refactor: consolidate VLM classes into single QwenVL implementation

Remove Qwen2VL, Qwen3VL, Qwen35VL in favor of one QwenVL class that
uses AutoModelForImageTextToText and works with the whole Qwen VL
family. Moves shared _parse_skills_response to BaseVLM and extracts
_build_messages/_prepare_inputs/_decode helpers to reduce duplication.

Made-with: Cursor

.github/workflows/fast_tests.yml

@@ -44,7 +44,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
 
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -61,6 +61,7 @@ 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:
@@ -89,5 +90,11 @@ 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: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -60,6 +60,7 @@ 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:
@@ -87,6 +88,12 @@ 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
 
@@ -101,9 +108,11 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      (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.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'
+      )
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:
@@ -160,6 +169,7 @@ 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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -171,6 +181,13 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$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: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
   DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
 
@@ -119,6 +119,7 @@ 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: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --shm-size "16gb"
@@ -130,6 +131,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$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
@@ -146,6 +152,7 @@ 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: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -157,6 +164,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$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
@@ -174,6 +186,7 @@ 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: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -185,12 +198,15 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$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
-      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
-        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
-        timeout-minutes: 10
+        run: pytest -vv tests/training/

.github/workflows/quality.yml

@@ -50,7 +50,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.10'
+          python-version: '3.12'
 
       - 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: "3.10"
+  PYTHON_VERSION: "3.12"
 
 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: '3.10'
+          python-version: '3.12'
 
       - name: Extract Version
         id: extract_info
@@ -83,14 +83,6 @@ 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: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -48,6 +48,7 @@ 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:
@@ -79,7 +80,11 @@ 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
 
@@ -91,6 +96,7 @@ 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:
@@ -136,6 +142,7 @@ 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: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -147,6 +154,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$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: python3.10
+    python: python3.12
 
 exclude: "tests/artifacts/.*\\.safetensors$"
 
@@ -55,7 +55,7 @@ repos:
     rev: v3.21.0
     hooks:
     -   id: pyupgrade
-        args: [--py310-plus]
+        args: [--py312-plus]
 
   ##### Markdown Quality #####
   - repo: https://github.com/rbubley/mirrors-prettier

AI_POLICY.md

@@ -0,0 +1,25 @@
+# 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](./CODE_OF_CONDUCT.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
 
 ## Ways to Contribute
 

MANIFEST.in

@@ -1,2 +1,3 @@
 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 research, please cite:
+If you use LeRobot in your project, please cite the GitHub repository to acknowledge the ongoing development and contributors:
 
 ```bibtex
 @misc{cadene2024lerobot,
@@ -146,6 +146,23 @@ If you use LeRobot in your research, please cite:
 }
 ```
 
+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](./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!

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.
-- `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.
+- `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.
 
 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 \
-        aliberts/aloha_mobile_shrimp_image \
+        lerobot/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 \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/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 \
-        aliberts/aloha_mobile_shrimp_image \
-        aliberts/paris_street \
-        aliberts/kitchen \
+        lerobot/aloha_mobile_shrimp_image \
+        lerobot/paris_street \
+        lerobot/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%    |
-| 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_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_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**  |
+| 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**  |
 
-|                                    |          | 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%**   |
+|                                   |          | 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%**   |

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=3.10
+ARG PYTHON_VERSION=3.12
 
 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -85,6 +85,8 @@ 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

@@ -19,7 +19,7 @@
 # docker run -it --rm lerobot-user
 
 # Configure the base image
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 FROM python:${PYTHON_VERSION}-slim
 
 # Configure environment variables

docs/source/_toctree.yml

@@ -29,6 +29,8 @@
     title: Using the Dataset 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

@@ -88,5 +88,8 @@ 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/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
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
     --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

docs/source/bring_your_own_policies.mdx

@@ -32,7 +32,7 @@ version = "0.1.0"
 dependencies = [
     # your policy-specific dependencies
 ]
-requires-python = ">= 3.11"
+requires-python = ">= 3.12"
 
 [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 Dict, Any
+from typing import 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 Dict, Any
+from typing import 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 3.10 or newer
+- Computer with Python 3.12 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
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -185,13 +185,16 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-python src/lerobot/scripts/lerobot_record.py \
+lerobot-record \
     --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
 ```
 

docs/source/envhub.mdx

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

docs/source/groot.mdx

@@ -120,9 +120,12 @@ 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"
-  --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" \
+  --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.reset_time_s=10
 ```
 

docs/source/hope_jr.mdx

@@ -224,12 +224,15 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+    --dataset.repo_id=<USER>/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
 ```
 
@@ -241,7 +244,7 @@ lerobot-replay \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
     --robot.side=right \
-    --dataset.repo_id=nepyope/hand_record_test_with_camera \
+    --dataset.repo_id=<USER>/hand_record_test_with_camera \
     --dataset.episode=0
 ```
 
@@ -249,13 +252,13 @@ lerobot-replay \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
+  --dataset.repo_id=<USER>/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=nepyope/hand_test_policy
+  --policy.repo_id=<USER>/hand_test_policy
 ```
 
 ### Evaluate
@@ -270,8 +273,11 @@ 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=nepyope/eval_hopejr \
+  --dataset.repo_id=<USER>/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
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | head -n 1)
+HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -185,7 +185,10 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube"
+    --dataset.single_task="Grab the black cube" \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -324,7 +327,7 @@ 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
-huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```
 
 #### Record function
@@ -488,7 +491,7 @@ 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
-huggingface-cli upload ${HF_USER}/act_so101_test \
+hf upload ${HF_USER}/act_so101_test \
   outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```
 
@@ -496,7 +499,7 @@ You can also upload intermediate checkpoints with:
 
 ```bash
 CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
+hf upload ${HF_USER}/act_so101_test${CKPT} \
   outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```
 
@@ -515,6 +518,9 @@ 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

@@ -1,30 +1,57 @@
 # Installation
 
-## Install [`miniforge`](https://conda-forge.org/download/)
+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/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```
 
-## Environment Setup
+## Step 2: Environment Setup
 
-Create a virtual environment with Python 3.10, using conda:
+Create a virtual environment with Python 3.12:
 
+<!-- prettier-ignore-start -->
+<hfoptions id="create_venv">
+<hfoption id="conda">
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 ```
-
-Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
-
+</hfoption>
+<hfoption id="uv">
 ```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]
@@ -38,7 +65,17 @@ conda install ffmpeg -c conda-forge
 >
 > - _[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`.
 
-## Install LeRobot 🤗
+> [!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 🤗
 
 ### From Source
 
@@ -51,23 +88,45 @@ 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:
+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.):
 
 ```bash
 pip install 'lerobot[all]'          # All available features
@@ -81,13 +140,10 @@ _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
@@ -97,7 +153,7 @@ 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`.
+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.
 
 ### 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
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 

docs/source/lerobot-dataset-v3.mdx

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

docs/source/phone_teleop.mdx

@@ -66,12 +66,13 @@ 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 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)
+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`.
 
 - Run this example to teleoperate:
 
   ```bash
-  python examples/phone_to_so100/teleoperate.py
+  cd examples/phone_to_so100
+  python teleoperate.py
   ```
 
 After running the example:
@@ -84,19 +85,22 @@ 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
-  python examples/phone_to_so100/record.py
+  cd examples/phone_to_so100
+  python record.py
   ```
 
 - Run this example to replay recorded episodes:
 
   ```bash
-  python examples/phone_to_so100/replay.py
+  cd examples/phone_to_so100
+  python replay.py
   ```
 
 - Run this example to evaluate a pretrained policy:
 
   ```bash
-  python examples/phone_to_so100/evaluate.py
+  cd examples/phone_to_so100
+  python evaluate.py
   ```
 
 ### Important pipeline steps and options

docs/source/pi0.mdx

@@ -34,11 +34,6 @@ 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:
@@ -60,7 +55,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \

docs/source/pi05.mdx

@@ -36,11 +36,6 @@ 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:
@@ -56,7 +51,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py\
+lerobot-train \
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \

docs/source/pi0fast.mdx

@@ -43,16 +43,11 @@ 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 `physical-intelligence/fast` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose 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.
 
 2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
 
@@ -114,15 +109,15 @@ 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                              | `physical-intelligence/fast` |
-| `--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                              | `lerobot/fast-action-tokenizer` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
 
 ## Inference
 

docs/source/reachy2.mdx

@@ -159,6 +159,9 @@ 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
 ```
 
@@ -198,6 +201,9 @@ 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/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
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **dense annotations only** (sparse auto-generated):
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ python src/lerobot/scripts/lerobot_train.py \
 Train with **both sparse and dense annotations**:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --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
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
   --use_rabc=true \

docs/source/smolvla.mdx

@@ -106,6 +106,9 @@ 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

@@ -0,0 +1,155 @@
+# 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

@@ -1,23 +1,72 @@
 # Unitree G1
 
-This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/unitree_thumbnail.jpg"
+  alt="Unitree G1 locomanipulation demo"
+  style={{ width: "100%" }}
+/>
 
-## 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
+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.
 
 ---
 
-## Connection guide
+## Part 1: Getting Started
 
-### Step 1: Configure Ethernet Interface
+### Install the Unitree SDK
 
-Set a static IP on the same subnet as the robot:
+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`.
 
 ```bash
 # Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
@@ -26,47 +75,23 @@ sudo ip addr add 192.168.123.200/24 dev enp131s0
 sudo ip link set enp131s0 up
 ```
 
-**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
+### SSH into the Robot
 
 ```bash
 ssh unitree@192.168.123.164
 # Password: 123
 ```
 
-You should now be connected to the G1's Orin.
+### Share Internet via 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
+The G1 needs internet access to clone repos and install packages. Share your laptop's connection over Ethernet:
 
 **On your laptop:**
 
 ```bash
-# Enable IP forwarding
 sudo sysctl -w net.ipv4.ip_forward=1
 
-# Set up NAT (replace wlp132s0f0 with your WiFi interface)
+# Replace wlp132s0f0 with your WiFi interface name
 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
@@ -75,217 +100,193 @@ 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
 
-# Test connection
+# Verify
 ping -c 3 8.8.8.8
 ```
 
-### Step 3: Connect to WiFi Network
+### 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:**
 
 ```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
 ```
 
-### 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:
+You can then SSH over WiFi instead of Ethernet:
 
 ```bash
-ssh unitree@<YOUR_ROBOT_IP>
+ssh unitree@<ROBOT_WIFI_IP>
 # Password: 123
 ```
 
-Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address.
+---
+
+## 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`.
 
 ---
 
-## Part 3: Robot Server Setup
+## Part 3: Loco-Manipulation with the Homunculus Exoskeleton
 
-### Step 1: Install LeRobot on the Orin
+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).
 
-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
+### Calibrate
 
 ```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
 ```
 
-### Teleoperate in Simulation
+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
 
 ```bash
-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
+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
 ```
 
-### Record Dataset in Simulation
+> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
 
-```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)
+Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)
 
 ---
 
-## Running on Real Robot
+## Part 4: Training & Inference
 
-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:
+### Train
 
 ```bash
-python src/lerobot/cameras/zmq/image_server.py
+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
 ```
 
-Keep this running in a separate terminal for camera streaming during recording.
+### Inference with RTC
 
-### Teleoperate Real Robot
+Once trained, we recommend deploying policies using inference-time RTC:
 
 ```bash
-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
+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
 ```
 
-### 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
@@ -294,8 +295,8 @@ Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/da
 - [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: December 2025_
+_Last updated: March 2026_

docs/source/using_dataset_tools.mdx

@@ -12,6 +12,7 @@ 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`.
@@ -156,6 +157,30 @@ 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
-python src/lerobot/scripts/lerobot_train.py \
+lerobot-train \
     --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=pepijn223/bimanual-so100-handover-cube \
+  --dataset.repo_id=<USER>/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=aliberts/record-test \
+    --dataset.repo_id=<USER>/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').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

examples/dataset/slurm_compute_rabc.py

@@ -0,0 +1,490 @@
+#!/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/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=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/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=lipsop/reuben_pi0 \
-        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --policy.path=<USER>/reuben_pi0 \
+        --dataset.repo_id=<USER>/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=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/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=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/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=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \

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=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/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=helper2424/smolvla_check_rtc_last3 \
+        --policy.path=<USER>/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=helper2424/pi05_check_rtc \
+        --policy.path=<USER>/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -78,6 +78,7 @@ 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
@@ -97,6 +98,7 @@ 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

pyproject.toml

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.4"
+version = "0.5.1"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.10"
+requires-python = ">=3.12"
 authors = [
     { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
     { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,7 +50,8 @@ classifiers = [
     "Intended Audience :: Education",
     "Intended Audience :: Science/Research",
     "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
     "Topic :: Software Development :: Build Tools",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -59,28 +60,30 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<4.2.0",
+    "datasets>=4.0.0,<5.0.0",
     "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
+    "huggingface-hub>=1.0.0,<2.0.0",
     "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,<4.13.0",
+    "opencv-python-headless>=4.9.0,<4.14.0",
     "av>=15.0.0,<16.0.0",
     "jsonlines>=4.0.0,<5.0.0",
-    "packaging>=24.2,<26.0",
-    "pynput>=1.7.7,<1.9.0",
+    "pynput>=1.7.8,<1.9.0",
     "pyserial>=3.5,<4.0",
+
     "wandb>=0.24.0,<0.25.0",
-
-    "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 ==
+    "draccus==0.10.0", # TODO: Relax version constraint
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -95,14 +98,20 @@ dependencies = [
 
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.57.1,<5.0.0"]
+placo-dep = ["placo>=0.9.6,<0.9.17"]
+transformers-dep = ["transformers>=5.3.0,<6.0.0"]
 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 = ["python-can>=4.2.0,<5.0.0"]
+damiao = ["lerobot[can-dep]"]
+robstride = ["lerobot[can-dep]"]
 
 # Robots
 openarms = ["lerobot[damiao]"]
@@ -110,34 +119,35 @@ 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",
-    "pin>=3.0.0,<4.0.0",
+    "onnx>=1.16.0,<2.0.0",
     "meshcat>=0.3.0,<0.4.0",
-    "matplotlib>=3.9.0,<4.0.0",
-    "casadi>=3.6.0,<4.0.0",
+    "lerobot[matplotlib-dep]",
+    "lerobot[pygame-dep]",
 ]
 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,<2.55.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
 
 # Policies
 wallx = [
-    "transformers==4.49.0",
-    "peft==0.17.1",
-    "scipy==1.15.3",
-    "torchdiffeq==0.2.5",
-    "qwen_vl_utils==0.0.11"
+    "lerobot[transformers-dep]",
+    "lerobot[peft]",
+    "lerobot[scipy-dep]",
+    "torchdiffeq>=0.2.4,<0.3.0",
+    "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
 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]",
-    "peft>=0.13.0,<1.0.0",
+    "lerobot[peft]",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
     "safetensors>=0.4.3,<1.0.0",
@@ -146,13 +156,13 @@ 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", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 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]", "matplotlib>=3.10.3,<4.0.0"]
-peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
+async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
+peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # 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"]
@@ -160,13 +170,19 @@ 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
-aloha = ["gym-aloha>=0.1.2,<0.2.0"]
+# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
+aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 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"]
-metaworld = ["metaworld==3.0.0"]
+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]"]
 
 # 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]",
@@ -174,8 +190,8 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    # "lerobot[wallx]",
-    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
+    "lerobot[wallx]",
+    "lerobot[pi]",
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
@@ -187,10 +203,11 @@ all = [
     "lerobot[aloha]",
     "lerobot[pusht]",
     "lerobot[phone]",
-    "lerobot[libero]",
+    "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
     "lerobot[peft]",
+    # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]
 
 [project.scripts]
@@ -205,6 +222,7 @@ lerobot-eval="lerobot.scripts.lerobot_eval:main"
 lerobot-train="lerobot.scripts.lerobot_train:main"
 lerobot-train-tokenizer="lerobot.scripts.lerobot_train_tokenizer:main"
 lerobot-dataset-viz="lerobot.scripts.lerobot_dataset_viz:main"
+lerobot-dataset-subtask-annotate="lerobot.scripts.lerobot_subtask_annotate:main"
 lerobot-info="lerobot.scripts.lerobot_info:main"
 lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
@@ -212,11 +230,14 @@ 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 = "py310"
+target-version = "py312"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -308,7 +329,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
 
 [tool.mypy]
-python_version = "3.10"
+python_version = "3.12"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -360,9 +381,9 @@ 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.*"
@@ -392,85 +413,3 @@ 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

