improve script, time saving subtask array

Signed-off-by: Jade Choghari <chogharijade@gmail.com>

.github/workflows/fast_tests.yml

@@ -44,7 +44,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
 
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -61,7 +61,6 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -90,11 +89,5 @@ jobs:
       - name: Install lerobot with test extras
         run: uv sync --extra "test"
 
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          uv run hf auth whoami
-
       - name: Run pytest
         run: uv run pytest tests -vv --maxfail=10

.github/workflows/full_tests.yml

@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -60,7 +60,6 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -88,12 +87,6 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          uv run hf auth whoami
-
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
 
@@ -108,11 +101,9 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      github.repository == 'huggingface/lerobot' && (
-        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-        github.event_name == 'push' ||
-        github.event_name == 'workflow_dispatch'
-      )
+      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch'
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:
@@ -169,7 +160,6 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -181,13 +171,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
-      - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/nightly.yml

@@ -28,7 +28,7 @@ on:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
   DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
   DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
 
@@ -119,7 +119,6 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --shm-size "16gb"
@@ -131,11 +130,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
       - name: Run pytest on CPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -152,7 +146,6 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -164,11 +157,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -186,7 +174,6 @@ jobs:
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
       CUDA_VISIBLE_DEVICES: "0,1,2,3"
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -198,15 +185,12 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
       - name: Verify GPU availability
         run: |
           nvidia-smi
           python -c "import torch; print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
 
       - name: Run multi-GPU training tests
-        run: pytest -vv tests/training/
+      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
+        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
+        timeout-minutes: 10

.github/workflows/quality.yml

@@ -50,7 +50,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.12'
+          python-version: '3.10'
 
       - name: Run pre-commit hooks
         uses: pre-commit/action@v3.0.1 # zizmor: ignore[unpinned-uses]

.github/workflows/release.yml

@@ -22,7 +22,7 @@ on:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
 
 jobs:
   # This job builds the Python package and publishes it to PyPI
@@ -45,7 +45,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.12'
+          python-version: '3.10'
 
       - name: Extract Version
         id: extract_info
@@ -83,6 +83,14 @@ jobs:
             exit 1
           fi
 
+      - name: Remove Tags with Git dependencies
+        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
+        run: |
+          echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
+          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
+          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
+          echo "::info:: Git dependencies removed. Proceeding with build."
+
       - name: Install build dependencies
         run: python -m pip install build
 

.github/workflows/unbound_deps_tests.yml

@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -48,7 +48,6 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -80,11 +79,7 @@ jobs:
 
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          uv run hf auth whoami
+
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv
 
@@ -96,7 +91,6 @@ jobs:
     name: Build and Push Docker
     runs-on:
       group: aws-general-8-plus
-    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
     env:
@@ -142,7 +136,6 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -154,11 +147,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv
       - name: Run end-to-end tests

.pre-commit-config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 default_language_version:
-    python: python3.12
+    python: python3.10
 
 exclude: "tests/artifacts/.*\\.safetensors$"
 
@@ -55,7 +55,7 @@ repos:
     rev: v3.21.0
     hooks:
     -   id: pyupgrade
-        args: [--py312-plus]
+        args: [--py310-plus]
 
   ##### Markdown Quality #####
   - repo: https://github.com/rbubley/mirrors-prettier

AI_POLICY.md

@@ -1,25 +0,0 @@
-# AI Usage Policy
-
-The LeRobot project welcomes contributions from everyone, and we have a few guidelines regarding AI usage to ensure high code quality, clear communication, and a healthy open-source ecosystem:
-
-- **Please disclose significant AI assistance.** If you used AI tools (e.g., Copilot, Claude, Cursor, ChatGPT) to generate a substantial portion of your code or text, let us know in your PR description. Transparency helps us review your changes more effectively.
-- **Own your code (The Human-in-the-Loop).** You must fully understand all the changes you are proposing. If you cannot explain what your AI-assisted code does or how it interacts with LeRobot's broader architecture, please take the time to learn and test it before submitting.
-- **Keep issues and discussions focused.** You are welcome to use AI to help draft issues or PR descriptions, but please review and edit them carefully before posting. AI can often be overly verbose; trimming the noise and getting straight to the point helps our maintainers address your needs faster.
-
-Our core maintainers also use AI tools to aid their workflows, but they do so while bringing deep contextual knowledge of the LeRobot codebase to validate the output. We ask all contributors to apply that same level of rigor.
-
-## Remember the Human Maintainers
-
-Please remember that LeRobot is maintained by a dedicated team of humans.
-
-Every discussion, issue, and pull request is read and reviewed by real people. While AI tools can generate thousands of lines of code in seconds, reviewing that code still takes human time and energy. Submitting unverified or low-effort AI output puts an unfair burden on our maintainers.
-
-Today, the quality of the AI output still heavily depends on the developer driving the tool. We ask that you respect our maintainers' time by thoroughly vetting, testing, and refining your submissions.
-
-## AI is Welcome Here
-
-LeRobot operates at the cutting edge of AI and robotics, and many of our maintainers actively embrace AI coding assistants as valuable productivity tools. We are a pro-AI project!
-
-Our reason for having an AI policy is not an anti-AI stance. Rather, it exists to ensure that AI is used to enhance human contributions, not replace them with unverified noise. It's about how the tools are used, not the tools themselves.
-
-We value the unique human insight you bring to the LeRobot community. Let AI empower your workflow, but always let your own judgment take the wheel.

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md) and our [AI policy](https://github.com/huggingface/lerobot/blob/main/AI_POLICY.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
 
 ## Ways to Contribute
 
@@ -32,7 +32,7 @@ git remote add upstream https://github.com/huggingface/lerobot.git
 
 ### 2. Environment Installation
 
-Please follow our [Installation Guide](https://huggingface.co/docs/lerobot/installation) for the environment setup & installation from source.
+Please follow our [Installation Guide](./docs/source/installation.mdx) for the environment setup & installation from source.
 
 ## Running Tests & Quality Checks
 
@@ -75,8 +75,8 @@ pytest -sv tests/test_specific_feature.py
 
 Use the templates for required fields and examples.
 
-- **Issues:** Follow the [ticket template](https://github.com/huggingface/lerobot/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml).
-- **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md).
+- **Issues:** Follow the [ticket template](./.github/ISSUE_TEMPLATE/bug-report.yml).
+- **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](./.github/PULL_REQUEST_TEMPLATE.md).
 
 One member of the LeRobot team will then review your contribution.
 

MANIFEST.in

@@ -1,3 +1,2 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
-include src/lerobot/envs/metaworld_config.json

README.md

@@ -135,7 +135,7 @@ Learn how to implement your own simulation environment or benchmark and distribu
 
 ## Citation
 
-If you use LeRobot in your project, please cite the GitHub repository to acknowledge the ongoing development and contributors:
+If you use LeRobot in your research, please cite:
 
 ```bibtex
 @misc{cadene2024lerobot,
@@ -146,26 +146,9 @@ If you use LeRobot in your project, please cite the GitHub repository to acknowl
 }
 ```
 
-If you are referencing our research or the academic paper, please also cite our ICLR publication:
-
-<details>
-<summary><b>ICLR 2026 Paper</b></summary>
-
-```bibtex
-@inproceedings{cadenelerobot,
-  title={LeRobot: An Open-Source Library for End-to-End Robot Learning},
-  author={Cadene, Remi and Alibert, Simon and Capuano, Francesco and Aractingi, Michel and Zouitine, Adil and Kooijmans, Pepijn and Choghari, Jade and Russi, Martino and Pascal, Caroline and Palma, Steven and Shukor, Mustafa and Moss, Jess and Soare, Alexander and Aubakirova, Dana and Lhoest, Quentin and Gallou\'edec, Quentin and Wolf, Thomas},
-  booktitle={The Fourteenth International Conference on Learning Representations},
-  year={2026},
-  url={https://arxiv.org/abs/2602.22818}
-}
-```
-
-</details>
-
 ## Contribute
 
-We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
+We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](./CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
 
 <p align="center">
   <img alt="SO101 Video" src="./media/readme/so100_video.webp" width="640px">

benchmarks/video/README.md

@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 For these reasons, we run this benchmark on four representative datasets:
 
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
-- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
-- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
 
 Note: The datasets used for this benchmark need to be image datasets, not video datasets.
 
@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
     --vcodec libsvtav1 \
     --pix-fmt yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read
 
 These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
 
-| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
+| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
 
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
+| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
 
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
+| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

docker/Dockerfile.internal

@@ -24,7 +24,7 @@ ARG OS_VERSION=22.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 
 # Define Python version argument
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.10
 
 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -85,8 +85,6 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
 
 RUN uv pip install --no-cache ".[all]"
 
-RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
-
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .

docker/Dockerfile.user

@@ -18,10 +18,8 @@
 # docker build -f docker/Dockerfile.user -t lerobot-user .
 # docker run -it --rm lerobot-user
 
-# With USB physical access : docker run -it --device=/dev/ -v /dev/:/dev/ --rm lerobot-user
-
 # Configure the base image
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.10
 FROM python:${PYTHON_VERSION}-slim
 
 # Configure environment variables

docs/source/_toctree.yml

@@ -19,8 +19,6 @@
     title: Multi GPU training
   - local: peft_training
     title: Training with PEFT (e.g., LoRA)
-  - local: rename_map
-    title: Using Rename Map and Empty Cameras
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3
@@ -29,10 +27,10 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
+  - local: annotation_tools
+    title: Using the Annotation Tools
   - local: dataset_subtask
     title: Using Subtasks in the Dataset
-  - local: streaming_video_encoding
-    title: Streaming Video Encoding
   title: "Datasets"
 - sections:
   - local: act

docs/source/act.mdx

@@ -88,8 +88,5 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
   --dataset.num_episodes=10 \
   --dataset.single_task="Your task description" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=${HF_USER}/act_policy
 ```

docs/source/annotation_tools.mdx

@@ -0,0 +1,425 @@
+# Dataset Annotation Tools
+
+This guide explains how to use the automatic annotation tools to add skill labels and synthetic dialogue to your LeRobot datasets.
+
+## Overview
+
+The annotation pipeline consists of two main components:
+
+1. **Subtask Annotation** (`subtask_annotate.py`): Automatically segments robot demonstrations into atomic skills using Vision-Language Models (VLMs)
+2. **High-Level Annotation** (`high_level_annotate.py`): Generates synthetic user prompts and robot utterances for hierarchical policy training
+
+These tools enable you to transform raw robot demonstration data into richly annotated datasets suitable for training hierarchical policies.
+
+## Installation Requirements
+
+Before using the annotation tools, ensure you have the required dependencies:
+
+```bash
+pip install transformers qwen-vl-utils opencv-python rich pandas pyarrow
+```
+
+You'll also need FFmpeg for video processing:
+
+```bash
+# Ubuntu/Debian
+sudo apt-get install ffmpeg
+
+# macOS
+brew install ffmpeg
+```
+
+## Part 1: Subtask Annotation
+
+### What It Does
+
+The subtask annotator segments each episode into short atomic manipulation skills (1-3 seconds each). For example, a "pick and place" episode might be segmented into:
+- "reach towards object" (0.0s - 1.2s)
+- "grasp object" (1.2s - 2.1s)
+- "lift object" (2.1s - 3.5s)
+- "move to target" (3.5s - 5.0s)
+- "release object" (5.0s - 6.2s)
+
+### Usage
+
+#### Basic Example
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id your-username/your-dataset \
+    --video-key observation.images.base \
+    --output-dir /path/to/output
+```
+
+#### With Local Dataset
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --data-dir /path/to/local/dataset \
+    --video-key observation.images.base \
+    --output-dir /path/to/output
+```
+
+#### Advanced Options
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id your-username/your-dataset \
+    --video-key observation.images.base \
+    --model Qwen/Qwen2-VL-7B-Instruct \
+    --batch-size 16 \
+    --output-dir /path/to/output \
+    --push-to-hub
+```
+
+### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--repo-id` | HuggingFace Hub dataset ID | Required (or use --data-dir) |
+| `--data-dir` | Path to local dataset | Required (or use --repo-id) |
+| `--video-key` | Video observation key | Required |
+| `--model` | VLM model to use | `Qwen/Qwen2-VL-7B-Instruct` |
+| `--device` | Device to run model on | `cuda` |
+| `--dtype` | Model dtype | `bfloat16` |
+| `--batch-size` | Episodes per batch | `8` |
+| `--episodes` | Specific episodes to annotate | All episodes |
+| `--output-dir` | Output directory | Auto-generated |
+| `--push-to-hub` | Push to HuggingFace Hub | `False` |
+
+### Supported Models
+
+- **Qwen2-VL**: `Qwen/Qwen2-VL-2B-Instruct`, `Qwen/Qwen2-VL-7B-Instruct`, `Qwen/Qwen2-VL-72B-Instruct`
+- **Qwen3-VL**: `Qwen/Qwen3-VL-30B-A3B-Instruct`
+
+### Output Files
+
+The subtask annotation creates the following files in your dataset:
+
+1. **`meta/subtasks.parquet`**: DataFrame with unique subtask names
+   ```python
+   # Structure:
+   # Index: subtask name (string)
+   # Column: subtask_index (int64)
+   ```
+
+2. **`meta/skills.json`**: Raw skill annotations with timestamps
+   ```json
+   {
+     "coarse_description": "Pick and place the object",
+     "skill_to_subtask_index": {
+       "reach towards object": 0,
+       "grasp object": 1,
+       ...
+     },
+     "episodes": {
+       "0": {
+         "episode_index": 0,
+         "description": "Pick and place the object",
+         "skills": [
+           {"name": "reach towards object", "start": 0.0, "end": 1.2},
+           {"name": "grasp object", "start": 1.2, "end": 2.1},
+           ...
+         ]
+       }
+     }
+   }
+   ```
+
+3. **`subtask_index` feature**: Added to each frame in the dataset
+   - Type: `int64`
+   - Shape: `(1,)`
+   - Maps each frame to its corresponding subtask
+
+### Accessing Subtask Annotations
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Load annotated dataset
+dataset = LeRobotDataset(repo_id="your/dataset_with_subtasks")
+
+# Get a frame
+frame = dataset[100]
+
+# Get the subtask for this frame
+subtask_idx = frame["subtask_index"].item()
+subtask_name = dataset.meta.subtasks.iloc[subtask_idx].name
+
+print(f"Frame 100 is performing: {subtask_name}")
+
+# Load all subtasks
+subtasks_df = dataset.meta.subtasks
+print(subtasks_df)
+```
+
+## Part 2: High-Level Annotation
+
+### What It Does
+
+The high-level annotator generates synthetic dialogue for hierarchical policy training. For each skill, it creates:
+- **User Prompt** (`ℓ_t`): A natural language request from the user
+- **Robot Utterance** (`u_t`): A natural language response from the robot
+
+This enables training policies that can understand and respond to human instructions in natural dialogue.
+
+### Prerequisites
+
+**Important**: You must run subtask annotation first! High-level annotation requires the `skills.json` file generated by subtask annotation.
+
+### Usage
+
+#### Image Mode (Default)
+
+Samples frames at regular intervals and passes images to the VLM:
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --repo-id your/dataset_with_subtasks \
+    --model Qwen/Qwen2-VL-7B-Instruct \
+    --image-key observation.images.base \
+    --output-dir /path/to/output
+```
+
+#### Video Mode
+
+Passes entire episode videos to the VLM for better temporal understanding:
+
+```bash
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --repo-id your/dataset_with_subtasks \
+    --model Qwen/Qwen2-VL-7B-Instruct \
+    --video-mode \
+    --video-key observation.images.base \
+    --video-batch-size 4 \
+    --output-dir /path/to/output
+```
+
+### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--repo-id` | HuggingFace Hub dataset ID | Required (or use --data-dir) |
+| `--data-dir` | Path to local dataset | Required (or use --repo-id) |
+| `--model` | VLM model to use | `Qwen/Qwen2-VL-7B-Instruct` |
+| `--image-key` | Image observation key (image mode) | First camera key |
+| `--video-mode` | Use video instead of images | `False` |
+| `--video-key` | Video observation key (video mode) | Auto-detected |
+| `--video-batch-size` | Episodes per batch (video mode) | `1` |
+| `--sample-interval` | Sampling interval in seconds | `1.0` |
+| `--temperature` | Sampling temperature | `0.7` |
+| `--output-dir` | Output directory | Auto-generated |
+| `--push-to-hub` | Push to HuggingFace Hub | `False` |
+
+### Output Files
+
+The high-level annotation creates:
+
+1. **`meta/tasks_high_level.parquet`**: DataFrame with high-level tasks
+   ```python
+   # Structure:
+   # Index: task string (concatenated user_prompt | robot_utterance)
+   # Columns:
+   #   - task_index: int64
+   #   - user_prompt: string
+   #   - robot_utterance: string
+   #   - skill: string (associated subtask)
+   #   - scenario_type: string
+   #   - response_type: string
+   ```
+
+2. **`meta/syn_annotations.jsonl`**: Debug annotations (JSONL format)
+   ```json
+   {"episode_id": 0, "timestamp": 1.5, "skill_current": "grasp object", "user_prompt": "Can you pick that up?", "robot_utterance": "Sure, I'll grasp it now", ...}
+   ```
+
+3. **`task_index_high_level` feature**: Added to each frame
+   - Type: `int64`
+   - Shape: `(1,)`
+   - Maps each frame to its high-level task
+
+### Dialogue Types Generated
+
+The system generates diverse interaction types:
+
+**Scenario Types:**
+- `specific_object`: "Pick up the red block"
+- `negative_task`: "Don't touch the blue one"
+- `situated_correction`: "Actually, move to the other box instead"
+- `implicit_request`: "I need something red for the tower"
+- `constraint_based`: "Make sure to handle it gently"
+
+**Response Types:**
+- `confirmation`: "OK, I'll pick it up"
+- `clarification`: "Just to confirm, you want me to pick up the red block?"
+- `acknowledgment`: "Got it, picking up the red block"
+- `constraint_acknowledgment`: "Sure, I'll pick it up gently"
+
+### Accessing High-Level Annotations
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+import pandas as pd
+
+# Load annotated dataset
+dataset = LeRobotDataset(repo_id="your/dataset_with_high_level_tasks")
+
+# Get a frame
+frame = dataset[100]
+
+# Get the high-level task
+task_idx = frame["task_index_high_level"].item()
+
+# Load tasks metadata
+tasks_df = pd.read_parquet(dataset.root / "meta" / "tasks_high_level.parquet")
+task_row = tasks_df[tasks_df["task_index"] == task_idx].iloc[0]
+
+print(f"User: {task_row['user_prompt']}")
+print(f"Robot: {task_row['robot_utterance']}")
+print(f"Skill: {task_row['skill']}")
+
+# Use in a DataLoader
+import torch
+from torch.utils.data import DataLoader
+
+dataloader = DataLoader(dataset, batch_size=32, shuffle=True)
+batch = next(iter(dataloader))
+
+print(f"Task indices: {batch['task_index_high_level']}")
+print(f"User prompts: {batch['user_prompt'][0]}")
+print(f"Robot utterances: {batch['robot_utterance'][0]}")
+```
+
+## Complete Pipeline Example
+
+Here's how to run both annotation stages:
+
+```bash
+#!/bin/bash
+
+REPO_ID="your-username/your-dataset"
+MODEL="Qwen/Qwen2-VL-7B-Instruct"
+OUTPUT_DIR="/path/to/output"
+
+# Step 1: Subtask Annotation
+python src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+    --repo-id "$REPO_ID" \
+    --video-key observation.images.base \
+    --model "$MODEL" \
+    --batch-size 8 \
+    --output-dir "${OUTPUT_DIR}/subtasks"
+
+# Step 2: High-Level Annotation (Image Mode)
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --data-dir "${OUTPUT_DIR}/subtasks" \
+    --model "$MODEL" \
+    --image-key observation.images.base \
+    --sample-interval 1.0 \
+    --output-dir "${OUTPUT_DIR}/final"
+
+# Or Step 2: High-Level Annotation (Video Mode - Recommended)
+python src/lerobot/policies/pi05_full/annotate/high_level_annotate.py \
+    --data-dir "${OUTPUT_DIR}/subtasks" \
+    --model "$MODEL" \
+    --video-mode \
+    --video-key observation.images.base \
+    --video-batch-size 4 \
+    --output-dir "${OUTPUT_DIR}/final"
+```
+
+## Performance Tips
+
+### For Faster Processing
+
+1. **Increase batch size**: Use `--batch-size 16` or higher (subtask annotation)
+2. **Increase video batch size**: Use `--video-batch-size 8` (high-level annotation in video mode)
+3. **Larger sampling interval**: Use `--sample-interval 5.0` for testing (samples every 5 seconds instead of 1)
+4. **Use smaller models**: `Qwen/Qwen2-VL-2B-Instruct` is faster than `Qwen2-VL-7B-Instruct`
+5. **Process specific episodes**: Use `--episodes 0 1 2 3` to annotate only a subset
+
+### For Better Quality
+
+1. **Use larger models**: `Qwen/Qwen3-VL-30B-A3B-Instruct` or `Qwen/Qwen2-VL-72B-Instruct`
+2. **Use video mode**: Provides better temporal context
+3. **Smaller sampling intervals**: `--sample-interval 0.5` for dense annotations
+4. **Adjust temperature**: Use `--temperature 0.9` for more diverse dialogue
+
+## Memory Requirements
+
+| Model | GPU Memory | Recommended Batch Size |
+|-------|------------|------------------------|
+| Qwen2-VL-2B | ~8 GB | 16-32 |
+| Qwen2-VL-7B | ~16 GB | 8-16 |
+| Qwen2-VL-72B | ~80 GB | 1-2 |
+| Qwen3-VL-30B | ~40 GB | 4-8 |
+
+## Troubleshooting
+
+### "FFmpeg not found"
+```bash
+# Install FFmpeg
+sudo apt-get install ffmpeg  # Ubuntu/Debian
+brew install ffmpeg          # macOS
+```
+
+### "CUDA out of memory"
+- Reduce batch size: `--batch-size 1` or `--video-batch-size 1`
+- Use smaller model: `Qwen/Qwen2-VL-2B-Instruct`
+- Use CPU: `--device cpu` (much slower)
+
+### "No skills.json found"
+Run subtask annotation first before high-level annotation.
+
+### "Video key not found"
+List available keys:
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+dataset = LeRobotDataset(repo_id="your/dataset")
+print("Video keys:", dataset.meta.video_keys)
+print("Camera keys:", dataset.meta.camera_keys)
+```
+
+## Dataset Structure After Annotation
+
+```
+your_dataset_with_high_level_tasks/
+├── meta/
+│   ├── info.json                      # Original metadata
+│   ├── tasks.parquet                  # Original tasks (preserved)
+│   ├── subtasks.parquet              # NEW: Subtask names and indices
+│   ├── skills.json                    # NEW: Raw skill annotations with timestamps
+│   ├── tasks_high_level.parquet      # NEW: High-level tasks with dialogue
+│   └── syn_annotations.jsonl         # NEW: Debug annotations
+├── data/
+│   └── chunk-000/
+│       ├── observation.images.base.mp4
+│       ├── action.safetensors
+│       ├── subtask_index.safetensors # NEW: Subtask per frame
+│       └── task_index_high_level.safetensors # NEW: High-level task per frame
+└── videos/
+    └── ...
+```
+
+## Citation
+
+If you use these annotation tools in your research, please cite:
+
+```bibtex
+@article{lerobot2024,
+  title={LeRobot: State-of-the-art Machine Learning for Real-World Robotics},
+  author={LeRobot Contributors},
+  year={2024},
+  url={https://github.com/huggingface/lerobot}
+}
+```
+
+## Next Steps
+
+After annotation, you can:
+1. Train hierarchical policies using the subtask and high-level annotations
+2. Use the synthetic dialogue for instruction-following policy training
+3. Analyze skill distributions and dialogue patterns
+4. Share your annotated dataset on HuggingFace Hub with `--push-to-hub`
+
+For training examples, see the [training documentation](../training/).
+

docs/source/async.mdx

@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
     --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
     --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
     --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
     --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
     --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
     --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
@@ -310,4 +310,4 @@ Asynchronous inference represents a significant advancement in real-time robotic
 - **Universal Compatibility**: Works with all LeRobot-supported policies, from lightweight ACT models to vision-language models like SmolVLA
 
 Start experimenting with the default parameters, monitor your action queue sizes, and iteratively refine your setup to achieve optimal performance for your specific use case.
-If you want to discuss this further, hop into our [Discord community](https://discord.gg/s3KuuzsPFb), or open an issue on our [GitHub repository](https://github.com/huggingface/lerobot/issues).
+If you want to discuss this further, hop into our [Discord community](https://discord.gg/s3KuuzsPFb), or open an issue on our [GitHub repository](https://github.com/lerobot/lerobot/issues).

docs/source/bring_your_own_policies.mdx

@@ -32,7 +32,7 @@ version = "0.1.0"
 dependencies = [
     # your policy-specific dependencies
 ]
-requires-python = ">= 3.12"
+requires-python = ">= 3.11"
 
 [build-system]
 build-backend = # your-build-backend
@@ -82,7 +82,7 @@ Create your policy implementation by inheriting from LeRobot's base `PreTrainedP
 # modeling_my_custom_policy.py
 import torch
 import torch.nn as nn
-from typing import Any
+from typing import Dict, Any
 
 from lerobot.policies.pretrained import PreTrainedPolicy
 from .configuration_my_custom_policy import MyCustomPolicyConfig
@@ -91,7 +91,7 @@ class MyCustomPolicy(PreTrainedPolicy):
     config_class = MyCustomPolicyConfig
     name = "my_custom_policy"
 
-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: Dict[str, Any] = None):
         super().__init__(config, dataset_stats)
         ...
 ```
@@ -102,7 +102,7 @@ Create processor functions:
 
 ```python
 # processor_my_custom_policy.py
-from typing import Any
+from typing import Dict, Any
 import torch
 
 

docs/source/earthrover_mini_plus.mdx

@@ -13,7 +13,7 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 ### Hardware
 
 - EarthRover Mini robot
-- Computer with Python 3.12 or newer
+- Computer with Python 3.10 or newer
 - Internet connection
 
 ### Setting Up the Frodobots SDK
@@ -170,13 +170,13 @@ Once you can drive the robot well, you can start recording data to train AI mode
 We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,16 +185,13 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-lerobot-record \
+python src/lerobot/scripts/lerobot_record.py \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -204,26 +201,22 @@ Replace `your_username/dataset_name` with your Hugging Face username and a name
 
 Your dataset includes:
 
-**Your Actions (2 features)**:
+**Your Actions (2 things)**:
 
-- `linear_velocity`: How much you moved forward/backward
-- `angular_velocity`: How much you turned left/right
+- How much you moved forward/backward
+- How much you turned left/right
 
-**Robot Observations (24 features)**:
+**Robot Observations (12 things)**:
 
 - Front camera video
 - Rear camera video
 - Current speed
 - Battery level
-- Orientation
-- GPS (latitude, longitude, signal strength)
+- Which way the robot is facing
+- GPS location (latitude, longitude, signal strength)
 - Network signal strength
 - Vibration level
-- Lamp state (on/off)
-- Accelerometer (x, y, z)
-- Gyroscope (x, y, z)
-- Magnetometer (x, y, z)
-- Wheel RPMs (4 wheels)
+- Lamp status (on/off)
 
 ### Where Your Data Goes
 

docs/source/envhub.mdx

@@ -155,10 +155,10 @@ Upload your repository to Hugging Face:
 pip install huggingface_hub
 
 # Login to Hugging Face
-hf auth login
+huggingface-cli login
 
 # Create a new repository
-hf repo create my-org/my-custom-env
+huggingface-cli repo create my-custom-env --type space --org my-org
 
 # Initialize git and push
 git init

docs/source/groot.mdx

@@ -120,12 +120,9 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=<user>/eval_groot-bimanual  \
   --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
-  --policy.path=<user>/groot-bimanual \ # your trained model
-  --dataset.episode_time_s=30 \
+  --dataset.single_task="Grab and handover the red cube to the other arm"
+  --policy.path=<user>/groot-bimanual # your trained model
+  --dataset.episode_time_s=30
   --dataset.reset_time_s=10
 ```
 

