowcIFoyzpvgn88dQ6x6kVwNG/view?usp=drive_link
https://drive.google.com/file/d/1PNvMqG9tL7QxeLaYBGHiWYR6SYb5iIct/view?usp=drive_link
https://drive.google.com/file/d/1xkRtzbvIkUsylx9hrFLGQsJn0h1EYu-5/view?usp=drive_link
https://drive.google.com/file/d/1nxMRrJlSayjDIfr5CmHO1NzAw3COhsLi/view?usp=drive_link
https://drive.google.com/file/d/1Qs3WEyMGrmagiHIkkFEueWNnJhkUeR1s/view?usp=drive_link
https://drive.google.com/file/d/1D-G2_Q0SS3M8zyJbg_XzkF2ANPw1HTuX/view?usp=drive_link
https://drive.google.com/file/d/1mdmJsDGO-YtJAOF_yPKl6lq4PJOIbQhT/view?usp=drive_link
https://drive.google.com/file/d/11m9bwfop_sPmnQr_8amB6EEsrbAeG_z5/view?usp=drive_link
https://drive.google.com/file/d/19tyYt5FMn5kru0g9o2nMJhKPnsDqkIZv/view?usp=drive_link
https://drive.google.com/file/d/1XvTpUdsVTZ-vydvdYYmynbma--HfUGSl/view?usp=drive_link
https://drive.google.com/file/d/1MO3hFu68J6NohTzr9aB_fY02VA6QSOqj/view?usp=drive_link
https://drive.google.com/file/d/1Lh-UjwAk__04YOTWINF_QGVU8SjetVaY/view?usp=drive_link
https://drive.google.com/file/d/1jkSOUwZV5GJ7rZlVeErjcu0DBQs8Np0d/view?usp=drive_link
https://drive.google.com/file/d/1VIN1eLI-93WrVQwCjsv6XQr353DqqBYA/view?usp=drive_link

8
lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_ziploc_slide.txt
View File

@ -1,8 +0,0 @@
https://drive.google.com/drive/folders/1EgKar7rWBmTIRmeJYZciSwjZx3uP2mHO
https://drive.google.com/file/d/12eYWQO15atK2hBjXhynPJd9MKAj_42pz/view?usp=drive_link
https://drive.google.com/file/d/1Ul4oEeICJDjgfYTl4H1uaisTzVYIM6wd/view?usp=drive_link
https://drive.google.com/file/d/1WSF-OG8lKSe2wVYCv5D1aJNipxpgddk-/view?usp=drive_link
https://drive.google.com/file/d/1_ppD5j5sFh26aWW0JmhLzJMeNB-lCArk/view?usp=drive_link
https://drive.google.com/file/d/1WUp846dgWXYhu4oJfhHxiU6YL_7N6s4W/view?usp=drive_link
https://drive.google.com/file/d/1HRZNAIoAQw_uYiPwnBvtBioQoqiqoXdA/view?usp=drive_link
https://drive.google.com/file/d/1hedGq-QDMnIn8GlXXBC3GiEJ_Y-LTxyt/view?usp=drive_link

634
lerobot/common/datasets/push_dataset_to_hub/_diffusion_policy_replay_buffer.py
View File

@ -1,634 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Helper code for loading PushT dataset from Diffusion Policy (https://diffusion-policy.cs.columbia.edu/)
Copied from the original Diffusion Policy repository and used in our `download_and_upload_dataset.py` script.
"""
from __future__ import annotations
import math
import numbers
import os
from functools import cached_property
import numcodecs
import numpy as np
import zarr
def check_chunks_compatible(chunks: tuple, shape: tuple):
assert len(shape) == len(chunks)
for c in chunks:
assert isinstance(c, numbers.Integral)
assert c > 0
def rechunk_recompress_array(group, name, chunks=None, chunk_length=None, compressor=None, tmp_key="_temp"):
old_arr = group[name]
if chunks is None:
chunks = (chunk_length,) + old_arr.chunks[1:] if chunk_length is not None else old_arr.chunks
check_chunks_compatible(chunks, old_arr.shape)
if compressor is None:
compressor = old_arr.compressor
if (chunks == old_arr.chunks) and (compressor == old_arr.compressor):
# no change
return old_arr
# rechunk recompress
group.move(name, tmp_key)
old_arr = group[tmp_key]
n_copied, n_skipped, n_bytes_copied = zarr.copy(
source=old_arr,
dest=group,
name=name,
chunks=chunks,
compressor=compressor,
)
del group[tmp_key]
arr = group[name]
return arr
def get_optimal_chunks(shape, dtype, target_chunk_bytes=2e6, max_chunk_length=None):
"""
Common shapes
T,D
T,N,D
T,H,W,C
T,N,H,W,C
"""
itemsize = np.dtype(dtype).itemsize
# reversed
rshape = list(shape[::-1])
if max_chunk_length is not None:
rshape[-1] = int(max_chunk_length)
split_idx = len(shape) - 1
for i in range(len(shape) - 1):
this_chunk_bytes = itemsize * np.prod(rshape[:i])
next_chunk_bytes = itemsize * np.prod(rshape[: i + 1])
if this_chunk_bytes <= target_chunk_bytes and next_chunk_bytes > target_chunk_bytes:
split_idx = i
rchunks = rshape[:split_idx]
item_chunk_bytes = itemsize * np.prod(rshape[:split_idx])
this_max_chunk_length = rshape[split_idx]
next_chunk_length = min(this_max_chunk_length, math.ceil(target_chunk_bytes / item_chunk_bytes))
rchunks.append(next_chunk_length)
len_diff = len(shape) - len(rchunks)
rchunks.extend([1] * len_diff)
chunks = tuple(rchunks[::-1])
# print(np.prod(chunks) * itemsize / target_chunk_bytes)
return chunks
class ReplayBuffer:
"""
Zarr-based temporal datastructure.
Assumes first dimension to be time. Only chunk in time dimension.
"""
def __init__(self, root: zarr.Group | dict[str, dict]):
"""
Dummy constructor. Use copy_from* and create_from* class methods instead.
"""
assert "data" in root
assert "meta" in root
assert "episode_ends" in root["meta"]
for value in root["data"].values():
assert value.shape[0] == root["meta"]["episode_ends"][-1]
self.root = root
# ============= create constructors ===============
@classmethod
def create_empty_zarr(cls, storage=None, root=None):
if root is None:
if storage is None:
storage = zarr.MemoryStore()
root = zarr.group(store=storage)
root.require_group("data", overwrite=False)
meta = root.require_group("meta", overwrite=False)
if "episode_ends" not in meta:
meta.zeros("episode_ends", shape=(0,), dtype=np.int64, compressor=None, overwrite=False)
return cls(root=root)
@classmethod
def create_empty_numpy(cls):
root = {"data": {}, "meta": {"episode_ends": np.zeros((0,), dtype=np.int64)}}
return cls(root=root)
@classmethod
def create_from_group(cls, group, **kwargs):
if "data" not in group:
# create from stratch
buffer = cls.create_empty_zarr(root=group, **kwargs)
else:
# already exist
buffer = cls(root=group, **kwargs)
return buffer
@classmethod
def create_from_path(cls, zarr_path, mode="r", **kwargs):
"""
Open a on-disk zarr directly (for dataset larger than memory).
Slower.
"""
group = zarr.open(os.path.expanduser(zarr_path), mode)
return cls.create_from_group(group, **kwargs)
# ============= copy constructors ===============
@classmethod
def copy_from_store(
cls,
src_store,
store=None,
keys=None,
chunks: dict[str, tuple] | None = None,
compressors: dict | str | numcodecs.abc.Codec | None = None,
if_exists="replace",
**kwargs,
):
"""
Load to memory.
"""
src_root = zarr.group(src_store)
if chunks is None:
chunks = {}
if compressors is None:
compressors = {}
root = None
if store is None:
# numpy backend
meta = {}
for key, value in src_root["meta"].items():
if len(value.shape) == 0:
meta[key] = np.array(value)
else:
meta[key] = value[:]
if keys is None:
keys = src_root["data"].keys()
data = {}
for key in keys:
arr = src_root["data"][key]
data[key] = arr[:]
root = {"meta": meta, "data": data}
else:
root = zarr.group(store=store)
# copy without recompression
n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
source=src_store, dest=store, source_path="/meta", dest_path="/meta", if_exists=if_exists
)
data_group = root.create_group("data", overwrite=True)
if keys is None:
keys = src_root["data"].keys()
for key in keys:
value = src_root["data"][key]
cks = cls._resolve_array_chunks(chunks=chunks, key=key, array=value)
cpr = cls._resolve_array_compressor(compressors=compressors, key=key, array=value)
if cks == value.chunks and cpr == value.compressor:
# copy without recompression
this_path = "/data/" + key
n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
source=src_store,
dest=store,
source_path=this_path,
dest_path=this_path,
if_exists=if_exists,
)
else:
# copy with recompression
n_copied, n_skipped, n_bytes_copied = zarr.copy(
source=value,
dest=data_group,
name=key,
chunks=cks,
compressor=cpr,
if_exists=if_exists,
)
buffer = cls(root=root)
return buffer
@classmethod
def copy_from_path(
cls,
zarr_path,
backend=None,
store=None,
keys=None,
chunks: dict[str, tuple] | None = None,
compressors: dict | str | numcodecs.abc.Codec | None = None,
if_exists="replace",
**kwargs,
):
"""
Copy a on-disk zarr to in-memory compressed.
Recommended
"""
if chunks is None:
chunks = {}
if compressors is None:
compressors = {}
if backend == "numpy":
print("backend argument is deprecated!")
store = None
group = zarr.open(os.path.expanduser(zarr_path), "r")
return cls.copy_from_store(
src_store=group.store,
store=store,
keys=keys,
chunks=chunks,
compressors=compressors,
if_exists=if_exists,
**kwargs,
)
# ============= save methods ===============
def save_to_store(
self,
store,
chunks: dict[str, tuple] | None = None,
compressors: str | numcodecs.abc.Codec | dict | None = None,
if_exists="replace",
**kwargs,
):
root = zarr.group(store)
if chunks is None:
chunks = {}
if compressors is None:
compressors = {}
if self.backend == "zarr":
# recompression free copy
n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
source=self.root.store,
dest=store,
source_path="/meta",
dest_path="/meta",
if_exists=if_exists,
)
else:
meta_group = root.create_group("meta", overwrite=True)
# save meta, no chunking
for key, value in self.root["meta"].items():
_ = meta_group.array(name=key, data=value, shape=value.shape, chunks=value.shape)
# save data, chunk
data_group = root.create_group("data", overwrite=True)
for key, value in self.root["data"].items():
cks = self._resolve_array_chunks(chunks=chunks, key=key, array=value)
cpr = self._resolve_array_compressor(compressors=compressors, key=key, array=value)
if isinstance(value, zarr.Array):
if cks == value.chunks and cpr == value.compressor:
# copy without recompression
this_path = "/data/" + key
n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
source=self.root.store,
dest=store,
source_path=this_path,
dest_path=this_path,
if_exists=if_exists,
)
else:
# copy with recompression
n_copied, n_skipped, n_bytes_copied = zarr.copy(
source=value,
dest=data_group,
name=key,
chunks=cks,
compressor=cpr,
if_exists=if_exists,
)
else:
# numpy
_ = data_group.array(name=key, data=value, chunks=cks, compressor=cpr)
return store
def save_to_path(
self,
zarr_path,
chunks: dict[str, tuple] | None = None,
compressors: str | numcodecs.abc.Codec | dict | None = None,
if_exists="replace",
**kwargs,
):
if chunks is None:
chunks = {}
if compressors is None:
compressors = {}
store = zarr.DirectoryStore(os.path.expanduser(zarr_path))
return self.save_to_store(
store, chunks=chunks, compressors=compressors, if_exists=if_exists, **kwargs
)
@staticmethod
def resolve_compressor(compressor="default"):
if compressor == "default":
compressor = numcodecs.Blosc(cname="lz4", clevel=5, shuffle=numcodecs.Blosc.NOSHUFFLE)
elif compressor == "disk":
compressor = numcodecs.Blosc("zstd", clevel=5, shuffle=numcodecs.Blosc.BITSHUFFLE)
return compressor
@classmethod
def _resolve_array_compressor(cls, compressors: dict | str | numcodecs.abc.Codec, key, array):
# allows compressor to be explicitly set to None
cpr = "nil"
if isinstance(compressors, dict):
if key in compressors:
cpr = cls.resolve_compressor(compressors[key])
elif isinstance(array, zarr.Array):
cpr = array.compressor
else:
cpr = cls.resolve_compressor(compressors)
# backup default
if cpr == "nil":
cpr = cls.resolve_compressor("default")
return cpr
@classmethod
def _resolve_array_chunks(cls, chunks: dict | tuple, key, array):
cks = None
if isinstance(chunks, dict):
if key in chunks:
cks = chunks[key]
elif isinstance(array, zarr.Array):
cks = array.chunks
elif isinstance(chunks, tuple):
cks = chunks
else:
raise TypeError(f"Unsupported chunks type {type(chunks)}")
# backup default
if cks is None:
cks = get_optimal_chunks(shape=array.shape, dtype=array.dtype)
# check
check_chunks_compatible(chunks=cks, shape=array.shape)
return cks
# ============= properties =================
@cached_property
def data(self):
return self.root["data"]
@cached_property
def meta(self):
return self.root["meta"]
def update_meta(self, data):
# sanitize data
np_data = {}
for key, value in data.items():
if isinstance(value, np.ndarray):
np_data[key] = value
else:
arr = np.array(value)
if arr.dtype == object:
raise TypeError(f"Invalid value type {type(value)}")
np_data[key] = arr
meta_group = self.meta
if self.backend == "zarr":
for key, value in np_data.items():
_ = meta_group.array(
name=key, data=value, shape=value.shape, chunks=value.shape, overwrite=True
)
else:
meta_group.update(np_data)
return meta_group
@property
def episode_ends(self):
return self.meta["episode_ends"]
def get_episode_idxs(self):
import numba
numba.jit(nopython=True)
def _get_episode_idxs(episode_ends):
result = np.zeros((episode_ends[-1],), dtype=np.int64)
for i in range(len(episode_ends)):
start = 0
if i > 0:
start = episode_ends[i - 1]
end = episode_ends[i]
for idx in range(start, end):
result[idx] = i
return result
return _get_episode_idxs(self.episode_ends)
@property
def backend(self):
backend = "numpy"
if isinstance(self.root, zarr.Group):
backend = "zarr"
return backend
# =========== dict-like API ==============
def __repr__(self) -> str:
if self.backend == "zarr":
return str(self.root.tree())
else:
return super().__repr__()
def keys(self):
return self.data.keys()
def values(self):
return self.data.values()
def items(self):
return self.data.items()
def __getitem__(self, key):
return self.data[key]
def __contains__(self, key):
return key in self.data
# =========== our API ==============
@property
def n_steps(self):
if len(self.episode_ends) == 0:
return 0
return self.episode_ends[-1]
@property
def n_episodes(self):
return len(self.episode_ends)
@property
def chunk_size(self):
if self.backend == "zarr":
return next(iter(self.data.arrays()))[-1].chunks[0]
return None
@property
def episode_lengths(self):
ends = self.episode_ends[:]
ends = np.insert(ends, 0, 0)
lengths = np.diff(ends)
return lengths
def add_episode(
self,
data: dict[str, np.ndarray],
chunks: dict[str, tuple] | None = None,
compressors: str | numcodecs.abc.Codec | dict | None = None,
):
if chunks is None:
chunks = {}
if compressors is None:
compressors = {}
assert len(data) > 0
is_zarr = self.backend == "zarr"
curr_len = self.n_steps
episode_length = None
for value in data.values():
assert len(value.shape) >= 1
if episode_length is None:
episode_length = len(value)
else:
assert episode_length == len(value)
new_len = curr_len + episode_length
for key, value in data.items():
new_shape = (new_len,) + value.shape[1:]
# create array
if key not in self.data:
if is_zarr:
cks = self._resolve_array_chunks(chunks=chunks, key=key, array=value)
cpr = self._resolve_array_compressor(compressors=compressors, key=key, array=value)
arr = self.data.zeros(
name=key, shape=new_shape, chunks=cks, dtype=value.dtype, compressor=cpr
)
else:
# copy data to prevent modify
arr = np.zeros(shape=new_shape, dtype=value.dtype)
self.data[key] = arr
else:
arr = self.data[key]
assert value.shape[1:] == arr.shape[1:]
# same method for both zarr and numpy
if is_zarr:
arr.resize(new_shape)
else:
arr.resize(new_shape, refcheck=False)
# copy data
arr[-value.shape[0] :] = value
# append to episode ends
episode_ends = self.episode_ends
if is_zarr:
episode_ends.resize(episode_ends.shape[0] + 1)
else:
episode_ends.resize(episode_ends.shape[0] + 1, refcheck=False)
episode_ends[-1] = new_len
# rechunk
if is_zarr and episode_ends.chunks[0] < episode_ends.shape[0]:
rechunk_recompress_array(self.meta, "episode_ends", chunk_length=int(episode_ends.shape[0] * 1.5))
def drop_episode(self):
is_zarr = self.backend == "zarr"
episode_ends = self.episode_ends[:].copy()
assert len(episode_ends) > 0
start_idx = 0
if len(episode_ends) > 1:
start_idx = episode_ends[-2]
for value in self.data.values():
new_shape = (start_idx,) + value.shape[1:]
if is_zarr:
value.resize(new_shape)
else:
value.resize(new_shape, refcheck=False)
if is_zarr:
self.episode_ends.resize(len(episode_ends) - 1)
else:
self.episode_ends.resize(len(episode_ends) - 1, refcheck=False)
def pop_episode(self):
assert self.n_episodes > 0
episode = self.get_episode(self.n_episodes - 1, copy=True)
self.drop_episode()
return episode
def extend(self, data):
self.add_episode(data)
def get_episode(self, idx, copy=False):
idx = list(range(len(self.episode_ends)))[idx]
start_idx = 0
if idx > 0:
start_idx = self.episode_ends[idx - 1]
end_idx = self.episode_ends[idx]
result = self.get_steps_slice(start_idx, end_idx, copy=copy)
return result
def get_episode_slice(self, idx):
start_idx = 0
if idx > 0:
start_idx = self.episode_ends[idx - 1]
end_idx = self.episode_ends[idx]
return slice(start_idx, end_idx)
def get_steps_slice(self, start, stop, step=None, copy=False):
_slice = slice(start, stop, step)
result = {}
for key, value in self.data.items():
x = value[_slice]
if copy and isinstance(value, np.ndarray):
x = x.copy()
result[key] = x
return result
# =========== chunking =============
def get_chunks(self) -> dict:
assert self.backend == "zarr"
chunks = {}
for key, value in self.data.items():
chunks[key] = value.chunks
return chunks
def set_chunks(self, chunks: dict):
assert self.backend == "zarr"
for key, value in chunks.items():
if key in self.data:
arr = self.data[key]
if value != arr.chunks:
check_chunks_compatible(chunks=value, shape=arr.shape)
rechunk_recompress_array(self.data, key, chunks=value)
def get_compressors(self) -> dict:
assert self.backend == "zarr"
compressors = {}
for key, value in self.data.items():
compressors[key] = value.compressor
return compressors
def set_compressors(self, compressors: dict):
assert self.backend == "zarr"
for key, value in compressors.items():
if key in self.data:
arr = self.data[key]
compressor = self.resolve_compressor(value)
if compressor != arr.compressor:
rechunk_recompress_array(self.data, key, compressor=compressor)

