WIP

Co-authored-by: rj <rj@teleopstrio-razer.lan>
Co-authored-by: Remi <re.cadene@gmail.com>

.gitattributes

@@ -1,2 +1,6 @@
 *.memmap filter=lfs diff=lfs merge=lfs -text
 *.stl filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

.github/workflows/build-docker-images.yml

@@ -10,7 +10,6 @@ on:
 
 env:
   PYTHON_VERSION: "3.10"
-#   CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
 
 jobs:
   latest-cpu:
@@ -35,6 +34,8 @@ jobs:
 
       - name: Check out code
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Login to DockerHub
         uses: docker/login-action@v3
@@ -51,34 +52,50 @@ jobs:
           tags: huggingface/lerobot-cpu
           build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
 
-    #   - name: Post to a Slack channel
-    #     id: slack
-    #     #uses: slackapi/slack-github-action@v1.25.0
-    #     uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-    #     with:
-    #       # Slack channel id, channel name, or user id to post message.
-    #       # See also: https://api.slack.com/methods/chat.postMessage#channels
-    #       channel-id: ${{ env.CI_SLACK_CHANNEL }}
-    #       # For posting a rich message using Block Kit
-    #       payload: |
-    #         {
-    #           "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
-    #           "blocks": [
-    #             {
-    #               "type": "section",
-    #               "text": {
-    #                 "type": "mrkdwn",
-    #                 "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
-    #               }
-    #             }
-    #           ]
-    #         }
-    #     env:
-    #       SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
 
   latest-cuda:
     name: GPU
     runs-on: ubuntu-latest
+    steps:
+      - name: Cleanup disk
+        run: |
+          sudo df -h
+          # sudo ls -l /usr/local/lib/
+          # sudo ls -l /usr/share/
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /usr/share/dotnet
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo df -h
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+
+
+  latest-cuda-dev:
+    name: GPU Dev
+    runs-on: ubuntu-latest
     steps:
       - name: Cleanup disk
         run: |
@@ -104,36 +121,11 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
-      - name: Build and Push GPU
+      - name: Build and Push GPU dev
         uses: docker/build-push-action@v5
         with:
           context: .
-          file: ./docker/lerobot-gpu/Dockerfile
+          file: ./docker/lerobot-gpu-dev/Dockerfile
           push: true
-          tags: huggingface/lerobot-gpu
+          tags: huggingface/lerobot-gpu:dev
           build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
-
-      # - name: Post to a Slack channel
-      #   id: slack
-      #   #uses: slackapi/slack-github-action@v1.25.0
-      #   uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-      #   with:
-      #     # Slack channel id, channel name, or user id to post message.
-      #     # See also: https://api.slack.com/methods/chat.postMessage#channels
-      #     channel-id: ${{ env.CI_SLACK_CHANNEL }}
-      #     # For posting a rich message using Block Kit
-      #     payload: |
-      #       {
-      #         "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
-      #         "blocks": [
-      #           {
-      #             "type": "section",
-      #             "text": {
-      #               "type": "mrkdwn",
-      #               "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
-      #             }
-      #           }
-      #         ]
-      #       }
-      #   env:
-      #     SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}

.github/workflows/nightly-tests.yml

@@ -70,6 +70,8 @@ jobs:
       #     files: ./coverage.xml
       #     verbose: true
       - name: Tests end-to-end
+        env:
+          DEVICE: cuda
         run: make test-end-to-end
 
     #   - name: Generate Report

.github/workflows/test.yml

@@ -29,6 +29,8 @@ jobs:
       MUJOCO_GL: egl
     steps:
       - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
 
       - name: Install EGL
         run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
@@ -57,6 +59,40 @@ jobs:
             && rm -rf tests/outputs outputs
 
 
+  pytest-minimal:
+    name: Pytest (minimal install)
+    runs-on: ubuntu-latest
+    env:
+      DATA_DIR: tests/data
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
+
+      - name: Install poetry
+        run: |
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      - name: Set up Python 3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Install poetry dependencies
+        run: |
+          poetry install --extras "test"
+
+      - name: Test with pytest
+        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
+
+
   end-to-end:
     name: End-to-end
     runs-on: ubuntu-latest
@@ -65,6 +101,8 @@ jobs:
       MUJOCO_GL: egl
     steps:
       - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
 
       - name: Install EGL
         run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev

.gitignore

@@ -2,12 +2,17 @@
 logs
 tmp
 wandb
+
+# Data
 data
 outputs
-.vscode
-rl
+
+# Apple
 .DS_Store
 
+# VS Code
+.vscode
+
 # HPC
 nautilus/*.yaml
 *.key
@@ -90,6 +95,7 @@ instance/
 docs/_build/
 
 # PyBuilder
+.pybuilder/
 target/
 
 # Jupyter Notebook
@@ -102,13 +108,6 @@ ipython_config.py
 # pyenv
 .python-version
 
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow
 __pypackages__/
 
@@ -119,6 +118,14 @@ celerybeat.pid
 # SageMath parsed files
 *.sage.py
 
+# Environments
+.env
+.venv
+venv/
+ENV/
+env.bak/
+venv.bak/
+
 # Spyder project settings
 .spyderproject
 .spyproject
@@ -136,3 +143,9 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

CONTRIBUTING.md

@@ -139,13 +139,11 @@ Follow these steps to start contributing:
 
    To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
    ```bash
-   pip install poetry
    poetry install --sync --extras "dev test"
    ```
 
    You can also install the project with all its dependencies (including environments):
    ```bash
-   pip install poetry
    poetry install --sync --all-extras
    ```
 
@@ -197,6 +195,11 @@ Follow these steps to start contributing:
    git commit
    ```
 
+   Note, if you already commited some changes that have a wrong formatting, you can use:
+   ```bash
+   pre-commit run --all-files
+   ```
+
    Please write [good commit messages](https://chris.beams.io/posts/git-commit/).
 
    It is a good idea to sync your copy of the code with the original

Makefile

@@ -10,6 +10,7 @@ endif
 
 export PATH := $(dir $(PYTHON_PATH)):$(PATH)
 
+DEVICE ?= cpu
 
 build-cpu:
 	docker build -t lerobot:latest -f docker/lerobot-cpu/Dockerfile .
@@ -18,26 +19,29 @@ build-gpu:
 	docker build -t lerobot:latest -f docker/lerobot-gpu/Dockerfile .
 
 test-end-to-end:
-	${MAKE} test-act-ete-train
-	${MAKE} test-act-ete-eval
-	${MAKE} test-diffusion-ete-train
-	${MAKE} test-diffusion-ete-eval
-	# TODO(rcadene, alexander-soare): enable end-to-end tests for tdmpc
-	# ${MAKE} test-tdmpc-ete-train
-	# ${MAKE} test-tdmpc-ete-eval
-	${MAKE} test-default-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train
+	${MAKE} DEVICE=$(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-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=cpu \
-		training.save_model=true \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
 		training.save_freq=2 \
 		policy.n_action_steps=20 \
 		policy.chunk_size=20 \
@@ -46,61 +50,92 @@ test-act-ete-train:
 
 test-act-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/act/checkpoints/000002 \
+		-p tests/outputs/act/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		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/ \
+		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
 
 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=cpu \
-		training.save_model=true \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
 		training.save_freq=2 \
 		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/diffusion/checkpoints/000002 \
+		-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		device=$(DEVICE) \
 
+# TODO(alexander-soare): Restore online_steps to 2 when it is reinstated.
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
 		policy=tdmpc \
 		env=xarm \
 		env.task=XarmLift-v0 \
-		dataset_repo_id=lerobot/xarm_lift_medium_replay \
+		dataset_repo_id=lerobot/xarm_lift_medium \
 		wandb.enable=False \
 		training.offline_steps=2 \
-		training.online_steps=2 \
+		training.online_steps=0 \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=2 \
-		device=cpu \
-		training.save_model=true \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
 		training.save_freq=2 \
 		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/tdmpc/
 
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/tdmpc/checkpoints/000002 \
+		-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
-
+		device=$(DEVICE) \
 
 test-default-ete-eval:
 	python lerobot/scripts/eval.py \
@@ -108,4 +143,21 @@ test-default-ete-eval:
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		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 \
+		hydra.run.dir=tests/outputs/act_pusht/
+	rm lerobot/configs/policy/created_by_Makefile.yaml

README.md

@@ -57,7 +57,6 @@
 - Thanks to Tony Zaho, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
 - Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
 - Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
-- Thanks to Vincent Moens and colleagues for open sourcing [TorchRL](https://github.com/pytorch/rl). It allowed for quick experimentations on the design of `LeRobot`.
 - Thanks to Antonio Loquercio and Ashish Kumar for their early support.
 
 
@@ -78,6 +77,10 @@ Install 🤗 LeRobot:
 pip install .
 ```
 
+> **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`
+
 For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
 - [aloha](https://github.com/huggingface/gym-aloha)
 - [xarm](https://github.com/huggingface/gym-xarm)
@@ -93,11 +96,14 @@ To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tra
 wandb login
 ```
 
+(note: you will also need to enable WandB in the configuration. See below.)
+
 ## Walkthrough
 
 ```
 .
 ├── 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
@@ -121,13 +127,21 @@ wandb login
 
 Check out [example 1](./examples/1_load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically download data from the Hugging Face hub.
 
-You can also locally visualize episodes from a dataset by executing our script from the command line:
+You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
     --repo-id lerobot/pusht \
     --episode-index 0
 ```
 
+or from a dataset in a local folder with the root `DATA_DIR` environment variable
+```bash
+DATA_DIR='./my_local_data_dir' python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --episode-index 0
+```
+
+
 It will open `rerun.io` and display the camera streams, robot states and actions, like this:
 
 https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
@@ -135,6 +149,51 @@ https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-f
 
 Our script can also visualize datasets stored on a distant server. See `python lerobot/scripts/visualize_dataset.py --help` for more instructions.
 
+### The `LeRobotDataset` format
+
+A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and Pytorch dataset. For instance `dataset[0]` will retrieve a sample of the dataset observations and actions in pytorch tensors format ready to be fed to a model.
+
+A specificity of `LeRobotDataset` is that we can retrieve several frames for one sample query. By setting `delta_timestamps` to a list of delta timestamps, e.g. `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}`  one can retrieve, for each query, 4 images including one at -1 second before the current time step, the two others at -0.5 second and -0.2, and the final one at the current time step (0 second). See example [1_load_lerobot_dataset.py](examples/1_load_lerobot_dataset.py) for more details on `delta_timestamps`.
+
+Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states.
+
+Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
+
+```
+dataset attributes:
+  ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
+  │  ├ observation.images.cam_high: VideoFrame
+  │  │   VideoFrame = {'path': path to a mp4 video, 'timestamp': float32 timestamp in the video}
+  │  ├ observation.state: List of float32: position of an arm joints (for instance)
+  │  ... (more observations)
+  │  ├ action: List of float32
+  │  ├ episode_index: int64: index of the episode for this sample
+  │  ├ frame_index: int64: index of the frame for this sample in the episode ; starts at 0 for each episode
+  │  ├ timestamp: float32: timestamp in the episode
+  │  ├ next.done: bool: indicates the end of en episode ; True for the last frame in each episode
+  │  └ index: int64: general index in the whole dataset
+  ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
+  │  ├ from: 1D int64 tensor of first frame index for each episode: shape (num episodes,) starts with 0
+  │  └ to: 1D int64 tensor of last frame index for each episode: shape (num episodes,)
+  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
+  │  ├ observation.images.cam_high: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
+  │  ...
+  ├ info: a dictionary of metadata on the dataset
+  │  ├ fps: float - frame per second the dataset is recorded/synchronized to
+  │  └ video: bool - indicates if frames are encoded in mp4 video files to save space or stored as png files
+  ├ videos_dir: path to where the mp4 videos or png images are stored/accessed
+  └ camera_keys: List of string: the keys to access camera features in the item returned by the dataset (e.g. `["observation.images.cam_high", ...]`)
+```
+
+A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
+- hf_dataset stored using Hugging Face datasets library serialization to parquet
+- videos are stored in mp4 format to save space or png files
+- episode_data_index saved using `safetensor` tensor serializtion format
+- stats saved using `safetensor` tensor serializtion format
+- info are saved using JSON
+
+Dataset can uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can set the `DATA_DIR` environment variable to you root dataset folder as illustrated in the above section on dataset visualization.
+
 ### Evaluate a pretrained policy
 
 Check out [example 2](./examples/2_evaluate_pretrained_policy.py) that illustrates how to download a pretrained policy from Hugging Face hub, and run an evaluation on its corresponding environment.
@@ -148,24 +207,25 @@ python lerobot/scripts/eval.py \
 ```
 
 Note: After training your own policy, you can re-evaluate the checkpoints with:
+
 ```bash
-python lerobot/scripts/eval.py \
-    -p PATH/TO/TRAIN/OUTPUT/FOLDER
+python lerobot/scripts/eval.py -p {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 start training a model.
+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.
+
+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:
 
-In general, you can use our training script to easily train any policy. To use wandb for logging training and evaluation curves, make sure you ran `wandb login`. Here is an example of training the ACT policy on trajectories collected by humans on the Aloha simulation environment for the insertion task:
 ```bash
 python lerobot/scripts/train.py \
     policy=act \
     env=aloha \
     env.task=AlohaInsertion-v0 \
-    dataset_repo_id=lerobot/aloha_sim_insertion_human
+    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:
@@ -173,17 +233,42 @@ The experiment directory is automatically generated and will show up in yellow i
     hydra.run.dir=your/new/experiment/dir
 ```
 
-A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of logs from wandb:
-![](media/wandb.png)
+In the experiment directory there will be a folder called `checkpoints` which will have the following structure:
 
-You can deactivate wandb by adding these arguments to the `train.py` python command:
 ```bash
-    wandb.disable_artifact=true \
-    wandb.enable=false
+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
 ```
 
-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. 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.
+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:
+
+![](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.
+
+#### 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:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=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
 
@@ -196,11 +281,11 @@ To add a dataset to the hub, you need to login using a write-access token, which
 huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
-Then move your dataset folder in `data` directory (e.g. `data/aloha_ping_pong`), and push your dataset to the hub with:
+Then move your dataset folder in `data` directory (e.g. `data/aloha_static_pingpong_test`), and push your dataset to the hub with:
 ```bash
 python lerobot/scripts/push_dataset_to_hub.py \
 --data-dir data \
---dataset-id aloha_ping_ping \
+--dataset-id aloha_static_pingpong_test \
 --raw-format aloha_hdf5 \
 --community-id lerobot
 ```
@@ -214,14 +299,14 @@ If your dataset format is not supported, implement your own in `lerobot/common/d
 
 Once you have trained a policy you may upload it to the Hugging Face hub using a hub id that looks like `${hf_user}/${repo_name}` (e.g. [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)).
 
-You first need to find the checkpoint located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). It should contain:
+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.
 