docs/source/hope_jr.mdx

@@ -224,15 +224,12 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
     --dataset.single_task="Hand recording test with video data" \
     --dataset.num_episodes=1 \
     --dataset.episode_time_s=5 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -244,7 +241,7 @@ lerobot-replay \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
     --robot.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_camera \
+    --dataset.repo_id=nepyope/hand_record_test_with_camera \
     --dataset.episode=0
 ```
 
@@ -252,13 +249,13 @@ lerobot-replay \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
   --job_name=hopejr \
   --policy.device=mps \
   --wandb.enable=true \
-  --policy.repo_id=<USER>/hand_test_policy
+  --policy.repo_id=nepyope/hand_test_policy
 ```
 
 ### Evaluate
@@ -273,11 +270,8 @@ lerobot-record \
   --robot.side=right \
   --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
   --display_data=false \
-  --dataset.repo_id=<USER>/eval_hopejr \
+  --dataset.repo_id=nepyope/eval_hopejr \
   --dataset.single_task="Evaluate hopejr hand policy" \
   --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -159,13 +159,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,10 +185,7 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube" \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
+    --dataset.single_task="Grab the black cube"
 ```
 </hfoption>
 <hfoption id="API example">
@@ -327,7 +324,7 @@ You can look for other LeRobot datasets on the hub by searching for `LeRobot` [t
 You can also push your local dataset to the Hub manually, running:
 
 ```bash
-hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```
 
 #### Record function
@@ -491,7 +488,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Once training is done, upload the latest checkpoint with:
 
 ```bash
-hf upload ${HF_USER}/act_so101_test \
+huggingface-cli upload ${HF_USER}/act_so101_test \
   outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```
 
@@ -499,7 +496,7 @@ You can also upload intermediate checkpoints with:
 
 ```bash
 CKPT=010000
-hf upload ${HF_USER}/act_so101_test${CKPT} \
+huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
   outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```
 
@@ -518,9 +515,6 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${HF_USER}/eval_so100 \
   --dataset.single_task="Put lego brick into the transparent box" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/installation.mdx

@@ -1,57 +1,30 @@
 # Installation
 
-This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and `ffmpeg` installed with the `libsvtav1` encoder, then skip ahead to [Environment Setup](#step-2-environment-setup).
-
-## Step 1 (`conda` only): Install [`miniforge`](https://conda-forge.org/download/)
+## Install [`miniforge`](https://conda-forge.org/download/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```
 
-## Step 2: Environment Setup
+## Environment Setup
 
-Create a virtual environment with Python 3.12:
+Create a virtual environment with Python 3.10, using conda:
 
-<!-- prettier-ignore-start -->
-<hfoptions id="create_venv">
-<hfoption id="conda">
 ```bash
-conda create -y -n lerobot python=3.12
+conda create -y -n lerobot python=3.10
 ```
-</hfoption>
-<hfoption id="uv">
+
+Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
+
 ```bash
-uv python install 3.12
-uv venv --python 3.12
-```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
-
-Then activate your virtual environment, you have to do this each time you open a shell to use lerobot:
-
-<!-- prettier-ignore-start -->
-<hfoptions id="activate_venv">
-<hfoption id="conda">```bash
 conda activate lerobot
-```</hfoption>
-<hfoption id="uv">
-```bash
-# Linux/macOSsource
-source .venv/bin/activate
-# Windows PowerShell
-source .venv\Scripts\Activate.ps1
 ```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
 
 When using `conda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
-ffmpeg -version  # ffmpeg 8.X is not yet supported !
 ```
 
 > [!TIP]
@@ -65,17 +38,7 @@ ffmpeg -version  # ffmpeg 8.X is not yet supported !
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-> [!NOTE]
-> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
->
-> ```bash
-> conda install evdev -c conda-forge
-> ```
-
-> [!IMPORTANT]
-> If you are using `uv` you will have to install `ffmpeg` system-wide (outside of the virtual environment). You rely on `uv` and `torchcodec` ability to dynamically link to the system `ffmpeg`.
-
-## Step 3: Install LeRobot 🤗
+## Install LeRobot 🤗
 
 ### From Source
 
@@ -88,45 +51,23 @@ cd lerobot
 
 Then, install the library in editable mode. This is useful if you plan to contribute to the code.
 
-<!-- prettier-ignore-start -->
-<hfoptions id="install_lerobot_src">
-<hfoption id="conda">
 ```bash
 pip install -e .
 ```
-</hfoption>
-<hfoption id="uv">
-```bash
-uv pip install -e .
-```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
 
 ### Installation from PyPI
 
 **Core Library:**
 Install the base package with:
 
-<!-- prettier-ignore-start -->
-<hfoptions id="install_lerobot_pypi">
-<hfoption id="conda">
 ```bash
 pip install lerobot
 ```
-</hfoption>
-<hfoption id="uv">
-```bash
-uv pip install lerobot
-```
-</hfoption>
-</hfoptions>
-<!-- prettier-ignore-end -->
 
 _This installs only the default dependencies._
 
 **Extra Features:**
-To install additional functionality, use one of the following (If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.):
+To install additional functionality, use one of the following:
 
 ```bash
 pip install 'lerobot[all]'          # All available features
@@ -140,10 +81,13 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
+> [!NOTE]
+> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
+
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
-To install these for Linux run:
+To install these for linux run:
 
 ```bash
 sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
@@ -153,7 +97,7 @@ For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/
 
 ## Optional dependencies
 
-LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
+LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`.
 
 ### Simulations
 

docs/source/lekiwi.mdx

@@ -279,13 +279,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
 

docs/source/lerobot-dataset-v3.mdx

@@ -41,10 +41,7 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/record-test \
   --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube" \
-  --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
-  --dataset.encoder_threads=2
+  --dataset.single_task="Grab the black cube"
 ```
 
 See the [recording guide](./il_robots#record-a-dataset) for more details.

docs/source/phone_teleop.mdx

@@ -66,13 +66,12 @@ Run on of the examples scripts to teleoperate, record a dataset, replay a datase
 
 All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
 
-Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.
+Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
 
 - Run this example to teleoperate:
 
   ```bash
-  cd examples/phone_to_so100
-  python teleoperate.py
+  python examples/phone_to_so100/teleoperate.py
   ```
 
 After running the example:
@@ -85,22 +84,19 @@ Additionally you can customize mapping or safety limits by editing the processor
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 
   ```bash
-  cd examples/phone_to_so100
-  python record.py
+  python examples/phone_to_so100/record.py
   ```
 
 - Run this example to replay recorded episodes:
 
   ```bash
-  cd examples/phone_to_so100
-  python replay.py
+  python examples/phone_to_so100/replay.py
   ```
 
 - Run this example to evaluate a pretrained policy:
 
   ```bash
-  cd examples/phone_to_so100
-  python evaluate.py
+  python examples/phone_to_so100/evaluate.py
   ```
 
 ### Important pipeline steps and options

docs/source/pi0.mdx

@@ -34,6 +34,11 @@ As described by Physical Intelligence, while AI has achieved remarkable success
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Training Data and Capabilities
 
 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:
@@ -55,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \

docs/source/pi05.mdx

@@ -36,6 +36,11 @@ This diverse training mixture creates a "curriculum" that enables generalization
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Usage
 
 To use π₀.₅ in your LeRobot configuration, specify the policy type as:
@@ -51,7 +56,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py\
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \

docs/source/pi0fast.mdx

@@ -43,11 +43,16 @@ This approach can transform **any existing VLM** into a VLA by training it to pr
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install the pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Training a Custom FAST Tokenizer
 
 You have two options for the FAST tokenizer:
 
-1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
+1. **Use the pre-trained tokenizer**: The `physical-intelligence/fast` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
 
 2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
 
@@ -109,15 +114,15 @@ lerobot-train \
 
 ### Key Training Parameters
 
-| Parameter                              | Description                                        | Default                         |
-| -------------------------------------- | -------------------------------------------------- | ------------------------------- |
-| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                         |
-| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                       |
-| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                            |
-| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                            |
-| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                           |
-| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `lerobot/fast-action-tokenizer` |
-| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
+| Parameter                              | Description                                        | Default                      |
+| -------------------------------------- | -------------------------------------------------- | ---------------------------- |
+| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                      |
+| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                    |
+| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                         |
+| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                         |
+| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                        |
+| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `physical-intelligence/fast` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                      |
 
 ## Inference
 

docs/source/reachy2.mdx

@@ -159,9 +159,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -201,9 +198,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/rename_map.mdx

@@ -1,114 +0,0 @@
-# Rename Map and Empty Cameras
-
-When you train, evaluate, or record with a robot policy, your **dataset** or **environment** provides observations under one set of keys (e.g. `observation.images.front`, `observation.images.eagle`), while your **policy** expects another (e.g. `observation.images.image`, `observation.images.image2`). The **rename map** bridges that gap without changing the policy or data source.
-
-> **Scope:** The rename map only renames **observation** keys (images and state). Action keys are not affected.
-
-## Why observation keys don't always match
-
-Policies have a fixed set of **input feature names** baked into their pretrained config. For example:
-
-- [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero) expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb`.
-- [xvla-base](https://huggingface.co/lerobot/xvla-base) expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`.
-
-Your dataset might use different names entirely (e.g. `observation.images.front`, `observation.images.eagle`, `observation.images.glove`), and your eval environment might use yet another set. Rather than editing the policy config or renaming columns in the dataset, you pass a **rename map**: a JSON dictionary that maps source keys to the keys the policy expects. Renaming happens inside the preprocessor pipeline, so the policy always sees its expected keys.
-
-## Using the rename map
-
-Pass the mapping as a JSON string on the command line. The convention is always:
-
-```
---rename_map='{"source_key": "policy_key", ...}'
-```
-
-where **source_key** is what the dataset or environment provides, and **policy_key** is what the policy expects.
-
-Only listed keys are renamed; everything else passes through unchanged. Order of entries doesn't matter.
-
-Supported policies: **PI0**, **PI05**, **PI0Fast**, **SmolVLA**, and **XVLA**.
-
-### Training
-
-Suppose you fine-tune [lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base) on a dataset with images under `observation.images.front`, `observation.images.eagle`, and `observation.images.glove`. XVLA expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=YOUR_DATASET \
-  --output_dir=./outputs/xvla_training \
-  --job_name=xvla_training \
-  --policy.path="lerobot/xvla-base" \
-  --policy.repo_id="HF_USER/xvla-your-robot" \
-  --policy.dtype=bfloat16 \
-  --policy.action_mode=auto \
-  --steps=20000 \
-  --policy.device=cuda \
-  --policy.freeze_vision_encoder=false \
-  --policy.freeze_language_encoder=false \
-  --policy.train_policy_transformer=true \
-  --policy.train_soft_prompts=true \
-  --rename_map='{"observation.images.front": "observation.images.image", "observation.images.eagle": "observation.images.image2", "observation.images.glove": "observation.images.image3"}'
-```
-
-### Evaluation
-
-A policy that expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb` (e.g. [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)), but the LIBERO environment returns `observation.images.image` and `observation.images.image2`:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/pi0fast-libero \
-  --env.type=libero \
-  ... \
-  --rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
-```
-
-### Recording
-
-`lerobot-record` also supports rename maps, nested under the dataset config:
-
-```bash
-lerobot-record \ # When running inference
-  --policy.path="<user>/smolVLA_finetuned" \
-  ... \
-  --dataset.rename_map='{"observation.images.glove2": "observation.images.image"}'
-```
-
-## Alternative: edit the policy config directly
-
-If you always use the same dataset or environment, you can **edit the policy's `config.json`** so its observation keys match your data source. Then no rename map is needed.
-
-The tradeoff: modifying the policy config ties it to one data source. A rename map keeps one policy usable across many datasets and environments.
-
-## Empty cameras: fewer views than the policy expects
-
-Some policies are built for a fixed number of image inputs. If your dataset has fewer cameras, you can set **`empty_cameras`** in the policy config instead of modifying the model architecture.
-
-### How it works
-
-Setting `empty_cameras=N` adds N placeholder image features to the policy config, named:
-
-```
-observation.images.empty_camera_0
-observation.images.empty_camera_1
-...
-```
-
-At runtime, these keys have no corresponding data in the batch. The policy fills them with masked dummy tensors (padded with `-1` for SigLIP-based vision encoders, with a zero attention mask), so the extra image slots are effectively ignored during training and inference.
-
-### Example
-
-XVLA-base has three visual inputs and `empty_cameras=0` by default. Your dataset only has two cameras:
-
-1. Set `--policy.empty_cameras=1`.
-2. The config adds a third key: `observation.images.empty_camera_0`.
-3. Use the rename map for your two real cameras as usual.
-4. The third slot is masked out — no fake images needed in your dataset.
-
-## Quick reference
-
-| Goal                                      | What to do                                                                  |
-| ----------------------------------------- | --------------------------------------------------------------------------- |
-| Dataset keys ≠ policy keys                | `--rename_map='{"dataset_key": "policy_key", ...}'`                         |
-| Env keys ≠ policy keys (eval)             | `--rename_map='{"env_key": "policy_key", ...}'`                             |
-| Recording with different keys (inference) | `--dataset.rename_map='{"source_key": "policy_key", ...}'`.                 |
-| Fewer cameras than policy expects         | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
-| Avoid passing a rename map                | Edit the policy's `config.json` so its keys match your data source          |

docs/source/sarm.mdx

@@ -269,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ lerobot-train \
 Train with **dense annotations only** (sparse auto-generated):
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ lerobot-train \
 Train with **both sparse and dense annotations**:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dual \
@@ -468,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
   --use_rabc=true \

docs/source/smolvla.mdx

@@ -106,9 +106,6 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
   --dataset.episode_time_s=50 \
   --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/streaming_video_encoding.mdx

@@ -1,155 +0,0 @@
-# Streaming Video Encoding Guide
-
-## 1. Overview
-
-Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
-
-1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
-
-Frames can be encoded in real-time during capture:
-
-1. Capture frame -> queue to encoder thread -> encode to MP4 directly
-
-This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
-
-## 2. Tuning Parameters
-
-| Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
-| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
-| `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
-| `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
-| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
-
-## 3. Performance Considerations
-
-Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
-
-- **Control loop** (reading cameras, control the robot, writing non-video data)
-- **Encoder threads** (one pool per camera)
-- **Rerun visualization** (if enabled)
-- **OS and other processes**
-
-### Resolution & Number of Cameras Impact
-
-| Setup                     | Throughput (px/sec) | CPU Encoding Load | Notes                          |
-| ------------------------- | ------------------- | ----------------- | ------------------------------ |
-| 2camsx 640x480x3 @30fps   | 55M                 | Low               | Works on most systems          |
-| 2camsx 1280x720x3 @30fps  | 165M                | Moderate          | Comfortable on modern systems  |
-| 2camsx 1920x1080x3 @30fps | 373M                | High              | Requires powerful high-end CPU |
-
-### `encoder_threads` Tuning
-
-This parameter controls how many threads each encoder instance uses internally:
-
-- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
-- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
-- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
-
-### Backpressure and Frame Dropping
-
-Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
-
-1. The queue fills up (consuming RAM)
-2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
-3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
-4. At episode end, total dropped frames per camera are reported
-
-### Symptoms of Encoder Falling Behind
-
-- **System feels laggy and freezes**: all CPUs are at 100%
-- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
-- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
-- **Accumulated rerun lag**: Visualization falls behind real-time
-
-## 4. Hardware-Accelerated Encoding
-
-### When to Use
-
-Use HW encoding when:
-
-- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
-- You have compatible hardware (GPU or dedicated encoder)
-- You're recording at high throughput (high resolution or with many cameras)
-
-### Choosing a Codec
-
-| Codec                 | CPU Usage | File Size      | Quality | Notes                                                            |
-| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
-| `libsvtav1` (default) | High      | Smallest       | Best    | Default. Best compression but most CPU-intensive                 |
-| `h264`                | Medium    | ~30-50% larger | Good    | Software H.264. Lower CPU                                        |
-| HW encoders           | Very Low  | Largest        | Good    | Offloads to dedicated hardware. Best for CPU-constrained systems |
-
-### Available HW Encoders
-
-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
-
-> [!NOTE]
-> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
-
-> [!NOTE]
-> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
-
-## 5. Troubleshooting
-
-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
-| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
-
-## 6. Recommended Configurations
-
-These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
-
-### High-End Systems: modern 12+ cores (24+ threads)
-
-A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
-
-```bash
-# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-lerobot-record --dataset.encoder_threads=5 ...
-
-# 3camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
-
-A throughput between ~80-300M px/sec should be possible in CPU.
-
-```bash
-# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-lerobot-record --dataset.encoder_threads=2 ...
-
-# 2camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
-
-On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
-
-```bash
-# 2camsx 640x480x3 @30fps: Requires some tuning.
-
-# Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
-```
-
-## 7. Closing note
-
-Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
-`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.

docs/source/unitree_g1.mdx

@@ -1,72 +1,23 @@
 # Unitree G1
 
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/unitree_thumbnail.jpg"
-  alt="Unitree G1 locomanipulation demo"
-  style={{ width: "100%" }}
-/>
+This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
 
-The Unitree G1 humanoid is now supported in LeRobot! You can teleoperate, train locomanipulation policies, test in sim, and more. Both 29 and 23 DoF variants are supported.
+## About
+
+We support both 29 and 23 DOF G1 EDU version. We introduce:
+
+- **`unitree g1` robot class, handling low level read/write from/to the humanoid**
+- **ZMQ socket bridge** for remote communication and camera streaming, allowing for remote policy deployment over wlan, eth or directly on the robot
+- **Locomotion policies** from NVIDIA gr00t and Amazon FAR Holosoma
+- **Simulation mode** for testing policies without the physical robot in mujoco
 
 ---
 
-## Part 1: Getting Started
+## Connection guide
 
-### Install the Unitree SDK
+### Step 1: Configure Ethernet Interface
 
-Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation). Tested with `unitree_sdk2py==1.0.1` and `cyclonedds==0.10.2`:
-
-```bash
-conda create -y -n lerobot python=3.12
-conda activate lerobot
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python
-pip install -e .
-cd ..
-```
-
-### Install LeRobot
-
-```bash
-conda install ffmpeg -c conda-forge
-conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-```
-
-<Tip>
-  For now, pinocchio must be installed from conda-forge (not pip) to include the
-  CasADi bindings needed for arm IK.
-</Tip>
-
-### Test the Installation (Simulation)
-
-The simulation environment has its own dependencies. Check the Simulation environment dependencies: [Unitree G1 Mujoco EnvHub](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main).
-
-```bash
-pip install mujoco loguru msgpack msgpack-numpy
-```
-
-```bash
-lerobot-teleoperate \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=true \
-  --teleop.type=unitree_g1 \
-  --teleop.id=wbc_unitree \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 5}}' \
-  --display_data=true \
-  --robot.controller=GrootLocomotionController
-```
-
-This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) via `--robot.controller`.
-
-- Press `9` to release the robot
-- Press `7` / `8` to increase / decrease waist height
-
-### Connect to the Physical Robot
-
-The G1's Ethernet IP is fixed at `192.168.123.164`. Your machine must have a static IP on the same subnet: `192.168.123.x` where `x ≠ 164`.
+Set a static IP on the same subnet as the robot:
 
 ```bash
 # Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
@@ -75,23 +26,47 @@ sudo ip addr add 192.168.123.200/24 dev enp131s0
 sudo ip link set enp131s0 up
 ```
 
-### SSH into the Robot
+**Note**: The G1's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` with x ≠ 164.
+
+### Step 2: SSH into the Robot
 
 ```bash
 ssh unitree@192.168.123.164
 # Password: 123
 ```
 
-### Share Internet via Ethernet
+You should now be connected to the G1's Orin.
 