202
lerobot/common/datasets/push_dataset_to_hub/_download_raw.py
View File

@ -1,202 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This file contains download scripts for raw datasets.
Example of usage:
```
python lerobot/common/datasets/push_dataset_to_hub/_download_raw.py \
--raw-dir data/lerobot-raw/pusht_raw \
--repo-id lerobot-raw/pusht_raw
```
"""
import argparse
import logging
import warnings
from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
# {raw_repo_id: raw_format}
AVAILABLE_RAW_REPO_IDS = {
"lerobot-raw/aloha_mobile_cabinet_raw": "aloha_hdf5",
"lerobot-raw/aloha_mobile_chair_raw": "aloha_hdf5",
"lerobot-raw/aloha_mobile_elevator_raw": "aloha_hdf5",
"lerobot-raw/aloha_mobile_shrimp_raw": "aloha_hdf5",
"lerobot-raw/aloha_mobile_wash_pan_raw": "aloha_hdf5",
"lerobot-raw/aloha_mobile_wipe_wine_raw": "aloha_hdf5",
"lerobot-raw/aloha_sim_insertion_human_raw": "aloha_hdf5",
"lerobot-raw/aloha_sim_insertion_scripted_raw": "aloha_hdf5",
"lerobot-raw/aloha_sim_transfer_cube_human_raw": "aloha_hdf5",
"lerobot-raw/aloha_sim_transfer_cube_scripted_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_battery_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_candy_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_coffee_new_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_coffee_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_cups_open_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_fork_pick_up_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_pingpong_test_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_pro_pencil_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_screw_driver_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_tape_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_thread_velcro_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_towel_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_vinh_cup_left_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_vinh_cup_raw": "aloha_hdf5",
"lerobot-raw/aloha_static_ziploc_slide_raw": "aloha_hdf5",
"lerobot-raw/umi_cup_in_the_wild_raw": "umi_zarr",
"lerobot-raw/pusht_raw": "pusht_zarr",
"lerobot-raw/unitreeh1_fold_clothes_raw": "aloha_hdf5",
"lerobot-raw/unitreeh1_rearrange_objects_raw": "aloha_hdf5",
"lerobot-raw/unitreeh1_two_robot_greeting_raw": "aloha_hdf5",
"lerobot-raw/unitreeh1_warehouse_raw": "aloha_hdf5",
"lerobot-raw/xarm_lift_medium_raw": "xarm_pkl",
"lerobot-raw/xarm_lift_medium_replay_raw": "xarm_pkl",
"lerobot-raw/xarm_push_medium_raw": "xarm_pkl",
"lerobot-raw/xarm_push_medium_replay_raw": "xarm_pkl",
"lerobot-raw/fractal20220817_data_raw": "openx_rlds.fractal20220817_data",
"lerobot-raw/kuka_raw": "openx_rlds.kuka",
"lerobot-raw/bridge_openx_raw": "openx_rlds.bridge_openx",
"lerobot-raw/taco_play_raw": "openx_rlds.taco_play",
"lerobot-raw/jaco_play_raw": "openx_rlds.jaco_play",
"lerobot-raw/berkeley_cable_routing_raw": "openx_rlds.berkeley_cable_routing",
"lerobot-raw/roboturk_raw": "openx_rlds.roboturk",
"lerobot-raw/nyu_door_opening_surprising_effectiveness_raw": "openx_rlds.nyu_door_opening_surprising_effectiveness",
"lerobot-raw/viola_raw": "openx_rlds.viola",
"lerobot-raw/berkeley_autolab_ur5_raw": "openx_rlds.berkeley_autolab_ur5",
"lerobot-raw/toto_raw": "openx_rlds.toto",
"lerobot-raw/language_table_raw": "openx_rlds.language_table",
"lerobot-raw/columbia_cairlab_pusht_real_raw": "openx_rlds.columbia_cairlab_pusht_real",
"lerobot-raw/stanford_kuka_multimodal_dataset_raw": "openx_rlds.stanford_kuka_multimodal_dataset",
"lerobot-raw/nyu_rot_dataset_raw": "openx_rlds.nyu_rot_dataset",
"lerobot-raw/io_ai_tech_raw": "openx_rlds.io_ai_tech",
"lerobot-raw/stanford_hydra_dataset_raw": "openx_rlds.stanford_hydra_dataset",
"lerobot-raw/austin_buds_dataset_raw": "openx_rlds.austin_buds_dataset",
"lerobot-raw/nyu_franka_play_dataset_raw": "openx_rlds.nyu_franka_play_dataset",
"lerobot-raw/maniskill_dataset_raw": "openx_rlds.maniskill_dataset",
"lerobot-raw/furniture_bench_dataset_raw": "openx_rlds.furniture_bench_dataset",
"lerobot-raw/cmu_franka_exploration_dataset_raw": "openx_rlds.cmu_franka_exploration_dataset",
"lerobot-raw/ucsd_kitchen_dataset_raw": "openx_rlds.ucsd_kitchen_dataset",
"lerobot-raw/ucsd_pick_and_place_dataset_raw": "openx_rlds.ucsd_pick_and_place_dataset",
"lerobot-raw/spoc_raw": "openx_rlds.spoc",
"lerobot-raw/austin_sailor_dataset_raw": "openx_rlds.austin_sailor_dataset",
"lerobot-raw/austin_sirius_dataset_raw": "openx_rlds.austin_sirius_dataset",
"lerobot-raw/bc_z_raw": "openx_rlds.bc_z",
"lerobot-raw/utokyo_pr2_opening_fridge_raw": "openx_rlds.utokyo_pr2_opening_fridge",
"lerobot-raw/utokyo_pr2_tabletop_manipulation_raw": "openx_rlds.utokyo_pr2_tabletop_manipulation",
"lerobot-raw/utokyo_xarm_pick_and_place_raw": "openx_rlds.utokyo_xarm_pick_and_place",
"lerobot-raw/utokyo_xarm_bimanual_raw": "openx_rlds.utokyo_xarm_bimanual",
"lerobot-raw/utokyo_saytap_raw": "openx_rlds.utokyo_saytap",
"lerobot-raw/robo_net_raw": "openx_rlds.robo_net",
"lerobot-raw/robo_set_raw": "openx_rlds.robo_set",
"lerobot-raw/berkeley_mvp_raw": "openx_rlds.berkeley_mvp",
"lerobot-raw/berkeley_rpt_raw": "openx_rlds.berkeley_rpt",
"lerobot-raw/kaist_nonprehensile_raw": "openx_rlds.kaist_nonprehensile",
"lerobot-raw/stanford_mask_vit_raw": "openx_rlds.stanford_mask_vit",
"lerobot-raw/tokyo_u_lsmo_raw": "openx_rlds.tokyo_u_lsmo",
"lerobot-raw/dlr_sara_pour_raw": "openx_rlds.dlr_sara_pour",
"lerobot-raw/dlr_sara_grid_clamp_raw": "openx_rlds.dlr_sara_grid_clamp",
"lerobot-raw/dlr_edan_shared_control_raw": "openx_rlds.dlr_edan_shared_control",
"lerobot-raw/asu_table_top_raw": "openx_rlds.asu_table_top",
"lerobot-raw/stanford_robocook_raw": "openx_rlds.stanford_robocook",
"lerobot-raw/imperialcollege_sawyer_wrist_cam_raw": "openx_rlds.imperialcollege_sawyer_wrist_cam",
"lerobot-raw/iamlab_cmu_pickup_insert_raw": "openx_rlds.iamlab_cmu_pickup_insert",
"lerobot-raw/uiuc_d3field_raw": "openx_rlds.uiuc_d3field",
"lerobot-raw/utaustin_mutex_raw": "openx_rlds.utaustin_mutex",
"lerobot-raw/berkeley_fanuc_manipulation_raw": "openx_rlds.berkeley_fanuc_manipulation",
"lerobot-raw/cmu_playing_with_food_raw": "openx_rlds.cmu_playing_with_food",
"lerobot-raw/cmu_play_fusion_raw": "openx_rlds.cmu_play_fusion",
"lerobot-raw/cmu_stretch_raw": "openx_rlds.cmu_stretch",
"lerobot-raw/berkeley_gnm_recon_raw": "openx_rlds.berkeley_gnm_recon",
"lerobot-raw/berkeley_gnm_cory_hall_raw": "openx_rlds.berkeley_gnm_cory_hall",
"lerobot-raw/berkeley_gnm_sac_son_raw": "openx_rlds.berkeley_gnm_sac_son",
"lerobot-raw/droid_raw": "openx_rlds.droid",
"lerobot-raw/droid_100_raw": "openx_rlds.droid100",
"lerobot-raw/fmb_raw": "openx_rlds.fmb",
"lerobot-raw/dobbe_raw": "openx_rlds.dobbe",
"lerobot-raw/usc_cloth_sim_raw": "openx_rlds.usc_cloth_sim",
"lerobot-raw/plex_robosuite_raw": "openx_rlds.plex_robosuite",
"lerobot-raw/conq_hose_manipulation_raw": "openx_rlds.conq_hose_manipulation",
"lerobot-raw/vima_raw": "openx_rlds.vima",
"lerobot-raw/robot_vqa_raw": "openx_rlds.robot_vqa",
"lerobot-raw/mimic_play_raw": "openx_rlds.mimic_play",
"lerobot-raw/tidybot_raw": "openx_rlds.tidybot",
"lerobot-raw/eth_agent_affordances_raw": "openx_rlds.eth_agent_affordances",
}
def download_raw(raw_dir: Path, repo_id: str):
check_repo_id(repo_id)
user_id, dataset_id = repo_id.split("/")
if not dataset_id.endswith("_raw"):
warnings.warn(
f"""`dataset_id` ({dataset_id}) doesn't end with '_raw' (e.g. 'lerobot/pusht_raw'). Following this
naming convention by renaming your repository is advised, but not mandatory.""",
stacklevel=1,
)
# Send warning if raw_dir isn't well formated
if raw_dir.parts[-2] != user_id or raw_dir.parts[-1] != dataset_id:
warnings.warn(
f"""`raw_dir` ({raw_dir}) doesn't contain a community or user id `/` the name of the dataset that
match the `repo_id` (e.g. 'data/lerobot/pusht_raw'). Following this naming convention is advised,
but not mandatory.""",
stacklevel=1,
)
raw_dir.mkdir(parents=True, exist_ok=True)
logging.info(f"Start downloading from huggingface.co/{user_id} for {dataset_id}")
snapshot_download(repo_id, repo_type="dataset", local_dir=raw_dir)
logging.info(f"Finish downloading from huggingface.co/{user_id} for {dataset_id}")
def download_all_raw_datasets(data_dir: Path | None = None):
if data_dir is None:
data_dir = Path("data")
for repo_id in AVAILABLE_RAW_REPO_IDS:
raw_dir = data_dir / repo_id
download_raw(raw_dir, repo_id)
def main():
parser = argparse.ArgumentParser(
description=f"""A script to download raw datasets from Hugging Face hub to a local directory. Here is a
non exhaustive list of available repositories to use in `--repo-id`: {list(AVAILABLE_RAW_REPO_IDS.keys())}""",
)
parser.add_argument(
"--raw-dir",
type=Path,
required=True,
help="Directory containing input raw datasets (e.g. `data/aloha_mobile_chair_raw` or `data/pusht_raw).",
)
parser.add_argument(
"--repo-id",
type=str,
required=True,
help="""Repositery identifier on Hugging Face: a community or a user name `/` the name of
the dataset (e.g. `lerobot/pusht_raw`, `cadene/aloha_sim_insertion_human_raw`).""",
)
args = parser.parse_args()
download_raw(**vars(args))
if __name__ == "__main__":
main()

184
lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py
View File

@ -1,184 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Use this script to batch encode lerobot dataset from their raw format to LeRobotDataset and push their updated
version to the hub. Under the hood, this script reuses 'push_dataset_to_hub.py'. It assumes that you already
downloaded raw datasets, which you can do with the related '_download_raw.py' script.
For instance, for codebase_version = 'v1.6', the following command was run, assuming raw datasets from
lerobot-raw were downloaded in 'raw/datasets/directory':
```bash
python lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py \
--raw-dir raw/datasets/directory \
--raw-repo-ids lerobot-raw \
--local-dir push/datasets/directory \
--tests-data-dir tests/data \
--push-repo lerobot \
--vcodec libsvtav1 \
--pix-fmt yuv420p \
--g 2 \
--crf 30
```
"""
import argparse
from pathlib import Path
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub._download_raw import AVAILABLE_RAW_REPO_IDS
from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
from lerobot.scripts.push_dataset_to_hub import push_dataset_to_hub
def get_push_repo_id_from_raw(raw_repo_id: str, push_repo: str) -> str:
dataset_id_raw = raw_repo_id.split("/")[1]
dataset_id = dataset_id_raw.removesuffix("_raw")
return f"{push_repo}/{dataset_id}"
def encode_datasets(
raw_dir: Path,
raw_repo_ids: list[str],
push_repo: str,
vcodec: str,
pix_fmt: str,
g: int,
crf: int,
local_dir: Path | None = None,
tests_data_dir: Path | None = None,
raw_format: str | None = None,
dry_run: bool = False,
) -> None:
if len(raw_repo_ids) == 1 and raw_repo_ids[0].lower() == "lerobot-raw":
raw_repo_ids_format = AVAILABLE_RAW_REPO_IDS
else:
if raw_format is None:
raise ValueError(raw_format)
raw_repo_ids_format = {id_: raw_format for id_ in raw_repo_ids}
for raw_repo_id, repo_raw_format in raw_repo_ids_format.items():
check_repo_id(raw_repo_id)
dataset_repo_id_push = get_push_repo_id_from_raw(raw_repo_id, push_repo)
dataset_raw_dir = raw_dir / raw_repo_id
dataset_dir = local_dir / dataset_repo_id_push if local_dir is not None else None
encoding = {
"vcodec": vcodec,
"pix_fmt": pix_fmt,
"g": g,
"crf": crf,
}
if not (dataset_raw_dir).is_dir():
raise NotADirectoryError(dataset_raw_dir)
if not dry_run:
push_dataset_to_hub(
dataset_raw_dir,
raw_format=repo_raw_format,
repo_id=dataset_repo_id_push,
local_dir=dataset_dir,
resume=True,
encoding=encoding,
tests_data_dir=tests_data_dir,
)
else:
print(
f"DRY RUN: {dataset_raw_dir} --> {dataset_dir} --> {dataset_repo_id_push}@{CODEBASE_VERSION}"
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--raw-dir",
type=Path,
default=Path("data"),
help="Directory where raw datasets are located.",
)
parser.add_argument(
"--raw-repo-ids",
type=str,
nargs="*",
default=["lerobot-raw"],
help="""Raw dataset repo ids. if 'lerobot-raw', the keys from `AVAILABLE_RAW_REPO_IDS` will be
used and raw datasets will be fetched from the 'lerobot-raw/' repo and pushed with their
associated format. It is assumed that each dataset is located at `raw_dir / raw_repo_id` """,
)
parser.add_argument(
"--raw-format",
type=str,
default=None,
help="""Raw format to use for the raw repo-ids. Must be specified if --raw-repo-ids is not
'lerobot-raw'""",
)
parser.add_argument(
"--local-dir",
type=Path,
default=None,
help="""When provided, writes the dataset converted to LeRobotDataset format in this directory
(e.g. `data/lerobot/aloha_mobile_chair`).""",
)
parser.add_argument(
"--push-repo",
type=str,
default="lerobot",
help="Repo to upload datasets to",
)
parser.add_argument(
"--vcodec",
type=str,
default="libsvtav1",
help="Codec to use for encoding videos",
)
parser.add_argument(
"--pix-fmt",
type=str,
default="yuv420p",
help="Pixel formats (chroma subsampling) to be used for encoding",
)
parser.add_argument(
"--g",
type=int,
default=2,
help="Group of pictures sizes to be used for encoding.",
)
parser.add_argument(
"--crf",
type=int,
default=30,
help="Constant rate factors to be used for encoding.",
)
parser.add_argument(
"--tests-data-dir",
type=Path,
default=None,
help=(
"When provided, save tests artifacts into the given directory "
"(e.g. `--tests-data-dir tests/data` will save to tests/data/{--repo-id})."
),
)
parser.add_argument(
"--dry-run",
type=int,
default=0,
help="If not set to 0, this script won't download or upload anything.",
)
args = parser.parse_args()
encode_datasets(**vars(args))
if __name__ == "__main__":
main()

326
lerobot/common/datasets/push_dataset_to_hub/_umi_imagecodecs_numcodecs.py
View File

@ -1,326 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# imagecodecs/numcodecs.py
# Copyright (c) 2021-2022, Christoph Gohlke
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its
# contributors may be used to endorse or promote products derived from
# this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
# Copied from: https://github.com/real-stanford/universal_manipulation_interface/blob/298776ce251f33b6b3185a98d6e7d1f9ad49168b/diffusion_policy/codecs/imagecodecs_numcodecs.py#L1
"""Additional numcodecs implemented using imagecodecs."""
__version__ = "2022.9.26"
__all__ = ("register_codecs",)
import imagecodecs
import numpy
from numcodecs.abc import Codec
from numcodecs.registry import get_codec, register_codec
# TODO (azouitine): Remove useless codecs
def protective_squeeze(x: numpy.ndarray):
"""
Squeeze dim only if it's not the last dim.
Image dim expected to be *, H, W, C
"""
img_shape = x.shape[-3:]
if len(x.shape) > 3:
n_imgs = numpy.prod(x.shape[:-3])
if n_imgs > 1:
img_shape = (-1,) + img_shape
return x.reshape(img_shape)
def get_default_image_compressor(**kwargs):
if imagecodecs.JPEGXL:
# has JPEGXL
this_kwargs = {
"effort": 3,
"distance": 0.3,
# bug in libjxl, invalid codestream for non-lossless
# when decoding speed > 1
"decodingspeed": 1,
}
this_kwargs.update(kwargs)
return JpegXl(**this_kwargs)
else:
this_kwargs = {"level": 50}
this_kwargs.update(kwargs)
return Jpeg2k(**this_kwargs)
class Jpeg2k(Codec):
"""JPEG 2000 codec for numcodecs."""
codec_id = "imagecodecs_jpeg2k"
def __init__(
self,
level=None,
codecformat=None,
colorspace=None,
tile=None,
reversible=None,
bitspersample=None,
resolutions=None,
numthreads=None,
verbose=0,
):
self.level = level
self.codecformat = codecformat
self.colorspace = colorspace
self.tile = None if tile is None else tuple(tile)
self.reversible = reversible
self.bitspersample = bitspersample
self.resolutions = resolutions
self.numthreads = numthreads
self.verbose = verbose
def encode(self, buf):
buf = protective_squeeze(numpy.asarray(buf))
return imagecodecs.jpeg2k_encode(
buf,
level=self.level,
codecformat=self.codecformat,
colorspace=self.colorspace,
tile=self.tile,
reversible=self.reversible,
bitspersample=self.bitspersample,
resolutions=self.resolutions,
numthreads=self.numthreads,
verbose=self.verbose,
)
def decode(self, buf, out=None):
return imagecodecs.jpeg2k_decode(buf, verbose=self.verbose, numthreads=self.numthreads, out=out)
class JpegXl(Codec):
"""JPEG XL codec for numcodecs."""
codec_id = "imagecodecs_jpegxl"
def __init__(
self,
# encode
level=None,
effort=None,
distance=None,
lossless=None,
decodingspeed=None,
photometric=None,
planar=None,
usecontainer=None,
# decode
index=None,
keeporientation=None,
# both
numthreads=None,
):
"""
Return JPEG XL image from numpy array.
Float must be in nominal range 0..1.
Currently L, LA, RGB, RGBA images are supported in contig mode.
Extra channels are only supported for grayscale images in planar mode.
Parameters
----------
level : Default to None, i.e. not overwriting lossess and decodingspeed options.
When < 0: Use lossless compression
When in [0,1,2,3,4]: Sets the decoding speed tier for the provided options.
Minimum is 0 (slowest to decode, best quality/density), and maximum
is 4 (fastest to decode, at the cost of some quality/density).
effort : Default to 3.
Sets encoder effort/speed level without affecting decoding speed.
Valid values are, from faster to slower speed: 1:lightning 2:thunder
3:falcon 4:cheetah 5:hare 6:wombat 7:squirrel 8:kitten 9:tortoise.
Speed: lightning, thunder, falcon, cheetah, hare, wombat, squirrel, kitten, tortoise
control the encoder effort in ascending order.
This also affects memory usage: using lower effort will typically reduce memory
consumption during encoding.
lightning and thunder are fast modes useful for lossless mode (modular).
falcon disables all of the following tools.
cheetah enables coefficient reordering, context clustering, and heuristics for selecting DCT sizes and quantization steps.
hare enables Gaborish filtering, chroma from luma, and an initial estimate of quantization steps.
wombat enables error diffusion quantization and full DCT size selection heuristics.
squirrel (default) enables dots, patches, and spline detection, and full context clustering.
kitten optimizes the adaptive quantization for a psychovisual metric.
tortoise enables a more thorough adaptive quantization search.
distance : Default to 1.0
Sets the distance level for lossy compression: target max butteraugli distance,
lower = higher quality. Range: 0 .. 15. 0.0 = mathematically lossless
(however, use JxlEncoderSetFrameLossless instead to use true lossless,
as setting distance to 0 alone is not the only requirement).
1.0 = visually lossless. Recommended range: 0.5 .. 3.0.
lossess : Default to False.
Use lossess encoding.
decodingspeed : Default to 0.
Duplicate to level. [0,4]
photometric : Return JxlColorSpace value.
Default logic is quite complicated but works most of the time.
Accepted value:
int: [-1,3]
str: ['RGB',
'WHITEISZERO', 'MINISWHITE',
'BLACKISZERO', 'MINISBLACK', 'GRAY',
'XYB', 'KNOWN']
planar : Enable multi-channel mode.
Default to false.
usecontainer :
Forces the encoder to use the box-based container format (BMFF)
even when not necessary.
When using JxlEncoderUseBoxes, JxlEncoderStoreJPEGMetadata or
JxlEncoderSetCodestreamLevel with level 10, the encoder will
automatically also use the container format, it is not necessary
to use JxlEncoderUseContainer for those use cases.
By default this setting is disabled.
index : Selectively decode frames for animation.
Default to 0, decode all frames.
When set to > 0, decode that frame index only.
keeporientation :
Enables or disables preserving of as-in-bitstream pixeldata orientation.
Some images are encoded with an Orientation tag indicating that the
decoder must perform a rotation and/or mirroring to the encoded image data.
If skip_reorientation is JXL_FALSE (the default): the decoder will apply
the transformation from the orientation setting, hence rendering the image
according to its specified intent. When producing a JxlBasicInfo, the decoder
will always set the orientation field to JXL_ORIENT_IDENTITY (matching the
returned pixel data) and also align xsize and ysize so that they correspond
to the width and the height of the returned pixel data.
If skip_reorientation is JXL_TRUE: the decoder will skip applying the
transformation from the orientation setting, returning the image in
the as-in-bitstream pixeldata orientation. This may be faster to decode
since the decoder doesnt have to apply the transformation, but can
cause wrong display of the image if the orientation tag is not correctly
taken into account by the user.
By default, this option is disabled, and the returned pixel data is
re-oriented according to the images Orientation setting.
threads : Default to 1.
If <= 0, use all cores.
If > 32, clipped to 32.
"""
self.level = level
self.effort = effort
self.distance = distance
self.lossless = bool(lossless)
self.decodingspeed = decodingspeed
self.photometric = photometric
self.planar = planar
self.usecontainer = usecontainer
self.index = index
self.keeporientation = keeporientation
self.numthreads = numthreads
def encode(self, buf):
# TODO: only squeeze all but last dim
buf = protective_squeeze(numpy.asarray(buf))
return imagecodecs.jpegxl_encode(
buf,
level=self.level,
effort=self.effort,
distance=self.distance,
lossless=self.lossless,
decodingspeed=self.decodingspeed,
photometric=self.photometric,
planar=self.planar,
usecontainer=self.usecontainer,
numthreads=self.numthreads,
)
def decode(self, buf, out=None):
return imagecodecs.jpegxl_decode(
buf,
index=self.index,
keeporientation=self.keeporientation,
numthreads=self.numthreads,
out=out,
)
def _flat(out):
"""Return numpy array as contiguous view of bytes if possible."""
if out is None:
return None
view = memoryview(out)
if view.readonly or not view.contiguous:
return None
return view.cast("B")
def register_codecs(codecs=None, force=False, verbose=True):
"""Register codecs in this module with numcodecs."""
for name, cls in globals().items():
if not hasattr(cls, "codec_id") or name == "Codec":
continue
if codecs is not None and cls.codec_id not in codecs:
continue
try:
try: # noqa: SIM105
get_codec({"id": cls.codec_id})
except TypeError:
# registered, but failed
pass
except ValueError:
# not registered yet
pass
else:
if not force:
if verbose:
log_warning(f"numcodec {cls.codec_id!r} already registered")
continue
if verbose:
log_warning(f"replacing registered numcodec {cls.codec_id!r}")
register_codec(cls)
def log_warning(msg, *args, **kwargs):
"""Log message with level WARNING."""
import logging
logging.getLogger(__name__).warning(msg, *args, **kwargs)

