Merge 6b808dfbe2 into 1c873df5c0

benchmarks/video/capture_camera_feed.py

@@ -17,12 +17,19 @@
 
 import argparse
 import datetime as dt
+import os
 from pathlib import Path
 
 import cv2
+import keyboard
+import rerun as rr
 
 
 def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
+    rr.init("lerobot_capture_camera_feed")
+    memory_limit = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
+    rr.spawn(memory_limit=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():
@@ -45,18 +52,18 @@ def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height
         if not ret:
             print("Error: Could not read frame.")
             break
-
+        rr.log("video/stream", rr.Image(frame.numpy()), static=True)
-        cv2.imshow("Video Stream", frame)
         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"):
+        if keyboard.is_pressed("q"):
             break
 
     # Release the capture and destroy all windows
     cap.release()
-    cv2.destroyAllWindows()
+    # TODO(Steven): Find a way to close visualizer: https://github.com/rerun-io/rerun/pull/9400
+    # cv2.destroyAllWindows()
 
 
 if __name__ == "__main__":

examples/10_use_so100.md

@@ -491,6 +491,13 @@ python lerobot/scripts/control_robot.py \
 
 #### 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`. You can adjust the viewer's behavior with these environment variables:
+> - `RERUN_FLUSH_NUM_BYTES`
+> - `LEROBOT_RERUN_MEMORY_LIMIT`
+> - `LEROBOT_VIEWER_IP` (only for remote robots)
+> - `LEROBOT_VIEWER_PORT` (only for remote robots)
+
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=so100 \

examples/11_use_lekiwi.md

@@ -412,6 +412,12 @@ python lerobot/scripts/control_robot.py \
   --control.fps=30
 ```
 
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`. You can adjust the viewer's behavior with these environment variables:
+> - `RERUN_FLUSH_NUM_BYTES`
+> - `LEROBOT_RERUN_MEMORY_LIMIT`
+> - `LEROBOT_VIEWER_IP` (only for remote robots)
+> - `LEROBOT_VIEWER_PORT` (only for remote robots)
+
 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) |
 | ---------- | ------------------ | ---------------------- |

examples/11_use_moss.md

@@ -212,6 +212,13 @@ python lerobot/scripts/control_robot.py \
 
 **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`. You can adjust the viewer's behavior with these environment variables:
+> - `RERUN_FLUSH_NUM_BYTES`
+> - `LEROBOT_RERUN_MEMORY_LIMIT`
+> - `LEROBOT_VIEWER_IP` (only for remote robots)
+> - `LEROBOT_VIEWER_PORT` (only for remote robots)
+
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=moss \

examples/7_get_started_with_real_robot.md

@@ -55,6 +55,13 @@ Finally, connect both arms to your computer via USB. Note that the USB doesn't p
 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`. You can adjust the viewer's behavior with these environment variables:
+> - `RERUN_FLUSH_NUM_BYTES`
+> - `LEROBOT_RERUN_MEMORY_LIMIT`
+> - `LEROBOT_VIEWER_IP` (only for remote robots)
+> - `LEROBOT_VIEWER_PORT` (only for remote robots)
+
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=koch \

examples/8_use_stretch.md

@@ -97,6 +97,13 @@ This is equivalent to running `stretch_robot_home.py`
 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`. You can adjust the viewer's behavior with these environment variables:
+> - `RERUN_FLUSH_NUM_BYTES`
+> - `LEROBOT_RERUN_MEMORY_LIMIT`
+> - `LEROBOT_VIEWER_IP` (only for remote robots)
+> - `LEROBOT_VIEWER_PORT` (only for remote robots)
+
 ```bash
 python lerobot/scripts/control_robot.py \
     --robot.type=stretch \

examples/9_use_aloha.md

@@ -43,6 +43,13 @@ 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`. You can adjust the viewer's behavior with these environment variables:
+> - `RERUN_FLUSH_NUM_BYTES`
+> - `LEROBOT_RERUN_MEMORY_LIMIT`
+> - `LEROBOT_VIEWER_IP` (only for remote robots)
+> - `LEROBOT_VIEWER_PORT` (only for remote robots)
+
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=aloha \

lerobot/common/robot_devices/control_configs.py

@@ -41,7 +41,7 @@ class TeleoperateControlConfig(ControlConfig):
     fps: int | None = None
     teleop_time_s: float | None = None
     # Display all cameras on screen
-    display_cameras: bool = True
+    display_data: bool = False
 