-The G1 needs internet access to clone repos and install packages. Share your laptop's connection over Ethernet:
+---
+
+## Part 2: Enable WiFi on the Robot
+
+Wlan0 is disabled by default on the G1. To enable it:
+
+### Step 1: Enable WiFi Hardware
+
+```bash
+sudo rfkill unblock wifi
+sudo rfkill unblock all
+
+# Bring up wlan0
+sudo ip link set wlan0 up
+
+# Enable NetworkManager control of wlan0
+sudo nmcli radio wifi on
+sudo nmcli device set wlan0 managed yes
+sudo systemctl restart NetworkManager
+```
+
+### Step 2: Enable Internet Forwarding
 
 **On your laptop:**
 
 ```bash
+# Enable IP forwarding
 sudo sysctl -w net.ipv4.ip_forward=1
 
-# Replace wlp132s0f0 with your WiFi interface name
+# Set up NAT (replace wlp132s0f0 with your WiFi interface)
 sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
 sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
 sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
@@ -100,193 +75,217 @@ sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
 **On the G1:**
 
 ```bash
+# Add laptop as default gateway
 sudo ip route del default 2>/dev/null || true
 sudo ip route add default via 192.168.123.200 dev eth0
 echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
 
-# Verify
+# Test connection
 ping -c 3 8.8.8.8
 ```
 
-### Install the Unitree SDK on the G1
-
-Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation):
-
-```bash
-conda create -y -n lerobot python=3.12
-conda activate lerobot
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python
-python -m pip install -e .
-cd ..
-```
-
-### Install LeRobot on the G1
-
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
-python -m pip install -e '.[unitree_g1]'
-```
-
-<Tip>
-  For now, pinocchio must be installed from conda-forge (not pip) to include the
-  CasADi bindings needed for arm IK.
-</Tip>
-
-### (Optional) Enable WiFi on the Robot
-
-For wireless SSH access, you can enable WiFi on the G1 (it's blocked by default):
-
-```bash
-sudo rfkill unblock all
-sudo ip link set wlan0 up
-sudo nmcli radio wifi on
-sudo nmcli device set wlan0 managed yes
-sudo systemctl restart NetworkManager
-```
-
-**Connect to a WiFi network:**
+### Step 3: Connect to WiFi Network
 
 ```bash
+# List available networks
 nmcli device wifi list
 
+# Connect to your WiFi (example)
 sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
 sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
 sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
 sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
 sudo nmcli connection up "YourNetwork"
 
+# Check WiFi IP address
 ip a show wlan0
 ```
 
-You can then SSH over WiFi instead of Ethernet:
+### Step 4: SSH Over WiFi
+
+Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:
 
 ```bash
-ssh unitree@<ROBOT_WIFI_IP>
+ssh unitree@<YOUR_ROBOT_IP>
 # Password: 123
 ```
 
----
-
-## Part 2: Teleoperation & Locomotion
-
-### Run the Robot Server
-
-On the robot (from `~/lerobot`):
-
-```bash
-cd ~/lerobot
-python src/lerobot/robots/unitree_g1/run_g1_server.py --camera
-```
-
-### Run the Locomotion Policy
-
-You can run the teleoperation client from your laptop over Ethernet, over WiFi (experimental), or directly on the robot itself. Mind potential latency introduced by your network.
-
-**From your laptop:**
-
-```bash
-lerobot-teleoperate \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.robot_ip=<ROBOT_IP> \
-  --teleop.type=unitree_g1 \
-  --teleop.id=wbc_unitree \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --display_data=true \
-  --robot.controller=HolosomaLocomotionController
-```
-
-We support both [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) and [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) via `--robot.controller`.
+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address.
 
 ---
 
-## Part 3: Loco-Manipulation with the Homunculus Exoskeleton
+## Part 3: Robot Server Setup
 
-We provide a loco-manipulation solution via the Homunculus Exoskeleton — an open-source 7 DoF exoskeleton for whole-body control. Check it out [here](https://github.com/nepyope/hmc_exo).
+### Step 1: Install LeRobot on the Orin
 
-### Calibrate
+SSH into the robot and install LeRobot:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+
+### Step 2: Run the Robot Server
+
+On the robot:
+
+```bash
+python src/lerobot/robots/unitree_g1/run_g1_server.py
+```
+
+**Important**: Keep this terminal running. The server must be active for remote control.
+
+---
+
+## Part 4: Controlling the robot
+
+With the robot server running, you can now control the robot remotely. Let's launch a locomotion policy
+
+### Step 1: Install LeRobot on your machine
+
+```bash
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+### Step 2: Update Robot IP in Config
+
+Edit the config file to match your robot's WiFi IP:
+
+```python
+# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
+robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
+```
+
+### Step 3: Run the Locomotion Policy
+
+```bash
+# Run GR00T locomotion controller
+python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
+
+# Run Holosoma locomotion controller
+python examples/unitree_g1/holosoma_locomotion.py
+
+```
+
+Press `Ctrl+C` to stop the policy.
+
+---
+
+## Running in Simulation Mode (MuJoCo)
+
+You can test policies before deploying on the physical robot using MuJoCo simulation. Set `is_simulation=True` in config or pass `--robot.is_simulation=true` via CLI.
+
+### Calibrate Exoskeleton Teleoperator
 
 ```bash
 lerobot-calibrate \
-  --teleop.type=unitree_g1 \
-  --teleop.left_arm_config.port=/dev/ttyACM1 \
-  --teleop.right_arm_config.port=/dev/ttyACM0 \
-  --teleop.id=exo
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo
 ```
 
-During calibration move each joint through its entire range. After fitting, move the joint in a neutral position and press `n` to advance.
-
-### Record a Dataset
+### Teleoperate in Simulation
 
 ```bash
-lerobot-record \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=true \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --teleop.type=unitree_g1 \
-  --teleop.left_arm_config.port=/dev/ttyACM1 \
-  --teleop.right_arm_config.port=/dev/ttyACM0 \
-  --teleop.id=exo \
-  --dataset.repo_id=your-username/dataset-name \
-  --dataset.single_task="Test" \
-  --dataset.num_episodes=2 \
-  --dataset.episode_time_s=5 \
-  --dataset.reset_time_s=5 \
-  --dataset.push_to_hub=true \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
 ```
 
-> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
+### Record Dataset in Simulation
 
-Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)
+```bash
+python -m lerobot.scripts.lerobot_record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true
+```
+
+Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
 
 ---
 
-## Part 4: Training & Inference
+## Running on Real Robot
 
-### Train
+Once the robot server is running on the G1 (see Part 3), you can teleoperate and record on the real robot.
+
+### Start the Camera Server
+
+On the robot, start the ZMQ image server:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
-  --dataset.repo_id=your-username/dataset-name  \
-  --policy.type=pi05 \
-  --output_dir=./outputs/pi05_training \
-  --job_name=pi05_training \
-  --policy.repo_id=your-username/your-repo-id \
-  --policy.pretrained_path=lerobot/pi05_base \
-  --policy.compile_model=true \
-  --policy.gradient_checkpointing=true \
-  --wandb.enable=true \
-  --policy.dtype=bfloat16 \
-  --policy.freeze_vision_encoder=false \
-  --policy.train_expert_only=false \
-  --steps=3000 \
-  --policy.device=cuda \
-  --batch_size=32
+python src/lerobot/cameras/zmq/image_server.py
 ```
 
-### Inference with RTC
+Keep this running in a separate terminal for camera streaming during recording.
 
-Once trained, we recommend deploying policies using inference-time RTC:
+### Teleoperate Real Robot
 
 ```bash
-python examples/rtc/eval_with_real_robot.py \
-  --policy.path=your-username/your-repo-id \
-  --policy.device=cuda \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.controller=HolosomaLocomotionController \
-  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --task="task_description" \
-  --duration=1000 \
-  --fps=30 \
-  --rtc.enabled=true
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
 ```
 
+### Record Dataset on Real Robot
+
+```bash
+python -m lerobot.scripts.lerobot_record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true
+```
+
+**Note**: Update `server_address` to match your robot's camera server IP.
+
+Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
+
 ---
 
 ## Additional Resources
@@ -295,8 +294,8 @@ python examples/rtc/eval_with_real_robot.py \
 - [GR00T-WholeBodyControl](https://github.com/NVlabs/GR00T-WholeBodyControl)
 - [Holosoma](https://github.com/amazon-far/holosoma)
 - [LeRobot Documentation](https://github.com/huggingface/lerobot)
-- [Unitree IL LeRobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
+- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
 
 ---
 
-_Last updated: March 2026_
+_Last updated: December 2025_

docs/source/using_dataset_tools.mdx

@@ -12,7 +12,6 @@ LeRobot provides several utilities for manipulating datasets:
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
 6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
-7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -157,30 +156,6 @@ lerobot-edit-dataset \
 
 **Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
 
-### Show the information of datasets
-
-Show the information of datasets such as number of episode, number of frame, File size and so on.
-No change will be made to the dataset
-
-```bash
-
-# Show dataset information without feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-
-# Show dataset information with feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-    --operation.show_features true
-
-```
-
-**Parameters:**
-
-- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
-
 ### Push to Hub
 
 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:

docs/source/walloss.mdx

@@ -45,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=wall_x \
     --output_dir=./outputs/wallx_training \

docs/source/xvla.mdx

@@ -154,7 +154,7 @@ lerobot-train \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
   --output_dir=./outputs/xvla_bimanual \
   --job_name=xvla_so101_training \
   --policy.path="lerobot/xvla-base" \