233
lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py
View File

@ -1,233 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Contains utilities to process raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act
"""
import gc
import shutil
from pathlib import Path
import h5py
import numpy as np
import torch
import tqdm
from datasets import Dataset, Features, Image, Sequence, Value
from PIL import Image as PILImage
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub.utils import (
calculate_episode_data_index,
concatenate_episodes,
get_default_encoding,
save_images_concurrently,
)
from lerobot.common.datasets.utils import (
hf_transform_to_torch,
)
from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
def get_cameras(hdf5_data):
# ignore depth channel, not currently handled
# TODO(rcadene): add depth
rgb_cameras = [key for key in hdf5_data["/observations/images"].keys() if "depth" not in key] # noqa: SIM118
return rgb_cameras
def check_format(raw_dir) -> bool:
# only frames from simulation are uncompressed
compressed_images = "sim" not in raw_dir.name
hdf5_paths = list(raw_dir.glob("episode_*.hdf5"))
assert len(hdf5_paths) != 0
for hdf5_path in hdf5_paths:
with h5py.File(hdf5_path, "r") as data:
assert "/action" in data
assert "/observations/qpos" in data
assert data["/action"].ndim == 2
assert data["/observations/qpos"].ndim == 2
num_frames = data["/action"].shape[0]
assert num_frames == data["/observations/qpos"].shape[0]
for camera in get_cameras(data):
assert num_frames == data[f"/observations/images/{camera}"].shape[0]
if compressed_images:
assert data[f"/observations/images/{camera}"].ndim == 2
else:
assert data[f"/observations/images/{camera}"].ndim == 4
b, h, w, c = data[f"/observations/images/{camera}"].shape
assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
def load_from_raw(
raw_dir: Path,
videos_dir: Path,
fps: int,
video: bool,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
# only frames from simulation are uncompressed
compressed_images = "sim" not in raw_dir.name
hdf5_files = sorted(raw_dir.glob("episode_*.hdf5"))
num_episodes = len(hdf5_files)
ep_dicts = []
ep_ids = episodes if episodes else range(num_episodes)
for ep_idx in tqdm.tqdm(ep_ids):
ep_path = hdf5_files[ep_idx]
with h5py.File(ep_path, "r") as ep:
num_frames = ep["/action"].shape[0]
# last step of demonstration is considered done
done = torch.zeros(num_frames, dtype=torch.bool)
done[-1] = True
state = torch.from_numpy(ep["/observations/qpos"][:])
action = torch.from_numpy(ep["/action"][:])
if "/observations/qvel" in ep:
velocity = torch.from_numpy(ep["/observations/qvel"][:])
if "/observations/effort" in ep:
effort = torch.from_numpy(ep["/observations/effort"][:])
ep_dict = {}
for camera in get_cameras(ep):
img_key = f"observation.images.{camera}"
if compressed_images:
import cv2
# load one compressed image after the other in RAM and uncompress
imgs_array = []
for data in ep[f"/observations/images/{camera}"]:
imgs_array.append(cv2.imdecode(data, 1))
imgs_array = np.array(imgs_array)
else:
# load all images in RAM
imgs_array = ep[f"/observations/images/{camera}"][:]
if video:
# save png images in temporary directory
tmp_imgs_dir = videos_dir / "tmp_images"
save_images_concurrently(imgs_array, tmp_imgs_dir)
# encode images to a mp4 video
fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
video_path = videos_dir / fname
encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
# clean temporary images directory
shutil.rmtree(tmp_imgs_dir)
# store the reference to the video frame
ep_dict[img_key] = [
{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
]
else:
ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
ep_dict["observation.state"] = state
if "/observations/velocity" in ep:
ep_dict["observation.velocity"] = velocity
if "/observations/effort" in ep:
ep_dict["observation.effort"] = effort
ep_dict["action"] = action
ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
ep_dict["next.done"] = done
# TODO(rcadene): add reward and success by computing them in sim
assert isinstance(ep_idx, int)
ep_dicts.append(ep_dict)
gc.collect()
data_dict = concatenate_episodes(ep_dicts)
total_frames = data_dict["frame_index"].shape[0]
data_dict["index"] = torch.arange(0, total_frames, 1)
return data_dict
def to_hf_dataset(data_dict, video) -> Dataset:
features = {}
keys = [key for key in data_dict if "observation.images." in key]
for key in keys:
if video:
features[key] = VideoFrame()
else:
features[key] = Image()
features["observation.state"] = Sequence(
length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
)
if "observation.velocity" in data_dict:
features["observation.velocity"] = Sequence(
length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
)
if "observation.effort" in data_dict:
features["observation.effort"] = Sequence(
length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
)
features["action"] = Sequence(
length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
)
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["next.done"] = Value(dtype="bool", id=None)
features["index"] = Value(dtype="int64", id=None)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
# sanity check
check_format(raw_dir)
if fps is None:
fps = 50
data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
hf_dataset = to_hf_dataset(data_dict, video)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video,
}
if video:
info["encoding"] = get_default_encoding()
return hf_dataset, episode_data_index, info

107
lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py
View File

@ -1,107 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Contains utilities to process raw data format of png images files recorded with capture_camera_feed.py
"""
from pathlib import Path
import torch
from datasets import Dataset, Features, Image, Value
from PIL import Image as PILImage
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub.utils import (
calculate_episode_data_index,
concatenate_episodes,
)
from lerobot.common.datasets.utils import hf_transform_to_torch
from lerobot.common.datasets.video_utils import VideoFrame
def check_format(raw_dir: Path) -> bool:
image_paths = list(raw_dir.glob("frame_*.png"))
if len(image_paths) == 0:
raise ValueError
def load_from_raw(raw_dir: Path, fps: int, episodes: list[int] | None = None):
if episodes is not None:
# TODO(aliberts): add support for multi-episodes.
raise NotImplementedError()
ep_dict = {}
ep_idx = 0
image_paths = sorted(raw_dir.glob("frame_*.png"))
num_frames = len(image_paths)
ep_dict["observation.image"] = [PILImage.open(x) for x in image_paths]
ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
ep_dicts = [ep_dict]
data_dict = concatenate_episodes(ep_dicts)
total_frames = data_dict["frame_index"].shape[0]
data_dict["index"] = torch.arange(0, total_frames, 1)
return data_dict
def to_hf_dataset(data_dict, video) -> Dataset:
features = {}
if video:
features["observation.image"] = VideoFrame()
else:
features["observation.image"] = Image()
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["index"] = Value(dtype="int64", id=None)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
if video or episodes or encoding is not None:
# TODO(aliberts): support this
raise NotImplementedError
# sanity check
check_format(raw_dir)
if fps is None:
fps = 30
data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
hf_dataset = to_hf_dataset(data_dict, video)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video,
}
return hf_dataset, episode_data_index, info

233
lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py
View File

@ -1,233 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Contains utilities to process raw data format from dora-record
"""
import re
import warnings
from pathlib import Path
import pandas as pd
import torch
from datasets import Dataset, Features, Image, Sequence, Value
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub.utils import calculate_episode_data_index
from lerobot.common.datasets.utils import (
hf_transform_to_torch,
)
from lerobot.common.datasets.video_utils import VideoFrame
def check_format(raw_dir) -> bool:
assert raw_dir.exists()
leader_file = list(raw_dir.glob("*.parquet"))
if len(leader_file) == 0:
raise ValueError(f"Missing parquet files in '{raw_dir}'")
return True
def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
# Load data stream that will be used as reference for the timestamps synchronization
reference_files = list(raw_dir.glob("observation.images.cam_*.parquet"))
if len(reference_files) == 0:
raise ValueError(f"Missing reference files for camera, starting with in '{raw_dir}'")
# select first camera in alphanumeric order
reference_key = sorted(reference_files)[0].stem
reference_df = pd.read_parquet(raw_dir / f"{reference_key}.parquet")
reference_df = reference_df[["timestamp_utc", reference_key]]
# Merge all data stream using nearest backward strategy
df = reference_df
for path in raw_dir.glob("*.parquet"):
key = path.stem # action or observation.state or ...
if key == reference_key:
continue
if "failed_episode_index" in key:
# TODO(rcadene): add support for removing episodes that are tagged as "failed"
continue
modality_df = pd.read_parquet(path)
modality_df = modality_df[["timestamp_utc", key]]
df = pd.merge_asof(
df,
modality_df,
on="timestamp_utc",
# "nearest" is the best option over "backward", since the latter can desynchronizes camera timestamps by
# matching timestamps that are too far appart, in order to fit the backward constraints. It's not the case for "nearest".
# However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
# are too far appart.
direction="nearest",
tolerance=pd.Timedelta(f"{1/fps} seconds"),
)
# Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
df = df[df["episode_index"] != -1]
image_keys = [key for key in df if "observation.images." in key]
def get_episode_index(row):
episode_index_per_cam = {}
for key in image_keys:
path = row[key][0]["path"]
match = re.search(r"_(\d{6}).mp4", path)
if not match:
raise ValueError(path)
episode_index = int(match.group(1))
episode_index_per_cam[key] = episode_index
if len(set(episode_index_per_cam.values())) != 1:
raise ValueError(
f"All cameras are expected to belong to the same episode, but getting {episode_index_per_cam}"
)
return episode_index
df["episode_index"] = df.apply(get_episode_index, axis=1)
# dora only use arrays, so single values are encapsulated into a list
df["frame_index"] = df.groupby("episode_index").cumcount()
df = df.reset_index()
df["index"] = df.index
# set 'next.done' to True for the last frame of each episode
df["next.done"] = False
df.loc[df.groupby("episode_index").tail(1).index, "next.done"] = True
df["timestamp"] = df["timestamp_utc"].map(lambda x: x.timestamp())
# each episode starts with timestamp 0 to match the ones from the video
df["timestamp"] = df.groupby("episode_index")["timestamp"].transform(lambda x: x - x.iloc[0])
del df["timestamp_utc"]
# sanity check
has_nan = df.isna().any().any()
if has_nan:
raise ValueError("Dataset contains Nan values.")
# sanity check episode indices go from 0 to n-1
ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
expected_ep_ids = list(range(df["episode_index"].max() + 1))
if ep_ids != expected_ep_ids:
raise ValueError(f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}")
# Create symlink to raw videos directory (that needs to be absolute not relative)
videos_dir.parent.mkdir(parents=True, exist_ok=True)
videos_dir.symlink_to((raw_dir / "videos").absolute())
# sanity check the video paths are well formated
for key in df:
if "observation.images." not in key:
continue
for ep_idx in ep_ids:
video_path = videos_dir / f"{key}_episode_{ep_idx:06d}.mp4"
if not video_path.exists():
raise ValueError(f"Video file not found in {video_path}")
data_dict = {}
for key in df:
# is video frame
if "observation.images." in key:
# we need `[0] because dora only use arrays, so single values are encapsulated into a list.
# it is the case for video_frame dictionary = [{"path": ..., "timestamp": ...}]
data_dict[key] = [video_frame[0] for video_frame in df[key].values]
# sanity check the video path is well formated
video_path = videos_dir.parent / data_dict[key][0]["path"]
if not video_path.exists():
raise ValueError(f"Video file not found in {video_path}")
# is number
elif df[key].iloc[0].ndim == 0 or df[key].iloc[0].shape[0] == 1:
data_dict[key] = torch.from_numpy(df[key].values)
# is vector
elif df[key].iloc[0].shape[0] > 1:
data_dict[key] = torch.stack([torch.from_numpy(x.copy()) for x in df[key].values])
else:
raise ValueError(key)
return data_dict
def to_hf_dataset(data_dict, video) -> Dataset:
features = {}
keys = [key for key in data_dict if "observation.images." in key]
for key in keys:
if video:
features[key] = VideoFrame()
else:
features[key] = Image()
features["observation.state"] = Sequence(
length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
)
if "observation.velocity" in data_dict:
features["observation.velocity"] = Sequence(
length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
)
if "observation.effort" in data_dict:
features["observation.effort"] = Sequence(
length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
)
features["action"] = Sequence(
length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
)
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["next.done"] = Value(dtype="bool", id=None)
features["index"] = Value(dtype="int64", id=None)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
# sanity check
check_format(raw_dir)
if fps is None:
fps = 30
else:
raise NotImplementedError()
if not video:
raise NotImplementedError()
if encoding is not None:
warnings.warn(
"Video encoding is currently done outside of LeRobot for the dora_parquet format.",
stacklevel=1,
)
data_df = load_from_raw(raw_dir, videos_dir, fps, episodes)
hf_dataset = to_hf_dataset(data_df, video)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video,
}
if video:
info["encoding"] = "unknown"
return hf_dataset, episode_data_index, info