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

requirements-ubuntu.txt

@@ -1,12 +1,12 @@
 #
-# This file is autogenerated by pip-compile with Python 3.10
+# This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-ubuntu.txt requirements.in
 #
 -e .[all]
     # via -[all]
-absl-py==2.3.1
+absl-py==2.4.0
     # via
     #   dm-control
     #   dm-env
@@ -14,30 +14,33 @@ absl-py==2.3.1
     #   labmaze
     #   mujoco
     #   tensorboard
-accelerate==1.11.0
+accelerate==1.13.0
     # via
     #   lerobot
     #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.13.3
     # 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==4.11.0
+anyio==4.12.1
     # via
+    #   httpx
     #   starlette
     #   watchfiles
-asttokens==3.0.0
+asttokens==3.0.1
     # via stack-data
-async-timeout==5.0.1
-    # via aiohttp
 attrs==25.4.0
     # via
     #   aiohttp
@@ -47,30 +50,35 @@ attrs==25.4.0
     #   referencing
     #   rerun-sdk
 av==15.1.0
-    # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
     # via
+    #   lerobot
+    #   qwen-vl-utils
+bddl==1.0.1
+    # via hf-libero
+certifi==2026.2.25
+    # via
+    #   httpcore
+    #   httpx
     #   requests
     #   sentry-sdk
 cffi==2.0.0
     # via pymunk
-cfgv==3.4.0
+cfgv==3.5.0
     # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.5
     # via requests
-click==8.3.0
+click==8.3.1
     # via
+    #   typer
     #   uvicorn
     #   wandb
-cloudpickle==3.1.1
+cloudpickle==3.1.2
     # via
     #   gymnasium
-    #   libero
-cmake==4.1.0
+    #   hf-libero
+cmake==4.1.3
     # via lerobot
-cmeel==0.57.3
+cmeel==0.59.0
     # via
     #   cmeel-assimp
     #   cmeel-boost
@@ -108,20 +116,24 @@ cmeel-zlib==1.3.1
     # via cmeel-assimp
 coal-library==3.0.1
     # via pin
-contourpy==1.3.2
-    # via matplotlib
-coverage[toml]==7.11.0
+contourpy==1.3.3
+    # via
+    #   lerobot
+    #   matplotlib
+coverage[toml]==7.13.4
     # 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==4.1.1
+datasets==4.6.1
     # via lerobot
-debugpy==1.8.17
+debugpy==1.8.20
     # via lerobot
 decorator==5.2.1
     # via ipython
-decord==0.6.0
-    # via lerobot
 deepdiff==8.6.1
     # via lerobot
 diffusers==0.35.2
@@ -132,7 +144,7 @@ dill==0.4.0
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.37
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -140,7 +152,6 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
-    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
@@ -148,66 +159,60 @@ draccus==0.10.0
 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
+    # via hf-libero
+egl-probe==1.0.2
+    # via robomimic
 eigenpy==3.10.3
     # via coal-library
-einops==0.8.1
+einops==0.8.2
     # via
-    #   flash-attn
+    #   hf-libero
     #   lerobot
-    #   libero
 eiquadprog==1.2.9
     # via placo
-etils[epath,epy]==1.13.0
+etils[epath,epy]==1.14.0
     # via mujoco
-evdev==1.9.2
+evdev==1.9.3
     # 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==0.119.1
-    # via teleop
+fastapi==0.135.1
+    # via
+    #   lerobot
+    #   teleop
 fastjsonschema==2.21.2
     # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.20.0
+filelock==3.25.0
     # via
     #   datasets
     #   diffusers
     #   huggingface-hub
+    #   python-discovery
     #   torch
-    #   transformers
     #   virtualenv
-flash-attn==2.8.3
-    # via lerobot
-fonttools==4.60.1
+fonttools==4.61.1
     # via matplotlib
 frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2026.2.0
     # via
     #   datasets
     #   etils
     #   huggingface-hub
     #   torch
 future==1.0.0
-    # via libero
+    # via hf-libero
 gitdb==4.0.12
     # via gitpython
-gitpython==3.1.45
+gitpython==3.1.46
     # via wandb
 glfw==2.10.0
     # via
@@ -230,50 +235,60 @@ gym-hil==0.1.13
     # via lerobot
 gym-pusht==0.1.6
     # via lerobot
-gymnasium==1.2.1
+gymnasium==1.2.3
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
+    #   hf-libero
     #   lerobot
-    #   libero
     #   metaworld
 h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
+    # via
+    #   httpcore
+    #   uvicorn
+h5py==3.16.0
     # via robomimic
 hebi-py==2.11.0
     # via lerobot
-hf-transfer==0.1.9
-    # via huggingface-hub
-hf-xet==1.1.10
+hf-egl-probe==1.0.2
+    # via hf-libero
+hf-libero==0.1.3
+    # via lerobot
+hf-xet==1.3.2
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
+httpcore==1.0.9
+    # via httpx
 httptools==0.7.1
     # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+httpx==0.28.1
+    # via
+    #   datasets
+    #   huggingface-hub
+huggingface-hub==1.6.0
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
     #   peft
-    #   timm
     #   tokenizers
     #   transformers
 hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+    # via hf-libero
+identify==2.6.17
     # via pre-commit
 idna==3.11
     # via
     #   anyio
+    #   httpx
     #   requests
     #   yarl
-imageio[ffmpeg]==2.37.0
+imageio[ffmpeg]==2.37.2
     # via
     #   gym-aloha
     #   gym-hil
@@ -285,16 +300,14 @@ imageio-ffmpeg==0.6.0
     # via
     #   imageio
     #   robomimic
-importlib-metadata==8.7.0
+importlib-metadata==8.7.1
     # via diffusers
-importlib-resources==6.5.2
-    # via etils
 iniconfig==2.3.0
     # via pytest
-inquirerpy==0.3.4
-    # via huggingface-hub
-ipython==8.37.0
+ipython==9.11.0
     # via meshcat
+ipython-pygments-lexers==1.1.1
+    # via ipython
 ischedule==1.2.7
     # via placo
 jedi==0.19.2
@@ -303,40 +316,41 @@ jinja2==3.1.6
     # via torch
 jsonlines==4.0.0
     # via lerobot
-jsonschema==4.25.1
+jsonschema==4.26.0
     # via nbformat
 jsonschema-specifications==2025.9.1
     # via jsonschema
 jupyter-core==5.9.1
     # via nbformat
-jupytext==1.18.1
+jupytext==1.19.1
     # via bddl
 kiwisolver==1.4.9
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
-lazy-loader==0.4
+lazy-loader==0.5
     # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
+librt==0.8.1
+    # via mypy
+llvmlite==0.46.0
     # via numba
 lxml==6.0.2
     # via dm-control
-markdown==3.9
+markdown==3.10.2
     # via tensorboard
 markdown-it-py==4.0.0
     # via
     #   jupytext
     #   mdit-py-plugins
+    #   rich
 markupsafe==3.0.3
     # via
     #   jinja2
     #   werkzeug
-matplotlib==3.10.7
+matplotlib==3.10.8
     # via
+    #   hf-libero
     #   lerobot
-    #   libero
 matplotlib-inline==0.2.1
     # via ipython
 mdit-py-plugins==0.5.0
@@ -353,36 +367,38 @@ mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.3.7
+mujoco==3.5.0
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   libero
+    #   hf-libero
     #   metaworld
     #   robosuite
-multidict==6.7.0
+multidict==6.7.1
     # via
     #   aiohttp
     #   yarl
-multiprocess==0.70.16
+multiprocess==0.70.18
     # via datasets
+mypy==1.19.1
+    # via lerobot
 mypy-extensions==1.1.0
-    # via typing-inspect
+    # via
+    #   mypy
+    #   typing-inspect
 nbformat==5.10.4
     # via jupytext
-networkx==3.4.2
+networkx==3.6.1
     # via
     #   bddl
     #   scikit-image
     #   torch
-ninja==1.13.0
-    # via lerobot
-nodeenv==1.9.1
+nodeenv==1.10.0
     # via pre-commit
 num2words==0.5.14
     # via lerobot
-numba==0.62.1
+numba==0.64.0
     # via robosuite
 numpy==2.2.6
     # via
@@ -391,7 +407,6 @@ numpy==2.2.6
     #   cmeel-boost
     #   contourpy
     #   datasets
-    #   decord
     #   diffusers
     #   dm-control
     #   dm-env
@@ -399,9 +414,10 @@ numpy==2.2.6
     #   gymnasium
     #   h5py
     #   hebi-py
+    #   hf-libero
     #   imageio
     #   labmaze
-    #   libero
+    #   lerobot
     #   matplotlib
     #   meshcat
     #   metaworld
@@ -426,49 +442,51 @@ numpy==2.2.6
     #   torchvision
     #   transformers
     #   transforms3d
-nvidia-cublas-cu12==12.6.4.1
+nvidia-cublas-cu12==12.8.4.1
     # via
     #   nvidia-cudnn-cu12
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cuda-cupti-cu12==12.6.80
+nvidia-cuda-cupti-cu12==12.8.90
     # via torch
-nvidia-cuda-nvrtc-cu12==12.6.77
+nvidia-cuda-nvrtc-cu12==12.8.93
     # via torch
-nvidia-cuda-runtime-cu12==12.6.77
+nvidia-cuda-runtime-cu12==12.8.90
     # via torch
-nvidia-cudnn-cu12==9.5.1.17
+nvidia-cudnn-cu12==9.10.2.21
     # via torch
-nvidia-cufft-cu12==11.3.0.4
+nvidia-cufft-cu12==11.3.3.83
     # via torch
-nvidia-cufile-cu12==1.11.1.6
+nvidia-cufile-cu12==1.13.1.3
     # via torch
-nvidia-curand-cu12==10.3.7.77
+nvidia-curand-cu12==10.3.9.90
     # via torch
-nvidia-cusolver-cu12==11.7.1.2
+nvidia-cusolver-cu12==11.7.3.90
     # via torch
-nvidia-cusparse-cu12==12.5.4.2
+nvidia-cusparse-cu12==12.5.8.93
     # via
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cusparselt-cu12==0.6.3
+nvidia-cusparselt-cu12==0.7.1
     # via torch
-nvidia-nccl-cu12==2.26.2
+nvidia-nccl-cu12==2.27.5
     # via torch
-nvidia-nvjitlink-cu12==12.6.85
+nvidia-nvjitlink-cu12==12.8.93
     # via
     #   nvidia-cufft-cu12
     #   nvidia-cusolver-cu12
     #   nvidia-cusparse-cu12
     #   torch
-nvidia-nvtx-cu12==12.6.77
+nvidia-nvshmem-cu12==3.4.5
+    # via torch
+nvidia-nvtx-cu12==12.8.90
     # via torch
 omegaconf==2.3.0
     # via hydra-core
-opencv-python==4.12.0.88
+opencv-python==4.13.0.92
     # via
     #   gym-pusht
-    #   libero
+    #   hf-libero
     #   reachy2-sdk
     #   robosuite
 opencv-python-headless==4.12.0.88
@@ -487,6 +505,7 @@ packaging==25.0
     #   matplotlib
     #   peft
     #   pytest
+    #   qwen-vl-utils
     #   reachy2-sdk
     #   scikit-image
     #   tensorboard
@@ -497,21 +516,21 @@ pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.5
+parso==0.8.6
     # via jedi
-peft==0.17.1
+pathspec==1.0.4
+    # via mypy
+peft==0.18.1
     # via lerobot
 pexpect==4.9.0
     # via ipython
-pfzy==0.3.4
-    # via inquirerpy
-pillow==12.0.0
+pillow==12.1.1
     # via
     #   diffusers
     #   imageio
-    #   lerobot
     #   matplotlib
     #   meshcat
+    #   qwen-vl-utils
     #   rerun-sdk
     #   robosuite
     #   scikit-image
@@ -519,28 +538,27 @@ pillow==12.0.0
     #   torchvision
 pin==3.4.0
     # via placo
-placo==0.9.14
+placo==0.9.16
     # via lerobot
-platformdirs==4.5.0
+platformdirs==4.9.4
     # via
     #   jupyter-core
+    #   python-discovery
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.5.1
     # via lerobot
 prompt-toolkit==3.0.52
-    # via
-    #   inquirerpy
-    #   ipython
+    # via ipython
 propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
-protobuf==6.31.0
+protobuf==6.31.1
     # via
     #   dm-control
     #   grpcio-tools
@@ -550,7 +568,7 @@ protobuf==6.31.0
     #   tensorboard
     #   tensorboardx
     #   wandb
-psutil==7.1.1
+psutil==7.2.2
     # via
     #   accelerate
     #   imageio
@@ -560,17 +578,17 @@ ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
     # via stack-data
-pyarrow==21.0.0
+pyarrow==23.0.1
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.23
+pycparser==3.0
     # via cffi
-pydantic==2.12.3
+pydantic==2.12.5
     # via
     #   fastapi
     #   wandb
-pydantic-core==2.41.4
+pydantic-core==2.41.5
     # via pydantic
 pygame==2.6.1
     # via
@@ -580,12 +598,14 @@ pygame==2.6.1
 pygments==2.19.2
     # via
     #   ipython
+    #   ipython-pygments-lexers
     #   pytest
+    #   rich
 pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.4.1
+pyngrok==7.5.1
     # via meshcat
 pynput==1.8.1
     # via
@@ -595,7 +615,7 @@ pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.5
+pyparsing==3.3.2
     # via
     #   dm-control
     #   matplotlib
@@ -621,13 +641,16 @@ pytest-timeout==2.4.0
     # via lerobot
 python-dateutil==2.9.0.post0
     # via
+    #   faker
     #   matplotlib
     #   pandas
-python-dotenv==1.1.1
+python-discovery==1.1.1
+    # via virtualenv
+python-dotenv==1.2.2
     # via uvicorn
 python-xlib==0.33
     # via pynput
-pytz==2025.2
+pytz==2026.1.post1
     # via pandas
 pyyaml==6.0.3
     # via
@@ -642,7 +665,6 @@ pyyaml==6.0.3
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
-    #   timm
     #   transformers
     #   uvicorn
     #   wandb
@@ -652,7 +674,9 @@ pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-reachy2-sdk==1.0.14
+qwen-vl-utils==0.0.14
+    # via lerobot
+reachy2-sdk==1.0.15
     # via lerobot
 reachy2-sdk-api==1.0.21
     # via reachy2-sdk
@@ -660,7 +684,7 @@ referencing==0.37.0
     # via
     #   jsonschema
     #   jsonschema-specifications
-regex==2025.10.23
+regex==2026.2.28
     # via
     #   diffusers
     #   transformers
@@ -669,60 +693,62 @@ requests==2.32.5
     #   datasets
     #   diffusers
     #   dm-control
-    #   huggingface-hub
+    #   qwen-vl-utils
     #   teleop
-    #   transformers
     #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.26.2
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
+rich==14.3.3
+    # via typer
 robomimic==0.2.0
-    # via libero
+    # via hf-libero
 robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
+    # via hf-libero
+rpds-py==0.30.0
     # via
     #   jsonschema
     #   referencing
-safetensors==0.6.2
+safetensors==0.7.0
     # via
     #   accelerate
     #   diffusers
     #   lerobot
     #   peft
-    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
     #   gym-pusht
     #   lerobot
-scipy==1.15.3
+scipy==1.17.1
     # via
     #   dm-control
+    #   lerobot
     #   metaworld
     #   robosuite
     #   scikit-image
-sentry-sdk==2.42.1
+    #   torchdiffeq
+sentry-sdk==2.54.0
     # 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==5.0.2
+smmap==5.0.3
     # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.48.0
+starlette==0.52.1
     # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.2
+teleop==0.1.4
     # via lerobot
 tensorboard==2.20.0
     # via robomimic
@@ -730,46 +756,38 @@ tensorboard-data-server==0.7.2
     # via tensorboard
 tensorboardx==2.6.4
     # via robomimic
-termcolor==3.1.0
+termcolor==3.3.0
     # via
     #   lerobot
     #   robomimic
 thop==0.1.1.post2209072238
-    # via libero
-tifffile==2025.5.10
+    # via hf-libero
+tifffile==2026.3.3
     # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.22.2
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.3.0
-    # via
-    #   cmeel
-    #   coverage
-    #   jupytext
-    #   pytest
-torch==2.7.1
+torch==2.10.0
     # via
     #   accelerate
-    #   flash-attn
     #   lerobot
     #   peft
     #   robomimic
     #   thop
-    #   timm
+    #   torchdiffeq
     #   torchvision
-torchcodec==0.5
+torchcodec==0.10.0
     # via lerobot
-torchvision==0.22.1
+torchdiffeq==0.2.5
+    # via lerobot
+torchvision==0.25.0
     # via
     #   lerobot
     #   robomimic
-    #   timm
-tornado==6.5.2
+tornado==6.5.4
     # via meshcat
-tqdm==4.67.1
+tqdm==4.67.3
     # via
     #   datasets
     #   dm-control
@@ -783,26 +801,29 @@ traitlets==5.14.3
     #   jupyter-core
     #   matplotlib-inline
     #   nbformat
-transformers==4.57.1
+transformers==5.3.0
     # via
+    #   hf-libero
     #   lerobot
-    #   libero
     #   peft
 transforms3d==0.4.2
     # via teleop
-triton==3.3.1
+triton==3.6.0
     # via torch
+typer==0.24.1
+    # via
+    #   huggingface-hub
+    #   transformers
 typing-extensions==4.15.0
     # via
     #   aiosignal
     #   anyio
     #   etils
-    #   exceptiongroup
+    #   faker
     #   fastapi
     #   gymnasium
     #   huggingface-hub
-    #   ipython
-    #   multidict
+    #   mypy
     #   pydantic
     #   pydantic-core
     #   referencing
@@ -811,46 +832,46 @@ 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 pydantic
-tzdata==2025.2
+    # via
+    #   fastapi
+    #   pydantic
+tzdata==2025.3
     # via pandas
 u-msgpack-python==2.8.0
     # via meshcat
-urllib3==2.5.0
+urllib3==2.6.3
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.38.0
+uvicorn[standard]==0.41.0
     # via teleop
 uvloop==0.22.1
     # via uvicorn
-virtualenv==20.35.3
+virtualenv==21.1.0
     # via pre-commit
-wandb==0.21.4
+wandb==0.24.2
     # via
+    #   hf-libero
     #   lerobot
-    #   libero
 watchfiles==1.1.1
     # via uvicorn
-wcwidth==0.2.14
+wcwidth==0.6.0
     # via prompt-toolkit
 websocket-client==1.9.0
     # via teleop
-websockets==15.0.1
+websockets==16.0
     # via uvicorn
-werkzeug==3.1.3
+werkzeug==3.1.6
     # via tensorboard
-wrapt==2.0.0
+wrapt==2.1.2
     # via dm-tree
 xxhash==3.6.0
     # via datasets
-yarl==1.22.0
+yarl==1.23.0
     # 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 26.0.1 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-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-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
+# 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
 
 -e .[all]

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,
@@ -485,8 +485,9 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    if cfg.robot.type not in SUPPORTED_ROBOTS:
-        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    # 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!")
 
     client = RobotClient(cfg)
 
@@ -512,4 +513,5 @@ 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, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Backends, 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 = 1000) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 500) -> 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

@@ -25,6 +25,10 @@ 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
@@ -32,6 +36,25 @@ 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

@@ -32,10 +32,11 @@ 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.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_backend, get_cv2_rotation
+from ..utils import get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -117,7 +118,7 @@ class OpenCVCamera(Camera):
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = get_cv2_backend()
+        self.backend: int = config.backend
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -132,6 +133,7 @@ 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.
@@ -148,8 +150,6 @@ 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.
@@ -178,6 +178,7 @@ 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.
@@ -197,8 +198,6 @@ 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:
@@ -348,6 +347,7 @@ 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.
@@ -374,9 +374,6 @@ 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.")
 
@@ -490,6 +487,7 @@ 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.
@@ -512,8 +510,6 @@ 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.")
@@ -533,7 +529,8 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -548,8 +545,6 @@ 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, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]
 
 
 @CameraConfig.register_subclass("opencv")
@@ -50,6 +50,7 @@ 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.
@@ -62,22 +63,12 @@ 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:
-        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."
-            )
+        self.color_mode = ColorMode(self.color_mode)
+        self.rotation = Cv2Rotation(self.rotation)
+        self.backend = Cv2Backends(self.backend)
 
         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

@@ -74,7 +74,4 @@ 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."
             )
 
-        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."
-            )
+        self.color_mode = ColorMode(self.color_mode)

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -32,6 +32,7 @@ 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:
@@ -123,6 +124,7 @@ 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.
@@ -136,9 +138,6 @@ 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.")
 
@@ -184,6 +183,7 @@ class Reachy2Camera(Camera):
 
         return frame
 
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Same as read()
@@ -197,12 +197,11 @@ 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()
 
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -219,8 +218,6 @@ 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.")
@@ -233,6 +230,7 @@ class Reachy2Camera(Camera):
 
         return self.latest_frame
 
+    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -240,8 +238,6 @@ 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

@@ -30,7 +30,8 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -152,6 +153,7 @@ 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.
@@ -169,8 +171,6 @@ 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()
@@ -290,6 +290,7 @@ 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.
 
@@ -299,8 +300,6 @@ 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.")
@@ -320,6 +319,7 @@ 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.
@@ -345,9 +345,6 @@ 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.")
 
@@ -374,6 +371,7 @@ 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.
@@ -403,9 +401,6 @@ 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.")
 
@@ -534,6 +529,7 @@ 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.
@@ -556,8 +552,6 @@ 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.")
@@ -578,7 +572,8 @@ class RealSenseCamera(Camera):
         return frame
 
     # NOTE(Steven): Missing implementation for depth for now
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the
@@ -593,8 +588,6 @@ 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

@@ -60,20 +60,8 @@ class RealSenseCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        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."
-            )
+        self.color_mode = ColorMode(self.color_mode)
+        self.rotation = Cv2Rotation(self.rotation)
 
         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

@@ -14,7 +14,6 @@
 # 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
@@ -68,14 +67,3 @@ 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

@@ -34,7 +34,8 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -104,6 +105,7 @@ 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.
 
@@ -111,8 +113,6 @@ 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:
-            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
+            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
             if type(e).__name__ == "Again":
                 raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
             raise
@@ -211,6 +211,7 @@ 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.
@@ -228,9 +229,6 @@ 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.")
 
@@ -301,6 +299,7 @@ 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.
@@ -317,8 +316,6 @@ 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.")
@@ -335,6 +332,7 @@ 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).
 
@@ -350,8 +348,6 @@ 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

@@ -32,10 +32,7 @@ class ZMQCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        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."
-            )
+        self.color_mode = ColorMode(self.color_mode)
 
         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

@@ -23,6 +23,7 @@ import base64
 import contextlib
 import json
 import logging
+import threading
 import time
 from collections import deque
 
@@ -42,10 +43,57 @@ 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])
@@ -61,6 +109,10 @@ 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)
@@ -73,6 +125,18 @@ 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:
@@ -80,10 +144,12 @@ class ImageServer:
 
                 # Build message
                 message = {"timestamps": {}, "images": {}}
-                for name, cam in self.cameras.items():
-                    frame = cam.read()  # Returns RGB
-                    message["timestamps"][name] = time.time()
-                    message["images"][name] = encode_image(frame)
+                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
 
                 # Send as JSON string (suppress if buffer full)
                 with contextlib.suppress(zmq.Again):
@@ -102,6 +168,8 @@ 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').
+    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/configs/policies.py

@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-        input_normalization_modes: A dictionary with key representing the modality and the value specifies the
-            normalization mode to apply.
-        output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
-            the original scale.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
     """
 
     n_obs_steps: int = 1

src/lerobot/configs/train.py

@@ -50,6 +50,9 @@ 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%.
+    cudnn_deterministic: bool = False
     # Number of workers for the dataloader.
     num_workers: int = 4
     batch_size: int = 8

src/lerobot/data_processing/data_annotations/__init__.py

@@ -0,0 +1,2 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+# Data annotations for subtasks and VLM-based labeling.