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=<USER>/record-test \
+    --dataset.repo_id=aliberts/record-test \
     --dataset.episode=2
 ```
 """
@@ -57,7 +57,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

examples/dataset/load_lerobot_dataset.py

@@ -32,8 +32,7 @@ import torch
 from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 
 
 def main():

examples/dataset/slurm_compute_rabc.py

@@ -1,490 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 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.
-
-"""
-SLURM-distributed SARM RA-BC annotation pipeline.
-
-Computes SARM progress values for all frames in a dataset, distributed across
-SLURM workers, then merges the shards into a single sarm_progress.parquet.
-
-Two subcommands, each a separate SLURM submission:
-
-  compute    – N workers, each computes progress for a subset of episodes
-  aggregate  – 1 worker, merges N shards into sarm_progress.parquet, pushes to hub
-
-Usage:
-    python slurm_compute_rabc.py compute \\
-        --repo-id user/dataset --reward-model-path user/sarm_model \\
-        --stride 10 --device cpu --workers 50 --partition cpu
-
-    python slurm_compute_rabc.py aggregate \\
-        --repo-id user/dataset --reward-model-path user/sarm_model \\
-        --partition cpu --push-to-hub
-"""
-
-import argparse
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-
-
-class ComputeProgressShards(PipelineStep):
-    """Each worker computes SARM progress for its assigned episodes."""
-
-    def __init__(
-        self, repo_id, reward_model_path, stride=1, head_mode="sparse", device="cpu", shard_dir="rabc_shards"
-    ):
-        super().__init__()
-        if stride < 1:
-            raise ValueError(f"stride must be >= 1, got {stride}")
-        self.repo_id = repo_id
-        self.reward_model_path = reward_model_path
-        self.stride = stride
-        self.head_mode = head_mode
-        self.device = device
-        self.shard_dir = shard_dir
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import logging
-        from pathlib import Path
-
-        import numpy as np
-        import pyarrow as pa
-        import pyarrow.parquet as pq
-        import torch
-        from tqdm import tqdm
-
-        from lerobot.policies.sarm.compute_rabc_weights import (
-            generate_all_frame_indices,
-            interpolate_progress,
-            load_sarm_resources,
-        )
-        from lerobot.utils.utils import init_logging
-
-        init_logging()
-
-        dataset, reward_model, preprocess = load_sarm_resources(
-            self.repo_id,
-            self.reward_model_path,
-            self.device,
-        )
-
-        if hasattr(preprocess, "eval"):
-            preprocess.eval()
-        for step in preprocess.steps:
-            if hasattr(step, "eval"):
-                step.eval()
-
-        image_key = reward_model.config.image_key
-        state_key = reward_model.config.state_key
-        frame_gap = reward_model.config.frame_gap
-        center_idx = reward_model.config.n_obs_steps // 2
-
-        dual_mode = reward_model.config.uses_dual_heads
-        compute_sparse = self.head_mode in ("sparse", "both") or not dual_mode
-        compute_dense = self.head_mode in ("dense", "both") and dual_mode
-
-        my_episodes = list(range(dataset.num_episodes))[rank::world_size]
-        if not my_episodes:
-            logging.info(f"Rank {rank}: no episodes assigned")
-            return
-        logging.info(f"Rank {rank}: {len(my_episodes)} / {dataset.num_episodes} episodes")
-
-        all_rows = []
-
-        for ep_idx in tqdm(my_episodes, desc=f"Rank {rank}"):
-            ep = dataset.meta.episodes[ep_idx]
-            ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
-            task = dataset[ep_start].get("task", "perform the task")
-
-            all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
-            if self.stride > 1:
-                compute_indices = [i for i in all_ep_indices if (i - ep_start) % self.stride == 0]
-                if (ep_end - 1) not in compute_indices:
-                    compute_indices.append(ep_end - 1)
-                compute_indices = sorted(set(compute_indices))
-            else:
-                compute_indices = all_ep_indices
-
-            frame_results = {}
-            for qi in tqdm(compute_indices, desc=f"  Ep {ep_idx}", leave=False):
-                try:
-                    sample = dataset[qi]
-                    batch = {
-                        image_key: sample[image_key],
-                        "task": task,
-                        "index": qi,
-                        "episode_index": ep_idx,
-                    }
-                    if state_key in sample:
-                        batch[state_key] = sample[state_key]
-
-                    with torch.no_grad():
-                        processed = preprocess(batch)
-                        vf = processed["video_features"].to(self.device)
-                        tf = processed["text_features"].to(self.device)
-                        sf = processed.get("state_features")
-                        if sf is not None:
-                            sf = sf.to(self.device)
-                        lengths = processed.get("lengths")
-
-                        sparse_val = dense_val = np.nan
-                        if compute_sparse:
-                            r = reward_model.calculate_rewards(
-                                text_embeddings=tf,
-                                video_embeddings=vf,
-                                state_features=sf,
-                                lengths=lengths,
-                                return_all_frames=True,
-                                head_mode="sparse",
-                            )
-                            sparse_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
-                        if compute_dense:
-                            r = reward_model.calculate_rewards(
-                                text_embeddings=tf,
-                                video_embeddings=vf,
-                                state_features=sf,
-                                lengths=lengths,
-                                return_all_frames=True,
-                                head_mode="dense",
-                            )
-                            dense_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
-
-                        frame_results[qi] = (sparse_val, dense_val)
-                except Exception as e:
-                    logging.warning(f"Failed frame {qi}: {e}")
-
-            if not frame_results:
-                logging.warning(f"Episode {ep_idx}: all frames failed, skipping")
-                continue
-
-            # Interpolate to all frames in this episode
-            computed_idx = np.array(sorted(frame_results.keys()))
-            all_frame_arr = np.arange(ep_start, ep_end)
-
-            sparse_vals = np.array([frame_results[i][0] for i in computed_idx]) if compute_sparse else None
-            dense_vals = np.array([frame_results[i][1] for i in computed_idx]) if compute_dense else None
-
-            if self.stride > 1 and len(computed_idx) > 1:
-                if compute_sparse:
-                    sparse_vals = interpolate_progress(computed_idx, sparse_vals, all_frame_arr)
-                if compute_dense:
-                    dense_vals = interpolate_progress(computed_idx, dense_vals, all_frame_arr)
-                output_frames = all_frame_arr
-            else:
-                # Use only successfully computed frames to avoid indexing mismatch on failures
-                output_frames = computed_idx
-
-            for i, fi in enumerate(output_frames):
-                row = {"index": int(fi), "episode_index": ep_idx, "frame_index": int(fi - ep_start)}
-                if compute_sparse:
-                    row["progress_sparse"] = float(sparse_vals[i])
-                if compute_dense:
-                    row["progress_dense"] = float(dense_vals[i])
-                all_rows.append(row)
-
-        if all_rows:
-            import pandas as pd
-
-            df = pd.DataFrame(all_rows).sort_values("index").reset_index(drop=True)
-            table = pa.Table.from_pandas(df, preserve_index=False)
-            table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
-            shard_dir = Path(self.shard_dir)
-            shard_dir.mkdir(parents=True, exist_ok=True)
-            out = shard_dir / f"shard_{rank:05d}.parquet"
-            pq.write_table(table, out)
-            logging.info(f"Rank {rank}: saved {len(df)} rows to {out}")
-
-
-class AggregateProgress(PipelineStep):
-    """Merge all shard parquets into final sarm_progress.parquet."""
-
-    def __init__(self, repo_id, reward_model_path, shard_dir="rabc_shards", push_to_hub=False):
-        super().__init__()
-        self.repo_id = repo_id
-        self.reward_model_path = reward_model_path
-        self.shard_dir = shard_dir
-        self.push_to_hub = push_to_hub
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import datetime
-        import logging
-        import os
-        from pathlib import Path
-
-        import pandas as pd
-        import pyarrow as pa
-        import pyarrow.parquet as pq
-
-        from lerobot.datasets.lerobot_dataset import LeRobotDataset
-        from lerobot.utils.utils import init_logging
-
-        init_logging()
-        if rank != 0:
-            return
-
-        shard_dir = Path(self.shard_dir)
-        shards = sorted(shard_dir.glob("shard_*.parquet"))
-        if not shards:
-            raise FileNotFoundError(f"No shards found in {shard_dir}")
-
-        # Log shard modification time range to help detect stale files
-        mtimes = [os.path.getmtime(s) for s in shards]
-        oldest = datetime.datetime.fromtimestamp(min(mtimes)).isoformat(timespec="seconds")
-        newest = datetime.datetime.fromtimestamp(max(mtimes)).isoformat(timespec="seconds")
-        logging.info(f"Aggregating {len(shards)} shards (oldest: {oldest}, newest: {newest})")
-
-        df = pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True)
-        df = df.sort_values("index").reset_index(drop=True)
-
-        table = pa.Table.from_pandas(df, preserve_index=False)
-        table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
-
-        temp_ds = LeRobotDataset(self.repo_id, download_videos=False)
-        out_path = Path(temp_ds.root) / "sarm_progress.parquet"
-        out_path.parent.mkdir(parents=True, exist_ok=True)
-        pq.write_table(table, out_path)
-        logging.info(f"Saved {len(df)} rows to {out_path}")
-
-        for col in ["progress_sparse", "progress_dense"]:
-            if col in df.columns:
-                v = df[col].dropna()
-                logging.info(
-                    f"{col}: mean={v.mean():.4f} std={v.std():.4f} min={v.min():.4f} max={v.max():.4f}"
-                )
-
-        if self.push_to_hub:
-            from huggingface_hub import HfApi
-
-            api = HfApi()
-            hub_path = "sarm_progress.parquet"
-            logging.info(f"Uploading to {self.repo_id}/{hub_path}")
-            api.upload_file(
-                path_or_fileobj=str(out_path),
-                path_in_repo=hub_path,
-                repo_id=self.repo_id,
-                repo_type="dataset",
-            )
-            logging.info(f"Uploaded: https://huggingface.co/datasets/{self.repo_id}/blob/main/{hub_path}")
-
-
-def make_compute_executor(
-    repo_id,
-    reward_model_path,
-    stride,
-    head_mode,
-    device,
-    shard_dir,
-    logs_dir,
-    job_name,
-    slurm,
-    workers,
-    partition,
-    cpus_per_task,
-    mem_per_cpu,
-):
-    kwargs = {
-        "pipeline": [
-            ComputeProgressShards(repo_id, reward_model_path, stride, head_mode, device, str(shard_dir)),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": workers,
-                "workers": workers,
-                "time": "24:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        return SlurmPipelineExecutor(**kwargs)
-
-    kwargs.update({"tasks": workers, "workers": 1})
-    return LocalPipelineExecutor(**kwargs)
-
-
-def make_aggregate_executor(
-    repo_id,
-    reward_model_path,
-    shard_dir,
-    logs_dir,
-    job_name,
-    slurm,
-    partition,
-    cpus_per_task,
-    mem_per_cpu,
-    push_to_hub,
-):
-    kwargs = {
-        "pipeline": [
-            AggregateProgress(repo_id, reward_model_path, str(shard_dir), push_to_hub),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": 1,
-                "workers": 1,
-                "time": "02:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        return SlurmPipelineExecutor(**kwargs)
-
-    kwargs.update({"tasks": 1, "workers": 1})
-    return LocalPipelineExecutor(**kwargs)
-
-
-def _add_shared_args(p):
-    p.add_argument(
-        "--repo-id",
-        type=str,
-        required=True,
-        help="Hugging Face repository identifier, e.g. 'user/dataset'.",
-    )
-    p.add_argument(
-        "--shard-dir",
-        type=Path,
-        default=Path("rabc_shards"),
-        help="Directory to read/write per-rank parquet shards.",
-    )
-    p.add_argument(
-        "--logs-dir",
-        type=Path,
-        default=Path("logs"),
-        help="Directory for datatrove logs.",
-    )
-    p.add_argument(
-        "--job-name",
-        type=str,
-        default=None,
-        help="SLURM job name (defaults to rabc_<subcommand>).",
-    )
-    p.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="1 = submit via SLURM; 0 = run locally (useful for debugging).",
-    )
-    p.add_argument(
-        "--partition",
-        type=str,
-        default=None,
-        help="SLURM partition to submit to.",
-    )
-    p.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=4,
-        help="Number of CPUs per SLURM task.",
-    )
-    p.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="4G",
-        help="Memory per CPU, e.g. '4G' or '1950M'.",
-    )
-
-
-def main():
-    parser = argparse.ArgumentParser(
-        description="SLURM-distributed SARM RA-BC annotation pipeline",
-        formatter_class=argparse.RawDescriptionHelpFormatter,
-    )
-    sub = parser.add_subparsers(dest="command", required=True)
-
-    # compute subcommand
-    cp = sub.add_parser(
-        "compute",
-        help="Distribute progress computation across SLURM workers.",
-    )
-    _add_shared_args(cp)
-    cp.add_argument(
-        "--reward-model-path",
-        type=str,
-        required=True,
-        help="Path or HF repo id of the SARM reward model.",
-    )
-    cp.add_argument(
-        "--stride",
-        type=int,
-        default=1,
-        help="Compute every Nth frame; intermediate frames are interpolated (must be >= 1).",
-    )
-    cp.add_argument(
-        "--head-mode",
-        type=str,
-        default="sparse",
-        choices=["sparse", "dense", "both"],
-        help="Which reward head(s) to compute.",
-    )
-    cp.add_argument(
-        "--device",
-        type=str,
-        default="cpu",
-        help="Device for reward model inference, e.g. 'cpu' or 'cuda'.",
-    )
-    cp.add_argument(
-        "--workers",
-        type=int,
-        default=50,
-        help="Number of parallel SLURM tasks (one shard per worker).",
-    )
-
-    # aggregate subcommand
-    ap = sub.add_parser(
-        "aggregate",
-        help="Merge per-rank shards into a single sarm_progress.parquet.",
-    )
-    _add_shared_args(ap)
-    ap.add_argument(
-        "--reward-model-path",
-        type=str,
-        required=True,
-        help="Path or HF repo id of the SARM reward model (stored in parquet metadata).",
-    )
-    ap.add_argument(
-        "--push-to-hub",
-        action="store_true",
-        help="Upload sarm_progress.parquet to the Hugging Face Hub after aggregation.",
-    )
-
-    args = parser.parse_args()
-    job_name = args.job_name or f"rabc_{args.command}"
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    kwargs["job_name"] = job_name
-    command = kwargs.pop("command")
-
-    executor = make_compute_executor(**kwargs) if command == "compute" else make_aggregate_executor(**kwargs)
-
-    executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/lekiwi/evaluate.py

@@ -14,8 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.processor import make_default_processors

examples/lekiwi/record.py

@@ -14,8 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient

examples/phone_to_so100/evaluate.py

@@ -16,13 +16,15 @@
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.processor import (
+    RobotAction,
+    RobotObservation,
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
 )
@@ -38,7 +40,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/phone_to_so100/record.py

@@ -15,11 +15,11 @@
 # limitations under the License.
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
 from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
@@ -38,7 +38,6 @@ from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/phone_to_so100/replay.py

@@ -18,7 +18,7 @@ import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
@@ -27,7 +27,6 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/phone_to_so100/teleoperate.py

@@ -16,7 +16,7 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
@@ -31,7 +31,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/port_datasets/port_droid.py

@@ -22,8 +22,7 @@ from pathlib import Path
 import numpy as np
 import tensorflow_datasets as tfds
 
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
 
 DROID_SHARDS = 2048

examples/port_datasets/slurm_upload.py

@@ -26,7 +26,7 @@ from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
 from port_droid import DROID_SHARDS
 
-from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import create_lerobot_dataset_card
 from lerobot.utils.utils import init_logging
 
@@ -155,7 +155,7 @@ class UploadDataset(PipelineStep):
         from datasets.utils.tqdm import disable_progress_bars
         from huggingface_hub import CommitOperationAdd, preupload_lfs_files
 
-        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+        from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/rtc/eval_dataset.py

@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
     # Basic usage with smolvla policy
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/reuben_pi0 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --policy.path=lipsop/reuben_pi0 \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
         --rtc.execution_horizon=8 \
         --device=cuda
 
     # With torch.compile for faster inference (PyTorch 2.0+)
     # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \
@@ -113,9 +113,8 @@ from lerobot.configs import parser
 from lerobot.configs.default import DatasetConfig
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
 from lerobot.datasets.factory import resolve_delta_timestamps
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
     # Run RTC with Real robot with RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
+        --policy.path=helper2424/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -78,11 +78,10 @@ from torch import Tensor
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.cameras.zmq.configuration_zmq import ZMQCameraConfig  # noqa: F401
 from lerobot.configs import parser
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
 from lerobot.policies.rtc.action_queue import ActionQueue
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
@@ -98,7 +97,6 @@ from lerobot.robots import (  # noqa: F401
     bi_so_follower,
     koch_follower,
     so_follower,
-    unitree_g1,
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES

examples/so100_to_so100_EE/evaluate.py

@@ -16,13 +16,15 @@
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.processor import (
+    RobotAction,
+    RobotObservation,
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
 )
@@ -38,7 +40,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.scripts.lerobot_record import record_loop
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/so100_to_so100_EE/record.py

@@ -16,11 +16,11 @@
 
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
 from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
@@ -35,7 +35,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/so100_to_so100_EE/replay.py

@@ -19,7 +19,7 @@ import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
@@ -28,7 +28,6 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/so100_to_so100_EE/teleoperate.py

@@ -17,7 +17,7 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotProcessorPipeline
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     robot_action_to_transition,
@@ -30,7 +30,6 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
-from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/training/train_policy.py

@@ -19,9 +19,8 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
 from lerobot.policies.factory import make_pre_post_processors

examples/training/train_with_streaming.py

@@ -20,9 +20,9 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
+from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors

examples/tutorial/act/act_training_example.py

@@ -5,9 +5,8 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors

examples/tutorial/act/act_using_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.utils import build_inference_frame, make_robot_action

examples/tutorial/diffusion/diffusion_training_example.py

@@ -5,9 +5,8 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import FeatureType
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import dataset_to_policy_features
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
 from lerobot.policies.factory import make_pre_post_processors

examples/tutorial/diffusion/diffusion_using_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.utils import build_inference_frame, make_robot_action

examples/tutorial/pi0/using_pi0_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import hw_to_dataset_features
+from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.pi0.modeling_pi0 import PI0Policy
 from lerobot.policies.utils import build_inference_frame, make_robot_action

examples/tutorial/rl/hilserl_example.py

@@ -6,8 +6,8 @@ from queue import Empty, Full
 import torch
 import torch.optim as optim
 
-from lerobot.datasets.feature_utils import hw_to_dataset_features
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy

examples/tutorial/smolvla/using_smolvla_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.feature_utils import hw_to_dataset_features
+from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 from lerobot.policies.utils import build_inference_frame, make_robot_action

examples/unitree_g1/gr00t_locomotion.py

@@ -14,20 +14,20 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import argparse
 import logging
+import time
 from collections import deque
 
 import numpy as np
 import onnxruntime as ort
 from huggingface_hub import hf_hub_download
 
-from lerobot.robots.unitree_g1.g1_utils import (
-    REMOTE_AXES,
-    REMOTE_BUTTONS,
-    G1_29_JointIndex,
-    get_gravity_orientation,
-)
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
 
+logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
 
@@ -36,13 +36,18 @@ GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # Hip pitch
 GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # Knee
 GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # Ankle pitch
 
+MISSING_JOINTS = []
+G1_MODEL = "g1_23"  # Or "g1_29"
+if G1_MODEL == "g1_23":
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # Waist yaw/pitch, wrist pitch/yaw
+
 # Control parameters
 ACTION_SCALE = 0.25
 CONTROL_DT = 0.02  # 50Hz
 ANG_VEL_SCALE: float = 0.25
 DOF_POS_SCALE: float = 1.0
 DOF_VEL_SCALE: float = 0.05
-CMD_SCALE: list[float] = [2.0, 2.0, 0.25]
+CMD_SCALE: list = [2.0, 2.0, 0.25]
 
 
 DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
@@ -80,11 +85,11 @@ def load_groot_policies(
 class GrootLocomotionController:
     """GR00T lower-body locomotion controller for the Unitree G1."""
 
-    control_dt = CONTROL_DT  # Expose for unitree_g1.py
-
-    def __init__(self):
-        # Load policies
-        self.policy_balance, self.policy_walk = load_groot_policies()
+    def __init__(self, policy_balance, policy_walk, robot, config):
+        self.policy_balance = policy_balance
+        self.policy_walk = policy_walk
+        self.robot = robot
+        self.config = config
 
         self.cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
 
@@ -104,60 +109,45 @@ class GrootLocomotionController:
 
         logger.info("GrootLocomotionController initialized")
 
-    def reset(self) -> None:
-        """Reset internal state for a new episode."""
-        self.cmd[:] = 0.0
-        self.groot_qj_all[:] = 0.0
-        self.groot_dqj_all[:] = 0.0
-        self.groot_action[:] = 0.0
-        self.groot_obs_single[:] = 0.0
-        self.groot_obs_stacked[:] = 0.0
-        self.groot_height_cmd = 0.74
-        self.groot_orientation_cmd[:] = 0.0
-        self.groot_obs_history.clear()
-        for _ in range(6):
-            self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
+    def run_step(self):
+        # Get current observation
+        obs = self.robot.get_observation()
 
-    def run_step(self, action: dict, lowstate) -> dict:
-        """Run one step of the locomotion controller.
+        if not obs:
+            return
 
-        Args:
-            action: Action dict containing remote.lx/ly/rx/ry and buttons
-            lowstate: Robot lowstate containing motor positions/velocities and IMU
-
-        Returns:
-            Action dict for lower body joints (0-14)
-        """
-        if lowstate is None:
-            return {}
-
-        buttons = [int(action.get(k, 0)) for k in REMOTE_BUTTONS]
-        if buttons[0]:  # R1 - raise waist
+        # Get command from remote controller
+        if obs["remote.buttons"][0]:  # R1 - raise waist
             self.groot_height_cmd += 0.001
             self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
-        if buttons[4]:  # R2 - lower waist
+        if obs["remote.buttons"][4]:  # R2 - lower waist
             self.groot_height_cmd -= 0.001
             self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
 
-        lx, ly, rx, _ry = (action.get(k, 0.0) for k in REMOTE_AXES)
-        self.cmd[0] = ly  # Forward/backward
-        self.cmd[1] = -lx  # Left/right (negated)