 To upload these to the hub, run the following:
 ```bash
-huggingface-cli upload ${hf_user}/${repo_name} path/to/checkpoint/dir
+huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
 See [eval.py](https://github.com/huggingface/lerobot/blob/main/lerobot/scripts/eval.py) for an example of how other people may use your policy.

docker/lerobot-gpu-dev/Dockerfile

@@ -0,0 +1,40 @@
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04
+
+# Configure image
+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 \
+    git git-lfs openssh-client \
+    nano vim less util-linux \
+    htop atop nvtop \
+    sed gawk grep curl wget \
+    tcpdump sysstat screen tmux \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Install gh cli tool
+RUN (type -p wget >/dev/null || (apt update && apt-get install wget -y)) \
+    && mkdir -p -m 755 /etc/apt/keyrings \
+    && wget -qO- https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg > /dev/null \
+    && chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg \
+    && echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" | tee /etc/apt/sources.list.d/github-cli.list > /dev/null \
+    && apt update \
+    && apt install gh -y \
+    && apt clean && rm -rf /var/lib/apt/lists/*
+
+# Setup `python`
+RUN ln -s /usr/bin/python3 /usr/bin/python
+
+# Install poetry
+RUN curl -sSL https://install.python-poetry.org | python -
+ENV PATH="/root/.local/bin:$PATH"
+RUN echo 'if [ "$HOME" != "/root" ]; then ln -sf /root/.local/bin/poetry $HOME/.local/bin/poetry; fi' >> /root/.bashrc
+RUN poetry config virtualenvs.create false
+RUN poetry config virtualenvs.in-project true
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"

docker/lerobot-gpu/Dockerfile

@@ -4,6 +4,7 @@ FROM nvidia/cuda:12.4.1-base-ubuntu22.04
 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 \
@@ -11,6 +12,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     python${PYTHON_VERSION} python${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

examples/4_train_policy_with_script.md

@@ -0,0 +1,183 @@
+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.
+
+## 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:
+
+- 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.
+- 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
+
+```python
+python lerobot/scripts/train.py
+```
+
+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:
+
+```yaml
+defaults:
+  - _self_
+  - env: pusht
+  - policy: diffusion
+```
+
+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 overriden by `env/pusht.yaml`  which is overidden by `policy/diffusion.yaml`_.
+
+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`.
+
+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:
+
+```bash
+python lerobot/scripts/train.py \
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0
+```
+
+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.
+
+```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 \
+```
+
+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:
+
+```bash
+python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
+```
+
+Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
+
+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:
+
+```bash
+python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
+```
+
+Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
+
+---
+
+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):
+
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+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.
+
+Or in the meantime, happy coding! 🤗

examples/5_resume_training.md

@@ -0,0 +1,37 @@
+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! 🤗

examples/advanced/1_train_act_pusht/act_pusht.yaml

@@ -0,0 +1,87 @@
+# @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_momentum: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

examples/advanced/1_train_act_pusht/train_act_pusht.md

@@ -0,0 +1,70 @@
+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! 🤗

examples/advanced/2_calculate_validation_loss.py

@@ -0,0 +1,90 @@
+"""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
+is learning effectively.
+
+Furthermore, relying on validation loss to evaluate performance is generally not considered a good practice,
+especially in the context of imitation learning. The most reliable approach is to evaluate the policy directly
+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
+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")
+
+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],
+    "observation.state": [-0.1, 0.0],
+    # Load the previous action (-0.1), the next action to be executed (0.0),
+    # 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 full dataset
+full_dataset = LeRobotDataset("lerobot/pusht", split="train")
+# - Calculate train and val subsets
+num_train_episodes = math.floor(full_dataset.num_episodes * 90 / 100)
+num_val_episodes = full_dataset.num_episodes - num_train_episodes
+print(f"Number of episodes in full dataset: {full_dataset.num_episodes}")
+print(f"Number of episodes in training dataset (90% subset): {num_train_episodes}")
+print(f"Number of episodes in validation dataset (10% subset): {num_val_episodes}")
+# - Get first frame index of the validation set
+first_val_frame_index = full_dataset.episode_data_index["from"][num_train_episodes].item()
+# - Load frames subset belonging to validation set using the `split` argument.
+#   It utilizes the `datasets` library's syntax for slicing datasets.
+#   For more information on the Slice API, please see:
+#   https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+train_dataset = LeRobotDataset(
+    "lerobot/pusht", split=f"train[:{first_val_frame_index}]", delta_timestamps=delta_timestamps
+)
+val_dataset = LeRobotDataset(
+    "lerobot/pusht", split=f"train[{first_val_frame_index}:]", 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(
+    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:
+    batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+    output_dict = policy.forward(batch)
+
+    loss_cumsum += output_dict["loss"].item()
+    n_examples_evaluated += batch["index"].shape[0]
+
+# 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}")

examples/real_robot_example/README.md

@@ -0,0 +1,89 @@
+# Using `lerobot`  on a real world arm
+
+
+In this example, we'll be using `lerobot` on a real world arm to:
+- record a dataset in the `lerobot` format
+- (soon) train a policy on it
+- (soon) run the policy in the real-world
+
+## Which robotic arm to use
+
+In this example we're using the [open-source low-cost arm from Alexander Koch](https://github.com/AlexanderKoch-Koch/low_cost_robot) in the specific setup of:
+- having 6 servos per arm, i.e. using the elbow-to-wrist extension
+- adding two cameras around it, one on top and one in the front
+- having a teleoperation arm as well (build the leader and the follower arms in A. Koch repo, both with elbow-to-wrist extensions)
+
+I'm using these cameras (but the setup should not be sensitive to the exact cameras you're using):
+- C922 Pro Stream Webcam
+- Intel(R) RealSense D455 (using only the RGB input)
+
+
+In general, this example should be very easily extendable to any type of arm using Dynamixel servos with at least one camera by changing a couple of configuration in the gym env.
+
+## Install the example
+
+Follow these steps:
+- install `lerobot`
+- install the Dynamixel-sdk: `pip install dynamixel-sdk`
+
+## Usage
+
+### 0 - record examples
+
+Run the `record_training_data.py` example, selecting the duration and number of episodes you want to record, e.g.
+```
+DATA_DIR='./data' python record_training_data.py \
+--repo-id=thomwolf/blue_red_sort \
+--num-episodes=50 \
+--num-frames=400
+```
+
+TODO:
+- various length episodes
+- being able to drop episodes
+- checking uploading to the hub
+
+### 1 - visualize the dataset
+
+Use the standard dataset visualization script pointing it to the right folder:
+```
+DATA_DIR='./data' python ../../lerobot/scripts/visualize_dataset.py \
+    --repo-id thomwolf/blue_red_sort \
+    --episode-index 0
+```
+
+### 2 - Train a policy
+
+From the example directory let's run this command to train a model using ACT
+
+```
+DATA_DIR='./data' python ../../lerobot/scripts/train.py \
+    device=cuda \
+    hydra.searchpath=[file://./train_config/] \
+    hydra.run.dir=./outputs/train/blue_red_sort \
+    dataset_repo_id=thomwolf/blue_red_sort \
+    env=gym_real_world \
+    policy=act_real_world \
+    wandb.enable=false
+```
+
+### 3 - Evaluate the policy in the real world
+
+From the example directory let's run this command to evaluate our policy.
+The configuration for running the policy is in the checkpoint of the model.
+You can override parameters as follow:
+
+```
+python run_policy.py \
+    -p ./outputs/train/blue_red_sort/checkpoints/last/pretrained_model/
+    env.episode_length=1000
+```
+
+
+## Convert a hdf5 dataset recorded with the original ACT repo
+
+You can convert a dataset from the raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act with the following command:
+
+```
+python ./lerobot/scripts/push_dataset_to_hub.py
+```

examples/real_robot_example/convert_original_act_checkpoint.ipynb

@@ -0,0 +1,840 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 48,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from safetensors.torch import load_file, save_file\n",
+    "from pprint import pprint"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 52,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "original_ckpt_path = \"/home/thomwolf/Documents/Github/ACT/checkpoints/blue_red_sort/policy_last.ckpt\"\n",
+    "converted_ckpt_path = \"/home/thomwolf/Documents/Github/ACT/checkpoints/blue_red_sort/model.safetensors\"\n",
+    "\n",
+    "comparison_main_path = \"/home/thomwolf/Documents/Github/lerobot/examples/real_robot_example/outputs/train/blue_red_debug_no_masking/checkpoints/last/pretrained_model/\"\n",
+    "comparison_safetensor_path = comparison_main_path + \"model.safetensors\"\n",
+    "comparison_config_json_path = comparison_main_path + \"config.json\"\n",
+    "comparison_config_yaml_path = comparison_main_path + \"config.yaml\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 28,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "a = torch.load(original_ckpt_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "b = load_file(comparison_safetensor_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 17,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "['model.action_head.bias',\n",
+      " 'model.action_head.weight',\n",
+      " 'model.backbone.bn1.bias',\n",
+      " 'model.backbone.bn1.running_mean',\n",
+      " 'model.backbone.bn1.running_var',\n",
+      " 'model.backbone.bn1.weight',\n",
+      " 'model.backbone.conv1.weight',\n",
+      " 'model.backbone.layer1.0.bn1.bias',\n",
+      " 'model.backbone.layer1.0.bn1.running_mean',\n",
+      " 'model.backbone.layer1.0.bn1.running_var',\n",
+      " 'model.backbone.layer1.0.bn1.weight',\n",
+      " 'model.backbone.layer1.0.bn2.bias',\n",
+      " 'model.backbone.layer1.0.bn2.running_mean',\n",
+      " 'model.backbone.layer1.0.bn2.running_var',\n",
+      " 'model.backbone.layer1.0.bn2.weight',\n",
+      " 'model.backbone.layer1.0.conv1.weight',\n",
+      " 'model.backbone.layer1.0.conv2.weight',\n",
+      " 'model.backbone.layer1.1.bn1.bias',\n",
+      " 'model.backbone.layer1.1.bn1.running_mean',\n",
+      " 'model.backbone.layer1.1.bn1.running_var',\n",
+      " 'model.backbone.layer1.1.bn1.weight',\n",
+      " 'model.backbone.layer1.1.bn2.bias',\n",
+      " 'model.backbone.layer1.1.bn2.running_mean',\n",
+      " 'model.backbone.layer1.1.bn2.running_var',\n",
+      " 'model.backbone.layer1.1.bn2.weight',\n",
+      " 'model.backbone.layer1.1.conv1.weight',\n",
+      " 'model.backbone.layer1.1.conv2.weight',\n",
+      " 'model.backbone.layer2.0.bn1.bias',\n",
+      " 'model.backbone.layer2.0.bn1.running_mean',\n",
+      " 'model.backbone.layer2.0.bn1.running_var',\n",
+      " 'model.backbone.layer2.0.bn1.weight',\n",
+      " 'model.backbone.layer2.0.bn2.bias',\n",
+      " 'model.backbone.layer2.0.bn2.running_mean',\n",
+      " 'model.backbone.layer2.0.bn2.running_var',\n",
+      " 'model.backbone.layer2.0.bn2.weight',\n",
+      " 'model.backbone.layer2.0.conv1.weight',\n",
+      " 'model.backbone.layer2.0.conv2.weight',\n",
+      " 'model.backbone.layer2.0.downsample.0.weight',\n",
+      " 'model.backbone.layer2.0.downsample.1.bias',\n",
+      " 'model.backbone.layer2.0.downsample.1.running_mean',\n",
+      " 'model.backbone.layer2.0.downsample.1.running_var',\n",
+      " 'model.backbone.layer2.0.downsample.1.weight',\n",
+      " 'model.backbone.layer2.1.bn1.bias',\n",
+      " 'model.backbone.layer2.1.bn1.running_mean',\n",
+      " 'model.backbone.layer2.1.bn1.running_var',\n",
+      " 'model.backbone.layer2.1.bn1.weight',\n",
+      " 'model.backbone.layer2.1.bn2.bias',\n",
+      " 'model.backbone.layer2.1.bn2.running_mean',\n",
+      " 'model.backbone.layer2.1.bn2.running_var',\n",
+      " 'model.backbone.layer2.1.bn2.weight',\n",
+      " 'model.backbone.layer2.1.conv1.weight',\n",
+      " 'model.backbone.layer2.1.conv2.weight',\n",
+      " 'model.backbone.layer3.0.bn1.bias',\n",
+      " 'model.backbone.layer3.0.bn1.running_mean',\n",
+      " 'model.backbone.layer3.0.bn1.running_var',\n",
+      " 'model.backbone.layer3.0.bn1.weight',\n",
+      " 'model.backbone.layer3.0.bn2.bias',\n",
+      " 'model.backbone.layer3.0.bn2.running_mean',\n",
+      " 'model.backbone.layer3.0.bn2.running_var',\n",
+      " 'model.backbone.layer3.0.bn2.weight',\n",
+      " 'model.backbone.layer3.0.conv1.weight',\n",
+      " 'model.backbone.layer3.0.conv2.weight',\n",
+      " 'model.backbone.layer3.0.downsample.0.weight',\n",
+      " 'model.backbone.layer3.0.downsample.1.bias',\n",
+      " 'model.backbone.layer3.0.downsample.1.running_mean',\n",
+      " 'model.backbone.layer3.0.downsample.1.running_var',\n",
+      " 'model.backbone.layer3.0.downsample.1.weight',\n",
+      " 'model.backbone.layer3.1.bn1.bias',\n",
+      " 'model.backbone.layer3.1.bn1.running_mean',\n",
+      " 'model.backbone.layer3.1.bn1.running_var',\n",
+      " 'model.backbone.layer3.1.bn1.weight',\n",
+      " 'model.backbone.layer3.1.bn2.bias',\n",
+      " 'model.backbone.layer3.1.bn2.running_mean',\n",
+      " 'model.backbone.layer3.1.bn2.running_var',\n",
+      " 'model.backbone.layer3.1.bn2.weight',\n",
+      " 'model.backbone.layer3.1.conv1.weight',\n",
+      " 'model.backbone.layer3.1.conv2.weight',\n",
+      " 'model.backbone.layer4.0.bn1.bias',\n",
+      " 'model.backbone.layer4.0.bn1.running_mean',\n",
+      " 'model.backbone.layer4.0.bn1.running_var',\n",
+      " 'model.backbone.layer4.0.bn1.weight',\n",
+      " 'model.backbone.layer4.0.bn2.bias',\n",
+      " 'model.backbone.layer4.0.bn2.running_mean',\n",
+      " 'model.backbone.layer4.0.bn2.running_var',\n",
+      " 'model.backbone.layer4.0.bn2.weight',\n",
+      " 'model.backbone.layer4.0.conv1.weight',\n",
+      " 'model.backbone.layer4.0.conv2.weight',\n",
+      " 'model.backbone.layer4.0.downsample.0.weight',\n",
+      " 'model.backbone.layer4.0.downsample.1.bias',\n",
+      " 'model.backbone.layer4.0.downsample.1.running_mean',\n",
+      " 'model.backbone.layer4.0.downsample.1.running_var',\n",
+      " 'model.backbone.layer4.0.downsample.1.weight',\n",
+      " 'model.backbone.layer4.1.bn1.bias',\n",
+      " 'model.backbone.layer4.1.bn1.running_mean',\n",
+      " 'model.backbone.layer4.1.bn1.running_var',\n",
+      " 'model.backbone.layer4.1.bn1.weight',\n",
+      " 'model.backbone.layer4.1.bn2.bias',\n",
+      " 'model.backbone.layer4.1.bn2.running_mean',\n",
+      " 'model.backbone.layer4.1.bn2.running_var',\n",
+      " 'model.backbone.layer4.1.bn2.weight',\n",
+      " 'model.backbone.layer4.1.conv1.weight',\n",
+      " 'model.backbone.layer4.1.conv2.weight',\n",
+      " 'model.decoder.layers.0.linear1.bias',\n",
+      " 'model.decoder.layers.0.linear1.weight',\n",
+      " 'model.decoder.layers.0.linear2.bias',\n",
+      " 'model.decoder.layers.0.linear2.weight',\n",
+      " 'model.decoder.layers.0.multihead_attn.in_proj_bias',\n",
+      " 'model.decoder.layers.0.multihead_attn.in_proj_weight',\n",
+      " 'model.decoder.layers.0.multihead_attn.out_proj.bias',\n",
+      " 'model.decoder.layers.0.multihead_attn.out_proj.weight',\n",
+      " 'model.decoder.layers.0.norm1.bias',\n",
+      " 'model.decoder.layers.0.norm1.weight',\n",
+      " 'model.decoder.layers.0.norm2.bias',\n",
+      " 'model.decoder.layers.0.norm2.weight',\n",
+      " 'model.decoder.layers.0.norm3.bias',\n",
+      " 'model.decoder.layers.0.norm3.weight',\n",
+      " 'model.decoder.layers.0.self_attn.in_proj_bias',\n",
+      " 'model.decoder.layers.0.self_attn.in_proj_weight',\n",
+      " 'model.decoder.layers.0.self_attn.out_proj.bias',\n",
+      " 'model.decoder.layers.0.self_attn.out_proj.weight',\n",
+      " 'model.decoder_pos_embed.weight',\n",
+      " 'model.encoder.layers.0.linear1.bias',\n",
+      " 'model.encoder.layers.0.linear1.weight',\n",
+      " 'model.encoder.layers.0.linear2.bias',\n",
+      " 'model.encoder.layers.0.linear2.weight',\n",
+      " 'model.encoder.layers.0.norm1.bias',\n",
+      " 'model.encoder.layers.0.norm1.weight',\n",
+      " 'model.encoder.layers.0.norm2.bias',\n",
+      " 'model.encoder.layers.0.norm2.weight',\n",
+      " 'model.encoder.layers.0.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.0.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.0.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.0.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.1.linear1.bias',\n",
+      " 'model.encoder.layers.1.linear1.weight',\n",
+      " 'model.encoder.layers.1.linear2.bias',\n",
+      " 'model.encoder.layers.1.linear2.weight',\n",
+      " 'model.encoder.layers.1.norm1.bias',\n",
+      " 'model.encoder.layers.1.norm1.weight',\n",
+      " 'model.encoder.layers.1.norm2.bias',\n",
+      " 'model.encoder.layers.1.norm2.weight',\n",
+      " 'model.encoder.layers.1.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.1.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.1.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.1.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.2.linear1.bias',\n",
+      " 'model.encoder.layers.2.linear1.weight',\n",
+      " 'model.encoder.layers.2.linear2.bias',\n",
+      " 'model.encoder.layers.2.linear2.weight',\n",
+      " 'model.encoder.layers.2.norm1.bias',\n",
+      " 'model.encoder.layers.2.norm1.weight',\n",
+      " 'model.encoder.layers.2.norm2.bias',\n",
+      " 'model.encoder.layers.2.norm2.weight',\n",
+      " 'model.encoder.layers.2.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.2.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.2.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.2.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.3.linear1.bias',\n",
+      " 'model.encoder.layers.3.linear1.weight',\n",
+      " 'model.encoder.layers.3.linear2.bias',\n",
+      " 'model.encoder.layers.3.linear2.weight',\n",
+      " 'model.encoder.layers.3.norm1.bias',\n",
+      " 'model.encoder.layers.3.norm1.weight',\n",
+      " 'model.encoder.layers.3.norm2.bias',\n",
+      " 'model.encoder.layers.3.norm2.weight',\n",
+      " 'model.encoder.layers.3.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.3.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.3.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.3.self_attn.out_proj.weight',\n",
+      " 'model.encoder_img_feat_input_proj.bias',\n",
+      " 'model.encoder_img_feat_input_proj.weight',\n",
+      " 'model.encoder_latent_input_proj.bias',\n",
+      " 'model.encoder_latent_input_proj.weight',\n",
+      " 'model.encoder_robot_and_latent_pos_embed.weight',\n",
+      " 'model.encoder_robot_state_input_proj.bias',\n",
+      " 'model.encoder_robot_state_input_proj.weight',\n",
+      " 'model.vae_encoder.layers.0.linear1.bias',\n",
+      " 'model.vae_encoder.layers.0.linear1.weight',\n",
+      " 'model.vae_encoder.layers.0.linear2.bias',\n",
+      " 'model.vae_encoder.layers.0.linear2.weight',\n",
+      " 'model.vae_encoder.layers.0.norm1.bias',\n",
+      " 'model.vae_encoder.layers.0.norm1.weight',\n",
+      " 'model.vae_encoder.layers.0.norm2.bias',\n",
+      " 'model.vae_encoder.layers.0.norm2.weight',\n",
+      " 'model.vae_encoder.layers.0.self_attn.in_proj_bias',\n",
+      " 'model.vae_encoder.layers.0.self_attn.in_proj_weight',\n",
+      " 'model.vae_encoder.layers.0.self_attn.out_proj.bias',\n",
+      " 'model.vae_encoder.layers.0.self_attn.out_proj.weight',\n",
+      " 'model.vae_encoder.layers.1.linear1.bias',\n",
+      " 'model.vae_encoder.layers.1.linear1.weight',\n",
+      " 'model.vae_encoder.layers.1.linear2.bias',\n",
+      " 'model.vae_encoder.layers.1.linear2.weight',\n",
+      " 'model.vae_encoder.layers.1.norm1.bias',\n",
+      " 'model.vae_encoder.layers.1.norm1.weight',\n",
+      " 'model.vae_encoder.layers.1.norm2.bias',\n",
+      " 'model.vae_encoder.layers.1.norm2.weight',\n",
+      " 'model.vae_encoder.layers.1.self_attn.in_proj_bias',\n",
+      " 'model.vae_encoder.layers.1.self_attn.in_proj_weight',\n",
+      " 'model.vae_encoder.layers.1.self_attn.out_proj.bias',\n",
+      " 'model.vae_encoder.layers.1.self_attn.out_proj.weight',\n",
+      " 'model.vae_encoder.layers.2.linear1.bias',\n",
+      " 'model.vae_encoder.layers.2.linear1.weight',\n",
+      " 'model.vae_encoder.layers.2.linear2.bias',\n",
+      " 'model.vae_encoder.layers.2.linear2.weight',\n",
+      " 'model.vae_encoder.layers.2.norm1.bias',\n",
+      " 'model.vae_encoder.layers.2.norm1.weight',\n",
+      " 'model.vae_encoder.layers.2.norm2.bias',\n",
+      " 'model.vae_encoder.layers.2.norm2.weight',\n",
+      " 'model.vae_encoder.layers.2.self_attn.in_proj_bias',\n",
+      " 'model.vae_encoder.layers.2.self_attn.in_proj_weight',\n",
+      " 'model.vae_encoder.layers.2.self_attn.out_proj.bias',\n",
+      " 'model.vae_encoder.layers.2.self_attn.out_proj.weight',\n",
+      " 'model.vae_encoder.layers.3.linear1.bias',\n",
+      " 'model.vae_encoder.layers.3.linear1.weight',\n",
+      " 'model.vae_encoder.layers.3.linear2.bias',\n",
+      " 'model.vae_encoder.layers.3.linear2.weight',\n",
+      " 'model.vae_encoder.layers.3.norm1.bias',\n",
+      " 'model.vae_encoder.layers.3.norm1.weight',\n",
+      " 'model.vae_encoder.layers.3.norm2.bias',\n",
+      " 'model.vae_encoder.layers.3.norm2.weight',\n",
+      " 'model.vae_encoder.layers.3.self_attn.in_proj_bias',\n",
+      " 'model.vae_encoder.layers.3.self_attn.in_proj_weight',\n",
+      " 'model.vae_encoder.layers.3.self_attn.out_proj.bias',\n",
+      " 'model.vae_encoder.layers.3.self_attn.out_proj.weight',\n",
+      " 'model.vae_encoder_action_input_proj.bias',\n",
+      " 'model.vae_encoder_action_input_proj.weight',\n",
+      " 'model.vae_encoder_cls_embed.weight',\n",
+      " 'model.vae_encoder_latent_output_proj.bias',\n",
+      " 'model.vae_encoder_latent_output_proj.weight',\n",
+      " 'model.vae_encoder_pos_enc',\n",
+      " 'model.vae_encoder_robot_state_input_proj.bias',\n",
+      " 'model.vae_encoder_robot_state_input_proj.weight',\n",
+      " 'normalize_inputs.buffer_observation_images_front.mean',\n",
+      " 'normalize_inputs.buffer_observation_images_front.std',\n",
+      " 'normalize_inputs.buffer_observation_images_top.mean',\n",
+      " 'normalize_inputs.buffer_observation_images_top.std',\n",
+      " 'normalize_inputs.buffer_observation_state.mean',\n",
+      " 'normalize_inputs.buffer_observation_state.std',\n",
+      " 'normalize_targets.buffer_action.mean',\n",
+      " 'normalize_targets.buffer_action.std',\n",
+      " 'unnormalize_outputs.buffer_action.mean',\n",
+      " 'unnormalize_outputs.buffer_action.std']\n"
+     ]
+    }
+   ],
+   "source": [
+    "dest = list(b.keys())\n",
+    "pprint(dest)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 29,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "['model.pos_table',\n",
+      " 'model.transformer.encoder.layers.0.self_attn.in_proj_weight',\n",
+      " 'model.transformer.encoder.layers.0.self_attn.in_proj_bias',\n",
+      " 'model.transformer.encoder.layers.0.self_attn.out_proj.weight',\n",
+      " 'model.transformer.encoder.layers.0.self_attn.out_proj.bias',\n",
+      " 'model.transformer.encoder.layers.0.linear1.weight',\n",
+      " 'model.transformer.encoder.layers.0.linear1.bias',\n",
+      " 'model.transformer.encoder.layers.0.linear2.weight',\n",
+      " 'model.transformer.encoder.layers.0.linear2.bias',\n",
+      " 'model.transformer.encoder.layers.0.norm1.weight',\n",
+      " 'model.transformer.encoder.layers.0.norm1.bias',\n",
+      " 'model.transformer.encoder.layers.0.norm2.weight',\n",
+      " 'model.transformer.encoder.layers.0.norm2.bias',\n",
+      " 'model.transformer.encoder.layers.1.self_attn.in_proj_weight',\n",
+      " 'model.transformer.encoder.layers.1.self_attn.in_proj_bias',\n",
+      " 'model.transformer.encoder.layers.1.self_attn.out_proj.weight',\n",
+      " 'model.transformer.encoder.layers.1.self_attn.out_proj.bias',\n",
+      " 'model.transformer.encoder.layers.1.linear1.weight',\n",
+      " 'model.transformer.encoder.layers.1.linear1.bias',\n",
+      " 'model.transformer.encoder.layers.1.linear2.weight',\n",
+      " 'model.transformer.encoder.layers.1.linear2.bias',\n",
+      " 'model.transformer.encoder.layers.1.norm1.weight',\n",
+      " 'model.transformer.encoder.layers.1.norm1.bias',\n",
+      " 'model.transformer.encoder.layers.1.norm2.weight',\n",
+      " 'model.transformer.encoder.layers.1.norm2.bias',\n",
+      " 'model.transformer.encoder.layers.2.self_attn.in_proj_weight',\n",
+      " 'model.transformer.encoder.layers.2.self_attn.in_proj_bias',\n",
+      " 'model.transformer.encoder.layers.2.self_attn.out_proj.weight',\n",
+      " 'model.transformer.encoder.layers.2.self_attn.out_proj.bias',\n",
+      " 'model.transformer.encoder.layers.2.linear1.weight',\n",
+      " 'model.transformer.encoder.layers.2.linear1.bias',\n",
+      " 'model.transformer.encoder.layers.2.linear2.weight',\n",
+      " 'model.transformer.encoder.layers.2.linear2.bias',\n",
+      " 'model.transformer.encoder.layers.2.norm1.weight',\n",
+      " 'model.transformer.encoder.layers.2.norm1.bias',\n",
+      " 'model.transformer.encoder.layers.2.norm2.weight',\n",
+      " 'model.transformer.encoder.layers.2.norm2.bias',\n",
+      " 'model.transformer.encoder.layers.3.self_attn.in_proj_weight',\n",
+      " 'model.transformer.encoder.layers.3.self_attn.in_proj_bias',\n",
+      " 'model.transformer.encoder.layers.3.self_attn.out_proj.weight',\n",
+      " 'model.transformer.encoder.layers.3.self_attn.out_proj.bias',\n",
+      " 'model.transformer.encoder.layers.3.linear1.weight',\n",
+      " 'model.transformer.encoder.layers.3.linear1.bias',\n",
+      " 'model.transformer.encoder.layers.3.linear2.weight',\n",
+      " 'model.transformer.encoder.layers.3.linear2.bias',\n",
+      " 'model.transformer.encoder.layers.3.norm1.weight',\n",
+      " 'model.transformer.encoder.layers.3.norm1.bias',\n",
+      " 'model.transformer.encoder.layers.3.norm2.weight',\n",
+      " 'model.transformer.encoder.layers.3.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.0.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.0.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.0.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.0.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.0.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.0.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.0.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.0.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.0.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.0.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.0.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.0.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.0.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.0.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.0.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.0.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.0.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.0.norm3.bias',\n",
+      " 'model.transformer.decoder.layers.1.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.1.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.1.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.1.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.1.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.1.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.1.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.1.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.1.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.1.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.1.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.1.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.1.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.1.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.1.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.1.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.1.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.1.norm3.bias',\n",
+      " 'model.transformer.decoder.layers.2.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.2.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.2.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.2.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.2.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.2.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.2.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.2.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.2.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.2.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.2.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.2.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.2.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.2.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.2.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.2.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.2.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.2.norm3.bias',\n",
+      " 'model.transformer.decoder.layers.3.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.3.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.3.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.3.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.3.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.3.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.3.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.3.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.3.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.3.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.3.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.3.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.3.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.3.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.3.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.3.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.3.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.3.norm3.bias',\n",
+      " 'model.transformer.decoder.layers.4.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.4.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.4.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.4.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.4.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.4.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.4.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.4.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.4.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.4.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.4.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.4.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.4.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.4.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.4.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.4.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.4.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.4.norm3.bias',\n",
+      " 'model.transformer.decoder.layers.5.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.5.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.5.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.5.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.5.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.5.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.5.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.5.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.5.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.5.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.5.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.5.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.5.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.5.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.5.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.5.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.5.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.5.norm3.bias',\n",
+      " 'model.transformer.decoder.layers.6.self_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.6.self_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.6.self_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.6.self_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.6.multihead_attn.in_proj_weight',\n",
+      " 'model.transformer.decoder.layers.6.multihead_attn.in_proj_bias',\n",
+      " 'model.transformer.decoder.layers.6.multihead_attn.out_proj.weight',\n",
+      " 'model.transformer.decoder.layers.6.multihead_attn.out_proj.bias',\n",
+      " 'model.transformer.decoder.layers.6.linear1.weight',\n",
+      " 'model.transformer.decoder.layers.6.linear1.bias',\n",
+      " 'model.transformer.decoder.layers.6.linear2.weight',\n",
+      " 'model.transformer.decoder.layers.6.linear2.bias',\n",
+      " 'model.transformer.decoder.layers.6.norm1.weight',\n",
+      " 'model.transformer.decoder.layers.6.norm1.bias',\n",
+      " 'model.transformer.decoder.layers.6.norm2.weight',\n",
+      " 'model.transformer.decoder.layers.6.norm2.bias',\n",
+      " 'model.transformer.decoder.layers.6.norm3.weight',\n",
+      " 'model.transformer.decoder.layers.6.norm3.bias',\n",
+      " 'model.transformer.decoder.norm.weight',\n",
+      " 'model.transformer.decoder.norm.bias',\n",
+      " 'model.encoder.layers.0.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.0.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.0.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.0.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.0.linear1.weight',\n",
+      " 'model.encoder.layers.0.linear1.bias',\n",
+      " 'model.encoder.layers.0.linear2.weight',\n",
+      " 'model.encoder.layers.0.linear2.bias',\n",
+      " 'model.encoder.layers.0.norm1.weight',\n",
+      " 'model.encoder.layers.0.norm1.bias',\n",
+      " 'model.encoder.layers.0.norm2.weight',\n",
+      " 'model.encoder.layers.0.norm2.bias',\n",
+      " 'model.encoder.layers.1.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.1.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.1.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.1.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.1.linear1.weight',\n",
+      " 'model.encoder.layers.1.linear1.bias',\n",
+      " 'model.encoder.layers.1.linear2.weight',\n",
+      " 'model.encoder.layers.1.linear2.bias',\n",
+      " 'model.encoder.layers.1.norm1.weight',\n",
+      " 'model.encoder.layers.1.norm1.bias',\n",
+      " 'model.encoder.layers.1.norm2.weight',\n",
+      " 'model.encoder.layers.1.norm2.bias',\n",
+      " 'model.encoder.layers.2.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.2.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.2.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.2.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.2.linear1.weight',\n",
+      " 'model.encoder.layers.2.linear1.bias',\n",
+      " 'model.encoder.layers.2.linear2.weight',\n",
+      " 'model.encoder.layers.2.linear2.bias',\n",
+      " 'model.encoder.layers.2.norm1.weight',\n",
+      " 'model.encoder.layers.2.norm1.bias',\n",
+      " 'model.encoder.layers.2.norm2.weight',\n",
+      " 'model.encoder.layers.2.norm2.bias',\n",
+      " 'model.encoder.layers.3.self_attn.in_proj_weight',\n",
+      " 'model.encoder.layers.3.self_attn.in_proj_bias',\n",
+      " 'model.encoder.layers.3.self_attn.out_proj.weight',\n",
+      " 'model.encoder.layers.3.self_attn.out_proj.bias',\n",
+      " 'model.encoder.layers.3.linear1.weight',\n",
+      " 'model.encoder.layers.3.linear1.bias',\n",
+      " 'model.encoder.layers.3.linear2.weight',\n",
+      " 'model.encoder.layers.3.linear2.bias',\n",
+      " 'model.encoder.layers.3.norm1.weight',\n",
+      " 'model.encoder.layers.3.norm1.bias',\n",
+      " 'model.encoder.layers.3.norm2.weight',\n",
+      " 'model.encoder.layers.3.norm2.bias',\n",
+      " 'model.action_head.weight',\n",
+      " 'model.action_head.bias',\n",
+      " 'model.is_pad_head.weight',\n",
+      " 'model.is_pad_head.bias',\n",
+      " 'model.query_embed.weight',\n",
+      " 'model.input_proj.weight',\n",
+      " 'model.input_proj.bias',\n",
+      " 'model.backbones.0.0.body.conv1.weight',\n",
+      " 'model.backbones.0.0.body.bn1.weight',\n",
+      " 'model.backbones.0.0.body.bn1.bias',\n",
+      " 'model.backbones.0.0.body.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer1.0.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer1.0.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer1.0.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer1.1.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer1.1.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer1.1.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer2.0.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer2.0.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer2.0.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer2.0.downsample.0.weight',\n",
+      " 'model.backbones.0.0.body.layer2.0.downsample.1.weight',\n",
+      " 'model.backbones.0.0.body.layer2.0.downsample.1.bias',\n",
+      " 'model.backbones.0.0.body.layer2.0.downsample.1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer2.0.downsample.1.running_var',\n",
+      " 'model.backbones.0.0.body.layer2.0.downsample.1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer2.1.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer2.1.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer2.1.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer3.0.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer3.0.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer3.0.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer3.0.downsample.0.weight',\n",
+      " 'model.backbones.0.0.body.layer3.0.downsample.1.weight',\n",
+      " 'model.backbones.0.0.body.layer3.0.downsample.1.bias',\n",
+      " 'model.backbones.0.0.body.layer3.0.downsample.1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer3.0.downsample.1.running_var',\n",
+      " 'model.backbones.0.0.body.layer3.0.downsample.1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer3.1.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer3.1.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer3.1.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer4.0.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer4.0.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer4.0.bn2.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer4.0.downsample.0.weight',\n",
+      " 'model.backbones.0.0.body.layer4.0.downsample.1.weight',\n",
+      " 'model.backbones.0.0.body.layer4.0.downsample.1.bias',\n",
+      " 'model.backbones.0.0.body.layer4.0.downsample.1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer4.0.downsample.1.running_var',\n",
+      " 'model.backbones.0.0.body.layer4.0.downsample.1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer4.1.conv1.weight',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn1.weight',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn1.bias',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn1.running_mean',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn1.running_var',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn1.num_batches_tracked',\n",
+      " 'model.backbones.0.0.body.layer4.1.conv2.weight',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn2.weight',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn2.bias',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn2.running_mean',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn2.running_var',\n",
+      " 'model.backbones.0.0.body.layer4.1.bn2.num_batches_tracked',\n",
+      " 'model.input_proj_robot_state.weight',\n",
+      " 'model.input_proj_robot_state.bias',\n",
+      " 'model.cls_embed.weight',\n",
+      " 'model.encoder_action_proj.weight',\n",
+      " 'model.encoder_action_proj.bias',\n",
+      " 'model.encoder_joint_proj.weight',\n",
+      " 'model.encoder_joint_proj.bias',\n",
+      " 'model.latent_proj.weight',\n",
+      " 'model.latent_proj.bias',\n",
+      " 'model.latent_out_proj.weight',\n",
+      " 'model.latent_out_proj.bias',\n",
+      " 'model.additional_pos_embed.weight']\n"
+     ]
+    }
+   ],
+   "source": [
+    "orig = list(a.keys())\n",
+    "pprint(orig)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 45,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "a = torch.load(original_ckpt_path)\n",
+    "\n",
+    "to_remove_startswith = ['model.transformer.decoder.layers.1.',\n",
+    "             'model.transformer.decoder.layers.2.',\n",
+    "             'model.transformer.decoder.layers.3.',\n",
+    "             'model.transformer.decoder.layers.4.',\n",
+    "             'model.transformer.decoder.layers.5.',\n",
+    "             'model.transformer.decoder.layers.6.',\n",
+    "             'model.transformer.decoder.norm.',\n",
+    "             'model.is_pad_head']\n",
+    "\n",
+    "to_remove_in = ['num_batches_tracked',]\n",
+    "\n",
+    "conv = {}\n",
+    "\n",
+    "keys = list(a.keys())\n",
+    "for k in keys:\n",
+    "    if any(k.startswith(tr) for tr in to_remove_startswith):\n",
+    "        a.pop(k)\n",
+    "        continue\n",
+    "    if any(tr in k for tr in to_remove_in):\n",
+    "        a.pop(k)\n",
+    "        continue\n",
+    "    if k.startswith('model.transformer.encoder.layers.'):\n",
+    "        conv[k.replace('transformer.', '')] = a.pop(k)\n",
+    "    if k.startswith('model.transformer.decoder.layers.0.'):\n",
+    "        conv[k.replace('transformer.', '')] = a.pop(k)\n",
+    "    if k.startswith('model.encoder.layers.'):\n",
+    "        conv[k.replace('encoder.', 'vae_encoder.')] = a.pop(k)\n",
+    "    if k.startswith('model.action_head.'):\n",
+    "        conv[k] = a.pop(k)\n",
+    "    if k.startswith('model.pos_table'):\n",
+    "        conv[k.replace('pos_table', 'vae_encoder_pos_enc')] = a.pop(k)\n",
+    "    if k.startswith('model.query_embed.'):\n",
+    "        conv[k.replace('query_embed', 'decoder_pos_embed')] = a.pop(k)\n",
+    "    if k.startswith('model.input_proj.'):\n",
+    "        conv[k.replace('input_proj.', 'encoder_img_feat_input_proj.')] = a.pop(k)\n",
+    "    if k.startswith('model.input_proj_robot_state.'):\n",
+    "        conv[k.replace('input_proj_robot_state.', 'encoder_robot_state_input_proj.')] = a.pop(k)\n",
+    "    if k.startswith('model.backbones.0.0.body.'):\n",
+    "        conv[k.replace('backbones.0.0.body', 'backbone')] = a.pop(k)\n",
+    "    if k.startswith('model.cls_embed.'):\n",
+    "        conv[k.replace('cls_embed', 'vae_encoder_cls_embed')] = a.pop(k)\n",
+    "    if k.startswith('model.encoder_action_proj.'):\n",
+    "        conv[k.replace('encoder_action_proj', 'vae_encoder_action_input_proj')] = a.pop(k)\n",
+    "    if k.startswith('model.encoder_joint_proj.'):\n",
+    "        conv[k.replace('encoder_joint_proj', 'vae_encoder_robot_state_input_proj')] = a.pop(k)\n",
+    "    if k.startswith('model.latent_proj.'):\n",
+    "        conv[k.replace('latent_proj', 'vae_encoder_latent_output_proj')] = a.pop(k)\n",
+    "    if k.startswith('model.latent_out_proj.'):\n",
+    "        conv[k.replace('latent_out_proj', 'encoder_latent_input_proj')] = a.pop(k)\n",
+    "    if k.startswith('model.additional_pos_embed.'):\n",
+    "        conv[k.replace('additional_pos_embed', 'encoder_robot_and_latent_pos_embed')] = a.pop(k)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 46,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "OrderedDict()"
+      ]
+     },
+     "execution_count": 46,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "a"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 47,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for k, v in conv.items():\n",
+    "    assert b[k].shape == v.shape\n",
+    "    b[k] = v"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 53,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_file(b, converted_ckpt_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 54,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'/home/thomwolf/Documents/Github/ACT/checkpoints/blue_red_sort/config.yaml'"
+      ]
+     },
+     "execution_count": 54,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# Now also copy the config files\n",
+    "import shutil\n",
+    "shutil.copy(comparison_config_json_path, converted_ckpt_path.replace('model.safetensors', 'config.json'))\n",
+    "shutil.copy(comparison_config_yaml_path, converted_ckpt_path.replace('model.safetensors', 'config.yaml'))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "lerobot",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.14"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

examples/real_robot_example/gym_real_world/__init__.py

@@ -0,0 +1,8 @@
+from gymnasium.envs.registration import register
+
+register(
+    id="gym_real_world/RealEnv-v0",
+    entry_point="gym_real_world.gym_environment:RealEnv",
+    max_episode_steps=300,
+    nondeterministic=True,
+)

examples/real_robot_example/gym_real_world/dynamixel.py

@@ -0,0 +1,363 @@
+# ruff: noqa
+"""From Alexander Koch low_cost_robot codebase at https://github.com/AlexanderKoch-Koch/low_cost_robot
+Dynamixel class to control the dynamixel servos
+"""
+
+from __future__ import annotations
+
+import enum
+import math
+import os
+from dataclasses import dataclass
+
+import numpy as np
+from dynamixel_sdk import *  # Uses Dynamixel SDK library
+
+
+def pos2pwm(pos: np.ndarray) -> np.ndarray:
+    """
+    :param pos: numpy array of joint positions in range [-pi, pi]
+    :return: numpy array of pwm values in range [0, 4096]
+    """
+    return ((pos / 3.14 + 1.0) * 2048).astype(np.int64)
+
+
+def pwm2pos(pwm: np.ndarray) -> np.ndarray:
+    """
+    :param pwm: numpy array of pwm values in range [0, 4096]
+    :return: numpy array of joint positions in range [-pi, pi]
+    """
+    return (pwm / 2048 - 1) * 3.14
+
+
+def pwm2vel(pwm: np.ndarray) -> np.ndarray:
+    """
+    :param pwm: numpy array of pwm/s joint velocities
+    :return: numpy array of rad/s joint velocities
+    """
+    return pwm * 3.14 / 2048
+
+
+def vel2pwm(vel: np.ndarray) -> np.ndarray:
+    """
+    :param vel: numpy array of rad/s joint velocities
+    :return: numpy array of pwm/s joint velocities
+    """
+    return (vel * 2048 / 3.14).astype(np.int64)
+
+
+class ReadAttribute(enum.Enum):