src/lerobot/data_processing/data_annotations/subtask_annotations.py

@@ -0,0 +1,671 @@
+# 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.
+
+import json
+import subprocess
+import tempfile
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import cv2
+
+from lerobot.datasets.dataset_tools import add_features
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import (
+    create_subtask_index_array,
+    create_subtasks_dataframe,
+    save_subtasks,
+)
+
+if TYPE_CHECKING:
+    from lerobot.data_processing.data_annotations.vlm_annotations import BaseVLM
+
+
+# Skill Annotation Data Structures
+class Skill:
+    """Represents a single atomic skill/subtask in a demonstration."""
+
+    def __init__(self, name: str, start: float, end: float):
+        self.name = name
+        self.start = start  # Start timestamp in seconds
+        self.end = end  # End timestamp in seconds
+
+    def to_dict(self) -> dict:
+        return {"name": self.name, "start": self.start, "end": self.end}
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "Skill":
+        return cls(name=data["name"], start=data["start"], end=data["end"])
+
+    def __repr__(self) -> str:
+        return f"Skill(name='{self.name}', start={self.start:.2f}, end={self.end:.2f})"
+
+
+class EpisodeSkills:
+    """Container for all skills in an episode."""
+
+    def __init__(self, episode_index: int, description: str, skills: list[Skill]):
+        self.episode_index = episode_index
+        self.description = description
+        self.skills = skills
+
+    def to_dict(self) -> dict:
+        return {
+            "episode_index": self.episode_index,
+            "description": self.description,
+            "skills": [s.to_dict() for s in self.skills],
+        }
+
+
+# Video Extraction Utilities
+
+
+class VideoExtractor:
+    """Utilities for extracting and processing video segments from LeRobot datasets."""
+
+    def __init__(self) -> None:
+        pass
+
+    def extract_episode_video(
+        self,
+        video_path: Path,
+        start_timestamp: float,
+        end_timestamp: float,
+        target_fps: int = 1,
+    ) -> Path:
+        """
+        Extract a specific episode segment from a concatenated video file.
+
+        Args:
+            video_path: Path to the source video file
+            start_timestamp: Start time in seconds
+            end_timestamp: End time in seconds
+            target_fps: Target frames per second for output
+
+        Returns:
+            Path to the extracted temporary video file
+        """
+        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_file:
+            tmp_path = Path(tmp_file.name)
+
+        duration = end_timestamp - start_timestamp
+
+        print(f"Extracting: {start_timestamp:.1f}s - {end_timestamp:.1f}s ({duration:.1f}s)")
+
+        cmd = [
+            "ffmpeg",
+            "-i",
+            str(video_path),
+            "-ss",
+            str(start_timestamp),
+            "-t",
+            str(duration),
+            "-r",
+            str(target_fps),
+            "-c:v",
+            "libx264",
+            "-preset",
+            "ultrafast",
+            "-crf",
+            "23",
+            "-an",
+            "-y",
+            str(tmp_path),
+        ]
+
+        try:
+            subprocess.run(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, check=True)
+        except subprocess.CalledProcessError as e:
+            raise RuntimeError(f"FFmpeg failed: {e}") from e
+        except FileNotFoundError as e:
+            raise RuntimeError("FFmpeg not found. Please install ffmpeg.") from e
+
+        if not tmp_path.exists() or tmp_path.stat().st_size < 1024:
+            if tmp_path.exists():
+                tmp_path.unlink()
+            raise RuntimeError("Video extraction produced invalid file")
+
+        return tmp_path
+
+    def add_timer_overlay(self, video_path: Path) -> Path:
+        """
+        Add a visible timer overlay to each frame (elapsed time in seconds) in one corner.
+        Used so the VLM can read the timestamp from the image instead of relying on file metadata.
+        Draws a black box with white text at top-right. Writes to a new temporary file and returns its path.
+        """
+        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as out_file:
+            out_path = Path(out_file.name)
+
+        cap = cv2.VideoCapture(str(video_path))
+        if not cap.isOpened():
+            raise RuntimeError("Failed to open video")
+
+        fps = cap.get(cv2.CAP_PROP_FPS) or 1.0
+        w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+        writer = cv2.VideoWriter(str(out_path), fourcc, fps, (w, h))
+
+        font = cv2.FONT_HERSHEY_SIMPLEX
+        font_scale = max(1.2, min(h, w) / 350.0)
+        thickness = max(2, int(font_scale))
+
+        padding = 15
+        margin = 30
+
+        frame_idx = 0
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+
+            t_sec = frame_idx / fps
+            text = f"{t_sec:.2f} s"
+
+            (tw, th), baseline = cv2.getTextSize(text, font, font_scale, thickness)
+
+            # Top-right placement
+            x_text = w - tw - margin - padding
+            y_text = margin + th + padding
+
+            # Rectangle coordinates (black box behind text)
+            x1 = x_text - padding
+            y1 = y_text - th - padding
+            x2 = x_text + tw + padding
+            y2 = y_text + baseline + padding
+
+            # Draw black filled rectangle
+            cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 0, 0), -1)
+
+            # Draw white text
+            cv2.putText(
+                frame,
+                text,
+                (x_text, y_text),
+                font,
+                font_scale,
+                (255, 255, 255),
+                thickness,
+                lineType=cv2.LINE_AA,
+            )
+
+            writer.write(frame)
+            frame_idx += 1
+
+        cap.release()
+        writer.release()
+        if not out_path.exists() or out_path.stat().st_size < 1024:
+            if out_path.exists():
+                out_path.unlink()
+            raise RuntimeError("Timer overlay produced invalid file")
+        return out_path
+
+    def get_video_duration(self, video_path: Path) -> float:
+        """Get duration of a video file in seconds."""
+        cap = cv2.VideoCapture(str(video_path))
+        fps = cap.get(cv2.CAP_PROP_FPS) or 30
+        frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        cap.release()
+        return frame_count / fps
+
+
+# Skill Annotation Pipeline
+class SkillAnnotator:
+    """
+    Main class for annotating LeRobot datasets with skill labels.
+
+    This class orchestrates the full annotation pipeline:
+    1. Load dataset
+    2. Extract video segments for each episode
+    3. Run VLM-based skill segmentation
+    4. Update dataset task metadata
+    """
+
+    def __init__(
+        self,
+        vlm: "BaseVLM",
+        video_extractor: VideoExtractor | None = None,
+        batch_size: int = 8,
+        add_timer_overlay: bool = True,
+    ):
+        self.vlm = vlm
+        self.video_extractor = video_extractor or VideoExtractor()
+        self.batch_size = batch_size
+        self.add_timer_overlay = add_timer_overlay
+
+    def annotate_dataset(
+        self,
+        dataset: LeRobotDataset,
+        video_key: str,
+        episodes: list[int] | None = None,
+        skip_existing: bool = False,
+        subtask_labels: list[str] | None = None,
+    ) -> dict[int, EpisodeSkills]:
+        """
+        Annotate all episodes in a dataset with skill labels using batched processing.
+
+        Args:
+            dataset: LeRobot dataset to annotate
+            video_key: Key for video observations (e.g., "observation.images.base")
+            episodes: Specific episode indices to annotate (None = all)
+            skip_existing: Skip episodes that already have skill annotations
+            subtask_labels: If provided, model must choose only from these labels (closed vocabulary)
+
+        Returns:
+            Dictionary mapping episode index to EpisodeSkills
+        """
+        episode_indices = episodes or list(range(dataset.meta.total_episodes))
+        annotations: dict[int, EpisodeSkills] = {}
+        failed_episodes: dict[int, str] = {}  # Track failed episodes with error messages
+
+        # Get coarse task description if available
+        coarse_goal = self._get_coarse_goal(dataset)
+
+        # Filter out episodes that already have annotations if skip_existing is True
+        if skip_existing:
+            existing_annotations = load_skill_annotations(dataset.root)
+            if existing_annotations and "episodes" in existing_annotations:
+                # Only skip episodes that exist AND have non-empty skills
+                existing_episode_indices = set()
+                for idx_str, episode_data in existing_annotations["episodes"].items():
+                    idx = int(idx_str)
+                    # Check if skills list exists and is not empty
+                    if "skills" in episode_data and episode_data["skills"]:
+                        existing_episode_indices.add(idx)
+
+                original_count = len(episode_indices)
+                episode_indices = [ep for ep in episode_indices if ep not in existing_episode_indices]
+                skipped_count = original_count - len(episode_indices)
+                if skipped_count > 0:
+                    print(f"Skipping {skipped_count} episodes with existing non-empty annotations")
+
+        if not episode_indices:
+            print("No episodes to annotate (all already annotated)")
+            return annotations
+
+        print(f"Annotating {len(episode_indices)} episodes in batches of {self.batch_size}...")
+
+        # Process episodes in batches
+        for batch_start in range(0, len(episode_indices), self.batch_size):
+            batch_end = min(batch_start + self.batch_size, len(episode_indices))
+            batch_episodes = episode_indices[batch_start:batch_end]
+
+            print(
+                f"Processing batch {batch_start // self.batch_size + 1}/{(len(episode_indices) + self.batch_size - 1) // self.batch_size} (episodes {batch_episodes[0]} to {batch_episodes[-1]})..."
+            )
+
+            try:
+                batch_annotations = self._annotate_episodes_batch(
+                    dataset, batch_episodes, video_key, coarse_goal, subtask_labels
+                )
+
+                for ep_idx in batch_episodes:
+                    if ep_idx in batch_annotations and batch_annotations[ep_idx]:
+                        skills = batch_annotations[ep_idx]
+                        annotations[ep_idx] = EpisodeSkills(
+                            episode_index=ep_idx,
+                            description=coarse_goal,
+                            skills=skills,
+                        )
+                        print(f" Episode {ep_idx}: {len(skills)} skills identified")
+                    else:
+                        failed_episodes[ep_idx] = "Empty or missing skills from batch processing"
+                        print(f"⚠ Episode {ep_idx}: No skills extracted, will retry")
+            except Exception as e:
+                print(f"✗ Batch failed: {e}. Falling back to single-episode processing...")
+                # Fallback: process episodes one by one
+                for ep_idx in batch_episodes:
+                    try:
+                        skills = self._annotate_episode(
+                            dataset, ep_idx, video_key, coarse_goal, subtask_labels
+                        )
+                        if skills:
+                            annotations[ep_idx] = EpisodeSkills(
+                                episode_index=ep_idx,
+                                description=coarse_goal,
+                                skills=skills,
+                            )
+                            print(f" Episode {ep_idx}: {len(skills)} skills identified")
+                        else:
+                            failed_episodes[ep_idx] = "Empty skills list from single-episode processing"
+                            print(f"⚠ Episode {ep_idx}: No skills extracted, will retry")
+                    except Exception as ep_error:
+                        failed_episodes[ep_idx] = str(ep_error)
+                        print(f"⚠ Episode {ep_idx} failed: {ep_error}, will retry")
+
+        # Retry failed episodes one more time
+        if failed_episodes:
+            print(f"\nRetrying {len(failed_episodes)} failed episodes...")
+            retry_count = 0
+            for ep_idx, error_msg in list(failed_episodes.items()):
+                print(f"Retry attempt for episode {ep_idx} (previous error: {error_msg})")
+                try:
+                    skills = self._annotate_episode(dataset, ep_idx, video_key, coarse_goal, subtask_labels)
+                    if skills:
+                        annotations[ep_idx] = EpisodeSkills(
+                            episode_index=ep_idx,
+                            description=coarse_goal,
+                            skills=skills,
+                        )
+                        print(f" Episode {ep_idx} (retry): {len(skills)} skills identified")
+                        del failed_episodes[ep_idx]
+                        retry_count += 1
+                    else:
+                        print(f"✗ Episode {ep_idx} (retry): Still no skills extracted")
+                except Exception as retry_error:
+                    failed_episodes[ep_idx] = str(retry_error)
+                    print(f"✗ Episode {ep_idx} (retry) failed: {retry_error}")
+
+            if retry_count > 0:
+                print(f"Successfully recovered {retry_count} episodes on retry")
+
+            if failed_episodes:
+                print(f"\n⚠ Warning: {len(failed_episodes)} episodes still failed after retry:")
+                for ep_idx, error_msg in failed_episodes.items():
+                    print(f"  Episode {ep_idx}: {error_msg}")
+
+        return annotations
+
+    def _get_coarse_goal(self, dataset: LeRobotDataset) -> str:
+        """Extract or generate the coarse task description."""
+        # Try to get from existing task metadata
+        if dataset.meta.tasks is not None and len(dataset.meta.tasks) > 0:
+            # Get the first task description
+            first_task = dataset.meta.tasks.index[0]
+            if first_task:
+                return str(first_task)
+
+        return "Perform the demonstrated manipulation task."
+
+    def _annotate_episodes_batch(
+        self,
+        dataset: LeRobotDataset,
+        episode_indices: list[int],
+        video_key: str,
+        coarse_goal: str,
+        subtask_labels: list[str] | None = None,
+    ) -> dict[int, list[Skill]]:
+        """Annotate multiple episodes with skill labels in a batch."""
+        # Extract all videos for this batch
+        extracted_paths = []
+        timer_paths = []
+        paths_for_vlm = []
+        durations = []
+        valid_episode_indices = []
+
+        for ep_idx in episode_indices:
+            try:
+                # Get video path and timestamps
+                video_path = dataset.root / dataset.meta.get_video_file_path(ep_idx, video_key)
+
+                if not video_path.exists():
+                    print(f"Warning: Video not found for episode {ep_idx}")
+                    continue
+
+                # Get episode timestamps from metadata
+                ep = dataset.meta.episodes[ep_idx]
+                start_ts = float(ep[f"videos/{video_key}/from_timestamp"])
+                end_ts = float(ep[f"videos/{video_key}/to_timestamp"])
+                duration = end_ts - start_ts
+
+                # Extract episode segment to temporary file
+                extracted_path = self.video_extractor.extract_episode_video(
+                    video_path, start_ts, end_ts, target_fps=dataset.meta.fps
+                )
+
+                if self.add_timer_overlay:
+                    video_for_vlm = self.video_extractor.add_timer_overlay(extracted_path)
+                    extracted_paths.append(extracted_path)
+                    timer_paths.append(video_for_vlm)
+                else:
+                    video_for_vlm = extracted_path
+                    extracted_paths.append(extracted_path)
+                    timer_paths.append(None)
+
+                paths_for_vlm.append(video_for_vlm)
+                durations.append(duration)
+                valid_episode_indices.append(ep_idx)
+
+            except Exception as e:
+                print(f"Warning: Failed to extract video for episode {ep_idx}: {e}")
+                continue
+
+        if not paths_for_vlm:
+            return {}
+
+        try:
+            # Run VLM skill segmentation in batch
+            all_skills = self.vlm.segment_skills_batch(paths_for_vlm, durations, coarse_goal, subtask_labels)
+
+            # Map results back to episode indices
+            results = {}
+            for ep_idx, skills in zip(valid_episode_indices, all_skills, strict=True):
+                results[ep_idx] = skills
+
+            return results
+
+        finally:
+            # Clean up all temporary files (extracted and timer-overlay)
+            for path in extracted_paths:
+                if path.exists():
+                    path.unlink()
+            for path in timer_paths:
+                if path is not None and path.exists():
+                    path.unlink()
+
+    def _annotate_episode(
+        self,
+        dataset: LeRobotDataset,
+        episode_index: int,
+        video_key: str,
+        coarse_goal: str,
+        subtask_labels: list[str] | None = None,
+    ) -> list[Skill]:
+        """Annotate a single episode with skill labels."""
+        # Get video path and timestamps for this episode
+        video_path = dataset.root / dataset.meta.get_video_file_path(episode_index, video_key)
+
+        if not video_path.exists():
+            raise FileNotFoundError(f"Video not found: {video_path}")
+
+        # Get episode timestamps from metadata
+        ep = dataset.meta.episodes[episode_index]
+        start_ts = float(ep[f"videos/{video_key}/from_timestamp"])
+        end_ts = float(ep[f"videos/{video_key}/to_timestamp"])
+        duration = end_ts - start_ts
+
+        # Extract episode segment to temporary file
+        extracted_path = self.video_extractor.extract_episode_video(
+            video_path, start_ts, end_ts, target_fps=1
+        )
+        if self.add_timer_overlay:
+            video_for_vlm = self.video_extractor.add_timer_overlay(extracted_path)
+        else:
+            video_for_vlm = extracted_path
+
+        try:
+            # Run VLM skill segmentation
+            skills = self.vlm.segment_skills(video_for_vlm, duration, coarse_goal, subtask_labels)
+            return skills
+        finally:
+            # Clean up temporary files (extracted and optionally timer-overlay)
+            if extracted_path.exists():
+                extracted_path.unlink()
+            if self.add_timer_overlay and video_for_vlm != extracted_path and video_for_vlm.exists():
+                video_for_vlm.unlink()
+
+
+# Metadata Writer - Updates per-frame task_index based on skills
+
+
+def get_skill_for_timestamp(skills: list[Skill], timestamp: float) -> Skill | None:
+    """
+    Find which skill covers a given timestamp.
+
+    Args:
+        skills: List of skills with start/end times
+        timestamp: Frame timestamp in seconds
+
+    Returns:
+        The Skill that covers this timestamp, or None if not found
+    """
+    for skill in skills:
+        if skill.start <= timestamp < skill.end:
+            return skill
+        # Handle the last frame (end boundary)
+        if timestamp >= skill.end and skill == skills[-1]:
+            return skill
+    return skills[-1] if skills else None  # Fallback to last skill
+
+
+def save_skill_annotations(
+    dataset: LeRobotDataset,
+    annotations: dict[int, EpisodeSkills],
+    output_dir: Path | None = None,
+    repo_id: str | None = None,
+) -> LeRobotDataset:
+    """
+    Save skill annotations to the dataset by:
+    1. Creating a subtasks.parquet file with unique subtasks
+    2. Adding a subtask_index feature to the dataset
+    3. Saving raw skill annotations as JSON for reference
+
+    This function does NOT modify tasks.parquet - it keeps the original tasks intact
+    and creates a separate subtask hierarchy.
+
+    Args:
+        dataset: The LeRobot dataset to annotate
+        annotations: Dictionary of episode skills
+        output_dir: Optional directory to save the modified dataset
+        repo_id: Optional repository ID for the new dataset
+
+    Returns:
+        New dataset with subtask_index feature added
+    """
+    if not annotations:
+        print("No annotations to save")
+        return dataset
+
+    # Step 1: Create subtasks DataFrame
+    print("Creating subtasks DataFrame...")
+    subtasks_df, skill_to_subtask_idx = create_subtasks_dataframe(annotations)
+
+    # Step 2: Create subtask_index array for all frames
+    print("Creating subtask_index array...")
+    subtask_indices = create_subtask_index_array(dataset, annotations, skill_to_subtask_idx)
+
+    # Step 3: Save subtasks.parquet to the original dataset root
+    save_subtasks(subtasks_df, dataset.root)
+
+    # Step 4: Save the raw skill annotations as JSON for reference
+    skills_path = dataset.root / "meta" / "skills.json"
+    skills_path.parent.mkdir(parents=True, exist_ok=True)
+
+    # Load existing skills data if it exists and is not empty
+    existing_skills_data = None
+    if skills_path.exists():
+        try:
+            with open(skills_path) as f:
+                existing_skills_data = json.load(f)
+                if existing_skills_data and len(existing_skills_data.get("episodes", {})) > 0:
+                    print(
+                        f"Found existing skills.json with {len(existing_skills_data.get('episodes', {}))} episodes, merging..."
+                    )
+        except (OSError, json.JSONDecodeError):
+            print("Warning: Could not load existing skills.json, will create new file")
+            existing_skills_data = None
+
+    # Prepare new annotations
+    new_episodes = {str(ep_idx): ann.to_dict() for ep_idx, ann in annotations.items()}
+
+    # Merge with existing data if available
+    if existing_skills_data:
+        # Preserve existing episodes that are not being updated
+        merged_episodes = existing_skills_data.get("episodes", {}).copy()
+        merged_episodes.update(new_episodes)
+
+        # Merge skill_to_subtask_index mappings
+        merged_skill_to_subtask = existing_skills_data.get("skill_to_subtask_index", {}).copy()
+        merged_skill_to_subtask.update(skill_to_subtask_idx)
+
+        # Use existing coarse_description if available, otherwise use new one
+        coarse_desc = existing_skills_data.get(
+            "coarse_description", annotations[next(iter(annotations))].description
+        )
+
+        skills_data = {
+            "coarse_description": coarse_desc,
+            "skill_to_subtask_index": merged_skill_to_subtask,
+            "episodes": merged_episodes,
+        }
+        print(
+            f"Updated {len(new_episodes)} episode(s), total episodes in skills.json: {len(merged_episodes)}"
+        )
+    else:
+        # No existing data, create new
+        skills_data = {
+            "coarse_description": annotations[next(iter(annotations))].description,
+            "skill_to_subtask_index": skill_to_subtask_idx,
+            "episodes": new_episodes,
+        }
+
+    with open(skills_path, "w") as f:
+        json.dump(skills_data, f, indent=2)
+
+    print(f" Saved skill annotations to {skills_path}")
+
+    # Step 5: Add subtask_index feature to dataset using add_features
+    print("Adding subtask_index feature to dataset...")
+
+    # Determine output directory and repo_id
+    output_dir = dataset.root.parent / f"{dataset.root.name}" if output_dir is None else Path(output_dir)
+
+    if repo_id is None:
+        repo_id = f"{dataset.repo_id}"
+
+    # Add feature using dataset_tools
+    feature_info = {
+        "dtype": "int64",
+        "shape": (1,),
+        "names": None,
+    }
+    new_dataset = add_features(
+        dataset=dataset,
+        features={
+            "subtask_index": (subtask_indices, feature_info),
+        },
+        output_dir=output_dir,
+        repo_id=repo_id,
+    )
+
+    # Copy subtasks.parquet to new output directory
+    import shutil
+
+    shutil.copy(dataset.root / "meta" / "subtasks.parquet", output_dir / "meta" / "subtasks.parquet")
+    shutil.copy(dataset.root / "meta" / "skills.json", output_dir / "meta" / "skills.json")
+
+    print(" Successfully added subtask_index feature!")
+    print(f"  New dataset saved to: {new_dataset.root}")
+    print(f"  Total subtasks: {len(subtasks_df)}")
+
+    return new_dataset
+
+
+def load_skill_annotations(dataset_root: Path) -> dict | None:
+    """Load existing skill annotations from a dataset."""
+    skills_path = dataset_root / "meta" / "skills.json"
+    if skills_path.exists():
+        with open(skills_path) as f:
+            return json.load(f)
+    return None