-        self.cmd[2] = -rx  # Rotation rate (negated)
+        self.cmd[0] = obs["remote.ly"]  # Forward/backward
+        self.cmd[1] = obs["remote.lx"] * -1  # Left/right
+        self.cmd[2] = obs["remote.rx"] * -1  # Rotation rate
 
-        # Get joint positions and velocities from lowstate
+        # Get joint positions and velocities from flat dict
         for motor in G1_29_JointIndex:
+            name = motor.name
             idx = motor.value
-            self.groot_qj_all[idx] = lowstate.motor_state[idx].q
-            self.groot_dqj_all[idx] = lowstate.motor_state[idx].dq
+            self.groot_qj_all[idx] = obs[f"{name}.q"]
+            self.groot_dqj_all[idx] = obs[f"{name}.dq"]
+
+        # Adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.groot_qj_all[idx] = 0.0
+            self.groot_dqj_all[idx] = 0.0
 
         # Scale joint positions and velocities
         qj_obs = self.groot_qj_all.copy()
         dqj_obs = self.groot_dqj_all.copy()
 
         # Express IMU data in gravity frame of reference
-        quat = lowstate.imu_state.quaternion
-        ang_vel = np.array(lowstate.imu_state.gyroscope, dtype=np.float32)
-        gravity_orientation = get_gravity_orientation(quat)
+        quat = [obs["imu.quat.w"], obs["imu.quat.x"], obs["imu.quat.y"], obs["imu.quat.z"]]
+        ang_vel = np.array([obs["imu.gyro.x"], obs["imu.gyro.y"], obs["imu.gyro.z"]], dtype=np.float32)
+        gravity_orientation = self.robot.get_gravity_orientation(quat)
 
         # Scale joint positions and velocities before policy inference
         qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
@@ -196,10 +186,73 @@ class GrootLocomotionController:
         # Transform action back to target joint positions
         target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * ACTION_SCALE
 
-        # Build action dict
+        # Build action dict (only first 15 joints for GR00T)
         action_dict = {}
         for i in range(15):
             motor_name = G1_29_JointIndex(i).name
             action_dict[f"{motor_name}.q"] = float(target_dof_pos_15[i])
 
-        return action_dict
+        # Zero out missing joints for g1_23dof
+        for joint_idx in MISSING_JOINTS:
+            motor_name = G1_29_JointIndex(joint_idx).name
+            action_dict[f"{motor_name}.q"] = 0.0
+
+        # Send action to robot
+        self.robot.send_action(action_dict)
+
+
+def run(repo_id: str = DEFAULT_GROOT_REPO_ID) -> None:
+    """Main function to run the GR00T locomotion controller.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID for GR00T policies.
+    """
+    # Load policies
+    policy_balance, policy_walk = load_groot_policies(repo_id=repo_id)
+
+    # Initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    robot.connect()
+
+    # Initialize gr00T locomotion controller
+    groot_controller = GrootLocomotionController(
+        policy_balance=policy_balance,
+        policy_walk=policy_walk,
+        robot=robot,
+        config=config,
+    )
+
+    try:
+        robot.reset(CONTROL_DT, GROOT_DEFAULT_ANGLES)
+
+        logger.info("Use joystick: LY=fwd/back, LX=left/right, RX=rotate, R1=raise waist, R2=lower waist")
+        logger.info("Press Ctrl+C to stop")
+
+        # Run step
+        while not robot._shutdown_event.is_set():
+            start_time = time.time()
+            groot_controller.run_step()
+            elapsed = time.time() - start_time
+            sleep_time = max(0, CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+    except KeyboardInterrupt:
+        logger.info("Stopping locomotion...")
+    finally:
+        if robot.is_connected:
+            robot.disconnect()
+        logger.info("Done!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="GR00T Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_GROOT_REPO_ID,
+        help=f"Hugging Face Hub repo ID for GR00T policies (default: {DEFAULT_GROOT_REPO_ID})",
+    )
+    args = parser.parse_args()
+
+    run(repo_id=args.repo_id)

examples/unitree_g1/holosoma_locomotion.py

@@ -14,21 +14,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import argparse
 import json
 import logging
+import time
 
 import numpy as np
 import onnx
 import onnxruntime as ort
 from huggingface_hub import hf_hub_download
 
-from lerobot.robots.unitree_g1.g1_utils import (
-    REMOTE_AXES,
-    G1_29_JointArmIndex,
-    G1_29_JointIndex,
-    get_gravity_orientation,
-)
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
 
+logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
 DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
@@ -40,13 +40,18 @@ DEFAULT_ANGLES[16] = 0.2  # Left shoulder roll
 DEFAULT_ANGLES[23] = -0.2  # Right shoulder roll
 DEFAULT_ANGLES[[18, 25]] = 0.6  # Elbow
 
+MISSING_JOINTS = []
+G1_MODEL = "g1_23"  # Or "g1_29"
+if G1_MODEL == "g1_23":
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # Waist yaw/pitch, wrist pitch/yaw
+
 # Control parameters
 ACTION_SCALE = 0.25
-CONTROL_DT = 0.005  # 200Hz
+CONTROL_DT = 0.02  # 50Hz
 ANG_VEL_SCALE = 0.25
 DOF_POS_SCALE = 1.0
 DOF_VEL_SCALE = 0.05
-GAIT_PERIOD = 0.5
+GAIT_PERIOD = 1.0
 
 
 DEFAULT_HOLOSOMA_REPO_ID = "nepyope/holosoma_locomotion"
@@ -82,7 +87,7 @@ def load_policy(
     logger.info(f"Policy loaded: {policy.get_inputs()[0].shape} → {policy.get_outputs()[0].shape}")
 
     # Extract KP/KD from ONNX metadata
-    model = onnx.load(policy_path, load_external_data=False)
+    model = onnx.load(policy_path)
     metadata = {prop.key: prop.value for prop in model.metadata_props}
 
     if "kp" not in metadata or "kd" not in metadata:
@@ -96,13 +101,15 @@ def load_policy(
 
 
 class HolosomaLocomotionController:
-    """Holosoma lower-body locomotion controller for Unitree G1."""
+    """Holosoma whole-body locomotion controller for Unitree G1."""
 
-    control_dt = CONTROL_DT  # Expose for unitree_g1.py
+    def __init__(self, policy, robot, kp: np.ndarray, kd: np.ndarray):
+        self.policy = policy
+        self.robot = robot
 
-    def __init__(self):
-        # Load policy and gains
-        self.policy, self.kp, self.kd = load_policy()
+        # Override robot's PD gains with policy gains
+        self.robot.kp = kp
+        self.robot.kd = kd
 
         self.cmd = np.zeros(3, dtype=np.float32)
 
@@ -117,55 +124,35 @@ class HolosomaLocomotionController:
         self.phase_dt = 2 * np.pi / ((1.0 / CONTROL_DT) * GAIT_PERIOD)
         self.is_standing = True
 
-        logger.info("HolosomaLocomotionController initialized")
+    def run_step(self):
+        # Get current observation
+        obs = self.robot.get_observation()
 
-    def reset(self) -> None:
-        """Reset internal state for a new episode."""
-        self.cmd[:] = 0.0
-        self.qj[:] = 0.0
-        self.dqj[:] = 0.0
-        self.obs[:] = 0.0
-        self.last_action[:] = 0.0
-        self.phase = np.array([[0.0, np.pi]], dtype=np.float32)
-        self.is_standing = True
+        if not obs:
+            return
 
-    def run_step(self, action: dict, lowstate) -> dict:
-        """Run one step of the locomotion controller.
-
-        Args:
-            action: Action dict containing remote.lx/ly/rx/ry
-            lowstate: Robot lowstate containing motor positions/velocities and IMU
-
-        Returns:
-            Action dict for lower body joints (0-14)
-        """
-        if lowstate is None:
-            return {}
-
-        lx, ly, rx, _ry = (action.get(k, 0.0) for k in REMOTE_AXES)
-        ly = ly if abs(ly) > 0.1 else 0.0
-        lx = lx if abs(lx) > 0.1 else 0.0
-        rx = rx if abs(rx) > 0.1 else 0.0
-        ly = np.clip(ly, -0.3, 0.3)
-        lx = np.clip(lx, -0.3, 0.3)
+        # Get command from remote controller
+        ly = obs["remote.ly"] if abs(obs["remote.ly"]) > 0.1 else 0.0
+        lx = obs["remote.lx"] if abs(obs["remote.lx"]) > 0.1 else 0.0
+        rx = obs["remote.rx"] if abs(obs["remote.rx"]) > 0.1 else 0.0
         self.cmd[:] = [ly, -lx, -rx]
 
-        # Get joint positions and velocities from lowstate
+        # Get joint positions and velocities
         for motor in G1_29_JointIndex:
+            name = motor.name
             idx = motor.value
-            self.qj[idx] = lowstate.motor_state[idx].q
-            self.dqj[idx] = lowstate.motor_state[idx].dq
+            self.qj[idx] = obs[f"{name}.q"]
+            self.dqj[idx] = obs[f"{name}.dq"]
 
-        # Hide arm positions from policy (show DEFAULT_ANGLES instead)
-        # This prevents policy from reacting to teleop arm movements
-        for arm_joint in G1_29_JointArmIndex:
-            self.qj[arm_joint.value] = DEFAULT_ANGLES[arm_joint.value]
-            self.dqj[arm_joint.value] = 0.0
+        # Adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.qj[idx] = 0.0
+            self.dqj[idx] = 0.0
 
         # Express IMU data in gravity frame of reference
-        quat = lowstate.imu_state.quaternion
-        ang_vel = np.array(lowstate.imu_state.gyroscope, dtype=np.float32)
-        gravity = get_gravity_orientation(quat)
+        quat = [obs["imu.quat.w"], obs["imu.quat.x"], obs["imu.quat.y"], obs["imu.quat.z"]]
+        ang_vel = np.array([obs["imu.gyro.x"], obs["imu.gyro.y"], obs["imu.gyro.z"]], dtype=np.float32)
+        gravity = self.robot.get_gravity_orientation(quat)
 
         # Scale joint positions and velocities before policy inference
         qj_obs = (self.qj - DEFAULT_ANGLES) * DOF_POS_SCALE
@@ -199,16 +186,79 @@ class HolosomaLocomotionController:
         # Run policy inference
         ort_in = {self.policy.get_inputs()[0].name: self.obs.reshape(1, -1).astype(np.float32)}
         raw_action = self.policy.run(None, ort_in)[0].squeeze()
-        policy_action = np.clip(raw_action, -100.0, 100.0)
-        self.last_action = policy_action.copy()
+        action = np.clip(raw_action, -100.0, 100.0)
+        self.last_action = action.copy()
 
         # Transform action back to target joint positions
-        target = DEFAULT_ANGLES + policy_action * ACTION_SCALE
+        target = DEFAULT_ANGLES + action * ACTION_SCALE
 
-        # Build action dict (first 15 joints only)
+        # Build action dict
         action_dict = {}
-        for i in range(15):
-            motor_name = G1_29_JointIndex(i).name
-            action_dict[f"{motor_name}.q"] = float(target[i])
+        for motor in G1_29_JointIndex:
+            action_dict[f"{motor.name}.q"] = float(target[motor.value])
 
-        return action_dict
+        # Zero out missing joints for g1_23dof
+        for joint_idx in MISSING_JOINTS:
+            motor_name = G1_29_JointIndex(joint_idx).name
+            action_dict[f"{motor_name}.q"] = 0.0
+
+        # Send action to robot
+        self.robot.send_action(action_dict)
+
+
+def run(repo_id: str = DEFAULT_HOLOSOMA_REPO_ID, policy_type: str = "fastsac") -> None:
+    """Main function to run the Holosoma locomotion controller.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID for Holosoma policies.
+        policy_type: Policy type to use ('fastsac' or 'ppo').
+    """
+    # Load policy and gains
+    policy, kp, kd = load_policy(repo_id=repo_id, policy_type=policy_type)
+
+    # Initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+    robot.connect()
+
+    holosoma_controller = HolosomaLocomotionController(policy, robot, kp, kd)
+
+    try:
+        robot.reset(CONTROL_DT, DEFAULT_ANGLES)
+
+        logger.info("Use joystick: LY=fwd/back, LX=left/right, RX=rotate")
+        logger.info("Press Ctrl+C to stop")
+
+        # Run step
+        while not robot._shutdown_event.is_set():
+            start_time = time.time()
+            holosoma_controller.run_step()
+            elapsed = time.time() - start_time
+            sleep_time = max(0, CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+    except KeyboardInterrupt:
+        logger.info("Stopping locomotion...")
+    finally:
+        if robot.is_connected:
+            robot.disconnect()
+        logger.info("Done!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Holosoma Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_HOLOSOMA_REPO_ID,
+        help=f"Hugging Face Hub repo ID for Holosoma policies (default: {DEFAULT_HOLOSOMA_REPO_ID})",
+    )
+    parser.add_argument(
+        "--policy",
+        type=str,
+        choices=["fastsac", "ppo"],
+        default="fastsac",
+        help="Policy type to use: 'fastsac' (default) or 'ppo'",
+    )
+    args = parser.parse_args()
+
+    run(repo_id=args.repo_id, policy_type=args.policy)

pyproject.toml

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.5.1"
+version = "0.4.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.12"
+requires-python = ">=3.10"
 authors = [
     { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
     { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,8 +50,7 @@ classifiers = [
     "Intended Audience :: Education",
     "Intended Audience :: Science/Research",
     "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.12",
-    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.10",
     "Topic :: Software Development :: Build Tools",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -60,30 +59,28 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.0.0,<4.2.0",
     "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub>=1.0.0,<2.0.0",
+    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
     "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
-    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "setuptools>=71.0.0,<81.0.0",
     "cmake>=3.29.0.1,<4.2.0",
-    "packaging>=24.2,<26.0",
-
-    "torch>=2.2.1,<2.11.0",
-    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')",
-    "torchvision>=0.21.0,<0.26.0",
-
     "einops>=0.8.0,<0.9.0",
-    "opencv-python-headless>=4.9.0,<4.14.0",
+    "opencv-python-headless>=4.9.0,<4.13.0",
     "av>=15.0.0,<16.0.0",
     "jsonlines>=4.0.0,<5.0.0",
-    "pynput>=1.7.8,<1.9.0",
+    "packaging>=24.2,<26.0",
+    "pynput>=1.7.7,<1.9.0",
     "pyserial>=3.5,<4.0",
-
     "wandb>=0.24.0,<0.25.0",
-    "draccus==0.10.0", # TODO: Relax version constraint
+
+    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
+    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
+
+    "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -98,20 +95,14 @@ dependencies = [
 
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.9.17"]
-transformers-dep = ["transformers>=5.3.0,<6.0.0"]
+placo-dep = ["placo>=0.9.6,<0.10.0"]
+transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
-can-dep = ["python-can>=4.2.0,<5.0.0"]
-peft-dep = ["peft>=0.18.0,<1.0.0"]
-scipy-dep = ["scipy>=1.14.0,<2.0.0"]
-qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
-matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["lerobot[can-dep]"]
-robstride = ["lerobot[can-dep]"]
+damiao = ["python-can>=4.2.0,<5.0.0"]
 
 # Robots
 openarms = ["lerobot[damiao]"]
@@ -119,35 +110,34 @@ gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
-    # "unitree-sdk2==1.0.1",
     "pyzmq>=26.2.1,<28.0.0",
     "onnxruntime>=1.16.0,<2.0.0",
-    "onnx>=1.16.0,<2.0.0",
+    "pin>=3.0.0,<4.0.0",
     "meshcat>=0.3.0,<0.4.0",
-    "lerobot[matplotlib-dep]",
-    "lerobot[pygame-dep]",
+    "matplotlib>=3.9.0,<4.0.0",
+    "casadi>=3.6.0,<4.0.0",
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
-    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
 
 # Policies
 wallx = [
-    "lerobot[transformers-dep]",
-    "lerobot[peft]",
-    "lerobot[scipy-dep]",
-    "torchdiffeq>=0.2.4,<0.3.0",
-    "lerobot[qwen-vl-utils-dep]",
+    "transformers==4.49.0",
+    "peft==0.17.1",
+    "scipy==1.15.3",
+    "torchdiffeq==0.2.5",
+    "qwen_vl_utils==0.0.11"
 ]
-pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
+pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
 groot = [
     "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "peft>=0.13.0,<1.0.0",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
     "safetensors>=0.4.3,<1.0.0",
@@ -156,13 +146,13 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
-async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
-peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
+async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
+peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -170,19 +160,13 @@ test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
-# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+aloha = ["gym-aloha>=0.1.2,<0.2.0"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
-metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
+metaworld = ["metaworld==3.0.0"]
 
 # All
 all = [
-    # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
-    # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
-    # helps pip's resolver converge by constraining scipy early, before it encounters
-    # the loose scipy requirements from transitive deps like dm-control and metaworld.
-    "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
@@ -190,8 +174,8 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    "lerobot[wallx]",
-    "lerobot[pi]",
+    # "lerobot[wallx]",
+    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
@@ -203,11 +187,10 @@ all = [
     "lerobot[aloha]",
     "lerobot[pusht]",
     "lerobot[phone]",
-    "lerobot[libero]; sys_platform == 'linux'",
+    "lerobot[libero]",
     "lerobot[metaworld]",
     "lerobot[sarm]",
     "lerobot[peft]",
-    # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]
 
 [project.scripts]
@@ -229,14 +212,11 @@ lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
-[tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
-
 [tool.setuptools.packages.find]
 where = ["src"]
 
 [tool.ruff]
-target-version = "py312"
+target-version = "py310"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -328,7 +308,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
 
 [tool.mypy]
-python_version = "3.12"
+python_version = "3.10"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -380,9 +360,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-[[tool.mypy.overrides]]
-module = "lerobot.motors.*"
-ignore_errors = false
+# [[tool.mypy.overrides]]
+# module = "lerobot.motors.*"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"
@@ -412,3 +392,85 @@ ignore_errors = false
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
+
+[tool.uv]
+# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
+conflicts = [
+    [
+        { extra = "wallx" },
+        { extra = "transformers-dep" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "pi" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "smolvla" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "groot" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "xvla" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "sarm" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "hilserl" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "libero" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "peft" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "all" },
+    ],
+    # pi uses custom branch which conflicts with transformers-dep
+    [
+        { extra = "pi" },
+        { extra = "transformers-dep" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "smolvla" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "groot" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "xvla" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "sarm" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "hilserl" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "libero" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "peft" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "all" },
+    ],
+]

requirements-macos.txt

@@ -1,73 +1,76 @@
 #
-# This file is autogenerated by pip-compile with Python 3.12
+# This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
 #    pip-compile --output-file=requirements-macos.txt requirements.in
 #
 -e .[all]
     # via -[all]
-absl-py==2.4.0
+absl-py==2.3.1
     # via
     #   dm-control
     #   dm-env
     #   dm-tree
     #   labmaze
     #   mujoco
-accelerate==1.13.0
+    #   tensorboard
+accelerate==1.11.0
     # via
     #   lerobot
     #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.3
+aiohttp==3.13.1
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
-annotated-doc==0.0.4
-    # via
-    #   fastapi
-    #   typer
 annotated-types==0.7.0
     # via pydantic
-anyio==4.12.1
+antlr4-python3-runtime==4.9.3
+    # via
+    #   hydra-core
+    #   omegaconf
+anyio==4.11.0
     # via
-    #   httpx
     #   starlette
     #   watchfiles
-asttokens==3.0.1
+asttokens==3.0.0
     # via stack-data
+async-timeout==5.0.1
+    # via aiohttp
 attrs==25.4.0
     # via
     #   aiohttp
     #   dm-tree
     #   jsonlines
+    #   jsonschema
+    #   referencing
     #   rerun-sdk
 av==15.1.0
+    # via lerobot
+bddl==1.0.1
+    # via libero
+certifi==2025.10.5
     # via
-    #   lerobot
-    #   qwen-vl-utils
-certifi==2026.2.25
-    # via
-    #   httpcore
-    #   httpx
     #   requests
     #   sentry-sdk
 cffi==2.0.0
     # via pymunk
-cfgv==3.5.0
+cfgv==3.4.0
     # via pre-commit
-charset-normalizer==3.4.5
+charset-normalizer==3.4.4
     # via requests
-click==8.3.1
+click==8.3.0
     # via
-    #   typer
     #   uvicorn
     #   wandb
-cloudpickle==3.1.2
-    # via gymnasium
-cmake==4.1.3
+cloudpickle==3.1.1
+    # via
+    #   gymnasium
+    #   libero
+cmake==4.1.0
     # via lerobot
-cmeel==0.59.0
+cmeel==0.57.3
     # via
     #   cmeel-assimp
     #   cmeel-boost
@@ -105,17 +108,15 @@ cmeel-zlib==1.3.1
     # via cmeel-assimp
 coal-library==3.0.1
     # via pin
-contourpy==1.3.3
-    # via
-    #   lerobot
-    #   matplotlib
-coverage[toml]==7.13.4
+contourpy==1.3.2
+    # via matplotlib
+coverage[toml]==7.11.0
     # via pytest-cov
 cycler==0.12.1
     # via matplotlib
-datasets==4.6.1
+datasets==4.1.1
     # via lerobot
-debugpy==1.8.20
+debugpy==1.8.17
     # via lerobot
 decorator==5.2.1
     # via ipython
@@ -129,7 +130,7 @@ dill==0.4.0
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.37
+dm-control==1.0.34
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -137,55 +138,69 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
+    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
     # via lerobot
 dynamixel-sdk==3.8.4
     # via lerobot
+easydict==1.13
+    # via libero
+egl-probe @ git+https://github.com/huggingface/egl_probe.git
+    # via
+    #   libero
+    #   robomimic
 eigenpy==3.10.3
     # via coal-library
-einops==0.8.2
-    # via lerobot
-eiquadprog==1.2.9
-    # via placo
-etils[epath,epy]==1.14.0
-    # via mujoco
-executing==2.2.1
-    # via stack-data
-faker==34.0.2
-    # via lerobot
-farama-notifications==0.0.4
-    # via gymnasium
-fastapi==0.135.1
+einops==0.8.1
     # via
     #   lerobot
-    #   teleop
+    #   libero
+eiquadprog==1.2.9
+    # via placo
+etils[epath,epy]==1.13.0
+    # via mujoco
+exceptiongroup==1.3.0
+    # via
+    #   anyio
+    #   ipython
+    #   pytest
+executing==2.2.1
+    # via stack-data
+farama-notifications==0.0.4
+    # via gymnasium
+fastapi==0.119.1
+    # via teleop
+fastjsonschema==2.21.2
+    # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.25.0
+filelock==3.20.0
     # via
     #   datasets
     #   diffusers
     #   huggingface-hub
-    #   python-discovery
     #   torch
+    #   transformers
     #   virtualenv
-fonttools==4.61.1
+fonttools==4.60.1
     # via matplotlib
 frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2026.2.0
+fsspec[http]==2025.9.0
     # via
     #   datasets
     #   etils
     #   huggingface-hub
     #   torch
+future==1.0.0
+    # via libero
 gitdb==4.0.12
     # via gitpython
-gitpython==3.1.46
+gitpython==3.1.45
     # via wandb
 glfw==2.10.0
     # via
@@ -197,6 +212,7 @@ grpcio==1.73.1
     #   lerobot
     #   reachy2-sdk
     #   reachy2-sdk-api
+    #   tensorboard
 grpcio-tools==1.73.1
     # via
     #   lerobot
@@ -207,67 +223,71 @@ gym-hil==0.1.13
     # via lerobot
 gym-pusht==0.1.6
     # via lerobot
-gymnasium==1.2.3
+gymnasium==1.2.1
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
     #   lerobot
+    #   libero
     #   metaworld
 h11==0.16.0
-    # via
-    #   httpcore
-    #   uvicorn
+    # via uvicorn
+h5py==3.15.1
+    # via robomimic
 hebi-py==2.11.0
     # via lerobot
-hf-xet==1.3.2
+hf-transfer==0.1.9
+    # via huggingface-hub
+hf-xet==1.1.10
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
-httpcore==1.0.9
-    # via httpx
 httptools==0.7.1
     # via uvicorn
-httpx==0.28.1
-    # via
-    #   datasets
-    #   huggingface-hub
-huggingface-hub==1.6.0
+huggingface-hub[cli,hf-transfer]==0.35.3
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
     #   peft
+    #   timm
     #   tokenizers
     #   transformers
-identify==2.6.17
+hydra-core==1.3.2
+    # via libero
+identify==2.6.15
     # via pre-commit
 idna==3.11
     # via
     #   anyio
-    #   httpx
     #   requests
     #   yarl
-imageio[ffmpeg]==2.37.2
+imageio[ffmpeg]==2.37.0
     # via
     #   gym-aloha
     #   gym-hil
     #   lerobot
     #   metaworld
+    #   robomimic
     #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via imageio
-importlib-metadata==8.7.1
+    # via
+    #   imageio
+    #   robomimic
+importlib-metadata==8.7.0
     # via diffusers
+importlib-resources==6.5.2
+    # via etils
 iniconfig==2.3.0
     # via pytest
-ipython==9.11.0
+inquirerpy==0.3.4
+    # via huggingface-hub
+ipython==8.37.0
     # via meshcat
-ipython-pygments-lexers==1.1.1
-    # via ipython
 ischedule==1.2.7
     # via placo
 jedi==0.19.2
@@ -276,24 +296,44 @@ jinja2==3.1.6
     # via torch
 jsonlines==4.0.0
     # via lerobot
+jsonschema==4.25.1
+    # via nbformat
+jsonschema-specifications==2025.9.1
+    # via jsonschema
+jupyter-core==5.9.1
+    # via nbformat
+jupytext==1.18.1
+    # via bddl
 kiwisolver==1.4.9
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
-lazy-loader==0.5
+lazy-loader==0.4
     # via scikit-image
-librt==0.8.1
-    # via mypy
+libero @ git+https://github.com/huggingface/lerobot-libero.git@main
+    # via lerobot
+llvmlite==0.45.1
+    # via numba
 lxml==6.0.2
     # via dm-control
+markdown==3.9
+    # via tensorboard
 markdown-it-py==4.0.0
-    # via rich
+    # via
+    #   jupytext
+    #   mdit-py-plugins
 markupsafe==3.0.3
-    # via jinja2
-matplotlib==3.10.8
-    # via lerobot
+    # via
+    #   jinja2
+    #   werkzeug
+matplotlib==3.10.7
+    # via
+    #   lerobot
+    #   libero
 matplotlib-inline==0.2.1
     # via ipython
+mdit-py-plugins==0.5.0