 
 @ControlConfig.register_subclass("record")
@@ -82,7 +82,7 @@ class RecordControlConfig(ControlConfig):
     # Not enough threads might cause low camera fps.
     num_image_writer_threads_per_camera: int = 4
     # Display all cameras on screen
-    display_cameras: bool = True
+    display_data: bool = False
     # Use vocal synthesis to read events.
     play_sounds: bool = True
     # Resume recording on an existing dataset.
@@ -116,6 +116,8 @@ class ReplayControlConfig(ControlConfig):
 @dataclass
 class RemoteRobotConfig(ControlConfig):
     log_interval: int = 100
+    # Display all cameras on screen
+    display_data: bool = False
 
 
 @dataclass

lerobot/common/robot_devices/control_utils.py

@@ -24,7 +24,7 @@ from contextlib import nullcontext
 from copy import copy
 from functools import cache
 
-import cv2
+import rerun as rr
 import torch
 from deepdiff import DeepDiff
 from termcolor import colored
@@ -174,13 +174,13 @@ def warmup_record(
     events,
     enable_teleoperation,
     warmup_time_s,
-    display_cameras,
+    display_data,
     fps,
 ):
     control_loop(
         robot=robot,
         control_time_s=warmup_time_s,
-        display_cameras=display_cameras,
+        display_data=display_data,
         events=events,
         fps=fps,
         teleoperate=enable_teleoperation,
@@ -192,7 +192,7 @@ def record_episode(
     dataset,
     events,
     episode_time_s,
-    display_cameras,
+    display_data,
     policy,
     fps,
     single_task,
@@ -200,7 +200,7 @@ def record_episode(
     control_loop(
         robot=robot,
         control_time_s=episode_time_s,
-        display_cameras=display_cameras,
+        display_data=display_data,
         dataset=dataset,
         events=events,
         policy=policy,
@@ -215,7 +215,7 @@ def control_loop(
     robot,
     control_time_s=None,
     teleoperate=False,
-    display_cameras=False,
+    display_data=False,
     dataset: LeRobotDataset | None = None,
     events=None,
     policy: PreTrainedPolicy = None,
@@ -264,11 +264,15 @@ def control_loop(
             frame = {**observation, **action, "task": single_task}
             dataset.add_frame(frame)
 
-        if display_cameras and not is_headless():
+        # TODO(Steven): This should be more general (for RemoteRobot instead of checking the name, but anyways it will change soon)
+        if (display_data and not is_headless()) or (display_data and robot.robot_type.startswith("lekiwi")):
+            for k, v in action.items():
+                for i, vv in enumerate(v):
+                    rr.log(f"sent_{k}_{i}", rr.Scalar(vv.numpy()))
+
             image_keys = [key for key in observation if "image" in key]
             for key in image_keys:
-                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
+                rr.log(key, rr.Image(observation[key].numpy()), static=True)
-            cv2.waitKey(1)
 
         if fps is not None:
             dt_s = time.perf_counter() - start_loop_t
@@ -297,15 +301,15 @@ def reset_environment(robot, events, reset_time_s, fps):
     )
 
 
-def stop_recording(robot, listener, display_cameras):
+def stop_recording(robot, listener, display_data):
     robot.disconnect()
 
-    if not is_headless():
+    if not is_headless() and listener is not None:
-        if listener is not None:
         listener.stop()
 