src/lerobot/data_processing/data_annotations/vlm_annotations.py

@@ -0,0 +1,271 @@
+# 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.
+
+import json
+import logging
+import re
+from abc import ABC, abstractmethod
+from pathlib import Path
+
+import torch
+
+from lerobot.data_processing.data_annotations.subtask_annotations import Skill
+from lerobot.utils.constants import (
+    SKILL_SEGMENTATION_PROMPT_TEMPLATE,
+    format_subtask_labels_section,
+)
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_MODEL = "Qwen/Qwen3.5-27B"
+
+
+def create_skill_segmentation_prompt(
+    coarse_goal: str | None = None,
+    subtask_labels: list[str] | None = None,
+    duration_seconds: float | None = None,
+) -> str:
+    """Create the prompt for skill segmentation using the template from constants."""
+    if duration_seconds is None:
+        raise ValueError("duration_seconds is required for skill segmentation prompt")
+    goal_context = f'The overall goal is: "{coarse_goal}"\n\n' if coarse_goal else ""
+    subtask_labels_section = format_subtask_labels_section(subtask_labels) if subtask_labels else ""
+    video_duration_mm_ss = f"{int(duration_seconds // 60):02d}:{int(duration_seconds % 60):02d}"
+    return SKILL_SEGMENTATION_PROMPT_TEMPLATE.format(
+        goal_context=goal_context,
+        subtask_labels_section=subtask_labels_section,
+        video_duration_seconds=duration_seconds,
+        video_duration_mm_ss=video_duration_mm_ss,
+    )
+
+
+class BaseVLM(ABC):
+    """
+    Abstract base class for Vision-Language Models used in skill segmentation.
+
+    To add a new VLM family:
+    1. Subclass BaseVLM
+    2. Implement __init__, segment_skills, and segment_skills_batch
+    3. Register it in get_vlm()
+    """
+
+    @abstractmethod
+    def __init__(self, model_name: str, device: str = "cuda", torch_dtype: torch.dtype = torch.bfloat16):
+        pass
+
+    @abstractmethod
+    def segment_skills(
+        self,
+        video_path: Path,
+        episode_duration: float,
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[Skill]:
+        """Segment a single video into atomic skills."""
+        pass
+
+    @abstractmethod
+    def segment_skills_batch(
+        self,
+        video_paths: list[Path],
+        episode_durations: list[float],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[list[Skill]]:
+        """Segment multiple videos into atomic skills in a single batch."""
+        pass
+
+    def _parse_skills_response(self, response: str) -> list[Skill]:
+        """Parse JSON skill list from VLM response text."""
+        if "```json" in response:
+            response = response.split("```json")[1].split("```")[0]
+        elif "```" in response:
+            response = response.split("```")[1].split("```")[0]
+
+        try:
+            data = json.loads(response)
+            skills_data = data.get("skills", data)
+            if isinstance(skills_data, list):
+                return [Skill.from_dict(s) for s in skills_data]
+        except json.JSONDecodeError:
+            match = re.search(r"\{.*\}", response, re.DOTALL)
+            if match:
+                try:
+                    data = json.loads(match.group())
+                    skills_data = data.get("skills", [])
+                    return [Skill.from_dict(s) for s in skills_data]
+                except json.JSONDecodeError as e:
+                    raise ValueError(f"Could not parse JSON from VLM response: {response[:200]}...") from e
+
+        raise ValueError(f"Could not parse skills from response: {response[:200]}...")
+
+
+class QwenVL(BaseVLM):
+    """Qwen VL model for skill segmentation (default: Qwen3.5 series).
+
+    Uses qwen-vl-utils for video processing and the HuggingFace transformers
+    Qwen3VLProcessor pipeline. Requires transformers >= 5.4.0 for correct
+    video position embeddings.
+    """
+
+    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 AutoModelForImageTextToText, AutoProcessor
+
+        self.device = device
+        self.model_name = model_name
+        self.process_vision_info = process_vision_info
+
+        logger.info(f"Loading model: {model_name}...")
+
+        self.model = AutoModelForImageTextToText.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.processor.tokenizer.padding_side = "left"
+
+        logger.info(f"Model loaded on {device}")
+
+    def _build_messages(self, video_path: Path, episode_duration: float, prompt: str) -> list[dict]:
+        duration_str = f"{int(episode_duration // 60):02d}:{int(episode_duration % 60):02d}"
+        return [
+            {"role": "system", "content": [{"type": "text", "text": prompt}]},
+            {
+                "role": "user",
+                "content": [
+                    {"type": "video", "video": str(video_path), "fps": 1.0},
+                    {
+                        "type": "text",
+                        "text": (
+                            f"Video duration: {duration_str} (exactly {episode_duration:.1f} seconds). "
+                            f"Segment into atomic skills. Last skill must end at {episode_duration:.1f}."
+                        ),
+                    },
+                ],
+            },
+        ]
+
+    def _prepare_inputs(self, messages: list[dict]) -> dict:
+        """Tokenize a single message and return processor inputs on device."""
+        text = self.processor.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
+        )
+        image_inputs, video_inputs = self.process_vision_info(messages, return_video_metadata=True)
+
+        videos, video_metadata = None, None
+        if video_inputs:
+            videos = [v[0] for v in video_inputs]
+            video_metadata = [v[1] for v in video_inputs]
+
+        return self.processor(
+            text=[text],
+            images=image_inputs,
+            videos=videos,
+            videos_kwargs={
+                "video_metadata": video_metadata,
+                "do_sample_frames": False,
+            },
+            padding=True,
+            return_tensors="pt",
+        ).to(self.device)
+
+    def _decode(self, inputs, generated_ids) -> list[str]:
+        return self.processor.batch_decode(
+            [out[len(inp) :] for inp, out in zip(inputs.input_ids, generated_ids, strict=True)],
+            skip_special_tokens=True,
+            clean_up_tokenization_spaces=False,
+        )
+
+    def segment_skills(
+        self,
+        video_path: Path,
+        episode_duration: float,
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[Skill]:
+        prompt = create_skill_segmentation_prompt(
+            coarse_goal, subtask_labels, duration_seconds=episode_duration
+        )
+        messages = self._build_messages(video_path, episode_duration, prompt)
+        inputs = self._prepare_inputs(messages)
+
+        with torch.no_grad():
+            generated_ids = self.model.generate(
+                **inputs, max_new_tokens=1024, do_sample=True, temperature=0.7
+            )
+
+        response = self._decode(inputs, generated_ids)[0].strip()
+        return self._parse_skills_response(response)
+
+    def segment_skills_batch(
+        self,
+        video_paths: list[Path],
+        episode_durations: list[float],
+        coarse_goal: str | None = None,
+        subtask_labels: list[str] | None = None,
+    ) -> list[list[Skill]]:
+        all_texts = []
+        all_video_tuples: list[tuple] = []
+
+        for video_path, duration in zip(video_paths, episode_durations, strict=True):
+            prompt = create_skill_segmentation_prompt(coarse_goal, subtask_labels, duration_seconds=duration)
+            messages = self._build_messages(video_path, duration, prompt)
+
+            text = self.processor.apply_chat_template(
+                messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
+            )
+            _image_inputs, video_inputs = self.process_vision_info(messages, return_video_metadata=True)
+            all_texts.append(text)
+            all_video_tuples.extend(video_inputs or [])
+
+        videos, video_metadata = None, None
+        if all_video_tuples:
+            videos = [v[0] for v in all_video_tuples]
+            video_metadata = [v[1] for v in all_video_tuples]
+
+        inputs = self.processor(
+            text=all_texts,
+            videos=videos,
+            videos_kwargs={
+                "video_metadata": video_metadata,
+                "do_sample_frames": False,
+            },
+            padding=True,
+            return_tensors="pt",
+        ).to(self.device)
+
+        with torch.no_grad():
+            generated_ids = self.model.generate(
+                **inputs, max_new_tokens=1024, do_sample=True, temperature=0.7
+            )
+
+        responses = self._decode(inputs, generated_ids)
+
+        all_skills = []
+        for idx, response in enumerate(responses):
+            try:
+                skills = self._parse_skills_response(response.strip())
+                if not skills:
+                    logger.warning(f"No skills parsed for video {idx}")
+                all_skills.append(skills)
+            except Exception as e:
+                logger.warning(f"Failed to parse response for video {idx}: {e}")
+                all_skills.append([])
+
+        return all_skills
+
+
+def get_vlm(model_name: str, device: str = "cuda", torch_dtype: torch.dtype = torch.bfloat16) -> BaseVLM:
+    """Create a VLM instance. Defaults to QwenVL which supports the Qwen3.5 series."""
+    return QwenVL(model_name, device, torch_dtype)

src/lerobot/datasets/aggregate.py

@@ -289,7 +289,9 @@ 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=unique_tasks)
+    dst_meta.tasks = pd.DataFrame(
+        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
+    )
 
     meta_idx = {"chunk": 0, "file": 0}
     data_idx = {"chunk": 0, "file": 0}

src/lerobot/datasets/card_template.md

@@ -7,6 +7,13 @@
 
 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
+{% if repo_id is defined and repo_id %}
+<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path={{ repo_id }}">
+<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
+<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
+</a>
+{% endif %}
+
 ## Dataset Description
 
 {{ dataset_description | default("", true) }}

src/lerobot/datasets/dataset_tools.py

@@ -89,8 +89,8 @@ def delete_episodes(
     Args:
         dataset: The source LeRobotDataset.
         episode_indices: List of episode indices to delete.
-        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
     """
     if not episode_indices:
         raise ValueError("No episodes to delete")
@@ -152,7 +152,7 @@ def split_dataset(
         dataset: The source LeRobotDataset to split.
         splits: Either a dict mapping split names to episode indices, or a dict mapping
                 split names to fractions (must sum to <= 1.0).
-        output_dir: Base directory for output datasets. If None, uses default location.
+        output_dir: Root directory where the split datasets will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
 
     Examples:
       Split by specific episodes
@@ -243,8 +243,8 @@ def merge_datasets(
 
     Args:
         datasets: List of LeRobotDatasets to merge.
-        output_repo_id: Repository ID for the merged dataset.
-        output_dir: Directory to save the merged dataset. If None, uses default location.
+        output_repo_id: Merged dataset identifier.
+        output_dir: Root directory where the merged dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/output_repo_id.
     """
     if not datasets:
         raise ValueError("No datasets to merge")
@@ -288,8 +288,8 @@ def modify_features(
         dataset: The source LeRobotDataset.
         add_features: Optional dict mapping feature names to (feature_values, feature_info) tuples.
         remove_features: Optional feature name(s) to remove. Can be a single string or list.
-        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
 
     Returns:
         New dataset with features modified.
@@ -390,8 +390,8 @@ def add_features(
     Args:
         dataset: The source LeRobotDataset.
         features: Dictionary mapping feature names to (feature_values, feature_info) tuples.
-        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
 
     Returns:
         New dataset with all features added.
@@ -427,8 +427,8 @@ def remove_feature(
     Args:
         dataset: The source LeRobotDataset.
         feature_names: Name(s) of features to remove. Can be a single string or list.
-        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
 
     Returns:
         New dataset with features removed.
@@ -567,20 +567,22 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
     input_path: Path,
     output_path: Path,
-    episodes_to_keep: list[tuple[float, float]],
+    episodes_to_keep: list[tuple[int, int]],
     fps: float,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
-    This function decodes frames from specified time ranges and re-encodes them with
+    This function decodes frames from specified frame ranges and re-encodes them with
     properly reset timestamps to ensure monotonic progression.
 
     Args:
         input_path: Source video file path.
         output_path: Destination video file path.
-        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
+        episodes_to_keep: List of (start_frame, end_frame) tuples for episodes to keep.
+            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
+            is inclusive and end_frame is exclusive.
         fps: Frame rate of the video.
         vcodec: Video codec to use for encoding.
         pix_fmt: Pixel format for output video.
@@ -622,9 +624,10 @@ def _keep_episodes_from_video_with_av(
 
     # Create set of (start, end) ranges for fast lookup.
     # Convert to a sorted list for efficient checking.
-    time_ranges = sorted(episodes_to_keep)
+    frame_ranges = sorted(episodes_to_keep)
 
     # Track frame index for setting PTS and current range being processed.
+    src_frame_count = 0
     frame_count = 0
     range_idx = 0
 
@@ -634,21 +637,20 @@ def _keep_episodes_from_video_with_av(
             if frame is None:
                 continue
 
-            # Get frame timestamp.
-            frame_time = float(frame.pts * frame.time_base) if frame.pts is not None else 0.0
-
-            # Check if frame is in any of our desired time ranges.
+            # Check if frame is in any of our desired frame ranges.
             # Skip ranges that have already passed.
-            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
+            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
                 range_idx += 1
 
             # If we've passed all ranges, stop processing.
-            if range_idx >= len(time_ranges):
+            if range_idx >= len(frame_ranges):
                 break
 
             # Check if frame is in current range.
-            start_ts, end_ts = time_ranges[range_idx]
-            if frame_time < start_ts:
+            start_frame = frame_ranges[range_idx][0]
+
+            if src_frame_count < start_frame:
+                src_frame_count += 1
                 continue
 
             # Frame is in range - create a new frame with reset timestamps.
@@ -661,6 +663,7 @@ def _keep_episodes_from_video_with_av(
             for pkt in v_out.encode(new_frame):
                 out.mux(pkt)
 
+            src_frame_count += 1
             frame_count += 1
 
     # Flush encoder.
@@ -749,15 +752,17 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of time ranges to keep, in sorted order.
+                # Build list of frame ranges to keep, in sorted order.
                 sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
-                episodes_to_keep_ranges: list[tuple[float, float]] = []
-
+                episodes_to_keep_ranges: list[tuple[int, int]] = []
                 for old_idx in sorted_keep_episodes:
                     src_ep = src_dataset.meta.episodes[old_idx]
-                    from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
-                    to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
-                    episodes_to_keep_ranges.append((from_ts, to_ts))
+                    from_frame = round(src_ep[f"videos/{video_key}/from_timestamp"] * src_dataset.meta.fps)
+                    to_frame = round(src_ep[f"videos/{video_key}/to_timestamp"] * src_dataset.meta.fps)
+                    assert src_ep["length"] == to_frame - from_frame, (
+                        f"Episode length mismatch: {src_ep['length']} vs {to_frame - from_frame}"
+                    )
+                    episodes_to_keep_ranges.append((from_frame, to_frame))
 
                 # Use PyAV filters to efficiently re-encode only the desired segments.
                 assert src_dataset.meta.video_path is not None
@@ -1470,7 +1475,9 @@ def modify_tasks(
 
     # Collect all unique tasks and create new task mapping
     unique_tasks = sorted(set(episode_to_task.values()))
-    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
+    new_task_df = pd.DataFrame(
+        {"task_index": list(range(len(unique_tasks)))}, index=pd.Index(unique_tasks, name="task")
+    )
     task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
 
     logging.info(f"Modifying tasks in {dataset.repo_id}")
@@ -1524,7 +1531,7 @@ def modify_tasks(
 
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
-    output_dir: Path,
+    output_dir: Path | None = None,
     repo_id: str | None = None,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -1543,8 +1550,8 @@ def convert_image_to_video_dataset(
 
     Args:
         dataset: The source LeRobot dataset with images
-        output_dir: Directory to save the new video dataset
-        repo_id: Repository ID for the new dataset (default: original_id + "_video")
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
         vcodec: Video codec (default: libsvtav1)
         pix_fmt: Pixel format (default: yuv420p)
         g: Group of pictures size (default: 2)
@@ -1595,6 +1602,7 @@ def convert_image_to_video_dataset(
             # Video info will be updated after episodes are encoded
 
     # Create new metadata for video dataset
+    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
     new_meta = LeRobotDatasetMetadata.create(
         repo_id=repo_id,
         fps=dataset.meta.fps,

src/lerobot/datasets/lerobot_dataset.py

@@ -68,6 +68,7 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
+    StreamingVideoEncoder,
     VideoFrame,
     concatenate_video_files,
     decode_video_frames,
@@ -75,11 +76,11 @@ from lerobot.datasets.video_utils import (
     get_safe_default_codec,
     get_video_duration_in_s,
     get_video_info,
+    resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -313,7 +314,7 @@ class LeRobotDatasetMetadata:
         if self.tasks is None:
             new_tasks = tasks
             task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(tasks, name="task"))
         else:
             new_tasks = [task for task in tasks if task not in self.tasks.index]
             new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
@@ -545,12 +546,19 @@ class LeRobotDatasetMetadata:
 
 
 def _encode_video_worker(
-    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
+    video_key: str,
+    episode_index: int,
+    root: Path,
+    fps: int,
+    vcodec: str = "libsvtav1",
+    encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
+    encode_video_frames(
+        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
+    )
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -570,6 +578,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
+        streaming_encoding: bool = False,
+        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -653,11 +664,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         for the README).
 
         Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
-                will be stored under root/repo_id.
-            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
-                '~/.cache/huggingface/lerobot'.
+            repo_id (str): This is the repo id that will be used to fetch the dataset.
+            root (Path | None, optional): Local directory where the dataset will be downloaded and
+                stored. If set, all dataset files will be stored directly under this path. If not set, the
+                dataset files will be stored under $HF_LEROBOT_HOME/repo_id (configurable via the
+                HF_LEROBOT_HOME environment variable).
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
             image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
@@ -683,12 +694,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
             vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
-                encoding is CPU-heavy.
+                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
+                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
+            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
+                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
+            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
+                streaming encoding. Defaults to 30 (~1s at 30fps).
+            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
+                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
+                libsvtav1 and 'threads' for h264/hevc.
         """
         super().__init__()
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -700,7 +716,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = vcodec
+        self.vcodec = resolve_vcodec(vcodec)
+        self._encoder_threads = encoder_threads
 
         # Unused attributes
         self.image_writer = None
@@ -708,6 +725,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
+        self._streaming_encoder = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -729,7 +747,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Check if cached dataset contains all requested episodes
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
-        except (AssertionError, FileNotFoundError, NotADirectoryError):
+        except (FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
                 self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
@@ -749,6 +767,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
+        # Initialize streaming encoder for resumed recording
+        if streaming_encoding and len(self.meta.video_keys) > 0:
+            self._streaming_encoder = StreamingVideoEncoder(
+                fps=self.meta.fps,
+                vcodec=self.vcodec,
+                pix_fmt="yuv420p",
+                g=2,
+                crf=30,
+                preset=None,
+                queue_maxsize=encoder_queue_maxsize,
+                encoder_threads=encoder_threads,
+            )
+
     def _close_writer(self) -> None:
         """Close and cleanup the parquet writer if it exists."""
         writer = getattr(self, "writer", None)
@@ -808,7 +839,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             hub_api.upload_folder(**upload_kwargs)
 
         card = create_lerobot_dataset_card(
-            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, repo_id=self.repo_id, **card_kwargs
         )
         card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
 
@@ -1104,6 +1135,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         self._close_writer()
         self.meta._close_writer()
+        if self._streaming_encoder is not None:
+            self._streaming_encoder.close()
 
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
@@ -1158,6 +1191,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing
 
+        # Start streaming encoder on first frame of episode (once, before iterating keys)
+        if frame_index == 0 and self._streaming_encoder is not None:
+            self._streaming_encoder.start_episode(
+                video_keys=list(self.meta.video_keys),
+                temp_dir=self.root,
+            )
+
         # Add frame features to episode_buffer
         for key in frame:
             if key not in self.features:
@@ -1165,7 +1205,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                 )
 
-            if self.features[key]["dtype"] in ["image", "video"]:
+            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
+                self._streaming_encoder.feed_frame(key, frame[key])
+                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
+            elif self.features[key]["dtype"] in ["image", "video"]:
                 img_path = self._get_image_file_path(
                     episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                 )
@@ -1226,13 +1269,38 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-        ep_stats = compute_episode_stats(episode_buffer, self.features)
 
-        ep_metadata = self._save_episode_data(episode_buffer)
         has_video_keys = len(self.meta.video_keys) > 0
+        use_streaming = self._streaming_encoder is not None and has_video_keys
         use_batched_encoding = self.batch_encoding_size > 1
 
-        if has_video_keys and not use_batched_encoding:
+        if use_streaming:
+            # Compute stats for non-video features only (video stats come from encoder)
+            non_video_buffer = {
+                k: v
+                for k, v in episode_buffer.items()
+                if self.features.get(k, {}).get("dtype") not in ("video",)
+            }
+            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
+            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
+        else:
+            ep_stats = compute_episode_stats(episode_buffer, self.features)
+
+        ep_metadata = self._save_episode_data(episode_buffer)
+
+        if use_streaming:
+            # Finish streaming encoding and collect results
+            streaming_results = self._streaming_encoder.finish_episode()
+            for video_key in self.meta.video_keys:
+                temp_path, video_stats = streaming_results[video_key]
+                if video_stats is not None:
+                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
+                    ep_stats[video_key] = {
+                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
+                        for k, v in video_stats.items()
+                    }
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
+        elif has_video_keys and not use_batched_encoding:
             num_cameras = len(self.meta.video_keys)
             if parallel_encoding and num_cameras > 1:
                 # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1246,6 +1314,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             self.root,
                             self.fps,
                             self.vcodec,
+                            self._encoder_threads,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1514,6 +1583,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
+        # Cancel streaming encoder if active
+        if self._streaming_encoder is not None:
+            self._streaming_encoder.cancel_episode()
+
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1561,7 +1634,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
+        return _encode_video_worker(
+            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
+        )
 
     @classmethod
     def create(
@@ -1578,10 +1653,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
+        metadata_buffer_size: int = 10,
+        streaming_encoding: bool = False,
+        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+        vcodec = resolve_vcodec(vcodec)
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1590,6 +1668,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             features=features,
             root=root,
             use_videos=use_videos,
+            metadata_buffer_size=metadata_buffer_size,
         )
         obj.repo_id = obj.meta.repo_id
         obj.root = obj.meta.root
@@ -1599,6 +1678,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
         obj.vcodec = vcodec
+        obj._encoder_threads = encoder_threads
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1620,6 +1700,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj._lazy_loading = False
         obj._recorded_frames = 0
         obj._writer_closed_for_reading = False
+
+        # Initialize streaming encoder
+        if streaming_encoding and len(obj.meta.video_keys) > 0:
+            obj._streaming_encoder = StreamingVideoEncoder(
+                fps=fps,
+                vcodec=vcodec,
+                pix_fmt="yuv420p",
+                g=2,
+                crf=30,
+                preset=None,
+                queue_maxsize=encoder_queue_maxsize,
+                encoder_threads=encoder_threads,
+            )
+        else:
+            obj._streaming_encoder = None
+
         return obj
 
 
@@ -1675,11 +1771,12 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
             )
         for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
             extra_keys = set(ds.features).difference(intersection_features)
-            logging.warning(
-                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                "other datasets."
-            )
-            self.disabled_features.update(extra_keys)
+            if extra_keys:
+                logging.warning(
+                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                    "other datasets."
+                )
+                self.disabled_features.update(extra_keys)
 
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps

src/lerobot/datasets/transforms.py

@@ -216,16 +216,17 @@ class ImageTransformsConfig:
 
 
 def make_transform_from_config(cfg: ImageTransformConfig):
-    if cfg.type == "Identity":
-        return v2.Identity(**cfg.kwargs)
-    elif cfg.type == "ColorJitter":
-        return v2.ColorJitter(**cfg.kwargs)
-    elif cfg.type == "SharpnessJitter":
+    if cfg.type == "SharpnessJitter":
         return SharpnessJitter(**cfg.kwargs)
-    elif cfg.type == "RandomAffine":
-        return v2.RandomAffine(**cfg.kwargs)
-    else:
-        raise ValueError(f"Transform '{cfg.type}' is not valid.")
+
+    transform_cls = getattr(v2, cfg.type, None)
+    if isinstance(transform_cls, type) and issubclass(transform_cls, Transform):
+        return transform_cls(**cfg.kwargs)
+
+    raise ValueError(
+        f"Transform '{cfg.type}' is not valid. It must be a class in "
+        f"torchvision.transforms.v2 or 'SharpnessJitter'."
+    )
 
 
 class ImageTransforms(Transform):

src/lerobot/datasets/utils.py

@@ -1,4 +1,5 @@
 #!/usr/bin/env python
+from __future__ import annotations
 
 # Copyright 2024 The HuggingFace Inc. team. All rights reserved.
 #
@@ -21,7 +22,11 @@ from collections import deque
 from collections.abc import Iterable, Iterator
 from pathlib import Path
 from pprint import pformat
-from typing import Any, Generic, TypeVar
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from lerobot.data_processing.data_annotations.subtask_annotations import EpisodeSkills
+    from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 import datasets
 import numpy as np
@@ -78,8 +83,6 @@ DEFAULT_FEATURES = {
     "task_index": {"dtype": "int64", "shape": (1,), "names": None},
 }
 