+    # via jupytext
 mdurl==0.1.2
     # via markdown-it-py
 mergedeep==1.3.4
@@ -306,35 +346,41 @@ mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.5.0
+mujoco==3.3.7
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
+    #   libero
     #   metaworld
-multidict==6.7.1
+    #   robosuite
+multidict==6.7.0
     # via
     #   aiohttp
     #   yarl
-multiprocess==0.70.18
+multiprocess==0.70.16
     # via datasets
-mypy==1.19.1
-    # via lerobot
 mypy-extensions==1.1.0
+    # via typing-inspect
+nbformat==5.10.4
+    # via jupytext
+networkx==3.4.2
     # via
-    #   mypy
-    #   typing-inspect
-networkx==3.6.1
-    # via
+    #   bddl
     #   scikit-image
     #   torch
-nodeenv==1.10.0
+ninja==1.13.0
+    # via lerobot
+nodeenv==1.9.1
     # via pre-commit
 num2words==0.5.14
     # via lerobot
+numba==0.62.1
+    # via robosuite
 numpy==2.2.6
     # via
     #   accelerate
+    #   bddl
     #   cmeel-boost
     #   contourpy
     #   datasets
@@ -343,14 +389,16 @@ numpy==2.2.6
     #   dm-env
     #   dm-tree
     #   gymnasium
+    #   h5py
     #   hebi-py
     #   imageio
     #   labmaze
-    #   lerobot
+    #   libero
     #   matplotlib
     #   meshcat
     #   metaworld
     #   mujoco
+    #   numba
     #   opencv-python
     #   opencv-python-headless
     #   pandas
@@ -358,18 +406,26 @@ numpy==2.2.6
     #   pyquaternion
     #   reachy2-sdk
     #   rerun-sdk
+    #   robomimic
+    #   robosuite
     #   scikit-image
     #   scipy
     #   shapely
     #   teleop
+    #   tensorboard
+    #   tensorboardx
     #   tifffile
     #   torchvision
     #   transformers
     #   transforms3d
-opencv-python==4.13.0.92
+omegaconf==2.3.0
+    # via hydra-core
+opencv-python==4.12.0.88
     # via
     #   gym-pusht
+    #   libero
     #   reachy2-sdk
+    #   robosuite
 opencv-python-headless==4.12.0.88
     # via lerobot
 orderly-set==5.5.0
@@ -379,87 +435,97 @@ packaging==25.0
     #   accelerate
     #   datasets
     #   huggingface-hub
+    #   hydra-core
+    #   jupytext
     #   lazy-loader
     #   lerobot
     #   matplotlib
     #   peft
     #   pytest
-    #   qwen-vl-utils
     #   reachy2-sdk
     #   scikit-image
+    #   tensorboard
+    #   tensorboardx
     #   transformers
     #   wandb
 pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.6
+parso==0.8.5
     # via jedi
-pathspec==1.0.4
-    # via mypy
-peft==0.18.1
+peft==0.17.1
     # via lerobot
 pexpect==4.9.0
     # via ipython
-pillow==12.1.1
+pfzy==0.3.4
+    # via inquirerpy
+pillow==12.0.0
     # via
     #   diffusers
     #   imageio
+    #   lerobot
     #   matplotlib
     #   meshcat
-    #   qwen-vl-utils
     #   rerun-sdk
+    #   robosuite
     #   scikit-image
+    #   tensorboard
     #   torchvision
 pin==3.4.0
     # via placo
-placo==0.9.16
+placo==0.9.14
     # via lerobot
-platformdirs==4.9.4
+platformdirs==4.5.0
     # via
-    #   python-discovery
+    #   jupyter-core
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.5.1
+pre-commit==4.3.0
     # via lerobot
 prompt-toolkit==3.0.52
-    # via ipython
+    # via
+    #   inquirerpy
+    #   ipython
 propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
-protobuf==6.31.1
+protobuf==6.31.0
     # via
     #   dm-control
     #   grpcio-tools
     #   lerobot
     #   reachy2-sdk
     #   reachy2-sdk-api
+    #   tensorboard
+    #   tensorboardx
     #   wandb
-psutil==7.2.2
+psutil==7.1.1
     # via
     #   accelerate
     #   imageio
     #   peft
+    #   robomimic
 ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
     # via stack-data
-pyarrow==23.0.1
+pyarrow==21.0.0
     # via
     #   datasets
     #   rerun-sdk
-pycparser==3.0
+pycparser==2.23
     # via cffi
-pydantic==2.12.5
+pydantic==2.12.3
     # via
     #   fastapi
     #   wandb
-pydantic-core==2.41.5
+pydantic-core==2.41.4
     # via pydantic
 pygame==2.6.1
     # via
@@ -469,35 +535,33 @@ pygame==2.6.1
 pygments==2.19.2
     # via
     #   ipython
-    #   ipython-pygments-lexers
     #   pytest
-    #   rich
 pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.5.1
+pyngrok==7.4.1
     # via meshcat
 pynput==1.8.1
     # via
     #   gym-hil
     #   lerobot
-pyobjc-core==12.1
+pyobjc-core==12.0
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-cocoa
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==12.1
+pyobjc-framework-applicationservices==12.0
     # via pynput
-pyobjc-framework-cocoa==12.1
+pyobjc-framework-cocoa==12.0
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-coretext==12.1
+pyobjc-framework-coretext==12.0
     # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==12.1
+pyobjc-framework-quartz==12.0
     # via
     #   pynput
     #   pyobjc-framework-applicationservices
@@ -506,13 +570,13 @@ pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.3.2
+pyparsing==3.2.5
     # via
     #   dm-control
     #   matplotlib
 pyquaternion==0.9.9
     # via reachy2-sdk
-pyrealsense2-macosx==2.56.5
+pyrealsense2-macosx==2.54.2
     # via lerobot
 pyserial==3.5
     # via
@@ -521,6 +585,7 @@ pyserial==3.5
     #   lerobot
 pytest==8.4.2
     # via
+    #   bddl
     #   lerobot
     #   pytest-cov
     #   pytest-timeout
@@ -531,14 +596,11 @@ pytest-timeout==2.4.0
     # via lerobot
 python-dateutil==2.9.0.post0
     # via
-    #   faker
     #   matplotlib
     #   pandas
-python-discovery==1.1.1
-    # via virtualenv
-python-dotenv==1.2.2
+python-dotenv==1.1.1
     # via uvicorn
-pytz==2026.1.post1
+pytz==2025.2
     # via pandas
 pyyaml==6.0.3
     # via
@@ -547,10 +609,13 @@ pyyaml==6.0.3
     #   draccus
     #   hebi-py
     #   huggingface-hub
+    #   jupytext
+    #   omegaconf
     #   peft
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
+    #   timm
     #   transformers
     #   uvicorn
     #   wandb
@@ -560,13 +625,15 @@ pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-qwen-vl-utils==0.0.14
-    # via lerobot
-reachy2-sdk==1.0.15
+reachy2-sdk==1.0.14
     # via lerobot
 reachy2-sdk-api==1.0.21
     # via reachy2-sdk
-regex==2026.2.28
+referencing==0.37.0
+    # via
+    #   jsonschema
+    #   jsonschema-specifications
+regex==2025.10.23
     # via
     #   diffusers
     #   transformers
@@ -575,150 +642,184 @@ requests==2.32.5
     #   datasets
     #   diffusers
     #   dm-control
-    #   qwen-vl-utils
+    #   huggingface-hub
     #   teleop
+    #   transformers
     #   wandb
-rerun-sdk==0.26.2
+rerun-sdk==0.26.1
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-rich==14.3.3
-    # via typer
-safetensors==0.7.0
+robomimic==0.2.0
+    # via libero
+robosuite==1.4.0
+    # via libero
+rpds-py==0.28.0
+    # via
+    #   jsonschema
+    #   referencing
+safetensors==0.6.2
     # via
     #   accelerate
     #   diffusers
     #   lerobot
     #   peft
+    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
     #   gym-pusht
     #   lerobot
-scipy==1.17.1
+scipy==1.15.3
     # via
     #   dm-control
-    #   lerobot
     #   metaworld
+    #   robosuite
     #   scikit-image
-    #   torchdiffeq
-sentry-sdk==2.54.0
+sentry-sdk==2.42.1
     # via wandb
 shapely==2.1.2
     # via gym-pusht
-shellingham==1.5.4
-    # via typer
 six==1.17.0
     # via
     #   pynput
     #   python-dateutil
-smmap==5.0.3
+smmap==5.0.2
     # via gitdb
+sniffio==1.3.1
+    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.52.1
+starlette==0.48.0
     # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.4
+teleop==0.1.2
     # via lerobot
-termcolor==3.3.0
-    # via lerobot
-tifffile==2026.3.3
+tensorboard==2.20.0
+    # via robomimic
+tensorboard-data-server==0.7.2
+    # via tensorboard
+tensorboardx==2.6.4
+    # via robomimic
+termcolor==3.1.0
+    # via
+    #   lerobot
+    #   robomimic
+thop==0.1.1.post2209072238
+    # via libero
+tifffile==2025.5.10
     # via scikit-image
-tokenizers==0.22.2
+timm==1.0.20
+    # via lerobot
+tokenizers==0.22.1
     # via transformers
 toml==0.10.2
     # via draccus
-torch==2.10.0
+tomli==2.3.0
+    # via
+    #   cmeel
+    #   coverage
+    #   jupytext
+    #   pytest
+torch==2.7.1
     # via
     #   accelerate
     #   lerobot
     #   peft
-    #   torchdiffeq
+    #   robomimic
+    #   thop
+    #   timm
     #   torchvision
-torchcodec==0.10.0
+torchcodec==0.5
     # via lerobot
-torchdiffeq==0.2.5
-    # via lerobot
-torchvision==0.25.0
-    # via lerobot
-tornado==6.5.4
+torchvision==0.22.1
+    # via
+    #   lerobot
+    #   robomimic
+    #   timm
+tornado==6.5.2
     # via meshcat
-tqdm==4.67.3
+tqdm==4.67.1
     # via
     #   datasets
     #   dm-control
     #   huggingface-hub
     #   peft
+    #   robomimic
     #   transformers
 traitlets==5.14.3
     # via
     #   ipython
+    #   jupyter-core
     #   matplotlib-inline
-transformers==5.3.0
+    #   nbformat
+transformers==4.57.1
     # via
     #   lerobot
+    #   libero
     #   peft
 transforms3d==0.4.2
     # via teleop
-typer==0.24.1
-    # via
-    #   huggingface-hub
-    #   transformers
 typing-extensions==4.15.0
     # via
     #   aiosignal
     #   anyio
     #   etils
-    #   faker
+    #   exceptiongroup
     #   fastapi
     #   gymnasium
     #   huggingface-hub
-    #   mypy
+    #   ipython
+    #   multidict
     #   pydantic
     #   pydantic-core
+    #   referencing
     #   rerun-sdk
     #   starlette
     #   torch
     #   typing-inspect
     #   typing-inspection
+    #   uvicorn
+    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
 typing-inspection==0.4.2
-    # via
-    #   fastapi
-    #   pydantic
-tzdata==2025.3
+    # via pydantic
+tzdata==2025.2
     # via pandas
 u-msgpack-python==2.8.0
     # via meshcat
-urllib3==2.6.3
+urllib3==2.5.0
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.41.0
+uvicorn[standard]==0.38.0
     # via teleop
 uvloop==0.22.1
     # via uvicorn
-virtualenv==21.1.0
+virtualenv==20.35.3
     # via pre-commit
-wandb==0.24.2
-    # via lerobot
+wandb==0.21.4
+    # via
+    #   lerobot
+    #   libero
 watchfiles==1.1.1
     # via uvicorn
-wcwidth==0.6.0
+wcwidth==0.2.14
     # via prompt-toolkit
 websocket-client==1.9.0
     # via teleop
-websockets==16.0
+websockets==15.0.1
     # via uvicorn
-wrapt==2.1.2
+werkzeug==3.1.3
+    # via tensorboard
+wrapt==2.0.0
     # via dm-tree
 xxhash==3.6.0
     # via datasets
-yarl==1.23.0
+yarl==1.22.0
     # via aiohttp
 zipp==3.23.0
     # via

requirements-ubuntu.txt

@@ -1,12 +1,12 @@
 #
-# This file is autogenerated by pip-compile with Python 3.12
+# This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
 #    pip-compile --output-file=requirements-ubuntu.txt requirements.in
 #
 -e .[all]
     # via -[all]
-absl-py==2.4.0
+absl-py==2.3.1
     # via
     #   dm-control
     #   dm-env
@@ -14,33 +14,30 @@ absl-py==2.4.0
     #   labmaze
     #   mujoco
     #   tensorboard
-accelerate==1.13.0
+accelerate==1.11.0
     # via
     #   lerobot
     #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.3
+aiohttp==3.13.1
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
-annotated-doc==0.0.4
-    # via
-    #   fastapi
-    #   typer
 annotated-types==0.7.0
     # via pydantic
 antlr4-python3-runtime==4.9.3
     # via
     #   hydra-core
     #   omegaconf
-anyio==4.12.1
+anyio==4.11.0
     # via
-    #   httpx
     #   starlette
     #   watchfiles
-asttokens==3.0.1
+asttokens==3.0.0
     # via stack-data
+async-timeout==5.0.1
+    # via aiohttp
 attrs==25.4.0
     # via
     #   aiohttp
@@ -50,35 +47,30 @@ attrs==25.4.0
     #   referencing
     #   rerun-sdk
 av==15.1.0
-    # via
-    #   lerobot
-    #   qwen-vl-utils
+    # via lerobot
 bddl==1.0.1
-    # via hf-libero
-certifi==2026.2.25
+    # via libero
+certifi==2025.10.5
     # via
-    #   httpcore
-    #   httpx
     #   requests
     #   sentry-sdk
 cffi==2.0.0
     # via pymunk
-cfgv==3.5.0
+cfgv==3.4.0
     # via pre-commit
-charset-normalizer==3.4.5
+charset-normalizer==3.4.4
     # via requests
-click==8.3.1
+click==8.3.0
     # via
-    #   typer
     #   uvicorn
     #   wandb
-cloudpickle==3.1.2
+cloudpickle==3.1.1
     # via
     #   gymnasium
-    #   hf-libero
-cmake==4.1.3
+    #   libero
+cmake==4.1.0
     # via lerobot
-cmeel==0.59.0
+cmeel==0.57.3
     # via
     #   cmeel-assimp
     #   cmeel-boost
@@ -116,24 +108,20 @@ cmeel-zlib==1.3.1
     # via cmeel-assimp
 coal-library==3.0.1
     # via pin
-contourpy==1.3.3
-    # via
-    #   lerobot
-    #   matplotlib
-coverage[toml]==7.13.4
+contourpy==1.3.2
+    # via matplotlib
+coverage[toml]==7.11.0
     # via pytest-cov
-cuda-bindings==12.9.4
-    # via torch
-cuda-pathfinder==1.4.1
-    # via cuda-bindings
 cycler==0.12.1
     # via matplotlib
-datasets==4.6.1
+datasets==4.1.1
     # via lerobot
-debugpy==1.8.20
+debugpy==1.8.17
     # via lerobot
 decorator==5.2.1
     # via ipython
+decord==0.6.0
+    # via lerobot
 deepdiff==8.6.1
     # via lerobot
 diffusers==0.35.2
@@ -144,7 +132,7 @@ dill==0.4.0
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.37
+dm-control==1.0.34
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -152,6 +140,7 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
+    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
@@ -159,60 +148,66 @@ draccus==0.10.0
 dynamixel-sdk==3.8.4
     # via lerobot
 easydict==1.13
-    # via hf-libero
-egl-probe==1.0.2
-    # via robomimic
+    # via libero
+egl-probe @ git+https://github.com/huggingface/egl_probe.git
+    # via
+    #   libero
+    #   robomimic
 eigenpy==3.10.3
     # via coal-library
-einops==0.8.2
+einops==0.8.1
     # via
-    #   hf-libero
+    #   flash-attn
     #   lerobot
+    #   libero
 eiquadprog==1.2.9
     # via placo
-etils[epath,epy]==1.14.0
+etils[epath,epy]==1.13.0
     # via mujoco
-evdev==1.9.3
+evdev==1.9.2
     # via pynput
+exceptiongroup==1.3.0
+    # via
+    #   anyio
+    #   ipython
+    #   pytest
 executing==2.2.1
     # via stack-data
-faker==34.0.2
-    # via lerobot
 farama-notifications==0.0.4
     # via gymnasium
-fastapi==0.135.1
-    # via
-    #   lerobot
-    #   teleop
+fastapi==0.119.1
+    # via teleop
 fastjsonschema==2.21.2
     # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.25.0
+filelock==3.20.0
     # via
     #   datasets
     #   diffusers
     #   huggingface-hub
-    #   python-discovery
     #   torch
+    #   transformers
     #   virtualenv
-fonttools==4.61.1
+flash-attn==2.8.3
+    # via lerobot
+fonttools==4.60.1
     # via matplotlib
 frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2026.2.0
+fsspec[http]==2025.9.0
     # via
     #   datasets
     #   etils
     #   huggingface-hub
     #   torch
 future==1.0.0
-    # via hf-libero
+    # via libero
 gitdb==4.0.12
     # via gitpython
-gitpython==3.1.46
+gitpython==3.1.45
     # via wandb
 glfw==2.10.0
     # via
@@ -235,60 +230,50 @@ gym-hil==0.1.13
     # via lerobot
 gym-pusht==0.1.6
     # via lerobot
-gymnasium==1.2.3
+gymnasium==1.2.1
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
-    #   hf-libero
     #   lerobot
+    #   libero
     #   metaworld
 h11==0.16.0
-    # via
-    #   httpcore
-    #   uvicorn
-h5py==3.16.0
+    # via uvicorn
+h5py==3.15.1
     # via robomimic
 hebi-py==2.11.0
     # via lerobot
-hf-egl-probe==1.0.2
-    # via hf-libero
-hf-libero==0.1.3
-    # via lerobot
-hf-xet==1.3.2
+hf-transfer==0.1.9
+    # via huggingface-hub
+hf-xet==1.1.10
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
-httpcore==1.0.9
-    # via httpx
 httptools==0.7.1
     # via uvicorn
-httpx==0.28.1
-    # via
-    #   datasets
-    #   huggingface-hub
-huggingface-hub==1.6.0
+huggingface-hub[cli,hf-transfer]==0.35.3
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
     #   peft
+    #   timm
     #   tokenizers
     #   transformers
 hydra-core==1.3.2
-    # via hf-libero
-identify==2.6.17
+    # via libero
+identify==2.6.15
     # via pre-commit
 idna==3.11
     # via
     #   anyio
-    #   httpx
     #   requests
     #   yarl
-imageio[ffmpeg]==2.37.2
+imageio[ffmpeg]==2.37.0
     # via
     #   gym-aloha
     #   gym-hil
@@ -300,14 +285,16 @@ imageio-ffmpeg==0.6.0
     # via
     #   imageio
     #   robomimic
-importlib-metadata==8.7.1
+importlib-metadata==8.7.0
     # via diffusers
+importlib-resources==6.5.2
+    # via etils
 iniconfig==2.3.0
     # via pytest
-ipython==9.11.0
+inquirerpy==0.3.4
+    # via huggingface-hub
+ipython==8.37.0
     # via meshcat
-ipython-pygments-lexers==1.1.1
-    # via ipython
 ischedule==1.2.7
     # via placo
 jedi==0.19.2
@@ -316,41 +303,40 @@ jinja2==3.1.6
     # via torch
 jsonlines==4.0.0
     # via lerobot
-jsonschema==4.26.0
+jsonschema==4.25.1
     # via nbformat
 jsonschema-specifications==2025.9.1
     # via jsonschema
 jupyter-core==5.9.1
     # via nbformat
-jupytext==1.19.1
+jupytext==1.18.1
     # via bddl
 kiwisolver==1.4.9
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
-lazy-loader==0.5
+lazy-loader==0.4
     # via scikit-image
-librt==0.8.1
-    # via mypy
-llvmlite==0.46.0
+libero @ git+https://github.com/huggingface/lerobot-libero.git@main
+    # via lerobot
+llvmlite==0.45.1
     # via numba
 lxml==6.0.2
     # via dm-control
-markdown==3.10.2
+markdown==3.9
     # via tensorboard
 markdown-it-py==4.0.0
     # via
     #   jupytext
     #   mdit-py-plugins
-    #   rich
 markupsafe==3.0.3
     # via
     #   jinja2
     #   werkzeug
-matplotlib==3.10.8
+matplotlib==3.10.7
     # via
-    #   hf-libero
     #   lerobot
+    #   libero
 matplotlib-inline==0.2.1
     # via ipython
 mdit-py-plugins==0.5.0
@@ -367,38 +353,36 @@ mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.5.0
+mujoco==3.3.7
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   hf-libero
+    #   libero
     #   metaworld
     #   robosuite
-multidict==6.7.1
+multidict==6.7.0
     # via
     #   aiohttp
     #   yarl
-multiprocess==0.70.18
+multiprocess==0.70.16
     # via datasets
-mypy==1.19.1
-    # via lerobot
 mypy-extensions==1.1.0
-    # via
-    #   mypy
-    #   typing-inspect
+    # via typing-inspect
 nbformat==5.10.4
     # via jupytext
-networkx==3.6.1
+networkx==3.4.2
     # via
     #   bddl
     #   scikit-image
     #   torch
-nodeenv==1.10.0
+ninja==1.13.0
+    # via lerobot
+nodeenv==1.9.1
     # via pre-commit
 num2words==0.5.14
     # via lerobot
-numba==0.64.0
+numba==0.62.1
     # via robosuite
 numpy==2.2.6
     # via
@@ -407,6 +391,7 @@ numpy==2.2.6
     #   cmeel-boost
     #   contourpy
     #   datasets
+    #   decord
     #   diffusers
     #   dm-control
     #   dm-env
@@ -414,10 +399,9 @@ numpy==2.2.6
     #   gymnasium
     #   h5py
     #   hebi-py
-    #   hf-libero
     #   imageio
     #   labmaze
-    #   lerobot
+    #   libero
     #   matplotlib
     #   meshcat
     #   metaworld
@@ -442,51 +426,49 @@ numpy==2.2.6
     #   torchvision
     #   transformers
     #   transforms3d
-nvidia-cublas-cu12==12.8.4.1
+nvidia-cublas-cu12==12.6.4.1
     # via
     #   nvidia-cudnn-cu12
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cuda-cupti-cu12==12.8.90
+nvidia-cuda-cupti-cu12==12.6.80
     # via torch
-nvidia-cuda-nvrtc-cu12==12.8.93
+nvidia-cuda-nvrtc-cu12==12.6.77
     # via torch
-nvidia-cuda-runtime-cu12==12.8.90
+nvidia-cuda-runtime-cu12==12.6.77
     # via torch
-nvidia-cudnn-cu12==9.10.2.21
+nvidia-cudnn-cu12==9.5.1.17
     # via torch
-nvidia-cufft-cu12==11.3.3.83
+nvidia-cufft-cu12==11.3.0.4
     # via torch
-nvidia-cufile-cu12==1.13.1.3
+nvidia-cufile-cu12==1.11.1.6
     # via torch
-nvidia-curand-cu12==10.3.9.90
+nvidia-curand-cu12==10.3.7.77
     # via torch
-nvidia-cusolver-cu12==11.7.3.90
+nvidia-cusolver-cu12==11.7.1.2
     # via torch
-nvidia-cusparse-cu12==12.5.8.93
+nvidia-cusparse-cu12==12.5.4.2
     # via
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cusparselt-cu12==0.7.1
+nvidia-cusparselt-cu12==0.6.3
     # via torch
-nvidia-nccl-cu12==2.27.5
+nvidia-nccl-cu12==2.26.2
     # via torch
-nvidia-nvjitlink-cu12==12.8.93
+nvidia-nvjitlink-cu12==12.6.85
     # via
     #   nvidia-cufft-cu12
     #   nvidia-cusolver-cu12
     #   nvidia-cusparse-cu12
     #   torch
-nvidia-nvshmem-cu12==3.4.5
-    # via torch
-nvidia-nvtx-cu12==12.8.90
+nvidia-nvtx-cu12==12.6.77
     # via torch
 omegaconf==2.3.0
     # via hydra-core
-opencv-python==4.13.0.92
+opencv-python==4.12.0.88
     # via
     #   gym-pusht
-    #   hf-libero
+    #   libero
     #   reachy2-sdk
     #   robosuite
 opencv-python-headless==4.12.0.88
@@ -505,7 +487,6 @@ packaging==25.0
     #   matplotlib
     #   peft
     #   pytest
-    #   qwen-vl-utils
     #   reachy2-sdk
     #   scikit-image
     #   tensorboard
@@ -516,21 +497,21 @@ pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.6
+parso==0.8.5
     # via jedi
-pathspec==1.0.4
-    # via mypy
-peft==0.18.1
+peft==0.17.1
     # via lerobot
 pexpect==4.9.0
     # via ipython
-pillow==12.1.1
+pfzy==0.3.4
+    # via inquirerpy
+pillow==12.0.0
     # via
     #   diffusers
     #   imageio
+    #   lerobot
     #   matplotlib
     #   meshcat
-    #   qwen-vl-utils
     #   rerun-sdk
     #   robosuite
     #   scikit-image
@@ -538,27 +519,28 @@ pillow==12.1.1
     #   torchvision
 pin==3.4.0
     # via placo
-placo==0.9.16
+placo==0.9.14
     # via lerobot
-platformdirs==4.9.4
+platformdirs==4.5.0
     # via
     #   jupyter-core
-    #   python-discovery
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.5.1
+pre-commit==4.3.0
     # via lerobot
 prompt-toolkit==3.0.52
-    # via ipython
+    # via
+    #   inquirerpy
+    #   ipython
 propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
-protobuf==6.31.1
+protobuf==6.31.0
     # via
     #   dm-control
     #   grpcio-tools
@@ -568,7 +550,7 @@ protobuf==6.31.1
     #   tensorboard
     #   tensorboardx
     #   wandb
-psutil==7.2.2
+psutil==7.1.1
     # via
     #   accelerate
     #   imageio
@@ -578,17 +560,17 @@ ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
     # via stack-data
-pyarrow==23.0.1
+pyarrow==21.0.0
     # via
     #   datasets
     #   rerun-sdk
-pycparser==3.0
+pycparser==2.23
     # via cffi
-pydantic==2.12.5
+pydantic==2.12.3
     # via
     #   fastapi
     #   wandb
-pydantic-core==2.41.5
+pydantic-core==2.41.4
     # via pydantic
 pygame==2.6.1
     # via
@@ -598,14 +580,12 @@ pygame==2.6.1
 pygments==2.19.2
     # via
     #   ipython
-    #   ipython-pygments-lexers
     #   pytest
-    #   rich
 pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.5.1
+pyngrok==7.4.1
     # via meshcat
 pynput==1.8.1
     # via
@@ -615,7 +595,7 @@ pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.3.2
+pyparsing==3.2.5
     # via
     #   dm-control
     #   matplotlib
@@ -641,16 +621,13 @@ pytest-timeout==2.4.0
     # via lerobot
 python-dateutil==2.9.0.post0
     # via
-    #   faker
     #   matplotlib
     #   pandas
-python-discovery==1.1.1
-    # via virtualenv
-python-dotenv==1.2.2
+python-dotenv==1.1.1
     # via uvicorn
 python-xlib==0.33
     # via pynput
-pytz==2026.1.post1
+pytz==2025.2
     # via pandas
 pyyaml==6.0.3
     # via
@@ -665,6 +642,7 @@ pyyaml==6.0.3
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
+    #   timm
     #   transformers
     #   uvicorn
     #   wandb
@@ -674,9 +652,7 @@ pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-qwen-vl-utils==0.0.14
-    # via lerobot
-reachy2-sdk==1.0.15
+reachy2-sdk==1.0.14
     # via lerobot
 reachy2-sdk-api==1.0.21
     # via reachy2-sdk
@@ -684,7 +660,7 @@ referencing==0.37.0
     # via
     #   jsonschema
     #   jsonschema-specifications
-regex==2026.2.28
+regex==2025.10.23
     # via
     #   diffusers
     #   transformers
@@ -693,62 +669,60 @@ requests==2.32.5
     #   datasets
     #   diffusers
     #   dm-control
-    #   qwen-vl-utils
+    #   huggingface-hub
     #   teleop
+    #   transformers
     #   wandb
-rerun-sdk==0.26.2
+rerun-sdk==0.26.1
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-rich==14.3.3
-    # via typer
 robomimic==0.2.0
-    # via hf-libero
+    # via libero
 robosuite==1.4.0
-    # via hf-libero
-rpds-py==0.30.0
+    # via libero
+rpds-py==0.28.0
     # via
     #   jsonschema
     #   referencing
-safetensors==0.7.0
+safetensors==0.6.2
     # via
     #   accelerate
     #   diffusers
     #   lerobot
     #   peft
+    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
     #   gym-pusht
     #   lerobot
-scipy==1.17.1
+scipy==1.15.3
     # via
     #   dm-control
-    #   lerobot
     #   metaworld
     #   robosuite
     #   scikit-image
-    #   torchdiffeq
-sentry-sdk==2.54.0
+sentry-sdk==2.42.1
     # via wandb
 shapely==2.1.2
     # via gym-pusht
-shellingham==1.5.4
-    # via typer
 six==1.17.0
     # via
     #   pynput
     #   python-dateutil
     #   python-xlib
-smmap==5.0.3
+smmap==5.0.2
     # via gitdb
+sniffio==1.3.1
+    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.52.1
+starlette==0.48.0
     # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.4
+teleop==0.1.2
     # via lerobot
 tensorboard==2.20.0
     # via robomimic
@@ -756,38 +730,46 @@ tensorboard-data-server==0.7.2
     # via tensorboard
 tensorboardx==2.6.4
     # via robomimic
-termcolor==3.3.0
+termcolor==3.1.0
     # via
     #   lerobot
     #   robomimic
 thop==0.1.1.post2209072238
-    # via hf-libero
-tifffile==2026.3.3
+    # via libero
+tifffile==2025.5.10
     # via scikit-image
-tokenizers==0.22.2
+timm==1.0.20
+    # via lerobot
+tokenizers==0.22.1
     # via transformers
 toml==0.10.2
     # via draccus
-torch==2.10.0
+tomli==2.3.0
+    # via
+    #   cmeel
+    #   coverage
+    #   jupytext
+    #   pytest
+torch==2.7.1
     # via
     #   accelerate
+    #   flash-attn
     #   lerobot
     #   peft
     #   robomimic
     #   thop
-    #   torchdiffeq
+    #   timm
     #   torchvision
-torchcodec==0.10.0
+torchcodec==0.5
     # via lerobot
-torchdiffeq==0.2.5
-    # via lerobot
-torchvision==0.25.0
+torchvision==0.22.1
     # via
     #   lerobot
     #   robomimic
-tornado==6.5.4
+    #   timm
+tornado==6.5.2
     # via meshcat
-tqdm==4.67.3
+tqdm==4.67.1
     # via
     #   datasets
     #   dm-control
@@ -801,29 +783,26 @@ traitlets==5.14.3
     #   jupyter-core
     #   matplotlib-inline
     #   nbformat
-transformers==5.3.0
+transformers==4.57.1
     # via
-    #   hf-libero
     #   lerobot
+    #   libero
     #   peft
 transforms3d==0.4.2
     # via teleop
-triton==3.6.0
+triton==3.3.1
     # via torch
-typer==0.24.1
-    # via
-    #   huggingface-hub
-    #   transformers
 typing-extensions==4.15.0
     # via
     #   aiosignal
     #   anyio
     #   etils
-    #   faker
+    #   exceptiongroup
     #   fastapi
     #   gymnasium
     #   huggingface-hub
-    #   mypy