+    TEMPERATURE = 146
+    VOLTAGE = 145
+    VELOCITY = 128
+    POSITION = 132
+    CURRENT = 126
+    PWM = 124
+    HARDWARE_ERROR_STATUS = 70
+    HOMING_OFFSET = 20
+    BAUDRATE = 8
+
+
+class OperatingMode(enum.Enum):
+    VELOCITY = 1
+    POSITION = 3
+    CURRENT_CONTROLLED_POSITION = 5
+    PWM = 16
+    UNKNOWN = -1
+
+
+class Dynamixel:
+    ADDR_TORQUE_ENABLE = 64
+    ADDR_GOAL_POSITION = 116
+    ADDR_VELOCITY_LIMIT = 44
+    ADDR_GOAL_PWM = 100
+    OPERATING_MODE_ADDR = 11
+    POSITION_I = 82
+    POSITION_P = 84
+    ADDR_ID = 7
+
+    @dataclass
+    class Config:
+        def instantiate(self):
+            return Dynamixel(self)
+
+        baudrate: int = 57600
+        protocol_version: float = 2.0
+        device_name: str = ""  # /dev/tty.usbserial-1120'
+        dynamixel_id: int = 1
+
+    def __init__(self, config: Config):
+        self.config = config
+        self.connect()
+
+    def connect(self):
+        if self.config.device_name == "":
+            for port_name in os.listdir("/dev"):
+                if "ttyUSB" in port_name or "ttyACM" in port_name:
+                    self.config.device_name = "/dev/" + port_name
+                    print(f"using device {self.config.device_name}")
+        self.portHandler = PortHandler(self.config.device_name)
+        # self.portHandler.LA
+        self.packetHandler = PacketHandler(self.config.protocol_version)
+        if not self.portHandler.openPort():
+            raise Exception(f"Failed to open port {self.config.device_name}")
+
+        if not self.portHandler.setBaudRate(self.config.baudrate):
+            raise Exception(f"failed to set baudrate to {self.config.baudrate}")
+
+        # self.operating_mode = OperatingMode.UNKNOWN
+        # self.torque_enabled = False
+        # self._disable_torque()
+
+        self.operating_modes = [None for _ in range(32)]
+        self.torque_enabled = [None for _ in range(32)]
+        return True
+
+    def disconnect(self):
+        self.portHandler.closePort()
+
+    def set_goal_position(self, motor_id, goal_position):
+        # if self.operating_modes[motor_id] is not OperatingMode.POSITION:
+        #     self._disable_torque(motor_id)
+        #     self.set_operating_mode(motor_id, OperatingMode.POSITION)
+
+        # if not self.torque_enabled[motor_id]:
+        #     self._enable_torque(motor_id)
+
+        # self._enable_torque(motor_id)
+        dxl_comm_result, dxl_error = self.packetHandler.write4ByteTxRx(
+            self.portHandler, motor_id, self.ADDR_GOAL_POSITION, goal_position
+        )
+        # self._process_response(dxl_comm_result, dxl_error)
+        # print(f'set position of motor {motor_id} to {goal_position}')
+
+    def set_pwm_value(self, motor_id: int, pwm_value, tries=3):
+        if self.operating_modes[motor_id] is not OperatingMode.PWM:
+            self._disable_torque(motor_id)
+            self.set_operating_mode(motor_id, OperatingMode.PWM)
+
+        if not self.torque_enabled[motor_id]:
+            self._enable_torque(motor_id)
+            # print(f'enabling torque')
+        dxl_comm_result, dxl_error = self.packetHandler.write2ByteTxRx(
+            self.portHandler, motor_id, self.ADDR_GOAL_PWM, pwm_value
+        )
+        # self._process_response(dxl_comm_result, dxl_error)
+        # print(f'set pwm of motor {motor_id} to {pwm_value}')
+        if dxl_comm_result != COMM_SUCCESS:
+            if tries <= 1:
+                raise ConnectionError(f"dxl_comm_result: {self.packetHandler.getTxRxResult(dxl_comm_result)}")
+            else:
+                print(f"dynamixel pwm setting failure trying again with {tries - 1} tries")
+                self.set_pwm_value(motor_id, pwm_value, tries=tries - 1)
+        elif dxl_error != 0:
+            print(f"dxl error {dxl_error}")
+            raise ConnectionError(f"dynamixel error: {self.packetHandler.getTxRxResult(dxl_error)}")
+
+    def read_temperature(self, motor_id: int):
+        return self._read_value(motor_id, ReadAttribute.TEMPERATURE, 1)
+
+    def read_velocity(self, motor_id: int):
+        pos = self._read_value(motor_id, ReadAttribute.VELOCITY, 4)
+        if pos > 2**31:
+            pos -= 2**32
+        # print(f'read position {pos} for motor {motor_id}')
+        return pos
+
+    def read_position(self, motor_id: int):
+        pos = self._read_value(motor_id, ReadAttribute.POSITION, 4)
+        if pos > 2**31:
+            pos -= 2**32
+        # print(f'read position {pos} for motor {motor_id}')
+        return pos
+
+    def read_position_degrees(self, motor_id: int) -> float:
+        return (self.read_position(motor_id) / 4096) * 360
+
+    def read_position_radians(self, motor_id: int) -> float:
+        return (self.read_position(motor_id) / 4096) * 2 * math.pi
+
+    def read_current(self, motor_id: int):
+        current = self._read_value(motor_id, ReadAttribute.CURRENT, 2)
+        if current > 2**15:
+            current -= 2**16
+        return current
+
+    def read_present_pwm(self, motor_id: int):
+        return self._read_value(motor_id, ReadAttribute.PWM, 2)
+
+    def read_hardware_error_status(self, motor_id: int):
+        return self._read_value(motor_id, ReadAttribute.HARDWARE_ERROR_STATUS, 1)
+
+    def disconnect(self):
+        self.portHandler.closePort()
+
+    def set_id(self, old_id, new_id, use_broadcast_id: bool = False):
+        """
+        sets the id of the dynamixel servo
+        @param old_id: current id of the servo
+        @param new_id: new id
+        @param use_broadcast_id: set ids of all connected dynamixels if True.
+         If False, change only servo with self.config.id
+        @return:
+        """
+        if use_broadcast_id:
+            current_id = 254
+        else:
+            current_id = old_id
+        dxl_comm_result, dxl_error = self.packetHandler.write1ByteTxRx(
+            self.portHandler, current_id, self.ADDR_ID, new_id
+        )
+        self._process_response(dxl_comm_result, dxl_error, old_id)
+        self.config.id = id
+
+    def _enable_torque(self, motor_id):
+        dxl_comm_result, dxl_error = self.packetHandler.write1ByteTxRx(
+            self.portHandler, motor_id, self.ADDR_TORQUE_ENABLE, 1
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+        self.torque_enabled[motor_id] = True
+
+    def _disable_torque(self, motor_id):
+        dxl_comm_result, dxl_error = self.packetHandler.write1ByteTxRx(
+            self.portHandler, motor_id, self.ADDR_TORQUE_ENABLE, 0
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+        self.torque_enabled[motor_id] = False
+
+    def _process_response(self, dxl_comm_result: int, dxl_error: int, motor_id: int):
+        if dxl_comm_result != COMM_SUCCESS:
+            raise ConnectionError(
+                f"dxl_comm_result for motor {motor_id}: {self.packetHandler.getTxRxResult(dxl_comm_result)}"
+            )
+        elif dxl_error != 0:
+            print(f"dxl error {dxl_error}")
+            raise ConnectionError(
+                f"dynamixel error for motor {motor_id}: {self.packetHandler.getTxRxResult(dxl_error)}"
+            )
+
+    def set_operating_mode(self, motor_id: int, operating_mode: OperatingMode):
+        dxl_comm_result, dxl_error = self.packetHandler.write2ByteTxRx(
+            self.portHandler, motor_id, self.OPERATING_MODE_ADDR, operating_mode.value
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+        self.operating_modes[motor_id] = operating_mode
+
+    def set_pwm_limit(self, motor_id: int, limit: int):
+        dxl_comm_result, dxl_error = self.packetHandler.write2ByteTxRx(self.portHandler, motor_id, 36, limit)
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+
+    def set_velocity_limit(self, motor_id: int, velocity_limit):
+        dxl_comm_result, dxl_error = self.packetHandler.write4ByteTxRx(
+            self.portHandler, motor_id, self.ADDR_VELOCITY_LIMIT, velocity_limit
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+
+    def set_P(self, motor_id: int, P: int):
+        dxl_comm_result, dxl_error = self.packetHandler.write2ByteTxRx(
+            self.portHandler, motor_id, self.POSITION_P, P
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+
+    def set_I(self, motor_id: int, I: int):
+        dxl_comm_result, dxl_error = self.packetHandler.write2ByteTxRx(
+            self.portHandler, motor_id, self.POSITION_I, I
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+
+    def read_home_offset(self, motor_id: int):
+        self._disable_torque(motor_id)
+        # dxl_comm_result, dxl_error = self.packetHandler.write4ByteTxRx(self.portHandler, motor_id,
+        #                                                                ReadAttribute.HOMING_OFFSET.value, home_position)
+        home_offset = self._read_value(motor_id, ReadAttribute.HOMING_OFFSET, 4)
+        # self._process_response(dxl_comm_result, dxl_error)
+        self._enable_torque(motor_id)
+        return home_offset
+
+    def set_home_offset(self, motor_id: int, home_position: int):
+        self._disable_torque(motor_id)
+        dxl_comm_result, dxl_error = self.packetHandler.write4ByteTxRx(
+            self.portHandler, motor_id, ReadAttribute.HOMING_OFFSET.value, home_position
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+        self._enable_torque(motor_id)
+
+    def set_baudrate(self, motor_id: int, baudrate):
+        # translate baudrate into dynamixel baudrate setting id
+        if baudrate == 57600:
+            baudrate_id = 1
+        elif baudrate == 1_000_000:
+            baudrate_id = 3
+        elif baudrate == 2_000_000:
+            baudrate_id = 4
+        elif baudrate == 3_000_000:
+            baudrate_id = 5
+        elif baudrate == 4_000_000:
+            baudrate_id = 6
+        else:
+            raise Exception("baudrate not implemented")
+
+        self._disable_torque(motor_id)
+        dxl_comm_result, dxl_error = self.packetHandler.write1ByteTxRx(
+            self.portHandler, motor_id, ReadAttribute.BAUDRATE.value, baudrate_id
+        )
+        self._process_response(dxl_comm_result, dxl_error, motor_id)
+
+    def _read_value(self, motor_id, attribute: ReadAttribute, num_bytes: int, tries=10):
+        try:
+            if num_bytes == 1:
+                value, dxl_comm_result, dxl_error = self.packetHandler.read1ByteTxRx(
+                    self.portHandler, motor_id, attribute.value
+                )
+            elif num_bytes == 2:
+                value, dxl_comm_result, dxl_error = self.packetHandler.read2ByteTxRx(
+                    self.portHandler, motor_id, attribute.value
+                )
+            elif num_bytes == 4:
+                value, dxl_comm_result, dxl_error = self.packetHandler.read4ByteTxRx(
+                    self.portHandler, motor_id, attribute.value
+                )
+        except Exception:
+            if tries == 0:
+                raise Exception
+            else:
+                return self._read_value(motor_id, attribute, num_bytes, tries=tries - 1)
+        if dxl_comm_result != COMM_SUCCESS:
+            if tries <= 1:
+                # print("%s" % self.packetHandler.getTxRxResult(dxl_comm_result))
+                raise ConnectionError(f"dxl_comm_result {dxl_comm_result} for servo {motor_id} value {value}")
+            else:
+                print(f"dynamixel read failure for servo {motor_id} trying again with {tries - 1} tries")
+                time.sleep(0.02)
+                return self._read_value(motor_id, attribute, num_bytes, tries=tries - 1)
+        elif dxl_error != 0:  # # print("%s" % self.packetHandler.getRxPacketError(dxl_error))
+            # raise ConnectionError(f'dxl_error {dxl_error} binary ' + "{0:b}".format(37))
+            if tries == 0 and dxl_error != 128:
+                raise Exception(f"Failed to read value from motor {motor_id} error is {dxl_error}")
+            else:
+                return self._read_value(motor_id, attribute, num_bytes, tries=tries - 1)
+        return value
+
+    def set_home_position(self, motor_id: int):
+        print(f"setting home position for motor {motor_id}")
+        self.set_home_offset(motor_id, 0)
+        current_position = self.read_position(motor_id)
+        print(f"position before {current_position}")
+        self.set_home_offset(motor_id, -current_position)
+        # dynamixel.set_home_offset(motor_id, -4096)
+        # dynamixel.set_home_offset(motor_id, -4294964109)
+        current_position = self.read_position(motor_id)
+        # print(f'signed position {current_position - 2** 32}')
+        print(f"position after {current_position}")
+
+
+if __name__ == "__main__":
+    dynamixel = Dynamixel.Config(baudrate=1_000_000, device_name="/dev/tty.usbmodem57380045631").instantiate()
+    motor_id = 1
+    pos = dynamixel.read_position(motor_id)
+    for i in range(10):
+        s = time.monotonic()
+        pos = dynamixel.read_position(motor_id)
+        delta = time.monotonic() - s
+        print(f"read position took {delta}")
+        print(f"position {pos}")

examples/real_robot_example/gym_real_world/gym_environment.py

@@ -0,0 +1,192 @@
+import time
+from unittest.mock import MagicMock
+
+import cv2
+import gymnasium as gym
+import numpy as np
+from gymnasium import spaces
+
+from .dynamixel import pos2pwm, pwm2pos
+from .robot import Robot
+
+FPS = 30
+
+CAMERAS_SHAPES = {
+    "images.high": (480, 640, 3),
+    "images.low": (480, 640, 3),
+}
+
+CAMERAS_PORTS = {
+    "images.high": "/dev/video6",
+    "images.low": "/dev/video0",
+}
+
+LEADER_PORT = "/dev/ttyACM1"
+FOLLOWER_PORT = "/dev/ttyACM0"
+
+MockRobot = MagicMock()
+MockRobot.read_position = MagicMock()
+MockRobot.read_position.return_value = np.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0])
+
+MockCamera = MagicMock()
+MockCamera.isOpened = MagicMock(return_value=True)
+MockCamera.read = MagicMock(return_value=(True, np.zeros((480, 640, 3), dtype=np.uint8)))
+
+
+def capture_image(cam, cam_width, cam_height):
+    # Capture a single frame
+    _, frame = cam.read()
+    image = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+    # # Define your crop coordinates (top left corner and bottom right corner)
+    # x1, y1 = 400, 0  # Example starting coordinates (top left of the crop rectangle)
+    # x2, y2 = 1600, 900  # Example ending coordinates (bottom right of the crop rectangle)
+    # # Crop the image
+    # image = image[y1:y2, x1:x2]
+    # Resize the image
+    image = cv2.resize(image, (cam_width, cam_height), interpolation=cv2.INTER_AREA)
+
+    return image
+
+
+class RealEnv(gym.Env):
+    metadata = {}
+
+    def __init__(
+        self,
+        record: bool = False,
+        num_joints: int = 6,
+        cameras_shapes: dict = CAMERAS_SHAPES,
+        cameras_ports: dict = CAMERAS_PORTS,
+        follower_port: str = FOLLOWER_PORT,
+        leader_port: str = LEADER_PORT,
+        warmup_steps: int = 100,
+        trigger_torque=70,
+        fps: int = FPS,
+        fps_tolerance: float = 0.1,
+        mock: bool = False,
+    ):
+        self.num_joints = num_joints
+        self.cameras_shapes = cameras_shapes
+        self.cameras_ports = cameras_ports
+        self.warmup_steps = warmup_steps
+        assert len(self.cameras_shapes) == len(self.cameras_ports), "Number of cameras and shapes must match."
+
+        self.follower_port = follower_port
+        self.leader_port = leader_port
+        self.record = record
+        self.fps = fps
+        self.fps_tolerance = fps_tolerance
+
+        # Initialize the robot
+        self.follower = Robot(device_name=self.follower_port) if not mock else MockRobot
+        if self.record:
+            self.leader = Robot(device_name=self.leader_port) if not mock else MockRobot
+            self.leader.set_trigger_torque(trigger_torque)
+
+        # Initialize the cameras - sorted by camera names
+        self.cameras = {}
+        for cn, p in sorted(self.cameras_ports.items()):
+            self.cameras[cn] = cv2.VideoCapture(p) if not mock else MockCamera
+            if not self.cameras[cn].isOpened():
+                raise OSError(
+                    f"Cannot open camera port {p} for {cn}."
+                    f" Make sure the camera is connected and the port is correct."
+                    f"Also check you are not spinning several instances of the same environment (eval.batch_size)"
+                )
+
+        # Specify gym action and observation spaces
+        observation_space = {}
+
+        if self.num_joints > 0:
+            observation_space["agent_pos"] = spaces.Box(
+                low=-1000.0,
+                high=1000.0,
+                shape=(num_joints,),
+                dtype=np.float64,
+            )
+        if self.record:
+            observation_space["leader_pos"] = spaces.Box(
+                low=-1000.0,
+                high=1000.0,
+                shape=(num_joints,),
+                dtype=np.float64,
+            )
+
+        if self.cameras_shapes:
+            for cn, hwc_shape in self.cameras_shapes.items():
+                # Assumes images are unsigned int8 in [0,255]
+                observation_space[cn] = spaces.Box(
+                    low=0,
+                    high=255,
+                    # height x width x channels (e.g. 480 x 640 x 3)
+                    shape=hwc_shape,
+                    dtype=np.uint8,
+                )
+
+        self.observation_space = spaces.Dict(observation_space)
+        self.action_space = spaces.Box(low=-1, high=1, shape=(num_joints,), dtype=np.float32)
+
+        self._observation = {}
+        self._terminated = False
+        self.timestamps = []
+
+    def _get_obs(self):
+        qpos = self.follower.read_position()
+        self._observation["agent_pos"] = pwm2pos(qpos)
+        for cn, c in self.cameras.items():
+            self._observation[cn] = capture_image(c, self.cameras_shapes[cn][1], self.cameras_shapes[cn][0])
+
+        if self.record:
+            action = self.leader.read_position()
+            self._observation["leader_pos"] = pwm2pos(action)
+
+    def reset(self, seed: int | None = None):
+        # Reset the robot and sync the leader and follower if we are recording
+        for _ in range(self.warmup_steps):
+            self._get_obs()
+            if self.record:
+                self.follower.set_goal_pos(pos2pwm(self._observation["leader_pos"]))
+        self._terminated = False
+        info = {}
+        self.timestamps = []
+        return self._observation, info
+
+    def step(self, action: np.ndarray = None):
+        if self.timestamps:
+            # wait the right amount of time to stay at the desired fps
+            time.sleep(max(0, 1 / self.fps - (time.time() - self.timestamps[-1])))
+
+        self.timestamps.append(time.time())
+
+        # Get the observation
+        self._get_obs()
+        if self.record:
+            # Teleoperate the leader
+            self.follower.set_goal_pos(pos2pwm(self._observation["leader_pos"]))
+        else:
+            # Apply the action to the follower
+            self.follower.set_goal_pos(pos2pwm(action))
+
+        reward = 0
+        terminated = truncated = self._terminated
+        info = {"timestamp": self.timestamps[-1] - self.timestamps[0], "fps_error": False}
+
+        # Check if we are able to keep up with the desired fps
+        if len(self.timestamps) > 1 and (self.timestamps[-1] - self.timestamps[-2]) > 1 / (
+            self.fps - self.fps_tolerance
+        ):
+            print(
+                f"Error: recording fps {1 / (self.timestamps[-1] - self.timestamps[-2]):.5f} is lower"
+                f" than min admited fps {(self.fps - self.fps_tolerance):.5f}"
+                f" at frame {len(self.timestamps)}"
+            )
+            info["fps_error"] = True
+
+        return self._observation, reward, terminated, truncated, info
+
+    def render(self): ...
+
+    def close(self):
+        self.follower._disable_torque()
+        if self.record:
+            self.leader._disable_torque()

examples/real_robot_example/gym_real_world/robot.py

@@ -0,0 +1,173 @@
+# ruff: noqa
+"""From Alexander Koch low_cost_robot codebase at https://github.com/AlexanderKoch-Koch/low_cost_robot
+Class to control the robot using dynamixel servos.
+"""
+
+from enum import Enum, auto
+from typing import Union
+
+import numpy as np
+from dynamixel_sdk import DXL_HIBYTE, DXL_HIWORD, DXL_LOBYTE, DXL_LOWORD, GroupSyncRead, GroupSyncWrite
+
+from .dynamixel import Dynamixel, OperatingMode, ReadAttribute
+
+
+class MotorControlType(Enum):
+    PWM = auto()
+    POSITION_CONTROL = auto()
+    DISABLED = auto()
+    UNKNOWN = auto()
+
+
+class Robot:
+    def __init__(self, device_name: str, baudrate=1_000_000, servo_ids=[1, 2, 3, 4, 5, 6]) -> None:
+        self.servo_ids = servo_ids
+        self.dynamixel = Dynamixel.Config(baudrate=baudrate, device_name=device_name).instantiate()
+        self._init_motors()
+
+    def _init_motors(self):
+        self.position_reader = GroupSyncRead(
+            self.dynamixel.portHandler, self.dynamixel.packetHandler, ReadAttribute.POSITION.value, 4
+        )
+        for id in self.servo_ids:
+            self.position_reader.addParam(id)
+
+        self.velocity_reader = GroupSyncRead(
+            self.dynamixel.portHandler, self.dynamixel.packetHandler, ReadAttribute.VELOCITY.value, 4
+        )
+        for id in self.servo_ids:
+            self.velocity_reader.addParam(id)
+
+        self.pos_writer = GroupSyncWrite(
+            self.dynamixel.portHandler, self.dynamixel.packetHandler, self.dynamixel.ADDR_GOAL_POSITION, 4
+        )
+        for id in self.servo_ids:
+            self.pos_writer.addParam(id, [2048])
+
+        self.pwm_writer = GroupSyncWrite(
+            self.dynamixel.portHandler, self.dynamixel.packetHandler, self.dynamixel.ADDR_GOAL_PWM, 2
+        )
+        for id in self.servo_ids:
+            self.pwm_writer.addParam(id, [2048])
+        # self._disable_torque()
+        self.motor_control_state = MotorControlType.DISABLED
+
+    def read_position(self, tries=2):
+        """
+        Reads the joint positions of the robot. 2048 is the center position. 0 and 4096 are 180 degrees in each direction.
+        :param tries: maximum number of tries to read the position
+        :return: list of joint positions in range [0, 4096]
+        """
+        result = self.position_reader.txRxPacket()
+        if result != 0:
+            if tries > 0:
+                return self.read_position(tries=tries - 1)
+            else:
+                print("failed to read position!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+        positions = []
+        for id in self.servo_ids:
+            position = self.position_reader.getData(id, ReadAttribute.POSITION.value, 4)
+            if position > 2**31:
+                position -= 2**32
+            positions.append(position)
+        return np.array(positions)
+
+    def read_velocity(self):
+        """
+        Reads the joint velocities of the robot.
+        :return: list of joint velocities,
+        """
+        self.velocity_reader.txRxPacket()
+        velocties = []
+        for id in self.servo_ids:
+            velocity = self.velocity_reader.getData(id, ReadAttribute.VELOCITY.value, 4)
+            if velocity > 2**31:
+                velocity -= 2**32
+            velocties.append(velocity)
+        return np.array(velocties)
+
+    def set_goal_pos(self, action):
+        """
+        :param action: list or numpy array of target joint positions in range [0, 4096[
+        """
+        if self.motor_control_state is not MotorControlType.POSITION_CONTROL:
+            self._set_position_control()
+        for i, motor_id in enumerate(self.servo_ids):
+            data_write = [
+                DXL_LOBYTE(DXL_LOWORD(action[i])),
+                DXL_HIBYTE(DXL_LOWORD(action[i])),
+                DXL_LOBYTE(DXL_HIWORD(action[i])),
+                DXL_HIBYTE(DXL_HIWORD(action[i])),
+            ]
+            self.pos_writer.changeParam(motor_id, data_write)
+
+        self.pos_writer.txPacket()
+
+    def set_pwm(self, action):
+        """
+        Sets the pwm values for the servos.
+        :param action: list or numpy array of pwm values in range [0, 885]
+        """
+        if self.motor_control_state is not MotorControlType.PWM:
+            self._set_pwm_control()
+        for i, motor_id in enumerate(self.servo_ids):
+            data_write = [
+                DXL_LOBYTE(DXL_LOWORD(action[i])),
+                DXL_HIBYTE(DXL_LOWORD(action[i])),
+            ]
+            self.pwm_writer.changeParam(motor_id, data_write)
+
+        self.pwm_writer.txPacket()
+
+    def set_trigger_torque(self, torque: int):
+        """
+        Sets a constant torque torque for the last servo in the chain. This is useful for the trigger of the leader arm
+        """
+        self.dynamixel._enable_torque(self.servo_ids[-1])
+        self.dynamixel.set_pwm_value(self.servo_ids[-1], torque)
+
+    def limit_pwm(self, limit: Union[int, list, np.ndarray]):
+        """
+        Limits the pwm values for the servos in for position control
+        @param limit: 0 ~ 885
+        @return:
+        """
+        if isinstance(limit, int):
+            limits = [
+                limit,
+            ] * 5
+        else:
+            limits = limit
+        self._disable_torque()
+        for motor_id, limit in zip(self.servo_ids, limits, strict=False):
+            self.dynamixel.set_pwm_limit(motor_id, limit)
+        self._enable_torque()
+
+    def _disable_torque(self):
+        print(f"disabling torque for servos {self.servo_ids}")
+        for motor_id in self.servo_ids:
+            self.dynamixel._disable_torque(motor_id)
+
+    def _enable_torque(self):
+        print(f"enabling torque for servos {self.servo_ids}")
+        for motor_id in self.servo_ids:
+            self.dynamixel._enable_torque(motor_id)
+
+    def _set_pwm_control(self):
+        self._disable_torque()
+        for motor_id in self.servo_ids:
+            self.dynamixel.set_operating_mode(motor_id, OperatingMode.PWM)
+        self._enable_torque()
+        self.motor_control_state = MotorControlType.PWM
+
+    def _set_position_control(self):
+        self._disable_torque()
+        for motor_id in self.servo_ids:
+            # TODO(rcadene): redesign
+            if motor_id == 9:
+                self.dynamixel.set_operating_mode(9, OperatingMode.CURRENT_CONTROLLED_POSITION)
+            else:
+                self.dynamixel.set_operating_mode(motor_id, OperatingMode.POSITION)
+
+        self._enable_torque()
+        self.motor_control_state = MotorControlType.POSITION_CONTROL

examples/real_robot_example/record_training_data.py

@@ -0,0 +1,237 @@
+"""This script demonstrates how to record a LeRobot dataset of training data
+using a very simple gym environment (see in examples/real_robot_example/gym_real_world/gym_environment.py).
+
+"""
+
+import argparse
+import copy
+import os
+from pathlib import Path
+
+import gym_real_world  # noqa: F401
+import gymnasium as gym
+import numpy as np
+import torch
+from datasets import Dataset, Features, Sequence, Value
+from omegaconf import OmegaConf
+from tqdm import tqdm
+
+from lerobot.common.datasets.compute_stats import compute_stats
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, DATA_DIR, LeRobotDataset
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+from lerobot.scripts.push_dataset_to_hub import push_meta_data_to_hub, push_videos_to_hub, save_meta_data
+
+# parse the repo_id name via command line
+parser = argparse.ArgumentParser()
+parser.add_argument("--repo-id", type=str, default="thomwolf/blue_red_sort")
+parser.add_argument("--num-episodes", type=int, default=2)
+parser.add_argument("--num-frames", type=int, default=400)
+parser.add_argument("--num-workers", type=int, default=16)
+parser.add_argument("--keep-last", action="store_true")
+parser.add_argument("--data_dir", type=str, default=None)
+parser.add_argument("--push-to-hub", action="store_true")
+parser.add_argument("--fps", type=int, default=30, help="Frames per second of the recording.")
+parser.add_argument(
+    "--fps_tolerance",
+    type=float,
+    default=0.5,
+    help="Tolerance in fps for the recording before dropping episodes.",
+)
+parser.add_argument(
+    "--revision", type=str, default=CODEBASE_VERSION, help="Codebase version used to generate the dataset."
+)
+parser.add_argument("--gym-config", type=str, default=None, help="Path to the gym config file.")
+parser.add_argument("--mock_robot", action="store_true")
+args = parser.parse_args()
+
+repo_id = args.repo_id
+num_episodes = args.num_episodes
+num_frames = args.num_frames
+revision = args.revision
+fps = args.fps
+fps_tolerance = args.fps_tolerance
+
+out_data = DATA_DIR / repo_id if args.data_dir is None else Path(args.data_dir)
+
+# During data collection, frames are stored as png images in `images_dir`
+images_dir = out_data / "images"
+# After data collection, png images of each episode are encoded into a mp4 file stored in `videos_dir`
+videos_dir = out_data / "videos"
+meta_data_dir = out_data / "meta_data"
+
+gym_config = None
+if args.config is not None:
+    gym_config = OmegaConf.load(args.config)
+
+# Create image and video directories
+if not os.path.exists(images_dir):
+    os.makedirs(images_dir, exist_ok=True)
+if not os.path.exists(videos_dir):
+    os.makedirs(videos_dir, exist_ok=True)
+
+if __name__ == "__main__":
+    # Create the gym environment - check the kwargs in gym_real_world/gym_environment.py
+    gym_handle = "gym_real_world/RealEnv-v0"
+    gym_kwargs = {}
+    if gym_config is not None:
+        gym_kwargs = OmegaConf.to_container(gym_config.gym_kwargs)
+    env = gym.make(
+        gym_handle, disable_env_checker=True, record=True, fps=fps, fps_tolerance=fps_tolerance, mock=True
+    )
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+    ep_fps = []
+    id_from = 0
+    id_to = 0
+    os.system('spd-say "gym environment created"')
+
+    ep_idx = 0
+    while ep_idx < num_episodes:
+        # bring the follower to the leader and start camera
+        env.reset()
+
+        os.system(f'spd-say "go {ep_idx}"')
+        # init buffers
+        obs_replay = {k: [] for k in env.observation_space}
+
+        drop_episode = False
+        timestamps = []
+        for _ in tqdm(range(num_frames)):
+            # Apply the next action
+            observation, _, _, _, info = env.step(action=None)
+            # images_stacked = np.hstack(list(observation['pixels'].values()))
+            # images_stacked = cv2.cvtColor(images_stacked, cv2.COLOR_RGB2BGR)
+            # cv2.imshow('frame', images_stacked)
+
+            if info["fps_error"]:
+                os.system(f'spd-say "Error fps too low, dropping episode {ep_idx}"')
+                drop_episode = True
+                break
+
+            # store data
+            for key in observation:
+                obs_replay[key].append(copy.deepcopy(observation[key]))
+            timestamps.append(info["timestamp"])
+
+            # if cv2.waitKey(1) & 0xFF == ord('q'):
+            #     break
+
+        os.system('spd-say "stop"')
+
+        if not drop_episode:
+            os.system(f'spd-say "saving episode {ep_idx}"')
+            ep_dict = {}
+            # store images in png and create the video
+            for img_key in env.cameras:
+                save_images_concurrently(
+                    obs_replay[img_key],
+                    images_dir / f"{img_key}_episode_{ep_idx:06d}",
+                    args.num_workers,
+                )
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                # store the reference to the video frame
+                ep_dict[f"observation.{img_key}"] = [
+                    {"path": f"videos/{fname}", "timestamp": tstp} for tstp in timestamps
+                ]
+
+            state = torch.tensor(np.array(obs_replay["agent_pos"]))
+            action = torch.tensor(np.array(obs_replay["leader_pos"]))
+            next_done = torch.zeros(num_frames, dtype=torch.bool)
+            next_done[-1] = True
+
+            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.tensor(timestamps)
+            ep_dict["next.done"] = next_done
+            ep_fps.append(num_frames / timestamps[-1])
+            ep_dicts.append(ep_dict)
+            print(f"Episode {ep_idx} done, fps: {ep_fps[-1]:.2f}")
+
+            episode_data_index["from"].append(id_from)
+            episode_data_index["to"].append(
+                id_from + num_frames if args.keep_last else id_from + num_frames - 1
+            )
+
+            id_to = id_from + num_frames if args.keep_last else id_from + num_frames - 1
+            id_from = id_to
+
+            ep_idx += 1
+
+    env.close()
+
+    os.system('spd-say "encode video frames"')
+    for ep_idx in range(num_episodes):
+        for img_key in env.cameras:
+            # If necessary, we may want to encode the video
+            # with variable frame rate: https://superuser.com/questions/1661901/encoding-video-from-vfr-still-images
+            encode_video_frames(
+                images_dir / f"{img_key}_episode_{ep_idx:06d}",
+                videos_dir / f"{img_key}_episode_{ep_idx:06d}.mp4",
+                ep_fps[ep_idx],
+            )
+
+    os.system('spd-say "concatenate episodes"')
+    data_dict = concatenate_episodes(
+        ep_dicts, drop_episodes_last_frame=not args.keep_last
+    )  # Since our fps varies we are sometimes off tolerance for the last frame
+
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        features[key] = VideoFrame()
+
+    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.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)
+
+    info = {
+        "fps": sum(ep_fps) / len(ep_fps),  # to have a good tolerance in data processing for the slowest video
+        "video": 1,
+    }
+
+    os.system('spd-say "from preloaded"')
+    lerobot_dataset = LeRobotDataset.from_preloaded(
+        repo_id=repo_id,
+        version=revision,
+        hf_dataset=hf_dataset,
+        episode_data_index=episode_data_index,
+        info=info,
+        videos_dir=videos_dir,
+    )
+    os.system('spd-say "compute stats"')
+    stats = compute_stats(lerobot_dataset)
+
+    os.system('spd-say "save to disk"')
+    hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
+    hf_dataset.save_to_disk(str(out_data / "train"))
+
+    save_meta_data(info, stats, episode_data_index, meta_data_dir)
+
+    if args.push_to_hub:
+        hf_dataset.push_to_hub(repo_id, token=True, revision="main")
+        hf_dataset.push_to_hub(repo_id, token=True, revision=revision)
+
+        push_meta_data_to_hub(repo_id, meta_data_dir, revision="main")
+        push_meta_data_to_hub(repo_id, meta_data_dir, revision=revision)
+
+        push_videos_to_hub(repo_id, videos_dir, revision="main")
+        push_videos_to_hub(repo_id, videos_dir, revision=revision)

examples/real_robot_example/run_policy.py

@@ -0,0 +1,60 @@
+import argparse
+import logging
+from pathlib import Path
+
+import gym_real_world  # noqa: F401
+import gymnasium as gym  # noqa: F401
+from huggingface_hub import snapshot_download
+from huggingface_hub.utils._errors import RepositoryNotFoundError
+from huggingface_hub.utils._validators import HFValidationError
+
+from lerobot.common.utils.utils import init_logging
+from lerobot.scripts.eval import eval
+
+if __name__ == "__main__":
+    init_logging()
+
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument(
+        "-p",
+        "--pretrained-policy-name-or-path",
+        help=(
+            "Either the repo ID of a model hosted on the Hub or a path to a directory containing weights "
+            "saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch "
+            "(useful for debugging). This argument is mutually exclusive with `--config`."
+        ),
+    )
+    parser.add_argument("--revision", help="Optionally provide the Hugging Face Hub revision ID.")
+    parser.add_argument(
+        "overrides",
+        nargs="*",
+        help="Any key=value arguments to override config values (use dots for.nested=overrides)",
+    )
+    args = parser.parse_args()
+
+    try:
+        pretrained_policy_path = Path(
+            snapshot_download(args.pretrained_policy_name_or_path, revision=args.revision)
+        )
+    except (HFValidationError, RepositoryNotFoundError) as e:
+        if isinstance(e, HFValidationError):
+            error_message = (
+                "The provided pretrained_policy_name_or_path is not a valid Hugging Face Hub repo ID."
+            )
+        else:
+            error_message = (
+                "The provided pretrained_policy_name_or_path was not found on the Hugging Face Hub."
+            )
+
+        logging.warning(f"{error_message} Treating it as a local directory.")
+        pretrained_policy_path = Path(args.pretrained_policy_name_or_path)
+    if not pretrained_policy_path.is_dir() or not pretrained_policy_path.exists():
+        raise ValueError(
+            "The provided pretrained_policy_name_or_path is not a valid/existing Hugging Face Hub "
+            "repo ID, nor is it an existing local directory."
+        )
+
+    eval(pretrained_policy_path=pretrained_policy_path, config_overrides=args.overrides)

examples/real_robot_example/train_config/policy/act_real_world.yaml

@@ -0,0 +1,103 @@
+# @package _global_
+
+# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. right_wrist, left_wrist, images,
+# 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: ???
+
+override_dataset_stats:
+  observation.images.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.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: 1000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 1000
+  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(1, ${policy.chunk_size} + 1)]"
+
+eval:
+  n_episodes: 1
+  batch_size: 1
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.high: [3, 480, 640]
+    observation.images.low: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.high: mean_std
+    observation.images.low: mean_std
+    observation.state: 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

examples/real_robot_example/train_config/policy/act_real_world_debug.yaml

@@ -0,0 +1,103 @@
+# @package _global_
+
+# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. right_wrist, left_wrist, images,
+# front) 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: ???
+
+override_dataset_stats:
+  observation.images.top:
+    # 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.front:
+    # 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: 1000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 1000
+  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(1, ${policy.chunk_size} + 1)]"
+
+eval:
+  n_episodes: 1
+  batch_size: 1
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.top: [3, 480, 640]
+    observation.images.front: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.top: mean_std
+    observation.images.front: mean_std
+    observation.state: 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

lerobot/__init__.py

@@ -1,3 +1,18 @@
+#!/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 lists of available environments, dataset and policies to reflect the current state of LeRobot library.
 We do not want to import all the dependencies, but instead we keep it lightweight to ensure fast access to these variables.
@@ -30,6 +45,9 @@ import itertools
 
 from lerobot.__version__ import __version__  # noqa: F401
 
+# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
+# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
+# a yaml file AND a environment name. The difference should be more obvious.
 available_tasks_per_env = {
     "aloha": [
         "AlohaInsertion-v0",
@@ -37,6 +55,7 @@ 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())
 
@@ -46,13 +65,38 @@ available_datasets_per_env = {
         "lerobot/aloha_sim_insertion_scripted",
         "lerobot/aloha_sim_transfer_cube_human",
         "lerobot/aloha_sim_transfer_cube_scripted",
+        "lerobot/aloha_sim_insertion_human_image",
+        "lerobot/aloha_sim_insertion_scripted_image",
+        "lerobot/aloha_sim_transfer_cube_human_image",
+        "lerobot/aloha_sim_transfer_cube_scripted_image",
     ],
-    "pusht": ["lerobot/pusht"],
+    "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
     "xarm": [
         "lerobot/xarm_lift_medium",
         "lerobot/xarm_lift_medium_replay",
         "lerobot/xarm_push_medium",
         "lerobot/xarm_push_medium_replay",
+        "lerobot/xarm_lift_medium_image",
+        "lerobot/xarm_lift_medium_replay_image",
+        "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",
     ],
 }
 
@@ -85,16 +129,19 @@ available_datasets = list(
     itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
 )
 
+# lists all available policies from `lerobot/common/policies` by their class attribute: `name`.
 available_policies = [
     "act",
     "diffusion",
     "tdmpc",
 ]
 
+# keys and values refer to yaml files
 available_policies_per_env = {
     "aloha": ["act"],
     "pusht": ["diffusion"],
     "xarm": ["tdmpc"],
+    "dora_aloha_real": ["act_real"],
 }
 
 env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

lerobot/__version__.py

@@ -1,3 +1,18 @@
+#!/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.
 """To enable `lerobot.__version__`"""
 
 from importlib.metadata import PackageNotFoundError, version

lerobot/common/datasets/_video_benchmark/run_video_benchmark.py

@@ -1,3 +1,18 @@
+#!/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 json
 import random
 import shutil

lerobot/common/datasets/compute_stats.py

@@ -1,17 +1,30 @@
+#!/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 copy import deepcopy
 from math import ceil
 
-import datasets
 import einops
 import torch
 import tqdm
 from datasets import Image
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.datasets.video_utils import VideoFrame
 
 
-def get_stats_einops_patterns(dataset: LeRobotDataset | datasets.Dataset, num_workers=0):
+def get_stats_einops_patterns(dataset, num_workers=0):
     """These einops patterns will be used to aggregate batches and compute statistics.
 
     Note: We assume the images are in channel first format
@@ -51,9 +64,8 @@ def get_stats_einops_patterns(dataset: LeRobotDataset | datasets.Dataset, num_wo
     return stats_patterns
 
 
-def compute_stats(
-    dataset: LeRobotDataset | datasets.Dataset, batch_size=32, num_workers=16, max_num_samples=None
-):
+def compute_stats(dataset, batch_size=32, num_workers=16, 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)
 
@@ -144,3 +156,54 @@ def compute_stats(
             "min": min[key],
         }
     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.
+
+    The final stats will have the union of all data keys from each of the datasets.
+
+    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, ...)
+    - new_min = min(min_dataset_0, min_dataset_1, ...)
+    - new_mean = (mean of all data)
+    - new_std = (std of all data)
+    """
+    data_keys = set()
+    for dataset in ls_datasets:
+        data_keys.update(dataset.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([d.stats[data_key][stat_key] for d in ls_datasets if data_key in d.stats], dim=0),
+                "n ... -> ...",
+                stat_key,
+            )
+        total_samples = sum(d.num_samples for d in ls_datasets if data_key in d.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_samples / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["mean"] = sum(
+            d.stats[data_key]["mean"] * (d.num_samples / total_samples)
+            for d in ls_datasets
+            if data_key in d.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_samples / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["std"] = torch.sqrt(
+            sum(
+                (d.stats[data_key]["std"] ** 2 + (d.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2)
+                * (d.num_samples / total_samples)
+                for d in ls_datasets
+                if data_key in d.stats
+            )
+        )
+    return stats

lerobot/common/datasets/factory.py

@@ -1,34 +1,88 @@
+#!/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 logging
 
 import torch
-from omegaconf import OmegaConf
+from omegaconf import ListConfig, OmegaConf
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
 
 
-def make_dataset(
-    cfg,
-    split="train",
-):
-    if cfg.env.name not in cfg.dataset_repo_id:
-        logging.warning(
-            f"There might be a mismatch between your training dataset ({cfg.dataset_repo_id=}) and your "
-            f"environment ({cfg.env.name=})."
-        )
+def resolve_delta_timestamps(cfg):
+    """Resolves delta_timestamps config key (in-place) by using `eval`.
 
+    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):
-                delta_timestamps[key] = eval(delta_timestamps[key])
+                # 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]"`.
+    Returns:
+        The LeRobotDataset.
+    """
+    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."
+        )
+
+    # A soft check to warn if the environment matches the dataset. Don't check if we are using a real world env (dora).
+    if not cfg.env.real_world:
+        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=})."
+                )
+
+    resolve_delta_timestamps(cfg)
 
     # TODO(rcadene): add data augmentations
 
-    dataset = LeRobotDataset(
-        cfg.dataset_repo_id,
-        split=split,
-        delta_timestamps=delta_timestamps,
-    )
+    if isinstance(cfg.dataset_repo_id, str):
+        dataset = LeRobotDataset(
+            cfg.dataset_repo_id,
+            split=split,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
+        )
+    else:
+        dataset = MultiLeRobotDataset(
+            cfg.dataset_repo_id, split=split, delta_timestamps=cfg.training.get("delta_timestamps")
+        )
 
     if cfg.get("override_dataset_stats"):
         for key, stats_dict in cfg.override_dataset_stats.items():

lerobot/common/datasets/lerobot_dataset.py

@@ -1,21 +1,42 @@
+#!/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 logging
 import os
 from pathlib import Path
+from typing import Callable
 
 import datasets
 import torch
+import torch.utils
 
+from lerobot.common.datasets.compute_stats import aggregate_stats
 from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
     load_episode_data_index,
     load_hf_dataset,
     load_info,
     load_previous_and_future_frames,
     load_stats,
     load_videos,
+    reset_episode_index,
 )
 from lerobot.common.datasets.video_utils import VideoFrame, load_from_videos
 
 DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
-CODEBASE_VERSION = "v1.3"
+CODEBASE_VERSION = "v1.4"
 
 
 class LeRobotDataset(torch.utils.data.Dataset):
@@ -25,7 +46,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         version: str | None = CODEBASE_VERSION,
         root: Path | None = DATA_DIR,
         split: str = "train",
-        transform: callable = None,
+        transform: Callable | None = None,
         delta_timestamps: dict[list[float]] | None = None,
     ):
         super().__init__()
@@ -39,7 +60,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # TODO(rcadene, aliberts): implement faster transfer
         # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
         self.hf_dataset = load_hf_dataset(repo_id, version, root, split)
-        self.episode_data_index = load_episode_data_index(repo_id, version, root)
+        if split == "train":
+            self.episode_data_index = load_episode_data_index(repo_id, version, root)
+        else:
+            self.episode_data_index = calculate_episode_data_index(self.hf_dataset)