312
lerobot/common/datasets/push_dataset_to_hub/openx_rlds_format.py
View File

@ -1,312 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
For all datasets in the RLDS format.
For https://github.com/google-deepmind/open_x_embodiment (OPENX) datasets.
NOTE: You need to install tensorflow and tensorflow_datsets before running this script.
Example:
python lerobot/scripts/push_dataset_to_hub.py \
--raw-dir /path/to/data/bridge_dataset/1.0.0/ \
--repo-id your_hub/sampled_bridge_data_v2 \
--raw-format rlds \
--episodes 3 4 5 8 9
Exact dataset fps defined in openx/config.py, obtained from:
https://docs.google.com/spreadsheets/d/1rPBD77tk60AEIGZrGSODwyyzs5FgCU9Uz3h-3_t2A9g/edit?gid=0#gid=0&range=R:R
"""
import shutil
from pathlib import Path
import numpy as np
import tensorflow as tf
import tensorflow_datasets as tfds
import torch
import tqdm
from datasets import Dataset, Features, Image, Sequence, Value
from PIL import Image as PILImage
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub.utils import (
calculate_episode_data_index,
concatenate_episodes,
get_default_encoding,
save_images_concurrently,
)
from lerobot.common.datasets.utils import (
hf_transform_to_torch,
)
from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
np.set_printoptions(precision=2)
def tf_to_torch(data):
return torch.from_numpy(data.numpy())
def tf_img_convert(img):
if img.dtype == tf.string:
img = tf.io.decode_image(img, expand_animations=False, dtype=tf.uint8)
elif img.dtype != tf.uint8:
raise ValueError(f"Unsupported image dtype: found with dtype {img.dtype}")
return img.numpy()
def _broadcast_metadata_rlds(i: tf.Tensor, traj: dict) -> dict:
"""
In the RLDS format, each trajectory has some top-level metadata that is explicitly separated out, and a "steps"
entry. This function moves the "steps" entry to the top level, broadcasting any metadata to the length of the
trajectory. This function also adds the extra metadata fields `_len`, `_traj_index`, and `_frame_index`.
NOTE: adapted from DLimp library https://github.com/kvablack/dlimp/
"""
steps = traj.pop("steps")
traj_len = tf.shape(tf.nest.flatten(steps)[0])[0]
# broadcast metadata to the length of the trajectory
metadata = tf.nest.map_structure(lambda x: tf.repeat(x, traj_len), traj)
# put steps back in
assert "traj_metadata" not in steps
traj = {**steps, "traj_metadata": metadata}
assert "_len" not in traj
assert "_traj_index" not in traj
assert "_frame_index" not in traj
traj["_len"] = tf.repeat(traj_len, traj_len)
traj["_traj_index"] = tf.repeat(i, traj_len)
traj["_frame_index"] = tf.range(traj_len)
return traj
def load_from_raw(
raw_dir: Path,
videos_dir: Path,
fps: int,
video: bool,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
"""
Args:
raw_dir (Path): _description_
videos_dir (Path): _description_
fps (int): _description_
video (bool): _description_
episodes (list[int] | None, optional): _description_. Defaults to None.
"""
ds_builder = tfds.builder_from_directory(str(raw_dir))
dataset = ds_builder.as_dataset(
split="all",
decoders={"steps": tfds.decode.SkipDecoding()},
)
dataset_info = ds_builder.info
print("dataset_info: ", dataset_info)
ds_length = len(dataset)
dataset = dataset.take(ds_length)
# "flatten" the dataset as such we can apply trajectory level map() easily
# each [obs][key] has a shape of (frame_size, ...)
dataset = dataset.enumerate().map(_broadcast_metadata_rlds)
# we will apply the standardization transform if the dataset_name is provided
# if the dataset name is not provided and the goal is to convert any rlds formatted dataset
# search for 'image' keys in the observations
image_keys = []
state_keys = []
observation_info = dataset_info.features["steps"]["observation"]
for key in observation_info:
# check whether the key is for an image or a vector observation
if len(observation_info[key].shape) == 3:
# only adding uint8 images discards depth images
if observation_info[key].dtype == tf.uint8:
image_keys.append(key)
else:
state_keys.append(key)
lang_key = "language_instruction" if "language_instruction" in dataset.element_spec else None
print(" - image_keys: ", image_keys)
print(" - lang_key: ", lang_key)
it = iter(dataset)
ep_dicts = []
# Init temp path to save ep_dicts in case of crash
tmp_ep_dicts_dir = videos_dir.parent.joinpath("ep_dicts")
tmp_ep_dicts_dir.mkdir(parents=True, exist_ok=True)
# check if ep_dicts have already been saved in /tmp
starting_ep_idx = 0
saved_ep_dicts = [ep.__str__() for ep in tmp_ep_dicts_dir.iterdir()]
if len(saved_ep_dicts) > 0:
saved_ep_dicts.sort()
# get last ep_idx number
starting_ep_idx = int(saved_ep_dicts[-1][-13:-3]) + 1
for i in range(starting_ep_idx):
episode = next(it)
ep_dicts.append(torch.load(saved_ep_dicts[i]))
# if we user specified episodes, skip the ones not in the list
if episodes is not None:
if ds_length == 0:
raise ValueError("No episodes found.")
# convert episodes index to sorted list
episodes = sorted(episodes)
for ep_idx in tqdm.tqdm(range(starting_ep_idx, ds_length)):
episode = next(it)
# if user specified episodes, skip the ones not in the list
if episodes is not None:
if len(episodes) == 0:
break
if ep_idx == episodes[0]:
# process this episode
print(" selecting episode idx: ", ep_idx)
episodes.pop(0)
else:
continue # skip
num_frames = episode["action"].shape[0]
ep_dict = {}
for key in state_keys:
ep_dict[f"observation.{key}"] = tf_to_torch(episode["observation"][key])
ep_dict["action"] = tf_to_torch(episode["action"])
ep_dict["next.reward"] = tf_to_torch(episode["reward"]).float()
ep_dict["next.done"] = tf_to_torch(episode["is_last"])
ep_dict["is_terminal"] = tf_to_torch(episode["is_terminal"])
ep_dict["is_first"] = tf_to_torch(episode["is_first"])
ep_dict["discount"] = tf_to_torch(episode["discount"])
# If lang_key is present, convert the entire tensor at once
if lang_key is not None:
ep_dict["language_instruction"] = [x.numpy().decode("utf-8") for x in episode[lang_key]]
ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
image_array_dict = {key: [] for key in image_keys}
for im_key in image_keys:
imgs = episode["observation"][im_key]
image_array_dict[im_key] = [tf_img_convert(img) for img in imgs]
# loop through all cameras
for im_key in image_keys:
img_key = f"observation.images.{im_key}"
imgs_array = image_array_dict[im_key]
imgs_array = np.array(imgs_array)
if video:
# save png images in temporary directory
tmp_imgs_dir = videos_dir / "tmp_images"
save_images_concurrently(imgs_array, tmp_imgs_dir)
# encode images to a mp4 video
fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
video_path = videos_dir / fname
encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
# clean temporary images directory
shutil.rmtree(tmp_imgs_dir)
# store the reference to the video frame
ep_dict[img_key] = [
{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
]
else:
ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
path_ep_dict = tmp_ep_dicts_dir.joinpath(
"ep_dict_" + "0" * (10 - len(str(ep_idx))) + str(ep_idx) + ".pt"
)
torch.save(ep_dict, path_ep_dict)
ep_dicts.append(ep_dict)
data_dict = concatenate_episodes(ep_dicts)
total_frames = data_dict["frame_index"].shape[0]
data_dict["index"] = torch.arange(0, total_frames, 1)
return data_dict
def to_hf_dataset(data_dict, video) -> Dataset:
features = {}
for key in data_dict:
# check if vector state obs
if key.startswith("observation.") and "observation.images." not in key:
features[key] = Sequence(length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None))
# check if image obs
elif "observation.images." in key:
if video:
features[key] = VideoFrame()
else:
features[key] = Image()
if "language_instruction" in data_dict:
features["language_instruction"] = Value(dtype="string", id=None)
features["action"] = Sequence(
length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
)
features["is_terminal"] = Value(dtype="bool", id=None)
features["is_first"] = Value(dtype="bool", id=None)
features["discount"] = Value(dtype="float32", id=None)
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["next.reward"] = Value(dtype="float32", id=None)
features["next.done"] = Value(dtype="bool", id=None)
features["index"] = Value(dtype="int64", id=None)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
hf_dataset = to_hf_dataset(data_dict, video)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video,
}
if video:
info["encoding"] = get_default_encoding()
return hf_dataset, episode_data_index, info

275
lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py
View File

@ -1,275 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Process zarr files formatted like in: https://github.com/real-stanford/diffusion_policy"""
import shutil
from pathlib import Path
import numpy as np
import torch
import tqdm
import zarr
from datasets import Dataset, Features, Image, Sequence, Value
from PIL import Image as PILImage
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub.utils import (
calculate_episode_data_index,
concatenate_episodes,
get_default_encoding,
save_images_concurrently,
)
from lerobot.common.datasets.utils import (
hf_transform_to_torch,
)
from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
def check_format(raw_dir):
zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
zarr_data = zarr.open(zarr_path, mode="r")
required_datasets = {
"data/action",
"data/img",
"data/keypoint",
"data/n_contacts",
"data/state",
"meta/episode_ends",
}
for dataset in required_datasets:
assert dataset in zarr_data
nb_frames = zarr_data["data/img"].shape[0]
required_datasets.remove("meta/episode_ends")
assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
def load_from_raw(
raw_dir: Path,
videos_dir: Path,
fps: int,
video: bool,
episodes: list[int] | None = None,
keypoints_instead_of_image: bool = False,
encoding: dict | None = None,
):
try:
import pymunk
from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
ReplayBuffer as DiffusionPolicyReplayBuffer,
)
except ModuleNotFoundError as e:
print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
raise e
# as define in gmy-pusht env: https://github.com/huggingface/gym-pusht/blob/e0684ff988d223808c0a9dcfaba9dc4991791370/gym_pusht/envs/pusht.py#L174
success_threshold = 0.95 # 95% coverage,
zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
episode_ids = torch.from_numpy(zarr_data.get_episode_idxs())
assert len(
{zarr_data[key].shape[0] for key in zarr_data.keys()} # noqa: SIM118
), "Some data type dont have the same number of total frames."
# TODO(rcadene): verify that goal pose is expected to be fixed
goal_pos_angle = np.array([256, 256, np.pi / 4]) # x, y, theta (in radians)
goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
imgs = torch.from_numpy(zarr_data["img"]) # b h w c
states = torch.from_numpy(zarr_data["state"])
actions = torch.from_numpy(zarr_data["action"])
# load data indices from which each episode starts and ends
from_ids, to_ids = [], []
from_idx = 0
for to_idx in zarr_data.meta["episode_ends"]:
from_ids.append(from_idx)
to_ids.append(to_idx)
from_idx = to_idx
num_episodes = len(from_ids)
ep_dicts = []
ep_ids = episodes if episodes else range(num_episodes)
for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
from_idx = from_ids[selected_ep_idx]
to_idx = to_ids[selected_ep_idx]
num_frames = to_idx - from_idx
# sanity check
assert (episode_ids[from_idx:to_idx] == ep_idx).all()
# get image
if not keypoints_instead_of_image:
image = imgs[from_idx:to_idx]
assert image.min() >= 0.0
assert image.max() <= 255.0
image = image.type(torch.uint8)
# get state
state = states[from_idx:to_idx]
agent_pos = state[:, :2]
block_pos = state[:, 2:4]
block_angle = state[:, 4]
# get reward, success, done, and (maybe) keypoints
reward = torch.zeros(num_frames)
success = torch.zeros(num_frames, dtype=torch.bool)
if keypoints_instead_of_image:
keypoints = torch.zeros(num_frames, 16) # 8 keypoints each with 2 coords
done = torch.zeros(num_frames, dtype=torch.bool)
for i in range(num_frames):
space = pymunk.Space()
space.gravity = 0, 0
space.damping = 0
# Add walls.
walls = [
PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
]
space.add(*walls)
block_body, block_shapes = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
block_geom = pymunk_to_shapely(block_body, block_body.shapes)
intersection_area = goal_geom.intersection(block_geom).area
goal_area = goal_geom.area
coverage = intersection_area / goal_area
reward[i] = np.clip(coverage / success_threshold, 0, 1)
success[i] = coverage > success_threshold
if keypoints_instead_of_image:
keypoints[i] = torch.from_numpy(PushTEnv.get_keypoints(block_shapes).flatten())
# last step of demonstration is considered done
done[-1] = True
ep_dict = {}
if not keypoints_instead_of_image:
imgs_array = [x.numpy() for x in image]
img_key = "observation.image"
if video:
# save png images in temporary directory
tmp_imgs_dir = videos_dir / "tmp_images"
save_images_concurrently(imgs_array, tmp_imgs_dir)
# encode images to a mp4 video
fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
video_path = videos_dir / fname
encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
# clean temporary images directory
shutil.rmtree(tmp_imgs_dir)
# store the reference to the video frame
ep_dict[img_key] = [
{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
]
else:
ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
ep_dict["observation.state"] = agent_pos
if keypoints_instead_of_image:
ep_dict["observation.environment_state"] = keypoints
ep_dict["action"] = actions[from_idx:to_idx]
ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
# ep_dict["next.observation.image"] = image[1:],
# ep_dict["next.observation.state"] = agent_pos[1:],
# TODO(rcadene)] = verify that reward and done are aligned with image and agent_pos
ep_dict["next.reward"] = torch.cat([reward[1:], reward[[-1]]])
ep_dict["next.done"] = torch.cat([done[1:], done[[-1]]])
ep_dict["next.success"] = torch.cat([success[1:], success[[-1]]])
ep_dicts.append(ep_dict)
data_dict = concatenate_episodes(ep_dicts)
total_frames = data_dict["frame_index"].shape[0]
data_dict["index"] = torch.arange(0, total_frames, 1)
return data_dict
def to_hf_dataset(data_dict, video, keypoints_instead_of_image: bool = False):
features = {}
if not keypoints_instead_of_image:
if video:
features["observation.image"] = VideoFrame()
else:
features["observation.image"] = Image()
features["observation.state"] = Sequence(
length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
)
if keypoints_instead_of_image:
features["observation.environment_state"] = Sequence(
length=data_dict["observation.environment_state"].shape[1],
feature=Value(dtype="float32", id=None),
)
features["action"] = Sequence(
length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
)
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["next.reward"] = Value(dtype="float32", id=None)
features["next.done"] = Value(dtype="bool", id=None)
features["next.success"] = Value(dtype="bool", id=None)
features["index"] = Value(dtype="int64", id=None)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
# Manually change this to True to use keypoints of the T instead of an image observation (but don't merge
# with True). Also make sure to use video = 0 in the `push_dataset_to_hub.py` script.
keypoints_instead_of_image = False
# sanity check
check_format(raw_dir)
if fps is None:
fps = 10
data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, keypoints_instead_of_image, encoding)
hf_dataset = to_hf_dataset(data_dict, video, keypoints_instead_of_image)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video if not keypoints_instead_of_image else 0,
}
if video:
info["encoding"] = get_default_encoding()
return hf_dataset, episode_data_index, info

234
lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py
View File

@ -1,234 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Process UMI (Universal Manipulation Interface) data stored in Zarr format like in: https://github.com/real-stanford/universal_manipulation_interface"""
import logging
import shutil
from pathlib import Path
import torch
import tqdm
import zarr
from datasets import Dataset, Features, Image, Sequence, Value
from PIL import Image as PILImage
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
from lerobot.common.datasets.push_dataset_to_hub.utils import (
calculate_episode_data_index,
concatenate_episodes,
get_default_encoding,
save_images_concurrently,
)
from lerobot.common.datasets.utils import (
hf_transform_to_torch,
)
from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
def check_format(raw_dir) -> bool:
zarr_path = raw_dir / "cup_in_the_wild.zarr"
zarr_data = zarr.open(zarr_path, mode="r")
required_datasets = {
"data/robot0_demo_end_pose",
"data/robot0_demo_start_pose",
"data/robot0_eef_pos",
"data/robot0_eef_rot_axis_angle",
"data/robot0_gripper_width",
"meta/episode_ends",
"data/camera0_rgb",
}
for dataset in required_datasets:
if dataset not in zarr_data:
return False
# mandatory to access zarr_data
register_codecs()
nb_frames = zarr_data["data/camera0_rgb"].shape[0]
required_datasets.remove("meta/episode_ends")
assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
def load_from_raw(
raw_dir: Path,
videos_dir: Path,
fps: int,
video: bool,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
zarr_path = raw_dir / "cup_in_the_wild.zarr"
zarr_data = zarr.open(zarr_path, mode="r")
# We process the image data separately because it is too large to fit in memory
end_pose = torch.from_numpy(zarr_data["data/robot0_demo_end_pose"][:])
start_pos = torch.from_numpy(zarr_data["data/robot0_demo_start_pose"][:])
eff_pos = torch.from_numpy(zarr_data["data/robot0_eef_pos"][:])
eff_rot_axis_angle = torch.from_numpy(zarr_data["data/robot0_eef_rot_axis_angle"][:])
gripper_width = torch.from_numpy(zarr_data["data/robot0_gripper_width"][:])
states_pos = torch.cat([eff_pos, eff_rot_axis_angle], dim=1)
states = torch.cat([states_pos, gripper_width], dim=1)
episode_ends = zarr_data["meta/episode_ends"][:]
num_episodes = episode_ends.shape[0]
# We convert it in torch tensor later because the jit function does not support torch tensors
episode_ends = torch.from_numpy(episode_ends)
# load data indices from which each episode starts and ends
from_ids, to_ids = [], []
from_idx = 0
for to_idx in episode_ends:
from_ids.append(from_idx)
to_ids.append(to_idx)
from_idx = to_idx
ep_dicts_dir = videos_dir / "ep_dicts"
ep_dicts_dir.mkdir(exist_ok=True, parents=True)
ep_dicts = []
ep_ids = episodes if episodes else range(num_episodes)
for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
ep_dict_path = ep_dicts_dir / f"{ep_idx}"
if not ep_dict_path.is_file():
from_idx = from_ids[selected_ep_idx]
to_idx = to_ids[selected_ep_idx]
num_frames = to_idx - from_idx
# TODO(rcadene): save temporary images of the episode?
state = states[from_idx:to_idx]
ep_dict = {}
# load 57MB of images in RAM (400x224x224x3 uint8)
imgs_array = zarr_data["data/camera0_rgb"][from_idx:to_idx]
img_key = "observation.image"
if video:
fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
video_path = videos_dir / fname
if not video_path.is_file():
# save png images in temporary directory
tmp_imgs_dir = videos_dir / "tmp_images"
save_images_concurrently(imgs_array, tmp_imgs_dir)
# encode images to a mp4 video
encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
# clean temporary images directory
shutil.rmtree(tmp_imgs_dir)
# store the reference to the video frame
ep_dict[img_key] = [
{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
]
else:
ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
ep_dict["observation.state"] = state
ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
ep_dict["episode_data_index_from"] = torch.tensor([from_idx] * num_frames)
ep_dict["episode_data_index_to"] = torch.tensor([from_idx + num_frames] * num_frames)
ep_dict["end_pose"] = end_pose[from_idx:to_idx]
ep_dict["start_pos"] = start_pos[from_idx:to_idx]
ep_dict["gripper_width"] = gripper_width[from_idx:to_idx]
torch.save(ep_dict, ep_dict_path)
else:
ep_dict = torch.load(ep_dict_path)
ep_dicts.append(ep_dict)
data_dict = concatenate_episodes(ep_dicts)
total_frames = data_dict["frame_index"].shape[0]
data_dict["index"] = torch.arange(0, total_frames, 1)
return data_dict
def to_hf_dataset(data_dict, video):
features = {}
if video:
features["observation.image"] = VideoFrame()
else:
features["observation.image"] = Image()
features["observation.state"] = Sequence(
length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
)
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["index"] = Value(dtype="int64", id=None)
features["episode_data_index_from"] = Value(dtype="int64", id=None)
features["episode_data_index_to"] = Value(dtype="int64", id=None)
# `start_pos` and `end_pos` respectively represent the positions of the end-effector
# at the beginning and the end of the episode.
# `gripper_width` indicates the distance between the grippers, and this value is included
# in the state vector, which comprises the concatenation of the end-effector position
# and gripper width.
features["end_pose"] = Sequence(
length=data_dict["end_pose"].shape[1], feature=Value(dtype="float32", id=None)
)
features["start_pos"] = Sequence(
length=data_dict["start_pos"].shape[1], feature=Value(dtype="float32", id=None)
)
features["gripper_width"] = Sequence(
length=data_dict["gripper_width"].shape[1], feature=Value(dtype="float32", id=None)
)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
# sanity check
check_format(raw_dir)
if fps is None:
# For umi cup in the wild: https://arxiv.org/pdf/2402.10329#table.caption.16
fps = 10
if not video:
logging.warning(
"Generating UMI dataset without `video=True` creates ~150GB on disk and requires ~80GB in RAM."
)
data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
hf_dataset = to_hf_dataset(data_dict, video)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video,
}
if video:
info["encoding"] = get_default_encoding()
return hf_dataset, episode_data_index, info