+    #   ipython
+    #   multidict
     #   pydantic
     #   pydantic-core
     #   referencing
@@ -832,46 +811,46 @@ typing-extensions==4.15.0
     #   torch
     #   typing-inspect
     #   typing-inspection
+    #   uvicorn
+    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
 typing-inspection==0.4.2
-    # via
-    #   fastapi
-    #   pydantic
-tzdata==2025.3
+    # via pydantic
+tzdata==2025.2
     # via pandas
 u-msgpack-python==2.8.0
     # via meshcat
-urllib3==2.6.3
+urllib3==2.5.0
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.41.0
+uvicorn[standard]==0.38.0
     # via teleop
 uvloop==0.22.1
     # via uvicorn
-virtualenv==21.1.0
+virtualenv==20.35.3
     # via pre-commit
-wandb==0.24.2
+wandb==0.21.4
     # via
-    #   hf-libero
     #   lerobot
+    #   libero
 watchfiles==1.1.1
     # via uvicorn
-wcwidth==0.6.0
+wcwidth==0.2.14
     # via prompt-toolkit
 websocket-client==1.9.0
     # via teleop
-websockets==16.0
+websockets==15.0.1
     # via uvicorn
-werkzeug==3.1.6
+werkzeug==3.1.3
     # via tensorboard
-wrapt==2.1.2
+wrapt==2.0.0
     # via dm-tree
 xxhash==3.6.0
     # via datasets
-yarl==1.23.0
+yarl==1.22.0
     # via aiohttp
 zipp==3.23.0
     # via

requirements.in

@@ -1,9 +1,9 @@
 # requirements.in
 
-# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64).
-# Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64
+# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.0.1 25A362 arm64).
+# Darwin MacBook-Pro.local 25.0.0 Darwin Kernel Version 25.0.0: Wed Sep 17 21:42:08 PDT 2025; root:xnu-12377.1.9~141/RELEASE_ARM64_T8132 arm64
 
-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64).
-# Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
+# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.3 LTS x86_64).
+# Linux mlerobot-linux 6.14.0-33-generic #33~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Sep 19 17:02:30 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
 
 -e .[all]

src/lerobot/async_inference/helpers.py

@@ -23,7 +23,7 @@ from typing import Any
 import torch
 
 from lerobot.configs.types import PolicyFeature
-from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
 
 # NOTE: Configs need to be loaded for the client to be able to instantiate the policy config
 from lerobot.policies import (  # noqa: F401

src/lerobot/async_inference/policy_server.py

@@ -39,13 +39,15 @@ import grpc
 import torch
 
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.processor import PolicyProcessorPipeline
+from lerobot.processor import (
+    PolicyAction,
+    PolicyProcessorPipeline,
+)
 from lerobot.transport import (
     services_pb2,  # type: ignore
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import receive_bytes_in_chunks
-from lerobot.types import PolicyAction
 
 from .configs import PolicyServerConfig
 from .constants import SUPPORTED_POLICIES

src/lerobot/async_inference/robot_client.py

@@ -63,9 +63,9 @@ from lerobot.transport import (
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
-from lerobot.utils.import_utils import register_third_party_plugins
 
 from .configs import RobotClientConfig
+from .constants import SUPPORTED_ROBOTS
 from .helpers import (
     Action,
     FPSTracker,
@@ -485,9 +485,8 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    # TODO: Assert if checking robot support is still needed with the plugin system
-    # if cfg.robot.type not in SUPPORTED_ROBOTS:
-    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    if cfg.robot.type not in SUPPORTED_ROBOTS:
+        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -513,5 +512,4 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
-    register_third_party_plugins()
     async_client()  # run the client

src/lerobot/cameras/__init__.py

@@ -13,5 +13,5 @@
 # limitations under the License.
 
 from .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Rotation
 from .utils import make_cameras_from_configs

src/lerobot/cameras/camera.py

@@ -150,7 +150,7 @@ class Camera(abc.ABC):
         """
         pass
 
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/configs.py

@@ -25,10 +25,6 @@ class ColorMode(str, Enum):
     RGB = "rgb"
     BGR = "bgr"
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 class Cv2Rotation(int, Enum):
     NO_ROTATION = 0
@@ -36,25 +32,6 @@ class Cv2Rotation(int, Enum):
     ROTATE_180 = 180
     ROTATE_270 = -90
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")
-
-
-# Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
-class Cv2Backends(int, Enum):
-    ANY = 0
-    V4L2 = 200
-    DSHOW = 700
-    PVAPI = 800
-    ANDROID = 1000
-    AVFOUNDATION = 1200
-    MSMF = 1400
-
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 @dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus

src/lerobot/cameras/opencv/camera_opencv.py

@@ -32,11 +32,10 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_rotation
+from ..utils import get_cv2_backend, get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -118,7 +117,7 @@ class OpenCVCamera(Camera):
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = config.backend
+        self.backend: int = get_cv2_backend()
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -133,7 +132,6 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
@@ -150,6 +148,8 @@ class OpenCVCamera(Camera):
             ConnectionError: If the specified camera index/path is not found or fails to open.
             RuntimeError: If the camera opens but fails to apply requested settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         # Use 1 thread for OpenCV operations to avoid potential conflicts or
         # blocking in multi-threaded applications, especially during data collection.
@@ -178,7 +178,6 @@ class OpenCVCamera(Camera):
 
         logger.info(f"{self} connected.")
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """
         Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -198,6 +197,8 @@ class OpenCVCamera(Camera):
                           to the requested value.
             DeviceNotConnectedError: If the camera is not connected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
         # Set FOURCC first (if specified) as it can affect available FPS/resolution options
         if self.config.fourcc is not None:
@@ -347,7 +348,6 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -374,6 +374,9 @@ class OpenCVCamera(Camera):
                 f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -487,7 +490,6 @@ class OpenCVCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -510,6 +512,8 @@ class OpenCVCamera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
@@ -529,8 +533,7 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -545,6 +548,8 @@ class OpenCVCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -15,9 +15,9 @@
 from dataclasses import dataclass
 from pathlib import Path
 
-from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
 
 
 @CameraConfig.register_subclass("opencv")
@@ -50,7 +50,6 @@ class OpenCVCameraConfig(CameraConfig):
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
         fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
-        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -63,12 +62,22 @@ class OpenCVCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
     fourcc: str | None = None
-    backend: Cv2Backends = Cv2Backends.ANY
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
-        self.rotation = Cv2Rotation(self.rotation)
-        self.backend = Cv2Backends(self.backend)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
             raise ValueError(

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -74,4 +74,7 @@ class Reachy2CameraConfig(CameraConfig):
                 f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
             )
 
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -32,7 +32,6 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
-from lerobot.utils.decorators import check_if_not_connected
 from lerobot.utils.import_utils import _reachy2_sdk_available
 
 if TYPE_CHECKING or _reachy2_sdk_available:
@@ -124,7 +123,6 @@ class Reachy2Camera(Camera):
         """
         raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -138,6 +136,9 @@ class Reachy2Camera(Camera):
         """
         start_time = time.perf_counter()
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.cam_manager is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
@@ -183,7 +184,6 @@ class Reachy2Camera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Same as read()
@@ -197,11 +197,12 @@ class Reachy2Camera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         return self.read()
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -218,6 +219,8 @@ class Reachy2Camera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.latest_frame is None or self.latest_timestamp is None:
             raise RuntimeError(f"{self} has not captured any frames yet.")
@@ -230,7 +233,6 @@ class Reachy2Camera(Camera):
 
         return self.latest_frame
 
-    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -238,6 +240,8 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} not connected.")
 
         if self.cam_manager is not None:
             self.cam_manager.disconnect()

src/lerobot/cameras/realsense/camera_realsense.py

@@ -30,8 +30,7 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -153,7 +152,6 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
@@ -171,6 +169,8 @@ class RealSenseCamera(Camera):
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         self.rs_pipeline = rs.pipeline()
         rs_config = rs.config()
@@ -290,7 +290,6 @@ class RealSenseCamera(Camera):
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
@@ -300,6 +299,8 @@ class RealSenseCamera(Camera):
         Raises:
             DeviceNotConnectedError: If device is not connected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -319,7 +320,6 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
@@ -345,6 +345,9 @@ class RealSenseCamera(Camera):
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -371,7 +374,6 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
@@ -401,6 +403,9 @@ class RealSenseCamera(Camera):
                 f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -529,7 +534,6 @@ class RealSenseCamera(Camera):
             self.new_frame_event.clear()
 
     # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
@@ -552,6 +556,8 @@ class RealSenseCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
@@ -572,8 +578,7 @@ class RealSenseCamera(Camera):
         return frame
 
     # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -588,6 +593,8 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -60,8 +60,20 @@ class RealSenseCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
-        self.rotation = Cv2Rotation(self.rotation)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):

src/lerobot/cameras/utils.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import platform
 from typing import cast
 
 from lerobot.utils.import_utils import make_device_from_device_class
@@ -67,3 +68,14 @@ def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
         return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
+
+
+def get_cv2_backend() -> int:
+    import cv2
+
+    if platform.system() == "Windows":
+        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
+    # elif platform.system() == "Darwin":  # macOS
+    #     return cv2.CAP_AVFOUNDATION
+    else:  # Linux and others
+        return int(cv2.CAP_ANY)

src/lerobot/cameras/zmq/camera_zmq.py

@@ -34,8 +34,7 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -105,7 +104,6 @@ class ZMQCamera(Camera):
         """Checks if the ZMQ socket is initialized and connected."""
         return self._connected and self.context is not None and self.socket is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """Connect to ZMQ camera server.
 
@@ -113,6 +111,8 @@ class ZMQCamera(Camera):
             warmup (bool): If True, waits for the camera to provide at least one
                            valid frame before returning. Defaults to True.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         logger.info(f"Connecting to {self}...")
 
@@ -181,7 +181,7 @@ class ZMQCamera(Camera):
         try:
             message = self.socket.recv_string()
         except Exception as e:
-            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
+            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
             if type(e).__name__ == "Again":
                 raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
             raise
@@ -211,7 +211,6 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -229,6 +228,9 @@ class ZMQCamera(Camera):
                 f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
             )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
@@ -299,7 +301,6 @@ class ZMQCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -316,6 +317,8 @@ class ZMQCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread is not running.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
@@ -332,7 +335,6 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
     def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
@@ -348,6 +350,8 @@ class ZMQCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -32,7 +32,10 @@ class ZMQCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
 
         if self.timeout_ms <= 0:
             raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")

src/lerobot/cameras/zmq/image_server.py

@@ -23,7 +23,6 @@ import base64
 import contextlib
 import json
 import logging
-import threading
 import time
 from collections import deque
 
@@ -43,57 +42,10 @@ def encode_image(image: np.ndarray, quality: int = 80) -> str:
     return base64.b64encode(buffer).decode("utf-8")
 
 
-class CameraCaptureThread:
-    """Background thread that continuously captures and encodes frames from a camera."""
-
-    def __init__(self, camera: OpenCVCamera, name: str):
-        self.camera = camera
-        self.name = name
-        self.latest_encoded: str | None = None  # Pre-encoded JPEG as base64
-        self.latest_timestamp: float = 0.0
-        self.frame_lock = threading.Lock()
-        self.running = False
-        self.thread: threading.Thread | None = None
-
-    def start(self):
-        """Start the capture thread."""
-        self.running = True
-        self.thread = threading.Thread(target=self._capture_loop, daemon=True)
-        self.thread.start()
-
-    def stop(self):
-        """Stop the capture thread."""
-        self.running = False
-        if self.thread:
-            self.thread.join(timeout=1.0)
-
-    def _capture_loop(self):
-        """Continuously capture and encode frames at the camera's native rate."""
-        while self.running:
-            try:
-                frame = self.camera.read()  # Blocks at camera's native rate
-                timestamp = time.time()
-                # Encode immediately in capture thread (this is the slow part)
-                encoded = encode_image(frame)
-                with self.frame_lock:
-                    self.latest_encoded = encoded
-                    self.latest_timestamp = timestamp
-            except Exception as e:
-                logger.warning(f"Camera {self.name} capture error: {e}")
-                time.sleep(0.01)
-
-    def get_latest(self) -> tuple[str | None, float]:
-        """Get the latest encoded frame and its timestamp."""
-        with self.frame_lock:
-            return self.latest_encoded, self.latest_timestamp
-
-
 class ImageServer:
     def __init__(self, config: dict, port: int = 5555):
-        # fps controls the publish loop rate (how often frames are sent over ZMQ), not the camera capture rate
         self.fps = config.get("fps", 30)
         self.cameras: dict[str, OpenCVCamera] = {}
-        self.capture_threads: dict[str, CameraCaptureThread] = {}
 
         for name, cfg in config.get("cameras", {}).items():
             shape = cfg.get("shape", [480, 640])
@@ -109,10 +61,6 @@ class ImageServer:
             self.cameras[name] = camera
             logger.info(f"Camera {name}: {shape[1]}x{shape[0]}")
 
-            # Create capture thread for this camera
-            capture_thread = CameraCaptureThread(camera, name)
-            self.capture_threads[name] = capture_thread
-
         # ZMQ PUB socket
         self.context = zmq.Context()
         self.socket = self.context.socket(zmq.PUB)
@@ -125,18 +73,6 @@ class ImageServer:
     def run(self):
         frame_count = 0
         frame_times = deque(maxlen=60)
-        last_published_ts: dict[str, float] = {}
-
-        # Start all capture threads
-        for capture_thread in self.capture_threads.values():
-            capture_thread.start()
-
-        # Wait for first frames to be captured and encoded
-        logger.info("Waiting for cameras to start capturing...")
-        for name, capture_thread in self.capture_threads.items():
-            while capture_thread.get_latest()[0] is None:
-                time.sleep(0.01)
-            logger.info(f"Camera {name} ready (capture + encode in background)")
 
         try:
             while True:
@@ -144,12 +80,10 @@ class ImageServer:
 
                 # Build message
                 message = {"timestamps": {}, "images": {}}
-                for name, capture_thread in self.capture_threads.items():
-                    encoded, timestamp = capture_thread.get_latest()
-                    if encoded is not None and timestamp > last_published_ts.get(name, 0.0):
-                        message["timestamps"][name] = timestamp
-                        message["images"][name] = encoded
-                        last_published_ts[name] = timestamp
+                for name, cam in self.cameras.items():
+                    frame = cam.read()  # Returns RGB
+                    message["timestamps"][name] = time.time()
+                    message["images"][name] = encode_image(frame)
 
                 # Send as JSON string (suppress if buffer full)
                 with contextlib.suppress(zmq.Again):
@@ -168,8 +102,6 @@ class ImageServer:
         except KeyboardInterrupt:
             pass
         finally:
-            for capture_thread in self.capture_threads.values():
-                capture_thread.stop()
             for cam in self.cameras.values():
                 cam.disconnect()
             self.socket.close()

src/lerobot/configs/default.py

@@ -27,7 +27,7 @@ class DatasetConfig:
     # "dataset_index" into the returned item. The index mapping is made according to the order in which the
     # datasets are provided.
     repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
@@ -36,16 +36,6 @@ class DatasetConfig:
     video_backend: str = field(default_factory=get_safe_default_codec)
     streaming: bool = False
 
-    def __post_init__(self) -> None:
-        if self.episodes is not None:
-            if any(ep < 0 for ep in self.episodes):
-                raise ValueError(
-                    f"Episode indices must be non-negative, got: {[ep for ep in self.episodes if ep < 0]}"
-                )
-            if len(self.episodes) != len(set(self.episodes)):
-                duplicates = sorted({ep for ep in self.episodes if self.episodes.count(ep) > 1})
-                raise ValueError(f"Episode indices contain duplicates: {duplicates}")
-
 
 @dataclass
 class WandBConfig:
@@ -57,7 +47,6 @@ class WandBConfig:
     notes: str | None = None
     run_id: str | None = None
     mode: str | None = None  # Allowed values: 'online', 'offline' 'disabled'. Defaults to 'online'
-    add_tags: bool = True  # If True, save configuration as tags in the WandB run.
 
 
 @dataclass

src/lerobot/configs/policies.py

@@ -30,8 +30,8 @@ from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
 from lerobot.utils.constants import ACTION, OBS_STATE
-from lerobot.utils.device_utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 
 T = TypeVar("T", bound="PreTrainedConfig")
 logger = getLogger(__name__)

src/lerobot/configs/train.py

@@ -50,9 +50,6 @@ class TrainPipelineConfig(HubMixin):
     # `seed` is used for training (eg: model initialization, dataset shuffling)
     # AND for the evaluation environments.
     seed: int | None = 1000
-    # Set to True to use deterministic cuDNN algorithms for reproducibility.
-    # This disables cudnn.benchmark and may reduce training speed by ~10-20 percent.
-    cudnn_deterministic: bool = False
     # Number of workers for the dataloader.
     num_workers: int = 4
     batch_size: int = 8

src/lerobot/data_processing/annotations/annotate_libero.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+# Example script to run synthetic data generation with Qwen VLM
+# This generates user prompts and robot utterances for hierarchical policy training
+
+# Configuration
+REPO_ID="lerobot/libero_10"
+MODEL="Qwen/Qwen3-VL-30B-A3B-Instruct"
+# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
+
+
+OUTPUT_DIR="/fsx/jade_choghari/outputs/libero-10-annotate-high"
+
+BATCH_SIZE=16
+TEMPERATURE=0.9
+SAMPLE_INTERVAL=5.0  # generate dialogue every 1 second (all episodes processed)
+
+# Run subtask annotation
+# python /admin/home/jade_choghari/lerobot/src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
+#     --repo-id "$REPO_ID" \
+#     --video-key observation.images.image \
+#     --output-dir "$OUTPUT_DIR" \
+#     --skip-existing \
+#     --output-repo-id "jadechoghari/libero10-annotate" \
+#     --batch-size "$BATCH_SIZE" \
+# run synthetic data generation (all episodes processed)
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --temperature "$TEMPERATURE" \
+#     --batch-size "$BATCH_SIZE" \
+#     --sample-interval "$SAMPLE_INTERVAL" \
+#     --image-key observation.images.base \
+#     --num-image-views-per-sample 1
+
+# for faster testing, increase sample interval:
+# --sample-interval 5.0  # Samples every 5 seconds (much faster)
+
+# to push to hub after generation:
+# add --push-to-hub flag
+
+# efficient batch processing: 4 episodes at once
+python src/lerobot/data_processing/annotations/high_level_annotate.py \
+    --data-dir "/fsx/jade_choghari/outputs/libero-10-annotate" \
+    --output-dir "$OUTPUT_DIR" \
+    --video-mode \
+    --video-key observation.images.image \
+    --video-batch-size "$BATCH_SIZE" \
+    --sample-interval 5.0

src/lerobot/data_processing/annotations/load_lerobot_high.py

@@ -0,0 +1,52 @@
+import torch
+from huggingface_hub import HfApi
+
+import lerobot
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.configs.policies import PreTrainedConfig
+
+# /fsx/jade_choghari/data/libero_10_subtasks_kw_converted
+dataset = LeRobotDataset(repo_id="lerobot/libero_10_image_subtask")
+
+dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=0,
+        batch_size=2,
+        shuffle=True,
+)
+
+cfg = PreTrainedConfig.from_pretrained(
+    pretrained_name_or_path="/fsx/jade_choghari/models/pi05-base",
+)
+cfg.dtype = "bfloat16"
+
+pre_processor, post_processor = make_pre_post_processors(
+    policy_cfg=cfg,
+    pretrained_path="/fsx/jade_choghari/models/pi05-base",
+)
+batch = next(iter(dataloader))
+breakpoint()
+batch1 = pre_processor(batch)
+breakpoint()
+print(batch.keys())
+# print(batch['task_index_high_level'].shape)
+# print(batch['task_index_high_level'])
+# print(batch['user_prompt'][0])
+# print(batch['robot_utterance'][0])
+# print(batch['task'][0])
+
+valid_episode_list = []
+for episode_idx in range(len(dataset.meta.episodes)):
+        subtask_index = dataset[episode_idx]["subtask_index"]
+        valid_episode_list.append(episode_idx)
+
+print(len(valid_episode_list))
+
+# read this parquet /fsx/jade_choghari/outputs/pgen_annotations1/meta/tasks.parquett
+# import pandas as pd
+# tasks_df = pd.read_parquet('/fsx/jade_choghari/outputs/pgen_annotations1/meta/tasks.parquet')
+
+# # print all
+# print(tasks_df.columns)
+# breakpoint()

src/lerobot/data_processing/annotations/run_pgen.sh

@@ -0,0 +1,74 @@
+#!/bin/bash
+
+# Example script to run synthetic data generation with Qwen VLM
+# This generates user prompts and robot utterances for hierarchical policy training
+
+# Configuration
+REPO_ID="jadechoghari/piper-demo-20260205_103303"
+# MODEL="Qwen/Qwen3-VL-30B-A3B-Thinking"
+MODEL="Qwen/Qwen3.5-27B"
+# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
+
+
+OUTPUT_DIR="/fsx/jade_choghari/outputs/collect-data-pgen_new"
+
+BATCH_SIZE=2
+TEMPERATURE=0.9
+SAMPLE_INTERVAL=5.0  # generate dialogue every 1 second (all episodes processed)
+
+# Run subtask annotation.
+# To use closed-vocabulary labels, add a line: --subtask-labels "label1" "label2" ...
+# Example (add backslash after "$MODEL" and uncomment the next line):
+#   --model "$MODEL" \
+#   --subtask-labels "pick_up_yellow_nut_bar" "pick_up_cake" "pick_up_biscuit_pack" "pick_up_soda_can"
+python /home/lerobot/src/lerobot/data_processing/annotations/subtask_annotate.py \
+    --repo-id "$REPO_ID" \
+    --video-key observation.images.top \
+    --output-dir "$OUTPUT_DIR" \
+    --output-repo-id "jadechoghari/piper-demo-annotated1" \
+    --push-to-hub \
+    --no-timer-overlay \
+    --model "$MODEL" \
+    --subtask-labels "pick_up_yellow_nut_bar" "pick_up_cake" "pick_up_biscuit_pack" "pick_up_soda_can" \
+    --batch-size 2
+
+# Run subtask annotation (image-window: frames as images for better accuracy)
+# python /admin/home/jade_choghari/lerobot/src/lerobot/data_processing/annotations/subtask_annotate_image.py \
+#     --repo-id "$REPO_ID" \
+#     --camera-key observation.images.wrist \
+#     --output-dir "$OUTPUT_DIR" \
+#     --output-repo-id "jadechoghari/piper-demo-annotated1-image" \
+#     --push-to-hub \
+#     --model "$MODEL" \
+#     --window-size 184 \
+#     --max-frames-per-window 16 \
+#     --subtask-labels "pick_up_yellow_nut_bar" "pick_up_cake" "pick_up_biscuit_pack" "pick_up_soda_can" \
+#     --batch-size 2
+
+    
+# run synthetic data generation (all episodes processed)
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --temperature "$TEMPERATURE" \
+#     --batch-size "$BATCH_SIZE" \
+#     --sample-interval "$SAMPLE_INTERVAL" \
+#     --image-key observation.images.base \
+#     --num-image-views-per-sample 1
+
+# for faster testing, increase sample interval:
+# --sample-interval 5.0  # Samples every 5 seconds (much faster)
+
+# to push to hub after generation:
+# add --push-to-hub flag
+
+# efficient batch processing: 4 episodes at once
+# python examples/dataset/annotate_pgen.py \