+            self.hf_dataset = reset_episode_index(self.hf_dataset)
         self.stats = load_stats(repo_id, version, root)
         self.info = load_info(repo_id, version, root)
         if self.video:
@@ -150,7 +175,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
     @classmethod
     def from_preloaded(
         cls,
-        repo_id: str,
+        repo_id: str = "from_preloaded",
         version: str | None = CODEBASE_VERSION,
         root: Path | None = None,
         split: str = "train",
@@ -162,7 +187,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
         stats=None,
         info=None,
         videos_dir=None,
-    ):
+    ) -> "LeRobotDataset":
+        """Create a LeRobot Dataset from existing data and attributes instead of loading from the filesystem.
+
+        It is especially useful when converting raw data into LeRobotDataset before saving the dataset
+        on the filesystem or uploading to the hub.
+
+        Note: Meta-data attributes like `repo_id`, `version`, `root`, etc are optional and potentially
+        meaningless depending on the downstream usage of the return dataset.
+        """
         # create an empty object of type LeRobotDataset
         obj = cls.__new__(cls)
         obj.repo_id = repo_id
@@ -174,6 +207,193 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.hf_dataset = hf_dataset
         obj.episode_data_index = episode_data_index
         obj.stats = stats
-        obj.info = info
+        obj.info = info if info is not None else {}
         obj.videos_dir = videos_dir
         return obj
+
+
+class MultiLeRobotDataset(torch.utils.data.Dataset):
+    """A dataset consisting of multiple underlying `LeRobotDataset`s.
+
+    The underlying `LeRobotDataset`s are effectively concatenated, and this class adopts much of the API
+    structure of `LeRobotDataset`.
+    """
+
+    def __init__(
+        self,
+        repo_ids: list[str],
+        version: str | None = CODEBASE_VERSION,
+        root: Path | None = DATA_DIR,
+        split: str = "train",
+        transform: Callable | None = None,
+        delta_timestamps: dict[list[float]] | None = None,
+    ):
+        super().__init__()
+        self.repo_ids = repo_ids
+        # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
+        # are handled by this class.
+        self._datasets = [
+            LeRobotDataset(
+                repo_id,
+                version=version,
+                root=root,
+                split=split,
+                delta_timestamps=delta_timestamps,
+                transform=transform,
+            )
+            for repo_id in repo_ids
+        ]
+        # Check that some properties are consistent across datasets. Note: We may relax some of these
+        # consistency requirements in future iterations of this class.
+        for repo_id, dataset in zip(self.repo_ids, self._datasets, strict=True):
+            if dataset.info != self._datasets[0].info:
+                raise ValueError(
+                    f"Detected a mismatch in dataset info between {self.repo_ids[0]} and {repo_id}. This is "
+                    "not yet supported."
+                )
+        # Disable any data keys that are not common across all of the datasets. Note: we may relax this
+        # restriction in future iterations of this class. For now, this is necessary at least for being able
+        # to use PyTorch's default DataLoader collate function.
+        self.disabled_data_keys = set()
+        intersection_data_keys = set(self._datasets[0].hf_dataset.features)
+        for dataset in self._datasets:
+            intersection_data_keys.intersection_update(dataset.hf_dataset.features)
+        if len(intersection_data_keys) == 0:
+            raise RuntimeError(
+                "Multiple datasets were provided but they had no keys common to all of them. The "
+                "multi-dataset functionality currently only keeps common keys."
+            )
+        for repo_id, dataset in zip(self.repo_ids, self._datasets, strict=True):
+            extra_keys = set(dataset.hf_dataset.features).difference(intersection_data_keys)
+            logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
+            )
+            self.disabled_data_keys.update(extra_keys)
+
+        self.version = version
+        self.root = root
+        self.split = split
+        self.transform = transform
+        self.delta_timestamps = delta_timestamps
+        self.stats = aggregate_stats(self._datasets)
+
+    @property
+    def repo_id_to_index(self):
+        """Return a mapping from dataset repo_id to a dataset index automatically created by this class.
+
+        This index is incorporated as a data key in the dictionary returned by `__getitem__`.
+        """
+        return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
+
+    @property
+    def repo_index_to_id(self):
+        """Return the inverse mapping if repo_id_to_index."""
+        return {v: k for k, v in self.repo_id_to_index}
+
+    @property
+    def fps(self) -> int:
+        """Frames per second used during data collection.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].info["fps"]
+
+    @property
+    def video(self) -> bool:
+        """Returns True if this dataset loads video frames from mp4 files.
+
+        Returns False if it only loads images from png files.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].info.get("video", False)
+
+    @property
+    def features(self) -> datasets.Features:
+        features = {}
+        for dataset in self._datasets:
+            features.update({k: v for k, v in dataset.features.items() if k not in self.disabled_data_keys})
+        return features
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access image and video stream from cameras."""
+        keys = []
+        for key, feats in self.features.items():
+            if isinstance(feats, (datasets.Image, VideoFrame)):
+                keys.append(key)
+        return keys
+
+    @property
+    def video_frame_keys(self) -> list[str]:
+        """Keys to access video frames that requires to be decoded into images.
+
+        Note: It is empty if the dataset contains images only,
+        or equal to `self.cameras` if the dataset contains videos only,
+        or can even be a subset of `self.cameras` in a case of a mixed image/video dataset.
+        """
+        video_frame_keys = []
+        for key, feats in self.features.items():
+            if isinstance(feats, VideoFrame):
+                video_frame_keys.append(key)
+        return video_frame_keys
+
+    @property
+    def num_samples(self) -> int:
+        """Number of samples/frames."""
+        return sum(d.num_samples for d in self._datasets)
+
+    @property
+    def num_episodes(self) -> int:
+        """Number of episodes."""
+        return sum(d.num_episodes for d in self._datasets)
+
+    @property
+    def tolerance_s(self) -> float:
+        """Tolerance in seconds used to discard loaded frames when their timestamps
+        are not close enough from the requested frames. It is only used when `delta_timestamps`
+        is provided or when loading video frames from mp4 files.
+        """
+        # 1e-4 to account for possible numerical error
+        return 1 / self.fps - 1e-4
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self):
+            raise IndexError(f"Index {idx} out of bounds.")
+        # Determine which dataset to get an item from based on the index.
+        start_idx = 0
+        dataset_idx = 0
+        for dataset in self._datasets:
+            if idx >= start_idx + dataset.num_samples:
+                start_idx += dataset.num_samples
+                dataset_idx += 1
+                continue
+            break
+        else:
+            raise AssertionError("We expect the loop to break out as long as the index is within bounds.")
+        item = self._datasets[dataset_idx][idx - start_idx]
+        item["dataset_index"] = torch.tensor(dataset_idx)
+        for data_key in self.disabled_data_keys:
+            if data_key in item:
+                del item[data_key]
+        return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository IDs: '{self.repo_ids}',\n"
+            f"  Version: '{self.version}',\n"
+            f"  Split: '{self.split}',\n"
+            f"  Number of Samples: {self.num_samples},\n"
+            f"  Number of Episodes: {self.num_episodes},\n"
+            f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"
+            f"  Recorded Frames per Second: {self.fps},\n"
+            f"  Camera Keys: {self.camera_keys},\n"
+            f"  Video Frame Keys: {self.video_frame_keys if self.video else 'N/A'},\n"
+            f"  Transformations: {self.transform},\n"
+            f")"
+        )

lerobot/common/datasets/push_dataset_to_hub/_diffusion_policy_replay_buffer.py

@@ -1,3 +1,18 @@
+#!/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.

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -1,3 +1,18 @@
+#!/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 all obsolete download scripts. They are centralized here to not have to load
 useless dependencies when using datasets.
@@ -9,17 +24,16 @@ import shutil
 from pathlib import Path
 
 import tqdm
-
-ALOHA_RAW_URLS_DIR = "lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls"
+from huggingface_hub import snapshot_download
 
 
 def download_raw(raw_dir, dataset_id):
-    if "pusht" in dataset_id:
+    if "aloha" in dataset_id or "image" in dataset_id:
+        download_hub(raw_dir, dataset_id)
+    elif "pusht" in dataset_id:
         download_pusht(raw_dir)
     elif "xarm" in dataset_id:
         download_xarm(raw_dir)
-    elif "aloha" in dataset_id:
-        download_aloha(raw_dir, dataset_id)
     elif "umi" in dataset_id:
         download_umi(raw_dir)
     else:
@@ -88,37 +102,13 @@ def download_xarm(raw_dir: Path):
     zip_path.unlink()
 
 
-def download_aloha(raw_dir: Path, dataset_id: str):
-    import gdown
-
-    subset_id = dataset_id.replace("aloha_", "")
-    urls_path = Path(ALOHA_RAW_URLS_DIR) / f"{subset_id}.txt"
-    assert urls_path.exists(), f"{subset_id}.txt not found in '{ALOHA_RAW_URLS_DIR}' directory."
-
-    with open(urls_path) as f:
-        # strip lines and ignore empty lines
-        urls = [url.strip() for url in f if url.strip()]
-
-    # sanity check
-    for url in urls:
-        assert (
-            "drive.google.com/drive/folders" in url or "drive.google.com/file" in url
-        ), f"Wrong url provided '{url}' in file '{urls_path}'."
-
+def download_hub(raw_dir: Path, dataset_id: str):
     raw_dir = Path(raw_dir)
     raw_dir.mkdir(parents=True, exist_ok=True)
 
-    logging.info(f"Start downloading from google drive for {dataset_id}")
-    for url in urls:
-        if "drive.google.com/drive/folders" in url:
-            # when a folder url is given, download up to 50 files from the folder
-            gdown.download_folder(url, output=str(raw_dir), remaining_ok=True)
-
-        elif "drive.google.com/file" in url:
-            # because of the 50 files limit per folder, we download the remaining files (file by file)
-            gdown.download(url, output=str(raw_dir), fuzzy=True)
-
-    logging.info(f"End downloading from google drive for {dataset_id}")
+    logging.info(f"Start downloading from huggingface.co/cadene for {dataset_id}")
+    snapshot_download(f"cadene/{dataset_id}_raw", repo_type="dataset", local_dir=raw_dir)
+    logging.info(f"Finish downloading from huggingface.co/cadene for {dataset_id}")
 
 
 def download_umi(raw_dir: Path):
@@ -133,21 +123,30 @@ def download_umi(raw_dir: Path):
 if __name__ == "__main__":
     data_dir = Path("data")
     dataset_ids = [
+        "pusht_image",
+        "xarm_lift_medium_image",
+        "xarm_lift_medium_replay_image",
+        "xarm_push_medium_image",
+        "xarm_push_medium_replay_image",
+        "aloha_sim_insertion_human_image",
+        "aloha_sim_insertion_scripted_image",
+        "aloha_sim_transfer_cube_human_image",
+        "aloha_sim_transfer_cube_scripted_image",
         "pusht",
         "xarm_lift_medium",
         "xarm_lift_medium_replay",
         "xarm_push_medium",
         "xarm_push_medium_replay",
+        "aloha_sim_insertion_human",
+        "aloha_sim_insertion_scripted",
+        "aloha_sim_transfer_cube_human",
+        "aloha_sim_transfer_cube_scripted",
         "aloha_mobile_cabinet",
         "aloha_mobile_chair",
         "aloha_mobile_elevator",
         "aloha_mobile_shrimp",
         "aloha_mobile_wash_pan",
         "aloha_mobile_wipe_wine",
-        "aloha_sim_insertion_human",
-        "aloha_sim_insertion_scripted",
-        "aloha_sim_transfer_cube_human",
-        "aloha_sim_transfer_cube_scripted",
         "aloha_static_battery",
         "aloha_static_candy",
         "aloha_static_coffee",

lerobot/common/datasets/push_dataset_to_hub/_umi_imagecodecs_numcodecs.py

@@ -1,3 +1,18 @@
+#!/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

lerobot/common/datasets/push_dataset_to_hub/aloha_dora_format.py

@@ -0,0 +1,230 @@
+#!/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 logging
+import re
+from pathlib import Path
+
+import pandas as pd
+import torch
+from datasets import Dataset, Features, Image, Sequence, Value
+
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame
+from lerobot.common.utils.utils import init_logging
+
+
+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, out_dir: Path, fps: int):
+    # 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)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    videos_dir = out_dir / "videos"
+    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)
+
+    # Get the episode index containing for each unique episode index
+    first_ep_index_df = df.groupby("episode_index").agg(start_index=("index", "first")).reset_index()
+    from_ = first_ep_index_df["start_index"].tolist()
+    to_ = from_[1:] + [len(df)]
+    episode_data_index = {
+        "from": from_,
+        "to": to_,
+    }
+
+    return data_dict, episode_data_index
+
+
+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, out_dir: Path, fps=None, video=True, debug=False):
+    init_logging()
+
+    if debug:
+        logging.warning("debug=True not implemented. Falling back to debug=False.")
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+    else:
+        raise NotImplementedError()
+
+    if not video:
+        raise NotImplementedError()
+
+    data_df, episode_data_index = load_from_raw(raw_dir, out_dir, fps)
+    hf_dataset = to_hf_dataset(data_df, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -1,8 +1,23 @@
+#!/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 re
+import gc
 import shutil
 from pathlib import Path
 
@@ -28,9 +43,6 @@ def get_cameras(hdf5_data):
 
 
 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:
@@ -47,27 +59,23 @@ def check_format(raw_dir) -> bool:
             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
+                # ndim 2 when image are compressed and 4 when uncompressed
+                assert data[f"/observations/images/{camera}"].ndim in [2, 4]
+                if 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, out_dir, fps, video, debug):
     # only frames from simulation are uncompressed
-    compressed_images = "sim" not in raw_dir.name
 
     hdf5_files = list(raw_dir.glob("*.hdf5"))
     ep_dicts = []
     episode_data_index = {"from": [], "to": []}
 
     id_from = 0
-
-    for ep_path in tqdm.tqdm(hdf5_files, total=len(hdf5_files)):
+    for ep_idx, ep_path in tqdm.tqdm(enumerate(hdf5_files), total=len(hdf5_files)):
         with h5py.File(ep_path, "r") as ep:
-            ep_idx = int(re.search(r"episode_(\d+)", ep_path.name).group(1))
             num_frames = ep["/action"].shape[0]
 
             # last step of demonstration is considered done
@@ -76,13 +84,17 @@ def load_from_raw(raw_dir, out_dir, fps, video, debug):
 
             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:
+                if ep[f"/observations/images/{camera}"].ndim == 2:
                     import cv2
 
                     # load one compressed image after the other in RAM and uncompress
@@ -116,6 +128,10 @@ def load_from_raw(raw_dir, out_dir, fps, video, debug):
                     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)
@@ -131,6 +147,8 @@ def load_from_raw(raw_dir, out_dir, fps, video, debug):
 
         id_from += num_frames
 
+        gc.collect()
+
         # process first episode only
         if debug:
             break
@@ -152,6 +170,14 @@ def to_hf_dataset(data_dict, video) -> Dataset:
     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)
     )

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -1,3 +1,18 @@
+#!/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

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -1,3 +1,18 @@
+#!/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

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -1,3 +1,18 @@
+#!/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 concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 
@@ -6,19 +21,24 @@ import PIL
 import torch
 
 
-def concatenate_episodes(ep_dicts):
+def concatenate_episodes(ep_dicts, drop_episodes_last_frame=False):
     data_dict = {}
 
     keys = ep_dicts[0].keys()
     for key in keys:
         if torch.is_tensor(ep_dicts[0][key][0]):
-            data_dict[key] = torch.cat([ep_dict[key] for ep_dict in ep_dicts])
+            if drop_episodes_last_frame:
+                data_dict[key] = torch.cat([ep_dict[key][:-1] for ep_dict in ep_dicts])
+            else:
+                data_dict[key] = torch.cat([ep_dict[key] for ep_dict in ep_dicts])
         else:
             if key not in data_dict:
                 data_dict[key] = []
             for ep_dict in ep_dicts:
                 for x in ep_dict[key]:
                     data_dict[key].append(x)
+                if drop_episodes_last_frame:
+                    data_dict[key].pop()
 
     total_frames = data_dict["frame_index"].shape[0]
     data_dict["index"] = torch.arange(0, total_frames, 1)

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -1,3 +1,18 @@
+#!/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

lerobot/common/datasets/sampler.py

@@ -0,0 +1,61 @@
+#!/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 typing import Iterator, Union
+
+import torch
+
+
+class EpisodeAwareSampler:
+    def __init__(
+        self,
+        episode_data_index: dict,
+        episode_indices_to_use: Union[list, None] = None,
+        drop_n_first_frames: int = 0,
+        drop_n_last_frames: int = 0,
+        shuffle: bool = False,
+    ):
+        """Sampler that optionally incorporates episode boundary information.
+
+        Args:
+            episode_data_index: Dictionary with keys 'from' and 'to' containing the start and end indices of each episode.
+            episode_indices_to_use: List of episode indices to use. If None, all episodes are used.
+                                    Assumes that episodes are indexed from 0 to N-1.
+            drop_n_first_frames: Number of frames to drop from the start of each episode.
+            drop_n_last_frames: Number of frames to drop from the end of each episode.
+            shuffle: Whether to shuffle the indices.
+        """
+        indices = []
+        for episode_idx, (start_index, end_index) in enumerate(
+            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
+        ):
+            if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
+                indices.extend(
+                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
+                )
+
+        self.indices = indices
+        self.shuffle = shuffle
+
+    def __iter__(self) -> Iterator[int]:
+        if self.shuffle:
+            for i in torch.randperm(len(self.indices)):
+                yield self.indices[i]
+        else:
+            for i in self.indices:
+                yield i
+
+    def __len__(self) -> int:
+        return len(self.indices)

lerobot/common/datasets/utils.py

@@ -1,5 +1,22 @@
+#!/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 json
+import re
 from pathlib import Path
+from typing import Dict
 
 import datasets
 import torch
@@ -42,7 +59,7 @@ def unflatten_dict(d, sep="/"):
     return outdict
 
 
-def hf_transform_to_torch(items_dict):
+def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
     a channel last representation (h w c) of uint8 type, to a torch image representation
@@ -56,6 +73,8 @@ def hf_transform_to_torch(items_dict):
         elif isinstance(first_item, dict) and "path" in first_item and "timestamp" in first_item:
             # video frame will be processed downstream
             pass
+        elif first_item is None:
+            pass
         else:
             items_dict[key] = [torch.tensor(x) for x in items_dict[key]]
     return items_dict
@@ -64,7 +83,23 @@ def hf_transform_to_torch(items_dict):
 def load_hf_dataset(repo_id, version, root, split) -> datasets.Dataset:
     """hf_dataset contains all the observations, states, actions, rewards, etc."""
     if root is not None:
-        hf_dataset = load_from_disk(str(Path(root) / repo_id / split))
+        hf_dataset = load_from_disk(str(Path(root) / repo_id / "train"))
+        # TODO(rcadene): clean this which enables getting a subset of dataset
+        if split != "train":
+            if "%" in split:
+                raise NotImplementedError(f"We dont support splitting based on percentage for now ({split}).")
+            match_from = re.search(r"train\[(\d+):\]", split)
+            match_to = re.search(r"train\[:(\d+)\]", split)
+            if match_from:
+                from_frame_index = int(match_from.group(1))
+                hf_dataset = hf_dataset.select(range(from_frame_index, len(hf_dataset)))
+            elif match_to:
+                to_frame_index = int(match_to.group(1))
+                hf_dataset = hf_dataset.select(range(to_frame_index))
+            else:
+                raise ValueError(
+                    f'`split` ({split}) should either be "train", "train[INT:]", or "train[:INT]"'
+                )
     else:
         hf_dataset = load_dataset(repo_id, revision=version, split=split)
     hf_dataset.set_transform(hf_transform_to_torch)
@@ -230,6 +265,84 @@ def load_previous_and_future_frames(
     return item
 
 
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
+    """
+    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
+
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
+
+    Returns:
+    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
+        - "from": A tensor containing the starting index of each episode.
+        - "to": A tensor containing the ending index of each episode.
+    """
+    episode_data_index = {"from": [], "to": []}
+
+    current_episode = None
+    """
+    The episode_index is a list of integers, each representing the episode index of the corresponding example.
+    For instance, the following is a valid episode_index:
+      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
+
+    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
+    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
+        {
+            "from": [0, 3, 7],
+            "to": [3, 7, 12]
+        }
+    """
+    if len(hf_dataset) == 0:
+        episode_data_index = {
+            "from": torch.tensor([]),
+            "to": torch.tensor([]),
+        }
+        return episode_data_index
+    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
+        if episode_idx != current_episode:
+            # We encountered a new episode, so we append its starting location to the "from" list
+            episode_data_index["from"].append(idx)
+            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
+            if current_episode is not None:
+                episode_data_index["to"].append(idx)
+            # Let's keep track of the current episode index
+            current_episode = episode_idx
+        else:
+            # We are still in the same episode, so there is nothing for us to do here
+            pass
+    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
+    episode_data_index["to"].append(idx + 1)
+
+    for k in ["from", "to"]:
+        episode_data_index[k] = torch.tensor(episode_data_index[k])
+
+    return episode_data_index
+
+
+def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
+    """Reset the `episode_index` of the provided HuggingFace Dataset.
+
+    `episode_data_index` (and related functionality such as `load_previous_and_future_frames`) requires the
+    `episode_index` to be sorted, continuous (1,1,1 and not 1,2,1) and start at 0.
+
+    This brings the `episode_index` to the required format.
+    """
+    if len(hf_dataset) == 0:
+        return hf_dataset
+    unique_episode_idxs = torch.stack(hf_dataset["episode_index"]).unique().tolist()
+    episode_idx_to_reset_idx_mapping = {
+        ep_id: reset_ep_id for reset_ep_id, ep_id in enumerate(unique_episode_idxs)
+    }
+
+    def modify_ep_idx_func(example):
+        example["episode_index"] = episode_idx_to_reset_idx_mapping[example["episode_index"].item()]
+        return example
+
+    hf_dataset = hf_dataset.map(modify_ep_idx_func)
+
+    return hf_dataset
+
+
 def cycle(iterable):
     """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
 

lerobot/common/datasets/video_utils.py

@@ -1,3 +1,18 @@
+#!/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 logging
 import subprocess
 import warnings

lerobot/common/envs/factory.py

@@ -1,3 +1,18 @@
+#!/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 importlib
 
 import gymnasium as gym
@@ -12,14 +27,6 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
     if n_envs is not None and n_envs < 1:
         raise ValueError("`n_envs must be at least 1")
 
-    kwargs = {
-        "obs_type": "pixels_agent_pos",
-        "render_mode": "rgb_array",
-        "max_episode_steps": cfg.env.episode_length,
-        "visualization_width": 384,
-        "visualization_height": 384,
-    }
-
     package_name = f"gym_{cfg.env.name}"
 
     try:
@@ -31,12 +38,16 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
         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
 
     # 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 = env_cls(
         [
-            lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
+            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)
         ]
     )

lerobot/common/envs/utils.py

@@ -1,3 +1,18 @@
+#!/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 einops
 import numpy as np
 import torch
@@ -14,10 +29,12 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     # map to expected inputs for the policy
     return_observations = {}
 
-    if isinstance(observations["pixels"], dict):
+    if "pixels" in observations and isinstance(observations["pixels"], dict):
         imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
-    else:
+    elif "pixels" in observations and isinstance(observations["pixels"], np.ndarray):
         imgs = {"observation.image": observations["pixels"]}
+    else:
+        imgs = {f"observation.{key}": img for key, img in observations.items() if "images" in key}
 
     for imgkey, img in imgs.items():
         img = torch.from_numpy(img)

lerobot/common/logger.py

@@ -1,22 +1,45 @@
+#!/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
-"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py"""
+"""
 
 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 OmegaConf
+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, return_list=False):
+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}",
@@ -27,22 +50,54 @@ def cfg_to_group(cfg, return_list=False):
     return lst if return_list else "-".join(lst)
 
 
-class Logger:
-    """Primary logger object. Logs either locally or using wandb."""
+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
 
-    def __init__(self, log_dir, job_name, cfg):
-        self._log_dir = Path(log_dir)
-        self._log_dir.mkdir(parents=True, exist_ok=True)
-        self._job_name = job_name
-        self._model_dir = self._log_dir / "checkpoints"
-        self._buffer_dir = self._log_dir / "buffers"
-        self._save_model = cfg.training.save_model
-        self._disable_wandb_artifact = cfg.wandb.disable_artifact
-        self._save_buffer = cfg.training.get("save_buffer", False)
-        self._group = cfg_to_group(cfg)
-        self._seed = cfg.seed
+
+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._eval = []
+        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)
@@ -54,70 +109,134 @@ class Logger:
             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=job_name,
+                name=wandb_job_name,
                 notes=cfg.get("wandb", {}).get("notes"),
-                # group=self._group,
                 tags=cfg_to_group(cfg, return_list=True),
-                dir=self._log_dir,
+                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",
-                # TODO(rcadene): add resume option
-                resume=None,
+                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
 
-    def save_model(self, policy: Policy, identifier):
-        if self._save_model:
-            self._model_dir.mkdir(parents=True, exist_ok=True)
-            save_dir = self._model_dir / str(identifier)
-            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._disable_wandb_artifact:
-                # note wandb artifact does not accept ":" in its name
-                artifact = self._wandb.Artifact(
-                    self._group.replace(":", "_").replace("/", "__")
-                    + "-"
-                    + str(self._seed)
-                    + "-"
-                    + str(identifier),
-                    type="model",
-                )
-                artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
-                self._wandb.log_artifact(artifact)
+    @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"
 
-    def save_buffer(self, buffer, identifier):
-        self._buffer_dir.mkdir(parents=True, exist_ok=True)
-        fp = self._buffer_dir / f"{str(identifier)}.pkl"
-        buffer.save(fp)
-        if self._wandb:
-            artifact = self._wandb.Artifact(
-                self._group + "-" + str(self._seed) + "-" + str(identifier),
-                type="buffer",
-            )
-            artifact.add_file(fp)
+    @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 finish(self, agent, buffer):
-        if self._save_model:
-            self.save_model(agent, identifier="final")
-        if self._save_buffer:
-            self.save_buffer(buffer, identifier="buffer")
-        if self._wandb:
-            self._wandb.finish()
+    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_checkpont(
+        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"):

lerobot/common/policies/act/configuration_act.py

@@ -1,3 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and 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 dataclasses import dataclass, field
 
 
@@ -10,6 +25,13 @@ class ACTConfig:
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
     Those are: `input_shapes` and 'output_shapes`.
 
+    Notes on the inputs and outputs:
+        - At least one key starting with "observation.image is required as an input.
+        - If there are multiple keys beginning with "observation.images." they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - May optionally work without an "observation.state" key for the proprioceptive robot state.
+        - "action" is required as an output key.
+
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
@@ -18,15 +40,15 @@ class ACTConfig:
             This should be no greater than the chunk size. For example, if the chunk size size 100, you may
             set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
             environment, and throws the other 50 out.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            The key represents the input data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "observation.images.top" refers to an input from the
-            "top" camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, shapes doesn't include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-            The key represents the output data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "action" refers to an output shape of [14], indicating
-            14-dimensional actions. Importantly, shapes doesn't include batch dimension or temporal dimension.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
         input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
             and the value specifies the normalization mode to apply. The two available modes are "mean_std"
             which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
@@ -51,8 +73,12 @@ class ACTConfig:
             documentation in the policy class).
         latent_dim: The VAE's latent dimension.
         n_vae_encoder_layers: The number of transformer layers to use for the VAE's encoder.
-        use_temporal_aggregation: Whether to blend the actions of multiple policy invocations for any given
-            environment step.
+        temporal_ensemble_momentum: Exponential moving average (EMA) momentum parameter (α) for ensembling
+            actions for a given time step over multiple policy invocations. Updates are calculated as:
+            x⁻ₙ = αx⁻ₙ₋₁ + (1-α)xₙ. Note that the ACT paper and original ACT code describes a different
+            parameter here: they refer to a weighting scheme wᵢ = exp(-m⋅i) and set m = 0.01. With our
+            formulation, this is equivalent to α = exp(-0.01) ≈ 0.99. When this parameter is provided, we
+            require `n_action_steps == 1` (since we need to query the policy every step anyway).
         dropout: Dropout to use in the transformer layers (see code for details).
         kl_weight: The weight to use for the KL-divergence component of the loss if the variational objective
             is enabled. Loss is then calculated as: `reconstruction_loss + kl_weight * kld_loss`.
@@ -100,14 +126,19 @@ class ACTConfig:
     dim_feedforward: int = 3200
     feedforward_activation: str = "relu"
     n_encoder_layers: int = 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.
+    # As a consequence we also remove the final, unused layer normalization, by default
     n_decoder_layers: int = 1
+    decoder_norm: bool = False
     # VAE.
     use_vae: bool = True
     latent_dim: int = 32
     n_vae_encoder_layers: int = 4
 
     # Inference.
-    use_temporal_aggregation: bool = False
+    temporal_ensemble_momentum: float | None = None
 
     # Training and loss computation.
     dropout: float = 0.1
@@ -119,8 +150,11 @@ class ACTConfig:
             raise ValueError(
                 f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
             )
-        if self.use_temporal_aggregation:
-            raise NotImplementedError("Temporal aggregation is not yet implemented.")
+        if self.temporal_ensemble_momentum is not None and self.n_action_steps > 1:
+            raise NotImplementedError(
+                "`n_action_steps` must be 1 when using temporal ensembling. This is "
+                "because the policy needs to be queried every step to compute the ensembled action."
+            )
         if self.n_action_steps > self.chunk_size:
             raise ValueError(
                 f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
@@ -130,10 +164,3 @@ class ACTConfig:
             raise ValueError(
                 f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
             )
-        # Check that there is only one image.
-        # TODO(alexander-soare): generalize this to multiple images.
-        if (
-            sum(k.startswith("observation.images.") for k in self.input_shapes) != 1
-            or "observation.images.top" not in self.input_shapes
-        ):
-            raise ValueError('For now, only "observation.images.top" is accepted for an image input.')

lerobot/common/policies/act/modeling_act.py

@@ -1,3 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and 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.
 """Action Chunking Transformer Policy
 
 As per Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (https://arxiv.org/abs/2304.13705).
@@ -46,7 +61,8 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
         super().__init__()
         if config is None:
             config = ACTConfig()
-        self.config = config
+        self.config: ACTConfig = config
+
         self.normalize_inputs = Normalize(
             config.input_shapes, config.input_normalization_modes, dataset_stats
         )
@@ -56,11 +72,18 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
         self.unnormalize_outputs = Unnormalize(
             config.output_shapes, config.output_normalization_modes, dataset_stats
         )
+
         self.model = ACT(config)
 
+        self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+
+        self.reset()
+
     def reset(self):
         """This should be called whenever the environment is reset."""
-        if self.config.n_action_steps is not None:
+        if self.config.temporal_ensemble_momentum is not None:
+            self._ensembled_actions = None
+        else:
             self._action_queue = deque([], maxlen=self.config.n_action_steps)
 
     @torch.no_grad
@@ -71,37 +94,56 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
         environment. It works by managing the actions in a queue and only calling `select_actions` when the
         queue is empty.
         """
-        assert "observation.images.top" in batch
-        assert "observation.state" in batch
-
         self.eval()
 
         batch = self.normalize_inputs(batch)
-        self._stack_images(batch)
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
 
+        # If we are doing temporal ensembling, keep track of the exponential moving average (EMA), and return
+        # the first action.
+        if self.config.temporal_ensemble_momentum is not None:
+            actions = self.model(batch)[0]  # (batch_size, chunk_size, action_dim)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+            if self._ensembled_actions is None:
+                # Initializes `self._ensembled_action` to the sequence of actions predicted during the first
+                # time step of the episode.
+                self._ensembled_actions = actions.clone()
+            else:
+                # self._ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
+                # the EMA update for those entries.
+                alpha = self.config.temporal_ensemble_momentum
+                self._ensembled_actions = alpha * self._ensembled_actions + (1 - alpha) * actions[:, :-1]
+                # The last action, which has no prior moving average, needs to get concatenated onto the end.
+                self._ensembled_actions = torch.cat([self._ensembled_actions, actions[:, -1:]], dim=1)
+            # "Consume" the first action.
+            action, self._ensembled_actions = self._ensembled_actions[:, 0], self._ensembled_actions[:, 1:]
+            return action
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
         if len(self._action_queue) == 0:
-            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
-            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
-            actions = self.model(batch)[0][: self.config.n_action_steps]
+            actions = self.model(batch)[0][:, : self.config.n_action_steps]
 
             # TODO(rcadene): make _forward return output dictionary?
             actions = self.unnormalize_outputs({"action": actions})["action"]
 
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
             self._action_queue.extend(actions.transpose(0, 1))
         return self._action_queue.popleft()
 
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
-        self._stack_images(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
         l1_loss = (
             F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
         ).mean()
 
-        loss_dict = {"l1_loss": l1_loss}
+        loss_dict = {"l1_loss": l1_loss.item()}
         if self.config.use_vae:
             # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
             # each dimension independently, we sum over the latent dimension to get the total
@@ -110,28 +152,13 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
             mean_kld = (
                 (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
             )
-            loss_dict["kld_loss"] = mean_kld
+            loss_dict["kld_loss"] = mean_kld.item()
             loss_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
         else:
             loss_dict["loss"] = l1_loss
 
         return loss_dict
 
-    def _stack_images(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        """Stacks all the images in a batch and puts them in a new key: "observation.images".
-
-        This function expects `batch` to have (at least):
-        {
-            "observation.state": (B, state_dim) batch of robot states.
-            "observation.images.{name}": (B, C, H, W) tensor of images.
-        }
-        """
-        # Stack images in the order dictated by input_shapes.
-        batch["observation.images"] = torch.stack(
-            [batch[k] for k in self.config.input_shapes if k.startswith("observation.images.")],
-            dim=-4,
-        )
-
 
 class ACT(nn.Module):
     """Action Chunking Transformer: The underlying neural network for ACTPolicy.
@@ -161,37 +188,41 @@ class ACT(nn.Module):
               │ encoder │ │     │ │Transf.│             │
               │         │ │     │ │encoder│             │
               └───▲─────┘ │     │ │       │             │
-                  │       │     │ └───▲───┘             │
-                  │       │     │     │                 │
-                inputs    └─────┼─────┘                 │
-                                │                       │
+                  │       │     │ └▲──▲─▲─┘             │
+                  │       │     │  │  │ │               │
+                inputs    └─────┼──┘  │ image emb.      │
+                                │    state emb.         │
                                 └───────────────────────┘
     """
 
     def __init__(self, config: ACTConfig):
         super().__init__()
         self.config = config
-        # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
+        # BERT style VAE encoder with input tokens [cls, robot_state, *action_sequence].
         # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
+        self.use_input_state = "observation.state" in config.input_shapes
         if self.config.use_vae:
             self.vae_encoder = ACTEncoder(config)
             self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
             # Projection layer for joint-space configuration to hidden dimension.
-            self.vae_encoder_robot_state_input_proj = nn.Linear(
-                config.input_shapes["observation.state"][0], config.dim_model
-            )
+            if self.use_input_state:
+                self.vae_encoder_robot_state_input_proj = nn.Linear(
+                    config.input_shapes["observation.state"][0], config.dim_model
+                )
             # Projection layer for action (joint-space target) to hidden dimension.
             self.vae_encoder_action_input_proj = nn.Linear(
-                config.input_shapes["observation.state"][0], config.dim_model
+                config.output_shapes["action"][0], config.dim_model
             )
-            self.latent_dim = config.latent_dim
             # Projection layer from the VAE encoder's output to the latent distribution's parameter space.
-            self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, self.latent_dim * 2)
-            # Fixed sinusoidal positional embedding the whole input to the VAE encoder. Unsqueeze for batch
+            self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
+            # Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
             # dimension.
+            num_input_token_encoder = 1 + config.chunk_size
+            if self.use_input_state:
+                num_input_token_encoder += 1
             self.register_buffer(
                 "vae_encoder_pos_enc",
-                create_sinusoidal_pos_embedding(1 + 1 + config.chunk_size, config.dim_model).unsqueeze(0),
+                create_sinusoidal_pos_embedding(num_input_token_encoder, config.dim_model).unsqueeze(0),
             )
 
         # Backbone for image feature extraction.
@@ -211,15 +242,17 @@ class ACT(nn.Module):
 
         # Transformer encoder input projections. The tokens will be structured like
         # [latent, robot_state, image_feature_map_pixels].
-        self.encoder_robot_state_input_proj = nn.Linear(
-            config.input_shapes["observation.state"][0], config.dim_model
-        )
-        self.encoder_latent_input_proj = nn.Linear(self.latent_dim, config.dim_model)
+        if self.use_input_state:
+            self.encoder_robot_state_input_proj = nn.Linear(
+                config.input_shapes["observation.state"][0], config.dim_model
+            )
+        self.encoder_latent_input_proj = nn.Linear(config.latent_dim, config.dim_model)
         self.encoder_img_feat_input_proj = nn.Conv2d(
             backbone_model.fc.in_features, config.dim_model, kernel_size=1
         )
         # Transformer encoder positional embeddings.
-        self.encoder_robot_and_latent_pos_embed = nn.Embedding(2, config.dim_model)
+        num_input_token_decoder = 2 if self.use_input_state else 1
+        self.encoder_robot_and_latent_pos_embed = nn.Embedding(num_input_token_decoder, config.dim_model)
         self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
 
         # Transformer decoder.
@@ -258,7 +291,7 @@ class ACT(nn.Module):
                 "action" in batch
             ), "actions must be provided when using the variational objective in training mode."
 
-        batch_size = batch["observation.state"].shape[0]
+        batch_size = batch["observation.images"].shape[0]
 
         # Prepare the latent for input to the transformer encoder.
         if self.config.use_vae and "action" in batch:
@@ -266,31 +299,43 @@ class ACT(nn.Module):
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
             )  # (B, 1, D)
-            robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"]).unsqueeze(
-                1
-            )  # (B, 1, D)
+            if self.use_input_state:
+                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
+                robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
             action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
-            vae_encoder_input = torch.cat([cls_embed, robot_state_embed, action_embed], axis=1)  # (B, S+2, D)
+
+            if self.use_input_state:
+                vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
+            else:
+                vae_encoder_input = [cls_embed, action_embed]
+            vae_encoder_input = torch.cat(vae_encoder_input, axis=1)
 
             # Prepare fixed positional embedding.
             # Note: detach() shouldn't be necessary but leaving it the same as the original code just in case.
             pos_embed = self.vae_encoder_pos_enc.clone().detach()  # (1, S+2, D)
 
             # Forward pass through VAE encoder to get the latent PDF parameters.
+            cls_joint_is_pad = torch.full((batch_size, 2), False).to(
+                batch["observation.state"].device