-T = TypeVar("T")
-
 
 def get_parquet_file_size_in_mb(parquet_path: str | Path) -> float:
     metadata = pq.read_metadata(parquet_path)
@@ -122,19 +125,9 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
-        # PyArrow loads the entire dataset into memory
-        if episodes is None:
-            return Dataset.from_parquet([str(path) for path in paths], features=features)
-
-        arrow_dataset = pa_ds.dataset(paths, format="parquet")
-        filter_expr = pa_ds.field("episode_index").isin(episodes)
-        table = arrow_dataset.to_table(filter=filter_expr)
-
-        if features is not None:
-            table = table.cast(features.arrow_schema)
-
-        return Dataset(table)
+        # We use .from_parquet() memory-mapped loading for efficiency
+        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -351,6 +344,7 @@ def write_tasks(tasks: pandas.DataFrame, local_dir: Path) -> None:
 
 def load_tasks(local_dir: Path) -> pandas.DataFrame:
     tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
+    tasks.index.name = "task"
     return tasks
 
 
@@ -1227,6 +1221,111 @@ def find_float_index(target, float_list, threshold=1e-6):
     return -1
 
 
+def create_subtasks_dataframe(
+    annotations: dict[int, EpisodeSkills],
+) -> tuple[pd.DataFrame, dict[str, int]]:
+    """
+    Create a subtasks DataFrame from skill annotations.
+
+    Args:
+        annotations: Dictionary of episode skills
+
+    Returns:
+        Tuple of (subtasks_df, skill_to_subtask_idx mapping)
+    """
+    # Collect all unique skill names
+    all_skill_names: set[str] = set()
+    for episode_skills in annotations.values():
+        for skill in episode_skills.skills:
+            all_skill_names.add(skill.name)
+    # Build subtasks DataFrame
+    subtask_data = []
+    for i, skill_name in enumerate(sorted(all_skill_names)):
+        subtask_data.append(
+            {
+                "subtask": skill_name,
+                "subtask_index": i,
+            }
+        )
+
+    if not subtask_data:
+        subtasks_df = pd.DataFrame(columns=["subtask", "subtask_index"]).set_index("subtask")
+    else:
+        subtasks_df = pd.DataFrame(subtask_data).set_index("subtask")
+
+    # Build skill name to subtask_index mapping
+    skill_to_subtask_idx = {
+        skill_name: int(subtasks_df.loc[skill_name, "subtask_index"]) for skill_name in all_skill_names
+    }
+
+    return subtasks_df, skill_to_subtask_idx
+
+
+def save_subtasks(
+    subtasks_df: pd.DataFrame,
+    dataset_root: Path,
+) -> None:
+    """Save subtasks to subtasks.parquet."""
+    output_path = dataset_root / "meta" / "subtasks.parquet"
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    subtasks_df.to_parquet(output_path, engine="pyarrow", compression="snappy")
+
+
+def create_subtask_index_array(
+    dataset: LeRobotDataset,
+    annotations: dict[int, EpisodeSkills],
+    skill_to_subtask_idx: dict[str, int],
+) -> np.ndarray:
+    """
+    Create a subtask_index array for each frame based on skill annotations.
+
+    Args:
+        dataset: The LeRobot dataset
+        annotations: Dictionary of episode skills
+        skill_to_subtask_idx: Mapping from skill name to subtask_index
+
+    Returns:
+        Array of subtask indices for each frame in the dataset
+    """
+    # Array to store subtask index for each frame
+    # Initialize with -1 to indicate unannotated frames
+    full_dataset_length = len(dataset)
+    subtask_indices = np.full(full_dataset_length, -1, dtype=np.int64)
+
+    # Assign subtask_index for each annotated episode
+    fps = float(dataset.meta.fps)
+    for ep_idx, episode_skills in annotations.items():
+        skills = episode_skills.skills
+
+        # Get episode frame range
+        ep = dataset.meta.episodes[ep_idx]
+        ep_from = int(ep["dataset_from_index"])
+        ep_to = int(ep["dataset_to_index"])
+
+        # Process each frame in the episode (compute timestamp from index to avoid loading video)
+        for frame_idx in range(ep_from, ep_to):
+            timestamp = (frame_idx - ep_from) / fps
+
+            # Find which skill covers this timestamp (inline to avoid circular import)
+            skill = None
+            for s in skills:
+                if s.start <= timestamp < s.end:
+                    skill = s
+                    break
+                if timestamp >= s.end and s == skills[-1]:
+                    skill = s
+                    break
+            if not skill and skills:
+                skill = skills[-1]
+
+            if skill and skill.name in skill_to_subtask_idx:
+                subtask_idx = skill_to_subtask_idx[skill.name]
+                subtask_indices[frame_idx] = subtask_idx
+
+    return subtask_indices
+
+
 class LookBackError(Exception):
     """
     Exception raised when trying to look back in the history of a Backtrackable object.
@@ -1243,7 +1342,7 @@ class LookAheadError(Exception):
     pass
 
 
-class Backtrackable(Generic[T]):
+class Backtrackable[T]:
     """
     Wrap any iterator/iterable so you can step back up to `history` items
     and look ahead up to `lookahead` items.
@@ -1290,7 +1389,7 @@ class Backtrackable(Generic[T]):
         self._history = history
         self._lookahead = lookahead
 
-    def __iter__(self) -> "Backtrackable[T]":
+    def __iter__(self) -> Backtrackable[T]:
         return self
 
     def __next__(self) -> T:

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -36,8 +36,11 @@ Convert a local dataset (works in place):
 ```bash
 python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
     --repo-id=lerobot/pusht \
-    --root=/path/to/local/dataset/directory
+    --root=/path/to/local/dataset/directory \
     --push-to-hub=false
+
+N.B. Path semantics (v2): --root is the exact dataset folder containing
+meta/, data/, videos/. When omitted, defaults to $HF_LEROBOT_HOME/{repo_id}.
 ```
 
 """
@@ -105,7 +108,7 @@ episodes.jsonl
 {"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
 
 NEW
-meta/episodes/chunk-000/episodes_000.parquet
+meta/episodes/chunk-000/file_000.parquet
 episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
 -------------------------
 OLD
@@ -113,15 +116,16 @@ tasks.jsonl
 {"task_index": 1, "task": "Put the blue block in the green bowl"}
 
 NEW
-meta/tasks/chunk-000/file_000.parquet
+meta/tasks.parquet
 task_index | task
 -------------------------
 OLD
 episodes_stats.jsonl
+{"episode_index": 1, "stats": {"feature_name": {"min": ..., "max": ..., "mean": ..., "std": ..., "count": ...}}}
 
 NEW
-meta/episodes_stats/chunk-000/file_000.parquet
-episode_index | mean | std | min | max
+meta/episodes/chunk-000/file_000.parquet
+episode_index | feature_name/min | feature_name/max | feature_name/mean | feature_name/std | feature_name/count
 -------------------------
 UPDATE
 meta/info.json
@@ -170,7 +174,7 @@ def convert_tasks(root, new_root):
     tasks, _ = legacy_load_tasks(root)
     task_indices = tasks.keys()
     task_strings = tasks.values()
-    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
+    df_tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(task_strings, name="task"))
     write_tasks(df_tasks, new_root)
 
 
@@ -201,7 +205,6 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
 
     image_keys = get_image_keys(root)
 
-    ep_idx = 0
     chunk_idx = 0
     file_idx = 0
     size_in_mb = 0
@@ -211,9 +214,23 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
 
     logging.info(f"Converting data files from {len(ep_paths)} episodes")
 
-    for ep_path in tqdm.tqdm(ep_paths, desc="convert data files"):
+    for ep_idx, ep_path in enumerate(tqdm.tqdm(ep_paths, desc="convert data files")):
         ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
         ep_num_frames = get_parquet_num_frames(ep_path)
+
+        # Check if we need to start a new file BEFORE creating metadata
+        if size_in_mb + ep_size_in_mb >= data_file_size_in_mb and len(paths_to_cat) > 0:
+            # Write the accumulated data files
+            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
+
+            # Move to next file
+            chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
+
+            # Reset for the next file
+            size_in_mb = 0
+            paths_to_cat = []
+
+        # Now create metadata with correct chunk/file indices
         ep_metadata = {
             "episode_index": ep_idx,
             "data/chunk_index": chunk_idx,
@@ -224,20 +241,7 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
         size_in_mb += ep_size_in_mb
         num_frames += ep_num_frames
         episodes_metadata.append(ep_metadata)
-        ep_idx += 1
-
-        if size_in_mb < data_file_size_in_mb:
-            paths_to_cat.append(ep_path)
-            continue
-
-        if paths_to_cat:
-            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-        # Reset for the next file
-        size_in_mb = ep_size_in_mb
-        paths_to_cat = [ep_path]
-
-        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
+        paths_to_cat.append(ep_path)
 
     # Write remaining data if any
     if paths_to_cat:
@@ -469,7 +473,7 @@ def convert_dataset(
 
     # Set root based on whether local dataset path is provided
     use_local_dataset = False
-    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root) / repo_id
+    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root)
     if root.exists():
         validate_local_dataset_version(root)
         use_local_dataset = True
@@ -529,7 +533,7 @@ if __name__ == "__main__":
         type=str,
         required=True,
         help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
     )
     parser.add_argument(
         "--branch",
@@ -553,7 +557,7 @@ if __name__ == "__main__":
         "--root",
         type=str,
         default=None,
-        help="Local directory to use for downloading/writing the dataset.",
+        help="Local directory to use for downloading/writing the dataset. Defaults to $HF_LEROBOT_HOME/repo_id.",
     )
     parser.add_argument(
         "--push-to-hub",

src/lerobot/datasets/video_utils.py

@@ -13,25 +13,106 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import contextlib
 import glob
 import importlib
 import logging
+import queue
 import shutil
 import tempfile
+import threading
 import warnings
 from dataclasses import dataclass, field
+from fractions import Fraction
 from pathlib import Path
 from threading import Lock
 from typing import Any, ClassVar
 
 import av
 import fsspec
+import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
+# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
+# Determines the order of preference for auto-selection when vcodec="auto" is used.
+HW_ENCODERS = [
+    "h264_videotoolbox",  # macOS
+    "hevc_videotoolbox",  # macOS
+    "h264_nvenc",  # NVIDIA GPU
+    "hevc_nvenc",  # NVIDIA GPU
+    "h264_vaapi",  # Linux Intel/AMD
+    "h264_qsv",  # Intel Quick Sync
+]
+
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
+
+
+def _get_codec_options(
+    vcodec: str,
+    g: int | None = 2,
+    crf: int | None = 30,
+    preset: int | None = None,
+) -> dict:
+    """Build codec-specific options dict for video encoding."""
+    options = {}
+
+    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
+    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
+        options["g"] = str(g)
+
+    # Quality control (codec-specific parameter names)
+    if crf is not None:
+        if vcodec in ("h264", "hevc", "libsvtav1"):
+            options["crf"] = str(crf)
+        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
+            quality = max(1, min(100, int(100 - crf * 2)))
+            options["q:v"] = str(quality)
+        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
+            options["rc"] = "constqp"
+            options["qp"] = str(crf)
+        elif vcodec in ("h264_vaapi",):
+            options["qp"] = str(crf)
+        elif vcodec in ("h264_qsv",):
+            options["global_quality"] = str(crf)
+
+    # Preset (only for libsvtav1)
+    if vcodec == "libsvtav1":
+        options["preset"] = str(preset) if preset is not None else "12"
+
+    return options
+
+
+def detect_available_hw_encoders() -> list[str]:
+    """Probe PyAV/FFmpeg for available hardware video encoders."""
+    available = []
+    for codec_name in HW_ENCODERS:
+        try:
+            av.codec.Codec(codec_name, "w")
+            available.append(codec_name)
+        except Exception:  # nosec B110
+            pass  # nosec B110
+    return available
+
+
+def resolve_vcodec(vcodec: str) -> str:
+    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
+    if vcodec not in VALID_VIDEO_CODECS:
+        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+    if vcodec != "auto":
+        logging.info(f"Using video codec: {vcodec}")
+        return vcodec
+    available = detect_available_hw_encoders()
+    for encoder in HW_ENCODERS:
+        if encoder in available:
+            logging.info(f"Auto-selected video codec: {encoder}")
+            return encoder
+    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
+    return "libsvtav1"
+
 
 def get_safe_default_codec():
     if importlib.util.find_spec("torchcodec"):
@@ -146,16 +227,17 @@ def decode_video_frames_torchvision(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    assert is_within_tol.all(), (
-        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-        "It means that the closest frame that can be loaded from the video is too far away in time."
-        "This might be due to synchronization issues with timestamps during data collection."
-        "To be safe, we advise to ignore this item during training."
-        f"\nqueried timestamps: {query_ts}"
-        f"\nloaded timestamps: {loaded_ts}"
-        f"\nvideo: {video_path}"
-        f"\nbackend: {backend}"
-    )
+    if not is_within_tol.all():
+        raise FrameTimestampError(
+            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+            " It means that the closest frame that can be loaded from the video is too far away in time."
+            " This might be due to synchronization issues with timestamps during data collection."
+            " To be safe, we advise to ignore this item during training."
+            f"\nqueried timestamps: {query_ts}"
+            f"\nloaded timestamps: {loaded_ts}"
+            f"\nvideo: {video_path}"
+            f"\nbackend: {backend}"
+        )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -167,7 +249,11 @@ def decode_video_frames_torchvision(
     # convert to the pytorch format which is float32 in [0,1] range (and channel first)
     closest_frames = closest_frames.type(torch.float32) / 255
 
-    assert len(timestamps) == len(closest_frames)
+    if len(timestamps) != len(closest_frames):
+        raise FrameTimestampError(
+            f"Number of retrieved frames ({len(closest_frames)}) does not match "
+            f"number of queried timestamps ({len(timestamps)})"
+        )
     return closest_frames
 
 
@@ -272,15 +358,16 @@ def decode_video_frames_torchcodec(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    assert is_within_tol.all(), (
-        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-        "It means that the closest frame that can be loaded from the video is too far away in time."
-        "This might be due to synchronization issues with timestamps during data collection."
-        "To be safe, we advise to ignore this item during training."
-        f"\nqueried timestamps: {query_ts}"
-        f"\nloaded timestamps: {loaded_ts}"
-        f"\nvideo: {video_path}"
-    )
+    if not is_within_tol.all():
+        raise FrameTimestampError(
+            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+            " It means that the closest frame that can be loaded from the video is too far away in time."
+            " This might be due to synchronization issues with timestamps during data collection."
+            " To be safe, we advise to ignore this item during training."
+            f"\nqueried timestamps: {query_ts}"
+            f"\nloaded timestamps: {loaded_ts}"
+            f"\nvideo: {video_path}"
+        )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -309,14 +396,13 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.ERROR,
+    log_level: int | None = av.logging.WARNING,
     overwrite: bool = False,
     preset: int | None = None,
+    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    # Check encoder availability
-    if vcodec not in ["h264", "hevc", "libsvtav1"]:
-        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
+    vcodec = resolve_vcodec(vcodec)
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -347,21 +433,22 @@ def encode_video_frames(
         width, height = dummy_image.size
 
     # Define video codec options
-    video_options = {}
-
-    if g is not None:
-        video_options["g"] = str(g)
-
-    if crf is not None:
-        video_options["crf"] = str(crf)
+    video_options = _get_codec_options(vcodec, g, crf, preset)
 
     if fast_decode:
         key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         video_options[key] = value
 
-    if vcodec == "libsvtav1":
-        video_options["preset"] = str(preset) if preset is not None else "12"
+    if encoder_threads is not None:
+        if vcodec == "libsvtav1":
+            lp_param = f"lp={encoder_threads}"
+            if "svtav1-params" in video_options:
+                video_options["svtav1-params"] += f":{lp_param}"
+            else:
+                video_options["svtav1-params"] = lp_param
+        else:
+            video_options["threads"] = str(encoder_threads)
 
     # Set logging level
     if log_level is not None:
@@ -480,6 +567,348 @@ def concatenate_video_files(
     Path(tmp_concatenate_path).unlink()
 