200
lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py
View File

@ -1,200 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Process pickle files formatted like in: https://github.com/fyhMer/fowm"""
import pickle
import shutil
from pathlib import Path
import einops
import torch
import tqdm
from datasets import Dataset, Features, Image, Sequence, Value
from PIL import Image as PILImage
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
from lerobot.common.datasets.push_dataset_to_hub.utils import (
calculate_episode_data_index,
concatenate_episodes,
get_default_encoding,
save_images_concurrently,
)
from lerobot.common.datasets.utils import (
hf_transform_to_torch,
)
from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
def check_format(raw_dir):
keys = {"actions", "rewards", "dones"}
nested_keys = {"observations": {"rgb", "state"}, "next_observations": {"rgb", "state"}}
xarm_files = list(raw_dir.glob("*.pkl"))
assert len(xarm_files) > 0
with open(xarm_files[0], "rb") as f:
dataset_dict = pickle.load(f)
assert isinstance(dataset_dict, dict)
assert all(k in dataset_dict for k in keys)
# Check for consistent lengths in nested keys
expected_len = len(dataset_dict["actions"])
assert all(len(dataset_dict[key]) == expected_len for key in keys if key in dataset_dict)
for key, subkeys in nested_keys.items():
nested_dict = dataset_dict.get(key, {})
assert all(len(nested_dict[subkey]) == expected_len for subkey in subkeys if subkey in nested_dict)
def load_from_raw(
raw_dir: Path,
videos_dir: Path,
fps: int,
video: bool,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
pkl_path = raw_dir / "buffer.pkl"
with open(pkl_path, "rb") as f:
pkl_data = pickle.load(f)
# load data indices from which each episode starts and ends
from_ids, to_ids = [], []
from_idx, to_idx = 0, 0
for done in pkl_data["dones"]:
to_idx += 1
if not done:
continue
from_ids.append(from_idx)
to_ids.append(to_idx)
from_idx = to_idx
num_episodes = len(from_ids)
ep_dicts = []
ep_ids = episodes if episodes else range(num_episodes)
for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
from_idx = from_ids[selected_ep_idx]
to_idx = to_ids[selected_ep_idx]
num_frames = to_idx - from_idx
image = torch.tensor(pkl_data["observations"]["rgb"][from_idx:to_idx])
image = einops.rearrange(image, "b c h w -> b h w c")
state = torch.tensor(pkl_data["observations"]["state"][from_idx:to_idx])
action = torch.tensor(pkl_data["actions"][from_idx:to_idx])
# TODO(rcadene): we have a missing last frame which is the observation when the env is done
# it is critical to have this frame for tdmpc to predict a "done observation/state"
# next_image = torch.tensor(pkl_data["next_observations"]["rgb"][from_idx:to_idx])
# next_state = torch.tensor(pkl_data["next_observations"]["state"][from_idx:to_idx])
next_reward = torch.tensor(pkl_data["rewards"][from_idx:to_idx])
next_done = torch.tensor(pkl_data["dones"][from_idx:to_idx])
ep_dict = {}
imgs_array = [x.numpy() for x in image]
img_key = "observation.image"
if video:
# save png images in temporary directory
tmp_imgs_dir = videos_dir / "tmp_images"
save_images_concurrently(imgs_array, tmp_imgs_dir)
# encode images to a mp4 video
fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
video_path = videos_dir / fname
encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
# clean temporary images directory
shutil.rmtree(tmp_imgs_dir)
# store the reference to the video frame
ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
else:
ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
ep_dict["observation.state"] = state
ep_dict["action"] = action
ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
# ep_dict["next.observation.image"] = next_image
# ep_dict["next.observation.state"] = next_state
ep_dict["next.reward"] = next_reward
ep_dict["next.done"] = next_done
ep_dicts.append(ep_dict)
data_dict = concatenate_episodes(ep_dicts)
total_frames = data_dict["frame_index"].shape[0]
data_dict["index"] = torch.arange(0, total_frames, 1)
return data_dict
def to_hf_dataset(data_dict, video):
features = {}
if video:
features["observation.image"] = VideoFrame()
else:
features["observation.image"] = Image()
features["observation.state"] = Sequence(
length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
)
features["action"] = Sequence(
length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
)
features["episode_index"] = Value(dtype="int64", id=None)
features["frame_index"] = Value(dtype="int64", id=None)
features["timestamp"] = Value(dtype="float32", id=None)
features["next.reward"] = Value(dtype="float32", id=None)
features["next.done"] = Value(dtype="bool", id=None)
features["index"] = Value(dtype="int64", id=None)
# TODO(rcadene): add success
# features["next.success"] = Value(dtype='bool', id=None)
hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
hf_dataset.set_transform(hf_transform_to_torch)
return hf_dataset
def from_raw_to_lerobot_format(
raw_dir: Path,
videos_dir: Path,
fps: int | None = None,
video: bool = True,
episodes: list[int] | None = None,
encoding: dict | None = None,
):
# sanity check
check_format(raw_dir)
if fps is None:
fps = 15
data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
hf_dataset = to_hf_dataset(data_dict, video)
episode_data_index = calculate_episode_data_index(hf_dataset)
info = {
"codebase_version": CODEBASE_VERSION,
"fps": fps,
"video": video,
}
if video:
info["encoding"] = get_default_encoding()
return hf_dataset, episode_data_index, info

174
lerobot/common/datasets/transforms.py
View File

@ -14,7 +14,8 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import collections
from typing import Any, Callable, Dict, Sequence
from dataclasses import dataclass, field
from typing import Any, Callable, Sequence
import torch
from torchvision.transforms import v2
@ -65,6 +66,8 @@ class RandomSubsetApply(Transform):
self.n_subset = n_subset
self.random_order = random_order
self.selected_transforms = None
def forward(self, *inputs: Any) -> Any:
needs_unpacking = len(inputs) > 1
@ -72,9 +75,9 @@ class RandomSubsetApply(Transform):
if not self.random_order:
selected_indices = selected_indices.sort().values
selected_transforms = [self.transforms[i] for i in selected_indices]
self.selected_transforms = [self.transforms[i] for i in selected_indices]
for transform in selected_transforms:
for transform in self.selected_transforms:
outputs = transform(*inputs)
inputs = outputs if needs_unpacking else (outputs,)
@ -129,69 +132,118 @@ class SharpnessJitter(Transform):
return float(sharpness[0]), float(sharpness[1])
def _generate_value(self, left: float, right: float) -> float:
return torch.empty(1).uniform_(left, right).item()
def make_params(self, flat_inputs: list[Any]) -> dict[str, Any]:
sharpness_factor = torch.empty(1).uniform_(self.sharpness[0], self.sharpness[1]).item()
return {"sharpness_factor": sharpness_factor}
def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
sharpness_factor = self._generate_value(self.sharpness[0], self.sharpness[1])
def transform(self, inpt: Any, params: dict[str, Any]) -> Any:
sharpness_factor = params["sharpness_factor"]
return self._call_kernel(F.adjust_sharpness, inpt, sharpness_factor=sharpness_factor)
def get_image_transforms(
brightness_weight: float = 1.0,
brightness_min_max: tuple[float, float] | None = None,
contrast_weight: float = 1.0,
contrast_min_max: tuple[float, float] | None = None,
saturation_weight: float = 1.0,
saturation_min_max: tuple[float, float] | None = None,
hue_weight: float = 1.0,
hue_min_max: tuple[float, float] | None = None,
sharpness_weight: float = 1.0,
sharpness_min_max: tuple[float, float] | None = None,
max_num_transforms: int | None = None,
random_order: bool = False,
):
def check_value(name, weight, min_max):
if min_max is not None:
if len(min_max) != 2:
raise ValueError(
f"`{name}_min_max` is expected to be a tuple of 2 dimensions, but {min_max} provided."
)
if weight < 0.0:
raise ValueError(
f"`{name}_weight` is expected to be 0 or positive, but is negative ({weight})."
@dataclass
class ImageTransformConfig:
"""
For each transform, the following parameters are available:
weight: This represents the multinomial probability (with no replacement)
used for sampling the transform. If the sum of the weights is not 1,
they will be normalized.
type: The name of the class used. This is either a class available under torchvision.transforms.v2 or a
custom transform defined here.
kwargs: Lower & upper bound respectively used for sampling the transform's parameter
(following uniform distribution) when it's applied.
"""
weight: float = 1.0
type: str = "Identity"
kwargs: dict[str, Any] = field(default_factory=dict)
@dataclass
class ImageTransformsConfig:
"""
These transforms are all using standard torchvision.transforms.v2
You can find out how these transformations affect images here:
https://pytorch.org/vision/0.18/auto_examples/transforms/plot_transforms_illustrations.html
We use a custom RandomSubsetApply container to sample them.
"""
# Set this flag to `true` to enable transforms during training
enable: bool = False
# This is the maximum number of transforms (sampled from these below) that will be applied to each frame.
# It's an integer in the interval [1, number_of_available_transforms].
max_num_transforms: int = 3
# By default, transforms are applied in Torchvision's suggested order (shown below).
# Set this to True to apply them in a random order.
random_order: bool = False
tfs: dict[str, ImageTransformConfig] = field(
default_factory=lambda: {
"brightness": ImageTransformConfig(
weight=1.0,
type="ColorJitter",
kwargs={"brightness": (0.8, 1.2)},
),
"contrast": ImageTransformConfig(
weight=1.0,
type="ColorJitter",
kwargs={"contrast": (0.8, 1.2)},
),
"saturation": ImageTransformConfig(
weight=1.0,
type="ColorJitter",
kwargs={"saturation": (0.5, 1.5)},
),
"hue": ImageTransformConfig(
weight=1.0,
type="ColorJitter",
kwargs={"hue": (-0.05, 0.05)},
),
"sharpness": ImageTransformConfig(
weight=1.0,
type="SharpnessJitter",
kwargs={"sharpness": (0.5, 1.5)},
),
}
)
check_value("brightness", brightness_weight, brightness_min_max)
check_value("contrast", contrast_weight, contrast_min_max)
check_value("saturation", saturation_weight, saturation_min_max)
check_value("hue", hue_weight, hue_min_max)
check_value("sharpness", sharpness_weight, sharpness_min_max)
weights = []
transforms = []
if brightness_min_max is not None and brightness_weight > 0.0:
weights.append(brightness_weight)
transforms.append(v2.ColorJitter(brightness=brightness_min_max))
if contrast_min_max is not None and contrast_weight > 0.0:
weights.append(contrast_weight)
transforms.append(v2.ColorJitter(contrast=contrast_min_max))
if saturation_min_max is not None and saturation_weight > 0.0:
weights.append(saturation_weight)
transforms.append(v2.ColorJitter(saturation=saturation_min_max))
if hue_min_max is not None and hue_weight > 0.0:
weights.append(hue_weight)
transforms.append(v2.ColorJitter(hue=hue_min_max))
if sharpness_min_max is not None and sharpness_weight > 0.0:
weights.append(sharpness_weight)
transforms.append(SharpnessJitter(sharpness=sharpness_min_max))
n_subset = len(transforms)
if max_num_transforms is not None:
n_subset = min(n_subset, max_num_transforms)
if n_subset == 0:
return v2.Identity()
def make_transform_from_config(cfg: ImageTransformConfig):
if cfg.type == "Identity":
return v2.Identity(**cfg.kwargs)
elif cfg.type == "ColorJitter":
return v2.ColorJitter(**cfg.kwargs)
elif cfg.type == "SharpnessJitter":
return SharpnessJitter(**cfg.kwargs)
else:
# TODO(rcadene, aliberts): add v2.ToDtype float16?
return RandomSubsetApply(transforms, p=weights, n_subset=n_subset, random_order=random_order)
raise ValueError(f"Transform '{cfg.type}' is not valid.")
class ImageTransforms(Transform):
"""A class to compose image transforms based on configuration."""
def __init__(self, cfg: ImageTransformsConfig) -> None:
super().__init__()
self._cfg = cfg
self.weights = []
self.transforms = {}
for tf_name, tf_cfg in cfg.tfs.items():
if tf_cfg.weight <= 0.0:
continue
self.transforms[tf_name] = make_transform_from_config(tf_cfg)
self.weights.append(tf_cfg.weight)
n_subset = min(len(self.transforms), cfg.max_num_transforms)
if n_subset == 0 or not cfg.enable:
self.tf = v2.Identity()
else:
self.tf = RandomSubsetApply(
transforms=list(self.transforms.values()),
p=self.weights,
n_subset=n_subset,
random_order=cfg.random_order,
)
def forward(self, *inputs: Any) -> Any:
return self.tf(*inputs)