-        if display_cameras:
+        # TODO(Steven): Find a way to close visualizer: https://github.com/rerun-io/rerun/pull/9400
-            cv2.destroyAllWindows()
+        # if display_data:
+        #     cv2.destroyAllWindows()
 
 
 def sanity_check_dataset_name(repo_id, policy_cfg):

lerobot/scripts/control_robot.py

@@ -135,15 +135,19 @@ python lerobot/scripts/control_robot.py \
 """
 
 import logging
+import os
 import time
 from dataclasses import asdict
 from pprint import pformat
 
+import rerun as rr
+
 # from safetensors.torch import load_file, save_file
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.robot_devices.control_configs import (
     CalibrateControlConfig,
+    ControlConfig,
     ControlPipelineConfig,
     RecordControlConfig,
     RemoteRobotConfig,
@@ -153,6 +157,7 @@ from lerobot.common.robot_devices.control_configs import (
 from lerobot.common.robot_devices.control_utils import (
     control_loop,
     init_keyboard_listener,
+    is_headless,
     log_control_info,
     record_episode,
     reset_environment,
@@ -232,7 +237,7 @@ def teleoperate(robot: Robot, cfg: TeleoperateControlConfig):
         control_time_s=cfg.teleop_time_s,
         fps=cfg.fps,
         teleoperate=True,
-        display_cameras=cfg.display_cameras,
+        display_data=cfg.display_data,
     )
 
 
@@ -280,7 +285,7 @@ def record(
     # 3. place the cameras windows on screen
     enable_teleoperation = policy is None
     log_say("Warmup record", cfg.play_sounds)
-    warmup_record(robot, events, enable_teleoperation, cfg.warmup_time_s, cfg.display_cameras, cfg.fps)
+    warmup_record(robot, events, enable_teleoperation, cfg.warmup_time_s, cfg.display_data, cfg.fps)
 
     if has_method(robot, "teleop_safety_stop"):
         robot.teleop_safety_stop()
@@ -296,7 +301,7 @@ def record(
             dataset=dataset,
             events=events,
             episode_time_s=cfg.episode_time_s,
-            display_cameras=cfg.display_cameras,
+            display_data=cfg.display_data,
             policy=policy,
             fps=cfg.fps,
             single_task=cfg.single_task,
@@ -326,7 +331,7 @@ def record(
             break
 
     log_say("Stop recording", cfg.play_sounds, blocking=True)
-    stop_recording(robot, listener, cfg.display_cameras)
+    stop_recording(robot, listener, cfg.display_data)
 
     if cfg.push_to_hub:
         dataset.push_to_hub(tags=cfg.tags, private=cfg.private)
@@ -363,6 +368,42 @@ def replay(
         log_control_info(robot, dt_s, fps=cfg.fps)
 
 
+def _init_rerun(control_config: ControlConfig, session_name: str = "lerobot_control_loop") -> None:
+    """Initializes the Rerun SDK for visualizing the control loop.
+
+    Args:
+        control_config: Configuration determining data display and robot type.
+        session_name: Rerun session name. Defaults to "lerobot_control_loop".
+
+    Raises:
+        ValueError: If viewer IP is missing for non-remote configurations with display enabled.
+    """
+    if (control_config.display_data and not is_headless()) or (
+        control_config.display_data and isinstance(control_config, RemoteRobotConfig)
+    ):
+        # Configure Rerun flush batch size
+        batch_size = os.getenv("RERUN_FLUSH_NUM_BYTES", "8000")
+        os.environ["RERUN_FLUSH_NUM_BYTES"] = batch_size
+
+        # Get memory limit and viewer connection parameters
+        memory_limit = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
+        viewer_ip = os.getenv("LEROBOT_VIEWER_IP")
+        viewer_port = os.getenv("LEROBOT_VIEWER_PORT", "9876")
+
+        # Initialize Rerun based on configuration
+        rr.init(session_name)
+        if isinstance(control_config, RemoteRobotConfig):
+            if not viewer_ip:
+                raise ValueError(
+                    "Viewer IP required for remote config. Set LEROBOT_VIEWER_IP "
+                    "or disable control_config.display_data."
+                )
+            logging.info(f"Connecting to viewer at {viewer_ip}:{viewer_port}")
+            rr.connect_tcp(f"{viewer_ip}:{viewer_port}")
+        else:
+            rr.spawn(memory_limit=memory_limit)
+
+
 @parser.wrap()
 def control_robot(cfg: ControlPipelineConfig):
     init_logging()
@@ -370,17 +411,22 @@ def control_robot(cfg: ControlPipelineConfig):
 
     robot = make_robot_from_config(cfg.robot)
 
+    # TODO(Steven): Blueprint for fixed window size
+
     if isinstance(cfg.control, CalibrateControlConfig):
         calibrate(robot, cfg.control)
     elif isinstance(cfg.control, TeleoperateControlConfig):
+        _init_rerun(control_config=cfg.control, session_name="lerobot_control_loop_teleop")
         teleoperate(robot, cfg.control)
     elif isinstance(cfg.control, RecordControlConfig):
+        _init_rerun(control_config=cfg.control, session_name="lerobot_control_loop_record")
         record(robot, cfg.control)
     elif isinstance(cfg.control, ReplayControlConfig):
         replay(robot, cfg.control)
     elif isinstance(cfg.control, RemoteRobotConfig):
         from lerobot.common.robot_devices.robots.lekiwi_remote import run_lekiwi
 
+        _init_rerun(control_config=cfg.control, session_name="lerobot_control_loop_remote")
         run_lekiwi(cfg.robot)
 
     if robot.is_connected:

pyproject.toml

@@ -60,7 +60,8 @@ dependencies = [
     "jsonlines>=4.0.0",
     "numba>=0.59.0",
     "omegaconf>=2.3.0",
-    "opencv-python>=4.9.0",
+    "opencv-python-headless>=4.9.0",
+    "keyboard>=0.13.5",
     "packaging>=24.2",
     "av>=12.0.5,<13.0.0",
     "pymunk>=6.6.0",

tests/robots/test_control_robot.py

@@ -172,8 +172,7 @@ def test_record_and_replay_and_policy(tmp_path, request, robot_type, mock):
         push_to_hub=False,
         # TODO(rcadene, aliberts): test video=True
         video=False,
-        # TODO(rcadene): display cameras through cv2 sometimes crashes on mac
+        display_data=False,
-        display_cameras=False,
         play_sounds=False,
     )
     dataset = record(robot, rec_cfg)
@@ -226,7 +225,7 @@ def test_record_and_replay_and_policy(tmp_path, request, robot_type, mock):
         num_episodes=2,
         push_to_hub=False,
         video=False,
-        display_cameras=False,
+        display_data=False,
         play_sounds=False,
         num_image_writer_processes=num_image_writer_processes,
     )
@@ -273,7 +272,7 @@ def test_resume_record(tmp_path, request, robot_type, mock):
         episode_time_s=1,
         push_to_hub=False,
         video=False,
-        display_cameras=False,
+        display_data=False,
         play_sounds=False,
         num_episodes=1,
     )
@@ -330,7 +329,7 @@ def test_record_with_event_rerecord_episode(tmp_path, request, robot_type, mock)
             num_episodes=1,
             push_to_hub=False,
             video=False,
-            display_cameras=False,
+            display_data=False,
             play_sounds=False,
         )
         dataset = record(robot, rec_cfg)
@@ -380,7 +379,7 @@ def test_record_with_event_exit_early(tmp_path, request, robot_type, mock):
             num_episodes=1,
             push_to_hub=False,
             video=False,
-            display_cameras=False,
+            display_data=False,
             play_sounds=False,
         )
 
@@ -433,7 +432,7 @@ def test_record_with_event_stop_recording(tmp_path, request, robot_type, mock, n
             num_episodes=2,
             push_to_hub=False,
             video=False,
-            display_cameras=False,
+            display_data=False,
             play_sounds=False,
             num_image_writer_processes=num_image_writer_processes,
         )