 
+class _CameraEncoderThread(threading.Thread):
+    """A thread that encodes video frames streamed via a queue into an MP4 file.
+
+    One instance is created per camera per episode. Frames are received as numpy arrays
+    from the main thread, encoded in real-time using PyAV (which releases the GIL during
+    encoding), and written to disk. Stats are computed incrementally using
+    RunningQuantileStats and returned via result_queue.
+    """
+
+    def __init__(
+        self,
+        video_path: Path,
+        fps: int,
+        vcodec: str,
+        pix_fmt: str,
+        g: int | None,
+        crf: int | None,
+        preset: int | None,
+        frame_queue: queue.Queue,
+        result_queue: queue.Queue,
+        stop_event: threading.Event,
+        encoder_threads: int | None = None,
+    ):
+        super().__init__(daemon=True)
+        self.video_path = video_path
+        self.fps = fps
+        self.vcodec = vcodec
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.preset = preset
+        self.frame_queue = frame_queue
+        self.result_queue = result_queue
+        self.stop_event = stop_event
+        self.encoder_threads = encoder_threads
+
+    def run(self) -> None:
+        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
+
+        container = None
+        output_stream = None
+        stats_tracker = RunningQuantileStats()
+        frame_count = 0
+
+        try:
+            logging.getLogger("libav").setLevel(av.logging.WARNING)
+
+            while True:
+                try:
+                    frame_data = self.frame_queue.get(timeout=1)
+                except queue.Empty:
+                    if self.stop_event.is_set():
+                        break
+                    continue
+
+                if frame_data is None:
+                    # Sentinel: flush and close
+                    break
+
+                # Ensure HWC uint8 numpy array
+                if isinstance(frame_data, np.ndarray):
+                    if frame_data.ndim == 3 and frame_data.shape[0] == 3:
+                        # CHW -> HWC
+                        frame_data = frame_data.transpose(1, 2, 0)
+                    if frame_data.dtype != np.uint8:
+                        frame_data = (frame_data * 255).astype(np.uint8)
+
+                # Open container on first frame (to get width/height)
+                if container is None:
+                    height, width = frame_data.shape[:2]
+                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
+                    if self.encoder_threads is not None:
+                        if self.vcodec == "libsvtav1":
+                            lp_param = f"lp={self.encoder_threads}"
+                            if "svtav1-params" in video_options:
+                                video_options["svtav1-params"] += f":{lp_param}"
+                            else:
+                                video_options["svtav1-params"] = lp_param
+                        else:
+                            video_options["threads"] = str(self.encoder_threads)
+                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
+                    container = av.open(str(self.video_path), "w")
+                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
+                    output_stream.pix_fmt = self.pix_fmt
+                    output_stream.width = width
+                    output_stream.height = height
+                    output_stream.time_base = Fraction(1, self.fps)
+
+                # Encode frame with explicit timestamps
+                pil_img = Image.fromarray(frame_data)
+                video_frame = av.VideoFrame.from_image(pil_img)
+                video_frame.pts = frame_count
+                video_frame.time_base = Fraction(1, self.fps)
+                packet = output_stream.encode(video_frame)
+                if packet:
+                    container.mux(packet)
+
+                # Update stats with downsampled frame (per-channel stats like compute_episode_stats)
+                img_chw = frame_data.transpose(2, 0, 1)  # HWC -> CHW
+                img_downsampled = auto_downsample_height_width(img_chw)
+                # Reshape CHW to (H*W, C) for per-channel stats
+                channels = img_downsampled.shape[0]
+                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
+                stats_tracker.update(img_for_stats)
+
+                frame_count += 1
+
+            # Flush encoder
+            if output_stream is not None:
+                packet = output_stream.encode()
+                if packet:
+                    container.mux(packet)
+
+            if container is not None:
+                container.close()
+
+            av.logging.restore_default_callback()
+
+            # Get stats and put on result queue
+            if frame_count >= 2:
+                stats = stats_tracker.get_statistics()
+                self.result_queue.put(("ok", stats))
+            else:
+                self.result_queue.put(("ok", None))
+
+        except Exception as e:
+            logging.error(f"Encoder thread error: {e}")
+            if container is not None:
+                with contextlib.suppress(Exception):
+                    container.close()
+            self.result_queue.put(("error", str(e)))
+
+
+class StreamingVideoEncoder:
+    """Manages per-camera encoder threads for real-time video encoding during recording.
+
+    Instead of writing frames as PNG images and then encoding to MP4 at episode end,
+    this class streams frames directly to encoder threads, eliminating the
+    PNG round-trip and making save_episode() near-instant.
+
+    Uses threading instead of multiprocessing to avoid the overhead of pickling large
+    numpy arrays through multiprocessing.Queue. PyAV's encode() releases the GIL,
+    so encoding runs in parallel with the main recording loop.
+    """
+
+    def __init__(
+        self,
+        fps: int,
+        vcodec: str = "libsvtav1",
+        pix_fmt: str = "yuv420p",
+        g: int | None = 2,
+        crf: int | None = 30,
+        preset: int | None = None,
+        queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
+    ):
+        self.fps = fps
+        self.vcodec = resolve_vcodec(vcodec)
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.preset = preset
+        self.queue_maxsize = queue_maxsize
+        self.encoder_threads = encoder_threads
+
+        self._frame_queues: dict[str, queue.Queue] = {}
+        self._result_queues: dict[str, queue.Queue] = {}
+        self._threads: dict[str, _CameraEncoderThread] = {}
+        self._stop_events: dict[str, threading.Event] = {}
+        self._video_paths: dict[str, Path] = {}
+        self._dropped_frames: dict[str, int] = {}
+        self._episode_active = False
+
+    def start_episode(self, video_keys: list[str], temp_dir: Path) -> None:
+        """Start encoder threads for a new episode.
+
+        Args:
+            video_keys: List of video feature keys (e.g. ["observation.images.laptop"])
+            temp_dir: Base directory for temporary MP4 files
+        """
+        if self._episode_active:
+            self.cancel_episode()
+
+        self._dropped_frames.clear()
+
+        for video_key in video_keys:
+            frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize)
+            result_queue: queue.Queue = queue.Queue(maxsize=1)
+            stop_event = threading.Event()
+
+            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
+            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
+
+            encoder_thread = _CameraEncoderThread(
+                video_path=video_path,
+                fps=self.fps,
+                vcodec=self.vcodec,
+                pix_fmt=self.pix_fmt,
+                g=self.g,
+                crf=self.crf,
+                preset=self.preset,
+                frame_queue=frame_queue,
+                result_queue=result_queue,
+                stop_event=stop_event,
+                encoder_threads=self.encoder_threads,
+            )
+            encoder_thread.start()
+
+            self._frame_queues[video_key] = frame_queue
+            self._result_queues[video_key] = result_queue
+            self._threads[video_key] = encoder_thread
+            self._stop_events[video_key] = stop_event
+            self._video_paths[video_key] = video_path
+
+        self._episode_active = True
+
+    def feed_frame(self, video_key: str, image: np.ndarray) -> None:
+        """Feed a frame to the encoder for a specific camera.
+
+        A copy of the image is made before enqueueing to prevent race conditions
+        with camera drivers that may reuse buffers. If the encoder queue is full
+        (encoder can't keep up), the frame is dropped with a warning instead of
+        crashing the recording session.
+
+        Args:
+            video_key: The video feature key
+            image: numpy array in (H,W,C) or (C,H,W) format, uint8 or float
+
+        Raises:
+            RuntimeError: If the encoder thread has crashed
+        """
+        if not self._episode_active:
+            raise RuntimeError("No active episode. Call start_episode() first.")
+
+        thread = self._threads[video_key]
+        if not thread.is_alive():
+            # Check for error
+            try:
+                status, msg = self._result_queues[video_key].get_nowait()
+                if status == "error":
+                    raise RuntimeError(f"Encoder thread for {video_key} crashed: {msg}")
+            except queue.Empty:
+                pass
+            raise RuntimeError(f"Encoder thread for {video_key} is not alive")
+
+        try:
+            self._frame_queues[video_key].put(image.copy(), timeout=0.1)
+        except queue.Full:
+            self._dropped_frames[video_key] = self._dropped_frames.get(video_key, 0) + 1
+            count = self._dropped_frames[video_key]
+            # Log periodically to avoid spam (1st, then every 10th)
+            if count == 1 or count % 10 == 0:
+                logging.warning(
+                    f"Encoder queue full for {video_key}, dropped {count} frame(s). "
+                    f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
+                )
+
+    def finish_episode(self) -> dict[str, tuple[Path, dict | None]]:
+        """Finish encoding the current episode.
+
+        Sends sentinel values, waits for encoder threads to complete,
+        and collects results.
+
+        Returns:
+            Dict mapping video_key to (mp4_path, stats_dict_or_None)
+        """
+        if not self._episode_active:
+            raise RuntimeError("No active episode to finish.")
+
+        results = {}
+
+        # Report dropped frames
+        for video_key, count in self._dropped_frames.items():
+            if count > 0:
+                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
+
+        # Send sentinel to all queues
+        for video_key in self._frame_queues:
+            self._frame_queues[video_key].put(None)
+
+        # Wait for all threads and collect results
+        for video_key in self._threads:
+            self._threads[video_key].join(timeout=120)
+            if self._threads[video_key].is_alive():
+                logging.error(f"Encoder thread for {video_key} did not finish in time")
+                self._stop_events[video_key].set()
+                self._threads[video_key].join(timeout=5)
+                results[video_key] = (self._video_paths[video_key], None)
+                continue
+
+            try:
+                status, data = self._result_queues[video_key].get(timeout=5)
+                if status == "error":
+                    raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
+                results[video_key] = (self._video_paths[video_key], data)
+            except queue.Empty:
+                logging.error(f"No result from encoder thread for {video_key}")
+                results[video_key] = (self._video_paths[video_key], None)
+
+        self._cleanup()
+        self._episode_active = False
+        return results
+
+    def cancel_episode(self) -> None:
+        """Cancel the current episode, stopping encoder threads and cleaning up."""
+        if not self._episode_active:
+            return
+
+        # Signal all threads to stop
+        for video_key in self._stop_events:
+            self._stop_events[video_key].set()
+
+        # Wait for threads to finish
+        for video_key in self._threads:
+            self._threads[video_key].join(timeout=5)
+
+            # Clean up temp MP4 files
+            video_path = self._video_paths.get(video_key)
+            if video_path is not None and video_path.exists():
+                shutil.rmtree(str(video_path.parent), ignore_errors=True)
+
+        self._cleanup()
+        self._episode_active = False
+
+    def close(self) -> None:
+        """Close the encoder, canceling any in-progress episode."""
+        if self._episode_active:
+            self.cancel_episode()
+
+    def _cleanup(self) -> None:
+        """Clean up queues and thread tracking dicts."""
+        for q in self._frame_queues.values():
+            with contextlib.suppress(Exception):
+                while not q.empty():
+                    q.get_nowait()
+        self._frame_queues.clear()
+        self._result_queues.clear()
+        self._threads.clear()
+        self._stop_events.clear()
+        self._video_paths.clear()
+
+
 @dataclass
 class VideoFrame:
     # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -514,7 +943,7 @@ with warnings.catch_warnings():
 
 def get_audio_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    logging.getLogger("libav").setLevel(av.logging.WARNING)
 
     # Getting audio stream information
     audio_info = {}
@@ -546,7 +975,7 @@ def get_audio_info(video_path: Path | str) -> dict:
 
 def get_video_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    logging.getLogger("libav").setLevel(av.logging.WARNING)
 
     # Getting video stream information
     video_info = {}
@@ -632,8 +1061,15 @@ class VideoEncodingManager:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        # Handle any remaining episodes that haven't been batch encoded
-        if self.dataset.episodes_since_last_encoding > 0:
+        streaming_encoder = getattr(self.dataset, "_streaming_encoder", None)
+
+        if streaming_encoder is not None:
+            # Handle streaming encoder cleanup
+            if exc_type is not None:
+                streaming_encoder.cancel_episode()
+            streaming_encoder.close()
+        elif self.dataset.episodes_since_last_encoding > 0:
+            # Handle any remaining episodes that haven't been batch encoded
             if exc_type is not None:
                 logging.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
@@ -650,8 +1086,8 @@ class VideoEncodingManager:
         # Finalize the dataset to properly close all writers
         self.dataset.finalize()
 
-        # Clean up episode images if recording was interrupted
-        if exc_type is not None:
+        # Clean up episode images if recording was interrupted (only for non-streaming mode)
+        if exc_type is not None and streaming_encoder is None:
             interrupted_episode_index = self.dataset.num_episodes
             for key in self.dataset.meta.video_keys:
                 img_dir = self.dataset._get_image_file_path(
@@ -665,14 +1101,12 @@ class VideoEncodingManager:
 
         # Clean up any remaining images directory if it's empty
         img_dir = self.dataset.root / "images"
-        # Check for any remaining PNG files
-        png_files = list(img_dir.rglob("*.png"))
-        if len(png_files) == 0:
-            # Only remove the images directory if no PNG files remain
-            if img_dir.exists():
+        if img_dir.exists():
+            png_files = list(img_dir.rglob("*.png"))
+            if len(png_files) == 0:
                 shutil.rmtree(img_dir)
                 logging.debug("Cleaned up empty images directory")
-        else:
-            logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+            else:
+                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
 
         return False  # Don't suppress the original exception

src/lerobot/envs/configs.py

@@ -205,6 +205,7 @@ class ObservationConfig:
 
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
+    add_ee_pose_to_observation: bool = False
     display_cameras: bool = False
 
 

src/lerobot/envs/libero.py

@@ -112,6 +112,7 @@ class LiberoEnv(gym.Env):
         visualization_height: int = 480,
         init_states: bool = True,
         episode_index: int = 0,
+        n_envs: int = 1,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
         control_mode: str = "relative",
@@ -145,7 +146,9 @@ class LiberoEnv(gym.Env):
         self.episode_length = episode_length
         # Load once and keep
         self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
-        self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
+        self._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
+
+        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
 
         self._env = self._make_envs_task(task_suite, self.task_id)
         default_steps = 500
@@ -295,7 +298,8 @@ class LiberoEnv(gym.Env):
         self._env.seed(seed)
         raw_obs = self._env.reset()
         if self.init_states and self._init_states is not None:
-            raw_obs = self._env.set_init_state(self._init_states[self._init_state_id])
+            raw_obs = self._env.set_init_state(self._init_states[self.init_state_id % len(self._init_states)])
+            self.init_state_id += self._reset_stride  # Change init_state_id when reset
 
         # After reset, objects may be unstable (slightly floating, intersecting, etc.).
         # Step the simulator with a no-op action for a few frames so everything settles.
@@ -373,6 +377,7 @@ def _make_env_fns(
             init_states=init_states,
             episode_length=episode_length,
             episode_index=episode_index,
+            n_envs=n_envs,
             control_mode=control_mode,
             **local_kwargs,
         )

src/lerobot/motors/calibration_gui.py

@@ -221,7 +221,7 @@ class RangeFinderGUI:
 
         self.bus = bus
         self.groups = groups if groups is not None else {"all": list(bus.motors)}
-        self.group_names = list(groups)
+        self.group_names = list(self.groups)
         self.current_group = self.group_names[0]
 
         if not bus.is_connected:
@@ -230,18 +230,20 @@ class RangeFinderGUI:
         self.calibration = bus.read_calibration()
         self.res_table = bus.model_resolution_table
         self.present_cache = {
-            m: bus.read("Present_Position", m, normalize=False) for motors in groups.values() for m in motors
+            m: bus.read("Present_Position", m, normalize=False)
+            for motors in self.groups.values()
+            for m in motors
         }
 
         pygame.init()
         self.font = pygame.font.Font(None, FONT_SIZE)
 
-        label_pad = max(self.font.size(m)[0] for ms in groups.values() for m in ms)
+        label_pad = max(self.font.size(m)[0] for ms in self.groups.values() for m in ms)
         self.label_pad = label_pad
         width = 40 + label_pad + BAR_LEN + 6 + BTN_W + 10 + SAVE_W + 10
         self.controls_bottom = 10 + SAVE_H
         self.base_y = self.controls_bottom + TOP_GAP
-        height = self.base_y + PADDING_Y * len(groups[self.current_group]) + 40
+        height = self.base_y + PADDING_Y * len(self.groups[self.current_group]) + 40
 
         self.screen = pygame.display.set_mode((width, height))
         pygame.display.set_caption("Motors range finder")

src/lerobot/motors/damiao/damiao.py

@@ -23,6 +23,7 @@ from copy import deepcopy
 from functools import cached_property
 from typing import TYPE_CHECKING, Any, TypedDict
 
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.import_utils import _can_available
 
 if TYPE_CHECKING or _can_available:
@@ -36,7 +37,6 @@ else:
 
 import numpy as np
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import enter_pressed, move_cursor_up
 
@@ -155,6 +155,7 @@ class DamiaoMotorsBus(MotorsBusBase):
         """Check if the CAN bus is connected."""
         return self._is_connected and self.canbus is not None
 
+    @check_if_already_connected
     def connect(self, handshake: bool = True) -> None:
         """
         Open the CAN bus and initialize communication.
@@ -162,10 +163,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         Args:
             handshake: If True, ping all motors to verify they're present
         """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(
-                f"{self.__class__.__name__}('{self.port}') is already connected."
-            )
 
         try:
             # Auto-detect interface type based on port name
@@ -211,6 +208,9 @@ class DamiaoMotorsBus(MotorsBusBase):
         logger.info("Starting handshake with motors...")
 
         # Drain any pending messages
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         while self.canbus.recv(timeout=0.01):
             pass
 
@@ -246,6 +246,7 @@ class DamiaoMotorsBus(MotorsBusBase):
             )
         logger.info("Handshake successful. All motors ready.")
 
+    @check_if_not_connected
     def disconnect(self, disable_torque: bool = True) -> None:
         """
         Close the CAN bus connection.
@@ -253,8 +254,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         Args:
             disable_torque: If True, disable torque on all motors before disconnecting
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self.__class__.__name__}('{self.port}') is not connected.")
 
         if disable_torque:
             try:
@@ -283,6 +282,10 @@ class DamiaoMotorsBus(MotorsBusBase):
         recv_id = self._get_motor_recv_id(motor)
         data = [0xFF] * 7 + [command_byte]
         msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         self.canbus.send(msg)
         if msg := self._recv_motor_response(expected_recv_id=recv_id):
             self._process_response(motor_name, msg)
@@ -341,6 +344,10 @@ class DamiaoMotorsBus(MotorsBusBase):
         recv_id = self._get_motor_recv_id(motor)
         data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
         msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         self.canbus.send(msg)
         return self._recv_motor_response(expected_recv_id=recv_id)
 
@@ -356,6 +363,10 @@ class DamiaoMotorsBus(MotorsBusBase):
         Returns:
             CAN message if received, None otherwise
         """
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         try:
             start_time = time.time()
             messages_seen = []
@@ -394,10 +405,13 @@ class DamiaoMotorsBus(MotorsBusBase):
         Returns:
             Dictionary mapping recv_id to CAN message
         """
-        responses = {}
+        responses: dict[int, can.Message] = {}
         expected_set = set(expected_recv_ids)
         start_time = time.time()
 
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         try:
             while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
                 # 100us poll timeout
@@ -461,6 +475,9 @@ class DamiaoMotorsBus(MotorsBusBase):
         motor_name = self._get_motor_name(motor)
         motor_type = self._motor_types[motor_name]
 
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
         msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
         self.canbus.send(msg)
@@ -488,6 +505,9 @@ class DamiaoMotorsBus(MotorsBusBase):
 
         recv_id_to_motor: dict[int, str] = {}
 
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         # Step 1: Send all MIT control commands
         for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
             motor_id = self._get_motor_id(motor)
@@ -562,10 +582,9 @@ class DamiaoMotorsBus(MotorsBusBase):
         except Exception as e:
             logger.warning(f"Failed to decode response from {motor}: {e}")
 
+    @check_if_not_connected
     def read(self, data_name: str, motor: str) -> Value:
         """Read a value from a single motor. Positions are always in degrees."""
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         # Refresh motor to get latest state
         msg = self._refresh_motor(motor)
@@ -595,6 +614,7 @@ class DamiaoMotorsBus(MotorsBusBase):
             raise ValueError(f"Unknown data_name: {data_name}")
         return mapping[data_name]
 
+    @check_if_not_connected
     def write(
         self,
         data_name: str,
@@ -605,8 +625,6 @@ class DamiaoMotorsBus(MotorsBusBase):
         Write a value to a single motor. Positions are always in degrees.
         Can write 'Goal_Position', 'Kp', or 'Kd'.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if data_name in ("Kp", "Kd"):
             self._gains[motor][data_name.lower()] = float(value)
@@ -656,6 +674,10 @@ class DamiaoMotorsBus(MotorsBusBase):
 
     def _batch_refresh(self, motors: list[str]) -> None:
         """Internal helper to refresh a list of motors and update cache."""
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
         # Send refresh commands
         for motor in motors:
             motor_id = self._get_motor_id(motor)
@@ -678,10 +700,12 @@ class DamiaoMotorsBus(MotorsBusBase):
             else:
                 logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
 
-    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+    @check_if_not_connected
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
         """
         Write values to multiple motors simultaneously. Positions are always in degrees.
         """
+
         if data_name in ("Kp", "Kd"):
             key = data_name.lower()
             for motor, val in values.items():
@@ -690,6 +714,8 @@ class DamiaoMotorsBus(MotorsBusBase):
         elif data_name == "Goal_Position":
             # Step 1: Send all MIT control commands
             recv_id_to_motor: dict[int, str] = {}
+            if self.canbus is None:
+                raise RuntimeError("CAN bus is not initialized.")
             for motor, value_degrees in values.items():
                 motor_id = self._get_motor_id(motor)
                 motor_name = self._get_motor_name(motor)
@@ -732,9 +758,9 @@ class DamiaoMotorsBus(MotorsBusBase):
 
     def record_ranges_of_motion(
         self,
-        motors: NameOrID | list[NameOrID] | None = None,
+        motors: str | list[str] | None = None,
         display_values: bool = True,
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
         """
         Interactively record the min/max values of each motor in degrees.
 

src/lerobot/motors/dynamixel/dynamixel.py

@@ -181,10 +181,10 @@ class DynamixelMotorsBus(SerialMotorsBus):
         for motor, m in self.motors.items():
             calibration[motor] = MotorCalibration(
                 id=m.id,
-                drive_mode=drive_modes[motor],
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                drive_mode=int(drive_modes[motor]),
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
             )
 
         return calibration
@@ -198,7 +198,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         if cache:
             self.calibration = calibration_dict
 
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
@@ -206,7 +206,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
         self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
 
@@ -235,7 +235,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         On Dynamixel Motors:
         Present_Position = Actual_Position + Homing_Offset
         """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -258,6 +258,6 @@ class DynamixelMotorsBus(SerialMotorsBus):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
 
-            return
+            return None
 
         return {id_: data[0] for id_, data in data_list.items()}

src/lerobot/motors/feetech/feetech.py

@@ -126,7 +126,7 @@ class FeetechMotorsBus(SerialMotorsBus):
 
         self.port_handler = scs.PortHandler(self.port)
         # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(  # type: ignore[method-assign]
             self.port_handler, scs.PortHandler
         )
         self.packet_handler = scs.PacketHandler(protocol_version)
@@ -262,9 +262,9 @@ class FeetechMotorsBus(SerialMotorsBus):
             calibration[motor] = MotorCalibration(
                 id=m.id,
                 drive_mode=0,
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
             )
 
         return calibration
@@ -284,7 +284,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         On Feetech Motors:
         Present_Position = Actual_Position - Homing_Offset
         """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -292,7 +292,7 @@ class FeetechMotorsBus(SerialMotorsBus):
 
         return half_turn_homings
 
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 0, num_retry=num_retry)
@@ -303,7 +303,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         addr, length = get_address(self.model_ctrl_table, model, "Lock")
         self._write(addr, length, motor, 0, num_retry=num_retry)
 
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 1, num_retry=num_retry)
@@ -334,7 +334,7 @@ class FeetechMotorsBus(SerialMotorsBus):
     def _broadcast_ping(self) -> tuple[dict[int, int], int]:
         import scservo_sdk as scs
 
-        data_list = {}
+        data_list: dict[int, int] = {}
 
         status_length = 6
 
@@ -414,7 +414,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         if not self._is_comm_success(comm):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return
+            return None
 
         ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
         if ids_errors:

src/lerobot/motors/motors_bus.py

@@ -23,12 +23,13 @@ from __future__ import annotations
 
 import abc
 import logging
+from collections.abc import Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
 from enum import Enum
 from functools import cached_property
 from pprint import pformat
-from typing import Protocol, TypeAlias
+from typing import Protocol
 
 import serial
 from deepdiff import DeepDiff
@@ -37,8 +38,8 @@ from tqdm import tqdm
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.utils import enter_pressed, move_cursor_up
 
-NameOrID: TypeAlias = str | int
-Value: TypeAlias = int | float
+type NameOrID = str | int
+type Value = int | float
 
 logger = logging.getLogger(__name__)
 
@@ -93,7 +94,7 @@ class MotorsBusBase(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
         """Write values to multiple motors."""
         pass
 
@@ -179,15 +180,16 @@ class Motor:
 
 
 class PortHandler(Protocol):
-    def __init__(self, port_name):
-        self.is_open: bool
-        self.baudrate: int
-        self.packet_start_time: float
-        self.packet_timeout: float
-        self.tx_time_per_byte: float
-        self.is_using: bool
-        self.port_name: str
-        self.ser: serial.Serial
+    is_open: bool
+    baudrate: int
+    packet_start_time: float
+    packet_timeout: float
+    tx_time_per_byte: float
+    is_using: bool
+    port_name: str
+    ser: serial.Serial
+
+    def __init__(self, port_name: str) -> None: ...
 
     def openPort(self): ...
     def closePort(self): ...
@@ -240,19 +242,22 @@ class PacketHandler(Protocol):
     def regWriteTxRx(self, port, id, address, length, data): ...
     def syncReadTx(self, port, start_address, data_length, param, param_length): ...
     def syncWriteTxOnly(self, port, start_address, data_length, param, param_length): ...
+    def broadcastPing(self, port): ...
 
 
 class GroupSyncRead(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.last_result: bool
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    last_result: bool
+    is_param_changed: bool
+    param: list
+    data_dict: dict
 
+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id): ...
     def removeParam(self, id): ...
@@ -265,15 +270,17 @@ class GroupSyncRead(Protocol):
 
 
 class GroupSyncWrite(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    is_param_changed: bool
+    param: list
+    data_dict: dict
 
+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id, data): ...
     def removeParam(self, id): ...
@@ -400,7 +407,7 @@ class SerialMotorsBus(MotorsBusBase):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motor_model(self, motor: NameOrID) -> int:
+    def _get_motor_model(self, motor: NameOrID) -> str:
         if isinstance(motor, str):
             return self.motors[motor].model
         elif isinstance(motor, int):
@@ -408,17 +415,19 @@ class SerialMotorsBus(MotorsBusBase):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+    def _get_motors_list(self, motors: NameOrID | Sequence[NameOrID] | None) -> list[str]:
         if motors is None:
             return list(self.motors)
         elif isinstance(motors, str):
             return [motors]
-        elif isinstance(motors, list):
-            return motors.copy()
+        elif isinstance(motors, int):
+            return [self._id_to_name(motors)]
+        elif isinstance(motors, Sequence):
+            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
         else:
             raise TypeError(motors)
 
-    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
+    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> dict[int, Value]:
         if isinstance(values, (int | float)):
             return dict.fromkeys(self.ids, values)
         elif isinstance(values, dict):
@@ -640,18 +649,19 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     @abc.abstractmethod
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         """Enable torque on selected motors.
 