+            )  # False: not a padding
+            key_padding_mask = torch.cat([cls_joint_is_pad, batch["action_is_pad"]], axis=1)  # (bs, seq+1)
             cls_token_out = self.vae_encoder(
-                vae_encoder_input.permute(1, 0, 2), pos_embed=pos_embed.permute(1, 0, 2)
+                vae_encoder_input.permute(1, 0, 2),
+                pos_embed=pos_embed.permute(1, 0, 2),
+                key_padding_mask=key_padding_mask,
             )[0]  # select the class token, with shape (B, D)
             latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out)
-            mu = latent_pdf_params[:, : self.latent_dim]
+            mu = latent_pdf_params[:, : self.config.latent_dim]
             # This is 2log(sigma). Done this way to match the original implementation.
-            log_sigma_x2 = latent_pdf_params[:, self.latent_dim :]
+            log_sigma_x2 = latent_pdf_params[:, self.config.latent_dim :]
 
             # Sample the latent with the reparameterization trick.
             latent_sample = mu + log_sigma_x2.div(2).exp() * torch.randn_like(mu)
         else:
             # When not using the VAE encoder, we set the latent to be all zeros.
             mu = log_sigma_x2 = None
-            latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
+            # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
+            latent_sample = torch.zeros([batch_size, self.config.latent_dim], dtype=torch.float32).to(
                 batch["observation.state"].device
             )
 
@@ -299,25 +344,29 @@ class ACT(nn.Module):
         all_cam_features = []
         all_cam_pos_embeds = []
         images = batch["observation.images"]
+
         for cam_index in range(images.shape[-4]):
             cam_features = self.backbone(images[:, cam_index])["feature_map"]
+            # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
             cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
             cam_features = self.encoder_img_feat_input_proj(cam_features)  # (B, C, h, w)
             all_cam_features.append(cam_features)
             all_cam_pos_embeds.append(cam_pos_embed)
         # Concatenate camera observation feature maps and positional embeddings along the width dimension.
-        encoder_in = torch.cat(all_cam_features, axis=3)
-        cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=3)
+        encoder_in = torch.cat(all_cam_features, axis=-1)
+        cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=-1)
 
         # Get positional embeddings for robot state and latent.
-        robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])
-        latent_embed = self.encoder_latent_input_proj(latent_sample)
+        if self.use_input_state:
+            robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])  # (B, C)
+        latent_embed = self.encoder_latent_input_proj(latent_sample)  # (B, C)
 
         # Stack encoder input and positional embeddings moving to (S, B, C).
+        encoder_in_feats = [latent_embed, robot_state_embed] if self.use_input_state else [latent_embed]
         encoder_in = torch.cat(
             [
-                torch.stack([latent_embed, robot_state_embed], axis=0),
-                encoder_in.flatten(2).permute(2, 0, 1),
+                torch.stack(encoder_in_feats, axis=0),
+                einops.rearrange(encoder_in, "b c h w -> (h w) b c"),
             ]
         )
         pos_embed = torch.cat(
@@ -330,6 +379,7 @@ class ACT(nn.Module):
 
         # Forward pass through the transformer modules.
         encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
+        # TODO(rcadene, alexander-soare): remove call to `device` ; precompute and use buffer
         decoder_in = torch.zeros(
             (self.config.chunk_size, batch_size, self.config.dim_model),
             dtype=pos_embed.dtype,
@@ -358,9 +408,11 @@ class ACTEncoder(nn.Module):
         self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(config.n_encoder_layers)])
         self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
 
-    def forward(self, x: Tensor, pos_embed: Tensor | None = None) -> Tensor:
+    def forward(
+        self, x: Tensor, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
+    ) -> Tensor:
         for layer in self.layers:
-            x = layer(x, pos_embed=pos_embed)
+            x = layer(x, pos_embed=pos_embed, key_padding_mask=key_padding_mask)
         x = self.norm(x)
         return x
 
@@ -383,12 +435,14 @@ class ACTEncoderLayer(nn.Module):
         self.activation = get_activation_fn(config.feedforward_activation)
         self.pre_norm = config.pre_norm
 
-    def forward(self, x, pos_embed: Tensor | None = None) -> Tensor:
+    def forward(self, x, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None) -> Tensor:
         skip = x
         if self.pre_norm:
             x = self.norm1(x)
         q = k = x if pos_embed is None else x + pos_embed
-        x = self.self_attn(q, k, value=x)[0]  # select just the output, not the attention weights
+        x = self.self_attn(q, k, value=x, key_padding_mask=key_padding_mask)[
+            0
+        ]  # select just the output, not the attention weights
         x = skip + self.dropout1(x)
         if self.pre_norm:
             skip = x
@@ -408,7 +462,10 @@ class ACTDecoder(nn.Module):
         """Convenience module for running multiple decoder layers followed by normalization."""
         super().__init__()
         self.layers = nn.ModuleList([ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)])
-        self.norm = nn.LayerNorm(config.dim_model)
+        if config.decoder_norm:
+            self.norm = nn.LayerNorm(config.dim_model)
+        else:
+            self.norm = nn.Identity()
 
     def forward(
         self,
@@ -421,8 +478,7 @@ class ACTDecoder(nn.Module):
             x = layer(
                 x, encoder_out, decoder_pos_embed=decoder_pos_embed, encoder_pos_embed=encoder_pos_embed
             )
-        if self.norm is not None:
-            x = self.norm(x)
+        x = self.norm(x)
         return x
 
 

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -1,3 +1,19 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and 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 dataclasses import dataclass, field
 
 
@@ -10,21 +26,26 @@ class DiffusionConfig:
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
     Those are: `input_shapes` and `output_shapes`.
 
+    Notes on the inputs and outputs:
+        - "observation.state" is required as an input key.
+        - A key starting with "observation.image is required as an input.
+        - "action" is required as an output key.
+
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            The key represents the input data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "observation.image" refers to an input from
-            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, shapes doesnt include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-            The key represents the output data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "action" refers to an output shape of [14], indicating
-            14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
         input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
             and the value specifies the normalization mode to apply. The two available modes are "mean_std"
             which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
@@ -51,6 +72,7 @@ class DiffusionConfig:
         use_film_scale_modulation: FiLM (https://arxiv.org/abs/1709.07871) is used for the Unet conditioning.
             Bias modulation is used be default, while this parameter indicates whether to also use scale
             modulation.
+        noise_scheduler_type: Name of the noise scheduler to use. Supported options: ["DDPM", "DDIM"].
         num_train_timesteps: Number of diffusion steps for the forward diffusion schedule.
         beta_schedule: Name of the diffusion beta schedule as per DDPMScheduler from Hugging Face diffusers.
         beta_start: Beta value for the first forward-diffusion step.
@@ -110,6 +132,7 @@ class DiffusionConfig:
     diffusion_step_embed_dim: int = 128
     use_film_scale_modulation: bool = True
     # Noise scheduler.
+    noise_scheduler_type: str = "DDPM"
     num_train_timesteps: int = 100
     beta_schedule: str = "squaredcos_cap_v2"
     beta_start: float = 0.0001
@@ -130,17 +153,30 @@ class DiffusionConfig:
             raise ValueError(
                 f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
             )
-        if (
-            self.crop_shape[0] > self.input_shapes["observation.image"][1]
-            or self.crop_shape[1] > self.input_shapes["observation.image"][2]
+        # There should only be one image key.
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if len(image_keys) != 1:
+            raise ValueError(
+                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
+            )
+        image_key = next(iter(image_keys))
+        if self.crop_shape is not None and (
+            self.crop_shape[0] > self.input_shapes[image_key][1]
+            or self.crop_shape[1] > self.input_shapes[image_key][2]
         ):
             raise ValueError(
-                f'`crop_shape` should fit within `input_shapes["observation.image"]`. Got {self.crop_shape} '
-                f'for `crop_shape` and {self.input_shapes["observation.image"]} for '
-                '`input_shapes["observation.image"]`.'
+                f"`crop_shape` should fit within `input_shapes[{image_key}]`. Got {self.crop_shape} "
+                f"for `crop_shape` and {self.input_shapes[image_key]} for "
+                "`input_shapes[{image_key}]`."
             )
         supported_prediction_types = ["epsilon", "sample"]
         if self.prediction_type not in supported_prediction_types:
             raise ValueError(
                 f"`prediction_type` must be one of {supported_prediction_types}. Got {self.prediction_type}."
             )
+        supported_noise_schedulers = ["DDPM", "DDIM"]
+        if self.noise_scheduler_type not in supported_noise_schedulers:
+            raise ValueError(
+                f"`noise_scheduler_type` must be one of {supported_noise_schedulers}. "
+                f"Got {self.noise_scheduler_type}."
+            )

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -1,8 +1,24 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and 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.
 """Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
 
 TODO(alexander-soare):
-  - Remove reliance on Robomimic for SpatialSoftmax.
   - Remove reliance on diffusers for DDPMScheduler and LR scheduler.
+  - Make compatible with multiple image keys.
 """
 
 import math
@@ -10,12 +26,13 @@ from collections import deque
 from typing import Callable
 
 import einops
+import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
+from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from huggingface_hub import PyTorchModelHubMixin
-from robomimic.models.base_nets import SpatialSoftmax
 from torch import Tensor, nn
 
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
@@ -66,10 +83,18 @@ class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
 
         self.diffusion = DiffusionModel(config)
 
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: This check is covered in the post-init of the config but have a sanity check just in case.
+        if len(image_keys) != 1:
+            raise NotImplementedError(
+                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
+            )
+        self.input_image_key = image_keys[0]
+
+        self.reset()
+
     def reset(self):
-        """
-        Clear observation and action queues. Should be called on `env.reset()`
-        """
+        """Clear observation and action queues. Should be called on `env.reset()`"""
         self._queues = {
             "observation.image": deque(maxlen=self.config.n_obs_steps),
             "observation.state": deque(maxlen=self.config.n_obs_steps),
@@ -98,16 +123,14 @@ class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
         "horizon" may not the best name to describe what the variable actually means, because this period is
         actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
         """
-        assert "observation.image" in batch
-        assert "observation.state" in batch
-
         batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
 
         self._queues = populate_queues(self._queues, batch)
 
         if len(self._queues["action"]) == 0:
             # stack n latest observations from the queue
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
+            batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
             actions = self.diffusion.generate_actions(batch)
 
             # TODO(rcadene): make above methods return output dictionary?
@@ -121,11 +144,25 @@ class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
         batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
         return {"loss": loss}
 
 
+def _make_noise_scheduler(name: str, **kwargs: dict) -> DDPMScheduler | DDIMScheduler:
+    """
+    Factory for noise scheduler instances of the requested type. All kwargs are passed
+    to the scheduler.
+    """
+    if name == "DDPM":
+        return DDPMScheduler(**kwargs)
+    elif name == "DDIM":
+        return DDIMScheduler(**kwargs)
+    else:
+        raise ValueError(f"Unsupported noise scheduler type {name}")
+
+
 class DiffusionModel(nn.Module):
     def __init__(self, config: DiffusionConfig):
         super().__init__()
@@ -138,12 +175,12 @@ class DiffusionModel(nn.Module):
             * config.n_obs_steps,
         )
 
-        self.noise_scheduler = DDPMScheduler(
+        self.noise_scheduler = _make_noise_scheduler(
+            config.noise_scheduler_type,
             num_train_timesteps=config.num_train_timesteps,
             beta_start=config.beta_start,
             beta_end=config.beta_end,
             beta_schedule=config.beta_schedule,
-            variance_type="fixed_small",
             clip_sample=config.clip_sample,
             clip_sample_range=config.clip_sample_range,
             prediction_type=config.prediction_type,
@@ -185,13 +222,12 @@ class DiffusionModel(nn.Module):
 
     def generate_actions(self, batch: dict[str, Tensor]) -> Tensor:
         """
-        This function expects `batch` to have (at least):
+        This function expects `batch` to have:
         {
             "observation.state": (B, n_obs_steps, state_dim)
             "observation.image": (B, n_obs_steps, C, H, W)
         }
         """
-        assert set(batch).issuperset({"observation.state", "observation.image"})
         batch_size, n_obs_steps = batch["observation.state"].shape[:2]
         assert n_obs_steps == self.config.n_obs_steps
 
@@ -203,10 +239,8 @@ class DiffusionModel(nn.Module):
         global_cond = torch.cat([batch["observation.state"], img_features], dim=-1).flatten(start_dim=1)
 
         # run sampling
-        sample = self.conditional_sample(batch_size, global_cond=global_cond)
+        actions = self.conditional_sample(batch_size, global_cond=global_cond)
 
-        # `horizon` steps worth of actions (from the first observation).
-        actions = sample[..., : self.config.output_shapes["action"][0]]
         # Extract `n_action_steps` steps worth of actions (from the current observation).
         start = n_obs_steps - 1
         end = start + self.config.n_action_steps
@@ -268,13 +302,88 @@ class DiffusionModel(nn.Module):
         loss = F.mse_loss(pred, target, reduction="none")
 
         # Mask loss wherever the action is padded with copies (edges of the dataset trajectory).
-        if self.config.do_mask_loss_for_padding and "action_is_pad" in batch:
+        if self.config.do_mask_loss_for_padding:
+            if "action_is_pad" not in batch:
+                raise ValueError(
+                    f"You need to provide 'action_is_pad' in the batch when {self.config.do_mask_loss_for_padding=}."
+                )
             in_episode_bound = ~batch["action_is_pad"]
             loss = loss * in_episode_bound.unsqueeze(-1)
 
         return loss.mean()
 
 
+class SpatialSoftmax(nn.Module):
+    """
+    Spatial Soft Argmax operation described in "Deep Spatial Autoencoders for Visuomotor Learning" by Finn et al.
+    (https://arxiv.org/pdf/1509.06113). A minimal port of the robomimic implementation.
+
+    At a high level, this takes 2D feature maps (from a convnet/ViT) and returns the "center of mass"
+    of activations of each channel, i.e., keypoints in the image space for the policy to focus on.
+
+    Example: take feature maps of size (512x10x12). We generate a grid of normalized coordinates (10x12x2):
+    -----------------------------------------------------
+    | (-1., -1.)   | (-0.82, -1.)   | ... | (1., -1.)   |
+    | (-1., -0.78) | (-0.82, -0.78) | ... | (1., -0.78) |
+    | ...          | ...            | ... | ...         |
+    | (-1., 1.)    | (-0.82, 1.)    | ... | (1., 1.)    |
+    -----------------------------------------------------
+    This is achieved by applying channel-wise softmax over the activations (512x120) and computing the dot
+    product with the coordinates (120x2) to get expected points of maximal activation (512x2).
+
+    The example above results in 512 keypoints (corresponding to the 512 input channels). We can optionally
+    provide num_kp != None to control the number of keypoints. This is achieved by a first applying a learnable
+    linear mapping (in_channels, H, W) -> (num_kp, H, W).
+    """
+
+    def __init__(self, input_shape, num_kp=None):
+        """
+        Args:
+            input_shape (list): (C, H, W) input feature map shape.
+            num_kp (int): number of keypoints in output. If None, output will have the same number of channels as input.
+        """
+        super().__init__()
+
+        assert len(input_shape) == 3
+        self._in_c, self._in_h, self._in_w = input_shape
+
+        if num_kp is not None:
+            self.nets = torch.nn.Conv2d(self._in_c, num_kp, kernel_size=1)
+            self._out_c = num_kp
+        else:
+            self.nets = None
+            self._out_c = self._in_c
+
+        # we could use torch.linspace directly but that seems to behave slightly differently than numpy
+        # and causes a small degradation in pc_success of pre-trained models.
+        pos_x, pos_y = np.meshgrid(np.linspace(-1.0, 1.0, self._in_w), np.linspace(-1.0, 1.0, self._in_h))
+        pos_x = torch.from_numpy(pos_x.reshape(self._in_h * self._in_w, 1)).float()
+        pos_y = torch.from_numpy(pos_y.reshape(self._in_h * self._in_w, 1)).float()
+        # register as buffer so it's moved to the correct device.
+        self.register_buffer("pos_grid", torch.cat([pos_x, pos_y], dim=1))
+
+    def forward(self, features: Tensor) -> Tensor:
+        """
+        Args:
+            features: (B, C, H, W) input feature maps.
+        Returns:
+            (B, K, 2) image-space coordinates of keypoints.
+        """
+        if self.nets is not None:
+            features = self.nets(features)
+
+        # [B, K, H, W] -> [B * K, H * W] where K is number of keypoints
+        features = features.reshape(-1, self._in_h * self._in_w)
+        # 2d softmax normalization
+        attention = F.softmax(features, dim=-1)
+        # [B * K, H * W] x [H * W, 2] -> [B * K, 2] for spatial coordinate mean in x and y dimensions
+        expected_xy = attention @ self.pos_grid
+        # reshape to [B, K, 2]
+        feature_keypoints = expected_xy.view(-1, self._out_c, 2)
+
+        return feature_keypoints
+
+
 class DiffusionRgbEncoder(nn.Module):
     """Encoder an RGB image into a 1D feature vector.
 
@@ -315,11 +424,20 @@ class DiffusionRgbEncoder(nn.Module):
 
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
+        # The dummy input should take the number of image channels from `config.input_shapes` and it should
+        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
+        # height and width from `config.input_shapes`.
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        assert len(image_keys) == 1
+        image_key = image_keys[0]
+        dummy_input_h_w = (
+            config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
+        )
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
         with torch.inference_mode():
-            feat_map_shape = tuple(
-                self.backbone(torch.zeros(size=(1, *config.input_shapes["observation.image"]))).shape[1:]
-            )
-        self.pool = SpatialSoftmax(feat_map_shape, num_kp=config.spatial_softmax_num_keypoints)
+            dummy_feature_map = self.backbone(dummy_input)
+        feature_map_shape = tuple(dummy_feature_map.shape[1:])
+        self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
         self.feature_dim = config.spatial_softmax_num_keypoints * 2
         self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
         self.relu = nn.ReLU()

lerobot/common/policies/factory.py

@@ -1,4 +1,20 @@
+#!/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 inspect
+import logging
 
 from omegaconf import DictConfig, OmegaConf
 
@@ -8,9 +24,10 @@ from lerobot.common.utils.utils import get_safe_torch_device
 
 def _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg):
     expected_kwargs = set(inspect.signature(policy_cfg_class).parameters)
-    assert set(hydra_cfg.policy).issuperset(
-        expected_kwargs
-    ), f"Hydra config is missing arguments: {set(expected_kwargs).difference(hydra_cfg.policy)}"
+    if not set(hydra_cfg.policy).issuperset(expected_kwargs):
+        logging.warning(
+            f"Hydra config is missing arguments: {set(expected_kwargs).difference(hydra_cfg.policy)}"
+        )
     policy_cfg = policy_cfg_class(
         **{
             k: v
@@ -62,11 +79,18 @@ def make_policy(
 
     policy_cls, policy_cfg_class = get_policy_and_config_classes(hydra_cfg.policy.name)
 
+    policy_cfg = _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg)
     if pretrained_policy_name_or_path is None:
-        policy_cfg = _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg)
+        # Make a fresh policy.
         policy = policy_cls(policy_cfg, dataset_stats)
     else:
-        policy = policy_cls.from_pretrained(pretrained_policy_name_or_path)
+        # Load a pretrained policy and override the config if needed (for example, if there are inference-time
+        # hyperparameters that we want to vary).
+        # TODO(alexander-soare): This hack makes use of huggingface_hub's tooling to load the policy with, pretrained
+        # weights which are then loaded into a fresh policy with the desired config. This PR in huggingface_hub should
+        # make it possible to avoid the hack: https://github.com/huggingface/huggingface_hub/pull/2274.
+        policy = policy_cls(policy_cfg)
+        policy.load_state_dict(policy_cls.from_pretrained(pretrained_policy_name_or_path).state_dict())
 
     policy.to(get_safe_torch_device(hydra_cfg.device))
 

lerobot/common/policies/normalize.py

@@ -1,3 +1,18 @@
+#!/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 torch
 from torch import Tensor, nn
 
@@ -132,7 +147,7 @@ class Normalize(nn.Module):
                 assert not torch.isinf(min).any(), _no_stats_error_str("min")
                 assert not torch.isinf(max).any(), _no_stats_error_str("max")
                 # normalize to [0,1]
-                batch[key] = (batch[key] - min) / (max - min)
+                batch[key] = (batch[key] - min) / (max - min + 1e-8)
                 # normalize to [-1, 1]
                 batch[key] = batch[key] * 2 - 1
             else:

lerobot/common/policies/policy_protocol.py

@@ -1,3 +1,18 @@
+#!/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.
 """A protocol that all policies should follow.
 
 This provides a mechanism for type-hinting and isinstance checks without requiring the policies classes
@@ -38,7 +53,8 @@ class Policy(Protocol):
     def forward(self, batch: dict[str, Tensor]) -> dict:
         """Run the batch through the model and compute the loss for training or validation.
 
-        Returns a dictionary with "loss" and maybe other information.
+        Returns a dictionary with "loss" and potentially other information. Apart from "loss" which is a Tensor, all
+        other items should be logging-friendly, native Python types.
         """
 
     def select_action(self, batch: dict[str, Tensor]):

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -1,3 +1,19 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# and 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 dataclasses import dataclass, field
 
 
@@ -15,6 +31,15 @@ class TDMPCConfig:
         n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
             action repeats in Q-learning or ask your favorite chatbot)
         horizon: Horizon for model predictive control.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
         input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
             and the value specifies the normalization mode to apply. The two available modes are "mean_std"
             which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
@@ -47,7 +72,7 @@ class TDMPCConfig:
         elite_weighting_temperature: The temperature to use for softmax weighting (by trajectory value) of the
             elites, when updating the gaussian parameters for CEM.
         gaussian_mean_momentum: Momentum (α) used for EMA updates of the mean parameter μ of the gaussian
-            paramters optimized in CEM. Updates are calculated as μ⁻ ← αμ⁻ + (1-α)μ.
+            parameters optimized in CEM. Updates are calculated as μ⁻ ← αμ⁻ + (1-α)μ.
         max_random_shift_ratio: Maximum random shift (as a proportion of the image size) to apply to the
             image(s) (in units of pixels) for training-time augmentation. If set to 0, no such augmentation
             is applied. Note that the input images are assumed to be square for this augmentation.
@@ -131,12 +156,18 @@ class TDMPCConfig:
 
     def __post_init__(self):
         """Input validation (not exhaustive)."""
-        if self.input_shapes["observation.image"][-2] != self.input_shapes["observation.image"][-1]:
+        # There should only be one image key.
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if len(image_keys) != 1:
+            raise ValueError(
+                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
+            )
+        image_key = next(iter(image_keys))
+        if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
             # TODO(alexander-soare): This limitation is solely because of code in the random shift
             # augmentation. It should be able to be removed.
             raise ValueError(
-                "Only square images are handled now. Got image shape "
-                f"{self.input_shapes['observation.image']}."
+                f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
             )
         if self.n_gaussian_samples <= 0:
             raise ValueError(

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -1,3 +1,19 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# and 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.
 """Implementation of Finetuning Offline World Models in the Real World.
 
 The comments in this code may sometimes refer to these references:
@@ -96,13 +112,12 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             config.output_shapes, config.output_normalization_modes, dataset_stats
         )
 
-    def save(self, fp):
-        """Save state dict of TOLD model to filepath."""
-        torch.save(self.state_dict(), fp)
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: This check is covered in the post-init of the config but have a sanity check just in case.
+        assert len(image_keys) == 1
+        self.input_image_key = image_keys[0]
 
-    def load(self, fp):
-        """Load a saved state dict from filepath into current agent."""
-        self.load_state_dict(torch.load(fp))
+        self.reset()
 
     def reset(self):
         """
@@ -121,10 +136,8 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]):
         """Select a single action given environment observations."""
-        assert "observation.image" in batch
-        assert "observation.state" in batch
-
         batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
 
         self._queues = populate_queues(self._queues, batch)
 
@@ -303,13 +316,11 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         device = get_device_from_parameters(self)
 
         batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
         batch = self.normalize_targets(batch)
 
         info = {}
 
-        # TODO(alexander-soare): Refactor TDMPC and make it comply with the policy interface documentation.
-        batch_size = batch["index"].shape[0]
-
         # (b, t) -> (t, b)
         for key in batch:
             if batch[key].ndim > 1:
@@ -337,6 +348,7 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         # Run latent rollout using the latent dynamics model and policy model.
         # Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
         # gives us a next `z`.
+        batch_size = batch["index"].shape[0]
         z_preds = torch.empty(horizon + 1, batch_size, self.config.latent_dim, device=device)
         z_preds[0] = self.model.encode(current_observation)
         reward_preds = torch.empty_like(reward, device=device)

lerobot/common/policies/utils.py

@@ -1,9 +1,28 @@
+#!/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 torch
 from torch import nn
 
 
 def populate_queues(queues, batch):
     for key in batch:
+        # Ignore keys not in the queues already (leaving the responsibility to the caller to make sure the
+        # queues have the keys they want).
+        if key not in queues:
+            continue
         if len(queues[key]) != queues[key].maxlen:
             # initialize by copying the first observation several times until the queue is full
             while len(queues[key]) != queues[key].maxlen:

lerobot/common/robot_devices/cameras/intelrealsense.py

@@ -0,0 +1,325 @@
+
+
+import argparse
+from dataclasses import dataclass, replace
+from pathlib import Path
+from threading import Thread
+import time
+import traceback
+import cv2
+import numpy as np
+import pyrealsense2 as rs
+
+from lerobot.common.robot_devices.cameras.opencv import find_camera_indices
+from lerobot.common.robot_devices.cameras.utils import save_color_image, save_depth_image
+
+SERIAL_NUMBER_INDEX = 1
+
+def find_camera_indices(raise_when_empty=True):
+    camera_ids = []
+    for device in rs.context().query_devices():
+        serial_number = int(device.get_info(rs.camera_info(SERIAL_NUMBER_INDEX)))
+        camera_ids.append(serial_number)
+
+    if raise_when_empty and len(camera_ids) == 0:
+        raise OSError("Not a single camera was detected. Try re-plugging, or re-installing `librealsense` and its python wrapper `pyrealsense2`, or updating the firmware.")
+
+    return camera_ids
+
+def benchmark_cameras(cameras, out_dir=None, save_images=False):
+    if save_images:
+        out_dir = Path(out_dir)
+        out_dir.mkdir(parents=True, exist_ok=True)
+
+    while True:
+        now = time.time()
+        for camera in cameras:
+            if camera.use_depth:
+                color_image, depth_image = camera.capture_image("bgr" if save_images else "rgb")
+            else:
+                color_image = camera.capture_image("bgr" if save_images else "rgb")
+
+            if save_images:
+                image_path = out_dir / f"camera_{camera.camera_index:02}.png"
+                print(f"Write to {image_path}")
+                save_color_image(color_image, image_path, write_shape=True)
+
+                if camera.use_depth:
+                    # Apply colormap on depth image (image must be converted to 8-bit per pixel first)
+                    depth_image_path = out_dir / f"camera_{camera.camera_index:02}_depth.png"
+                    print(f"Write to {depth_image_path}")
+                    save_depth_image(depth_image_path, depth_image, write_shape=True)
+
+        dt_s = (time.time() - now)
+        dt_ms = dt_s * 1000
+        freq = 1 / dt_s
+        print(f"Latency (ms): {dt_ms:.2f}\tFrequency: {freq:.2f}")
+
+        if save_images:
+            break
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
+
+# Pre-defined configs that worked
+
+@dataclass
+class IntelRealSenseCameraConfig:
+    """
+    Example of tested options for Intel Real Sense D405:
+    
+    ```python
+    IntelRealSenseCameraConfig(30, 640, 480)
+    IntelRealSenseCameraConfig(60, 640, 480)
+    IntelRealSenseCameraConfig(90, 640, 480)
+    IntelRealSenseCameraConfig(30, 1280, 720)
+
+    IntelRealSenseCameraConfig(30, 640, 480, use_depth=True)
+    IntelRealSenseCameraConfig(60, 640, 480, use_depth=True)
+    IntelRealSenseCameraConfig(90, 640, 480, use_depth=True)
+    IntelRealSenseCameraConfig(30, 1280, 720, use_depth=True)
+    ```
+    """
+    fps: int | None = None
+    width: int | None = None
+    height: int | None = None
+    color: str = "rgb"
+    use_depth: bool = False
+    force_hardware_reset: bool = True
+
+
+
+class IntelRealSenseCamera():
+    # TODO(rcadene): improve dosctring
+    """
+    Using this class requires:
+    - [installing `librealsense` and its python wrapper `pyrealsense2`](https://github.com/IntelRealSense/librealsense/blob/master/doc/distribution_linux.md)
+    - [updating the camera(s) firmware](https://dev.intelrealsense.com/docs/firmware-releases-d400)
+
+    Example of getting the `camera_index` for your camera(s):
+    ```bash
+    rs-fw-update -l
+
+    > Connected devices:
+    > 1) [USB] Intel RealSense D405 s/n 128422270109, update serial number: 133323070634, firmware version: 5.16.0.1
+    > 2) [USB] Intel RealSense D405 s/n 128422271609, update serial number: 130523070758, firmware version: 5.16.0.1
+    > 3) [USB] Intel RealSense D405 s/n 128422271614, update serial number: 133323070576, firmware version: 5.16.0.1
+    > 4) [USB] Intel RealSense D405 s/n 128422271393, update serial number: 133323070271, firmware version: 5.16.0.1
+    ```
+
+    Example of uage:
+
+    ```python
+    camera = IntelRealSenseCamera(128422270109)  # serial number (s/n)
+    color_image = camera.capture_image()
+    ```
+
+    Example of capturing additional depth image:
+
+    ```python
+    config = IntelRealSenseCameraConfig(use_depth=True)
+    camera = IntelRealSenseCamera(128422270109, config)
+    color_image, depth_image = camera.capture_image()
+    ```
+    """
+    AVAILABLE_CAMERA_INDICES = find_camera_indices()
+
+    def __init__(self,
+            camera_index: int | None = None,
+            config: IntelRealSenseCameraConfig | None = None,
+            **kwargs,
+        ):
+        if config is None:
+            config = IntelRealSenseCameraConfig()
+        # Overwrite config arguments using kwargs
+        config = replace(config, **kwargs)
+
+        self.camera_index = camera_index
+        self.fps = config.fps
+        self.width = config.width
+        self.height = config.height
+        self.color = config.color
+        self.use_depth = config.use_depth
+        self.force_hardware_reset = config.force_hardware_reset
+
+        # TODO(rcadene): move these two check in config dataclass
+        if self.color not in ["rgb", "bgr"]:
+            raise ValueError(f"Expected color values are 'rgb' or 'bgr', but {self.color} is provided.")
+
+        if (self.fps or self.width or self.height) and not (self.fps and self.width and self.height):
+            raise ValueError(f"Expected all fps, width and height to be set, when one of them is set, but {self.fps=}, {self.width=}, {self.height=}.")
+
+        if self.camera_index is None:
+            raise ValueError(f"`camera_index` is expected to be a serial number of one of these available cameras ({IntelRealSenseCamera.AVAILABLE_CAMERA_INDICES}), but {camera_index} is provided instead.")
+
+        self.camera = None
+        self.is_connected = False
+
+        self.t = Thread(target=self.capture_image_loop, args=())
+        self.t.daemon = True
+        self._color_image = None
+
+    def connect(self):
+        if self.is_connected:
+            raise ValueError(f"Camera {self.camera_index} is already connected.")
+
+        config = rs.config()
+        config.enable_device(str(self.camera_index))
+
+        if self.fps and self.width and self.height:
+            # TODO(rcadene): can we set rgb8 directly?
+            config.enable_stream(rs.stream.color, self.width, self.height, rs.format.rgb8, self.fps)
+        else:
+            config.enable_stream(rs.stream.color)
+
+        if self.use_depth:
+            if self.fps and self.width and self.height:
+                config.enable_stream(rs.stream.depth, self.width, self.height, rs.format.z16, self.fps)
+            else:
+                config.enable_stream(rs.stream.depth)
+
+        self.camera = rs.pipeline()
+        try:
+            self.camera.start(config)
+        except RuntimeError:
+            # Verify that the provided `camera_index` is valid before printing the traceback
+            if self.camera_index not in IntelRealSenseCamera.AVAILABLE_CAMERA_INDICES:
+                raise ValueError(f"`camera_index` is expected to be a serial number of one of these available cameras {IntelRealSenseCamera.AVAILABLE_CAMERA_INDICES}, but {self.camera_index} is provided instead.")
+            traceback.print_exc()
+
+        self.is_connected = True
+        self.t.start()
+
+    def capture_image(self, temporary_color: str | None = None) -> np.ndarray | tuple[np.ndarray, np.ndarray]:
+        frame = self.camera.wait_for_frames()
+
+        color_frame = frame.get_color_frame()
+        if not color_frame:
+            raise OSError(f"Can't capture color image from camera {self.camera_index}.")
+        color_image = np.asanyarray(color_frame.get_data())
+
+        if temporary_color is None:
+            requested_color = self.color
+        else:
+            requested_color = temporary_color
+        
+        if requested_color not in ["rgb", "bgr"]:
+            raise ValueError(f"Expected color values are 'rgb' or 'bgr', but {requested_color} is provided.")
+        
+        # OpenCV uses BGR format as default (blue, green red) for all operations, including displaying images.
+        # However, Deep Learning framework such as LeRobot uses RGB format as default to train neural networks,
+        # so we convert the image color from BGR to RGB.
+        # if requested_color == "rgb":
+        #     color_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2RGB)
+
+        if self.use_depth:
+            depth_frame = frame.get_depth_frame()
+            if not depth_frame:
+                raise OSError(f"Can't capture depth image from camera {self.camera_index}.")
+            depth_image = np.asanyarray(depth_frame.get_data())
+
+            return color_image, depth_image
+        else:
+            return color_image
+
+    def capture_image_loop(self):
+        while True:
+            self._color_image = self.capture_image()
+
+    def read(self):
+        while self._color_image is None:
+            time.sleep(0.1)
+        return self._color_image
+
+    def disconnect(self):
+        if getattr(self, "camera", None):
+            try:
+                self.camera.stop()
+            except RuntimeError as e:
+                if "stop() cannot be called before start()" in str(e):
+                    # skip this runtime error
+                    return
+                traceback.print_exc()
+
+    def __del__(self):
+        self.disconnect()
+
+
+def save_images_config(config, out_dir: Path):
+    camera_ids = IntelRealSenseCamera.AVAILABLE_CAMERA_INDICES
+    cameras = []
+    print(f"Available camera indices: {camera_ids}")
+    for camera_idx in camera_ids:
+        camera = IntelRealSenseCamera(camera_idx, config)
+        cameras.append(camera)
+
+    out_dir = out_dir.parent / f"{out_dir.name}_{config.width}x{config.height}_{config.fps}_depth_{config.use_depth}"
+    benchmark_cameras(cameras, out_dir, save_images=True)
+
+def benchmark_config(config, camera_ids: list[int]):
+    cameras = [IntelRealSenseCamera(idx, config) for idx in camera_ids]
+    benchmark_cameras(cameras)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--mode", type=str, choices=["save_images", 'benchmark'], default="save_images")
+    parser.add_argument("--camera-ids", type=int, nargs="*", default=[128422271609, 128422271614, 128422271393])
+    parser.add_argument("--fps", type=int, default=30)
+    parser.add_argument("--width", type=str, default=640)
+    parser.add_argument("--height", type=str, default=480)
+    parser.add_argument("--use-depth", type=int, default=0)
+    parser.add_argument("--out-dir", type=Path, default="outputs/benchmark_cameras/intelrealsense/2024_06_22_1738")
+    args = parser.parse_args()
+
+    config = IntelRealSenseCameraConfig(args.fps, args.width, args.height, use_depth=bool(args.use_depth))
+    # config = IntelRealSenseCameraConfig()
+    # config = IntelRealSenseCameraConfig(60, 640, 480)
+    # config = IntelRealSenseCameraConfig(90, 640, 480)
+    # config = IntelRealSenseCameraConfig(30, 1280, 720)
+
+    if args.mode == "save_images":
+        save_images_config(config, args.out_dir)
+    elif args.mode == "benchmark":
+        benchmark_config(config, args.camera_ids)
+    else:
+        raise ValueError(args.mode)
+    
+
+# if __name__ == "__main__":
+#     # Works well!
+#     # use_depth = False
+#     # fps = 90
+#     # width = 640
+#     # height = 480
+
+#     # # Works well!
+#     # use_depth = True
+#     # fps = 90
+#     # width = 640
+#     # height = 480
+
+#     # # Doesn't work well, latency varies too much
+#     # use_depth = True
+#     # fps = 30
+#     # width = 1280
+#     # height = 720
+
+#     # Works well
+#     use_depth = False
+#     fps = 30
+#     width = 1280
+#     height = 720
+
+#     config = IntelRealSenseCameraConfig()
+#     # config = IntelRealSenseCameraConfig(fps, width, height, use_depth=use_depth)
+#     cameras = [
+#         # IntelRealSenseCamera(0, config),
+#         # IntelRealSenseCamera(128422270109, config),
+#         IntelRealSenseCamera(128422271609, config),
+#         IntelRealSenseCamera(128422271614, config),
+#         IntelRealSenseCamera(128422271393, config),
+#     ]
+
+#     out_dir = "outputs/benchmark_cameras/intelrealsense/2024_06_22_1729"
+#     out_dir += f"{config.width}x{config.height}_{config.fps}_depth_{config.use_depth}"
+#     benchmark_cameras(cameras, out_dir, save_images=False)

lerobot/common/robot_devices/cameras/opencv.py

@@ -0,0 +1,249 @@
+
+
+import argparse
+from dataclasses import dataclass, replace
+from pathlib import Path
+from threading import Thread
+import time
+import cv2
+import numpy as np
+
+from lerobot.common.robot_devices.cameras.utils import save_color_image
+
+
+def find_camera_indices(raise_when_empty=False, max_index_search_range=60):
+    camera_ids = []
+    for camera_idx in range(max_index_search_range):
+        camera = cv2.VideoCapture(camera_idx)
+        is_open = camera.isOpened()
+        camera.release()
+
+        if is_open:
+            print(f"Camera found at index {camera_idx}")
+            camera_ids.append(camera_idx)
+
+    if raise_when_empty and len(camera_ids) == 0:
+        raise OSError("Not a single camera was detected. Try re-plugging, or re-installing `opencv2`, or your camera driver, or make sure your camera is compatible with opencv2.")
+
+    return camera_ids
+
+def benchmark_cameras(cameras, out_dir=None, save_images=False, num_warmup_frames=4):
+    if out_dir:
+        out_dir = Path(out_dir)
+        out_dir.mkdir(parents=True, exist_ok=True)
+
+    for _ in range(num_warmup_frames):
+        for camera in cameras:
+            try:
+                camera.capture_image()
+                time.sleep(0.01)
+            except OSError as e:
+                print(e)
+
+    while True:
+        now = time.time()
+        for camera in cameras:
+            color_image = camera.capture_image("bgr" if save_images else "rgb")
+
+            if save_images:
+                image_path = out_dir / f"camera_{camera.camera_index:02}.png"
+                print(f"Write to {image_path}")
+                save_color_image(color_image, image_path, write_shape=True)
+
+        dt_s = (time.time() - now)
+        dt_ms = dt_s * 1000
+        freq = 1 / dt_s
+        print(f"Latency (ms): {dt_ms:.2f}\tFrequency: {freq:.2f}")
+
+        if save_images:
+            break
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
+
+
+@dataclass
+class OpenCVCameraConfig:
+    """
+    Example of tested options for Intel Real Sense D405:
+    
+    ```python
+    OpenCVCameraConfig(30, 640, 480)
+    OpenCVCameraConfig(60, 640, 480)
+    OpenCVCameraConfig(90, 640, 480)
+    OpenCVCameraConfig(30, 1280, 720)
+    ```
+    """
+    fps: int | None = None
+    width: int | None = None
+    height: int | None = None
+    color: str = "rgb"
+
+
+
+class OpenCVCamera():
+    # TODO(rcadene): improve dosctring
+    """
+    https://docs.opencv.org/4.x/d0/da7/videoio_overview.html
+    https://docs.opencv.org/4.x/d4/d15/group__videoio__flags__base.html#ga023786be1ee68a9105bf2e48c700294d
+
+    Example of uage:
+
+    ```python
+    camera = OpenCVCamera(2)
+    color_image = camera.capture_image()
+    ```
+    """
+    def __init__(self, camera_index: int, config: OpenCVCameraConfig | None = None, **kwargs):
+        if config is None:
+            config = OpenCVCameraConfig()
+        # Overwrite config arguments using kwargs
+        config = replace(config, **kwargs)
+
+        self.camera_index = camera_index
+        self.fps = config.fps
+        self.width = config.width
+        self.height = config.height
+        self.color = config.color
+
+        if self.color not in ["rgb", "bgr"]:
+            raise ValueError(f"Expected color values are 'rgb' or 'bgr', but {self.color} is provided.")
+
+        if self.camera_index is None:
+            raise ValueError(f"`camera_index` is expected to be one of these available cameras {OpenCVCamera.AVAILABLE_CAMERAS_INDICES}, but {camera_index} is provided instead.")
+        
+        self.camera = None
+        self.is_connected = False
+
+        self.t = Thread(target=self.capture_image_loop, args=())
+        self.t.daemon = True
+        self._color_image = None
+
+    def connect(self):
+        if self.is_connected:
+            raise ValueError(f"Camera {self.camera_index} is already connected.")
+
+        # First create a temporary camera trying to access `camera_index`,
+        # and verify it is a valid camera by calling `isOpened`.
+        tmp_camera = cv2.VideoCapture(self.camera_index)
+        is_camera_open = tmp_camera.isOpened()
+        # Release camera to make it accessible for `find_camera_indices`
+        del tmp_camera
+
+        # If the camera doesn't work, display the camera indices corresponding to
+        # valid cameras.
+        if not is_camera_open:
+            # Verify that the provided `camera_index` is valid before printing the traceback
+            if self.camera_index not in find_camera_indices():
+                raise ValueError(f"`camera_index` is expected to be one of these available cameras {OpenCVCamera.AVAILABLE_CAMERAS_INDICES}, but {self.camera_index} is provided instead.")
+
+            raise OSError(f"Can't access camera {self.camera_index}.")
+        
+        # Secondly, create the camera that will be used downstream.