501
lerobot/common/datasets/utils.py
View File

@ -13,10 +13,10 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import contextlib
import importlib.resources
import json
import logging
import textwrap
from collections.abc import Iterator
from itertools import accumulate
from pathlib import Path
@ -27,20 +27,29 @@ from typing import Any
import datasets
import jsonlines
import numpy as np
import pyarrow.compute as pc
import packaging.version
import torch
from datasets.table import embed_table_storage
from huggingface_hub import DatasetCard, DatasetCardData, HfApi
from huggingface_hub.errors import RevisionNotFoundError
from PIL import Image as PILImage
from torchvision import transforms
from lerobot.common.datasets.backward_compatibility import (
V21_MESSAGE,
BackwardCompatibilityError,
ForwardCompatibilityError,
)
from lerobot.common.robot_devices.robots.utils import Robot
from lerobot.common.utils.utils import is_valid_numpy_dtype_string
from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
DEFAULT_CHUNK_SIZE = 1000 # Max number of episodes per chunk
INFO_PATH = "meta/info.json"
EPISODES_PATH = "meta/episodes.jsonl"
STATS_PATH = "meta/stats.json"
EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
TASKS_PATH = "meta/tasks.jsonl"
DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
@ -98,18 +107,39 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
return outdict
def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
split_keys = flattened_key.split(sep)
getter = obj[split_keys[0]]
if len(split_keys) == 1:
return getter
for key in split_keys[1:]:
getter = getter[key]
return getter
def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
serialized_dict = {key: value.tolist() for key, value in flatten_dict(stats).items()}
serialized_dict = {}
for key, value in flatten_dict(stats).items():
if isinstance(value, (torch.Tensor, np.ndarray)):
serialized_dict[key] = value.tolist()
elif isinstance(value, np.generic):
serialized_dict[key] = value.item()
elif isinstance(value, (int, float)):
serialized_dict[key] = value
else:
raise NotImplementedError(f"The value '{value}' of type '{type(value)}' is not supported.")
return unflatten_dict(serialized_dict)
def write_parquet(dataset: datasets.Dataset, fpath: Path) -> None:
def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
# Embed image bytes into the table before saving to parquet
format = dataset.format
dataset = dataset.with_format("arrow")
dataset = dataset.map(embed_table_storage, batched=False)
dataset = dataset.with_format(**format)
dataset.to_parquet(fpath)
return dataset
def load_json(fpath: Path) -> Any:
@ -140,6 +170,10 @@ def append_jsonlines(data: dict, fpath: Path) -> None:
writer.write(data)
def write_info(info: dict, local_dir: Path):
write_json(info, local_dir / INFO_PATH)
def load_info(local_dir: Path) -> dict:
info = load_json(local_dir / INFO_PATH)
for ft in info["features"].values():
@ -147,29 +181,76 @@ def load_info(local_dir: Path) -> dict:
return info
def load_stats(local_dir: Path) -> dict:
if not (local_dir / STATS_PATH).exists():
return None
stats = load_json(local_dir / STATS_PATH)
stats = {key: torch.tensor(value) for key, value in flatten_dict(stats).items()}
def write_stats(stats: dict, local_dir: Path):
serialized_stats = serialize_dict(stats)
write_json(serialized_stats, local_dir / STATS_PATH)
def cast_stats_to_numpy(stats) -> dict[str, dict[str, np.ndarray]]:
stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
return unflatten_dict(stats)
def load_tasks(local_dir: Path) -> dict:
def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
if not (local_dir / STATS_PATH).exists():
return None
stats = load_json(local_dir / STATS_PATH)
return cast_stats_to_numpy(stats)
def write_task(task_index: int, task: dict, local_dir: Path):
task_dict = {
"task_index": task_index,
"task": task,
}
append_jsonlines(task_dict, local_dir / TASKS_PATH)
def load_tasks(local_dir: Path) -> tuple[dict, dict]:
tasks = load_jsonlines(local_dir / TASKS_PATH)
return {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
task_to_task_index = {task: task_index for task_index, task in tasks.items()}
return tasks, task_to_task_index
def write_episode(episode: dict, local_dir: Path):
append_jsonlines(episode, local_dir / EPISODES_PATH)
def load_episodes(local_dir: Path) -> dict:
return load_jsonlines(local_dir / EPISODES_PATH)
episodes = load_jsonlines(local_dir / EPISODES_PATH)
return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
def load_image_as_numpy(fpath: str | Path, dtype="float32", channel_first: bool = True) -> np.ndarray:
def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
# We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
# is a dictionary of stats and not an integer.
episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
def load_episodes_stats(local_dir: Path) -> dict:
episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
return {
item["episode_index"]: cast_stats_to_numpy(item["stats"])
for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
}
def backward_compatible_episodes_stats(
stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
) -> dict[str, dict[str, np.ndarray]]:
return dict.fromkeys(episodes, stats)
def load_image_as_numpy(
fpath: str | Path, dtype: np.dtype = np.float32, channel_first: bool = True
) -> np.ndarray:
img = PILImage.open(fpath).convert("RGB")
img_array = np.array(img, dtype=dtype)
if channel_first: # (H, W, C) -> (C, H, W)
img_array = np.transpose(img_array, (2, 0, 1))
if "float" in dtype:
if np.issubdtype(dtype, np.floating):
img_array /= 255.0
return img_array
@ -188,77 +269,95 @@ def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
elif first_item is None:
pass
else:
items_dict[key] = [torch.tensor(x) for x in items_dict[key]]
items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
return items_dict
def _get_major_minor(version: str) -> tuple[int]:
split = version.strip("v").split(".")
return int(split[0]), int(split[1])
class BackwardCompatibilityError(Exception):
def __init__(self, repo_id, version):
message = textwrap.dedent(f"""
BackwardCompatibilityError: The dataset you requested ({repo_id}) is in {version} format.
We introduced a new format since v2.0 which is not backward compatible with v1.x.
Please, use our conversion script. Modify the following command with your own task description:
```
python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \\
--repo-id {repo_id} \\
--single-task "TASK DESCRIPTION." # <---- /!\\ Replace TASK DESCRIPTION /!\\
```
A few examples to replace TASK DESCRIPTION: "Pick up the blue cube and place it into the bin.",
"Insert the peg into the socket.", "Slide open the ziploc bag.", "Take the elevator to the 1st floor.",
"Open the top cabinet, store the pot inside it then close the cabinet.", "Push the T-shaped block onto the T-shaped target.",
"Grab the spray paint on the shelf and place it in the bin on top of the robot dog.", "Fold the sweatshirt.", ...
If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
""")
super().__init__(message)
def is_valid_version(version: str) -> bool:
try:
packaging.version.parse(version)
return True
except packaging.version.InvalidVersion:
return False
def check_version_compatibility(
repo_id: str, version_to_check: str, current_version: str, enforce_breaking_major: bool = True
repo_id: str,
version_to_check: str | packaging.version.Version,
current_version: str | packaging.version.Version,
enforce_breaking_major: bool = True,
) -> None:
current_major, _ = _get_major_minor(current_version)
major_to_check, _ = _get_major_minor(version_to_check)
if major_to_check < current_major and enforce_breaking_major:
raise BackwardCompatibilityError(repo_id, version_to_check)
elif float(version_to_check.strip("v")) < float(current_version.strip("v")):
logging.warning(
f"""The dataset you requested ({repo_id}) was created with a previous version ({version_to_check}) of the
codebase. The current codebase version is {current_version}. You should be fine since
backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
v_check = (
packaging.version.parse(version_to_check)
if not isinstance(version_to_check, packaging.version.Version)
else version_to_check
)
v_current = (
packaging.version.parse(current_version)
if not isinstance(current_version, packaging.version.Version)
else current_version
)
if v_check.major < v_current.major and enforce_breaking_major:
raise BackwardCompatibilityError(repo_id, v_check)
elif v_check.minor < v_current.minor:
logging.warning(V21_MESSAGE.format(repo_id=repo_id, version=v_check))
def get_hub_safe_version(repo_id: str, version: str) -> str:
def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
"""Returns available valid versions (branches and tags) on given repo."""
api = HfApi()
dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
branches = [b.name for b in dataset_info.branches]
if version not in branches:
num_version = float(version.strip("v"))
hub_num_versions = [float(v.strip("v")) for v in branches if v.startswith("v")]
if num_version >= 2.0 and all(v < 2.0 for v in hub_num_versions):
raise BackwardCompatibilityError(repo_id, version)
repo_refs = api.list_repo_refs(repo_id, repo_type="dataset")
repo_refs = [b.name for b in repo_refs.branches + repo_refs.tags]
repo_versions = []
for ref in repo_refs:
with contextlib.suppress(packaging.version.InvalidVersion):
repo_versions.append(packaging.version.parse(ref))
logging.warning(
f"""You are trying to load a dataset from {repo_id} created with a previous version of the
codebase. The following versions are available: {branches}.
The requested version ('{version}') is not found. You should be fine since
backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
return repo_versions
def get_safe_version(repo_id: str, version: str | packaging.version.Version) -> str:
"""
Returns the version if available on repo or the latest compatible one.
Otherwise, will throw a `CompatibilityError`.
"""
target_version = (
packaging.version.parse(version) if not isinstance(version, packaging.version.Version) else version
)
if "main" not in branches:
raise ValueError(f"Version 'main' not found on {repo_id}")
return "main"
else:
return version
hub_versions = get_repo_versions(repo_id)
if not hub_versions:
raise RevisionNotFoundError(
f"""Your dataset must be tagged with a codebase version.
Assuming _version_ is the codebase_version value in the info.json, you can run this:
```python
from huggingface_hub import HfApi
hub_api = HfApi()
hub_api.create_tag("{repo_id}", tag="_version_", repo_type="dataset")
```
"""
)
if target_version in hub_versions:
return f"v{target_version}"
compatibles = [
v for v in hub_versions if v.major == target_version.major and v.minor <= target_version.minor
]
if compatibles:
return_version = max(compatibles)
if return_version < target_version:
logging.warning(f"Revision {version} for {repo_id} not found, using version v{return_version}")
return f"v{return_version}"
lower_major = [v for v in hub_versions if v.major < target_version.major]
if lower_major:
raise BackwardCompatibilityError(repo_id, max(lower_major))
upper_versions = [v for v in hub_versions if v > target_version]
assert len(upper_versions) > 0
raise ForwardCompatibilityError(repo_id, min(upper_versions))
def get_hf_features_from_features(features: dict) -> datasets.Features:
@ -270,11 +369,20 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
hf_features[key] = datasets.Image()
elif ft["shape"] == (1,):
hf_features[key] = datasets.Value(dtype=ft["dtype"])
else:
assert len(ft["shape"]) == 1
elif len(ft["shape"]) == 1:
hf_features[key] = datasets.Sequence(
length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
)
elif len(ft["shape"]) == 2:
hf_features[key] = datasets.Array2D(shape=ft["shape"], dtype=ft["dtype"])
elif len(ft["shape"]) == 3:
hf_features[key] = datasets.Array3D(shape=ft["shape"], dtype=ft["dtype"])
elif len(ft["shape"]) == 4:
hf_features[key] = datasets.Array4D(shape=ft["shape"], dtype=ft["dtype"])
elif len(ft["shape"]) == 5:
hf_features[key] = datasets.Array5D(shape=ft["shape"], dtype=ft["dtype"])
else:
raise ValueError(f"Corresponding feature is not valid: {ft}")
return datasets.Features(hf_features)
@ -289,6 +397,37 @@ def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
return {**robot.motor_features, **camera_ft, **DEFAULT_FEATURES}
def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
# TODO(aliberts): Implement "type" in dataset features and simplify this
policy_features = {}
for key, ft in features.items():
shape = ft["shape"]
if ft["dtype"] in ["image", "video"]:
type = FeatureType.VISUAL
if len(shape) != 3:
raise ValueError(f"Number of dimensions of {key} != 3 (shape={shape})")
names = ft["names"]
# Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
if names[2] in ["channel", "channels"]: # (h, w, c) -> (c, h, w)
shape = (shape[2], shape[0], shape[1])
elif key == "observation.environment_state":
type = FeatureType.ENV
elif key.startswith("observation"):
type = FeatureType.STATE
elif key == "action":
type = FeatureType.ACTION
else:
continue
policy_features[key] = PolicyFeature(
type=type,
shape=shape,
)
return policy_features
def create_empty_dataset_info(
codebase_version: str,
fps: int,
@ -314,88 +453,85 @@ def create_empty_dataset_info(
def get_episode_data_index(
episode_dicts: list[dict], episodes: list[int] | None = None
episode_dicts: dict[dict], episodes: list[int] | None = None
) -> dict[str, torch.Tensor]:
episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in enumerate(episode_dicts)}
episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
if episodes is not None:
episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
cumulative_lenghts = list(accumulate(episode_lengths.values()))
cumulative_lengths = list(accumulate(episode_lengths.values()))
return {
"from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
"to": torch.LongTensor(cumulative_lenghts),
}
def calculate_total_episode(
hf_dataset: datasets.Dataset, raise_if_not_contiguous: bool = True
) -> dict[str, torch.Tensor]:
episode_indices = sorted(hf_dataset.unique("episode_index"))
total_episodes = len(episode_indices)
if raise_if_not_contiguous and episode_indices != list(range(total_episodes)):
raise ValueError("episode_index values are not sorted and contiguous.")
return total_episodes
def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:
episode_lengths = []
table = hf_dataset.data.table
total_episodes = calculate_total_episode(hf_dataset)
for ep_idx in range(total_episodes):
ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
episode_lengths.insert(ep_idx, len(ep_table))
cumulative_lenghts = list(accumulate(episode_lengths))
return {
"from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
"to": torch.LongTensor(cumulative_lenghts),
"from": torch.LongTensor([0] + cumulative_lengths[:-1]),
"to": torch.LongTensor(cumulative_lengths),
}
def check_timestamps_sync(
hf_dataset: datasets.Dataset,
episode_data_index: dict[str, torch.Tensor],
timestamps: np.ndarray,
episode_indices: np.ndarray,
episode_data_index: dict[str, np.ndarray],
fps: int,
tolerance_s: float,
raise_value_error: bool = True,
) -> bool:
"""
This check is to make sure that each timestamps is separated to the next by 1/fps +/- tolerance to
account for possible numerical error.
"""
timestamps = torch.stack(hf_dataset["timestamp"])
diffs = torch.diff(timestamps)
within_tolerance = torch.abs(diffs - 1 / fps) <= tolerance_s
This check is to make sure that each timestamp is separated from the next by (1/fps) +/- tolerance
to account for possible numerical error.
# We mask differences between the timestamp at the end of an episode
# and the one at the start of the next episode since these are expected
# to be outside tolerance.
mask = torch.ones(len(diffs), dtype=torch.bool)
ignored_diffs = episode_data_index["to"][:-1] - 1
Args:
timestamps (np.ndarray): Array of timestamps in seconds.
episode_indices (np.ndarray): Array indicating the episode index for each timestamp.
episode_data_index (dict[str, np.ndarray]): A dictionary that includes 'to',
which identifies indices for the end of each episode.
fps (int): Frames per second. Used to check the expected difference between consecutive timestamps.
tolerance_s (float): Allowed deviation from the expected (1/fps) difference.
raise_value_error (bool): Whether to raise a ValueError if the check fails.
Returns:
bool: True if all checked timestamp differences lie within tolerance, False otherwise.
Raises:
ValueError: If the check fails and `raise_value_error` is True.
"""
if timestamps.shape != episode_indices.shape:
raise ValueError(
"timestamps and episode_indices should have the same shape. "
f"Found {timestamps.shape=} and {episode_indices.shape=}."
)
# Consecutive differences
diffs = np.diff(timestamps)
within_tolerance = np.abs(diffs - (1.0 / fps)) <= tolerance_s
# Mask to ignore differences at the boundaries between episodes
mask = np.ones(len(diffs), dtype=bool)
ignored_diffs = episode_data_index["to"][:-1] - 1 # indices at the end of each episode
mask[ignored_diffs] = False
filtered_within_tolerance = within_tolerance[mask]
if not torch.all(filtered_within_tolerance):
# Check if all remaining diffs are within tolerance
if not np.all(filtered_within_tolerance):
# Track original indices before masking
original_indices = torch.arange(len(diffs))
original_indices = np.arange(len(diffs))
filtered_indices = original_indices[mask]
outside_tolerance_filtered_indices = torch.nonzero(~filtered_within_tolerance) # .squeeze()
outside_tolerance_filtered_indices = np.nonzero(~filtered_within_tolerance)[0]
outside_tolerance_indices = filtered_indices[outside_tolerance_filtered_indices]
episode_indices = torch.stack(hf_dataset["episode_index"])
outside_tolerances = []
for idx in outside_tolerance_indices:
entry = {
"timestamps": [timestamps[idx], timestamps[idx + 1]],
"diff": diffs[idx],
"episode_index": episode_indices[idx].item(),
"episode_index": episode_indices[idx].item()
if hasattr(episode_indices[idx], "item")
else episode_indices[idx],
}
outside_tolerances.append(entry)
if raise_value_error:
raise ValueError(
f"""One or several timestamps unexpectedly violate the tolerance inside episode range.
This might be due to synchronization issues with timestamps during data collection.
This might be due to synchronization issues during data collection.
\n{pformat(outside_tolerances)}"""
)
return False
@ -436,7 +572,7 @@ def check_delta_timestamps(
def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
delta_indices = {}
for key, delta_ts in delta_timestamps.items():
delta_indices[key] = (torch.tensor(delta_ts) * fps).long().tolist()
delta_indices[key] = [round(d * fps) for d in delta_ts]
return delta_indices
@ -560,3 +696,118 @@ class IterableNamespace(SimpleNamespace):
def keys(self):
return vars(self).keys()
def validate_frame(frame: dict, features: dict):
optional_features = {"timestamp"}
expected_features = (set(features) - set(DEFAULT_FEATURES.keys())) | {"task"}
actual_features = set(frame.keys())
error_message = validate_features_presence(actual_features, expected_features, optional_features)
if "task" in frame:
error_message += validate_feature_string("task", frame["task"])
common_features = actual_features & (expected_features | optional_features)
for name in common_features - {"task"}:
error_message += validate_feature_dtype_and_shape(name, features[name], frame[name])
if error_message:
raise ValueError(error_message)
def validate_features_presence(
actual_features: set[str], expected_features: set[str], optional_features: set[str]
):
error_message = ""
missing_features = expected_features - actual_features
extra_features = actual_features - (expected_features | optional_features)
if missing_features or extra_features:
error_message += "Feature mismatch in `frame` dictionary:\n"
if missing_features:
error_message += f"Missing features: {missing_features}\n"
if extra_features:
error_message += f"Extra features: {extra_features}\n"
return error_message
def validate_feature_dtype_and_shape(name: str, feature: dict, value: np.ndarray | PILImage.Image | str):
expected_dtype = feature["dtype"]
expected_shape = feature["shape"]
if is_valid_numpy_dtype_string(expected_dtype):
return validate_feature_numpy_array(name, expected_dtype, expected_shape, value)
elif expected_dtype in ["image", "video"]:
return validate_feature_image_or_video(name, expected_shape, value)
elif expected_dtype == "string":
return validate_feature_string(name, value)
else:
raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")
def validate_feature_numpy_array(
name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
):
error_message = ""
if isinstance(value, np.ndarray):
actual_dtype = value.dtype
actual_shape = value.shape
if actual_dtype != np.dtype(expected_dtype):
error_message += f"The feature '{name}' of dtype '{actual_dtype}' is not of the expected dtype '{expected_dtype}'.\n"
if actual_shape != expected_shape:
error_message += f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{expected_shape}'.\n"
else:
error_message += f"The feature '{name}' is not a 'np.ndarray'. Expected type is '{expected_dtype}', but type '{type(value)}' provided instead.\n"
return error_message
def validate_feature_image_or_video(name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image):
# Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
error_message = ""
if isinstance(value, np.ndarray):
actual_shape = value.shape
c, h, w = expected_shape
if len(actual_shape) != 3 or (actual_shape != (c, h, w) and actual_shape != (h, w, c)):
error_message += f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{(c, h, w)}' or '{(h, w, c)}'.\n"
elif isinstance(value, PILImage.Image):
pass
else:
error_message += f"The feature '{name}' is expected to be of type 'PIL.Image' or 'np.ndarray' channel first or channel last, but type '{type(value)}' provided instead.\n"
return error_message
def validate_feature_string(name: str, value: str):
if not isinstance(value, str):
return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
return ""
def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict):
if "size" not in episode_buffer:
raise ValueError("size key not found in episode_buffer")
if "task" not in episode_buffer:
raise ValueError("task key not found in episode_buffer")
if episode_buffer["episode_index"] != total_episodes:
# TODO(aliberts): Add option to use existing episode_index
raise NotImplementedError(
"You might have manually provided the episode_buffer with an episode_index that doesn't "
"match the total number of episodes already in the dataset. This is not supported for now."
)
if episode_buffer["size"] == 0:
raise ValueError("You must add one or several frames with `add_frame` before calling `add_episode`.")
buffer_keys = set(episode_buffer.keys()) - {"task", "size"}
if not buffer_keys == set(features):
raise ValueError(
f"Features from `episode_buffer` don't match the ones in `features`."
f"In episode_buffer not in features: {buffer_keys - set(features)}"
f"In features not in episode_buffer: {set(features) - buffer_keys}"
)

14
lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py
View File

@ -26,13 +26,14 @@ from pathlib import Path
from textwrap import dedent
from lerobot import available_datasets
from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset, parse_robot_config
from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
from lerobot.common.robot_devices.robots.configs import AlohaRobotConfig
LOCAL_DIR = Path("data/")
ALOHA_CONFIG = Path("lerobot/configs/robot/aloha.yaml")
# spellchecker:off
ALOHA_MOBILE_INFO = {
"robot_config": parse_robot_config(ALOHA_CONFIG),
"robot_config": AlohaRobotConfig(),
"license": "mit",
"url": "https://mobile-aloha.github.io/",
"paper": "https://arxiv.org/abs/2401.02117",
@ -45,7 +46,7 @@ ALOHA_MOBILE_INFO = {
}""").lstrip(),
}
ALOHA_STATIC_INFO = {
"robot_config": parse_robot_config(ALOHA_CONFIG),
"robot_config": AlohaRobotConfig(),
"license": "mit",
"url": "https://tonyzhaozh.github.io/aloha/",
"paper": "https://arxiv.org/abs/2304.13705",
@ -159,11 +160,11 @@ DATASETS = {
**ALOHA_STATIC_INFO,
},
"aloha_static_vinh_cup": {
"single_task": "Pick up the platic cup with the right arm, then pop its lid open with the left arm.",
"single_task": "Pick up the plastic cup with the right arm, then pop its lid open with the left arm.",
**ALOHA_STATIC_INFO,
},
"aloha_static_vinh_cup_left": {
"single_task": "Pick up the platic cup with the left arm, then pop its lid open with the right arm.",
"single_task": "Pick up the plastic cup with the left arm, then pop its lid open with the right arm.",
**ALOHA_STATIC_INFO,
},
"aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
@ -856,6 +857,7 @@ DATASETS = {
}""").lstrip(),
},
}
# spellchecker:on
def batch_convert():

61
lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py
View File

@ -17,7 +17,7 @@
"""
This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 1.6 to
2.0. You will be required to provide the 'tasks', which is a short but accurate description in plain English
for each of the task performed in the dataset. This will allow to easily train models with task-conditionning.
for each of the task performed in the dataset. This will allow to easily train models with task-conditioning.
We support 3 different scenarios for these tasks (see instructions below):
1. Single task dataset: all episodes of your dataset have the same single task.
@ -130,7 +130,7 @@ from lerobot.common.datasets.utils import (
create_branch,
create_lerobot_dataset_card,
flatten_dict,
get_hub_safe_version,
get_safe_version,
load_json,
unflatten_dict,
write_json,
@ -141,7 +141,8 @@ from lerobot.common.datasets.video_utils import (
get_image_pixel_channels,
get_video_info,
)
from lerobot.common.utils.utils import init_hydra_config
from lerobot.common.robot_devices.robots.configs import RobotConfig
from lerobot.common.robot_devices.robots.utils import make_robot_config
V16 = "v1.6"
V20 = "v2.0"
@ -152,19 +153,18 @@ V1_INFO_PATH = "meta_data/info.json"
V1_STATS_PATH = "meta_data/stats.safetensors"
def parse_robot_config(config_path: Path, config_overrides: list[str] | None = None) -> tuple[str, dict]:
robot_cfg = init_hydra_config(config_path, config_overrides)
if robot_cfg["robot_type"] in ["aloha", "koch"]:
def parse_robot_config(robot_cfg: RobotConfig) -> tuple[str, dict]:
if robot_cfg.type in ["aloha", "koch"]:
state_names = [
f"{arm}_{motor}" if len(robot_cfg["follower_arms"]) > 1 else motor
for arm in robot_cfg["follower_arms"]
for motor in robot_cfg["follower_arms"][arm]["motors"]
f"{arm}_{motor}" if len(robot_cfg.follower_arms) > 1 else motor
for arm in robot_cfg.follower_arms
for motor in robot_cfg.follower_arms[arm].motors
]
action_names = [
# f"{arm}_{motor}" for arm in ["left", "right"] for motor in robot_cfg["leader_arms"][arm]["motors"]
f"{arm}_{motor}" if len(robot_cfg["leader_arms"]) > 1 else motor
for arm in robot_cfg["leader_arms"]
for motor in robot_cfg["leader_arms"][arm]["motors"]
f"{arm}_{motor}" if len(robot_cfg.leader_arms) > 1 else motor
for arm in robot_cfg.leader_arms
for motor in robot_cfg.leader_arms[arm].motors
]
# elif robot_cfg["robot_type"] == "stretch3": TODO
else:
@ -173,7 +173,7 @@ def parse_robot_config(config_path: Path, config_overrides: list[str] | None = N
)
return {
"robot_type": robot_cfg["robot_type"],
"robot_type": robot_cfg.type,
"names": {
"observation.state": state_names,
"observation.effort": state_names,
@ -203,7 +203,10 @@ def convert_stats_to_json(v1_dir: Path, v2_dir: Path) -> None:
torch.testing.assert_close(stats_json[key], stats[key])
def get_features_from_hf_dataset(dataset: Dataset, robot_config: dict | None = None) -> dict[str, list]:
def get_features_from_hf_dataset(
dataset: Dataset, robot_config: RobotConfig | None = None
) -> dict[str, list]:
robot_config = parse_robot_config(robot_config)
features = {}
for key, ft in dataset.features.items():
if isinstance(ft, datasets.Value):
@ -224,11 +227,11 @@ def get_features_from_hf_dataset(dataset: Dataset, robot_config: dict | None = N
image = dataset[0][key] # Assuming first row
channels = get_image_pixel_channels(image)
shape = (image.height, image.width, channels)
names = ["height", "width", "channel"]
names = ["height", "width", "channels"]
elif ft._type == "VideoFrame":
dtype = "video"
shape = None # Add shape later
names = ["height", "width", "channel"]
names = ["height", "width", "channels"]
features[key] = {
"dtype": dtype,
@ -436,11 +439,11 @@ def convert_dataset(
single_task: str | None = None,
tasks_path: Path | None = None,
tasks_col: Path | None = None,
robot_config: dict | None = None,
robot_config: RobotConfig | None = None,
test_branch: str | None = None,
**card_kwargs,
):
v1 = get_hub_safe_version(repo_id, V16)
v1 = get_safe_version(repo_id, V16)
v1x_dir = local_dir / V16 / repo_id
v20_dir = local_dir / V20 / repo_id
v1x_dir.mkdir(parents=True, exist_ok=True)
@ -478,7 +481,7 @@ def convert_dataset(
# Tasks
if single_task:
tasks_by_episodes = {ep_idx: single_task for ep_idx in episode_indices}
tasks_by_episodes = dict.fromkeys(episode_indices, single_task)
dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
elif tasks_path:
@ -532,7 +535,7 @@ def convert_dataset(
episode_lengths = split_parquet_by_episodes(dataset, total_episodes, total_chunks, v20_dir)
if robot_config is not None:
robot_type = robot_config["robot_type"]
robot_type = robot_config.type
repo_tags = [robot_type]
else:
robot_type = "unknown"
@ -621,16 +624,10 @@ def main():
help="The path to a .json file containing one language instruction for each episode_index",
)
parser.add_argument(
"--robot-config",
type=Path,
default=None,
help="Path to the robot's config yaml the dataset during conversion.",
)
parser.add_argument(
"--robot-overrides",
"--robot",
type=str,
nargs="*",
help="Any key=value arguments to override the robot config values (use dots for.nested=overrides)",
default=None,
help="Robot config used for the dataset during conversion (e.g. 'koch', 'aloha', 'so100', etc.)",
)
parser.add_argument(
"--local-dir",
@ -655,8 +652,10 @@ def main():
if not args.local_dir:
args.local_dir = Path("/tmp/lerobot_dataset_v2")
robot_config = parse_robot_config(args.robot_config, args.robot_overrides) if args.robot_config else None
del args.robot_config, args.robot_overrides
if args.robot is not None:
robot_config = make_robot_config(args.robot)
del args.robot
convert_dataset(**vars(args), robot_config=robot_config)

87
lerobot/common/datasets/v21/_remove_language_instruction.py Normal file
View File

@ -0,0 +1,87 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import traceback
from pathlib import Path
from datasets import get_dataset_config_info
from huggingface_hub import HfApi
from lerobot import available_datasets
from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
from lerobot.common.datasets.utils import INFO_PATH, write_info
from lerobot.common.datasets.v21.convert_dataset_v20_to_v21 import V20, SuppressWarnings
LOCAL_DIR = Path("data/")
hub_api = HfApi()
def fix_dataset(repo_id: str) -> str:
if not hub_api.revision_exists(repo_id, V20, repo_type="dataset"):
return f"{repo_id}: skipped (not in {V20})."
dataset_info = get_dataset_config_info(repo_id, "default")
with SuppressWarnings():
lerobot_metadata = LeRobotDatasetMetadata(repo_id, revision=V20, force_cache_sync=True)
meta_features = {key for key, ft in lerobot_metadata.features.items() if ft["dtype"] != "video"}
parquet_features = set(dataset_info.features)
diff_parquet_meta = parquet_features - meta_features
diff_meta_parquet = meta_features - parquet_features
if diff_parquet_meta:
raise ValueError(f"In parquet not in info.json: {parquet_features - meta_features}")
if not diff_meta_parquet:
return f"{repo_id}: skipped (no diff)"
if diff_meta_parquet:
logging.warning(f"In info.json not in parquet: {meta_features - parquet_features}")
assert diff_meta_parquet == {"language_instruction"}
lerobot_metadata.features.pop("language_instruction")
write_info(lerobot_metadata.info, lerobot_metadata.root)
commit_info = hub_api.upload_file(
path_or_fileobj=lerobot_metadata.root / INFO_PATH,
path_in_repo=INFO_PATH,
repo_id=repo_id,
repo_type="dataset",
revision=V20,
commit_message="Remove 'language_instruction'",
create_pr=True,
)
return f"{repo_id}: success - PR: {commit_info.pr_url}"
def batch_fix():
status = {}
LOCAL_DIR.mkdir(parents=True, exist_ok=True)
logfile = LOCAL_DIR / "fix_features_v20.txt"
for num, repo_id in enumerate(available_datasets):
print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
print("---------------------------------------------------------")
try:
status = fix_dataset(repo_id)
except Exception:
status = f"{repo_id}: failed\n {traceback.format_exc()}"
logging.info(status)
with open(logfile, "a") as file:
file.write(status + "\n")
if __name__ == "__main__":
batch_fix()

54
lerobot/common/datasets/v21/batch_convert_dataset_v20_to_v21.py Normal file
View File

@ -0,0 +1,54 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.1.
"""
import traceback
from pathlib import Path
from huggingface_hub import HfApi
from lerobot import available_datasets
from lerobot.common.datasets.v21.convert_dataset_v20_to_v21 import V21, convert_dataset
LOCAL_DIR = Path("data/")
def batch_convert():
status = {}
LOCAL_DIR.mkdir(parents=True, exist_ok=True)
logfile = LOCAL_DIR / "conversion_log_v21.txt"
hub_api = HfApi()
for num, repo_id in enumerate(available_datasets):
print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
print("---------------------------------------------------------")
try:
if hub_api.revision_exists(repo_id, V21, repo_type="dataset"):
status = f"{repo_id}: success (already in {V21})."
else:
convert_dataset(repo_id)
status = f"{repo_id}: success."
except Exception:
status = f"{repo_id}: failed\n {traceback.format_exc()}"
with open(logfile, "a") as file:
file.write(status + "\n")
if __name__ == "__main__":
batch_convert()

114
lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py Normal file
View File

@ -0,0 +1,114 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 2.0 to
2.1. It will:
- Generate per-episodes stats and writes them in `episodes_stats.jsonl`
- Check consistency between these new stats and the old ones.
- Remove the deprecated `stats.json`.
- Update codebase_version in `info.json`.
- Push this new version to the hub on the 'main' branch and tags it with "v2.1".
Usage:
```bash
python lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py \
--repo-id=aliberts/koch_tutorial
```
"""
import argparse
import logging
from huggingface_hub import HfApi
from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
from lerobot.common.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
from lerobot.common.datasets.v21.convert_stats import check_aggregate_stats, convert_stats
V20 = "v2.0"
V21 = "v2.1"
class SuppressWarnings:
def __enter__(self):
self.previous_level = logging.getLogger().getEffectiveLevel()
logging.getLogger().setLevel(logging.ERROR)
def __exit__(self, exc_type, exc_val, exc_tb):
logging.getLogger().setLevel(self.previous_level)
def convert_dataset(
repo_id: str,
branch: str | None = None,
num_workers: int = 4,
):
with SuppressWarnings():
dataset = LeRobotDataset(repo_id, revision=V20, force_cache_sync=True)
if (dataset.root / EPISODES_STATS_PATH).is_file():
(dataset.root / EPISODES_STATS_PATH).unlink()
convert_stats(dataset, num_workers=num_workers)
ref_stats = load_stats(dataset.root)
check_aggregate_stats(dataset, ref_stats)
dataset.meta.info["codebase_version"] = CODEBASE_VERSION
write_info(dataset.meta.info, dataset.root)
dataset.push_to_hub(branch=branch, tag_version=False, allow_patterns="meta/")
# delete old stats.json file
if (dataset.root / STATS_PATH).is_file:
(dataset.root / STATS_PATH).unlink()
hub_api = HfApi()
if hub_api.file_exists(
repo_id=dataset.repo_id, filename=STATS_PATH, revision=branch, repo_type="dataset"
):
hub_api.delete_file(
path_in_repo=STATS_PATH, repo_id=dataset.repo_id, revision=branch, repo_type="dataset"
)
hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--repo-id",
type=str,
required=True,
help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
"(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
)
parser.add_argument(
"--branch",
type=str,
default=None,
help="Repo branch to push your dataset. Defaults to the main branch.",
)
parser.add_argument(
"--num-workers",
type=int,
default=4,
help="Number of workers for parallelizing stats compute. Defaults to 4.",
)
args = parser.parse_args()
convert_dataset(**vars(args))

99
lerobot/common/datasets/v21/convert_stats.py Normal file
View File

@ -0,0 +1,99 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from concurrent.futures import ThreadPoolExecutor, as_completed
import numpy as np
from tqdm import tqdm
from lerobot.common.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
from lerobot.common.datasets.utils import write_episode_stats
def sample_episode_video_frames(dataset: LeRobotDataset, episode_index: int, ft_key: str) -> np.ndarray:
ep_len = dataset.meta.episodes[episode_index]["length"]
sampled_indices = sample_indices(ep_len)
query_timestamps = dataset._get_query_timestamps(0.0, {ft_key: sampled_indices})
video_frames = dataset._query_videos(query_timestamps, episode_index)
return video_frames[ft_key].numpy()
def convert_episode_stats(dataset: LeRobotDataset, ep_idx: int):
ep_start_idx = dataset.episode_data_index["from"][ep_idx]
ep_end_idx = dataset.episode_data_index["to"][ep_idx]
ep_data = dataset.hf_dataset.select(range(ep_start_idx, ep_end_idx))
ep_stats = {}
for key, ft in dataset.features.items():
if ft["dtype"] == "video":
# We sample only for videos
ep_ft_data = sample_episode_video_frames(dataset, ep_idx, key)
else:
ep_ft_data = np.array(ep_data[key])
axes_to_reduce = (0, 2, 3) if ft["dtype"] in ["image", "video"] else 0
keepdims = True if ft["dtype"] in ["image", "video"] else ep_ft_data.ndim == 1
ep_stats[key] = get_feature_stats(ep_ft_data, axis=axes_to_reduce, keepdims=keepdims)
if ft["dtype"] in ["image", "video"]: # remove batch dim
ep_stats[key] = {
k: v if k == "count" else np.squeeze(v, axis=0) for k, v in ep_stats[key].items()
}
dataset.meta.episodes_stats[ep_idx] = ep_stats
def convert_stats(dataset: LeRobotDataset, num_workers: int = 0):
assert dataset.episodes is None
print("Computing episodes stats")
total_episodes = dataset.meta.total_episodes
if num_workers > 0:
with ThreadPoolExecutor(max_workers=num_workers) as executor:
futures = {
executor.submit(convert_episode_stats, dataset, ep_idx): ep_idx
for ep_idx in range(total_episodes)
}
for future in tqdm(as_completed(futures), total=total_episodes):
future.result()
else:
for ep_idx in tqdm(range(total_episodes)):
convert_episode_stats(dataset, ep_idx)
for ep_idx in tqdm(range(total_episodes)):
write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
def check_aggregate_stats(
dataset: LeRobotDataset,
reference_stats: dict[str, dict[str, np.ndarray]],
video_rtol_atol: tuple[float] = (1e-2, 1e-2),
default_rtol_atol: tuple[float] = (5e-6, 6e-5),
):
"""Verifies that the aggregated stats from episodes_stats are close to reference stats."""
agg_stats = aggregate_stats(list(dataset.meta.episodes_stats.values()))
for key, ft in dataset.features.items():
# These values might need some fine-tuning
if ft["dtype"] == "video":
# to account for image sub-sampling
rtol, atol = video_rtol_atol
else:
rtol, atol = default_rtol_atol
for stat, val in agg_stats[key].items():
if key in reference_stats and stat in reference_stats[key]:
err_msg = f"feature='{key}' stats='{stat}'"
np.testing.assert_allclose(
val, reference_stats[key][stat], rtol=rtol, atol=atol, err_msg=err_msg
)

123
lerobot/common/datasets/video_utils.py
View File

@ -13,6 +13,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import importlib
import json
import logging
import subprocess
@ -29,6 +30,46 @@ from datasets.features.features import register_feature
from PIL import Image
def get_safe_default_codec():
if importlib.util.find_spec("torchcodec"):
return "torchcodec"
else:
logging.warning(
"'torchcodec' is not available in your platform, falling back to 'pyav' as a default decoder"
)
return "pyav"
def decode_video_frames(
video_path: Path | str,
timestamps: list[float],
tolerance_s: float,
backend: str | None = None,
) -> torch.Tensor:
"""
Decodes video frames using the specified backend.
Args:
video_path (Path): Path to the video file.
timestamps (list[float]): List of timestamps to extract frames.
tolerance_s (float): Allowed deviation in seconds for frame retrieval.
backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav"..
Returns:
torch.Tensor: Decoded frames.
Currently supports torchcodec on cpu and pyav.
"""
if backend is None:
backend = get_safe_default_codec()
if backend == "torchcodec":
return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s)
elif backend in ["pyav", "video_reader"]:
return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
else:
raise ValueError(f"Unsupported video backend: {backend}")
def decode_video_frames_torchvision(
video_path: Path | str,
timestamps: list[float],
@ -69,11 +110,11 @@ def decode_video_frames_torchvision(
# set the first and last requested timestamps
# Note: previous timestamps are usually loaded, since we need to access the previous key frame
first_ts = timestamps[0]
last_ts = timestamps[-1]
first_ts = min(timestamps)
last_ts = max(timestamps)
# access closest key frame of the first requested frame
# Note: closest key frame timestamp is usally smaller than `first_ts` (e.g. key frame can be the first frame of the video)
# Note: closest key frame timestamp is usually smaller than `first_ts` (e.g. key frame can be the first frame of the video)
# for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
reader.seek(first_ts, keyframes_only=keyframes_only)
@ -127,6 +168,81 @@ def decode_video_frames_torchvision(
return closest_frames
def decode_video_frames_torchcodec(
video_path: Path | str,
timestamps: list[float],
tolerance_s: float,
device: str = "cpu",
log_loaded_timestamps: bool = False,
) -> torch.Tensor:
"""Loads frames associated with the requested timestamps of a video using torchcodec.
Note: Setting device="cuda" outside the main process, e.g. in data loader workers, will lead to CUDA initialization errors.
Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
and all subsequent frames until reaching the requested frame. The number of key frames in a video
can be adjusted during encoding to take into account decoding time and video size in bytes.
"""
if importlib.util.find_spec("torchcodec"):
from torchcodec.decoders import VideoDecoder
else:
raise ImportError("torchcodec is required but not available.")
# initialize video decoder
decoder = VideoDecoder(video_path, device=device, seek_mode="approximate")
loaded_frames = []
loaded_ts = []
# get metadata for frame information
metadata = decoder.metadata
average_fps = metadata.average_fps
# convert timestamps to frame indices
frame_indices = [round(ts * average_fps) for ts in timestamps]
# retrieve frames based on indices
frames_batch = decoder.get_frames_at(indices=frame_indices)
for frame, pts in zip(frames_batch.data, frames_batch.pts_seconds, strict=False):
loaded_frames.append(frame)
loaded_ts.append(pts.item())
if log_loaded_timestamps:
logging.info(f"Frame loaded at timestamp={pts:.4f}")
query_ts = torch.tensor(timestamps)
loaded_ts = torch.tensor(loaded_ts)
# compute distances between each query timestamp and loaded timestamps
dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
min_, argmin_ = dist.min(1)
is_within_tol = min_ < tolerance_s
assert is_within_tol.all(), (
f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
"It means that the closest frame that can be loaded from the video is too far away in time."
"This might be due to synchronization issues with timestamps during data collection."
"To be safe, we advise to ignore this item during training."
f"\nqueried timestamps: {query_ts}"
f"\nloaded timestamps: {loaded_ts}"
f"\nvideo: {video_path}"
)
# get closest frames to the query timestamps
closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
closest_ts = loaded_ts[argmin_]
if log_loaded_timestamps:
logging.info(f"{closest_ts=}")
# convert to float32 in [0,1] range (channel first)
closest_frames = closest_frames.type(torch.float32) / 255
assert len(timestamps) == len(closest_frames)
return closest_frames
def encode_video_frames(
imgs_dir: Path | str,
video_path: Path | str,
@ -141,6 +257,7 @@ def encode_video_frames(
) -> None:
"""More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
video_path = Path(video_path)
imgs_dir = Path(imgs_dir)
video_path.parent.mkdir(parents=True, exist_ok=True)
ffmpeg_args = OrderedDict(

15
lerobot/common/envs/__init__.py Normal file
View File

@ -0,0 +1,15 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv # noqa: F401

156
lerobot/common/envs/configs.py Normal file
View File

@ -0,0 +1,156 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
from dataclasses import dataclass, field
import draccus
from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
from lerobot.configs.types import FeatureType, PolicyFeature
@dataclass
class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
task: str | None = None
fps: int = 30
features: dict[str, PolicyFeature] = field(default_factory=dict)
features_map: dict[str, str] = field(default_factory=dict)
@property
def type(self) -> str:
return self.get_choice_name(self.__class__)
@abc.abstractproperty
def gym_kwargs(self) -> dict:
raise NotImplementedError()
@EnvConfig.register_subclass("aloha")
@dataclass
class AlohaEnv(EnvConfig):
task: str = "AlohaInsertion-v0"
fps: int = 50
episode_length: int = 400
obs_type: str = "pixels_agent_pos"
render_mode: str = "rgb_array"
features: dict[str, PolicyFeature] = field(
default_factory=lambda: {
"action": PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
}
)
features_map: dict[str, str] = field(
default_factory=lambda: {
"action": ACTION,
"agent_pos": OBS_ROBOT,
"top": f"{OBS_IMAGE}.top",
"pixels/top": f"{OBS_IMAGES}.top",
}
)
def __post_init__(self):
if self.obs_type == "pixels":
self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
elif self.obs_type == "pixels_agent_pos":
self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(14,))
self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
@property
def gym_kwargs(self) -> dict:
return {
"obs_type": self.obs_type,
"render_mode": self.render_mode,
"max_episode_steps": self.episode_length,
}
@EnvConfig.register_subclass("pusht")
@dataclass
class PushtEnv(EnvConfig):
task: str = "PushT-v0"
fps: int = 10
episode_length: int = 300
obs_type: str = "pixels_agent_pos"
render_mode: str = "rgb_array"
visualization_width: int = 384
visualization_height: int = 384
features: dict[str, PolicyFeature] = field(
default_factory=lambda: {
"action": PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
"agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(2,)),
}
)
features_map: dict[str, str] = field(
default_factory=lambda: {
"action": ACTION,
"agent_pos": OBS_ROBOT,
"environment_state": OBS_ENV,
"pixels": OBS_IMAGE,
}
)
def __post_init__(self):
if self.obs_type == "pixels_agent_pos":
self.features["pixels"] = PolicyFeature(type=FeatureType.VISUAL, shape=(384, 384, 3))
elif self.obs_type == "environment_state_agent_pos":
self.features["environment_state"] = PolicyFeature(type=FeatureType.ENV, shape=(16,))
@property
def gym_kwargs(self) -> dict:
return {
"obs_type": self.obs_type,
"render_mode": self.render_mode,
"visualization_width": self.visualization_width,
"visualization_height": self.visualization_height,
"max_episode_steps": self.episode_length,
}
@EnvConfig.register_subclass("xarm")
@dataclass
class XarmEnv(EnvConfig):
task: str = "XarmLift-v0"
fps: int = 15
episode_length: int = 200
obs_type: str = "pixels_agent_pos"
render_mode: str = "rgb_array"
visualization_width: int = 384
visualization_height: int = 384
features: dict[str, PolicyFeature] = field(
default_factory=lambda: {
"action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
"pixels": PolicyFeature(type=FeatureType.VISUAL, shape=(84, 84, 3)),
}
)
features_map: dict[str, str] = field(
default_factory=lambda: {
"action": ACTION,
"agent_pos": OBS_ROBOT,
"pixels": OBS_IMAGE,
}
)
def __post_init__(self):
if self.obs_type == "pixels_agent_pos":
self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
@property
def gym_kwargs(self) -> dict:
return {
"obs_type": self.obs_type,
"render_mode": self.render_mode,
"visualization_width": self.visualization_width,
"visualization_height": self.visualization_height,
"max_episode_steps": self.episode_length,
}

55
lerobot/common/envs/factory.py
View File

@ -16,43 +16,54 @@
import importlib
import gymnasium as gym
from omegaconf import DictConfig
from lerobot.common.envs.configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv
def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv | None:
"""Makes a gym vector environment according to the evaluation config.
def make_env_config(env_type: str, **kwargs) -> EnvConfig:
if env_type == "aloha":
return AlohaEnv(**kwargs)
elif env_type == "pusht":
return PushtEnv(**kwargs)
elif env_type == "xarm":
return XarmEnv(**kwargs)
else:
raise ValueError(f"Policy type '{env_type}' is not available.")
n_envs can be used to override eval.batch_size in the configuration. Must be at least 1.
def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> gym.vector.VectorEnv | None:
"""Makes a gym vector environment according to the config.
Args:
cfg (EnvConfig): the config of the environment to instantiate.
n_envs (int, optional): The number of parallelized env to return. Defaults to 1.
use_async_envs (bool, optional): Whether to return an AsyncVectorEnv or a SyncVectorEnv. Defaults to
False.
Raises:
ValueError: if n_envs < 1
ModuleNotFoundError: If the requested env package is not installed
Returns:
gym.vector.VectorEnv: The parallelized gym.env instance.
"""
if n_envs is not None and n_envs < 1:
if n_envs < 1:
raise ValueError("`n_envs must be at least 1")
if cfg.env.name == "real_world":
return
package_name = f"gym_{cfg.env.name}"
package_name = f"gym_{cfg.type}"
try:
importlib.import_module(package_name)
except ModuleNotFoundError as e:
print(
f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.env.name}]'`"
)
print(f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.type}]'`")
raise e
gym_handle = f"{package_name}/{cfg.env.task}"
gym_kwgs = dict(cfg.env.get("gym", {}))
if cfg.env.get("episode_length"):
gym_kwgs["max_episode_steps"] = cfg.env.episode_length
gym_handle = f"{package_name}/{cfg.task}"
# batched version of the env that returns an observation of shape (b, c)
env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
env = env_cls(
[
lambda: gym.make(gym_handle, disable_env_checker=True, **gym_kwgs)
for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
]
[lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
)
return env

65
lerobot/common/envs/utils.py
View File

@ -13,13 +13,22 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import warnings
from typing import Any
import einops
import gymnasium as gym
import numpy as np
import torch
from torch import Tensor
from lerobot.common.envs.configs import EnvConfig
from lerobot.common.utils.utils import get_channel_first_image_shape
from lerobot.configs.types import FeatureType, PolicyFeature
def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
# TODO(aliberts, rcadene): refactor this to use features from the environment (no hardcoding)
"""Convert environment observation to LeRobot format observation.
Args:
observation: Dictionary of observation batches from a Gym vector environment.
@ -35,6 +44,7 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
imgs = {"observation.image": observations["pixels"]}
for imgkey, img in imgs.items():
# TODO(aliberts, rcadene): use transforms.ToTensor()?
img = torch.from_numpy(img)
# sanity check that images are channel last
@ -60,3 +70,58 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
# requirement for "agent_pos"
return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
return return_observations
def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
# TODO(aliberts, rcadene): remove this hardcoding of keys and just use the nested keys as is
# (need to also refactor preprocess_observation and externalize normalization from policies)
policy_features = {}
for key, ft in env_cfg.features.items():
if ft.type is FeatureType.VISUAL:
if len(ft.shape) != 3:
raise ValueError(f"Number of dimensions of {key} != 3 (shape={ft.shape})")
shape = get_channel_first_image_shape(ft.shape)
feature = PolicyFeature(type=ft.type, shape=shape)
else:
feature = ft
policy_key = env_cfg.features_map[key]
policy_features[policy_key] = feature
return policy_features
def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
first_type = type(env.envs[0]) # Get type of first env
return all(type(e) is first_type for e in env.envs) # Fast type check
def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
with warnings.catch_warnings():
warnings.simplefilter("once", UserWarning) # Apply filter only in this function
if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
warnings.warn(
"The environment does not have 'task_description' and 'task'. Some policies require these features.",
UserWarning,
stacklevel=2,
)
if not are_all_envs_same_type(env):
warnings.warn(
"The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
UserWarning,
stacklevel=2,
)
def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
"""Adds task feature to the observation dict with respect to the first environment attribute."""
if hasattr(env.envs[0], "task_description"):
observation["task"] = env.call("task_description")
elif hasattr(env.envs[0], "task"):
observation["task"] = env.call("task")
else: # For envs without language instructions, e.g. aloha transfer cube and etc.
num_envs = observation[list(observation.keys())[0]].shape[0]
observation["task"] = ["" for _ in range(num_envs)]
return observation

246
lerobot/common/logger.py
View File

@ -1,246 +0,0 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py
# TODO(rcadene, alexander-soare): clean this file
"""
import logging
import os
import re
from glob import glob
from pathlib import Path
import torch
from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
from omegaconf import DictConfig, OmegaConf
from termcolor import colored
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LRScheduler
from lerobot.common.policies.policy_protocol import Policy
from lerobot.common.utils.utils import get_global_random_state, set_global_random_state
def log_output_dir(out_dir):
logging.info(colored("Output dir:", "yellow", attrs=["bold"]) + f" {out_dir}")
def cfg_to_group(cfg: DictConfig, return_list: bool = False) -> list[str] | str:
"""Return a group name for logging. Optionally returns group name as list."""
lst = [
f"policy:{cfg.policy.name}",
f"dataset:{cfg.dataset_repo_id}",
f"env:{cfg.env.name}",
f"seed:{cfg.seed}",
]
return lst if return_list else "-".join(lst)
def get_wandb_run_id_from_filesystem(checkpoint_dir: Path) -> str:
# Get the WandB run ID.
paths = glob(str(checkpoint_dir / "../wandb/latest-run/run-*"))
if len(paths) != 1:
raise RuntimeError("Couldn't get the previous WandB run ID for run resumption.")
match = re.search(r"run-([^\.]+).wandb", paths[0].split("/")[-1])
if match is None:
raise RuntimeError("Couldn't get the previous WandB run ID for run resumption.")
wandb_run_id = match.groups(0)[0]
return wandb_run_id
class Logger:
"""Primary logger object. Logs either locally or using wandb.
The logger creates the following directory structure:
provided_log_dir
├── .hydra # hydra's configuration cache
├── checkpoints
│ ├── specific_checkpoint_name
│ │ ├── pretrained_model # Hugging Face pretrained model directory
│ │ │ ├── ...
│ │ └── training_state.pth # optimizer, scheduler, and random states + training step
| ├── another_specific_checkpoint_name
│ │ ├── ...
| ├── ...
│ └── last # a softlink to the last logged checkpoint
"""
pretrained_model_dir_name = "pretrained_model"
training_state_file_name = "training_state.pth"
def __init__(self, cfg: DictConfig, log_dir: str, wandb_job_name: str | None = None):
"""
Args:
log_dir: The directory to save all logs and training outputs to.
job_name: The WandB job name.
"""
self._cfg = cfg
self.log_dir = Path(log_dir)
self.log_dir.mkdir(parents=True, exist_ok=True)
self.checkpoints_dir = self.get_checkpoints_dir(log_dir)
self.last_checkpoint_dir = self.get_last_checkpoint_dir(log_dir)
self.last_pretrained_model_dir = self.get_last_pretrained_model_dir(log_dir)
# Set up WandB.
self._group = cfg_to_group(cfg)
project = cfg.get("wandb", {}).get("project")
entity = cfg.get("wandb", {}).get("entity")
enable_wandb = cfg.get("wandb", {}).get("enable", False)
run_offline = not enable_wandb or not project
if run_offline:
logging.info(colored("Logs will be saved locally.", "yellow", attrs=["bold"]))
self._wandb = None
else:
os.environ["WANDB_SILENT"] = "true"
import wandb
wandb_run_id = None
if cfg.resume:
wandb_run_id = get_wandb_run_id_from_filesystem(self.checkpoints_dir)
wandb.init(
id=wandb_run_id,
project=project,
entity=entity,
name=wandb_job_name,
notes=cfg.get("wandb", {}).get("notes"),
tags=cfg_to_group(cfg, return_list=True),
dir=log_dir,
config=OmegaConf.to_container(cfg, resolve=True),
# TODO(rcadene): try set to True
save_code=False,
# TODO(rcadene): split train and eval, and run async eval with job_type="eval"
job_type="train_eval",
resume="must" if cfg.resume else None,
)
print(colored("Logs will be synced with wandb.", "blue", attrs=["bold"]))
logging.info(f"Track this run --> {colored(wandb.run.get_url(), 'yellow', attrs=['bold'])}")
self._wandb = wandb
@classmethod
def get_checkpoints_dir(cls, log_dir: str | Path) -> Path:
"""Given the log directory, get the sub-directory in which checkpoints will be saved."""
return Path(log_dir) / "checkpoints"
@classmethod
def get_last_checkpoint_dir(cls, log_dir: str | Path) -> Path:
"""Given the log directory, get the sub-directory in which the last checkpoint will be saved."""
return cls.get_checkpoints_dir(log_dir) / "last"
@classmethod
def get_last_pretrained_model_dir(cls, log_dir: str | Path) -> Path:
"""
Given the log directory, get the sub-directory in which the last checkpoint's pretrained weights will
be saved.
"""
return cls.get_last_checkpoint_dir(log_dir) / cls.pretrained_model_dir_name
def save_model(self, save_dir: Path, policy: Policy, wandb_artifact_name: str | None = None):
"""Save the weights of the Policy model using PyTorchModelHubMixin.
The weights are saved in a folder called "pretrained_model" under the checkpoint directory.
Optionally also upload the model to WandB.
"""
self.checkpoints_dir.mkdir(parents=True, exist_ok=True)
policy.save_pretrained(save_dir)
# Also save the full Hydra config for the env configuration.
OmegaConf.save(self._cfg, save_dir / "config.yaml")
if self._wandb and not self._cfg.wandb.disable_artifact:
# note wandb artifact does not accept ":" or "/" in its name
artifact = self._wandb.Artifact(wandb_artifact_name, type="model")
artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
self._wandb.log_artifact(artifact)
if self.last_checkpoint_dir.exists():
os.remove(self.last_checkpoint_dir)
def save_training_state(
self,
save_dir: Path,
train_step: int,
optimizer: Optimizer,
scheduler: LRScheduler | None,
):
"""Checkpoint the global training_step, optimizer state, scheduler state, and random state.
All of these are saved as "training_state.pth" under the checkpoint directory.
"""
training_state = {
"step": train_step,
"optimizer": optimizer.state_dict(),
**get_global_random_state(),
}
if scheduler is not None:
training_state["scheduler"] = scheduler.state_dict()
torch.save(training_state, save_dir / self.training_state_file_name)
def save_checkpoint(
self,
train_step: int,
policy: Policy,
optimizer: Optimizer,
scheduler: LRScheduler | None,
identifier: str,
):
"""Checkpoint the model weights and the training state."""
checkpoint_dir = self.checkpoints_dir / str(identifier)
wandb_artifact_name = (
None
if self._wandb is None
else f"{self._group.replace(':', '_').replace('/', '_')}-{self._cfg.seed}-{identifier}"
)
self.save_model(
checkpoint_dir / self.pretrained_model_dir_name, policy, wandb_artifact_name=wandb_artifact_name
)
self.save_training_state(checkpoint_dir, train_step, optimizer, scheduler)
os.symlink(checkpoint_dir.absolute(), self.last_checkpoint_dir)
def load_last_training_state(self, optimizer: Optimizer, scheduler: LRScheduler | None) -> int:
"""
Given the last checkpoint in the logging directory, load the optimizer state, scheduler state, and
random state, and return the global training step.
"""
training_state = torch.load(self.last_checkpoint_dir / self.training_state_file_name)
optimizer.load_state_dict(training_state["optimizer"])
if scheduler is not None:
scheduler.load_state_dict(training_state["scheduler"])
elif "scheduler" in training_state:
raise ValueError(
"The checkpoint contains a scheduler state_dict, but no LRScheduler was provided."
)
# Small hack to get the expected keys: use `get_global_random_state`.
set_global_random_state({k: training_state[k] for k in get_global_random_state()})
return training_state["step"]
def log_dict(self, d, step, mode="train"):
assert mode in {"train", "eval"}
# TODO(alexander-soare): Add local text log.
if self._wandb is not None:
for k, v in d.items():
if not isinstance(v, (int, float, str)):
logging.warning(
f'WandB logging of key "{k}" was ignored as its type is not handled by this wrapper.'
)
continue
self._wandb.log({f"{mode}/{k}": v}, step=step)
def log_video(self, video_path: str, step: int, mode: str = "train"):
assert mode in {"train", "eval"}
assert self._wandb is not None
wandb_video = self._wandb.Video(video_path, fps=self._cfg.fps, format="mp4")
self._wandb.log({f"{mode}/video": wandb_video}, step=step)

15
lerobot/common/optim/__init__.py Normal file
View File

@ -0,0 +1,15 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .optimizers import OptimizerConfig as OptimizerConfig

40
lerobot/common/optim/factory.py Normal file
View File

@ -0,0 +1,40 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LRScheduler
from lerobot.common.policies.pretrained import PreTrainedPolicy
from lerobot.configs.train import TrainPipelineConfig
def make_optimizer_and_scheduler(
cfg: TrainPipelineConfig, policy: PreTrainedPolicy
) -> tuple[Optimizer, LRScheduler | None]:
"""Generates the optimizer and scheduler based on configs.
Args:
cfg (TrainPipelineConfig): The training config that contains optimizer and scheduler configs
policy (PreTrainedPolicy): The policy config from which parameters and presets must be taken from.
Returns:
tuple[Optimizer, LRScheduler | None]: The couple (Optimizer, Scheduler). Scheduler can be `None`.
"""
params = policy.get_optim_params() if cfg.use_policy_training_preset else policy.parameters()
optimizer = cfg.optimizer.build(params)
lr_scheduler = cfg.scheduler.build(optimizer, cfg.steps) if cfg.scheduler is not None else None
return optimizer, lr_scheduler

118
lerobot/common/optim/optimizers.py Normal file
View File

@ -0,0 +1,118 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
from dataclasses import asdict, dataclass
from pathlib import Path
import draccus
import torch
from safetensors.torch import load_file, save_file
from lerobot.common.constants import (
OPTIMIZER_PARAM_GROUPS,
OPTIMIZER_STATE,
)
from lerobot.common.datasets.utils import flatten_dict, unflatten_dict, write_json
from lerobot.common.utils.io_utils import deserialize_json_into_object
@dataclass
class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
lr: float
weight_decay: float
grad_clip_norm: float
@property
def type(self) -> str:
return self.get_choice_name(self.__class__)
@classmethod
def default_choice_name(cls) -> str | None:
return "adam"
@abc.abstractmethod
def build(self) -> torch.optim.Optimizer:
raise NotImplementedError
@OptimizerConfig.register_subclass("adam")
@dataclass
class AdamConfig(OptimizerConfig):
lr: float = 1e-3
betas: tuple[float, float] = (0.9, 0.999)
eps: float = 1e-8
weight_decay: float = 0.0
grad_clip_norm: float = 10.0
def build(self, params: dict) -> torch.optim.Optimizer:
kwargs = asdict(self)
kwargs.pop("grad_clip_norm")
return torch.optim.Adam(params, **kwargs)
@OptimizerConfig.register_subclass("adamw")
@dataclass
class AdamWConfig(OptimizerConfig):
lr: float = 1e-3
betas: tuple[float, float] = (0.9, 0.999)
eps: float = 1e-8
weight_decay: float = 1e-2
grad_clip_norm: float = 10.0
def build(self, params: dict) -> torch.optim.Optimizer:
kwargs = asdict(self)
kwargs.pop("grad_clip_norm")
return torch.optim.AdamW(params, **kwargs)
@OptimizerConfig.register_subclass("sgd")
@dataclass
class SGDConfig(OptimizerConfig):
lr: float = 1e-3
momentum: float = 0.0
dampening: float = 0.0
nesterov: bool = False
weight_decay: float = 0.0
grad_clip_norm: float = 10.0
def build(self, params: dict) -> torch.optim.Optimizer:
kwargs = asdict(self)
kwargs.pop("grad_clip_norm")
return torch.optim.SGD(params, **kwargs)
def save_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
state = optimizer.state_dict()
param_groups = state.pop("param_groups")
flat_state = flatten_dict(state)
save_file(flat_state, save_dir / OPTIMIZER_STATE)
write_json(param_groups, save_dir / OPTIMIZER_PARAM_GROUPS)
def load_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
current_state_dict = optimizer.state_dict()
flat_state = load_file(save_dir / OPTIMIZER_STATE)
state = unflatten_dict(flat_state)
loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
if "param_groups" in current_state_dict:
param_groups = deserialize_json_into_object(
save_dir / OPTIMIZER_PARAM_GROUPS, current_state_dict["param_groups"]
)
loaded_state_dict["param_groups"] = param_groups
optimizer.load_state_dict(loaded_state_dict)
return optimizer

122
lerobot/common/optim/schedulers.py Normal file
View File

@ -0,0 +1,122 @@
#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
import math
from dataclasses import asdict, dataclass
from pathlib import Path
import draccus
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LambdaLR, LRScheduler
from lerobot.common.constants import SCHEDULER_STATE
from lerobot.common.datasets.utils import write_json
from lerobot.common.utils.io_utils import deserialize_json_into_object
@dataclass
class LRSchedulerConfig(draccus.ChoiceRegistry, abc.ABC):
num_warmup_steps: int
@property
def type(self) -> str:
return self.get_choice_name(self.__class__)
@abc.abstractmethod
def build(self, optimizer: Optimizer, num_training_steps: int) -> LRScheduler | None:
raise NotImplementedError
@LRSchedulerConfig.register_subclass("diffuser")
@dataclass
class DiffuserSchedulerConfig(LRSchedulerConfig):
name: str = "cosine"
num_warmup_steps: int | None = None
def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
from diffusers.optimization import get_scheduler
kwargs = {**asdict(self), "num_training_steps": num_training_steps, "optimizer": optimizer}
return get_scheduler(**kwargs)
@LRSchedulerConfig.register_subclass("vqbet")
@dataclass
class VQBeTSchedulerConfig(LRSchedulerConfig):
num_warmup_steps: int
num_vqvae_training_steps: int
num_cycles: float = 0.5
def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
def lr_lambda(current_step):
if current_step < self.num_vqvae_training_steps:
return float(1)
else:
adjusted_step = current_step - self.num_vqvae_training_steps
if adjusted_step < self.num_warmup_steps:
return float(adjusted_step) / float(max(1, self.num_warmup_steps))
progress = float(adjusted_step - self.num_warmup_steps) / float(
max(1, num_training_steps - self.num_warmup_steps)
)
return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)))
return LambdaLR(optimizer, lr_lambda, -1)
@LRSchedulerConfig.register_subclass("cosine_decay_with_warmup")
@dataclass
class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
"""Used by Physical Intelligence to train Pi0"""
num_warmup_steps: int
num_decay_steps: int
peak_lr: float
decay_lr: float
def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
del num_training_steps
def lr_lambda(current_step):
def linear_warmup_schedule(current_step):
if current_step <= 0:
return 1 / (self.num_warmup_steps + 1)
frac = 1 - current_step / self.num_warmup_steps
return (1 / (self.num_warmup_steps + 1) - 1) * frac + 1
def cosine_decay_schedule(current_step):
step = min(current_step, self.num_decay_steps)
cosine_decay = 0.5 * (1 + math.cos(math.pi * step / self.num_decay_steps))
alpha = self.decay_lr / self.peak_lr
decayed = (1 - alpha) * cosine_decay + alpha
return decayed
if current_step < self.num_warmup_steps:
return linear_warmup_schedule(current_step)
return cosine_decay_schedule(current_step)
return LambdaLR(optimizer, lr_lambda, -1)
def save_scheduler_state(scheduler: LRScheduler, save_dir: Path) -> None:
state_dict = scheduler.state_dict()
write_json(state_dict, save_dir / SCHEDULER_STATE)
def load_scheduler_state(scheduler: LRScheduler, save_dir: Path) -> LRScheduler:
state_dict = deserialize_json_into_object(save_dir / SCHEDULER_STATE, scheduler.state_dict())
scheduler.load_state_dict(state_dict)
return scheduler

19
lerobot/common/policies/__init__.py Normal file
View File

@ -0,0 +1,19 @@
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .act.configuration_act import ACTConfig as ACTConfig
from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
from .pi0.configuration_pi0 import PI0Config as PI0Config
from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig

69
lerobot/common/policies/act/configuration_act.py
View File

@ -15,9 +15,14 @@
# limitations under the License.
from dataclasses import dataclass, field
from lerobot.common.optim.optimizers import AdamWConfig
from lerobot.configs.policies import PreTrainedConfig
from lerobot.configs.types import NormalizationMode
@PreTrainedConfig.register_subclass("act")
@dataclass
class ACTConfig:
class ACTConfig(PreTrainedConfig):
"""Configuration class for the Action Chunking Transformers policy.
Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
@ -59,7 +64,7 @@ class ACTConfig:
output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
original scale. Note that this is also used for normalizing the training targets.
vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
pretrained_backbone_weights: Pretrained weights from torchvision to initalize the backbone.
pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
`None` means no pretrained weights.
replace_final_stride_with_dilation: Whether to replace the ResNet's final 2x2 stride with a dilated
convolution.
@ -90,28 +95,11 @@ class ACTConfig:
chunk_size: int = 100
n_action_steps: int = 100
input_shapes: dict[str, list[int]] = field(
normalization_mapping: dict[str, NormalizationMode] = field(
default_factory=lambda: {
"observation.images.top": [3, 480, 640],
"observation.state": [14],
}
)
output_shapes: dict[str, list[int]] = field(
default_factory=lambda: {
"action": [14],
}
)
# Normalization / Unnormalization
input_normalization_modes: dict[str, str] = field(
default_factory=lambda: {
"observation.images.top": "mean_std",
"observation.state": "mean_std",
}
)
output_normalization_modes: dict[str, str] = field(
default_factory=lambda: {
"action": "mean_std",
"VISUAL": NormalizationMode.MEAN_STD,
"STATE": NormalizationMode.MEAN_STD,
"ACTION": NormalizationMode.MEAN_STD,
}
)
@ -144,7 +132,14 @@ class ACTConfig:
dropout: float = 0.1
kl_weight: float = 10.0
# Training preset
optimizer_lr: float = 1e-5
optimizer_weight_decay: float = 1e-4
optimizer_lr_backbone: float = 1e-5
def __post_init__(self):
super().__post_init__()
"""Input validation (not exhaustive)."""
if not self.vision_backbone.startswith("resnet"):
raise ValueError(
@ -164,8 +159,28 @@ class ACTConfig:
raise ValueError(
f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
)
if (
not any(k.startswith("observation.image") for k in self.input_shapes)
and "observation.environment_state" not in self.input_shapes
):
def get_optimizer_preset(self) -> AdamWConfig:
return AdamWConfig(
lr=self.optimizer_lr,
weight_decay=self.optimizer_weight_decay,
)
def get_scheduler_preset(self) -> None:
return None
def validate_features(self) -> None:
if not self.image_features and not self.env_state_feature:
raise ValueError("You must provide at least one image or the environment state among the inputs.")
@property
def observation_delta_indices(self) -> None:
return None
@property
def action_delta_indices(self) -> list:
return list(range(self.chunk_size))
@property
def reward_delta_indices(self) -> None:
return None

Some files were not shown because too many files have changed in this diff Show More