+#     --repo-id "$REPO_ID" \
+#     --model "$MODEL" \
+#     --output-dir "$OUTPUT_DIR" \
+#     --video-mode \
+#     --video-key observation.images.up \
+#     --video-batch-size "$BATCH_SIZE" \
+#     --sample-interval 1.0

src/lerobot/data_processing/annotations/subtask_annotate_image.py

@@ -0,0 +1,561 @@
+#!/usr/bin/env python
+
+# Copyright 2025 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.
+
+"""
+Image-window subtask annotation for LeRobot datasets using Qwen VLMs.
+
+This script assigns a subtask to each window of consecutive frames by sending
+those frames as images to the VLM (instead of a video) for better accuracy.
+Supports Qwen2-VL and Qwen3-VL (same models as subtask_annotate.py).
+
+Pipeline:
+1. Load a LeRobot dataset (local or Hub).
+2. For each episode, slide a window over frame indices.
+3. For each window, load the corresponding images (from image_key or decoded video_key).
+4. Send the window of images to Qwen2-VL with the same skill prompt; get one subtask name.
+5. Assign that subtask to all frames in the window.
+6. Write subtasks.parquet and add subtask_index via add_features (same as subtask_annotate).
+
+Usage:
+  python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+    --data-dir /path/to/dataset --camera-key observation.images.base \\
+    --window-size 8 --stride 8 --output-dir ./output
+"""
+
+from __future__ import annotations
+
+import argparse
+import random
+import textwrap
+from pathlib import Path
+
+import numpy as np
+import PIL.Image
+import torch
+from rich.console import Console
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Reuse data structures and save/load from the video-based annotator
+from lerobot.data_processing.annotations.subtask_annotate import (
+    EpisodeSkills,
+    Skill,
+    load_skill_annotations,
+    save_skill_annotations,
+)
+
+
+def create_window_skill_prompt(
+    coarse_goal: str | None = None,
+    subtask_labels: list[str] | None = None,
+) -> str:
+    """Prompt for labeling a single window of frames with one atomic skill.
+    If subtask_labels are provided, the model must choose exactly one from that list.
+    """
+    goal_context = f'The overall goal is: "{coarse_goal}".\n\n' if coarse_goal else ""
+    if subtask_labels:
+        labels_list = ", ".join(f'"{l}"' for l in subtask_labels)
+        label_instruction = (
+            f"You must choose exactly ONE skill from this list: [{labels_list}]. "
+            "Do not create new labels. Reply with only that label.\n\n"
+        )
+    else:
+        label_instruction = ""
+    return textwrap.dedent(f"""\
+        # Role
+        You are a Robotics Vision System that labels short clips from robot manipulation demonstrations.
+
+        # Task
+        {goal_context}{label_instruction}The following images are consecutive frames from a single short clip of a robot demonstration.
+        What single atomic manipulation skill is being performed in this clip?
+
+        # Requirements
+        - Reply with ONLY one short skill name (e.g. "pick up object", "move arm left", "release gripper").
+        - No explanation, no timestamps, no JSON. Just the skill name.
+        """).strip()
+
+
+def _run_image_segmenter(
+    self,
+    images: list[PIL.Image.Image],
+    coarse_goal: str | None,
+    subtask_labels: list[str] | None = None,
+) -> str:
+    """Shared inference for Qwen2-VL and Qwen3-VL image window labeling."""
+    prompt = create_window_skill_prompt(coarse_goal, subtask_labels)
+    content = []
+    for img in images:
+        content.append({"type": "image", "image": img})
+    content.append({"type": "text", "text": "What single atomic skill is shown in these frames? Reply with only the skill name."})
+
+    messages = [
+        {"role": "system", "content": [{"type": "text", "text": prompt}]},
+        {"role": "user", "content": content},
+    ]
+    text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    image_inputs, video_inputs = self.process_vision_info(messages)
+    inputs = self.processor(
+        text=[text],
+        images=image_inputs,
+        videos=video_inputs,
+        padding=True,
+        return_tensors="pt",
+    ).to(self.device)
+
+    with torch.no_grad():
+        generated_ids = self.model.generate(**inputs, max_new_tokens=128, do_sample=False)
+
+    response = self.processor.batch_decode(
+        [out[len(inp) :] for inp, out in zip(inputs.input_ids, generated_ids)],
+        skip_special_tokens=True,
+    )[0].strip()
+    skill_name = response.split("\n")[0].strip().strip('."')
+    return skill_name if skill_name else "unknown"
+
+
+def _run_image_segmenter_batch(
+    self,
+    batch_images: list[list[PIL.Image.Image]],
+    coarse_goal: str | None,
+    subtask_labels: list[str] | None = None,
+) -> list[str]:
+    """Run VLM on multiple windows at once; returns one skill name per window."""
+    if not batch_images:
+        return []
+    prompt = create_window_skill_prompt(coarse_goal, subtask_labels)
+    all_texts = []
+    all_image_inputs = []
+    all_video_inputs = []
+    for images in batch_images:
+        content = []
+        for img in images:
+            content.append({"type": "image", "image": img})
+        content.append({"type": "text", "text": "What single atomic skill is shown in these frames? Reply with only the skill name."})
+        messages = [
+            {"role": "system", "content": [{"type": "text", "text": prompt}]},
+            {"role": "user", "content": content},
+        ]
+        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        image_inputs, video_inputs = self.process_vision_info(messages)
+        all_texts.append(text)
+        if image_inputs is not None:
+            all_image_inputs.extend(image_inputs if isinstance(image_inputs, list) else [image_inputs])
+        if video_inputs is not None:
+            all_video_inputs.extend(video_inputs if isinstance(video_inputs, list) else [video_inputs])
+    inputs = self.processor(
+        text=all_texts,
+        images=all_image_inputs if all_image_inputs else None,
+        videos=all_video_inputs if all_video_inputs else None,
+        padding=True,
+        return_tensors="pt",
+    ).to(self.device)
+    with torch.no_grad():
+        generated_ids = self.model.generate(**inputs, max_new_tokens=128, do_sample=False)
+    responses = self.processor.batch_decode(
+        [out[len(inp) :] for inp, out in zip(inputs.input_ids, generated_ids)],
+        skip_special_tokens=True,
+    )
+    return [
+        (r.split("\n")[0].strip().strip('."') or "unknown")
+        for r in responses
+    ]
+
+
+class Qwen2VLImageSegmenter:
+    """Uses Qwen2-VL to assign one skill name to a window of images (same model as subtask_annotate)."""
+
+    def __init__(self, model_name: str, device: str = "cuda", torch_dtype: torch.dtype = torch.bfloat16):
+        from qwen_vl_utils import process_vision_info
+        from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+
+        self.console = Console()
+        self.device = device
+        self.process_vision_info = process_vision_info
+        self.console.print(f"[cyan]Loading Qwen2-VL for image-window labeling: {model_name}...[/cyan]")
+        self.model = Qwen2VLForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=torch_dtype, device_map=device, trust_remote_code=True
+        )
+        self.processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
+        self.console.print(f"[green]✓ Model loaded on {device}[/green]")
+
+    def segment_skill_from_images(
+        self,
+        images: list[PIL.Image.Image],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> str:
+        """Return a single skill name for the given window of images."""
+        return _run_image_segmenter(self, images, coarse_goal, subtask_labels)
+
+    def segment_skill_from_images_batch(
+        self,
+        batch_images: list[list[PIL.Image.Image]],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[str]:
+        """Return one skill name per window; processes multiple windows in one forward pass."""
+        return _run_image_segmenter_batch(self, batch_images, coarse_goal, subtask_labels)
+
+
+class Qwen3VLImageSegmenter:
+    """Uses Qwen3-VL (MoE) to assign one skill name to a window of images."""
+
+    def __init__(self, model_name: str, device: str = "cuda", torch_dtype: torch.dtype = torch.bfloat16):
+        from qwen_vl_utils import process_vision_info
+        from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration
+
+        self.console = Console()
+        self.device = device
+        self.process_vision_info = process_vision_info
+        self.console.print(f"[cyan]Loading Qwen3-VL for image-window labeling: {model_name}...[/cyan]")
+        self.model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=torch_dtype, device_map=device, trust_remote_code=True
+        )
+        self.processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
+        self.console.print(f"[green]✓ Model loaded on {device}[/green]")
+
+    def segment_skill_from_images(
+        self,
+        images: list[PIL.Image.Image],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> str:
+        """Return a single skill name for the given window of images."""
+        return _run_image_segmenter(self, images, coarse_goal, subtask_labels)
+
+    def segment_skill_from_images_batch(
+        self,
+        batch_images: list[list[PIL.Image.Image]],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[str]:
+        """Return one skill name per window; processes multiple windows in one forward pass."""
+        return _run_image_segmenter_batch(self, batch_images, coarse_goal, subtask_labels)
+
+
+def get_image_segmenter(
+    model_name: str,
+    device: str = "cuda",
+    torch_dtype: torch.dtype = torch.bfloat16,
+):
+    """Return the appropriate image-window segmenter for the model (Qwen2-VL or Qwen3-VL)."""
+    model_lower = model_name.lower()
+    if "qwen3" in model_lower:
+        return Qwen3VLImageSegmenter(model_name, device, torch_dtype)
+    return Qwen2VLImageSegmenter(model_name, device, torch_dtype)
+
+
+def frame_to_pil(frame_value) -> PIL.Image.Image:
+    """Convert a single frame from dataset (tensor or PIL or path) to PIL.Image."""
+    if isinstance(frame_value, PIL.Image.Image):
+        return frame_value
+    if isinstance(frame_value, (str, Path)):
+        return PIL.Image.open(frame_value).convert("RGB")
+    if hasattr(frame_value, "numpy"):
+        arr = frame_value.numpy()
+    else:
+        arr = np.asarray(frame_value)
+    if arr.ndim == 3 and arr.shape[0] in (1, 3, 4):
+        arr = np.transpose(arr, (1, 2, 0))
+    if arr.dtype == np.float32 or arr.dtype == np.float64:
+        arr = (np.clip(arr, 0, 1) * 255).astype(np.uint8)
+    elif arr.dtype != np.uint8:
+        arr = np.clip(arr, 0, 255).astype(np.uint8)
+    if arr.shape[-1] == 1:
+        arr = np.repeat(arr, 3, axis=-1)
+    return PIL.Image.fromarray(arr)
+
+
+def _sample_window_indices(window_length: int, max_frames: int) -> list[int]:
+    """Return indices into a window of length window_length, at most max_frames, in order.
+    If window_length <= max_frames, returns range(window_length).
+    Otherwise returns sorted random sample of max_frames indices (temporal order preserved).
+    """
+    if max_frames <= 0 or window_length <= max_frames:
+        return list(range(window_length))
+    return sorted(random.sample(range(window_length), max_frames))
+
+
+class SkillAnnotatorImage:
+    """Annotates episodes by sliding a window over frames and labeling each window with the VLM."""
+
+    def __init__(
+        self,
+        segmenter: Qwen2VLImageSegmenter | Qwen3VLImageSegmenter,
+        window_size: int = 8,
+        stride: int | None = None,
+        batch_size: int = 1,
+        max_frames_per_window: int | None = None,
+        console: Console | None = None,
+    ):
+        self.segmenter = segmenter
+        self.window_size = window_size
+        self.stride = stride if stride is not None else window_size
+        self.batch_size = max(1, batch_size)
+        self.max_frames_per_window = max_frames_per_window
+        self.console = console or Console()
+
+    def annotate_dataset(
+        self,
+        dataset: LeRobotDataset,
+        camera_key: str,
+        episodes: list[int] | None = None,
+        skip_existing: bool = False,
+        subtask_labels: list[str] | None = None,
+    ) -> dict[int, EpisodeSkills]:
+        """Annotate episodes using image windows. camera_key can be an image_key or video_key."""
+        episode_indices = episodes or list(range(dataset.meta.total_episodes))
+        coarse_goal = self._get_coarse_goal(dataset)
+        annotations: dict[int, EpisodeSkills] = {}
+
+        if skip_existing:
+            existing = load_skill_annotations(dataset.root)
+            if existing and existing.get("episodes"):
+                existing_eps = {int(k) for k in existing["episodes"] if existing["episodes"][k].get("skills")}
+                episode_indices = [i for i in episode_indices if i not in existing_eps]
+
+        for ep_idx in episode_indices:
+            try:
+                skills = self._annotate_episode(
+                    dataset, ep_idx, camera_key, coarse_goal, subtask_labels
+                )
+                if skills:
+                    annotations[ep_idx] = EpisodeSkills(
+                        episode_index=ep_idx,
+                        description=coarse_goal,
+                        skills=skills,
+                    )
+                    self.console.print(f"[green]✓ Episode {ep_idx}: {len(skills)} window skills[/green]")
+                else:
+                    self.console.print(f"[yellow]⚠ Episode {ep_idx}: no skills[/yellow]")
+            except Exception as e:
+                self.console.print(f"[red]Episode {ep_idx} failed: {e}[/red]")
+
+        return annotations
+
+    def _get_coarse_goal(self, dataset: LeRobotDataset) -> str:
+        if dataset.meta.tasks is not None and len(dataset.meta.tasks) > 0:
+            return str(dataset.meta.tasks.index[0])
+        return "Perform the demonstrated manipulation task."
+
+    def _annotate_episode(
+        self,
+        dataset: LeRobotDataset,
+        episode_index: int,
+        camera_key: str,
+        coarse_goal: str,
+        subtask_labels: list[str] | None = None,
+    ) -> list[Skill]:
+        ep = dataset.meta.episodes[episode_index]
+        ep_from = int(ep["dataset_from_index"])
+        ep_to = int(ep["dataset_to_index"])
+        length = ep_to - ep_from
+        fps = dataset.meta.fps
+        if length == 0:
+            return []
+
+        # Collect full windows: (images, t_start, t_end) using frame timestamps.
+        # If max_frames_per_window is set and window is larger, sample that many frames (order preserved).
+        window_specs: list[tuple[list[PIL.Image.Image], float, float]] = []
+        start = 0
+        while start + self.window_size <= length:
+            offsets = _sample_window_indices(
+                self.window_size,
+                self.max_frames_per_window or self.window_size,
+            )
+            frame_indices = [ep_from + start + i for i in offsets]
+            images = []
+            t_start = float(dataset[frame_indices[0]]["timestamp"].item())
+            for idx in frame_indices:
+                item = dataset[idx]
+                images.append(frame_to_pil(item[camera_key]))
+            t_end = t_start + self.window_size / fps
+            window_specs.append((images, t_start, t_end))
+            start += self.stride
+
+        # Last partial window
+        if start < length:
+            partial_len = ep_to - (ep_from + start)
+            offsets = _sample_window_indices(
+                partial_len,
+                self.max_frames_per_window or partial_len,
+            )
+            frame_indices = [ep_from + start + i for i in offsets]
+            images = []
+            t_start = float(dataset[frame_indices[0]]["timestamp"].item())
+            for idx in frame_indices:
+                item = dataset[idx]
+                images.append(frame_to_pil(item[camera_key]))
+            t_end = float(dataset[frame_indices[-1]]["timestamp"].item()) + 1.0 / fps
+            window_specs.append((images, t_start, t_end))
+
+        # Run in batches
+        skills: list[Skill] = []
+        for i in range(0, len(window_specs), self.batch_size):
+            chunk = window_specs[i : i + self.batch_size]
+            batch_images = [spec[0] for spec in chunk]
+            if len(batch_images) > 1:
+                skill_names = self.segmenter.segment_skill_from_images_batch(
+                    batch_images, coarse_goal, subtask_labels
+                )
+            else:
+                skill_names = [
+                    self.segmenter.segment_skill_from_images(
+                        batch_images[0], coarse_goal, subtask_labels
+                    )
+                ]
+            for (_, t_start, t_end), name in zip(chunk, skill_names, strict=True):
+                skills.append(Skill(name=name, start=t_start, end=t_end))
+
+        return skills
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Image-window subtask annotation using Qwen VLM (frames as images for better accuracy)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=textwrap.dedent("""\
+            Examples:
+              python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+                --data-dir /path/to/dataset --camera-key observation.images.base \\
+                --window-size 8 --output-dir ./output
+
+              python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+                --repo-id user/dataset --camera-key observation.images.base \\
+                --window-size 6 --stride 3 --model Qwen/Qwen2-VL-7B-Instruct
+
+              # Use Qwen3-VL (MoE)
+              python -m lerobot.data_processing.annotations.subtask_annotate_image \\
+                --data-dir /path/to/dataset --camera-key observation.images.base \\
+                --model Qwen/Qwen3-VL-30B-A3B-Instruct
+        """),
+    )
+    data_group = parser.add_mutually_exclusive_group(required=True)
+    data_group.add_argument("--data-dir", type=str, help="Path to local LeRobot dataset")
+    data_group.add_argument("--repo-id", type=str, help="HuggingFace Hub dataset repository ID")
+
+    parser.add_argument(
+        "--camera-key",
+        type=str,
+        required=True,
+        help="Image or video observation key (e.g. observation.images.base)",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="Qwen/Qwen2-VL-7B-Instruct",
+        help="VLM model: Qwen2-VL or Qwen3-VL (default: Qwen/Qwen2-VL-7B-Instruct)",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+    )
+    parser.add_argument(
+        "--window-size",
+        type=int,
+        default=8,
+        help="Number of frames per window (default: 8)",
+    )
+    parser.add_argument(
+        "--stride",
+        type=int,
+        default=None,
+        help="Stride for sliding window (default: window_size = non-overlapping)",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=1,
+        help="Number of windows to process in one VLM call (default: 1; increase for speed)",
+    )
+    parser.add_argument(
+        "--max-frames-per-window",
+        type=int,
+        default=None,
+        metavar="N",
+        help="If window has more than N frames, randomly sample N frames (order kept) to avoid OOM (e.g. 16)",
+    )
+    parser.add_argument("--episodes", type=int, nargs="+", help="Episode indices to annotate (default: all)")
+    parser.add_argument("--skip-existing", action="store_true", help="Skip episodes that already have annotations")
+    parser.add_argument(
+        "--subtask-labels",
+        type=str,
+        nargs="*",
+        default=None,
+        help="Closed vocabulary: model must choose only from these labels",
+    )
+    parser.add_argument("--output-dir", type=str, help="Output directory for dataset with subtask_index")
+    parser.add_argument("--output-repo-id", type=str, help="Output repo id (default: <repo_id>_with_subtasks)")
+    parser.add_argument("--push-to-hub", action="store_true")
+
+    args = parser.parse_args()
+    console = Console()
+
+    # Load dataset
+    console.print("[cyan]Loading dataset...[/cyan]")
+    if args.data_dir:
+        dataset = LeRobotDataset(repo_id="local/dataset", root=args.data_dir, download_videos=False)
+    else:
+        dataset = LeRobotDataset(repo_id=args.repo_id, download_videos=True)
+    camera_keys = dataset.meta.camera_keys
+    if args.camera_key not in camera_keys:
+        console.print(f"[red]Error: camera key '{args.camera_key}' not in {camera_keys}[/red]")
+        return
+    console.print(f"[green]✓ Loaded dataset, {dataset.meta.total_episodes} episodes[/green]")
+
+    # Same Qwen VLM as subtask_annotate (Qwen2-VL or Qwen3-VL), image windows instead of video
+    segmenter = get_image_segmenter(args.model, args.device, torch.bfloat16)
+
+    annotator = SkillAnnotatorImage(
+        segmenter=segmenter,
+        window_size=args.window_size,
+        stride=args.stride,
+        batch_size=args.batch_size,
+        max_frames_per_window=args.max_frames_per_window,
+        console=console,
+    )
+    annotations = annotator.annotate_dataset(
+        dataset=dataset,
+        camera_key=args.camera_key,
+        episodes=args.episodes,
+        skip_existing=args.skip_existing,
+        subtask_labels=args.subtask_labels,
+    )
+
+    if not annotations:
+        console.print("[yellow]No annotations to save.[/yellow]")
+        return
+
+    output_dir = Path(args.output_dir) if args.output_dir else None
+    output_repo_id = args.output_repo_id
+    new_dataset = save_skill_annotations(dataset, annotations, output_dir, output_repo_id)
+
+    total_skills = sum(len(a.skills) for a in annotations.values())
+    console.print(f"[bold green]✓ Done.[/bold green] Episodes: {len(annotations)}, total window skills: {total_skills}")
+    console.print(f"  Dataset with subtask_index: {new_dataset.root}")
+
+    if args.push_to_hub and not args.data_dir:
+        console.print("[cyan]Pushing to Hub...[/cyan]")
+        try:
+            new_dataset.push_to_hub(push_videos=False)
+            console.print("[green]✓ Pushed.[/green]")
+        except Exception as e:
+            console.print(f"[red]Push failed: {e}[/red]")
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/data_processing/sarm_annotations/subtask_annotation.py

@@ -746,8 +746,7 @@ def save_annotations_to_dataset(
     dataset_path: Path, annotations: dict[int, SubtaskAnnotation], fps: int, prefix: str = "sparse"
 ):
     """Save annotations to LeRobot dataset parquet format."""
-    from lerobot.datasets.io_utils import load_episodes
-    from lerobot.datasets.utils import DEFAULT_EPISODES_PATH
+    from lerobot.datasets.utils import DEFAULT_EPISODES_PATH, load_episodes
 
     episodes_dataset = load_episodes(dataset_path)
     if not episodes_dataset or len(episodes_dataset) == 0:
@@ -841,7 +840,7 @@ def generate_auto_sparse_annotations(
 
 def load_annotations_from_dataset(dataset_path: Path, prefix: str = "sparse") -> dict[int, SubtaskAnnotation]:
     """Load annotations from LeRobot dataset parquet files."""
-    from lerobot.datasets.io_utils import load_episodes
+    from lerobot.datasets.utils import load_episodes
 
     episodes_dataset = load_episodes(dataset_path)
     if not episodes_dataset or len(episodes_dataset) == 0:

src/lerobot/datasets/aggregate.py

@@ -24,16 +24,7 @@ import pandas as pd
 import tqdm
 
 from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.feature_utils import get_hf_features_from_features
-from lerobot.datasets.io_utils import (
-    get_file_size_in_mb,
-    get_parquet_file_size_in_mb,
-    to_parquet_with_hf_images,
-    write_info,
-    write_stats,
-    write_tasks,
-)
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -41,7 +32,14 @@ from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
     DEFAULT_VIDEO_FILE_SIZE_IN_MB,
     DEFAULT_VIDEO_PATH,
+    get_file_size_in_mb,
+    get_hf_features_from_features,
+    get_parquet_file_size_in_mb,
+    to_parquet_with_hf_images,
     update_chunk_file_indices,
+    write_info,
+    write_stats,
+    write_tasks,
 )
 from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
 
@@ -291,9 +289,7 @@ def aggregate_datasets(
 
     logging.info("Find all tasks")
     unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame(
-        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
-    )
+    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
 
     meta_idx = {"chunk": 0, "file": 0}
     data_idx = {"chunk": 0, "file": 0}

src/lerobot/datasets/backward_compatibility.py

@@ -0,0 +1,56 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import packaging.version
+
+V30_MESSAGE = """
+The dataset you requested ({repo_id}) is in {version} format.
+
+We introduced a new format since v3.0 which is not backward compatible with v2.1.
+Please, update your dataset to the new format using this command:
+```
+python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id={repo_id}
+```
+
+If you already have a converted version uploaded to the hub, then this error might be because of
+an older version in your local cache. Consider deleting the cached version and retrying.
+
+If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
+or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
+"""
+
+FUTURE_MESSAGE = """
+The dataset you requested ({repo_id}) is only available in {version} format.
+As we cannot ensure forward compatibility with it, please update your current version of lerobot.
+"""
+
+
+class CompatibilityError(Exception): ...
+
+
+class BackwardCompatibilityError(CompatibilityError):
+    def __init__(self, repo_id: str, version: packaging.version.Version):
+        if version.major == 2 and version.minor == 1:
+            message = V30_MESSAGE.format(repo_id=repo_id, version=version)
+        else:
+            raise NotImplementedError(
+                "Contact the maintainer on [Discord](https://discord.com/invite/s3KuuzsPFb)."
+            )
+        super().__init__(message)
+
+
+class ForwardCompatibilityError(CompatibilityError):
+    def __init__(self, repo_id: str, version: packaging.version.Version):
+        message = FUTURE_MESSAGE.format(repo_id=repo_id, version=version)
+        super().__init__(message)