+        # Note: For some unknown reason, calling `isOpened` blocks the camera which then
+        # needs to be re-created.
+        self.camera = cv2.VideoCapture(self.camera_index)
+
+        if self.fps:
+            self.camera.set(cv2.CAP_PROP_FPS, self.fps)
+        if self.width:
+            self.camera.set(cv2.CAP_PROP_FRAME_WIDTH, self.width)
+        if self.height:
+            self.camera.set(cv2.CAP_PROP_FRAME_HEIGHT, self.height)
+
+        actual_fps = self.camera.get(cv2.CAP_PROP_FPS)
+        actual_width = self.camera.get(cv2.CAP_PROP_FRAME_WIDTH)
+        actual_height = self.camera.get(cv2.CAP_PROP_FRAME_HEIGHT)
+
+        if self.fps and self.fps != actual_fps:
+            raise OSError(f"Can't set {self.fps=} for camera {self.camera_index}. Actual value is {actual_fps}.")
+        if self.width and self.width != actual_width:
+            raise OSError(f"Can't set {self.width=} for camera {self.camera_index}. Actual value is {actual_width}.")
+        if self.height and self.height != actual_height:
+            raise OSError(f"Can't set {self.height=} for camera {self.camera_index}. Actual value is {actual_height}.")
+        
+        self.is_connected = True
+        self.t.start()
+
+    def capture_image(self, temporary_color: str | None = None) -> np.ndarray:
+        if not self.is_connected:
+            self.connect()
+
+        ret, color_image = self.camera.read()
+        if not ret:
+            raise OSError(f"Can't capture color image from camera {self.camera_index}.")
+
+        if temporary_color is None:
+            requested_color = self.color
+        else:
+            requested_color = temporary_color
+        
+        if requested_color not in ["rgb", "bgr"]:
+            raise ValueError(f"Expected color values are 'rgb' or 'bgr', but {requested_color} is provided.")
+        
+        # OpenCV uses BGR format as default (blue, green red) for all operations, including displaying images.
+        # However, Deep Learning framework such as LeRobot uses RGB format as default to train neural networks,
+        # so we convert the image color from BGR to RGB.
+        if requested_color == "rgb":
+            color_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2RGB)
+        
+        return color_image
+    
+    def capture_image_loop(self):
+        while True:
+            self._color_image = self.capture_image()
+
+    def read(self):
+        while self._color_image is None:
+            time.sleep(0.1)
+        return self._color_image
+
+    def disconnect(self):
+        if getattr(self, "camera", None):
+            self.camera.release()
+
+    def __del__(self):
+        self.disconnect()
+
+
+def save_images_config(config: OpenCVCameraConfig, out_dir: Path):
+    cameras = []
+    print(f"Available camera indices: {OpenCVCamera.AVAILABLE_CAMERAS_INDICES}")
+    for camera_idx in OpenCVCamera.AVAILABLE_CAMERAS_INDICES:
+        camera = OpenCVCamera(camera_idx, config)
+        cameras.append(camera)
+
+    out_dir = out_dir.parent / f"{out_dir.name}_{config.width}x{config.height}_{config.fps}"
+    benchmark_cameras(cameras, out_dir, save_images=True)
+
+def benchmark_config(config: OpenCVCameraConfig, camera_ids: list[int]):
+    cameras = [OpenCVCamera(idx, config) for idx in camera_ids]
+    benchmark_cameras(cameras)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--mode", type=str, choices=["save_images", 'benchmark'], default="save_images")
+    parser.add_argument("--camera-ids", type=int, nargs="*", default=[16, 4, 22, 10])
+    parser.add_argument("--fps", type=int, default=30)
+    parser.add_argument("--width", type=str, default=640)
+    parser.add_argument("--height", type=str, default=480)
+    parser.add_argument("--out-dir", type=Path, default="outputs/benchmark_cameras/opencv/2024_06_22_1727")
+    args = parser.parse_args()
+
+    config = OpenCVCameraConfig(args.fps, args.width, args.height)
+    # config = OpenCVCameraConfig()
+    # config = OpenCVCameraConfig(60, 640, 480)
+    # config = OpenCVCameraConfig(90, 640, 480)
+    # config = OpenCVCameraConfig(30, 1280, 720)
+
+    if args.mode == "save_images":
+        save_images_config(config, args.out_dir)
+    elif args.mode == "benchmark":
+        benchmark_config(config, args.camera_ids)
+    else:
+        raise ValueError(args.mode)
+    
+    
+    
+
+    

lerobot/common/robot_devices/cameras/utils.py

@@ -0,0 +1,48 @@
+
+from pathlib import Path
+import time
+import cv2
+from typing import Protocol
+
+import numpy as np
+
+
+def write_shape_on_image_inplace(image):
+    height, width = image.shape[:2]
+    text = f'Width: {width} Height: {height}'
+
+    # Define the font, scale, color, and thickness
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    font_scale = 1
+    color = (255, 0, 0)  # Blue in BGR
+    thickness = 2
+
+    position = (10, height - 10)  # 10 pixels from the bottom-left corner
+    cv2.putText(image, text, position, font, font_scale, color, thickness)
+
+
+def save_color_image(image, path, write_shape=False):
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if write_shape:
+        write_shape_on_image_inplace(image)
+    cv2.imwrite(str(path), image)
+
+
+def save_depth_image(depth, path, write_shape=False):
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    # Apply colormap on depth image (image must be converted to 8-bit per pixel first)
+    depth_image = cv2.applyColorMap(cv2.convertScaleAbs(depth, alpha=0.03), cv2.COLORMAP_JET)
+
+    if write_shape:
+        write_shape_on_image_inplace(depth_image)
+    cv2.imwrite(str(path), depth_image)
+
+
+# Defines a camera type
+class Camera(Protocol):
+    def connect(self): ...
+    def read(self, temporary_color: str | None = None) -> np.ndarray: ...
+    def disconnect(self): ...

lerobot/common/robot_devices/motors/dynamixel.py

@@ -0,0 +1,405 @@
+from copy import deepcopy
+import enum
+from typing import Union
+import numpy as np
+
+from dynamixel_sdk import PacketHandler, PortHandler, COMM_SUCCESS, GroupSyncRead, GroupSyncWrite
+from dynamixel_sdk import DXL_HIBYTE, DXL_HIWORD, DXL_LOBYTE, DXL_LOWORD
+
+PROTOCOL_VERSION = 2.0
+BAUD_RATE = 1_000_000
+TIMEOUT_MS = 1000
+
+# https://emanual.robotis.com/docs/en/dxl/x/xl330-m077
+# https://emanual.robotis.com/docs/en/dxl/x/xl330-m288
+# https://emanual.robotis.com/docs/en/dxl/x/xl430-w250
+# https://emanual.robotis.com/docs/en/dxl/x/xm430-w350
+# https://emanual.robotis.com/docs/en/dxl/x/xm540-w270
+
+# data_name: (address, size_byte)
+X_SERIES_CONTROL_TABLE = {
+    "Model_Number": (0, 2),
+    "Model_Information": (2, 4),
+    "Firmware_Version": (6, 1),
+    "ID": (7, 1),
+    "Baud_Rate": (8, 1),
+    "Return_Delay_Time": (9, 1),
+    "Drive_Mode": (10, 1),
+    "Operating_Mode": (11, 1),
+    "Secondary_ID": (12, 1),
+    "Protocol_Type": (13, 1),
+    "Homing_Offset": (20, 4),
+    "Moving_Threshold": (24, 4),
+    "Temperature_Limit": (31, 1),
+    "Max_Voltage_Limit": (32, 2),
+    "Min_Voltage_Limit": (34, 2),
+    "PWM_Limit": (36, 2),
+    "Current_Limit": (38, 2),
+    "Acceleration_Limit": (40, 4),
+    "Velocity_Limit": (44, 4),
+    "Max_Position_Limit": (48, 4),
+    "Min_Position_Limit": (52, 4),
+    "Shutdown": (63, 1),
+    "Torque_Enable": (64, 1),
+    "LED": (65, 1),
+    "Status_Return_Level": (68, 1),
+    "Registered_Instruction": (69, 1),
+    "Hardware_Error_Status": (70, 1),
+    "Velocity_I_Gain": (76, 2),
+    "Velocity_P_Gain": (78, 2),
+    "Position_D_Gain": (80, 2),
+    "Position_I_Gain": (82, 2),
+    "Position_P_Gain": (84, 2),
+    "Feedforward_2nd_Gain": (88, 2),
+    "Feedforward_1st_Gain": (90, 2),
+    "Bus_Watchdog": (98, 1),
+    "Goal_PWM": (100, 2),
+    "Goal_Current": (102, 2),
+    "Goal_Velocity": (104, 4),
+    "Profile_Acceleration": (108, 4),
+    "Profile_Velocity": (112, 4),
+    "Goal_Position": (116, 4),
+    "Realtime_Tick": (120, 2),
+    "Moving": (122, 1),
+    "Moving_Status": (123, 1),
+    "Present_PWM": (124, 2),
+    "Present_Current": (126, 2),
+    "Present_Velocity": (128, 4),
+    "Present_Position": (132, 4),
+    "Velocity_Trajectory": (136, 4),
+    "Position_Trajectory": (140, 4),
+    "Present_Input_Voltage": (144, 2),
+    "Present_Temperature": (146, 1)
+}
+
+CALIBRATION_REQUIRED = ["Goal_Position", "Present_Position"]
+CONVERT_UINT32_TO_INT32_REQUIRED = ["Goal_Position", "Present_Position"]
+#CONVERT_POSITION_TO_ANGLE_REQUIRED = ["Goal_Position", "Present_Position"]
+CONVERT_POSITION_TO_ANGLE_REQUIRED = []
+
+MODEL_CONTROL_TABLE = {
+    "x_series": X_SERIES_CONTROL_TABLE,
+    "xl330-m077": X_SERIES_CONTROL_TABLE,
+    "xl330-m288": X_SERIES_CONTROL_TABLE,
+    "xl430-w250": X_SERIES_CONTROL_TABLE,
+    "xm430-w350": X_SERIES_CONTROL_TABLE,
+    "xm540-w270": X_SERIES_CONTROL_TABLE,
+}
+
+def uint32_to_int32(values: np.ndarray):
+    """
+    Convert an unsigned 32-bit integer array to a signed 32-bit integer array.
+    """
+    for i in range(len(values)):
+        if values[i] is not None and values[i] > 2147483647:
+            values[i] = values[i] - 4294967296
+    return values
+
+def int32_to_uint32(values: np.ndarray):
+    """
+    Convert a signed 32-bit integer array to an unsigned 32-bit integer array.
+    """
+    for i in range(len(values)):
+        if values[i] is not None and values[i] < 0:
+            values[i] = values[i] + 4294967296
+    return values
+
+def motor_position_to_angle(position: np.ndarray) -> np.ndarray:
+    """
+    Convert from motor position in [-2048, 2048] to radian in [-pi, pi]
+    """
+    return (position / 2048) * 3.14
+
+def motor_angle_to_position(angle: np.ndarray) -> np.ndarray:
+    """
+    Convert from radian in [-pi, pi] to motor position in [-2048, 2048]
+    """
+    return ((angle / 3.14) * 2048).astype(np.int64)
+
+
+# def pwm2vel(pwm: np.ndarray) -> np.ndarray:
+#     """
+#     :param pwm: numpy array of pwm/s joint velocities
+#     :return: numpy array of rad/s joint velocities
+#     """
+#     return pwm * 3.14 / 2048
+
+
+# def vel2pwm(vel: np.ndarray) -> np.ndarray:
+#     """
+#     :param vel: numpy array of rad/s joint velocities
+#     :return: numpy array of pwm/s joint velocities
+#     """
+#     return (vel * 2048 / 3.14).astype(np.int64)
+
+
+def get_group_sync_key(data_name, motor_names):
+    group_key = f"{data_name}_" + "_".join([name for name in motor_names])
+    return group_key
+
+
+class TorqueMode(enum.Enum):
+    ENABLED = 1
+    DISABLED = 0
+
+
+class OperatingMode(enum.Enum):
+    VELOCITY = 1
+    POSITION = 3
+    EXTENDED_POSITION = 4
+    CURRENT_CONTROLLED_POSITION = 5
+    PWM = 16
+    UNKNOWN = -1
+
+
+class DriveMode(enum.Enum):
+    NON_INVERTED = 0
+    INVERTED = 1
+
+
+class DynamixelMotorsBus:
+
+    def __init__(self, port: str, motors: dict[str, tuple[int, str]],
+                 extra_model_control_table: dict[str, list[tuple]] | None = None):
+        self.port = port
+        self.motors = motors
+
+        self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
+        if extra_model_control_table:
+            self.model_ctrl_table.update(extra_model_control_table)
+
+        self.port_handler = PortHandler(self.port)
+        self.packet_handler = PacketHandler(PROTOCOL_VERSION)
+
+        if not self.port_handler.openPort():
+            raise OSError(f"Failed to open port {self.port}")
+
+        self.port_handler.setBaudRate(BAUD_RATE)
+        self.port_handler.setPacketTimeoutMillis(TIMEOUT_MS)
+
+        self.group_readers = {}
+        self.group_writers = {}
+
+        self.calibration = None
+
+    @property
+    def motor_names(self) -> list[int]:
+        return list(self.motors.keys())
+    
+    def set_calibration(self, calibration: dict[str, tuple[int, bool]]):
+        self.calibration = calibration
+
+    def apply_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
+        if not self.calibration:
+            return values
+        
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        for i, name in enumerate(motor_names):
+            homing_offset, drive_mode = self.calibration[name]
+
+            if values[i] is not None:
+                if drive_mode:
+                    values[i] *= -1
+                values[i] += homing_offset
+
+        return values
+    
+    def revert_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
+        if not self.calibration:
+            return values
+
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        for i, name in enumerate(motor_names):
+            homing_offset, drive_mode = self.calibration[name]
+
+            if values[i] is not None:
+                values[i] -= homing_offset
+                if drive_mode:
+                    values[i] *= -1
+
+        return values
+
+    def read(self, data_name, motor_names: list[str] | None = None):
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        motor_ids = []
+        models = []
+        for name in motor_names:
+            motor_idx, model = self.motors[name]
+            motor_ids.append(motor_idx)
+            models.append(model)
+
+        # TODO(rcadene): assert all motors follow same address
+        addr, bytes = self.model_ctrl_table[model][data_name]
+        group_key = get_group_sync_key(data_name, motor_names)
+
+        if data_name not in self.group_readers:
+            # create new group reader
+            self.group_readers[group_key] = GroupSyncRead(self.port_handler, self.packet_handler, addr, bytes)
+            for idx in motor_ids:
+                self.group_readers[group_key].addParam(idx)
+
+        comm = self.group_readers[group_key].txRxPacket()
+        if comm != COMM_SUCCESS:
+            raise ConnectionError(
+                f"Read failed due to communication error on port {self.port} for group_key {group_key}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+        values = []
+        for idx in motor_ids:
+            value = self.group_readers[group_key].getData(idx, addr, bytes)
+            values.append(value)
+
+        values = np.array(values)
+
+        # TODO(rcadene): explain why
+        if data_name in CONVERT_UINT32_TO_INT32_REQUIRED:
+            values = uint32_to_int32(values)
+
+        if data_name in CALIBRATION_REQUIRED:
+            values = self.apply_calibration(values, motor_names)
+        
+        if data_name in CONVERT_POSITION_TO_ANGLE_REQUIRED:
+            values = motor_position_to_angle(values)
+
+        return values
+
+    def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
+        if motor_names is None:
+            motor_names = self.motor_names
+
+        if isinstance(motor_names, str):
+            motor_names = [motor_names]
+
+        motor_ids = []
+        models = []
+        for name in motor_names:
+            motor_idx, model = self.motors[name]
+            motor_ids.append(motor_idx)
+            models.append(model)
+
+        if isinstance(values, (int, float, np.integer)):
+            values = [int(values)] * len(motor_ids)
+
+        values = np.array(values)
+
+        if data_name in CONVERT_POSITION_TO_ANGLE_REQUIRED:
+            values = motor_angle_to_position(values)
+
+        if data_name in CALIBRATION_REQUIRED:
+            values = self.revert_calibration(values, motor_names)
+
+        # TODO(rcadene): why dont we do it?
+        # if data_name in CONVERT_INT32_TO_UINT32_REQUIRED:
+        #     values = int32_to_uint32(values)
+
+        values = values.tolist()
+
+        # TODO(rcadene): assert all motors follow same address
+        addr, bytes = self.model_ctrl_table[model][data_name]
+        group_key = get_group_sync_key(data_name, motor_names)
+
+        init_group = data_name not in self.group_readers
+        if init_group:
+            self.group_writers[group_key] = GroupSyncWrite(self.port_handler, self.packet_handler, addr, bytes)
+
+        for idx, value in zip(motor_ids, values):
+            if bytes == 1:
+                data = [
+                    DXL_LOBYTE(DXL_LOWORD(value)),
+                ]
+            elif bytes == 2:
+                data = [
+                    DXL_LOBYTE(DXL_LOWORD(value)),
+                    DXL_HIBYTE(DXL_LOWORD(value)),
+                ]
+            elif bytes == 4:
+                data = [
+                    DXL_LOBYTE(DXL_LOWORD(value)),
+                    DXL_HIBYTE(DXL_LOWORD(value)),
+                    DXL_LOBYTE(DXL_HIWORD(value)),
+                    DXL_HIBYTE(DXL_HIWORD(value)),
+                ]
+            else:
+                raise NotImplementedError(
+                    f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but "
+                    f"{bytes} is provided instead.")
+
+            if init_group:
+                self.group_writers[group_key].addParam(idx, data)
+            else:
+                self.group_writers[group_key].changeParam(idx, data)
+
+        comm = self.group_writers[group_key].txPacket()
+        if comm != COMM_SUCCESS:
+            raise ConnectionError(
+                f"Write failed due to communication error on port {self.port} for group_key {group_key}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+    # def read(self, data_name, motor_name: str):
+    #     motor_idx, model = self.motors[motor_name]
+    #     addr, bytes = self.model_ctrl_table[model][data_name]
+
+    #     args = (self.port_handler, motor_idx, addr)
+    #     if bytes == 1:
+    #         value, comm, err = self.packet_handler.read1ByteTxRx(*args)
+    #     elif bytes == 2:
+    #         value, comm, err = self.packet_handler.read2ByteTxRx(*args)
+    #     elif bytes == 4:
+    #         value, comm, err = self.packet_handler.read4ByteTxRx(*args)
+    #     else:
+    #         raise NotImplementedError(
+    #             f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but "
+    #             f"{bytes} is provided instead.")
+
+    #     if comm != COMM_SUCCESS:
+    #         raise ConnectionError(
+    #             f"Read failed due to communication error on port {self.port} for motor {motor_idx}: "
+    #             f"{self.packet_handler.getTxRxResult(comm)}"
+    #         )
+    #     elif err != 0:
+    #         raise ConnectionError(
+    #             f"Read failed due to error {err} on port {self.port} for motor {motor_idx}: "
+    #             f"{self.packet_handler.getTxRxResult(err)}"
+    #         )
+
+    #     if data_name in CALIBRATION_REQUIRED:
+    #         value = self.apply_calibration([value], [motor_name])[0]
+
+    #     return value
+
+    # def write(self, data_name, value, motor_name: str):
+    #     if data_name in CALIBRATION_REQUIRED:
+    #         value = self.revert_calibration([value], [motor_name])[0]
+
+    #     motor_idx, model = self.motors[motor_name]
+    #     addr, bytes = self.model_ctrl_table[model][data_name]
+    #     args = (self.port_handler, motor_idx, addr, value)
+    #     if bytes == 1:
+    #         comm, err = self.packet_handler.write1ByteTxRx(*args)
+    #     elif bytes == 2:
+    #         comm, err = self.packet_handler.write2ByteTxRx(*args)
+    #     elif bytes == 4:
+    #         comm, err = self.packet_handler.write4ByteTxRx(*args)
+    #     else:
+    #         raise NotImplementedError(
+    #             f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but {bytes} "
+    #             f"is provided instead.")
+
+    #     if comm != COMM_SUCCESS:
+    #         raise ConnectionError(
+    #             f"Write failed due to communication error on port {self.port} for motor {motor_idx}: "
+    #             f"{self.packet_handler.getTxRxResult(comm)}"
+    #         )
+    #     elif err != 0:
+    #         raise ConnectionError(
+    #             f"Write failed due to error {err} on port {self.port} for motor {motor_idx}: "
+    #             f"{self.packet_handler.getTxRxResult(err)}"
+    #         )

lerobot/common/robot_devices/motors/feetech.py

@@ -0,0 +1,713 @@
+from copy import deepcopy
+import enum
+import numpy as np
+
+from scservo_sdk import PacketHandler, PortHandler, COMM_SUCCESS, GroupSyncRead, GroupSyncWrite
+from scservo_sdk import SCS_HIBYTE, SCS_HIBYTE, SCS_LOBYTE, SCS_LOWORD
+
+PROTOCOL_VERSION = 0
+BAUD_RATE = 1_000_000
+TIMEOUT_MS = 1000
+
+
+def u32_to_i32(value: int | np.array) -> int | np.array:
+    """
+    Convert an unsigned 32-bit integer array to a signed 32-bit integer array.
+    """
+
+    if isinstance(value, int):
+        if value > 2147483647:
+            value = value - 4294967296
+    else:
+        for i in range(len(value)):
+            if value[i] is not None and value[i] > 2147483647:
+                value[i] = value[i] - 4294967296
+
+    return value
+
+
+def i32_to_u32(value: int | np.array) -> int | np.array:
+    """
+    Convert a signed 32-bit integer array to an unsigned 32-bit integer array.
+    """
+
+    if isinstance(value, int):
+        if value < 0:
+            value = value + 4294967296
+    else:
+        for i in range(len(value)):
+            if value[i] is not None and value[i] < 0:
+                value[i] = value[i] + 4294967296
+
+    return value
+
+
+def retrieve_ids_and_command(values: np.array, ids: np.array) -> (list[int], np.array):
+    """
+    Convert the values to a chain command. Skip the None values and return the ids and values.
+    """
+
+    non_none_values = np.array([value for value in values if value is not None])
+    non_none_values_ids = [ids[i] for i, value in enumerate(values) if value is not None]
+
+    return non_none_values_ids, non_none_values
+
+
+class TorqueMode(enum.Enum):
+    ENABLED = 1
+    DISABLED = 0
+
+
+class OperatingMode(enum.Enum):
+    pass
+
+
+class DriveMode(enum.Enum):
+    pass
+
+
+SCS_SERIES_CONTROL_TABLE = [
+    ("Model", 3, 2),
+    ("ID", 5, 1),
+    ("Baud_Rate", 6, 1),
+    ("Return_Delay", 7, 1),
+    ("Response_Status_Level", 8, 1),
+    ("Min_Angle_Limit", 9, 2),
+    ("Max_Angle_Limit", 11, 2),
+    ("Max_Temperature_Limit", 13, 1),
+    ("Max_Voltage_Limit", 14, 1),
+    ("Min_Voltage_Limit", 15, 1),
+    ("Max_Torque_Limit", 16, 2),
+    ("Phase", 18, 1),
+    ("Unloading_Condition", 19, 1),
+    ("LED_Alarm_Condition", 20, 1),
+    ("P_Coefficient", 21, 1),
+    ("D_Coefficient", 22, 1),
+    ("I_Coefficient", 23, 1),
+    ("Minimum_Startup_Force", 24, 2),
+    ("CW_Dead_Zone", 26, 1),
+    ("CCW_Dead_Zone", 27, 1),
+    ("Protection_Current", 28, 2),
+    ("Angular_Resolution", 30, 1),
+    ("Offset", 31, 2),
+    ("Mode", 33, 1),
+    ("Protective_Torque", 34, 1),
+    ("Protection_Time", 35, 1),
+    ("Overload_Torque", 36, 1),
+    ("Speed_closed_loop_P_proportional_coefficient", 37, 1),
+    ("Over_Current_Protection_Time", 38, 1),
+    ("Velocity_closed_loop_I_integral_coefficient", 39, 1),
+    ("Torque_Enable", 40, 1),
+    ("Acceleration", 41, 1),
+    ("Goal_Position", 42, 2),
+    ("Goal_Time", 44, 2),
+    ("Goal_Speed", 46, 2),
+    ("Lock", 55, 1),
+    ("Present_Position", 56, 2),
+    ("Present_Speed", 58, 2),
+    ("Present_Load", 60, 2),
+    ("Present_Voltage", 62, 1),
+    ("Present_Temperature", 63, 1),
+    ("Status", 65, 1),
+    ("Moving", 66, 1),
+    ("Present_Current", 69, 2)
+]
+
+MODEL_CONTROL_TABLE = {
+    "scs_series": SCS_SERIES_CONTROL_TABLE,
+    "sts3215": SCS_SERIES_CONTROL_TABLE,
+}
+
+
+class FeetechBus:
+
+    def __init__(self, port: str, motor_models: dict[int, str],
+                 extra_model_control_table: dict[str, list[tuple]] | None = None):
+        self.port = port
+        self.motor_models = motor_models
+
+        self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
+        if extra_model_control_table:
+            self.model_ctrl_table.update(extra_model_control_table)
+
+        # Find read/write addresses and number of bytes for each motor
+        self.motor_ctrl = {}
+        for idx, model in self.motor_models.items():
+            for data_name, addr, bytes in self.model_ctrl_table[model]:
+                if idx not in self.motor_ctrl:
+                    self.motor_ctrl[idx] = {}
+                self.motor_ctrl[idx][data_name] = {
+                    "addr": addr,
+                    "bytes": bytes,
+                }
+
+        self.port_handler = PortHandler(self.port)
+        self.packet_handler = PacketHandler(PROTOCOL_VERSION)
+
+        if not self.port_handler.openPort():
+            raise OSError(f"Failed to open port {self.port}")
+
+        self.port_handler.setBaudRate(BAUD_RATE)
+        self.port_handler.setPacketTimeoutMillis(TIMEOUT_MS)
+
+        self.group_readers = {}
+        self.group_writers = {}
+
+    @property
+    def motor_ids(self) -> list[int]:
+        return list(self.motor_models.keys())
+
+    def close(self):
+        self.port_handler.closePort()
+
+    def write(self, data_name, value, motor_idx: int):
+
+        addr = self.motor_ctrl[motor_idx][data_name]["addr"]
+        bytes = self.motor_ctrl[motor_idx][data_name]["bytes"]
+        args = (self.port_handler, motor_idx, addr, value)
+        if bytes == 1:
+            comm, err = self.packet_handler.write1ByteTxRx(*args)
+        elif bytes == 2:
+            comm, err = self.packet_handler.write2ByteTxRx(*args)
+        elif bytes == 4:
+            comm, err = self.packet_handler.write4ByteTxRx(*args)
+        else:
+            raise NotImplementedError(
+                f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but {bytes} "
+                f"is provided instead.")
+
+        if comm != COMM_SUCCESS:
+            raise ConnectionError(
+                f"Write failed due to communication error on port {self.port} for motor {motor_idx}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+        elif err != 0:
+            raise ConnectionError(
+                f"Write failed due to error {err} on port {self.port} for motor {motor_idx}: "
+                f"{self.packet_handler.getTxRxResult(err)}"
+            )
+
+    def read(self, data_name, motor_idx: int):
+        addr = self.motor_ctrl[motor_idx][data_name]["addr"]
+        bytes = self.motor_ctrl[motor_idx][data_name]["bytes"]
+        args = (self.port_handler, motor_idx, addr)
+        if bytes == 1:
+            value, comm, err = self.packet_handler.read1ByteTxRx(*args)
+        elif bytes == 2:
+            value, comm, err = self.packet_handler.read2ByteTxRx(*args)
+        elif bytes == 4:
+            value, comm, err = self.packet_handler.read4ByteTxRx(*args)
+        else:
+            raise NotImplementedError(
+                f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but "
+                f"{bytes} is provided instead.")
+
+        if comm != COMM_SUCCESS:
+            raise ConnectionError(
+                f"Read failed due to communication error on port {self.port} for motor {motor_idx}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+        elif err != 0:
+            raise ConnectionError(
+                f"Read failed due to error {err} on port {self.port} for motor {motor_idx}: "
+                f"{self.packet_handler.getTxRxResult(err)}"
+            )
+
+        return value
+
+    def sync_read(self, data_name, motor_ids: list[int] | None = None):
+        if motor_ids is None:
+            motor_ids = self.motor_ids
+
+        group_key = f"{data_name}_" + "_".join([str(idx) for idx in motor_ids])
+        first_motor_idx = list(self.motor_ctrl.keys())[0]
+        addr = self.motor_ctrl[first_motor_idx][data_name]["addr"]
+        bytes = self.motor_ctrl[first_motor_idx][data_name]["bytes"]
+
+        if data_name not in self.group_readers:
+            self.group_readers[group_key] = GroupSyncRead(self.port_handler, self.packet_handler, addr, bytes)
+            for idx in motor_ids:
+                self.group_readers[group_key].addParam(idx)
+
+        comm = self.group_readers[group_key].txRxPacket()
+        if comm != COMM_SUCCESS:
+            raise ConnectionError(
+                f"Read failed due to communication error on port {self.port} for group_key {group_key}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+        values = []
+        for idx in motor_ids:
+            value = self.group_readers[group_key].getData(idx, addr, bytes)
+            values.append(value)
+
+        return np.array(values)
+
+    def sync_write(self, data_name, values: int | list[int], motor_ids: int | list[int] | None = None):
+        if motor_ids is None:
+            motor_ids = self.motor_ids
+
+        if isinstance(motor_ids, int):
+            motor_ids = [motor_ids]
+
+        if isinstance(values, (int, np.integer)):
+            values = [int(values)] * len(motor_ids)
+
+        if isinstance(values, np.ndarray):
+            values = values.tolist()
+
+        group_key = f"{data_name}_" + "_".join([str(idx) for idx in motor_ids])
+
+        first_motor_idx = list(self.motor_ctrl.keys())[0]
+        addr = self.motor_ctrl[first_motor_idx][data_name]["addr"]
+        bytes = self.motor_ctrl[first_motor_idx][data_name]["bytes"]
+        init_group = data_name not in self.group_readers
+
+        if init_group:
+            self.group_writers[group_key] = GroupSyncWrite(self.port_handler, self.packet_handler, addr, bytes)
+
+        for idx, value in zip(motor_ids, values):
+            if bytes == 1:
+                data = [
+                    SCS_LOBYTE(SCS_LOWORD(value)),
+                ]
+            elif bytes == 2:
+                data = [
+                    SCS_LOBYTE(SCS_LOWORD(value)),
+                    SCS_HIBYTE(SCS_LOWORD(value)),
+                ]
+            elif bytes == 4:
+                data = [
+                    SCS_LOBYTE(SCS_LOWORD(value)),
+                    SCS_HIBYTE(SCS_LOWORD(value)),
+                    SCS_LOBYTE(SCS_HIBYTE(value)),
+                    SCS_HIBYTE(SCS_HIBYTE(value)),
+                ]
+            else:
+                raise NotImplementedError(
+                    f"Value of the number of bytes to be sent is expected to be in [1, 2, 4], but {bytes} "
+                    f"is provided instead.")
+
+            if init_group:
+                self.group_writers[group_key].addParam(idx, data)
+            else:
+                self.group_writers[group_key].changeParam(idx, data)
+
+        comm = self.group_writers[group_key].txPacket()
+        if comm != COMM_SUCCESS:
+            raise ConnectionError(
+                f"Write failed due to communication error on port {self.port} for group_key {group_key}: "
+                f"{self.packet_handler.getTxRxResult(comm)}"
+            )
+
+    def read_model(self, motor_idx: int):
+        return self.read("Model", motor_idx)
+
+    def sync_read_model(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Model", motor_ids)
+
+    def write_id(self, value, motor_idx: int):
+        self.write("ID", value, motor_idx)
+
+    def read_id(self, motor_idx: int):
+        return self.read("ID", motor_idx)
+
+    def sync_read_id(self, motor_ids: list[int] | None = None):
+        return self.sync_read("ID", motor_ids)
+
+    def sync_write_id(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("ID", values, motor_ids)
+
+    def write_baud_rate(self, value, motor_idx: int):
+        self.write("Baud_Rate", value, motor_idx)
+
+    def read_baud_rate(self, motor_idx: int):
+        return self.read("Baud_Rate", motor_idx)
+
+    def sync_read_baud_rate(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Baud_Rate", motor_ids)
+
+    def sync_write_baud_rate(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Baud_Rate", values, motor_ids)
+
+    def read_return_delay(self, motor_idx: int):
+        return self.read("Return_Delay", motor_idx)
+
+    def sync_read_return_delay(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Return_Delay", motor_ids)
+
+    def read_response_status_level(self, motor_idx: int):
+        return self.read("Response_Status_Level", motor_idx)
+
+    def sync_read_response_status_level(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Response_Status_Level", motor_ids)
+
+    def write_min_angle_limit(self, value, motor_idx: int):
+        self.write("Min_Angle_Limit", value, motor_idx)
+
+    def read_min_angle_limit(self, motor_idx: int):
+        return self.read("Min_Angle_Limit", motor_idx)
+
+    def sync_read_min_angle_limit(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Min_Angle_Limit", motor_ids)
+
+    def sync_write_min_angle_limit(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Min_Angle_Limit", values, motor_ids)
+
+    def write_max_angle_limit(self, value, motor_idx: int):
+        self.write("Max_Angle_Limit", value, motor_idx)
+
+    def read_max_angle_limit(self, motor_idx: int):
+        return self.read("Max_Angle_Limit", motor_idx)
+
+    def sync_read_max_angle_limit(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Max_Angle_Limit", motor_ids)
+
+    def sync_write_max_angle_limit(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Max_Angle_Limit", values, motor_ids)
+
+    def write_max_temperature_limit(self, value, motor_idx: int):
+        self.write("Max_Temperature_Limit", value, motor_idx)
+
+    def read_max_temperature_limit(self, motor_idx: int):
+        return self.read("Max_Temperature_Limit", motor_idx)
+
+    def sync_read_max_temperature_limit(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Max_Temperature_Limit", motor_ids)
+
+    def sync_write_max_temperature_limit(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Max_Temperature_Limit", values, motor_ids)
+
+    def write_max_voltage_limit(self, value, motor_idx: int):
+        self.write("Max_Voltage_Limit", value, motor_idx)
+
+    def read_max_voltage_limit(self, motor_idx: int):
+        return self.read("Max_Voltage_Limit", motor_idx)
+
+    def sync_read_max_voltage_limit(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Max_Voltage_Limit", motor_ids)
+
+    def sync_write_max_voltage_limit(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Max_Voltage_Limit", values, motor_ids)
+
+    def write_min_voltage_limit(self, value, motor_idx: int):
+        self.write("Min_Voltage_Limit", value, motor_idx)
+
+    def read_min_voltage_limit(self, motor_idx: int):
+        return self.read("Min_Voltage_Limit", motor_idx)
+
+    def sync_read_min_voltage_limit(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Min_Voltage_Limit", motor_ids)
+
+    def sync_write_min_voltage_limit(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Min_Voltage_Limit", values, motor_ids)
+
+    def write_max_torque_limit(self, value, motor_idx: int):
+        self.write("Max_Torque_Limit", value, motor_idx)
+
+    def read_max_torque_limit(self, motor_idx: int):
+        return self.read("Max_Torque_Limit", motor_idx)
+
+    def sync_read_max_torque_limit(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Max_Torque_Limit", motor_ids)
+
+    def sync_write_max_torque_limit(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Max_Torque_Limit", values, motor_ids)
+
+    def write_p_coefficient(self, value, motor_idx: int):
+        self.write("P_Coefficient", value, motor_idx)
+
+    def read_p_coefficient(self, motor_idx: int):
+        return self.read("P_Coefficient", motor_idx)
+
+    def sync_read_p_coefficient(self, motor_ids: list[int] | None = None):
+        return self.sync_read("P_Coefficient", motor_ids)
+
+    def sync_write_p_coefficient(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("P_Coefficient", values, motor_ids)
+
+    def write_d_coefficient(self, value, motor_idx: int):
+        self.write("D_Coefficient", value, motor_idx)
+
+    def read_d_coefficient(self, motor_idx: int):
+        return self.read("D_Coefficient", motor_idx)
+
+    def sync_read_d_coefficient(self, motor_ids: list[int] | None = None):
+        return self.sync_read("D_Coefficient", motor_ids)
+
+    def sync_write_d_coefficient(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("D_Coefficient", values, motor_ids)
+
+    def write_i_coefficient(self, value, motor_idx: int):
+        self.write("I_Coefficient", value, motor_idx)
+
+    def read_i_coefficient(self, motor_idx: int):
+        return self.read("I_Coefficient", motor_idx)
+
+    def sync_read_i_coefficient(self, motor_ids: list[int] | None = None):
+        return self.sync_read("I_Coefficient", motor_ids)
+
+    def sync_write_i_coefficient(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("I_Coefficient", values, motor_ids)
+
+    def write_minimum_startup_force(self, value, motor_idx: int):
+        self.write("Minimum_Startup_Force", value, motor_idx)
+
+    def read_minimum_startup_force(self, motor_idx: int):
+        return self.read("Minimum_Startup_Force", motor_idx)
+
+    def sync_read_minimum_startup_force(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Minimum_Startup_Force", motor_ids)
+
+    def sync_write_minimum_startup_force(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Minimum_Startup_Force", values, motor_ids)
+
+    def write_cw_dead_zone(self, value, motor_idx: int):
+        self.write("CW_Dead_Zone", value, motor_idx)
+
+    def read_cw_dead_zone(self, motor_idx: int):
+        return self.read("CW_Dead_Zone", motor_idx)
+
+    def sync_read_cw_dead_zone(self, motor_ids: list[int] | None = None):
+        return self.sync_read("CW_Dead_Zone", motor_ids)
+
+    def sync_write_cw_dead_zone(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("CW_Dead_Zone", values, motor_ids)
+
+    def write_ccw_dead_zone(self, value, motor_idx: int):
+        self.write("CCW_Dead_Zone", value, motor_idx)
+
+    def read_ccw_dead_zone(self, motor_idx: int):
+        return self.read("CCW_Dead_Zone", motor_idx)
+
+    def sync_read_ccw_dead_zone(self, motor_ids: list[int] | None = None):
+        return self.sync_read("CCW_Dead_Zone", motor_ids)
+
+    def sync_write_ccw_dead_zone(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("CCW_Dead_Zone", values, motor_ids)
+
+    def write_protection_current(self, value, motor_idx: int):
+        self.write("Protection_Current", value, motor_idx)
+
+    def read_protection_current(self, motor_idx: int):
+        return self.read("Protection_Current", motor_idx)
+
+    def sync_read_protection_current(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Protection_Current", motor_ids)
+
+    def sync_write_protection_current(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Protection_Current", values, motor_ids)
+
+    def read_angular_resolution(self, motor_idx: int):
+        return self.read("Angular_Resolution", motor_idx)
+
+    def sync_read_angular_resolution(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Angular_Resolution", motor_ids)
+
+    def write_offset(self, value, motor_idx: int):
+        self.write("Offset", value, motor_idx)
+
+    def read_offset(self, motor_idx: int):
+        return self.read("Offset", motor_idx)
+
+    def sync_read_offset(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Offset", motor_ids)
+
+    def sync_write_offset(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Offset", values, motor_ids)
+
+    def write_mode(self, value, motor_idx: int):
+        self.write("Mode", value, motor_idx)
+
+    def read_mode(self, motor_idx: int):
+        return self.read("Mode", motor_idx)
+
+    def sync_read_mode(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Mode", motor_ids)
+
+    def sync_write_mode(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Mode", values, motor_ids)
+
+    def write_protective_torque(self, value, motor_idx: int):
+        self.write("Protective_Torque", value, motor_idx)
+
+    def read_protective_torque(self, motor_idx: int):
+        return self.read("Protective_Torque", motor_idx)
+
+    def sync_read_protective_torque(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Protective_Torque", motor_ids)
+
+    def sync_write_protective_torque(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Protective_Torque", values, motor_ids)
+
+    def read_protection_time(self, motor_idx: int):
+        return self.read("Protection_Time", motor_idx)
+
+    def sync_read_protection_time(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Protection_Time", motor_ids)
+
+    def write_speed_closed_loop_p_proportional_coefficient(self, value, motor_idx: int):
+        self.write("Speed_closed_loop_P_proportional_coefficient", value, motor_idx)
+
+    def read_speed_closed_loop_p_proportional_coefficient(self, motor_idx: int):
+        return self.read("Speed_closed_loop_P_proportional_coefficient", motor_idx)
+
+    def sync_read_speed_closed_loop_p_proportional_coefficient(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Speed_closed_loop_P_proportional_coefficient", motor_ids)
+
+    def sync_write_speed_closed_loop_p_proportional_coefficient(self, values: int | list[int],
+                                                                motor_ids: list[int] | None = None):
+        self.sync_write("Speed_closed_loop_P_proportional_coefficient", values, motor_ids)
+
+    def write_over_current_protection_time(self, value, motor_idx: int):
+        self.write("Over_Current_Protection_Time", value, motor_idx)
+
+    def read_over_current_protection_time(self, motor_idx: int):
+        return self.read("Over_Current_Protection_Time", motor_idx)
+
+    def sync_read_over_current_protection_time(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Over_Current_Protection_Time", motor_ids)
+
+    def sync_write_over_current_protection_time(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Over_Current_Protection_Time", values, motor_ids)
+
+    def write_velocity_closed_loop_i_integral_coefficient(self, value, motor_idx: int):
+        self.write("Velocity_closed_loop_I_integral_coefficient", value, motor_idx)
+
+    def read_velocity_closed_loop_i_integral_coefficient(self, motor_idx: int):
+        return self.read("Velocity_closed_loop_I_integral_coefficient", motor_idx)
+
+    def sync_read_velocity_closed_loop_i_integral_coefficient(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Velocity_closed_loop_I_integral_coefficient", motor_ids)
+
+    def sync_write_velocity_closed_loop_i_integral_coefficient(self, values: int | list[int],
+                                                               motor_ids: list[int] | None = None):
+        self.sync_write("Velocity_closed_loop_I_integral_coefficient", values, motor_ids)
+
+    def write_torque_enable(self, value, motor_idx: int):
+        self.write("Torque_Enable", value, motor_idx)
+
+    def read_torque_enable(self, motor_idx: int):
+        return self.read("Torque_Enable", motor_idx)
+
+    def sync_read_torque_enable(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Torque_Enable", motor_ids)
+
+    def sync_write_torque_enable(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Torque_Enable", values, motor_ids)
+
+    def write_goal_position_u32(self, value, motor_idx: int):
+        self.write("Goal_Position", value, motor_idx)
+
+    def write_goal_position_i32(self, value, motor_idx: int):
+        self.write("Goal_Position", i32_to_u32(value), motor_idx)
+
+    def read_goal_position_u32(self, motor_idx: int):
+        return self.read("Goal_Position", motor_idx)
+
+    def read_goal_position_i32(self, motor_idx: int):
+        goal_position_u32 = self.read_goal_position_u32(motor_idx)
+
+        return u32_to_i32(goal_position_u32)
+
+    def sync_read_goal_position_u32(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Goal_Position", motor_ids)
+
+    def sync_read_goal_position_i32(self, motor_ids: list[int] | None = None):
+        goal_position_u32 = self.sync_read_goal_position_u32(motor_ids)
+
+        return u32_to_i32(goal_position_u32)
+
+    def sync_write_goal_position_u32(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Goal_Position", values, motor_ids)
+
+    def sync_write_goal_position_i32(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Goal_Position", i32_to_u32(values), motor_ids)
+
+    def write_goal_time(self, value, motor_idx: int):
+        self.write("Goal_Time", value, motor_idx)
+
+    def read_goal_time(self, motor_idx: int):
+        return self.read("Goal_Time", motor_idx)
+
+    def sync_read_goal_time(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Goal_Time", motor_ids)
+
+    def sync_write_goal_time(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Goal_Time", values, motor_ids)
+
+    def write_goal_speed(self, value, motor_idx: int):
+        self.write("Goal_Speed", value, motor_idx)
+
+    def read_goal_speed(self, motor_idx: int):
+        return self.read("Goal_Speed", motor_idx)
+
+    def sync_read_goal_speed(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Goal_Speed", motor_ids)
+
+    def sync_write_goal_speed(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Goal_Speed", values, motor_ids)
+
+    def write_lock(self, value, motor_idx: int):
+        self.write("Lock", value, motor_idx)
+
+    def read_lock(self, motor_idx: int):
+        return self.read("Lock", motor_idx)
+
+    def sync_read_lock(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Lock", motor_ids)
+
+    def sync_write_lock(self, values: int | list[int], motor_ids: list[int] | None = None):
+        self.sync_write("Lock", values, motor_ids)
+
+    def read_present_position_u32(self, motor_idx: int):
+        return self.read("Present_Position", motor_idx)
+
+    def read_present_position_i32(self, motor_idx: int):
+        present_position_u32 = self.read_present_position_u32(motor_idx)
+
+        return u32_to_i32(present_position_u32)
+
+    def sync_read_present_position_u32(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Present_Position", motor_ids)
+
+    def sync_read_present_position_i32(self, motor_ids: list[int] | None = None):
+        present_position_u32 = self.sync_read_present_position_u32(motor_ids)
+
+        return u32_to_i32(present_position_u32)
+
+    def read_present_speed(self, motor_idx: int):
+        return self.read("Present_Speed", motor_idx)
+
+    def sync_read_present_speed(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Present_Speed", motor_ids)
+
+    def read_present_load(self, motor_idx: int):
+        return self.read("Present_Load", motor_idx)
+
+    def sync_read_present_load(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Present_Load", motor_ids)
+
+    def read_present_voltage(self, motor_idx: int):
+        return self.read("Present_Voltage", motor_idx)
+
+    def sync_read_present_voltage(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Present_Voltage", motor_ids)
+
+    def read_present_temperature(self, motor_idx: int):
+        return self.read("Present_Temperature", motor_idx)
+
+    def sync_read_present_temperature(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Present_Temperature", motor_ids)
+
+    def read_moving(self, motor_idx: int):
+        return self.read("Moving", motor_idx)
+
+    def sync_read_moving(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Moving", motor_ids)
+
+    def read_present_current(self, motor_idx: int):
+        return self.read("Present_Current", motor_idx)
+
+    def sync_read_present_current(self, motor_ids: list[int] | None = None):
+        return self.sync_read("Present_Current", motor_ids)

lerobot/common/robot_devices/motors/utils.py

@@ -0,0 +1,9 @@
+from typing import Protocol
+
+class MotorsBus(Protocol):
+    def motor_names(self): ...
+    def set_calibration(self): ...
+    def apply_calibration(self): ...
+    def revert_calibration(self): ...
+    def read(self): ...
+    def write(self): ...

lerobot/common/robot_devices/robots/aloha.py

@@ -0,0 +1,204 @@
+import copy
+from dataclasses import dataclass, field, replace
+import numpy as np
+import torch
+from examples.real_robot_example.gym_real_world.robot import Robot
+from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
+from lerobot.common.robot_devices.cameras.utils import Camera
+
+MAX_LEADER_GRIPPER_RAD = 0.7761942786701344
+MAX_LEADER_GRIPPER_POS = 2567
+MAX_FOLLOWER_GRIPPER_RAD = 1.6827769243105486
+MAX_FOLLOWER_GRIPPER_POS = 3100
+
+MIN_LEADER_GRIPPER_RAD = -0.12732040539450828
+MIN_LEADER_GRIPPER_POS = 1984
+MIN_FOLLOWER_GRIPPER_RAD = 0.6933593161243099
+MIN_FOLLOWER_GRIPPER_POS = 2512
+
+GRIPPER_INDEX = -1
+
+def convert_gripper_range_from_leader_to_follower(leader_pos):
+    follower_goal_pos = copy.copy(leader_pos)
+    follower_goal_pos[GRIPPER_INDEX] = \
+        (leader_pos[GRIPPER_INDEX] - MIN_LEADER_GRIPPER_POS) \
+        / (MAX_LEADER_GRIPPER_POS - MIN_LEADER_GRIPPER_POS) \
+        * (MAX_FOLLOWER_GRIPPER_POS - MIN_FOLLOWER_GRIPPER_POS) \
+        + MIN_FOLLOWER_GRIPPER_POS
+    return follower_goal_pos
+
+
+
+@dataclass
+class AlohaRobotConfig:
+    """
+    Example of usage:
+    ```python
+    AlohaRobotConfig()
+    ```
+
+    Example of only using left arm:
+    ```python
+    AlohaRobotConfig(
+        activated_leaders=["left"],
+        activated_followers=["left"],
+    )
+    ```
+    """
+
+    # Define all the components of the robot
+    leader_devices: dict[str, str] = field(
+        default_factory=lambda: {
+            "right": {
+                #"port": "/dev/ttyDXL_master_right",
+                "port": "/dev/ttyDXL_master_left",
+                "servos": [1, 2, 3, 4, 5, 6, 7, 8, 9],
+            },
+            "left": {
+                "port": "/dev/ttyDXL_master_left",
+                "servos": [1, 2, 3, 4, 5, 6, 7, 8, 9],
+            },
+        }
+    )
+    follower_devices: dict[str, str] = field(
+        default_factory=lambda: {
+            "right": {
+                "port": "/dev/ttyDXL_puppet_right",
+                "servos": [1, 2, 3, 4, 5, 6, 7, 8, 9],
+            },
+            "left": {
+                "port": "/dev/ttyDXL_puppet_left",
+                "servos": [1, 2, 3, 4, 5, 6, 7, 8, 9],
+            },
+        }
+    )
+    camera_devices: dict[str, Camera] = field(
+        default_factory=lambda: {
+            # "cam_high": OpenCVCamera(16),
+            # "cam_low": OpenCVCamera(4),
+            # "cam_left_wrist": OpenCVCamera(10),
+            # "cam_right_wrist": OpenCVCamera(22),
+        }
+    )
+
+    # Allows to easily pick a subset of all devices
+    activated_leaders: list[str] | None = field(
+        default_factory=lambda: ["left", "right"]
+    )
+    activated_followers: list[str] | None = field(
+        default_factory=lambda: ["left", "right"]
+    )
+    activated_cameras: list[str] | None = field(
+        default_factory=lambda: ["cam_high", "cam_low", "cam_left_wrist", "cam_right_wrist"]
+    )
+
+
+class AlohaRobot():
+    """ Trossen Robotics
+
+    Example of usage:
+    ```python
+    robot = AlohaRobot()
+    ```
+    """
+
+    def __init__(self, config: AlohaRobotConfig | None = None, **kwargs):
+        if config is None:
+            config = AlohaRobotConfig()
+        # Overwrite config arguments using kwargs
+        config = replace(config, **kwargs)
+        self.config = config
+
+        self.leaders = {}
+        self.followers = {}
+        self.cameras = {}
+
+        if config.activated_leaders:
+            for name in config.activated_leaders:
+                info = config.leader_devices[name]
+                self.leaders[name] = Robot(info["port"], servo_ids=info["servos"])
+
+        if config.activated_followers:
+            for name in config.activated_followers:
+                info = config.follower_devices[name]
+                self.followers[name] = Robot(info["port"], servo_ids=info["servos"])
+
+        if config.activated_cameras:
+            for name in config.activated_cameras:
+                self.cameras[name] = config.camera_devices[name]
+
+    def init_teleop(self):
+        for name in self.followers:
+            self.followers[name]._enable_torque()
+        for name in self.cameras:
+            self.cameras[name].connect()
+
+    def teleop_step(self, record_data=False) -> None | tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
+        # Prepare to assign the positions of the leader to the follower
+        leader_pos = {}
+        for name in self.leaders:
+            leader_pos[name] = self.leaders[name].read_position()
+
+        # Update the position of the follower gripper to account for the different minimum and maximum range
+        # position in range [0, 4096[ which corresponds to 4096 bins of 360 degrees
+        # for all our dynamixel servors
+        # gripper id=8 has a different range from leader to follower
+        follower_goal_pos = {}
+        for name in self.leaders:
+            follower_goal_pos[name] = convert_gripper_range_from_leader_to_follower(leader_pos[name])
+
+        # Send action
+        for name in self.followers:
+            self.followers[name].set_goal_pos(follower_goal_pos[name])
+
+        # Early exit when recording data is not requested
+        if not record_data:
+            return
+
+        # Read follower position
+        follower_pos = {}
+        for name in self.followers:
+            follower_pos[name] = self.followers[name].read_position()
+
+        # Create state by concatenating follower current position
+        state = []
+        for name in ["left", "right"]:
+            if name in follower_pos:
+                state.append(follower_pos[name])
+        state = np.concatenate(state)
+        state = pwm2pos(state)
+
+        # Create action by concatenating follower goal position
+        action = []
+        for name in ["left", "right"]:
+            if name in follower_goal_pos:
+                action.append(follower_goal_pos[name])
+        action = np.concatenate(action)
+        action = pwm2pos(action)
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            images[name] = self.cameras[name].read()
+
+        # Populate output dictionnaries and format to pytorch
+        obs_dict, action_dict = {}, {}
+        obs_dict["observation.state"] = torch.from_numpy(state)
+        action_dict["action"] = torch.from_numpy(action)
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] =  torch.from_numpy(images[name])
+
+        return obs_dict, action_dict
+
+    def send_action(self, action):
+        from_idx = 0
+        to_idx = 0
+        follower_goal_pos = {}
+        for name in ["left", "right"]:
+            if name in self.followers:
+                to_idx += len(self.config.follower_devices[name]["servos"])
+                follower_goal_pos[name] = pos2pwm(action[from_idx:to_idx].numpy())
+                from_idx = to_idx
+
+        for name in self.followers:
+            self.followers[name].set_goal_pos(follower_goal_pos[name])

lerobot/common/robot_devices/robots/factory.py

@@ -0,0 +1,46 @@
+
+
+def make_robot(name):
+
+    if name == "koch":
+        from lerobot.common.robot_devices.robots.koch import KochRobot
+        from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
+        from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
+
+        robot = KochRobot(
+            leader_arms={
+                "main": DynamixelMotorsBus(
+                    port="/dev/tty.usbmodem575E0031751",
+                    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": DynamixelMotorsBus(
+                    port="/dev/tty.usbmodem575E0032081",
+                    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"),
+                    },
+                ),
+            },
+            cameras={
+                "main": OpenCVCamera(1, fps=30, width=640, height=480),
+            }
+        )
+    else:
+        raise ValueError(f"Robot '{name}' not found.")
+    
+    return robot

lerobot/common/robot_devices/robots/koch.py

@@ -0,0 +1,365 @@
+import copy
+from dataclasses import dataclass, field, replace
+from pathlib import Path
+import pickle
+import numpy as np
+import torch
+from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
+from lerobot.common.robot_devices.cameras.utils import Camera
+from lerobot.common.robot_devices.motors.dynamixel import DriveMode, DynamixelMotorsBus, OperatingMode, TorqueMode, motor_position_to_angle
+from lerobot.common.robot_devices.motors.utils import MotorsBus
+
+
+########################################################################
+# Calibration logic
+########################################################################
+
+# TARGET_HORIZONTAL_POSITION = motor_position_to_angle(np.array([0, -1024, 1024, 0, -1024, 0]))
+# TARGET_90_DEGREE_POSITION = motor_position_to_angle(np.array([1024, 0, 0, 1024, 0, -1024]))
+# GRIPPER_OPEN = motor_position_to_angle(np.array([-400]))
+
+TARGET_HORIZONTAL_POSITION = np.array([0, -1024, 1024, 0, -1024, 0])
+TARGET_90_DEGREE_POSITION = np.array([1024, 0, 0, 1024, 0, -1024])
+GRIPPER_OPEN = np.array([-400])
+
+def apply_homing_offset(values: np.array, homing_offset: np.array) -> np.array:
+    for i in range(len(values)):
+        if values[i] is not None:
+            values[i] += homing_offset[i]
+    return values
+
+
+def apply_drive_mode(values: np.array, drive_mode: np.array) -> np.array:
+    for i in range(len(values)):
+        if values[i] is not None and drive_mode[i]:
+            values[i] = -values[i]
+    return values
+
+def apply_calibration(values: np.array, homing_offset: np.array, drive_mode: np.array) -> np.array:
+    values = apply_drive_mode(values, drive_mode)
+    values = apply_homing_offset(values, homing_offset)
+    return values
+
+def revert_calibration(values: np.array, homing_offset: np.array, drive_mode: np.array) -> np.array:
+    """
+    Transform working position into real position for the robot.
+    """
+    values = apply_homing_offset(values, np.array([
+        -homing_offset if homing_offset is not None else None for homing_offset in homing_offset
+    ]))
+    values = apply_drive_mode(values, drive_mode)
+    return values
+
+
+def revert_appropriate_positions(positions: np.array, drive_mode: list[bool]) -> np.array:
+    for i, revert in enumerate(drive_mode):
+        if not revert and positions[i] is not None:
+            positions[i] = -positions[i]
+    return positions
+
+
+def compute_corrections(positions: np.array, drive_mode: list[bool], target_position: np.array) -> np.array:
+    correction = revert_appropriate_positions(positions, drive_mode)
+
+    for i in range(len(positions)):
+        if correction[i] is not None:
+            if drive_mode[i]:
+                correction[i] -= target_position[i]
+            else:
+                correction[i] += target_position[i]
+
+    return correction
+
+
+def compute_nearest_rounded_positions(positions: np.array) -> np.array:
+    return np.array(
+        [round(positions[i] / 1024) * 1024 if positions[i] is not None else None for i in range(len(positions))])
+
+
+def compute_homing_offset(arm: DynamixelMotorsBus, drive_mode: list[bool], target_position: np.array) -> np.array:
+    # Get the present positions of the servos
+    present_positions = apply_calibration(
+        arm.read("Present_Position"),
+        np.array([0, 0, 0, 0, 0, 0]),
+        drive_mode)
+
+    nearest_positions = compute_nearest_rounded_positions(present_positions)
+    correction = compute_corrections(nearest_positions, drive_mode, target_position)
+    return correction
+
+
+def compute_drive_mode(arm: DynamixelMotorsBus, offset: np.array):
+    # Get current positions
+    present_positions = apply_calibration(
+        arm.read("Present_Position"),
+        offset,
+        np.array([False, False, False, False, False, False]))
+
+    nearest_positions = compute_nearest_rounded_positions(present_positions)
+
+    # construct 'drive_mode' list comparing nearest_positions and TARGET_90_DEGREE_POSITION
+    drive_mode = []
+    for i in range(len(nearest_positions)):
+        drive_mode.append(nearest_positions[i] != TARGET_90_DEGREE_POSITION[i])
+    return drive_mode
+
+
+def reset_arm(arm: MotorsBus):
+    # To be configured, all servos must be in "torque disable" mode
+    arm.write("Torque_Enable", TorqueMode.DISABLED.value)
+
+    # Use 'extended position mode' for all motors except gripper, because in joint mode the servos can't
+    # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
+    # you could end up with a servo with a position 0 or 4095 at a crucial point See [
+    # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
+    all_motors_except_gripper = [name for name in arm.motor_names if name != "gripper"]
+    arm.write("Operating_Mode", OperatingMode.EXTENDED_POSITION.value, all_motors_except_gripper)
+
+    # TODO(rcadene): why?
+    # Use 'position control current based' for gripper
+    arm.write("Operating_Mode", OperatingMode.CURRENT_CONTROLLED_POSITION.value, "gripper")
+
+    # Make sure the native calibration (homing offset abd drive mode) is disabled, since we use our own calibration layer to be more generic
+    arm.write("Homing_Offset", 0)
+    arm.write("Drive_Mode", DriveMode.NON_INVERTED.value)
+
+
+def run_arm_calibration(arm: MotorsBus, name: str):
+    reset_arm(arm)
+
+    # TODO(rcadene): document what position 1 mean
+    print(f"Please move the '{name}' arm to the horizontal position (gripper fully closed)")
+    input("Press Enter to continue...")
+
+    horizontal_homing_offset = compute_homing_offset(arm, [False, False, False, False, False, False], TARGET_HORIZONTAL_POSITION)
+
+    # TODO(rcadene): document what position 2 mean
+    print(f"Please move the '{name}' arm to the 90 degree position (gripper fully open)")
+    input("Press Enter to continue...")
+
+    drive_mode = compute_drive_mode(arm, horizontal_homing_offset)
+    homing_offset = compute_homing_offset(arm, drive_mode, TARGET_90_DEGREE_POSITION)
+
+    # Invert offset for all drive_mode servos
+    for i in range(len(drive_mode)):
+        if drive_mode[i]:
+            homing_offset[i] = -homing_offset[i]
+
+    print("Calibration is done!")
+
+    print("=====================================")
+    print("      HOMING_OFFSET: ", " ".join([str(i) for i in homing_offset]))
+    print("      DRIVE_MODE: ", " ".join([str(i) for i in drive_mode]))
+    print("=====================================")
+
+    return homing_offset, drive_mode
+
+
+########################################################################
+# Alexander Koch robot arm
+########################################################################
+
+@dataclass
+class KochRobotConfig:
+    """
+    Example of usage:
+    ```python
+    KochRobotConfig()
+    ```
+    """
+
+    # Define all the components of the robot
+    leader_arms: dict[str, MotorsBus] = field(
+        default_factory=lambda: {
+            "main": DynamixelMotorsBus(
+                port="/dev/tty.usbmodem575E0031751",
+                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: dict[str, MotorsBus] = field(
+        default_factory=lambda: {
+            "main": DynamixelMotorsBus(
+                port="/dev/tty.usbmodem575E0032081",
+                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"),
+                },
+            ),
+        }
+    )
+    cameras: dict[str, Camera] = field(
+        default_factory=lambda: {}
+    )
+
+class KochRobot():
+    """ Tau Robotics: https://tau-robotics.com
+
+    Example of usage:
+    ```python
+    robot = KochRobot()
+    ```
+    """
+
+    def __init__(self, config: KochRobotConfig | None = None, calibration_path: Path = ".cache/calibration/koch.pkl", **kwargs):
+        if config is None:
+            config = KochRobotConfig()
+        # Overwrite config arguments using kwargs
+        self.config = replace(config, **kwargs)
+        self.calibration_path = Path(calibration_path)
+
+        self.leader_arms = self.config.leader_arms
+        self.follower_arms = self.config.follower_arms
+        self.cameras = self.config.cameras
+
+    def init_teleop(self):
+        if self.calibration_path.exists():
+            # Reset all arms before setting calibration
+            for name in self.follower_arms:
+                reset_arm(self.follower_arms[name])
+
+            for name in self.leader_arms:
+                reset_arm(self.leader_arms[name])
+
+            with open(self.calibration_path, 'rb') as f:
+                calibration = pickle.load(f)
+        else:
+            calibration = self.run_calibration()
+
+            self.calibration_path.parent.mkdir(parents=True, exist_ok=True)
+            with open(self.calibration_path, 'wb') as f:
+                pickle.dump(calibration, f)
+
+        for name in self.follower_arms:
+            self.follower_arms[name].set_calibration(calibration[f"follower_{name}"])
+            self.follower_arms[name].write("Torque_Enable", 1)
+
+        for name in self.leader_arms:
+            self.leader_arms[name].set_calibration(calibration[f"leader_{name}"])
+            # TODO(rcadene): add comments
+            self.leader_arms[name].write("Goal_Position", GRIPPER_OPEN, "gripper")
+            self.leader_arms[name].write("Torque_Enable", 1, "gripper")
+
+        for name in self.cameras:
+            self.cameras[name].connect()
+
+    def run_calibration(self):
+        calibration = {}
+
+        for name in self.follower_arms:
+            homing_offset, drive_mode = run_arm_calibration(self.follower_arms[name], f"{name} follower")
+            
+            calibration[f"follower_{name}"] = {}
+            for idx, motor_name in enumerate(self.follower_arms[name].motor_names):
+                calibration[f"follower_{name}"][motor_name] = (homing_offset[idx], drive_mode[idx])
+
+        for name in self.leader_arms:
+            homing_offset, drive_mode = run_arm_calibration(self.leader_arms[name], f"{name} leader")
+            
+            calibration[f"leader_{name}"] = {}
+            for idx, motor_name in enumerate(self.leader_arms[name].motor_names):
+                calibration[f"leader_{name}"][motor_name] = (homing_offset[idx], drive_mode[idx])
+
+        return calibration
+
+    def teleop_step(self, record_data=False) -> None | tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
+        # Prepare to assign the positions of the leader to the follower
+        leader_pos = {}
+        for name in self.leader_arms:
+            leader_pos[name] = self.leader_arms[name].read("Present_Position")
+
+        follower_goal_pos = {}
+        for name in self.leader_arms:
+            follower_goal_pos[name] = leader_pos[name]
+
+        # Send action
+        for name in self.follower_arms:
+            self.follower_arms[name].write("Goal_Position", follower_goal_pos[name])
+
+        # Early exit when recording data is not requested
+        if not record_data:
+            return
+
+        # Read follower position
+        follower_pos = {}
+        for name in self.follower_arms:
+            follower_pos[name] = self.follower_arms[name].read("Present_Position")
+
+        # Create state by concatenating follower current position
+        state = []
+        for name in self.follower_arms:
+            if name in follower_pos:
+                state.append(follower_pos[name])
+        state = np.concatenate(state)
+
+        # Create action by concatenating follower goal position
+        action = []
+        for name in self.follower_arms:
+            if name in follower_goal_pos:
+                action.append(follower_goal_pos[name])
+        action = np.concatenate(action)
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            images[name] = self.cameras[name].read()
+
+        # Populate output dictionnaries and format to pytorch
+        obs_dict, action_dict = {}, {}
+        obs_dict["observation.state"] = torch.from_numpy(state)
+        action_dict["action"] = torch.from_numpy(action)
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] =  torch.from_numpy(images[name])
+
+        return obs_dict, action_dict
+
+    def capture_observation(self):
+        # Read follower position
+        follower_pos = {}
+        for name in self.follower_arms:
+            follower_pos[name] = self.follower_arms[name].read("Present_Position")
+
+        # Create state by concatenating follower current position
+        state = []
+        for name in self.follower_arms:
+            if name in follower_pos:
+                state.append(follower_pos[name])
+        state = np.concatenate(state)
+
+        # Capture images from cameras
+        images = {}
+        for name in self.cameras:
+            images[name] = self.cameras[name].read()
+
+        # Populate output dictionnaries and format to pytorch
+        obs_dict = {}
+        obs_dict["observation.state"] = torch.from_numpy(state)
+        for name in self.cameras:
+            obs_dict[f"observation.images.{name}"] = torch.from_numpy(images[name])
+        return obs_dict
+
+    def send_action(self, action):
+        from_idx = 0
+        to_idx = 0
+        follower_goal_pos = {}
+        for name in self.follower_arms:
+            if name in self.follower_arms:
+                to_idx += len(self.follower_arms[name].motor_names)
+                follower_goal_pos[name] = action[from_idx:to_idx].numpy()
+                from_idx = to_idx
+
+        for name in self.follower_arms:
+            self.follower_arms[name].write("Goal_Position", follower_goal_pos[name].astype(np.int32))

lerobot/common/robot_devices/robots/utils.py

@@ -0,0 +1,8 @@
+from typing import Protocol
+
+class Robot(Protocol):
+    def init_teleop(self): ...
+    def run_calibration(self): ...
+    def teleop_step(self, record_data=False): ...
+    def capture_observation(self): ...
+    def send_action(self, action): ...

lerobot/common/utils/import_utils.py

@@ -1,3 +1,18 @@
+#!/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 importlib
 import logging
 

lerobot/common/utils/io_utils.py

@@ -1,3 +1,18 @@
+#!/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 warnings
 
 import imageio
@@ -10,3 +25,42 @@ def write_video(video_path, stacked_frames, fps):
             "ignore", "pkg_resources is deprecated as an API", category=DeprecationWarning
         )
         imageio.mimsave(video_path, stacked_frames, fps=fps)
+
+
+import serial
+import os
+import time
+
+def reset_usb_port(port):
+    try:
+        # Close the serial port if it's open
+        ser = serial.Serial(port)
+        ser.close()
+    except serial.serialutil.SerialException as e:
+        print(f"Exception while closing the port: {e}")
+
+    # Find the USB device path
+    usb_device_path = None
+    for root, dirs, files in os.walk('/sys/bus/usb/drivers/usb'):
+        for dir_name in dirs:
+            if port in dir_name:
+                usb_device_path = os.path.join(root, dir_name)
+                break
+
+    if usb_device_path:
+        # Unbind and rebind the USB device
+        try:
+            unbind_path = os.path.join(usb_device_path, 'unbind')
+            bind_path = os.path.join(usb_device_path, 'bind')
+            usb_id = os.path.basename(usb_device_path)
+
+            with open(unbind_path, 'w') as f:
+                f.write(usb_id)
+            time.sleep(1)  # Wait for a second
+            with open(bind_path, 'w') as f:
+                f.write(usb_id)
+            print(f"USB port {port} has been reset.")
+        except Exception as e:
+            print(f"Exception during USB reset: {e}")
+    else:
+        print(f"Could not find USB device path for port: {port}")

lerobot/common/utils/utils.py

@@ -1,8 +1,25 @@
+#!/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 logging
 import os.path as osp
 import random
+from contextlib import contextmanager
 from datetime import datetime
 from pathlib import Path
+from typing import Any, Generator
 
 import hydra
 import numpy as np
@@ -31,12 +48,57 @@ def get_safe_torch_device(cfg_device: str, log: bool = False) -> torch.device:
     return device
 
 
+def get_global_random_state() -> dict[str, Any]:
+    """Get the random state for `random`, `numpy`, and `torch`."""
+    random_state_dict = {
+        "random_state": random.getstate(),
+        "numpy_random_state": np.random.get_state(),
+        "torch_random_state": torch.random.get_rng_state(),
+    }
+    if torch.cuda.is_available():
+        random_state_dict["torch_cuda_random_state"] = torch.cuda.random.get_rng_state()
+    return random_state_dict
+
+
+def set_global_random_state(random_state_dict: dict[str, Any]):
+    """Set the random state for `random`, `numpy`, and `torch`.
+
+    Args:
+        random_state_dict: A dictionary of the form returned by `get_global_random_state`.
+    """
+    random.setstate(random_state_dict["random_state"])
+    np.random.set_state(random_state_dict["numpy_random_state"])
+    torch.random.set_rng_state(random_state_dict["torch_random_state"])
+    if torch.cuda.is_available():
+        torch.cuda.random.set_rng_state(random_state_dict["torch_cuda_random_state"])
+
+
 def set_global_seed(seed):
     """Set seed for reproducibility."""
     random.seed(seed)
     np.random.seed(seed)
     torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+@contextmanager
+def seeded_context(seed: int) -> Generator[None, None, None]:
+    """Set the seed when entering a context, and restore the prior random state at exit.
+
+    Example usage:
+
+    ```
+    a = random.random()  # produces some random number
+    with seeded_context(1337):
+        b = random.random()  # produces some other random number
+    c = random.random()  # produces yet another random number, but the same it would have if we never made `b`
+    ```
+    """
+    random_state_dict = get_global_random_state()
+    set_global_seed(seed)
+    yield None
+    set_global_random_state(random_state_dict)
 
 
 def init_logging():
@@ -58,13 +120,13 @@ def init_logging():
     logging.getLogger().addHandler(console_handler)
 
 
-def format_big_number(num):
+def format_big_number(num, precision=0):
     suffixes = ["", "K", "M", "B", "T", "Q"]
     divisor = 1000.0
 
     for suffix in suffixes:
         if abs(num) < divisor:
-            return f"{num:.0f}{suffix}"
+            return f"{num:.{precision}f}{suffix}"
         num /= divisor
 
     return num

lerobot/configs/default.yaml

@@ -5,18 +5,33 @@ defaults:
 
 hydra:
   run:
+    # Set `dir` to where you would like to save all of the run outputs. If you run another training session
+    # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
     dir: outputs/train/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${policy.name}_${hydra.job.name}
   job:
     name: default
 
+# Set `resume` to true to resume a previous run. In order for this to work, you will need to make sure
+# `hydra.run.dir` is the directory of an existing run with at least one checkpoint in it.
+# Note that when resuming a run, the default behavior is to use the configuration from the checkpoint,
+# regardless of what's provided with the training command at the time of resumption.
+resume: false
 device: cuda  # cpu
+# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
+# automatic gradient scaling is used.
+use_amp: false
 # `seed` is used for training (eg: model initialization, dataset shuffling)
 # AND for the evaluation environments.
 seed: ???
+# You may provide a list of datasets here. `train.py` creates them all and concatenates them. Note: only data
+# keys common between the datasets are kept. Each dataset gets and additional transform that inserts the
+# "dataset_index" into the returned item. The index mapping is made according to the order in which the
+# datsets are provided.
 dataset_repo_id: lerobot/pusht
 
 training:
   offline_steps: ???
+  # NOTE: `online_steps` is not implemented yet. It's here as a placeholder.
   online_steps: ???
   online_steps_between_rollouts: ???
   online_sampling_ratio: 0.5
@@ -25,7 +40,9 @@ training:
   eval_freq: ???
   save_freq: ???
   log_freq: 250
-  save_model: true
+  save_checkpoint: true
+  num_workers: 4
+  batch_size: ???
 
 eval:
   n_episodes: 1
@@ -33,10 +50,12 @@ eval:
   batch_size: 1
   # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
   use_async_envs: false
+  # Specify the number of episodes to render during evaluation.
+  max_episodes_rendered: 10
 
 wandb:
-  enable: true
-  # Set to true to disable saving an artifact despite save_model == True
+  enable: false
+  # Set to true to disable saving an artifact despite save_checkpoint == True
   disable_artifact: false
   project: lerobot
   notes: ""

lerobot/configs/env/aloha.yaml

@@ -5,10 +5,11 @@ fps: 50
 env:
   name: aloha
   task: AlohaInsertion-v0
-  from_pixels: True
-  pixels_only: False
-  image_size: [3, 480, 640]
-  episode_length: 400
-  fps: ${fps}
   state_dim: 14
   action_dim: 14
+  fps: ${fps}
+  episode_length: 400
+  real_world: false
+  gym:
+    obs_type: pixels_agent_pos
+    render_mode: rgb_array

lerobot/configs/env/dora_aloha_real.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+
+fps: 30
+
+env:
+  name: dora
+  task: DoraAloha-v0
+  state_dim: 14
+  action_dim: 14
+  fps: ${fps}
+  episode_length: 400
+  real_world: true
+  gym:
+    fps: ${fps}

lerobot/configs/env/pusht.yaml

@@ -5,10 +5,14 @@ fps: 10
 env:
   name: pusht
   task: PushT-v0
-  from_pixels: True
-  pixels_only: False
   image_size: 96
-  episode_length: 300
-  fps: ${fps}
   state_dim: 2
   action_dim: 2
+  fps: ${fps}
+  episode_length: 300
+  real_world: false
+  gym:
+    obs_type: pixels_agent_pos
+    render_mode: rgb_array
+    visualization_width: 384
+    visualization_height: 384

lerobot/configs/env/xarm.yaml

@@ -5,10 +5,14 @@ fps: 15
 env:
   name: xarm
   task: XarmLift-v0
-  from_pixels: True
-  pixels_only: False
   image_size: 84
-  episode_length: 25
-  fps: ${fps}
   state_dim: 4
   action_dim: 4
+  fps: ${fps}
+  episode_length: 25
+  real_world: false
+  gym:
+    obs_type: pixels_agent_pos
+    render_mode: rgb_array
+    visualization_width: 384
+    visualization_height: 384

lerobot/configs/policy/act.yaml

@@ -3,13 +3,19 @@
 seed: 1000
 dataset_repo_id: lerobot/aloha_sim_insertion_human
 
+override_dataset_stats:
+  observation.images.top:
+    # 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
+  save_checkpoint: true
 
   batch_size: 8
   lr: 1e-5
@@ -18,12 +24,6 @@ training:
   grad_clip_norm: 10
   online_steps_between_rollouts: 1
 
-  override_dataset_stats:
-    observation.images.top:
-      # 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)
-
   delta_timestamps:
     action: "[i / ${fps} for i in range(${policy.chunk_size})]"
 
@@ -66,6 +66,9 @@ policy:
   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
@@ -73,7 +76,7 @@ policy:
   n_vae_encoder_layers: 4
 
   # Inference.
-  use_temporal_aggregation: false
+  temporal_ensemble_momentum: null
 
   # Training and loss computation.
   dropout: 0.1

lerobot/configs/policy/act_real.yaml

@@ -0,0 +1,115 @@
+# @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=dora_aloha_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]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${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
+    observation.state: 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

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=dora_aloha_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

lerobot/configs/policy/diffusion.yaml

@@ -7,13 +7,27 @@
 seed: 100000
 dataset_repo_id: lerobot/pusht
 
+override_dataset_stats:
+  # TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?
+  observation.image:
+    mean: [[[0.5]], [[0.5]], [[0.5]]]  # (c,1,1)
+    std: [[[0.5]], [[0.5]], [[0.5]]]  # (c,1,1)
+  # TODO(rcadene, alexander-soare): we override state and action stats to use the same as the pretrained model
+  # from the original codebase, but we should remove these and train our own pretrained model
+  observation.state:
+    min: [13.456424, 32.938293]
+    max: [496.14618, 510.9579]
+  action:
+    min: [12.0, 25.0]
+    max: [511.0, 511.0]
+
 training:
   offline_steps: 200000
   online_steps: 0
   eval_freq: 5000
   save_freq: 5000
   log_freq: 250
-  save_model: true
+  save_checkpoint: true
 
   batch_size: 64
   grad_clip_norm: 10
@@ -30,24 +44,14 @@ training:
     observation.state: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
     action: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"
 
+  # The original implementation doesn't sample frames for the last 7 steps,
+  # which avoids excessive padding and leads to improved training results.
+  drop_n_last_frames: 7  # ${policy.horizon} - ${policy.n_action_steps} - ${policy.n_obs_steps} + 1
+
 eval:
   n_episodes: 50
   batch_size: 50
 
-override_dataset_stats:
-  # TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?
-  observation.image:
-    mean: [[[0.5]], [[0.5]], [[0.5]]]  # (c,1,1)
-    std: [[[0.5]], [[0.5]], [[0.5]]]  # (c,1,1)
-  # TODO(rcadene, alexander-soare): we override state and action stats to use the same as the pretrained model
-  # from the original codebase, but we should remove these and train our own pretrained model
-  observation.state:
-    min: [13.456424, 32.938293]
-    max: [496.14618, 510.9579]
-  action:
-    min: [12.0, 25.0]
-    max: [511.0, 511.0]
-
 policy:
   name: diffusion
 
@@ -85,6 +89,7 @@ policy:
   diffusion_step_embed_dim: 128
   use_film_scale_modulation: True
   # Noise scheduler.
+  noise_scheduler_type: DDPM
   num_train_timesteps: 100
   beta_schedule: squaredcos_cap_v2
   beta_start: 0.0001

lerobot/configs/policy/tdmpc.yaml

@@ -1,11 +1,12 @@
 # @package _global_
 
 seed: 1
-dataset_repo_id: lerobot/xarm_lift_medium_replay
+dataset_repo_id: lerobot/xarm_lift_medium
 
 training:
   offline_steps: 25000
-  online_steps: 25000
+  # TODO(alexander-soare): uncomment when online training gets reinstated
+  online_steps: 0  # 25000 not implemented yet
   eval_freq: 5000
   online_steps_between_rollouts: 1
   online_sampling_ratio: 0.5

lerobot/scripts/control_robot.py

@@ -0,0 +1,400 @@
+"""
+Example of usage:
+
+- Unlimited teleoperation at highest frequency (~200 Hz is expected), to exit with CTRL+C:
+```bash
+python lerobot/scripts/control_robot.py teleoperate
+```
+
+- Unlimited teleoperation at a limited frequency of 30 Hz, to simulate data recording frequency:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --fps 30
+```
+
+- Record one episode in order to test replay:
+```bash
+python lerobot/scripts/control_robot.py record_dataset \
+    --fps 30 \
+    --root tmp/data \
+    --repo-id $USER/koch_test \
+    --num-episodes 1 \
+    --run-compute-stats 0
+```
+
+- Visualize dataset:
+```bash
+python lerobot/scripts/visualize_dataset.py \
+    --root tmp/data \
+    --repo-id $USER/koch_test \
+    --episode-index 0
+```
+
+- Replay this test episode:
+```bash
+python lerobot/scripts/control_robot.py replay_episode \
+    --fps 30 \
+    --root tmp/data \
+    --repo-id $USER/koch_test \
+    --episode 0
+```
+
+- Record a full dataset in order to train a policy:
+```bash
+python lerobot/scripts/control_robot.py record_dataset \
+    --fps 30 \
+    --root data \
+    --repo-id $USER/koch_pick_place_lego \
+    --num-episodes 50 \
+    --run-compute-stats 1
+```
+
+- Train on this dataset (TODO(rcadene)):
+```bash
+python lerobot/scripts/train.py
+```
+
+- Run the pretrained policy on the robot:
+```bash
+python lerobot/scripts/control_robot.py run_policy \
+    -p TODO(rcadene)
+```
+"""
+
+import argparse
+from contextlib import nullcontext
+import os
+from pathlib import Path
+import shutil
+import time
+
+from PIL import Image
+from omegaconf import DictConfig
+import torch
+from lerobot.common.datasets.compute_stats import compute_stats
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import to_hf_dataset
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
+from lerobot.common.datasets.utils import calculate_episode_data_index, load_hf_dataset
+from lerobot.common.datasets.video_utils import encode_video_frames
+from lerobot.common.policies.factory import make_policy
+from lerobot.common.robot_devices.robots.factory import make_robot
+from lerobot.common.robot_devices.robots.utils import Robot
+from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, set_global_seed
+from lerobot.scripts.eval import get_pretrained_policy_path
+from lerobot.scripts.push_dataset_to_hub import save_meta_data
+from lerobot.scripts.robot_controls.record_dataset import record_dataset
+import concurrent.futures
+
+
+########################################################################################
+# Utilities
+########################################################################################
+
+def save_image(img_tensor, key, frame_index, episode_index, videos_dir):
+    img = Image.fromarray(img_tensor.numpy())
+    path = videos_dir / f"{key}_episode_{episode_index:06d}" / f"frame_{frame_index:06d}.png"
+    path.parent.mkdir(parents=True, exist_ok=True)
+    img.save(str(path), quality=100)
+
+
+def busy_wait(seconds):
+    # Significantly more accurate than `time.sleep`, and mendatory for our use case,
+    # but it consumes CPU cycles.
+    # TODO(rcadene): find an alternative
+    end_time = time.perf_counter() + seconds
+    while time.perf_counter() < end_time:
+        pass
+
+def none_or_int(value):
+    if value == 'None':
+        return None
+    return int(value)
+
+########################################################################################
+# Control modes
+########################################################################################
+
+def teleoperate(robot: Robot, fps: int | None = None):
+    robot.init_teleop()
+
+    while True:
+        now = time.perf_counter()
+        robot.teleop_step()
+
+        if fps is not None:
+            dt_s = (time.perf_counter() - now)
+            busy_wait(1 / fps - dt_s)
+
+        dt_s = (time.perf_counter() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+
+def record_dataset(robot: Robot, fps: int | None = None, root="data", repo_id="lerobot/debug", warmup_time_s=2, episode_time_s=10, num_episodes=50, video=True, run_compute_stats=True):
+    if not video:
+        raise NotImplementedError()
+
+    robot.init_teleop()
+
+    local_dir = Path(root) / repo_id
+    if local_dir.exists():
+        shutil.rmtree(local_dir)
+
+    videos_dir = local_dir / "videos"
+    videos_dir.mkdir(parents=True, exist_ok=True)
+
+
+    start_time = time.perf_counter()
+
+    is_warmup_print = False
+    is_record_print = False
+    ep_dicts = []
+
+    # Save images using threads to reach high fps (30 and more)
+    # Using `with` ensures the program exists smoothly if an execption is raised.
+    with concurrent.futures.ThreadPoolExecutor() as executor:
+        for episode_index in range(num_episodes):
+
+            ep_dict = {}
+            frame_index = 0
+
+            while True:
+                if not is_warmup_print:
+                    print("Warming up by skipping frames")
+                    os.system('say "Warmup"')
+                    is_warmup_print = True
+                now = time.perf_counter()
+
+                observation, action = robot.teleop_step(record_data=True)
+                timestamp = time.perf_counter() - start_time
+
+                if timestamp < warmup_time_s:
+                    dt_s = (time.perf_counter() - now)
+                    busy_wait(1 / fps - dt_s)
+
+                    dt_s = (time.perf_counter() - now)
+                    print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f} (Warmup)")
+                    continue
+
+                if not is_record_print:
+                    print("Recording")
+                    os.system(f'say "Recording episode {episode_index}"')
+                    is_record_print = True
+
+                image_keys = [key for key in observation if "image" in key]
+                not_image_keys = [key for key in observation if "image" not in key]
+
+                for key in image_keys:
+                    executor.submit(save_image, observation[key], key, frame_index, episode_index, videos_dir)
+
+                for key in not_image_keys:
+                    if key not in ep_dict:
+                        ep_dict[key] = []
+                    ep_dict[key].append(observation[key])
+
+                for key in action:
+                    if key not in ep_dict:
+                        ep_dict[key] = []
+                    ep_dict[key].append(action[key])
+
+                frame_index += 1
+
+                dt_s = (time.perf_counter() - now)
+                busy_wait(1 / fps - dt_s)
+
+                dt_s = (time.perf_counter() - now)
+                print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+                if timestamp > episode_time_s - warmup_time_s:
+                    break
+
+            print("Encoding to `LeRobotDataset` format")
+            os.system('say "Encoding"')
+
+            num_frames = frame_index
+
+            for key in image_keys:
+                tmp_imgs_dir = videos_dir / f"{key}_episode_{episode_index:06d}"
+                fname = f"{key}_episode_{episode_index:06d}.mp4"
+                video_path = local_dir / "videos" / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+                # clean temporary images directory
+                # shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[key] = []
+                for i in range(num_frames):
+                    ep_dict[key].append({"path": f"videos/{fname}", "timestamp": i / fps})
+
+            for key in not_image_keys:
+                ep_dict[key] = torch.stack(ep_dict[key])
+            
+            for key in action:
+                ep_dict[key] = torch.stack(ep_dict[key])
+
+            ep_dict["episode_index"] = torch.tensor([episode_index] * num_frames)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+
+            done = torch.zeros(num_frames, dtype=torch.bool)
+            done[-1] = True
+            ep_dict["next.done"] = 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)
+
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+
+    meta_data_dir = local_dir / "meta_data"
+
+    for key in image_keys:
+        time.sleep(10)
+        tmp_imgs_dir = videos_dir / f"{key}_episode_{episode_index:06d}"
+        # shutil.rmtree(tmp_imgs_dir)
+
+    lerobot_dataset = LeRobotDataset.from_preloaded(
+        repo_id=repo_id,
+        hf_dataset=hf_dataset,
+        episode_data_index=episode_data_index,
+        info=info,
+        videos_dir=videos_dir,
+    )
+    if run_compute_stats:
+        stats = compute_stats(lerobot_dataset)
+    else:
+        stats = {}
+
+    hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
+    hf_dataset.save_to_disk(str(local_dir / "train"))
+
+    save_meta_data(info, stats, episode_data_index, meta_data_dir)
+
+    # TODO(rcadene): push to hub
+
+
+def replay_episode(robot: Robot, episode: int, fps: int | None = None, root="data", repo_id="lerobot/debug"):
+    local_dir = Path(root) / repo_id
+    if not local_dir.exists():
+        raise ValueError(local_dir)
+
+    dataset = LeRobotDataset(repo_id, root=root)
+    items = dataset.hf_dataset.select_columns("action")
+    from_idx = dataset.episode_data_index["from"][episode].item()
+    to_idx = dataset.episode_data_index["to"][episode].item()
+
+    robot.init_teleop()
+    
+    print("Replaying episode")
+    os.system('say "Replaying episode"')
+
+    for idx in range(from_idx, to_idx):
+        now = time.perf_counter()
+
+        action = items[idx]["action"]
+        robot.send_action(action)
+
+        dt_s = (time.perf_counter() - now)
+        busy_wait(1 / fps - dt_s)
+
+        dt_s = (time.perf_counter() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+
+def run_policy(robot: Robot, policy: torch.nn.Module, hydra_cfg: DictConfig):
+    policy.eval()
+
+    # Check device is available
+    device = get_safe_torch_device(hydra_cfg.device, log=True)
+
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+    set_global_seed(hydra_cfg.seed)
+
+    fps = hydra_cfg.env.fps
+
+    while True:
+        now = time.perf_counter()
+
+        observation = robot.capture_observation()
+
+        with torch.inference_mode(), torch.autocast(device_type=device.type) if hydra_cfg.use_amp else nullcontext():
+            action = policy.select_action(observation)
+
+        robot.send_action(action)
+
+        dt_s = (time.perf_counter() - now)
+        busy_wait(1 / fps - dt_s)
+
+        dt_s = (time.perf_counter() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    subparsers = parser.add_subparsers(dest="mode", required=True)
+
+    # Set common options for all the subparsers
+    base_parser = argparse.ArgumentParser(add_help=False)
+    base_parser.add_argument("--robot", type=str, default="koch", help="Name of the robot provided to the `make_robot(name)` factory function.")
+
+    parser_teleop = subparsers.add_parser("teleoperate", parents=[base_parser])
+    parser_teleop.add_argument('--fps', type=none_or_int, default=None, help='Frames per second (set to None to disable)')
+
+    parser_record = subparsers.add_parser("record_dataset", parents=[base_parser])
+    parser_record.add_argument('--fps', type=none_or_int, default=None, help='Frames per second (set to None to disable)')
+    parser_record.add_argument('--root', type=Path, default="data", help='')
+    parser_record.add_argument('--repo-id', type=str, default="lerobot/test", help='')
+    parser_record.add_argument('--warmup-time-s', type=int, default=2, help='')
+    parser_record.add_argument('--episode-time-s', type=int, default=10, help='')
+    parser_record.add_argument('--num-episodes', type=int, default=50, help='')
+    parser_record.add_argument('--run-compute-stats', type=int, default=1, help='')
+
+    parser_replay = subparsers.add_parser("replay_episode", parents=[base_parser])
+    parser_replay.add_argument('--fps', type=none_or_int, default=None, help='Frames per second (set to None to disable)')
+    parser_replay.add_argument('--root', type=Path, default="data", help='')
+    parser_replay.add_argument('--repo-id', type=str, default="lerobot/test", help='')
+    parser_replay.add_argument('--episode', type=int, default=0, help='')
+
+    parser_policy = subparsers.add_parser("run_policy", parents=[base_parser])
+    parser_policy.add_argument('-p', '--pretrained-policy-name-or-path', type=str,
+        help=(
+            "Either the repo ID of a model hosted on the Hub or a path to a directory containing weights "
+            "saved using `Policy.save_pretrained`."
+        )
+    )
+    parser_policy.add_argument(
+        "overrides",
+        nargs="*",
+        help="Any key=value arguments to override config values (use dots for.nested=overrides)",
+    )
+    args = parser.parse_args()
+
+    control_mode = args.mode
+    robot_name = args.robot
+    kwargs = vars(args)
+    del kwargs["mode"]
+    del kwargs["robot"]
+
+    robot = make_robot(robot_name)
+    if control_mode == "teleoperate":
+        teleoperate(robot, **kwargs)
+    elif control_mode == "record_dataset":
+        record_dataset(robot, **kwargs)
+    elif control_mode == "replay_episode":
+        replay_episode(robot, **kwargs)
+
+    elif control_mode == "run_policy":
+        pretrained_policy_path = get_pretrained_policy_path(args.pretrained_policy_name_or_path)
+        hydra_cfg = init_hydra_config(pretrained_policy_path / "config.yaml", args.overrides)
+        policy = make_policy(hydra_cfg=hydra_cfg, pretrained_policy_name_or_path=pretrained_policy_path)
+        run_policy(robot, policy, hydra_cfg)

lerobot/scripts/display_sys_info.py

@@ -1,3 +1,18 @@
+#!/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 platform
 
 import huggingface_hub

lerobot/scripts/eval.py

@@ -1,3 +1,18 @@
+#!/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.
 """Evaluate a policy on an environment by running rollouts and computing metrics.
 
 Usage examples:
@@ -13,7 +28,7 @@ OR, you want to evaluate a model checkpoint from the LeRobot training script for
 