         Args:
-            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
+            motors (int | str | list[str] | None, optional): Same semantics as :pymeth:`disable_torque`.
+                Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
         """
         pass
 
     @contextmanager
-    def torque_disabled(self, motors: int | str | list[str] | None = None):
+    def torque_disabled(self, motors: str | list[str] | None = None):
         """Context-manager that guarantees torque is re-enabled.
 
         This helper is useful to temporarily disable torque when configuring motors.
@@ -728,24 +738,19 @@ class SerialMotorsBus(MotorsBusBase):
         """
         pass
 
-    def reset_calibration(self, motors: NameOrID | list[NameOrID] | None = None) -> None:
+    def reset_calibration(self, motors: NameOrID | Sequence[NameOrID] | None = None) -> None:
         """Restore factory calibration for the selected motors.
 
         Homing offset is set to ``0`` and min/max position limits are set to the full usable range.
         The in-memory :pyattr:`calibration` is cleared.
 
         Args:
-            motors (NameOrID | list[NameOrID] | None, optional): Selection of motors. `None` (default)
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Selection of motors. `None` (default)
                 resets every motor.
         """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)
 
-        for motor in motors:
+        for motor in motor_names:
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
             self.write("Homing_Offset", motor, 0, normalize=False)
@@ -754,7 +759,9 @@ class SerialMotorsBus(MotorsBusBase):
 
         self.calibration = {}
 
-    def set_half_turn_homings(self, motors: NameOrID | list[NameOrID] | None = None) -> dict[NameOrID, Value]:
+    def set_half_turn_homings(
+        self, motors: NameOrID | Sequence[NameOrID] | None = None
+    ) -> dict[NameOrID, Value]:
         """Centre each motor range around its current position.
 
         The function computes and writes a homing offset such that the present position becomes exactly one
@@ -764,17 +771,12 @@ class SerialMotorsBus(MotorsBusBase):
             motors (NameOrID | list[NameOrID] | None, optional): Motors to adjust. Defaults to all motors (`None`).
 
         Returns:
-            dict[NameOrID, Value]: Mapping *motor → written homing offset*.
+            dict[str, Value]: Mapping *motor name → written homing offset*.
         """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)
 
-        self.reset_calibration(motors)
-        actual_positions = self.sync_read("Present_Position", motors, normalize=False)
+        self.reset_calibration(motor_names)
+        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
         homing_offsets = self._get_half_turn_homings(actual_positions)
         for motor, offset in homing_offsets.items():
             self.write("Homing_Offset", motor, offset)
@@ -786,8 +788,8 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     def record_ranges_of_motion(
-        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        self, motors: NameOrID | Sequence[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
         """Interactively record the min/max encoder values of each motor.
 
         Move the joints by hand (with torque disabled) while the method streams live positions. Press
@@ -799,30 +801,25 @@ class SerialMotorsBus(MotorsBusBase):
             display_values (bool, optional): When `True` (default) a live table is printed to the console.
 
         Returns:
-            tuple[dict[NameOrID, Value], dict[NameOrID, Value]]: Two dictionaries *mins* and *maxes* with the
+            tuple[dict[str, Value], dict[str, Value]]: Two dictionaries *mins* and *maxes* with the
                 extreme values observed for each motor.
         """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)
 
-        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
         mins = start_positions.copy()
         maxes = start_positions.copy()
 
         user_pressed_enter = False
         while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motors, normalize=False)
+            positions = self.sync_read("Present_Position", motor_names, normalize=False)
             mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
             maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}
 
             if display_values:
                 print("\n-------------------------------------------")
                 print(f"{'NAME':<15} | {'MIN':>6} | {'POS':>6} | {'MAX':>6}")
-                for motor in motors:
+                for motor in motor_names:
                     print(f"{motor:<15} | {mins[motor]:>6} | {positions[motor]:>6} | {maxes[motor]:>6}")
 
             if enter_pressed():
@@ -830,9 +827,9 @@ class SerialMotorsBus(MotorsBusBase):
 
             if display_values and not user_pressed_enter:
                 # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motors) + 3)
+                move_cursor_up(len(motor_names) + 3)
 
-        same_min_max = [motor for motor in motors if mins[motor] == maxes[motor]]
+        same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
         if same_min_max:
             raise ValueError(f"Some motors have the same min and max values:\n{pformat(same_min_max)}")
 
@@ -955,12 +952,12 @@ class SerialMotorsBus(MotorsBusBase):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
             else:
-                return
+                return None
         if self._is_error(error):
             if raise_on_error:
                 raise RuntimeError(self.packet_handler.getRxPacketError(error))
             else:
-                return
+                return None
 
         return model_number
 
@@ -1007,12 +1004,13 @@ class SerialMotorsBus(MotorsBusBase):
         err_msg = f"Failed to read '{data_name}' on {id_=} after {num_retry + 1} tries."
         value, _, _ = self._read(addr, length, id_, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
-        id_value = self._decode_sign(data_name, {id_: value})
+        decoded = self._decode_sign(data_name, {id_: value})
 
         if normalize and data_name in self.normalized_data:
-            id_value = self._normalize(id_value)
+            normalized = self._normalize(decoded)
+            return normalized[id_]
 
-        return id_value[id_]
+        return decoded[id_]
 
     def _read(
         self,
@@ -1023,7 +1021,7 @@ class SerialMotorsBus(MotorsBusBase):
         num_retry: int = 0,
         raise_on_error: bool = True,
         err_msg: str = "",
-    ) -> tuple[int, int]:
+    ) -> tuple[int, int, int]:
         if length == 1:
             read_fn = self.packet_handler.read1ByteTxRx
         elif length == 2:
@@ -1073,13 +1071,14 @@ class SerialMotorsBus(MotorsBusBase):
         model = self.motors[motor].model
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
+        int_value = int(value)
         if normalize and data_name in self.normalized_data:
-            value = self._unnormalize({id_: value})[id_]
+            int_value = self._unnormalize({id_: value})[id_]
 
-        value = self._encode_sign(data_name, {id_: value})[id_]
+        int_value = self._encode_sign(data_name, {id_: int_value})[id_]
 
-        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
-        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{int_value}' after {num_retry + 1} tries."
+        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     def _write(
         self,
@@ -1113,7 +1112,7 @@ class SerialMotorsBus(MotorsBusBase):
     def sync_read(
         self,
         data_name: str,
-        motors: str | list[str] | None = None,
+        motors: NameOrID | Sequence[NameOrID] | None = None,
         *,
         normalize: bool = True,
         num_retry: int = 0,
@@ -1122,7 +1121,7 @@ class SerialMotorsBus(MotorsBusBase):
 
         Args:
             data_name (str): Register name.
-            motors (str | list[str] | None, optional): Motors to query. `None` (default) reads every motor.
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Motors to query. `None` (default) reads every motor.
             normalize (bool, optional): Normalisation flag.  Defaults to `True`.
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
 
@@ -1143,16 +1142,17 @@ class SerialMotorsBus(MotorsBusBase):
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
         err_msg = f"Failed to sync read '{data_name}' on {ids=} after {num_retry + 1} tries."
-        ids_values, _ = self._sync_read(
+        raw_ids_values, _ = self._sync_read(
             addr, length, ids, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
         )
 
-        ids_values = self._decode_sign(data_name, ids_values)
+        decoded = self._decode_sign(data_name, raw_ids_values)
 
         if normalize and data_name in self.normalized_data:
-            ids_values = self._normalize(ids_values)
+            normalized = self._normalize(decoded)
+            return {self._id_to_name(id_): value for id_, value in normalized.items()}
 
-        return {self._id_to_name(id_): value for id_, value in ids_values.items()}
+        return {self._id_to_name(id_): value for id_, value in decoded.items()}
 
     def _sync_read(
         self,
@@ -1224,21 +1224,24 @@ class SerialMotorsBus(MotorsBusBase):
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
         """
 
-        ids_values = self._get_ids_values_dict(values)
-        models = [self._id_to_model(id_) for id_ in ids_values]
+        raw_ids_values = self._get_ids_values_dict(values)
+        models = [self._id_to_model(id_) for id_ in raw_ids_values]
         if self._has_different_ctrl_tables:
             assert_same_address(self.model_ctrl_table, models, data_name)
 
         model = next(iter(models))
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
+        int_ids_values = {id_: int(val) for id_, val in raw_ids_values.items()}
         if normalize and data_name in self.normalized_data:
-            ids_values = self._unnormalize(ids_values)
+            int_ids_values = self._unnormalize(raw_ids_values)
 
-        ids_values = self._encode_sign(data_name, ids_values)
+        int_ids_values = self._encode_sign(data_name, int_ids_values)
 
-        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
-        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to sync write '{data_name}' with ids_values={int_ids_values} after {num_retry + 1} tries."
+        self._sync_write(
+            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
+        )
 
     def _sync_write(
         self,
@@ -1274,4 +1277,4 @@ class SerialMotorsBus(MotorsBusBase):
 
 
 # Backward compatibility alias
-MotorsBus: TypeAlias = SerialMotorsBus
+MotorsBus = SerialMotorsBus

src/lerobot/motors/robstride/__init__.py

@@ -0,0 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .robstride import RobstrideMotorsBus
+from .tables import *

src/lerobot/motors/robstride/tables.py

@@ -0,0 +1,120 @@
+# 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.
+
+"""Configuration tables for Damiao motors."""
+
+from enum import IntEnum
+
+
+# Motor type definitions
+class MotorType(IntEnum):
+    O0 = 0
+    O1 = 1
+    O2 = 2
+    O3 = 3
+    O4 = 4
+    O5 = 5
+    ELO5 = 6
+    O6 = 7
+
+
+class CommMode(IntEnum):
+    PrivateProtocole = 0
+    CANopen = 1
+    MIT = 2
+
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 0
+    POS_VEL = 1
+    VEL = 2
+
+
+# Motor limit parameters [PMAX, VMAX, TMAX]
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS: dict[MotorType, tuple[float, float, float]] = {
+    MotorType.O0: (12.57, 33, 14),
+    MotorType.O1: (12.57, 44, 17),
+    MotorType.O2: (12.57, 33, 20),
+    MotorType.O3: (12.57, 33, 60),
+    MotorType.O4: (12.57, 33, 120),
+    MotorType.O5: (12.57, 50, 5.5),
+    MotorType.ELO5: (12.57, 50, 6),
+    MotorType.O6: (112.5, 50, 36),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.O0: "O0",
+    MotorType.O1: "O1",
+    MotorType.O2: "O2",
+    MotorType.O3: "O3",
+    MotorType.O4: "O4",
+    MotorType.O5: "O5",
+    MotorType.ELO5: "ELO5",
+    MotorType.O6: "O6",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "O0": 65536,
+    "O1": 65536,
+    "O2": 65536,
+    "O3": 65536,
+    "O4": 65536,
+    "O5": 65536,
+    "ELO5": 65536,
+    "O6": 65536,
+}
+
+# CAN baudrates supported by Robstride motors
+AVAILABLE_BAUDRATES = [
+    1000000,  # 4: 1 mbps (default)
+]
+DEFAULT_BAUDRATE = 1000000
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 0  # disabled by default, otherwise 20000 is 1s
+
+
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
+
+# MIT control parameter ranges
+MIT_KP_RANGE = (0.0, 500.0)
+MIT_KD_RANGE = (0.0, 5.0)
+
+# CAN frame command IDs
+CAN_CMD_ENABLE = 0xFC
+CAN_CMD_DISABLE = 0xFD
+CAN_CMD_SET_ZERO = 0xFE
+CAN_CMD_CLEAR_FAULT = 0xFB
+
+
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF
+
+
+RUNNING_TIMEOUT = 0.001
+PARAM_TIMEOUT = 0.01
+
+STATE_CACHE_TTL_S = 0.02

src/lerobot/policies/act/configuration_act.py

@@ -28,7 +28,7 @@ class ACTConfig(PreTrainedConfig):
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and 'output_shapes`.
+    Those are: `input_features` and `output_features`.
 
     Notes on the inputs and outputs:
         - Either:
@@ -48,21 +48,12 @@ class ACTConfig(PreTrainedConfig):
             This should be no greater than the chunk size. For example, if the chunk size size 100, you may
             set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
             environment, and throws the other 50 out.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -30,7 +30,7 @@ class DiffusionConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and `output_shapes`.
+    Those are: `input_features` and `output_features`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -48,32 +48,29 @@ class DiffusionConfig(PreTrainedConfig):
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
-            within the image size. If None, no cropping is done.
-        crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
-            mode).
+        resize_shape: (H, W) shape to resize images to as a preprocessing step for the vision
+            backbone. If None, no resizing is done and the original image resolution is used.
+        crop_ratio: Ratio in (0, 1] used to derive the crop size from resize_shape
+            (crop_h = int(resize_shape[0] * crop_ratio), likewise for width).
+            Set to 1.0 to disable cropping. Only takes effect when resize_shape is not None.
+        crop_shape: (H, W) shape to crop images to. When resize_shape is set and crop_ratio < 1.0,
+            this is computed automatically. Can also be set directly for legacy configs that use
+            crop-only (without resize). If None and no derivation applies, no cropping is done.
+        crop_is_random: Whether the crop should be random at training time (it's always a center
+            crop in eval mode).
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
         spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
-        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
+        use_separate_rgb_encoder_per_camera: Whether to use a separate RGB encoder for each camera view.
         down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
             You may provide a variable number of dimensions, therefore also controlling the degree of
             downsampling.
@@ -123,7 +120,9 @@ class DiffusionConfig(PreTrainedConfig):
     # Architecture / modeling.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    crop_shape: tuple[int, int] | None = (84, 84)
+    resize_shape: tuple[int, int] | None = None
+    crop_ratio: float = 1.0
+    crop_shape: tuple[int, int] | None = None
     crop_is_random: bool = True
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
@@ -148,6 +147,10 @@ class DiffusionConfig(PreTrainedConfig):
     # Inference
     num_inference_steps: int | None = None
 
+    # Optimization
+    compile_model: bool = False
+    compile_mode: str = "reduce-overhead"
+
     # Loss computation
     do_mask_loss_for_padding: bool = False
 
@@ -180,6 +183,25 @@ class DiffusionConfig(PreTrainedConfig):
                 f"Got {self.noise_scheduler_type}."
             )
 
+        if self.resize_shape is not None and (
+            len(self.resize_shape) != 2 or any(d <= 0 for d in self.resize_shape)
+        ):
+            raise ValueError(f"`resize_shape` must be a pair of positive integers. Got {self.resize_shape}.")
+        if not (0 < self.crop_ratio <= 1.0):
+            raise ValueError(f"`crop_ratio` must be in (0, 1]. Got {self.crop_ratio}.")
+
+        if self.resize_shape is not None:
+            if self.crop_ratio < 1.0:
+                self.crop_shape = (
+                    int(self.resize_shape[0] * self.crop_ratio),
+                    int(self.resize_shape[1] * self.crop_ratio),
+                )
+            else:
+                # Explicitly disable cropping for resize+ratio path when crop_ratio == 1.0.
+                self.crop_shape = None
+        if self.crop_shape is not None and (self.crop_shape[0] <= 0 or self.crop_shape[1] <= 0):
+            raise ValueError(f"`crop_shape` must have positive dimensions. Got {self.crop_shape}.")
+
         # Check that the horizon size and U-Net downsampling is compatible.
         # U-Net downsamples by 2 with each stage.
         downsampling_factor = 2 ** len(self.down_dims)
@@ -207,13 +229,12 @@ class DiffusionConfig(PreTrainedConfig):
         if len(self.image_features) == 0 and self.env_state_feature is None:
             raise ValueError("You must provide at least one image or the environment state among the inputs.")
 
-        if self.crop_shape is not None:
+        if self.resize_shape is None and self.crop_shape is not None:
             for key, image_ft in self.image_features.items():
                 if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
                     raise ValueError(
-                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for "
-                        f"`{key}`."
+                        f"`crop_shape` should fit within the image shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
                     )
 
         # Check that all input images have the same shape.

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -142,6 +142,9 @@ class DiffusionPolicy(PreTrainedPolicy):
         """Run the batch through the model and compute the loss for training or validation."""
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            for key in self.config.image_features:
+                if self.config.n_obs_steps == 1 and batch[key].ndim == 4:
+                    batch[key] = batch[key].unsqueeze(1)
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
         loss = self.diffusion.compute_loss(batch)
         # no output_dict so returning None
@@ -182,6 +185,11 @@ class DiffusionModel(nn.Module):
 
         self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
 
+        if config.compile_model:
+            # Compile the U-Net. "reduce-overhead" is preferred for the small-batch repetitive loops
+            # common in diffusion inference.
+            self.unet = torch.compile(self.unet, mode=config.compile_mode)
+
         self.noise_scheduler = _make_noise_scheduler(
             config.noise_scheduler_type,
             num_train_timesteps=config.num_train_timesteps,
@@ -446,12 +454,18 @@ class DiffusionRgbEncoder(nn.Module):
     def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if config.crop_shape is not None:
+        if config.resize_shape is not None:
+            self.resize = torchvision.transforms.Resize(config.resize_shape)
+        else:
+            self.resize = None
+
+        crop_shape = config.crop_shape
+        if crop_shape is not None:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(crop_shape)
             if config.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(crop_shape)
             else:
                 self.maybe_random_crop = self.center_crop
         else:
@@ -477,13 +491,16 @@ class DiffusionRgbEncoder(nn.Module):
 
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
-        # The dummy input should take the number of image channels from `config.image_features` and it should
-        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
-        # height and width from `config.image_features`.
+        # The dummy shape mirrors the runtime preprocessing order: resize -> crop.
 
         # Note: we have a check in the config class to make sure all images have the same shape.
         images_shape = next(iter(config.image_features.values())).shape
-        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
+        if config.crop_shape is not None:
+            dummy_shape_h_w = config.crop_shape
+        elif config.resize_shape is not None:
+            dummy_shape_h_w = config.resize_shape
+        else:
+            dummy_shape_h_w = images_shape[1:]
         dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
         feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
 
@@ -499,7 +516,10 @@ class DiffusionRgbEncoder(nn.Module):
         Returns:
             (B, D) image feature.
         """
-        # Preprocess: maybe crop (if it was set up in the __init__).
+        # Preprocess: resize if configured, then crop if configured.
+
+        if self.resize is not None:
+            x = self.resize(x)
         if self.do_crop:
             if self.training:  # noqa: SIM108
                 x = self.maybe_random_crop(x)

src/lerobot/policies/factory.py

@@ -18,10 +18,9 @@ from __future__ import annotations
 
 import importlib
 import logging
-from typing import Any, TypedDict
+from typing import Any, TypedDict, Unpack
 
 import torch
-from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType

src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py

@@ -4,17 +4,16 @@
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
 
+from __future__ import annotations
 
 # copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py
-from typing import Optional
-
 from transformers.image_processing_utils import (
     BatchFeature,
     get_patch_output_size,
 )
 from transformers.image_processing_utils_fast import (
     BaseImageProcessorFast,
-    DefaultFastImageProcessorKwargs,
+    ImagesKwargs,
     group_images_by_shape,
     reorder_images,
 )
@@ -77,7 +76,7 @@ def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> tor
     return img[:, top:bottom, left:right]
 
 
-class Eagle25VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+class Eagle25VLFastImageProcessorKwargs(ImagesKwargs):
     max_dynamic_tiles: int | None
     min_dynamic_tiles: int | None
     use_thumbnail: bool | None
@@ -165,11 +164,11 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _resize_for_patching(
         self,
-        image: "torch.Tensor",
+        image: torch.Tensor,
         target_resolution: tuple,
-        interpolation: "F.InterpolationMode",
+        interpolation: F.InterpolationMode,
         input_data_format: ChannelDimension,
-    ) -> "torch.Tensor":
+    ) -> torch.Tensor:
         """
         Resizes an image to a target resolution while maintaining aspect ratio.
 
@@ -219,8 +218,8 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
         return best_ratio
 
     def _pad_for_patching(
-        self, image: "torch.Tensor", target_resolution: tuple, input_data_format: ChannelDimension
-    ) -> "torch.Tensor":
+        self, image: torch.Tensor, target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> torch.Tensor:
         """
         Pad an image to a target resolution while maintaining aspect ratio.
         """
@@ -236,15 +235,15 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _get_image_patches(
         self,
-        image: "torch.Tensor",
+        image: torch.Tensor,
         min_num: int,
         max_num: int,
         size: tuple,
         tile_size: int,
         use_thumbnail: bool,
-        interpolation: "F.InterpolationMode",
+        interpolation: F.InterpolationMode,
         pad_during_tiling: bool,
-    ) -> list["torch.Tensor"]:
+    ) -> list[torch.Tensor]:
         image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST)
         orig_height, orig_width = image_size
         aspect_ratio = orig_width / orig_height
@@ -305,8 +304,8 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _pad_for_batching(
         self,
-        pixel_values: list["torch.Tensor"],
-    ) -> list["torch.Tensor"]:
+        pixel_values: list[torch.Tensor],
+    ) -> list[torch.Tensor]:
         """
         Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.
 
