Rename Aloha2 to Aloha

lerobot/configs/env/aloha_real.yaml

@@ -4,7 +4,7 @@ fps: 30
 
 env:
   name: dora
-  task: DoraAloha2-v0
+  task: DoraAloha-v0
   state_dim: 14
   action_dim: 14
   fps: ${fps}

lerobot/configs/policy/act_real.yaml

@@ -1,7 +1,21 @@
 # @package _global_
 
+# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. cam_right_wrist, cam_left_wrist, images,
+# cam_low) instead of 1 camera (i.e. top). Also, `training.eval_freq` is set to -1. This config is used
+# to evaluate checkpoints at a certain frequency of training steps. When it is set to -1, it deactivates evaluation.
+# This is because real-world evaluation is done through [dora-lerobot](https://github.com/dora-rs/dora-lerobot).
+# Look at its README for more information on how to evaluate a checkpoint in the real-world.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_real \
+#   env=aloha_real
+# ```
+
 seed: 1000
-dataset_repo_id: cadene/wrist_gripper
+dataset_repo_id: lerobot/aloha_static_vinh_cup
 
 override_dataset_stats:
   observation.images.cam_right_wrist:

lerobot/configs/policy/act_real_no_state.yaml

@@ -0,0 +1,111 @@
+# @package _global_
+
+# Use `act_real_no_state.yaml` to train on real-world Aloha/Aloha2 datasets when cameras are moving (e.g. wrist cameras)
+# Compared to `act_real.yaml`, it is camera only and does not use the state as input which is vector of robot joint positions.
+# We validated experimentaly that not using state reaches better success rate. Our hypothesis is that `act_real.yaml` might
+# overfits to the state, because the images are more complex to learn from since they are moving.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_real_no_state \
+#   env=aloha2_real
+# ```
+
+seed: 1000
+dataset_repo_id: lerobot/aloha_static_vinh_cup
+
+override_dataset_stats:
+  observation.images.cam_right_wrist:
+    # 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.cam_left_wrist:
+    # 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.cam_high:
+    # 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.cam_low:
+    # 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: -1
+  save_freq: 10000
+  log_freq: 100
+  save_checkpoint: 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:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.cam_right_wrist: [3, 480, 640]
+    observation.images.cam_left_wrist: [3, 480, 640]
+    observation.images.cam_high: [3, 480, 640]
+    observation.images.cam_low: [3, 480, 640]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.cam_right_wrist: mean_std
+    observation.images.cam_left_wrist: mean_std
+    observation.images.cam_high: mean_std
+    observation.images.cam_low: mean_std
+  output_normalization_modes:
+    action: mean_std
+
+  # 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_momentum: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0