 ```
 python lerobot/scripts/eval.py \
-    -p outputs/train/diffusion_pusht/checkpoints/005000 \
+    -p outputs/train/diffusion_pusht/checkpoints/005000/pretrained_model \
     eval.n_episodes=10
 ```
 
@@ -29,8 +44,10 @@ https://huggingface.co/lerobot/diffusion_pusht/tree/main.
 import argparse
 import json
 import logging
+import os
 import threading
 import time
+from contextlib import nullcontext
 from copy import deepcopy
 from datetime import datetime as dt
 from pathlib import Path
@@ -148,7 +165,10 @@ def rollout(
         # VectorEnv stores is_success in `info["final_info"][env_index]["is_success"]`. "final_info" isn't
         # available of none of the envs finished.
         if "final_info" in info:
-            successes = [info["is_success"] if info is not None else False for info in info["final_info"]]
+            successes = [
+                i["is_success"] if (i is not None and "is_success" in i) else False
+                for i in info["final_info"]
+            ]
         else:
             successes = [False] * env.num_envs
 
@@ -500,12 +520,13 @@ def eval(
     out_dir = (
         f"outputs/eval/{dt.now().strftime('%Y-%m-%d/%H-%M-%S')}_{hydra_cfg.env.name}_{hydra_cfg.policy.name}"
     )
+    os.makedirs(out_dir, exist_ok=True)
 
     if out_dir is None:
         raise NotImplementedError()
 
     # Check device is available
-    get_safe_torch_device(hydra_cfg.device, log=True)
+    device = get_safe_torch_device(hydra_cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
@@ -524,16 +545,17 @@ def eval(
         policy = make_policy(hydra_cfg=hydra_cfg, dataset_stats=make_dataset(hydra_cfg).stats)
     policy.eval()
 
-    info = eval_policy(
-        env,
-        policy,
-        hydra_cfg.eval.n_episodes,
-        max_episodes_rendered=10,
-        video_dir=Path(out_dir) / "eval",
-        start_seed=hydra_cfg.seed,
-        enable_progbar=True,
-        enable_inner_progbar=True,
-    )
+    with torch.no_grad(), torch.autocast(device_type=device.type) if hydra_cfg.use_amp else nullcontext():
+        info = eval_policy(
+            env,
+            policy,
+            hydra_cfg.eval.n_episodes,
+            max_episodes_rendered=hydra_cfg.eval.max_episodes_rendered,
+            video_dir=Path(out_dir) / "eval",
+            start_seed=hydra_cfg.seed,
+            enable_progbar=True,
+            enable_inner_progbar=True,
+        )
     print(info["aggregated"])
 
     # Save info
@@ -545,6 +567,31 @@ def eval(
     logging.info("End of eval")
 
 
+def get_pretrained_policy_path(pretrained_policy_name_or_path, revision=None):
+    try:
+        pretrained_policy_path = Path(
+            snapshot_download(pretrained_policy_name_or_path, revision=revision)
+        )
+    except (HFValidationError, RepositoryNotFoundError) as e:
+        if isinstance(e, HFValidationError):
+            error_message = (
+                "The provided pretrained_policy_name_or_path is not a valid Hugging Face Hub repo ID."
+            )
+        else:
+            error_message = (
+                "The provided pretrained_policy_name_or_path was not found on the Hugging Face Hub."
+            )
+
+        logging.warning(f"{error_message} Treating it as a local directory.")
+        pretrained_policy_path = Path(pretrained_policy_name_or_path)
+    if not pretrained_policy_path.is_dir() or not pretrained_policy_path.exists():
+        raise ValueError(
+            "The provided pretrained_policy_name_or_path is not a valid/existing Hugging Face Hub "
+            "repo ID, nor is it an existing local directory."
+        )
+    return pretrained_policy_path
+
+
 if __name__ == "__main__":
     init_logging()
 
@@ -579,26 +626,6 @@ if __name__ == "__main__":
     if args.pretrained_policy_name_or_path is None:
         eval(hydra_cfg_path=args.config, config_overrides=args.overrides)
     else:
-        try:
-            pretrained_policy_path = Path(
-                snapshot_download(args.pretrained_policy_name_or_path, revision=args.revision)
-            )
-        except (HFValidationError, RepositoryNotFoundError) as e:
-            if isinstance(e, HFValidationError):
-                error_message = (
-                    "The provided pretrained_policy_name_or_path is not a valid Hugging Face Hub repo ID."
-                )
-            else:
-                error_message = (
-                    "The provided pretrained_policy_name_or_path was not found on the Hugging Face Hub."
-                )
-
-            logging.warning(f"{error_message} Treating it as a local directory.")
-            pretrained_policy_path = Path(args.pretrained_policy_name_or_path)
-        if not pretrained_policy_path.is_dir() or not pretrained_policy_path.exists():
-            raise ValueError(
-                "The provided pretrained_policy_name_or_path is not a valid/existing Hugging Face Hub "
-                "repo ID, nor is it an existing local directory."
-            )
+        pretrained_policy_path = get_pretrained_policy_path(args.pretrained_policy_name_or_path, revision=args.revision)
 
         eval(pretrained_policy_path=pretrained_policy_path, config_overrides=args.overrides)

lerobot/scripts/push_dataset_to_hub.py

@@ -1,3 +1,18 @@
+#!/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 convert your dataset into LeRobot dataset format and upload it to the Hugging Face hub,
 or store it locally. LeRobot dataset format is lightweight, fast to load from, and does not require any
@@ -10,7 +25,6 @@ python lerobot/scripts/push_dataset_to_hub.py \
 --dataset-id pusht \
 --raw-format pusht_zarr \
 --community-id lerobot \
---revision v1.2 \
 --dry-run 1 \
 --save-to-disk 1 \
 --save-tests-to-disk 0 \
@@ -21,7 +35,6 @@ python lerobot/scripts/push_dataset_to_hub.py \
 --dataset-id xarm_lift_medium \
 --raw-format xarm_pkl \
 --community-id lerobot \
---revision v1.2 \
 --dry-run 1 \
 --save-to-disk 1 \
 --save-tests-to-disk 0 \
@@ -32,7 +45,6 @@ python lerobot/scripts/push_dataset_to_hub.py \
 --dataset-id aloha_sim_insertion_scripted \
 --raw-format aloha_hdf5 \
 --community-id lerobot \
---revision v1.2 \
 --dry-run 1 \
 --save-to-disk 1 \
 --save-tests-to-disk 0 \
@@ -43,7 +55,6 @@ python lerobot/scripts/push_dataset_to_hub.py \
 --dataset-id umi_cup_in_the_wild \
 --raw-format umi_zarr \
 --community-id lerobot \
---revision v1.2 \
 --dry-run 1 \
 --save-to-disk 1 \
 --save-tests-to-disk 0 \
@@ -60,9 +71,9 @@ import torch
 from huggingface_hub import HfApi
 from safetensors.torch import save_file
 
+from lerobot.common.datasets.compute_stats import compute_stats
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
 from lerobot.common.datasets.push_dataset_to_hub._download_raw import download_raw
-from lerobot.common.datasets.push_dataset_to_hub.compute_stats import compute_stats
 from lerobot.common.datasets.utils import flatten_dict
 
 
@@ -73,10 +84,14 @@ def get_from_raw_to_lerobot_format_fn(raw_format):
         from lerobot.common.datasets.push_dataset_to_hub.umi_zarr_format import from_raw_to_lerobot_format
     elif raw_format == "aloha_hdf5":
         from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import from_raw_to_lerobot_format
+    elif raw_format == "aloha_dora":
+        from lerobot.common.datasets.push_dataset_to_hub.aloha_dora_format import from_raw_to_lerobot_format
     elif raw_format == "xarm_pkl":
         from lerobot.common.datasets.push_dataset_to_hub.xarm_pkl_format import from_raw_to_lerobot_format
     else:
-        raise ValueError(raw_format)
+        raise ValueError(
+            f"The selected {raw_format} can't be found. Did you add it to `lerobot/scripts/push_dataset_to_hub.py::get_from_raw_to_lerobot_format_fn`?"
+        )
 
     return from_raw_to_lerobot_format
 
@@ -212,8 +227,7 @@ def push_dataset_to_hub(
         test_hf_dataset = test_hf_dataset.with_format(None)
         test_hf_dataset.save_to_disk(str(tests_out_dir / "train"))
 
-        # copy meta data to tests directory
-        shutil.copytree(meta_data_dir, tests_meta_data_dir)
+        save_meta_data(info, stats, episode_data_index, tests_meta_data_dir)
 
         # copy videos of first episode to tests directory
         episode_index = 0
@@ -222,6 +236,10 @@ def push_dataset_to_hub(
             fname = f"{key}_episode_{episode_index:06d}.mp4"
             shutil.copy(videos_dir / fname, tests_videos_dir / fname)
 
+    if not save_to_disk and out_dir.exists():
+        # remove possible temporary files remaining in the output directory
+        shutil.rmtree(out_dir)
+
 
 def main():
     parser = argparse.ArgumentParser()
@@ -299,7 +317,7 @@ def main():
     parser.add_argument(
         "--num-workers",
         type=int,
-        default=16,
+        default=8,
         help="Number of processes of Dataloader for computing the dataset statistics.",
     )
     parser.add_argument(

lerobot/scripts/robot_controls/koch_configure.py

@@ -0,0 +1,211 @@
+"""
+LCR Auto Configure: This program is used to automatically configure the Low Cost Robot (LCR) for the user.
+
+The program will:
+1. Disable all torque motors of provided LCR.
+2. Ask the user to move the LCR to the position 1 (see CONFIGURING.md for more details).
+3. Record the position of the LCR.
+4. Ask the user to move the LCR to the position 2 (see CONFIGURING.md for more details).
+5. Record the position of the LCR.
+6. Ask the user to move back the LCR to the position 1.
+7. Record the position of the LCR.
+8. Calculate the offset of the LCR and save it to the configuration file.
+
+It will also enable all appropriate operating modes for the LCR according if the LCR is a puppet or a master.
+"""
+
+import argparse
+import time
+
+import numpy as np
+
+from lerobot.common.robot_devices.motors.dynamixel import DynamixelBus, OperatingMode, DriveMode, TorqueMode
+
+
+def pause():
+    """
+    Pause the program until the user presses the enter key.
+    """
+    input("Press Enter to continue...")
+
+
+def prepare_configuration(arm: DynamixelBus):
+    """
+    Prepare the configuration for the LCR.
+    :param arm: DynamixelBus
+    """
+
+    # To be configured, all servos must be in "torque disable" mode
+    arm.sync_write_torque_enable(TorqueMode.DISABLED.value)
+
+    # We need to work with 'extended position mode' (4) for all servos, because in joint mode (1) the servos can't
+    # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
+    # you could end up with a servo with a position 0 or 4095 at a crucial point See [
+    # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
+    arm.sync_write_operating_mode(OperatingMode.EXTENDED_POSITION.value, [1, 2, 3, 4, 5])
+
+    # Gripper is always 'position control current based' (5)
+    arm.write_operating_mode(OperatingMode.CURRENT_CONTROLLED_POSITION.value, 6)
+
+    # We need to reset the homing offset for all servos
+    arm.sync_write_homing_offset(0)
+
+    # We need to work with 'normal drive mode' (0) for all servos
+    arm.sync_write_drive_mode(DriveMode.NON_INVERTED.value)
+
+
+def invert_appropriate_positions(positions: np.array, inverted: list[bool]) -> np.array:
+    """
+    Invert the appropriate positions.
+    :param positions: numpy array of positions
+    :param inverted: list of booleans to determine if the position should be inverted
+    :return: numpy array of inverted positions
+    """
+    for i, invert in enumerate(inverted):
+        if not invert and positions[i] is not None:
+            positions[i] = -positions[i]
+
+    return positions
+
+
+def calculate_corrections(positions: np.array, inverted: list[bool]) -> np.array:
+    """
+    Calculate the corrections for the positions.
+    :param positions: numpy array of positions
+    :param inverted: list of booleans to determine if the position should be inverted
+    :return: numpy array of corrections
+    """
+
+    wanted = wanted_position_1()
+
+    correction = invert_appropriate_positions(positions, inverted)
+
+    for i in range(len(positions)):
+        if correction[i] is not None:
+            if inverted[i]:
+                correction[i] -= wanted[i]
+            else:
+                correction[i] += wanted[i]
+
+    return correction
+
+
+def calculate_nearest_rounded_positions(positions: np.array) -> np.array:
+    """
+    Calculate the nearest rounded positions.
+    :param positions: numpy array of positions
+    :return: numpy array of nearest rounded positions
+    """
+
+    return np.array(
+        [round(positions[i] / 1024) * 1024 if positions[i] is not None else None for i in range(len(positions))])
+
+
+def configure_homing(arm: DynamixelBus, inverted: list[bool]) -> np.array:
+    """
+    Configure the homing for the LCR.
+    :param arm: DynamixelBus
+    :param inverted: list of booleans to determine if the position should be inverted
+    """
+
+    # Reset homing offset for the servos
+    arm.sync_write_homing_offset(0)
+
+    # Get the present positions of the servos
+    present_positions = arm.sync_read_present_position_i32()
+
+    nearest_positions = calculate_nearest_rounded_positions(present_positions)
+
+    correction = calculate_corrections(nearest_positions, inverted)
+
+    # Write the homing offset for the servos
+    arm.sync_write_homing_offset(correction)
+
+
+def configure_drive_mode(arm: DynamixelBus):
+    """
+    Configure the drive mode for the LCR.
+    :param arm: DynamixelBus
+    :param homing: numpy array of homing
+    """
+
+    # Get current positions
+    present_positions = arm.sync_read_present_position_i32()
+
+    nearest_positions = calculate_nearest_rounded_positions(present_positions)
+
+    # construct 'inverted' list comparing nearest_positions and wanted_position_2
+    inverted = []
+
+    for i in range(len(nearest_positions)):
+        inverted.append(nearest_positions[i] != wanted_position_2()[i])
+
+    # Write the drive mode for the servos
+    arm.sync_write_drive_mode(
+        [DriveMode.INVERTED.value if i else DriveMode.NON_INVERTED.value for i in inverted])
+
+    return inverted
+
+
+def wanted_position_1() -> np.array:
+    """
+    The present position wanted in position 1 for the arm
+    """
+    return np.array([0, -1024, 1024, 0, 0, 0])
+
+
+def wanted_position_2() -> np.array:
+    """
+    The present position wanted in position 2 for the arm
+    """
+    return np.array([1024, 0, 0, 1024, 1024, -1024])
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="LCR Auto Configure: This program is used to automatically configure the Low Cost Robot (LCR) for "
+                    "the user.")
+
+    parser.add_argument("--port", type=str, required=True, help="The port of the LCR.")
+
+    args = parser.parse_args()
+
+    arm = DynamixelBus(
+        args.port, {
+            1: "x_series",
+            2: "x_series",
+            3: "x_series",
+            4: "x_series",
+            5: "x_series",
+            6: "x_series",
+        }
+    )
+
+    prepare_configuration(arm)
+
+    # Ask the user to move the LCR to the position 1
+    print("Please move the LCR to the position 1")
+    pause()
+
+    configure_homing(arm, [False, False, False, False, False, False])
+
+    # Ask the user to move the LCR to the position 2
+    print("Please move the LCR to the position 2")
+    pause()
+
+    inverted = configure_drive_mode(arm)
+
+    # Ask the user to move back the LCR to the position 1
+    print("Please move back the LCR to the position 1")
+    pause()
+
+    configure_homing(arm, inverted)
+
+    print("Configuration done!")
+    print("Make sure everything is working properly:")
+
+    while True:
+        positions = arm.sync_read_present_position_i32()
+        print(positions)
+
+        time.sleep(1)

lerobot/scripts/robot_controls/record_dataset.py

@@ -0,0 +1,20 @@
+
+
+import time
+from lerobot.common.robot_devices.robots.aloha import AlohaRobot
+import torch
+
+
+def record_dataset():
+    robot = AlohaRobot(use_cameras=True)
+    robot.init_teleop()
+
+    while True:
+        now = time.time()
+        observation, action = robot.teleop_step(record_data=True)
+
+        dt_s = (time.time() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+if __name__ == "__main__":
+    record_dataset()

lerobot/scripts/robot_controls/replay_dataset.py

@@ -0,0 +1,19 @@
+
+import time
+from lerobot.common.robot_devices.robots.aloha import AlohaRobot
+import torch
+
+
+def record_dataset():
+    robot = AlohaRobot(use_cameras=True)
+    robot.init_teleop()
+
+    while True:
+        now = time.time()
+        observation, action = robot.teleop_step(record_data=True)
+
+        dt_s = (time.time() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+if __name__ == "__main__":
+    record_dataset()

lerobot/scripts/robot_controls/run_policy.py

@@ -0,0 +1,47 @@
+
+
+import time
+from lerobot.common.robot_devices.robots.aloha import AlohaRobot
+import torch
+
+
+
+def teleoperate():
+    robot = AlohaRobot(use_cameras=False)
+    robot.init_teleop()
+
+    while True:
+        now = time.time()
+        robot.teleop_step(record_data=False)
+
+        dt_s = (time.time() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+
+def record_teleop_data():
+    robot = AlohaRobot(use_cameras=True)
+    robot.init_teleop()
+
+    while True:
+        now = time.time()
+        observation, action = robot.teleop_step(record_data=True)
+
+        dt_s = (time.time() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+
+
+def evaluate_policy(policy):
+    robot = AlohaRobot(use_cameras=True)
+    observation = robot.init_evaluate()
+
+    while True:
+        now = time.time()
+        with torch.inference_mode():
+            action = policy.select_action(observation)
+
+        observation, action = robot.step(action, record_data=False)
+
+        dt_s = (time.time() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+

lerobot/scripts/robot_controls/teleoperate.py

@@ -0,0 +1,20 @@
+
+
+import time
+from lerobot.common.robot_devices.robots.aloha import AlohaRobot
+
+
+def teleoperate():
+    robot = AlohaRobot(use_cameras=False)
+    robot.init_teleop()
+
+    while True:
+        now = time.time()
+        robot.teleop_step(record_data=False)
+
+        dt_s = (time.time() - now)
+        print(f"Latency (ms): {dt_s * 1000:.2f}\tFrequency: {1 / dt_s:.2f}")
+
+
+if __name__ == "__main__":
+    teleoperate()

lerobot/scripts/robot_controls/teleoperate_cv2.py

@@ -0,0 +1,320 @@
+import argparse
+from pathlib import Path
+import time
+import warnings
+
+import cv2
+import numpy as np
+from examples.real_robot_example.gym_real_world.robot import Robot
+import signal
+import sys
+# import pyrealsense2 as rs
+
+MAX_LEADER_GRIPPER_RAD = 0.7761942786701344
+MAX_LEADER_GRIPPER_POS = 2567
+MAX_FOLLOWER_GRIPPER_RAD = 1.6827769243105486
+MAX_FOLLOWER_GRIPPER_POS = 3100
+
+MIN_LEADER_GRIPPER_RAD = -0.12732040539450828
+MIN_LEADER_GRIPPER_POS = 1984
+MIN_FOLLOWER_GRIPPER_RAD = 0.6933593161243099
+MIN_FOLLOWER_GRIPPER_POS = 2512
+
+CAMERA_WIDTH = 640
+CAMERA_HEIGHT = 480
+
+def convert_gripper_range_from_leader_to_follower(leader_gripper_pos):
+    follower_gripper_pos = \
+        (leader_gripper_pos - MIN_LEADER_GRIPPER_POS) \
+        / (MAX_LEADER_GRIPPER_POS - MIN_LEADER_GRIPPER_POS) \
+        * (MAX_FOLLOWER_GRIPPER_POS - MIN_FOLLOWER_GRIPPER_POS) \
+        + MIN_FOLLOWER_GRIPPER_POS
+    return follower_gripper_pos
+
+# alexander koch
+# leader_port = "/dev/ttyACM1"
+# follower_port = "/dev/ttyACM0"
+
+def disable_torque():
+    leader_right_port = "/dev/ttyDXL_master_right"
+    leader_left_port = "/dev/ttyDXL_master_left"
+    follower_right_port = "/dev/ttyDXL_puppet_right"
+    follower_left_port = "/dev/ttyDXL_puppet_left"
+    # starts at 1
+    all_servo_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9]
+    leader_right = Robot(leader_right_port, servo_ids=all_servo_ids)
+    leader_left = Robot(leader_left_port, servo_ids=all_servo_ids)
+    follower_right = Robot(follower_right_port, servo_ids=all_servo_ids)
+    follower_left = Robot(follower_left_port, servo_ids=all_servo_ids)
+
+    leader_right._disable_torque()
+    leader_left._disable_torque()
+    follower_right._disable_torque()
+    follower_left._disable_torque()
+
+
+def teleoperate():
+    leader_right_port = "/dev/ttyDXL_master_right"
+    follower_right_port = "/dev/ttyDXL_puppet_right"
+    leader_left_port = "/dev/ttyDXL_master_left"
+    follower_left_port = "/dev/ttyDXL_puppet_left"
+    # starts at 1
+    all_servo_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9]
+    leader_right = Robot(leader_right_port, servo_ids=all_servo_ids)
+    leader_left = Robot(leader_left_port, servo_ids=all_servo_ids)
+    follower_right = Robot(follower_right_port, servo_ids=all_servo_ids)
+    follower_left = Robot(follower_left_port, servo_ids=all_servo_ids)
+
+    follower_right._enable_torque()
+    follower_left._enable_torque()
+
+    while True:
+        now = time.time()
+        # Prepare to assign the positions of the leader to the follower 
+        follower_right_pos = leader_right.read_position()
+        follower_left_pos = leader_left.read_position()
+
+        # Update the position of the follower gripper to account for the different minimum and maximum range
+        # position in range [0, 4096[ which corresponds to 4096 bins of 360 degrees
+        # for all our dynamixel servors
+        # gripper id=8 has a different range from leader to follower
+        follower_right_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_right_pos[-1])
+        follower_left_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_left_pos[-1])
+
+        # Assign
+        follower_right.set_goal_pos(follower_right_pos)
+        follower_left.set_goal_pos(follower_left_pos)
+
+        print(f"Time to send pos: {(time.time() - now) * 1000}")
+
+
+def capture_frame(camera: cv2.VideoCapture, output_color="rgb"):
+    # OpenCV acquires frames in BGR format (blue, green red)
+    ret, frame = camera.read()
+    if not ret:
+        raise OSError(f"Camera not found.")
+
+    if output_color == "rgb":
+        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+
+    return frame
+
+def find_camera_ids(out_dir="outputs/find_camera_ids/2024_06_19_cv2_1344"):
+    save_images = True
+    max_index_search_range = 60
+    num_warmup_frames = 4
+
+    # Works well
+    codec = "yuyv"
+    fps = 30
+    width = 640
+    height = 480
+
+    # # Works well
+    # codec = "yuyv"
+    # fps = 60
+    # width = 640
+    # height = 480
+
+    # # Works well
+    # codec = "yuyv"
+    # fps = 90
+    # width = 640
+    # height = 480
+
+    # # Works well
+    # codec = "yuyv"
+    # fps = 30
+    # width = 1280
+    # height = 720
+
+    # Doesn't work well (timeout)
+    # codec = "mjpg"
+    # fps = 30
+    # width = 1280
+    # height = 720
+
+    out_dir += f"_{width}x{height}_{fps}_{codec}"
+
+    camera_ids = []
+    for camera_idx in range(max_index_search_range):
+        camera = cv2.VideoCapture(camera_idx)
+        is_open = camera.isOpened()
+        camera.release()
+
+        if is_open:
+            print(f"Camera found at index {camera_idx}")
+            camera_ids.append(camera_idx)
+
+    if len(camera_ids) == 0:
+        raise OSError("No camera has been found")
+    
+    # Change camera settings
+    cameras = []
+    for camera_idx in camera_ids:
+        camera = cv2.VideoCapture(camera_idx)
+
+        camera.set(cv2.CAP_PROP_FPS, fps)
+        camera.set(cv2.CAP_PROP_FRAME_WIDTH, width)
+        camera.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
+        if codec == "mjpg":
+            camera.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*"MJPG"))
+
+        actual_fps = camera.get(cv2.CAP_PROP_FPS)
+        actual_width = camera.get(cv2.CAP_PROP_FRAME_WIDTH)
+        actual_height = camera.get(cv2.CAP_PROP_FRAME_HEIGHT)
+
+        if fps != actual_fps:
+            warnings.warn(f"{fps=} != {actual_fps=}", stacklevel=1)
+        if width != actual_width:
+            warnings.warn(f"{width=} != {actual_width=}", stacklevel=1)
+        if height != actual_height:
+            warnings.warn(f"{height=} != {actual_height=}", stacklevel=1)
+
+        cameras.append(camera)
+
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    print("Capturing a few frames to warmup")
+    for _ in range(num_warmup_frames):
+        for camera_idx, camera in zip(camera_ids, cameras):
+            print(f"Capturing camera {camera_idx}")
+            try:
+                frame = capture_frame(camera, output_color="bgr" if save_images else "rgb")
+                time.sleep(0.01)
+            except OSError as e:
+                print(e)
+        time.sleep(0.1)
+
+    print("Capturing frames")
+    try:
+        while True:
+            now = time.time()
+            for camera_idx, camera in zip(camera_ids, cameras):
+                try:
+                    frame = capture_frame(camera, output_color="bgr" if save_images else "rgb")
+                except OSError as e:
+                    print(e)    
+
+                def write_shape(frame):
+                    height, width = frame.shape[:2]
+                    text = f'Width: {width} Height: {height}'
+
+                    # Define the font, scale, color, and thickness
+                    font = cv2.FONT_HERSHEY_SIMPLEX
+                    font_scale = 1
+                    color = (255, 0, 0)  # Blue in BGR
+                    thickness = 2
+
+                    position = (10, height - 10)  # 10 pixels from the bottom-left corner
+                    cv2.putText(frame, text, position, font, font_scale, color, thickness)
+
+                if save_images:
+                    frame_path = out_dir / f"camera_{camera_idx:02}.png"
+                    print(f"Write to {frame_path}")
+                    write_shape(frame)
+                    cv2.imwrite(str(frame_path), frame)
+                    time.sleep(0.1)
+
+            dt_s = (time.time() - now)
+            dt_ms = dt_s * 1000
+            freq = 1 / dt_s
+            print(f"Latency (ms): {dt_ms:.2f}\tFrequency: {freq:.2f}")
+
+            if save_images:
+                break
+            if cv2.waitKey(1) & 0xFF == ord("q"):
+                break
+    finally:
+        # Stop streaming
+        for camera in cameras:
+            camera.release()
+
+    return camera_ids
+
+def record_data():
+    leader_right_port = "/dev/ttyDXL_master_right"
+    follower_right_port = "/dev/ttyDXL_puppet_right"
+    leader_left_port = "/dev/ttyDXL_master_left"
+    follower_left_port = "/dev/ttyDXL_puppet_left"
+    # starts at 1
+    all_servo_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9]
+    leader_right = Robot(leader_right_port, servo_ids=all_servo_ids)
+    leader_left = Robot(leader_left_port, servo_ids=all_servo_ids)
+    follower_right = Robot(follower_right_port, servo_ids=all_servo_ids)
+    follower_left = Robot(follower_left_port, servo_ids=all_servo_ids)
+
+    follower_right._enable_torque()
+    follower_left._enable_torque()
+
+    # To get the camera_ids, run: `find_camera_ids()`
+    camera_high = cv2.VideoCapture(10)
+    camera_low = cv2.VideoCapture(22)
+    camera_right_wrist = cv2.VideoCapture(16)
+    camera_left_wrist = cv2.VideoCapture(4)
+
+    if not camera_high.isOpened():
+        raise OSError("Camera high port can't be accessed.")
+    if not camera_low.isOpened():
+        raise OSError("Camera low port can't be accessed.")
+    if not camera_right_wrist.isOpened():
+        raise OSError("Camera right_wrist port can't be accessed.")
+    if not camera_left_wrist.isOpened():
+        raise OSError("Camera left_wrist port can't be accessed.")
+
+    while True:
+        now = time.time()
+
+        frame_high = capture_frame(camera_high)
+        frame_low = capture_frame(camera_low)
+        frame_right_wrist = capture_frame(camera_right_wrist)
+        frame_left_wrist = capture_frame(camera_left_wrist)
+
+        # cv2.imshow("high", frame_high)
+        # cv2.imshow("low", frame_low)
+        # cv2.imshow("right_wrist", frame_right_wrist)
+        # cv2.imshow("left_wrist", frame_left_wrist)
+
+        # Prepare to assign the positions of the leader to the follower 
+        follower_right_pos = leader_right.read_position()
+        follower_left_pos = leader_left.read_position()
+
+        # Update the position of the follower gripper to account for the different minimum and maximum range
+        # position in range [0, 4096[ which corresponds to 4096 bins of 360 degrees
+        # for all our dynamixel servors
+        # gripper id=8 has a different range from leader to follower
+        follower_right_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_right_pos[-1])
+        follower_left_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_left_pos[-1])
+
+        # Assign
+        follower_right.set_goal_pos(follower_right_pos)
+        follower_left.set_goal_pos(follower_left_pos)
+
+        dt_s = (time.time() - now)
+        dt_ms = dt_s * 1000
+        freq = 1 / dt_s
+        print(f"Latency (ms): {dt_ms:.2f}\tFrequency: {freq:.2f}")
+
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
+
+    camera_high.release()
+    camera_low.release()
+    camera_right_wrist.release()
+    camera_left_wrist.release()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--mode", type=str, choices=["teleoperate", "disable_torque", "record_data", "find_camera_ids"], default="teleoperate")
+    args = parser.parse_args()
+
+    if args.mode == "teleoperate":
+        teleoperate()
+    elif args.mode == "disable_torque":
+        disable_torque()
+    elif args.mode == "record_data":
+        record_data()
+    elif args.mode == "find_camera_ids":
+        find_camera_ids()

lerobot/scripts/robot_controls/teleoperate_intelrealsense.py

@@ -0,0 +1,338 @@
+import argparse
+from pathlib import Path
+import time
+import traceback
+
+import cv2
+import numpy as np
+from examples.real_robot_example.gym_real_world.robot import Robot
+import signal
+import sys
+import pyrealsense2 as rs
+
+MAX_LEADER_GRIPPER_RAD = 0.7761942786701344
+MAX_LEADER_GRIPPER_POS = 2567
+MAX_FOLLOWER_GRIPPER_RAD = 1.6827769243105486
+MAX_FOLLOWER_GRIPPER_POS = 3100
+
+MIN_LEADER_GRIPPER_RAD = -0.12732040539450828
+MIN_LEADER_GRIPPER_POS = 1984
+MIN_FOLLOWER_GRIPPER_RAD = 0.6933593161243099
+MIN_FOLLOWER_GRIPPER_POS = 2512
+
+CAMERA_WIDTH = 640
+CAMERA_HEIGHT = 480
+
+def convert_gripper_range_from_leader_to_follower(leader_gripper_pos):
+    follower_gripper_pos = \
+        (leader_gripper_pos - MIN_LEADER_GRIPPER_POS) \
+        / (MAX_LEADER_GRIPPER_POS - MIN_LEADER_GRIPPER_POS) \
+        * (MAX_FOLLOWER_GRIPPER_POS - MIN_FOLLOWER_GRIPPER_POS) \
+        + MIN_FOLLOWER_GRIPPER_POS
+    return follower_gripper_pos
+
+# alexander koch
+# leader_port = "/dev/ttyACM1"
+# follower_port = "/dev/ttyACM0"
+
+def disable_torque():
+    leader_right_port = "/dev/ttyDXL_master_right"
+    leader_left_port = "/dev/ttyDXL_master_left"
+    follower_right_port = "/dev/ttyDXL_puppet_right"
+    follower_left_port = "/dev/ttyDXL_puppet_left"
+    # starts at 1
+    all_servo_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9]
+    leader_right = Robot(leader_right_port, servo_ids=all_servo_ids)
+    leader_left = Robot(leader_left_port, servo_ids=all_servo_ids)
+    follower_right = Robot(follower_right_port, servo_ids=all_servo_ids)
+    follower_left = Robot(follower_left_port, servo_ids=all_servo_ids)
+
+    leader_right._disable_torque()
+    leader_left._disable_torque()
+    follower_right._disable_torque()
+    follower_left._disable_torque()
+
+
+def teleoperate():
+    leader_right_port = "/dev/ttyDXL_master_right"
+    follower_right_port = "/dev/ttyDXL_puppet_right"
+    leader_left_port = "/dev/ttyDXL_master_left"
+    follower_left_port = "/dev/ttyDXL_puppet_left"
+    # starts at 1
+    all_servo_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9]
+    leader_right = Robot(leader_right_port, servo_ids=all_servo_ids)
+    leader_left = Robot(leader_left_port, servo_ids=all_servo_ids)
+    follower_right = Robot(follower_right_port, servo_ids=all_servo_ids)
+    follower_left = Robot(follower_left_port, servo_ids=all_servo_ids)
+
+    follower_right._enable_torque()
+    follower_left._enable_torque()
+
+    while True:
+        now = time.time()
+        # Prepare to assign the positions of the leader to the follower 
+        follower_right_pos = leader_right.read_position()
+        follower_left_pos = leader_left.read_position()
+
+        # Update the position of the follower gripper to account for the different minimum and maximum range
+        # position in range [0, 4096[ which corresponds to 4096 bins of 360 degrees
+        # for all our dynamixel servors
+        # gripper id=8 has a different range from leader to follower
+        follower_right_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_right_pos[-1])
+        follower_left_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_left_pos[-1])
+
+        # Assign
+        follower_right.set_goal_pos(follower_right_pos)
+        follower_left.set_goal_pos(follower_left_pos)
+
+        print(f"Time to send pos: {(time.time() - now) * 1000}")
+
+
+def capture_frame(camera: cv2.VideoCapture, width=CAMERA_WIDTH, height=CAMERA_HEIGHT, output_color="rgb"):
+    # OpenCV acquires frames in BGR format (blue, green red)
+    ret, frame = camera.read()
+    if not ret:
+        raise OSError(f"Camera not found.")
+
+    if output_color == "rgb":
+        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+    # # Define your crop coordinates (top left corner and bottom right corner)
+    # x1, y1 = 400, 0  # Example starting coordinates (top left of the crop rectangle)
+    # x2, y2 = 1600, 900  # Example ending coordinates (bottom right of the crop rectangle)
+    # # Crop the image
+    # image = image[y1:y2, x1:x2]
+    # Resize the image
+    frame = cv2.resize(frame, (width, height), interpolation=cv2.INTER_AREA)
+    return frame
+
+def write_shape(frame):
+    height, width = frame.shape[:2]
+    text = f'Width: {width} Height: {height}'
+
+    # Define the font, scale, color, and thickness
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    font_scale = 1
+    color = (255, 0, 0)  # Blue in BGR
+    thickness = 2
+
+    position = (10, height - 10)  # 10 pixels from the bottom-left corner
+    cv2.putText(frame, text, position, font, font_scale, color, thickness)
+
+def find_camera_ids(out_dir="outputs/find_camera_ids/2024_06_19_1039"):
+    """
+        Install: https://github.com/IntelRealSense/librealsense/blob/master/doc/distribution_linux.md
+
+        List cameras and make sure the firmware is up-to-date: https://dev.intelrealsense.com/docs/firmware-releases-d400
+        ```bash
+        rs-fw-update -l
+
+        > Connected devices:
+        > 1) [USB] Intel RealSense D405 s/n 128422270109, update serial number: 133323070634, firmware version: 5.16.0.1
+        > 2) [USB] Intel RealSense D405 s/n 128422271609, update serial number: 130523070758, firmware version: 5.16.0.1
+        > 3) [USB] Intel RealSense D405 s/n 128422271614, update serial number: 133323070576, firmware version: 5.16.0.1
+        > 4) [USB] Intel RealSense D405 s/n 128422271393, update serial number: 133323070271, firmware version: 5.16.0.1
+        ```
+    """
+    save_images = False
+    # enable once, if you reach "Frame didn't arrive" exception
+    force_hardware_reset = False
+
+    # Works well!
+    # use_depth = False
+    # fps = 90
+    # width = 640
+    # height = 480
+
+    # # Works well!
+    # use_depth = True
+    # fps = 90
+    # width = 640
+    # height = 480
+
+    # # Doesn't work well, latency varies too much
+    # use_depth = True
+    # fps = 30
+    # width = 1280
+    # height = 720
+
+    # Works well
+    use_depth = False
+    fps = 30
+    width = 1280
+    height = 720
+
+    out_dir += f"_{width}x{height}_{fps}_depth_{use_depth}"
+
+    ctx = rs.context()
+
+    serials = []
+    cameras = []
+    for device in ctx.query_devices():
+        print(device)
+        if force_hardware_reset:
+            device.hardware_reset()
+
+        SERIAL_NUMBER_INDEX = 1
+        serial_number = device.get_info(rs.camera_info(SERIAL_NUMBER_INDEX))
+
+        config = rs.config()
+        config.enable_device(serial_number)
+
+        config.enable_stream(rs.stream.color, width, height, rs.format.bgr8, fps)
+
+        if use_depth:
+            config.enable_stream(rs.stream.depth, width, height, rs.format.z16, fps)
+
+        pipeline = rs.pipeline()
+        pipeline.start(config)
+
+        serials.append(serial_number)
+        cameras.append(pipeline)
+
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    try:
+        while True:
+            now = time.time()
+            for serial, camera in zip(serials, cameras):
+                # Wait for a coherent pair of frames: depth and color
+                try:
+                    frames = camera.wait_for_frames()
+                except RuntimeError as e:
+                    if "Frame didn't arrive" in str(e):
+                        print(f"{e}: Trying hardware_reset. If it still doesn't work, try `force_hardware_reset=True`.")
+                        device.hardware_reset()
+                    traceback.print_exc()
+                    continue
+
+                # acquire color image
+                color_frame = frames.get_color_frame()
+                if not color_frame:
+                    print("Empty color frame")
+                    continue
+                # to numpy
+                image = np.asanyarray(color_frame.get_data())
+
+                if save_images:
+                    image_path = out_dir / f"camera_{serial:02}.png"
+                    print(f"Write to {image_path}")
+                    write_shape(image)
+                    cv2.imwrite(str(image_path), image)
+
+                if use_depth:
+                    # acquire depth image
+                    depth_frame = frames.get_depth_frame()
+                    if not depth_frame:
+                        print("Empty depth frame")
+                        continue
+                    # to numpy
+                    depth = np.asanyarray(depth_frame.get_data())
+                    
+                    if save_images:
+                        # Apply colormap on depth image (image must be converted to 8-bit per pixel first)
+                        depth_image = cv2.applyColorMap(cv2.convertScaleAbs(depth, alpha=0.03), cv2.COLORMAP_JET)
+                        depth_image_path = out_dir / f"camera_{serial:02}_depth.png"
+                        print(f"Write to {depth_image_path}")
+                        write_shape(depth_image)
+                        cv2.imwrite(str(depth_image_path), depth_image)
+
+            dt_s = (time.time() - now)
+            dt_ms = dt_s * 1000
+            freq = 1 / dt_s
+            print(f"Latency (ms): {dt_ms:.2f}\tFrequency: {freq:.2f}")
+
+            if save_images:
+                break
+            if cv2.waitKey(1) & 0xFF == ord("q"):
+                break
+    finally:
+        # Stop streaming
+        for camera in cameras:
+            camera.stop()
+
+    return serials
+
+
+def record_data():
+    leader_right_port = "/dev/ttyDXL_master_right"
+    follower_right_port = "/dev/ttyDXL_puppet_right"
+    leader_left_port = "/dev/ttyDXL_master_left"
+    follower_left_port = "/dev/ttyDXL_puppet_left"
+    # starts at 1
+    all_servo_ids = [1, 2, 3, 4, 5, 6, 7, 8, 9]
+    leader_right = Robot(leader_right_port, servo_ids=all_servo_ids)
+    leader_left = Robot(leader_left_port, servo_ids=all_servo_ids)
+    follower_right = Robot(follower_right_port, servo_ids=all_servo_ids)
+    follower_left = Robot(follower_left_port, servo_ids=all_servo_ids)
+
+    follower_right._enable_torque()
+    follower_left._enable_torque()
+
+    # To get the camera_ids, run: `find_camera_ids()`
+    camera_high = cv2.VideoCapture(10)
+    camera_low = cv2.VideoCapture(22)
+    camera_right_wrist = cv2.VideoCapture(16)
+    camera_left_wrist = cv2.VideoCapture(4)
+
+    if not camera_high.isOpened():
+        raise OSError("Camera high port can't be accessed.")
+    if not camera_low.isOpened():
+        raise OSError("Camera low port can't be accessed.")
+    if not camera_right_wrist.isOpened():
+        raise OSError("Camera right_wrist port can't be accessed.")
+    if not camera_left_wrist.isOpened():
+        raise OSError("Camera left_wrist port can't be accessed.")
+
+    while True:
+        now = time.time()
+
+        frame_high = capture_frame(camera_high)
+        frame_low = capture_frame(camera_low)
+        frame_right_wrist = capture_frame(camera_right_wrist)
+        frame_left_wrist = capture_frame(camera_left_wrist)
+
+        # cv2.imshow("high", frame_high)
+        # cv2.imshow("low", frame_low)
+        # cv2.imshow("right_wrist", frame_right_wrist)
+        # cv2.imshow("left_wrist", frame_left_wrist)
+
+        # Prepare to assign the positions of the leader to the follower 
+        follower_right_pos = leader_right.read_position()
+        follower_left_pos = leader_left.read_position()
+
+        # Update the position of the follower gripper to account for the different minimum and maximum range
+        # position in range [0, 4096[ which corresponds to 4096 bins of 360 degrees
+        # for all our dynamixel servors
+        # gripper id=8 has a different range from leader to follower
+        follower_right_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_right_pos[-1])
+        follower_left_pos[-1] = convert_gripper_range_from_leader_to_follower(follower_left_pos[-1])
+
+        # Assign
+        follower_right.set_goal_pos(follower_right_pos)
+        follower_left.set_goal_pos(follower_left_pos)
+
+        print(f"Time to send pos: {(time.time() - now) * 1000}")
+
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
+
+    camera_high.release()
+    camera_low.release()
+    camera_right_wrist.release()
+    camera_left_wrist.release()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--mode", type=str, choices=["teleoperate", "disable_torque", "record_data", "find_camera_ids"], default="teleoperate")
+    args = parser.parse_args()
+
+    if args.mode == "teleoperate":
+        teleoperate()
+    elif args.mode == "disable_torque":
+        disable_torque()
+    elif args.mode == "record_data":
+        record_data()
+    elif args.mode == "find_camera_ids":
+        find_camera_ids()

lerobot/scripts/test.py

@@ -0,0 +1,10 @@
+import pyrealsense2 as rs
+pipe = rs.pipeline()
+profile = pipe.start()
+try:
+    for i in range(0, 100):
+        frames = pipe.wait_for_frames()
+        for f in frames:
+            print(f.profile)
+finally:
+    pipe.stop()

lerobot/scripts/train.py

@@ -1,23 +1,44 @@
+#!/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 logging
 import time
-from copy import deepcopy
+from contextlib import nullcontext
 from pathlib import Path
+from pprint import pformat
 
-import datasets
 import hydra
 import torch
-from datasets import concatenate_datasets
-from datasets.utils import disable_progress_bars, enable_progress_bars
+from deepdiff import DeepDiff
+from omegaconf import DictConfig, OmegaConf
+from termcolor import colored
+from torch.cuda.amp import GradScaler
 
-from lerobot.common.datasets.factory import make_dataset
+from lerobot.common.datasets.factory import make_dataset, resolve_delta_timestamps
+from lerobot.common.datasets.lerobot_dataset import MultiLeRobotDataset
+from lerobot.common.datasets.sampler import EpisodeAwareSampler
 from lerobot.common.datasets.utils import cycle
 from lerobot.common.envs.factory import make_env
 from lerobot.common.logger import Logger, log_output_dir
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.policies.policy_protocol import PolicyWithUpdate
+from lerobot.common.policies.utils import get_device_from_parameters
 from lerobot.common.utils.utils import (
     format_big_number,
     get_safe_torch_device,
+    init_hydra_config,
     init_logging,
     set_global_seed,
 )
@@ -53,7 +74,6 @@ def make_optimizer_and_scheduler(cfg, policy):
             cfg.training.adam_eps,
             cfg.training.adam_weight_decay,
         )
-        assert cfg.training.online_steps == 0, "Diffusion Policy does not handle online training."
         from diffusers.optimization import get_scheduler
 
         lr_scheduler = get_scheduler(
@@ -71,20 +91,40 @@ def make_optimizer_and_scheduler(cfg, policy):
     return optimizer, lr_scheduler
 
 
-def update_policy(policy, batch, optimizer, grad_clip_norm, lr_scheduler=None):
-    start_time = time.time()
+def update_policy(
+    policy,
+    batch,
+    optimizer,
+    grad_clip_norm,
+    grad_scaler: GradScaler,
+    lr_scheduler=None,
+    use_amp: bool = False,
+):
+    """Returns a dictionary of items for logging."""
+    start_time = time.perf_counter()
+    device = get_device_from_parameters(policy)
     policy.train()
-    output_dict = policy.forward(batch)
-    # TODO(rcadene): policy.unnormalize_outputs(out_dict)
-    loss = output_dict["loss"]
-    loss.backward()
+    with torch.autocast(device_type=device.type) if use_amp else nullcontext():
+        output_dict = policy.forward(batch)
+        # TODO(rcadene): policy.unnormalize_outputs(out_dict)
+        loss = output_dict["loss"]
+    grad_scaler.scale(loss).backward()
+
+    # Unscale the graident of the optimzer's assigned params in-place **prior to gradient clipping**.
+    grad_scaler.unscale_(optimizer)
+
     grad_norm = torch.nn.utils.clip_grad_norm_(
         policy.parameters(),
         grad_clip_norm,
         error_if_nonfinite=False,
     )
 
-    optimizer.step()
+    # Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
+    # although it still skips optimizer.step() if the gradients contain infs or NaNs.
+    grad_scaler.step(optimizer)
+    # Updates the scale for next iteration.
+    grad_scaler.update()
+
     optimizer.zero_grad()
 
     if lr_scheduler is not None:
@@ -98,31 +138,14 @@ def update_policy(policy, batch, optimizer, grad_clip_norm, lr_scheduler=None):
         "loss": loss.item(),
         "grad_norm": float(grad_norm),
         "lr": optimizer.param_groups[0]["lr"],
-        "update_s": time.time() - start_time,
+        "update_s": time.perf_counter() - start_time,
+        **{k: v for k, v in output_dict.items() if k != "loss"},
     }
 
     return info
 
 
-@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
-def train_cli(cfg: dict):
-    train(
-        cfg,
-        out_dir=hydra.core.hydra_config.HydraConfig.get().run.dir,
-        job_name=hydra.core.hydra_config.HydraConfig.get().job.name,
-    )
-
-
-def train_notebook(out_dir=None, job_name=None, config_name="default", config_path="../configs"):
-    from hydra import compose, initialize
-
-    hydra.core.global_hydra.GlobalHydra.instance().clear()
-    initialize(config_path=config_path)
-    cfg = compose(config_name=config_name)
-    train(cfg, out_dir=out_dir, job_name=job_name)
-
-
-def log_train_info(logger, info, step, cfg, dataset, is_offline):
+def log_train_info(logger: Logger, info, step, cfg, dataset, is_offline):
     loss = info["loss"]
     grad_norm = info["grad_norm"]
     lr = info["lr"]
@@ -193,104 +216,7 @@ def log_eval_info(logger, info, step, cfg, dataset, is_offline):
     logger.log_dict(info, step, mode="eval")
 
 
-def calculate_online_sample_weight(n_off: int, n_on: int, pc_on: float):
-    """
-    Calculate the sampling weight to be assigned to samples so that a specified percentage of the batch comes from online dataset (on average).
-
-    Parameters:
-    - n_off (int): Number of offline samples, each with a sampling weight of 1.
-    - n_on (int): Number of online samples.
-    - pc_on (float): Desired percentage of online samples in decimal form (e.g., 50% as 0.5).
-
-    The total weight of offline samples is n_off * 1.0.
-    The total weight of offline samples is n_on * w.
-    The total combined weight of all samples is n_off + n_on * w.
-    The fraction of the weight that is online is n_on * w / (n_off + n_on * w).
-    We want this fraction to equal pc_on, so we set up the equation n_on * w / (n_off + n_on * w) = pc_on.
-    The solution is w = - (n_off * pc_on) / (n_on * (pc_on - 1))
-    """
-    assert 0.0 <= pc_on <= 1.0
-    return -(n_off * pc_on) / (n_on * (pc_on - 1))
-
-
-def add_episodes_inplace(
-    online_dataset: torch.utils.data.Dataset,
-    concat_dataset: torch.utils.data.ConcatDataset,
-    sampler: torch.utils.data.WeightedRandomSampler,
-    hf_dataset: datasets.Dataset,
-    episode_data_index: dict[str, torch.Tensor],
-    pc_online_samples: float,
-):
-    """
-    Modifies the online_dataset, concat_dataset, and sampler in place by integrating
-    new episodes from hf_dataset into the online_dataset, updating the concatenated
-    dataset's structure and adjusting the sampling strategy based on the specified
-    percentage of online samples.
-
-    Parameters:
-    - online_dataset (torch.utils.data.Dataset): The existing online dataset to be updated.
-    - concat_dataset (torch.utils.data.ConcatDataset): The concatenated dataset that combines
-      offline and online datasets, used for sampling purposes.
-    - sampler (torch.utils.data.WeightedRandomSampler): A sampler that will be updated to
-      reflect changes in the dataset sizes and specified sampling weights.
-    - hf_dataset (datasets.Dataset): A Hugging Face dataset containing the new episodes to be added.
-    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
-      They indicate the start index and end index of each episode in the dataset.
-    - pc_online_samples (float): The target percentage of samples that should come from
-      the online dataset during sampling operations.
-
-    Raises:
-    - AssertionError: If the first episode_id or index in hf_dataset is not 0
-    """
-    first_episode_idx = hf_dataset.select_columns("episode_index")[0]["episode_index"].item()
-    last_episode_idx = hf_dataset.select_columns("episode_index")[-1]["episode_index"].item()
-    first_index = hf_dataset.select_columns("index")[0]["index"].item()
-    last_index = hf_dataset.select_columns("index")[-1]["index"].item()
-    # sanity check
-    assert first_episode_idx == 0, f"{first_episode_idx=} is not 0"
-    assert first_index == 0, f"{first_index=} is not 0"
-    assert first_index == episode_data_index["from"][first_episode_idx].item()
-    assert last_index == episode_data_index["to"][last_episode_idx].item() - 1
-
-    if len(online_dataset) == 0:
-        # initialize online dataset
-        online_dataset.hf_dataset = hf_dataset
-        online_dataset.episode_data_index = episode_data_index
-    else:
-        # get the starting indices of the new episodes and frames to be added
-        start_episode_idx = last_episode_idx + 1
-        start_index = last_index + 1
-
-        def shift_indices(episode_index, index):
-            # note: we dont shift "frame_index" since it represents the index of the frame in the episode it belongs to
-            example = {"episode_index": episode_index + start_episode_idx, "index": index + start_index}
-            return example
-
-        disable_progress_bars()  # map has a tqdm progress bar
-        hf_dataset = hf_dataset.map(shift_indices, input_columns=["episode_index", "index"])
-        enable_progress_bars()
-
-        episode_data_index["from"] += start_index
-        episode_data_index["to"] += start_index
-
-        # extend online dataset
-        online_dataset.hf_dataset = concatenate_datasets([online_dataset.hf_dataset, hf_dataset])
-
-    # update the concatenated dataset length used during sampling
-    concat_dataset.cumulative_sizes = concat_dataset.cumsum(concat_dataset.datasets)
-
-    # update the sampling weights for each frame so that online frames get sampled a certain percentage of times
-    len_online = len(online_dataset)
-    len_offline = len(concat_dataset) - len_online
-    weight_offline = 1.0
-    weight_online = calculate_online_sample_weight(len_offline, len_online, pc_online_samples)
-    sampler.weights = torch.tensor([weight_offline] * len_offline + [weight_online] * len(online_dataset))
-
-    # update the total number of samples used during sampling
-    sampler.num_samples = len(concat_dataset)
-
-
-def train(cfg: dict, out_dir=None, job_name=None):
+def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = None):
     if out_dir is None:
         raise NotImplementedError()
     if job_name is None:
@@ -298,35 +224,96 @@ def train(cfg: dict, out_dir=None, job_name=None):
 
     init_logging()
 
-    if cfg.training.online_steps > 0 and cfg.eval.batch_size > 1:
-        logging.warning("eval.batch_size > 1 not supported for online training steps")
+    # If we are resuming a run, we need to check that a checkpoint exists in the log directory, and we need
+    # to check for any differences between the provided config and the checkpoint's config.
+    if cfg.resume:
+        if not Logger.get_last_checkpoint_dir(out_dir).exists():
+            raise RuntimeError(
+                "You have set resume=True, but there is no model checkpoint in "
+                f"{Logger.get_last_checkpoint_dir(out_dir)}"
+            )
+        checkpoint_cfg_path = str(Logger.get_last_pretrained_model_dir(out_dir) / "config.yaml")
+        logging.info(
+            colored(
+                "You have set resume=True, indicating that you wish to resume a run",
+                color="yellow",
+                attrs=["bold"],
+            )
+        )
+        # Get the configuration file from the last checkpoint.
+        checkpoint_cfg = init_hydra_config(checkpoint_cfg_path)
+        # Check for differences between the checkpoint configuration and provided configuration.
+        # Hack to resolve the delta_timestamps ahead of time in order to properly diff.
+        resolve_delta_timestamps(cfg)
+        diff = DeepDiff(OmegaConf.to_container(checkpoint_cfg), OmegaConf.to_container(cfg))
+        # Ignore the `resume` and parameters.
+        if "values_changed" in diff and "root['resume']" in diff["values_changed"]:
+            del diff["values_changed"]["root['resume']"]
+        # Log a warning about differences between the checkpoint configuration and the provided
+        # configuration.
+        if len(diff) > 0:
+            logging.warning(
+                "At least one difference was detected between the checkpoint configuration and "
+                f"the provided configuration: \n{pformat(diff)}\nNote that the checkpoint configuration "
+                "takes precedence.",
+            )
+        # Use the checkpoint config instead of the provided config (but keep `resume` parameter).
+        cfg = checkpoint_cfg
+        cfg.resume = True
+    elif Logger.get_last_checkpoint_dir(out_dir).exists():
+        raise RuntimeError(
+            f"The configured output directory {Logger.get_last_checkpoint_dir(out_dir)} already exists."
+        )
+
+    # log metrics to terminal and wandb
+    logger = Logger(cfg, out_dir, wandb_job_name=job_name)
+
+    if cfg.training.online_steps > 0:
+        raise NotImplementedError("Online training is not implemented yet.")
+
+    set_global_seed(cfg.seed)
 
     # Check device is available
-    get_safe_torch_device(cfg.device, log=True)
+    device = get_safe_torch_device(cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
-    set_global_seed(cfg.seed)
 
     logging.info("make_dataset")
     offline_dataset = make_dataset(cfg)
+    if isinstance(offline_dataset, MultiLeRobotDataset):
+        logging.info(
+            "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
+            f"{pformat(offline_dataset.repo_id_to_index , indent=2)}"
+        )
 
-    logging.info("make_env")
-    eval_env = make_env(cfg)
+    # Create environment used for evaluating checkpoints during training on simulation data.
+    # On real-world data, no need to create an environment as evaluations are done outside train.py,
+    # using the eval.py instead, with gym_dora environment and dora-rs.
+    if cfg.training.eval_freq > 0:
+        logging.info("make_env")
+        eval_env = make_env(cfg)
 
     logging.info("make_policy")
-    policy = make_policy(hydra_cfg=cfg, dataset_stats=offline_dataset.stats)
+    policy = make_policy(
+        hydra_cfg=cfg,
+        dataset_stats=offline_dataset.stats if not cfg.resume else None,
+        pretrained_policy_name_or_path=str(logger.last_pretrained_model_dir) if cfg.resume else None,
+    )
 
     # Create optimizer and scheduler
     # Temporary hack to move optimizer out of policy
     optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
+    grad_scaler = GradScaler(enabled=cfg.use_amp)
+
+    step = 0  # number of policy updates (forward + backward + optim)
+
+    if cfg.resume:
+        step = logger.load_last_training_state(optimizer, lr_scheduler)
 
     num_learnable_params = sum(p.numel() for p in policy.parameters() if p.requires_grad)
     num_total_params = sum(p.numel() for p in policy.parameters())
 
-    # log metrics to terminal and wandb
-    logger = Logger(out_dir, job_name, cfg)
-
     log_output_dir(out_dir)
     logging.info(f"{cfg.env.task=}")
     logging.info(f"{cfg.training.offline_steps=} ({format_big_number(cfg.training.offline_steps)})")
@@ -338,27 +325,31 @@ def train(cfg: dict, out_dir=None, job_name=None):
 
     # Note: this helper will be used in offline and online training loops.
     def evaluate_and_checkpoint_if_needed(step):
-        if step % cfg.training.eval_freq == 0:
+        if cfg.training.eval_freq > 0 and step % cfg.training.eval_freq == 0:
             logging.info(f"Eval policy at step {step}")
-            eval_info = eval_policy(
-                eval_env,
-                policy,
-                cfg.eval.n_episodes,
-                video_dir=Path(out_dir) / "eval",
-                max_episodes_rendered=4,
-                start_seed=cfg.seed,
-            )
-            log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline)
+            with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.use_amp else nullcontext():
+                eval_info = eval_policy(
+                    eval_env,
+                    policy,
+                    cfg.eval.n_episodes,
+                    video_dir=Path(out_dir) / "eval",
+                    max_episodes_rendered=4,
+                    start_seed=cfg.seed,
+                )
+            log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline=True)
             if cfg.wandb.enable:
                 logger.log_video(eval_info["video_paths"][0], step, mode="eval")
             logging.info("Resume training")
 
-        if cfg.training.save_model and step % cfg.training.save_freq == 0:
+        if cfg.training.save_checkpoint and step % cfg.training.save_freq == 0:
             logging.info(f"Checkpoint policy after step {step}")
             # Note: Save with step as the identifier, and format it to have at least 6 digits but more if
             # needed (choose 6 as a minimum for consistency without being overkill).
-            logger.save_model(
+            logger.save_checkpont(
+                step,
                 policy,
+                optimizer,
+                lr_scheduler,
                 identifier=str(step).zfill(
                     max(6, len(str(cfg.training.offline_steps + cfg.training.online_steps)))
                 ),
@@ -366,32 +357,48 @@ def train(cfg: dict, out_dir=None, job_name=None):
             logging.info("Resume training")
 
     # create dataloader for offline training
+    if cfg.training.get("drop_n_last_frames"):
+        shuffle = False
+        sampler = EpisodeAwareSampler(
+            offline_dataset.episode_data_index,
+            drop_n_last_frames=cfg.training.drop_n_last_frames,
+            shuffle=True,
+        )
+    else:
+        shuffle = True
+        sampler = None
     dataloader = torch.utils.data.DataLoader(
         offline_dataset,
-        num_workers=4,
+        num_workers=cfg.training.num_workers,
         batch_size=cfg.training.batch_size,
-        shuffle=True,
-        pin_memory=cfg.device != "cpu",
+        shuffle=shuffle,
+        sampler=sampler,
+        pin_memory=device.type != "cpu",
         drop_last=False,
     )
     dl_iter = cycle(dataloader)
 
     policy.train()
-    step = 0  # number of policy update (forward + backward + optim)
-    is_offline = True
-    for offline_step in range(cfg.training.offline_steps):
-        if offline_step == 0:
+    for _ in range(step, cfg.training.offline_steps):
+        if step == 0:
             logging.info("Start offline training on a fixed dataset")
         batch = next(dl_iter)
 
         for key in batch:
-            batch[key] = batch[key].to(cfg.device, non_blocking=True)
+            batch[key] = batch[key].to(device, non_blocking=True)
 
-        train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
+        train_info = update_policy(
+            policy,
+            batch,
+            optimizer,
+            cfg.training.grad_clip_norm,
+            grad_scaler=grad_scaler,
+            lr_scheduler=lr_scheduler,
+            use_amp=cfg.use_amp,
+        )
 
-        # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
         if step % cfg.training.log_freq == 0:
-            log_train_info(logger, train_info, step, cfg, offline_dataset, is_offline)
+            log_train_info(logger, train_info, step, cfg, offline_dataset, is_offline=True)
 
         # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
         # so we pass in step + 1.
@@ -399,79 +406,28 @@ def train(cfg: dict, out_dir=None, job_name=None):
 
         step += 1
 
-    # create an env dedicated to online episodes collection from policy rollout
-    online_training_env = make_env(cfg, n_envs=1)
-
-    # create an empty online dataset similar to offline dataset
-    online_dataset = deepcopy(offline_dataset)
-    online_dataset.hf_dataset = {}
-    online_dataset.episode_data_index = {}
-
-    # create dataloader for online training
-    concat_dataset = torch.utils.data.ConcatDataset([offline_dataset, online_dataset])
-    weights = [1.0] * len(concat_dataset)
-    sampler = torch.utils.data.WeightedRandomSampler(
-        weights, num_samples=len(concat_dataset), replacement=True
-    )
-    dataloader = torch.utils.data.DataLoader(
-        concat_dataset,
-        num_workers=4,
-        batch_size=cfg.training.batch_size,
-        sampler=sampler,
-        pin_memory=cfg.device != "cpu",
-        drop_last=False,
-    )
-    dl_iter = cycle(dataloader)
-
-    online_step = 0
-    is_offline = False
-    for env_step in range(cfg.training.online_steps):
-        if env_step == 0:
-            logging.info("Start online training by interacting with environment")
-
-        policy.eval()
-        with torch.no_grad():
-            eval_info = eval_policy(
-                online_training_env,
-                policy,
-                n_episodes=1,
-                return_episode_data=True,
-                start_seed=cfg.training.online_env_seed,
-                enable_progbar=True,
-            )
-
-        add_episodes_inplace(
-            online_dataset,
-            concat_dataset,
-            sampler,
-            hf_dataset=eval_info["episodes"]["hf_dataset"],
-            episode_data_index=eval_info["episodes"]["episode_data_index"],
-            pc_online_samples=cfg.training.online_sampling_ratio,
-        )
-
-        policy.train()
-        for _ in range(cfg.training.online_steps_between_rollouts):
-            batch = next(dl_iter)
-
-            for key in batch:
-                batch[key] = batch[key].to(cfg.device, non_blocking=True)
-
-            train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
-
-            if step % cfg.training.log_freq == 0:
-                log_train_info(logger, train_info, step, cfg, online_dataset, is_offline)
-
-            # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
-            # so we pass in step + 1.
-            evaluate_and_checkpoint_if_needed(step + 1)
-
-            step += 1
-            online_step += 1
-
-    eval_env.close()
-    online_training_env.close()
+    if cfg.training.eval_freq > 0:
+        eval_env.close()
     logging.info("End of training")
 
 
+@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
+def train_cli(cfg: dict):
+    train(
+        cfg,
+        out_dir=hydra.core.hydra_config.HydraConfig.get().run.dir,
+        job_name=hydra.core.hydra_config.HydraConfig.get().job.name,
+    )
+
+
+def train_notebook(out_dir=None, job_name=None, config_name="default", config_path="../configs"):
+    from hydra import compose, initialize
+
+    hydra.core.global_hydra.GlobalHydra.instance().clear()
+    initialize(config_path=config_path)
+    cfg = compose(config_name=config_name)
+    train(cfg, out_dir=out_dir, job_name=job_name)
+
+
 if __name__ == "__main__":
     train_cli()

lerobot/scripts/visualize_dataset.py

@@ -1,3 +1,18 @@
+#!/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.
 """ Visualize data of **all** frames of any episode of a dataset of type LeRobotDataset.
 
 Note: The last frame of the episode doesnt always correspond to a final state.
@@ -47,6 +62,7 @@ local$ rerun ws://localhost:9087
 """
 
 import argparse
+import gc
 import logging
 import time
 from pathlib import Path
@@ -115,15 +131,17 @@ def visualize_dataset(
 
     spawn_local_viewer = mode == "local" and not save
     rr.init(f"{repo_id}/episode_{episode_index}", spawn=spawn_local_viewer)
+
+    # Manually call python garbage collector after `rr.init` to avoid hanging in a blocking flush
+    # when iterating on a dataloader with `num_workers` > 0
+    # TODO(rcadene): remove `gc.collect` when rerun version 0.16 is out, which includes a fix
+    gc.collect()
+
     if mode == "distant":
         rr.serve(open_browser=False, web_port=web_port, ws_port=ws_port)
 
     logging.info("Logging to Rerun")
 
-    if num_workers > 0:
-        # TODO(rcadene): fix data workers hanging when `rr.init` is called
-        logging.warning("If data loader is hanging, try `--num-workers 0`.")
-
     for batch in tqdm.tqdm(dataloader, total=len(dataloader)):
         # iterate over the batch
         for i in range(len(batch["index"])):
@@ -196,7 +214,7 @@ def main():
     parser.add_argument(
         "--num-workers",
         type=int,
-        default=0,
+        default=4,
         help="Number of processes of Dataloader for loading the data.",
     )
     parser.add_argument(
@@ -206,7 +224,8 @@ def main():
         help=(
             "Mode of viewing between 'local' or 'distant'. "
             "'local' requires data to be on a local machine. It spawns a viewer to visualize the data locally. "
-            "'distant' creates a server on the distant machine where the data is stored. Visualize the data by connecting to the server with `rerun ws://localhost:PORT` on the local machine."
+            "'distant' creates a server on the distant machine where the data is stored. "
+            "Visualize the data by connecting to the server with `rerun ws://localhost:PORT` on the local machine."
         ),
     )
     parser.add_argument(
@@ -227,8 +246,8 @@ def main():
         default=0,
         help=(
             "Save a .rrd file in the directory provided by `--output-dir`. "
-            "It also deactivates the spawning of a viewer. ",
-            "Visualize the data by running `rerun path/to/file.rrd` on your local machine.",
+            "It also deactivates the spawning of a viewer. "
+            "Visualize the data by running `rerun path/to/file.rrd` on your local machine."
         ),
     )
     parser.add_argument(

pyproject.toml

@@ -28,40 +28,44 @@ packages = [{include = "lerobot"}]
 
 [tool.poetry.dependencies]
 python = ">=3.10,<3.13"
-termcolor = "^2.4.0"
-omegaconf = "^2.3.0"
-wandb = "^0.16.3"
-imageio = {extras = ["ffmpeg"], version = "^2.34.0"}
-gdown = "^5.1.0"
-hydra-core = "^1.3.2"
-einops = "^0.8.0"
-pymunk = "^6.6.0"
-zarr = "^2.17.0"
-numba = "^0.59.0"
+termcolor = ">=2.4.0"
+omegaconf = ">=2.3.0"
+wandb = ">=0.16.3"
+imageio = {extras = ["ffmpeg"], version = ">=2.34.0"}
+gdown = ">=5.1.0"
+hydra-core = ">=1.3.2"
+einops = ">=0.8.0"
+pymunk = ">=6.6.0"
+zarr = ">=2.17.0"
+numba = ">=0.59.0"
 torch = "^2.2.1"
-opencv-python = "^4.9.0.80"
+opencv-python = ">=4.9.0"
 diffusers = "^0.27.2"
-torchvision = "^0.18.0"
-h5py = "^3.10.0"
-huggingface-hub = "^0.21.4"
-robomimic = "0.2.0"
-gymnasium = "^0.29.1"
-cmake = "^3.29.0.1"
-gym-pusht = { version = "^0.1.1", optional = true}
-gym-xarm = { version = "^0.1.0", optional = true}
-gym-aloha = { version = "^0.1.0", optional = true}
-pre-commit = {version = "^3.7.0", optional = true}
-debugpy = {version = "^1.8.1", optional = true}
-pytest = {version = "^8.1.0", optional = true}
-pytest-cov = {version = "^5.0.0", optional = true}
+torchvision = ">=0.17.1"
+h5py = ">=3.10.0"
+huggingface-hub = {extras = ["hf-transfer"], version = "^0.23.0"}
+gymnasium = ">=0.29.1"
+cmake = ">=3.29.0.1"
+gym-dora = { git = "https://github.com/dora-rs/dora-lerobot.git", subdirectory = "gym_dora", optional = true }
+gym-pusht = { version = ">=0.1.3", optional = true}
+gym-xarm = { version = ">=0.1.1", optional = true}
+gym-aloha = { version = ">=0.1.1", optional = true}
+pre-commit = {version = ">=3.7.0", optional = true}
+debugpy = {version = ">=1.8.1", optional = true}
+pytest = {version = ">=8.1.0", optional = true}
+pytest-cov = {version = ">=5.0.0", optional = true}
 datasets = "^2.19.0"
-imagecodecs = { version = "^2024.1.1", optional = true }
-pyav = "^12.0.5"
-moviepy = "^1.0.3"
-rerun-sdk = "^0.15.1"
+imagecodecs = { version = ">=2024.1.1", optional = true }
+pyav = ">=12.0.5"
+moviepy = ">=1.0.3"
+rerun-sdk = ">=0.15.1"
+deepdiff = ">=7.0.1"
+dynamixel-sdk = {version = "^3.7.31", optional = true}
+
 
 
 [tool.poetry.extras]
+dora = ["gym-dora"]
 pusht = ["gym-pusht"]
 xarm = ["gym-xarm"]
 aloha = ["gym-aloha"]
@@ -104,5 +108,5 @@ ignore-init-module-imports = true
 
 
 [build-system]
-requires = ["poetry-core>=1.5.0"]
+requires = ["poetry-core"]
 build-backend = "poetry.core.masonry.api"

tests/conftest.py

@@ -1,3 +1,18 @@
+#!/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 .utils import DEVICE
 
 

tests/data/lerobot/aloha_mobile_cabinet/meta_data/episode_data_index.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f9347c8d9ac90ee44e6dd86f65043438168df6bbe4bab2d2b875e55ef7376ef
+size 1488

tests/data/lerobot/aloha_mobile_cabinet/meta_data/info.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf148247bf191c7f7e8af738a7b9e147f9ffffeec0e4b9d1c4783c4e384da7eb
+size 33

tests/data/lerobot/aloha_mobile_cabinet/meta_data/stats.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02fc4ea25766269f65752a60b0594c43d799b0ae528cd773bf024b064b5aa329
+size 4344

tests/data/lerobot/aloha_mobile_cabinet/train/data-00000-of-00001.arrow

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55d7b1a06fe3e3051482752740074348bdb5fc98fb2e305b06d6203994117b27
+size 592448