@@ -327,14 +326,14 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _preprocess(
         self,
-        images: list["torch.Tensor"],
+        images: list[torch.Tensor],
         do_resize: bool,
         size: SizeDict,
         max_dynamic_tiles: int,
         min_dynamic_tiles: int,
         use_thumbnail: bool,
         pad_during_tiling: bool,
-        interpolation: Optional["F.InterpolationMode"],
+        interpolation: F.InterpolationMode | None,
         do_center_crop: bool,
         crop_size: SizeDict,
         do_rescale: bool,

src/lerobot/policies/pi0/modeling_pi0.py

@@ -15,16 +15,16 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -32,13 +32,21 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
-    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
+
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
@@ -191,7 +199,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -202,7 +210,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -221,14 +229,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
     layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.language_model, gemma_expert.model]
+    models = [paligemma.model.language_model, gemma_expert.model]
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
         layer = models[i].layers[layer_idx]
-        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -254,10 +262,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -265,7 +273,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
@@ -277,15 +285,15 @@ def compute_layer_complete(
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
         out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
         # first residual
-        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
         # Convert to bfloat16 if the next layer (mlp) uses bfloat16
         if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
             out_emb = out_emb.to(dtype=torch.bfloat16)
         out_emb = layer.mlp(out_emb)
         # second residual
-        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -358,7 +366,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -366,7 +374,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -377,13 +385,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -398,10 +406,11 @@ class PaliGemmaWithExpertModel(
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Align with PI05.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -413,8 +422,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -424,15 +433,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32). Align with PI05.
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -446,7 +463,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -470,7 +487,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.language_model, self.gemma_expert.model]
             num_layers = self.paligemma.config.text_config.num_hidden_layers
 
             # Check if gradient checkpointing is enabled for any of the models
@@ -510,7 +527,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -576,29 +593,19 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             # Also compile the main forward pass used during training
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
         logging.info("Enabled gradient checkpointing for PI0Pytorch model")
 
     def gradient_checkpointing_disable(self):
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI0Pytorch model")
 
@@ -760,7 +767,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(state, x_t, time)
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -834,7 +841,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
 
         prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -908,6 +915,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
+        past_key_values = copy.deepcopy(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,
@@ -997,14 +1005,12 @@ class PI0Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -1012,7 +1018,7 @@ class PI0Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -1025,7 +1031,7 @@ class PI0Policy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
 
             # Then add "model." prefix for all keys that don't already have it
@@ -1070,7 +1076,7 @@ class PI0Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1120,6 +1126,14 @@ class PI0Policy(PreTrainedPolicy):
                 # Some checkpoints might have this, but current model expects different structure
                 logging.warning(f"Vision embedding key might need handling: {key}")
 
+            if (
+                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
+                or key == "paligemma_with_expert.paligemma.lm_head.weight"
+            ):
+                fixed_state_dict[
+                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
+                ] = value.clone()
+
             fixed_state_dict[new_key] = value
 
         return fixed_state_dict

src/lerobot/policies/pi05/modeling_pi05.py

@@ -15,16 +15,16 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -32,14 +32,20 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
-    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
-
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
@@ -92,10 +98,11 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
-    alpha_t = torch.as_tensor(alpha, dtype=torch.float32, device=device)
-    beta_t = torch.as_tensor(beta, dtype=torch.float32, device=device)
+    # Beta sampling uses _sample_dirichlet which isn't implemented for MPS, so sample on CPU
+    alpha_t = torch.tensor(alpha, dtype=torch.float32)
+    beta_t = torch.tensor(beta, dtype=torch.float32)
     dist = torch.distributions.Beta(alpha_t, beta_t)
-    return dist.sample((bsize,))
+    return dist.sample((bsize,)).to(device)
 
 
 def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (exact copy)
@@ -189,7 +196,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -200,7 +207,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -219,14 +226,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
     layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.language_model, gemma_expert.model]
+    models = [paligemma.model.language_model, gemma_expert.model]
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
         layer = models[i].layers[layer_idx]
-        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -252,10 +259,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -263,7 +270,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
@@ -275,15 +282,15 @@ def compute_layer_complete(
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
         out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
         # first residual
-        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
         # Convert to bfloat16 if the next layer (mlp) uses bfloat16
         if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
             out_emb = out_emb.to(dtype=torch.bfloat16)
         out_emb = layer.mlp(out_emb)
         # second residual
-        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -356,7 +363,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -364,7 +371,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -375,13 +382,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -396,10 +403,11 @@ class PaliGemmaWithExpertModel(
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Saves memory vs full float32; more memory than only 3 params.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -411,8 +419,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -422,15 +430,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32).
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -444,7 +460,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -468,7 +484,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.language_model, self.gemma_expert.model]
             num_layers = self.paligemma.config.text_config.num_hidden_layers
 
             # Check if gradient checkpointing is enabled for any of the models
@@ -508,7 +524,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -573,29 +589,19 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             # Also compile the main forward pass used during training
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
         logging.info("Enabled gradient checkpointing for PI05Pytorch model")
 
     def gradient_checkpointing_disable(self):
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI05Pytorch model")
 
@@ -737,7 +743,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(x_t, time)
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -808,7 +814,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
 
         prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -880,6 +886,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
+        past_key_values = copy.deepcopy(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,
@@ -969,14 +976,12 @@ class PI05Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -984,7 +989,7 @@ class PI05Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -997,7 +1002,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
 
             # Then add "model." prefix for all keys that don't already have it
@@ -1009,8 +1014,6 @@ class PI05Policy(PreTrainedPolicy):
                     new_key = f"model.{key}"
                     remapped_state_dict[new_key] = value
                     remap_count += 1
-                    if remap_count <= 10:  # Only print first 10 to avoid spam
-                        print(f"Remapped: {key} -> {new_key}")
                 else:
                     remapped_state_dict[key] = value
 
@@ -1044,7 +1047,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1098,6 +1101,14 @@ class PI05Policy(PreTrainedPolicy):
                 # Some checkpoints might have this, but current model expects different structure
                 logging.warning(f"Vision embedding key might need handling: {key}")
 
+            if (
+                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
+                or key == "paligemma_with_expert.paligemma.lm_head.weight"
+            ):
+                fixed_state_dict[
+                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
+                ] = value.clone()
+
             fixed_state_dict[new_key] = value
 
         return fixed_state_dict

src/lerobot/policies/pi05/processor_pi05.py

@@ -23,7 +23,6 @@ import torch
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
-from lerobot.policies.pi05.modeling_pi05 import pad_vector
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -68,9 +67,6 @@ class Pi05PrepareStateTokenizerProcessorStep(ProcessorStep):
         # TODO: check if this necessary
         state = deepcopy(state)
 
-        # Prepare state (pad to max_state_dim)
-        state = pad_vector(state, self.max_state_dim)
-
         # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
         # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
         state_np = state.cpu().numpy()

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -54,7 +54,7 @@ class PI0FastConfig(PreTrainedConfig):
 
     tokenizer_max_length: int = 200  # see openpi `__post_init__`
     text_tokenizer_name: str = "google/paligemma-3b-pt-224"
-    action_tokenizer_name: str = "physical-intelligence/fast"
+    action_tokenizer_name: str = "lerobot/fast-action-tokenizer"
     temperature: float = 0.0
     max_decoding_steps: int = 256
     fast_skip_tokens: int = 128

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -19,13 +19,12 @@ import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _scipy_available, _transformers_available
 
@@ -38,11 +37,16 @@ else:
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoTokenizer
     from transformers.models.auto import CONFIG_MAPPING
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaModel,
+    )
 else:
     CONFIG_MAPPING = None
-    PaliGemmaForConditionalGeneration = None
     AutoTokenizer = None
+    PiGemmaModel = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
@@ -121,7 +125,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -132,7 +136,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -206,16 +210,22 @@ class PI0FastPaliGemma(nn.Module):
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+
+        # Use PI Gemma (AdaRMS) as language model when use_adarms[0] is True so that
+        # forward(..., adarms_cond=...) is supported (same as pi0/pi05).
+        if use_adarms[0]:
+            text_config = self.paligemma.config.text_config
+            self.paligemma.model.language_model = PiGemmaModel(text_config)
 
         self.to_bfloat16_for_selected_params(precision)
 
@@ -228,10 +238,11 @@ class PI0FastPaliGemma(nn.Module):
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Align with PI05.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -242,10 +253,18 @@ class PI0FastPaliGemma(nn.Module):
                 param.data = param.data.to(dtype=torch.float32)
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32). Align with PI05.
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -259,7 +278,7 @@ class PI0FastPaliGemma(nn.Module):
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -306,24 +325,14 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
             self.sample_actions_fast = torch.compile(self.sample_actions_fast, mode=config.compile_mode)
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
         # Call the proper gradient_checkpointing_enable() method with use_reentrant=False for better memory efficiency
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_enable(
             gradient_checkpointing_kwargs={"use_reentrant": False}
         )
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_enable(
             gradient_checkpointing_kwargs={"use_reentrant": False}
         )
         logging.info("Enabled gradient checkpointing for PI0FastPytorch model")
@@ -332,8 +341,8 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
         # Call the proper gradient_checkpointing_disable() method
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing_disable()
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_disable()
         logging.info("Disabled gradient checkpointing for PI0FastPytorch model")
 
     def _apply_checkpoint(self, func, *args, **kwargs):
@@ -523,7 +532,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Convert embeddings to bfloat16 if needed
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -616,7 +625,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -714,7 +723,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Ensure correct precision (bfloat16/float32)
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -897,14 +906,12 @@ class PI0FastPolicy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -912,7 +919,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -925,8 +932,9 @@ class PI0FastPolicy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
+
             # Then add "model." prefix for all keys that don't already have it
             remapped_state_dict = {}
             remap_count = 0
@@ -936,8 +944,6 @@ class PI0FastPolicy(PreTrainedPolicy):
                     new_key = f"model.{key}"
                     remapped_state_dict[new_key] = value
                     remap_count += 1
-                    if remap_count <= 10:  # Only print first 10 to avoid spam
-                        print(f"Remapped: {key} -> {new_key}")
                 else:
                     remapped_state_dict[key] = value
 
@@ -971,7 +977,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -23,7 +23,6 @@ import torch
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
-from lerobot.policies.pi0_fast.modeling_pi0_fast import pad_vector
 from lerobot.processor import (
     ActionTokenizerProcessorStep,
     AddBatchDimensionProcessorStep,
@@ -69,9 +68,6 @@ class Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(ProcessorStep):
         # TODO: check if this necessary
         state = deepcopy(state)
 
-        # Prepare state (pad to max_state_dim)
-        state = pad_vector(state, self.max_state_dim)
-
         # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
         # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
         state_np = state.cpu().numpy()

src/lerobot/policies/pi_gemma.py

@@ -0,0 +1,363 @@
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+from torch import nn
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers.cache_utils import DynamicCache
+    from transformers.masking_utils import create_causal_mask
+    from transformers.modeling_layers import GradientCheckpointingLayer
+    from transformers.modeling_outputs import BaseModelOutputWithPast
+    from transformers.models.gemma.modeling_gemma import (
+        GemmaAttention,
+        GemmaConfig,
+        GemmaForCausalLM,
+        GemmaMLP,
+        GemmaModel,
+    )
+    from transformers.models.paligemma.modeling_paligemma import (
+        PaliGemmaForConditionalGeneration,
+        PaliGemmaModel,
+    )
+else:
+    GemmaAttention = None
+    GemmaConfig = None
+    GemmaForCausalLM = None
+    GemmaMLP = None
+    GemmaModel = None
+    PaliGemmaModel = None
+    PaliGemmaForConditionalGeneration = None
+    DynamicCache = None
+    GradientCheckpointingLayer = None
+    BaseModelOutputWithPast = None
+    create_causal_mask = None
+
+
+def _gated_residual(
+    x: torch.Tensor | None,
+    y: torch.Tensor | None,
+    gate: torch.Tensor | None,
+) -> torch.Tensor | None:
+    """Gated residual: x + y when gate is None, else x + y * gate."""
+    if x is None and y is None:
+        return None
+    if x is None or y is None:
+        return x if x is not None else y
+    if gate is None:
+        return x + y
+    return x + y * gate
+
+
+def layernorm_forward(
+    layernorm: nn.Module,
+    x: torch.Tensor,
+    cond: torch.Tensor | None = None,
+):
+    """
+    call layernorm and return hidden states and gate
+    if cond is not None, use conditional norm
+    otherwise, use normal gemma norm
+    """
+    if cond is not None:
+        return layernorm(x, cond=cond)
+    else:
+        return layernorm(x)
+
+
+class PiGemmaRMSNorm(nn.Module):
+    """
+    Adaptive RMSNorm for PI Gemma (AdaRMS).
+    When cond_dim is set, uses cond to modulate scale/shift/gate; otherwise behaves like standard GemmaRMSNorm.
+    forward(x, cond=None) returns (output, gate) for use with _gated_residual.
+    """
+
+    def __init__(self, dim: int, eps: float = 1e-6, cond_dim: int | None = None):
+        super().__init__()
+        self.eps = eps
+        self.dim = dim
+        self.cond_dim = cond_dim
+        if cond_dim is not None:
+            self.dense = nn.Linear(cond_dim, dim * 3, bias=True)
+            nn.init.zeros_(self.dense.weight)
+        else:
+            self.weight = nn.Parameter(torch.zeros(dim))
+            self.dense = None
+
+    def _norm(self, x):
+        # Compute variance in float32 (like the source implementation)
+        var = torch.mean(torch.square(x.float()), dim=-1, keepdim=True)
+        # Compute normalization in float32
+        normed_inputs = x * torch.rsqrt(var + self.eps)
+        return normed_inputs
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cond: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        dtype = x.dtype
+        normed = self._norm(x)
+        if cond is None or self.dense is None:
+            normed = normed * (1.0 + self.weight.float())
+            return normed.type_as(x), None
+        if cond.shape[-1] != self.cond_dim:
+            raise ValueError(f"Expected cond dim {self.cond_dim}, got {cond.shape[-1]}")
+        modulation = self.dense(cond)
+        if len(x.shape) == 3:
+            modulation = modulation.unsqueeze(1)
+        scale, shift, gate = modulation.chunk(3, dim=-1)
+        normed = normed * (1 + scale.float()) + shift.float()
+        return normed.to(dtype), gate.to(dtype)
+
+    def extra_repr(self) -> str:
+        if self.dense is not None:
+            return f"dim={self.dim}, eps={self.eps}, adaptive=True, cond_dim={self.cond_dim}"
+        return f"dim={self.dim}, eps={self.eps}"
+
+
+def _get_pi_gemma_decoder_layer_base():
+    """base for PiGemmaDecoderLayer"""
+
+    class _PiGemmaDecoderLayerBase(GradientCheckpointingLayer):
+        """Decoder layer that uses PiGemmaRMSNorm and _gated_residual, compatible with v5 Gemma."""
+
+        def __init__(self, config: GemmaConfig, layer_idx: int):
+            super().__init__()
+            self.hidden_size = config.hidden_size
+            self.self_attn = GemmaAttention(config=config, layer_idx=layer_idx)
+            self.mlp = GemmaMLP(config)
+            cond_dim = (
+                getattr(config, "adarms_cond_dim", None) if getattr(config, "use_adarms", False) else None
+            )
+            self.input_layernorm = PiGemmaRMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim
+            )
+            self.post_attention_layernorm = PiGemmaRMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim
+            )
+
+        def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: torch.Tensor | None = None,
+            position_ids: torch.LongTensor | None = None,
+            past_key_values=None,
+            use_cache: bool = False,
+            cache_position: torch.LongTensor | None = None,
+            position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+            adarms_cond: torch.Tensor | None = None,
+            **kwargs,
+        ) -> torch.Tensor:
+            residual = hidden_states
+            hidden_states, gate = self.input_layernorm(hidden_states, cond=adarms_cond)
+            hidden_states, _ = self.self_attn(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+            hidden_states = _gated_residual(residual, hidden_states, gate)
+
+            residual = hidden_states
+            hidden_states, gate = self.post_attention_layernorm(hidden_states, cond=adarms_cond)
+            hidden_states = self.mlp(hidden_states)
+            hidden_states = _gated_residual(residual, hidden_states, gate)
+            return hidden_states
+
+    return _PiGemmaDecoderLayerBase
+
+
+class PiGemmaModel(GemmaModel):  # type: ignore[misc]
+    """
+    GemmaModel extended with AdaRMS (adaptive RMSNorm) and gated residuals when config.use_adarms is True.
+    """
+
+    def __init__(self, config: GemmaConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        # if not getattr(config, "use_adarms", False):
+        #     return
+        cond_dim = getattr(config, "adarms_cond_dim", None)
+        pi_gemma_decoder_layer_base = _get_pi_gemma_decoder_layer_base()
+        self.layers = nn.ModuleList(
+            [pi_gemma_decoder_layer_base(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = PiGemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim)
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: DynamicCache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        adarms_cond: torch.Tensor | None = None,
+        **kwargs,
+    ) -> BaseModelOutputWithPast:
+        """
+        adarms_cond (`torch.Tensor` of shape `(batch_size, cond_dim)`, *optional*):
+            Condition for ADARMS.
+        """
+        output_attentions = (
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            import logging
+
+            logging.warning(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        # embed positions
+        hidden_states = inputs_embeds
+        # Convert to bfloat16 if the first layer uses bfloat16
+        if len(self.layers) > 0 and self.layers[0].self_attn.q_proj.weight.dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.bfloat16)
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # normalized
+        # Gemma downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
+        # See https://github.com/huggingface/transformers/pull/29402
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                adarms_cond=adarms_cond,
+                **kwargs,
+            )
+
+            hidden_states = layer_outputs
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states, _ = self.norm(hidden_states, adarms_cond)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class PiGemmaForCausalLM(GemmaForCausalLM):  # type: ignore[misc]
+    """
+    Causal LM wrapper using PiGemmaModel as the backbone, for consistency with GemmaForCausalLM
+    and the language model used in pi0_fast. Use this for the action expert in pi0/pi05.
+    """
+
+    def __init__(self, config: GemmaConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.model = PiGemmaModel(config)
+
+
+class PaliGemmaModelWithPiGemma(PaliGemmaModel):
+    """PaliGemmaModel whose language_model is PiGemmaModel (custom decoder with PiGemmaRMSNorm and gated residuals)."""
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.language_model = PiGemmaModel(config.text_config)
+
+
+class PaliGemmaForConditionalGenerationWithPiGemma(PaliGemmaForConditionalGeneration):
+    """PaliGemmaForConditionalGeneration using PiGemma decoder for the language model."""
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = PaliGemmaModelWithPiGemma(config)
+
+    # Make modules available through conditional class for BC
+    @property
+    def language_model(self):
+        return self.model.language_model
+
+
+__all__ = [
+    "PiGemmaModel",
+    "PiGemmaForCausalLM",
+    "PiGemmaRMSNorm",
+    "_gated_residual",
+    "layernorm_forward",
+    "PaliGemmaModelWithPiGemma",
+    "PaliGemmaForConditionalGenerationWithPiGemma",
+]

src/lerobot/policies/pretrained.py

@@ -19,7 +19,7 @@ import os
 from importlib.resources import files
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import TypedDict, TypeVar
+from typing import TypedDict, TypeVar, Unpack
 
 import packaging
 import safetensors
@@ -28,7 +28,6 @@ from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
 from safetensors.torch import load_model as load_model_as_safetensor, save_model as save_model_as_safetensor
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig

src/lerobot/policies/sac/reward_model/configuration_classifier.py

@@ -33,7 +33,7 @@ class RewardClassifierConfig(PreTrainedConfig):
     latent_dim: int = 256
     image_embedding_pooling_dim: int = 8
     dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"
+    model_name: str = "helper2424/resnet10"  # TODO: This needs to be updated. The model on the Hub doesn't call self.post_init() in its __init__, which is required by transformers v5 to set all_tied_weights_keys. The from_pretrained call fails when it tries to access this attribute during _finalize_model_loading.
     device: str = "cpu"
     model_type: str = "cnn"  # "transformer" or "cnn"
     num_cameras: int = 2

src/lerobot/policies/sarm/compute_rabc_weights.py

@@ -27,18 +27,18 @@ Usage:
     # Full RA-BC computation with visualizations
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4
+        --reward-model-path <USER>/sarm_single_uni4
 
     # Faster computation with stride (compute every 5 frames, interpolate the rest)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --reward-model-path <USER>/sarm_single_uni4 \\
         --stride 5
 
     # Visualize predictions only (no RA-BC computation)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --reward-model-path <USER>/sarm_single_uni4 \\
         --visualize-only \\
         --num-visualizations 5
 
@@ -714,12 +714,12 @@ Examples:
     # Full RA-BC computation with visualizations
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4
+        --reward-model-path <USER>/sarm_single_uni4
 
     # Visualize predictions only (no RA-BC computation)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path pepijn223/sarm_single_uni4 \\
+        --reward-model-path <USER>/sarm_single_uni4 \\
         --visualize-only \\
         --num-visualizations 10
         """,

src/lerobot/policies/sarm/processor_sarm.py

@@ -277,9 +277,7 @@ class SARMEncodingProcessorStep(ProcessorStep):
 
         # When language is perturbed, targets are zero so perturbed samples don't contribute to progress loss
         if self.dataset_meta is not None:
-            episodes_df = None
-            if self.sparse_subtask_names != ["task"]:
-                episodes_df = self.dataset_meta.episodes.to_pandas()
+            episodes_df = self.dataset_meta.episodes.to_pandas()
 
             # Generate sparse targets
             if self.sparse_temporal_proportions is not None: