Merge branch 'feat/mirror' into openarms_wallx_rebased_3

* add visualize_dataset call from `lerobot_dataset_viz` in web record server

* add delete button

* fixes

* remove viz

* unused import

.github/workflows/fast_tests.yml

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

.github/workflows/full_tests.yml

@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -60,7 +60,6 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -88,12 +87,6 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          uv run hf auth whoami
-
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
 
@@ -169,7 +162,6 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -181,13 +173,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
-      - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/nightly.yml

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

.github/workflows/quality.yml

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

.github/workflows/release.yml

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

.github/workflows/unbound_deps_tests.yml

@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.12"
+  PYTHON_VERSION: "3.10"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -48,7 +48,6 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -80,11 +79,7 @@ jobs:
 
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          uv run hf auth whoami
+
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv
 
@@ -142,7 +137,6 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -154,11 +148,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Login to Hugging Face
-        if: env.HF_USER_TOKEN != ''
-        run: |
-          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
-          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv
       - name: Run end-to-end tests

.pre-commit-config.yaml

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

AI_POLICY.md

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

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
 
 ## Ways to Contribute
 

MANIFEST.in

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

README.md

@@ -100,11 +100,11 @@ lerobot-train \
   --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-| Category                   | Models                                                                                                                                                                                                                  |
-| -------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md) |
-| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                             |
-| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx)            |
+| Category                   | Models                                                                                                                                                                                                       |
+| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md)                                                             |
+| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                  |
+| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx) |
 
 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub
 

STREAMING_ENCODING_PR.md

@@ -0,0 +1,140 @@
+# Streaming Video Encoding — Encode on the fly during recording
+
+## Problem
+
+After each episode, `save_episode()` blocks for **~79 seconds** on a 3-camera setup (3197 frames, 107s episode):
+
+| Step | Time |
+|------|------|
+| Write 9591 PNGs to disk | ~19s |
+| Read PNGs back → compute image stats | ~15s |
+| Read PNGs again → encode 3× AV1 videos → delete PNGs | ~44.5s |
+| Save parquet + metadata | ~0.6s |
+| **Total** | **~79s** |
+
+The entire pipeline writes frames as temporary PNGs, reads them back twice (stats + encoding), then deletes them. This round-trip is the bottleneck.
+
+## Architecture
+
+### Before: sequential post-episode pipeline
+
+```
+  Recording loop                          save_episode() — BLOCKS ~79s
+ ┌─────────────┐          ┌──────────────────────────────────────────────────────────┐
+ │  30fps loop  │          │                                                          │
+ │              │  frames  │  frame_buffer ──► write PNGs ──► read PNGs ──► stats     │
+ │  camera ─►───┼──► list  │       (~19s)         │              (~15s)               │
+ │  teleop      │          │                      ▼                                   │
+ │  policy      │          │               read PNGs ──► AV1 encode ──► delete PNGs   │
+ │              │          │                        (~44.5s)                           │
+ └──────┬───────┘          └──────────────────────────────────────────────────────────┘
+        │                                         │
+        ▼                                         ▼
+   episode ends                             next episode
+   (~107s recording)                        (~79s blocked)
+```
+
+**Data path:** `frame → list → PNG disk → read → stats` + `PNG disk → read → encode → MP4 → delete PNGs`
+
+### After: streaming pipeline (encodes during recording)
+
+```
+  Recording loop (encoding happens HERE)            save_episode() — ~0.5s
+ ┌───────────────────────────────────────┐         ┌──────────────────┐
+ │  30fps control loop                   │         │                  │
+ │                                       │         │  flush encoders  │
+ │  camera ──► frame ─┬─► queue ──► [T1] ├── AV1 ─┤  (already done)  │
+ │                    │    queue ──► [T2] ├── AV1 ─┤  ~0.16s          │
+ │                    │    queue ──► [T3] ├── AV1 ─┤                  │
+ │                    │                  │         │  running stats   │
+ │                    └─► downsample ──► │─ stats ─┤  → finalize      │
+ │                       RunningQuantile │         │  ~0.01s          │
+ │  teleop / policy (never blocked)      │         │                  │
+ └───────────────────────────────────────┘         │  save parquet    │
+                                                   │  ~0.36s          │
+        [T1] [T2] [T3] = encoder threads           └──────────────────┘
+        (one per camera, GIL released by PyAV)
+```
+
+**Data path:** `frame → queue → encode → MP4` (zero PNGs, zero re-reads)
+
+## Stats computation changes
+
+| | Before | After |
+|---|---|---|
+| **Method** | `compute_episode_stats()` reads all PNGs from disk, decodes them, computes min/max/mean/std/quantiles | `RunningQuantileStats` accumulates stats incrementally per frame during recording |
+| **Input** | Full-resolution PNGs read back from disk | Downsampled frames (via `auto_downsample_height_width`, ~150×100px) directly from memory |
+| **When** | After episode ends, inside `save_episode()` | During recording, inside `add_frame()` (~2ms per frame) |
+| **Output** | `{mean, std, min, max, q01..q99}` shaped `(C,1,1)` in `[0,1]` | Identical shape and scale — `RunningQuantileStats.get_statistics()` → reshape `(C,1,1)` / 255 |
+| **I/O** | Reads 9591 PNGs (~15s) | Zero disk I/O |
+| **Numeric features** | Computed from episode buffer (unchanged) | Computed from episode buffer (unchanged) |
+
+The running stats use the same `auto_downsample_height_width` function and produce the same statistical keys (`mean`, `std`, `min`, `max`, `count`, `q01`, `q10`, `q50`, `q90`, `q99`). Video features are excluded from the post-episode `compute_episode_stats()` call when streaming is active — only numeric features go through that path.
+
+## Results
+
+Tested on the same 3-camera setup (2028 frames, 67.6s episode):
+
+| Step | Before | After | Speedup |
+|------|--------|-------|---------|
+| Frame writing (PNGs) | ~19s | **0s** | ∞ (eliminated) |
+| Episode stats | ~15s | **0.01s** | 1500× |
+| Video encoding | ~44.5s | **0.16s** | 278× |
+| Parquet + meta | ~0.6s | **0.36s** | ~same |
+| **Total `save_episode()`** | **~79s** | **0.55s** | **143×** |
+
+The video encoding time drops to near-zero because most encoding already happened during recording. `finish_episode()` only flushes the last few buffered frames.
+
+### Per-frame overhead during recording
+
+| Operation | Time |
+|-----------|------|
+| `queue.put(frame)` (non-blocking) | ~0.01ms |
+| `auto_downsample_height_width` | ~0.5ms |
+| `RunningQuantileStats.update` | ~1ms |
+| **Total per frame** | **~2ms** (well within 33ms budget at 30fps) |
+
+## Usage
+
+Streaming is **on by default**. Users on weaker PCs can disable it to fall back to the old post-episode pipeline:
+
+```bash
+# Default (streaming ON)
+lerobot-record --dataset.repo_id=user/dataset ...
+
+# Old behavior (streaming OFF)
+lerobot-record --dataset.repo_id=user/dataset --dataset.streaming_encoding=false
+```
+
+For the RaC data collection script, set `streaming_encoding: false` in the dataset config.
+
+## Files Changed
+
+### `src/lerobot/datasets/video_utils.py`
+- Added `StreamingVideoEncoder` — manages one `_CameraEncoder` thread per camera
+- Added `_CameraEncoder` — daemon thread that reads frames from a queue and encodes with PyAV
+- Non-blocking unbounded queue ensures the control loop is never delayed
+
+### `src/lerobot/datasets/lerobot_dataset.py`
+- `create()` / `start_streaming_encoder()`: new `streaming_encoding` parameter
+- `add_frame()`: when streaming, feeds frames to encoder + accumulates running stats instead of writing PNGs
+- `save_episode()`: when streaming, uses running stats and calls `finish_episode()` to get already-encoded video paths
+- `clear_episode_buffer()`: cancels in-progress encoding on re-record
+- `finalize()`: cleans up encoder on shutdown
+- **Full backward compatibility**: when `streaming_encoding=False`, all existing code paths are unchanged
+
+### `src/lerobot/scripts/lerobot_record.py`
+- Added `streaming_encoding: bool = True` to `DatasetRecordConfig`
+- Wired through to both `create()` and `resume` paths
+
+### `examples/rac/rac_data_collection_openarms_rtc.py`
+- Added `streaming_encoding: bool = True` to `RaCRTCDatasetConfig`
+- Frames are added inline during the control loop (streaming) or buffered for post-loop writing (old path)
+- Automatically detects mode and adjusts behavior
+
+## Design Notes
+
+- **Why threads, not processes?** PyAV/FFmpeg releases the GIL during encoding. Threads share memory (zero-copy frame passing), avoiding the serialization overhead of multiprocessing.
+- **Why unbounded queue?** At 30fps production vs ~72fps encoding throughput, the queue stays near-empty. Even during brief encoder stalls, memory growth is bounded by episode length. The control loop must never block.
+- **Why running stats?** Avoids the expensive read-back-from-disk step. `RunningQuantileStats` + `auto_downsample_height_width` compute identical statistics incrementally with ~2ms overhead per frame.
+- **Backward compatible**: Setting `streaming_encoding=false` restores the original PNG → encode pipeline exactly. No behavior changes for existing users who don't opt in.

benchmarks/video/README.md

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

docker/Dockerfile.internal

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

docker/Dockerfile.user

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

docs/source/_toctree.yml

@@ -29,8 +29,6 @@
     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
@@ -47,8 +45,6 @@
     title: NVIDIA GR00T N1.5
   - local: xvla
     title: X-VLA
-  - local: multi_task_dit
-    title: Multitask DiT Policy
   - local: walloss
     title: WALL-OSS
   title: "Policies"
@@ -61,6 +57,8 @@
     title: Use Async Inference
   - local: rtc
     title: Real-Time Chunking (RTC)
+  - local: training_time_rtc
+    title: Training-Time RTC
   title: "Inference"
 - sections:
   - local: envhub

docs/source/act.mdx

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

docs/source/async.mdx

@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
     --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
     --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
     --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
     --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
     --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
     --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions

docs/source/bring_your_own_policies.mdx

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

docs/source/earthrover_mini_plus.mdx

@@ -13,7 +13,7 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 ### Hardware
 
 - EarthRover Mini robot
-- Computer with Python 3.12 or newer
+- Computer with Python 3.10 or newer
 - Internet connection
 
 ### Setting Up the Frodobots SDK
@@ -170,13 +170,13 @@ Once you can drive the robot well, you can start recording data to train AI mode
 We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,16 +185,13 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-lerobot-record \
+python src/lerobot/scripts/lerobot_record.py \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/envhub.mdx

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

docs/source/groot.mdx

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

docs/source/hope_jr.mdx

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

docs/source/il_robots.mdx

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

docs/source/installation.mdx

@@ -1,6 +1,6 @@
 # Installation
 
-This guide uses conda (via miniforge) to manage environments. 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 [Install LeRobot](#step-3-install-lerobot-).
+This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
 
 ## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
 
@@ -11,10 +11,10 @@ bash Miniforge3-$(uname)-$(uname -m).sh
 
 ## Step 2: Environment Setup
 
-Create a virtual environment with Python 3.12, using conda:
+Create a virtual environment with Python 3.10, using conda:
 
 ```bash
-conda create -y -n lerobot python=3.12
+conda create -y -n lerobot python=3.10
 ```
 
 Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
@@ -40,13 +40,6 @@ 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`.
 
-> [!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
-> ```
-
 ## Step 3: Install LeRobot 🤗
 
 ### From Source
@@ -90,6 +83,9 @@ _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`.

docs/source/lekiwi.mdx

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

docs/source/lerobot-dataset-v3.mdx

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

docs/source/multi_task_dit.mdx

@@ -1,340 +0,0 @@
-# Multitask DiT Policy
-
-Multitask Diffusion Transformer (DiT) Policy is an evolution of the original Diffusion Policy architecture, which leverages a large DiT with text and vision conditioning for multitask robot learning. This implementation supports both diffusion and flow matching objectives for action generation, enabling robots to perform diverse manipulation tasks conditioned on language instructions.
-
-## Model Overview
-
-The model uses:
-
-- **CLIP Vision Encoder**: Processes RGB images from multiple camera views
-- **CLIP Text Encoder**: Encodes language task instructions (frozen weights with learnable projection)
-- **Diffusion Transformer**: Predicts action sequences conditioned on observations and language
-- **Two Objectives**: Supports both diffusion (DDPM/DDIM) and flow matching for action generation
-
-This model is exciting because you can achieve extremely high dexterity, competitive with multi-billion parameter
-VLAs, with only ~450M parameters and significantly less training.
-
-## Installation Requirements
-
-Multitask DiT Policy has additional dependencies. Install it with:
-
-```bash
-pip install lerobot[multi_task_dit]
-```
-
-This will install all necessary dependencies including the HuggingFace Transformers library for CLIP models.
-
-## Usage
-
-To use Multitask DiT in your LeRobot configuration, specify the policy type as:
-
-```python
-policy.type=multi_task_dit
-```
-
-## Training
-
-### Basic Training Command
-
-Here's a complete training command for training Multitask DiT on your dataset:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=YOUR_DATASET \
-  --output_dir=./outputs/multitask_dit_training \
-  --batch_size=32 \
-  --steps=5000 \
-  --save_freq=500 \
-  --log_freq=100 \
-  --policy.type=multi_task_dit \
-  --policy.device=cuda \
-  --policy.repo_id="HF_USER/multitask-dit-your-robot" \
-  --wandb.enable=true
-```
-
-### Recommended Hyperparameters and Dataset Details (30Hz Control Frequency)
-
-For reliable performance, start with these suggested default hyperparameters:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=YOUR_DATASET \
-  --output_dir=./outputs/mutitask_dit_training \
-  --batch_size=320 \
-  --steps=30000 \
-  --policy.type=multi_task_dit \
-  --policy.device=cuda \
-  --policy.horizon=32 \
-  --policy.n_action_steps=24 \
-  --policy.objective=diffusion \
-  --policy.noise_scheduler_type=DDPM \
-  --policy.num_train_timesteps=100 \
-  --policy.repo_id="HF_USER/multitask-dit-your-robot" \
-  --wandb.enable=true
-```
-
-**Key Parameters:**
-
-- **Batch Size**: 192-320 - If you have access to a GPU that can support this, you will get the best training dynamics
-- **Horizon**: 32 - number of action steps to predict, ~1.0 sec at 30Hz
-- **n_action_steps**: 24 - ~0.8 seconds at 30Hz
-- **Objective**: `diffusion` - start with diffusion and experiment with flow matching if generation quality is poor
-- **Training Steps**: >30k steps recommended for a single task
-
-### Training Configuration Parameters
-
-#### Objective Selection
-
-Choose between diffusion and flow matching:
-
-```bash
-# Diffusion objective (default)
---policy.objective=diffusion \
---policy.noise_scheduler_type=DDPM \  # or "DDIM"
---policy.num_train_timesteps=100 \
---policy.num_inference_steps=10 \  # For faster inference
---policy.beta_schedule=squaredcos_cap_v2 \  # Noise schedule type
---policy.prediction_type=epsilon \  # "epsilon" (predict noise) or "sample" (predict clean)
---policy.clip_sample=true \  # Clip samples during denoising
---policy.clip_sample_range=1.0  # Clipping range [-x, x]
-
-# Flow matching objective
---policy.objective=flow_matching \
---policy.timestep_sampling_strategy=beta \  # or "uniform" | the beta sampling strategy performance appears much better in practice
---policy.num_integration_steps=100 \
---policy.integration_method=euler \  # or "rk4"
---policy.sigma_min=0.0  # Minimum noise in flow interpolation path
-```
-
-#### Transformer Architecture
-
-Adjust model capacity based on dataset size:
-
-```bash
-# Small datasets (< 100 examples)
---policy.num_layers=4 \
---policy.hidden_dim=512 \
---policy.num_heads=8  # should ideally be hidden_dim // 64
-
-# Medium datasets (100-5k examples) - default
---policy.num_layers=6 \
---policy.hidden_dim=512 \
---policy.num_heads=8  # should ideally be hidden_dim // 64
-
-# Large datasets (> 5k examples)
---policy.num_layers=8 \
---policy.hidden_dim=512 \
---policy.num_heads=8   # should ideally be hidden_dim // 64
-```
-
-**Positional Encoding Options:**
-
-The model supports two positional encoding methods for action sequences:
-
-```bash
-# Rotary Position Embedding (RoPE) - default, recommended
---policy.use_rope=true \
---policy.rope_base=10000.0  # Base frequency for RoPE
-
-# Absolute positional encoding
---policy.use_positional_encoding=true  # Disables RoPE when true
-```
-
-**Other Transformer Parameters:**
-
-```bash
---policy.dropout=0.1  # Dropout rate for DiT blocks (0.0-1.0)
---policy.timestep_embed_dim=256  # Timestep embedding dimension
-```
-
-#### Vision Encoder Configuration
-
-```bash
-# Use different CLIP model for more expressivity at the cost of inference time
-# experiment with larger or smaller models depending on the complexity of your tasks and size of dataset
---policy.vision_encoder_name=openai/clip-vit-large-patch14
-
-# Use separate vision encoder per camera
-# This may be useful when cameras have significantly different characteristics, but
-# be wary of increased VRAM footprint.
---policy.use_separate_rgb_encoder_per_camera=true
-
-# Image preprocessing
---policy.image_resize_shape=[XXX,YYY] \ # you may need to resize your images for inference speed ups
---policy.image_crop_shape=[224,224] \
---policy.image_crop_is_random=true  # Random during training, center at inference
-```
-
-#### Text Encoder Configuration
-
-```bash
-# Use different CLIP text encoder model
-# same as vision: experiment with larger or smaller models depending on the
-# complexity of your tasks and size of dataset
---policy.text_encoder_name=openai/clip-vit-large-patch14
-```
-
-#### Learning Rate Configuration
-
-The vision encoder uses a separate learning rate multiplier, where 1/10th is suggested to be the ideal staritng point:
-
-```bash
---policy.optimizer_lr=2e-5 \
---policy.vision_encoder_lr_multiplier=0.1  # Vision encoder LR = 0.1 * optimizer_lr
-```
-
-### Training Tuning Guidelines
-
-#### 1. Flow Matching with Beta Sampling
-
-The original diffusion implementation here is based on the work described in [TRI's LBM paper](https://arxiv.org/abs/2507.05331)
-
-Additionally, we have implemented a flow-matching objective, which is described at a high-level in [Boston Dynamics blog post](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/).
-
-Consider testing the flow-matching objective and evaluating performance differences for your task:
-
-```bash
---policy.objective=flow_matching \
---policy.timestep_sampling_strategy=beta \
---policy.timestep_sampling_alpha=1.5 \
---policy.timestep_sampling_beta=1.0 \
---policy.timestep_sampling_s=0.999
-```
-
-This hasn't been shown to be a silver bullet across every user case, but it occasionally results in smoother and more consistent actions.
-
-#### 2. Number of Transformer Layers
-
-Match model capacity to your dataset size:
-
-- **Small datasets** (< 100 examples): Reduce to 4 layers
-- **Large datasets** (> 5k examples): Increase to 8 layers
-
-#### 3. `horizon` Tuning
-
-The model can be sensitive to the horizon you choose. Start with around a 1 second horizon based on your control frequency:
-
-- **30 Hz frequency**: `horizon=30`
-- **10 Hz frequency**: `horizon=10`
-
-Then experiment with increasing from there. The horizon determines how far into the future the model predicts actions.
-
-#### 4. `n_action_steps` Sensitivity
-
-The model can also be very sensitive to `n_action_steps`. Start with it being around 0.8 seconds based on your control frequency and tune from there:
-
-- **Lower values**: More reactive but potentially less stable for long-horizon tasks
-- **Higher values**: Better for long-horizon execution but open-loop failures are limited in their recovery
-
-### Inference Tuning
-
-For faster inference, use DDIM with fewer sampling steps:
-
-```bash
---policy.noise_scheduler_type=DDIM \
---policy.num_inference_steps=10
-```
-
-### Resuming Training
-
-To resume training from a checkpoint:
-
-```bash
-lerobot-train \
-  --config_path=./outputs/mutitask_dit_training/checkpoints/last/pretrained_model/train_config.json \
-  --resume=true
-```
-
-The checkpoint directory should contain `model.safetensors` and `config.json` files (saved automatically during training). When resuming, the configuration is loaded from the checkpoint, so you don't need to specify other parameters.
-
-## Common Failure Modes and Debugging
-
-Training these models can be finicky. Here are common failure modes and debugging approaches:
-
-### Idling / No Motion
-
-The model may "collapse" during inference, resulting in static or no motion. This can occur when:
-
-1. **Insufficient training data**: If you only have 20-50 examples, try to roughly double your dataset size. Once you have above 300 examples, if you're still seeing this, the task may be too complex.
-
-2. **Multiple similar tasks**: When your dataset contains multiple similar tasks (e.g., picking up 2 different objects), the model may rely too heavily on language conditioning which might not be rich enough.
-
-**Debugging tips:**
-
-- Increase dataset size (double until you get to over 300 examples)
-- Train for longer, up to 100k steps, even when the loss flatlines
-- Check if the model is receiving proper language instructions or increase diversity of instruction
-
-### Executing the Wrong Task
-
-Sometimes the robot will completely ignore your instruction and perform some other task. This generally only happens if you have trained on multiple tasks.
-
-**Potential causes:**
-
-- Language instruction ambiguity
-- Insufficient task-specific training data
-- Model confusion between similar tasks in the multitask dataset
-
-**Debugging tips:**
-
-- Verify language instruction specificity, especially if descriptions are similar between multiple tasks
-- Check task distribution in your training dataset and add weighting to the failing/ignored task
-- Consider task-specific fine-tuning
-
-### Training Instability
-
-If training loss is unstable or diverging:
-
-- Try adjusting learning rate between `1e-5` and `3e-4`
-- Increase batch size if possible
-- Check that your dataset normalization is correct
-- Verify image preprocessing is working correctly
-
-## Performance Considerations
-
-### GPU Requirements
-
-- **Inference**: At least an RTX 5070 Ti (or equivalent GPU) is recommended for reasonable speed performance
-- **Training**: A GPU with enough VRAM to load batch sizes of >64 is ideal, which will vary depending on the number of image observations, etc
-
-### Batch Size Recommendations
-
-- **Minimum**: 64 (less than this may result in unstable training)
-- **Recommended**: 256-320 (best performance, requires larger GPU)
-
-## Example: Training on Custom Dataset
-
-Here's a complete example training on a custom dataset:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=YOUR_DATASET \
-  --output_dir=./outputs/mutitask_dit_training \
-  --batch_size=320 \
-  --steps=30000 \
-  --save_freq=1000 \
-  --log_freq=100 \
-  --eval_freq=1000 \
-  --policy.type=multi_task_dit \
-  --policy.device=cuda \
-  --policy.horizon=32 \
-  --policy.n_action_steps=24 \
-  --policy.objective=diffusion \
-  --policy.noise_scheduler_type=DDPM \
-  --policy.num_layers=6 \
-  --policy.hidden_dim=512 \
-  --policy.vision_encoder_name=openai/clip-vit-base-patch16 \
-  --policy.image_resize_shape=[320,240] \
-  --policy.image_crop_shape=[224,224] \
-  --policy.repo_id="HF_USER/multitask-dit-your-robot" \
-  --wandb.enable=true \
-  --wandb.project=multitask_dit
-```
-
-## References
-
-For more details on the technical implementation and architecture, see:
-
-- [A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation](https://arxiv.org/abs/2507.05331)
-- [Large Behavior Models and Atlas Find New Footing](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/)
-- [Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy](https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy)

docs/source/openarms.mdx

@@ -0,0 +1,328 @@
+# OpenArms Robot
+
+OpenArms is a 7 DOF robotic arm with a gripper, designed by [Enactic, Inc.](https://www.enactic.com/) It uses Damiao motors controlled via CAN bus communication and MIT control mode for smooth, precise motion.
+
+## Hardware Overview
+
+- **7 DOF per arm** (14 DOF total for dual arm setup)
+- **1 gripper per arm** (2 grippers total)
+- **Damiao motors** with 4 different types:
+  - **DM8009** (DM-J8009P-2EC) for shoulders (J1, J2) - high torque
+  - **DM4340** for shoulder rotation and elbow (J3, J4)
+  - **DM4310** (DM-J4310-2EC V1.1) for wrist (J5, J6, J7) and gripper (J8)
+- **24V power supply** required
+- **CAN interface device**:
+  - **Linux**: Any SocketCAN-compatible adapter
+  - **macOS**: CANable, PEAK PCAN-USB, or Kvaser USBcan
+- Proper CAN wiring (CANH, CANL, 120Ω termination)
+
+
+## Motor Configuration
+
+Each arm has the following motor configuration based on the [OpenArm setup guide](https://docs.openarm.dev/software/setup/):
+
+| Joint | Motor | Motor Type | Sender CAN ID | Receiver ID | Description |
+|-------|-------|------------|---------------|-------------|-------------|
+| J1 | joint_1 | DM8009 | 0x01 | 0x11 | Shoulder pan |
+| J2 | joint_2 | DM8009 | 0x02 | 0x12 | Shoulder lift |
+| J3 | joint_3 | DM4340 | 0x03 | 0x13 | Shoulder rotation |
+| J4 | joint_4 | DM4340 | 0x04 | 0x14 | Elbow flex |
+| J5 | joint_5 | DM4310 | 0x05 | 0x15 | Wrist roll |
+| J6 | joint_6 | DM4310 | 0x06 | 0x16 | Wrist pitch |
+| J7 | joint_7 | DM4310 | 0x07 | 0x17 | Wrist rotation |
+| J8 | gripper | DM4310 | 0x08 | 0x18 | Gripper |
+
+For dual arm setups, the left arm uses IDs 0x09-0x10 for joints 1-8 with the same motor types.
+
+## Quick Start 
+
+```bash
+# Install system dependencies
+sudo apt install can-utils iproute2
+
+# Install LeRobot with OpenArms support
+pip install -e ".[openarms]"
+```
+
+## Setup Guide
+
+### Step 1: Motor ID Configuration
+
+**IMPORTANT**: Before using the robot, motors must be configured with the correct CAN IDs.
+
+Refer to the [OpenArm Motor ID Configuration Guide](https://docs.openarm.dev/software/setup/motor-id) for detailed instructions using the Damiao Debugging Tools on Windows.
+
+Key points:
+- Each motor needs a unique **Sender CAN ID** (0x01-0x08)
+- Each motor needs a unique **Receiver/Master ID** (0x11-0x18)
+- Use the Damiao Debugging Tools to set these IDs
+
+### Step 2: Setup CAN Interface
+
+Configure your CAN interface as described in the [OpenArm CAN Setup Guide](https://docs.openarm.dev/software/setup/can-setup):
+
+#### Linux (SocketCAN)
+
+```bash
+# Find your CAN interface
+ip link show
+
+# Configure can0, 1, 2, 3
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000
+sudo ip link set can0 up
+
+sudo ip link set can1 down
+sudo ip link set can1 type can bitrate 1000000
+sudo ip link set can1 up
+
+sudo ip link set can2 down
+sudo ip link set can2 type can bitrate 1000000
+sudo ip link set can2 up
+
+sudo ip link set can3 down
+sudo ip link set can3 type can bitrate 1000000
+sudo ip link set can3 up
+
+# Verify configuration
+ip link show can0
+```
+
+or run:
+
+`examples/openarms/setup_can.sh`
+
+### Testing canbus and motor connection
+
+Please run this script to check if all motors can be found and to find your can-fd speed: `python examples/openarms/debug_can_communication.py`
+
+## Usage
+
+### Basic Setup
+
+
+```python
+from lerobot.robots.openarms import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+# Configure for dual arm setup
+config = OpenArmsFollowerConfig(
+    port="can0",
+    can_interface="socketcan",  # Or "auto" for auto-detection
+    id="openarms_dual",
+    is_dual_arm=True,
+)
+
+robot = OpenArmsFollower(config)
+robot.connect()
+```
+
+### Calibration
+
+On first use, you'll need to calibrate the robot:
+
+```python
+robot.calibrate()
+```
+
+The calibration process will:
+1. Disable torque on all motors
+2. Ask you to position arms in **hanging position with grippers closed**
+3. Set this as the zero position
+4. Ask you to move each joint through its full range
+5. Record min/max positions for each joint
+6. Save calibration to file
+
+### Reading Observations
+
+The robot provides comprehensive state information:
+
+```python
+observation = robot.get_observation()
+
+# Observation includes for each motor:
+# - {motor_name}.pos: Position in degrees
+# - {motor_name}.vel: Velocity in degrees/second  
+# - {motor_name}.torque: Motor torque
+# - {camera_name}: Camera images (if configured)
+
+print(f"Right arm joint 1 position: {observation['right_joint_1.pos']:.1f}°")
+print(f"Right arm joint 1 velocity: {observation['right_joint_1.vel']:.1f}°/s")
+print(f"Right arm joint 1 torque: {observation['right_joint_1.torque']:.3f} N·m")
+```
+
+### Sending Actions
+
+```python
+# Send target positions (in degrees)
+action = {
+    "right_joint_1.pos": 45.0,
+    "right_joint_2.pos": -30.0,
+    # ... all joints
+    "right_gripper.pos": 45.0,  # Half-closed
+}
+
+actual_action = robot.send_action(action)
+```
+
+### Gripper Control
+
+```python
+# Open gripper
+robot.open_gripper(arm="right")
+
+# Close gripper  
+robot.close_gripper(arm="right")
+```
+
+## Safety Features
+
+### 1. Maximum Relative Target
+
+Limits how far a joint can move in a single command to prevent sudden movements:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    # Limit all joints to 10 degrees per command
+    max_relative_target=10.0,
+    
+    # Or set per-motor limits
+    max_relative_target={
+        "right_joint_1": 15.0,  # Slower moving joint
+        "right_joint_2": 10.0,
+        "right_gripper": 5.0,   # Very slow gripper
+    }
+)
+```
+
+**How it works**: If current position is 50° and you command 80°, with `max_relative_target=10.0`, the robot will only move to 60° in that step.
+
+### 2. Torque Limits
+
+Control maximum torque output, especially important for grippers and teleoperation:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    # Gripper torque limit (fraction of motor's max torque)
+    gripper_torque_limit=0.5,  # 50% of max torque
+)
+```
+
+Lower torque limits prevent damage when gripping delicate objects.
+
+### 3. MIT Control Gains
+
+Control responsiveness and stability via PID-like gains:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    position_kp=10.0,  # Position gain (higher = more responsive)
+    position_kd=0.5,   # Velocity damping (higher = more damped)
+)
+```
+
+**Guidelines**:
+- **For following (robot)**: Higher gains for responsiveness
+  - `position_kp=10.0`, `position_kd=0.5`
+- **For teleoperation (leader)**: Lower gains or disable torque for manual movement
+  - `manual_control=True` (torque disabled)
+
+### 4. Velocity Limits
+
+Velocity limits are enforced by the Damiao motors based on motor type. For DM4310:
+- Max velocity: 30 rad/s ≈ 1718°/s
+
+The motors will automatically limit velocity to safe values.
+
+## Teleoperation
+
+### Leader Arm Setup
+
+The leader arm is moved manually (torque disabled) to generate commands:
+
+```python
+from lerobot.teleoperators.openarms import OpenArmsLeader
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+
+config = OpenArmsLeaderConfig(
+    port="can1",  # Separate CAN interface for leader
+    id="openarms_leader",
+    manual_control=True,  # Torque disabled for manual movement
+    is_dual_arm=True,
+)
+
+leader = OpenArmsLeader(config)
+leader.connect()
+
+# Read current position as action
+action = leader.get_action()
+# action contains positions for all joints in degrees
+```
+
+### Safety Considerations for Teleoperation
+
+1. **Use separate CAN interfaces** for leader and follower to avoid conflicts
+2. **Enable max_relative_target** on follower to smooth abrupt movements
+3. **Lower torque limits** on follower to prevent damage from tracking errors
+4. **Test with one arm** before enabling dual arm teleoperation
+5. **Have emergency stop** ready (power switch or CAN disable)
+
+```python
+# Recommended follower config for teleoperation
+follower_config = OpenArmsFollowerConfig(
+    port="can0",
+    max_relative_target=5.0,  # Small steps for smooth following
+    gripper_torque_limit=0.3, # Low torque for safety
+    position_kp=5.0,  # Lower gains for gentler following
+    position_kd=0.3,
+)
+```
+
+## Troubleshooting
+
+### Motor Shaking/Unstable
+
+- **Lower control gains**: Reduce `position_kp` and `position_kd`
+- **Check calibration**: Re-run calibration procedure
+- **Verify power**: Insufficient current can cause instability
+- **Check mechanical**: Loose connections, binding, or damaged components
+
+### CAN Bus Errors
+
+```bash
+# Check for errors
+ip -s link show can0
+
+# Reset CAN interface
+sudo ip link set can0 down
+sudo ip link set can0 up
+```
+
+### Control Mode
+
+OpenArms uses **MIT control mode** which allows simultaneous control of:
+- Position (degrees)
+- Velocity (degrees/second)
+- Torque (N·m)
+- Position gain (Kp)
+- Velocity damping (Kd)
+
+### Communication
+
+- **Protocol**: CAN 2.0 at 1 Mbps (or CAN-FD at 5 Mbps)
+- **Frame format**: Standard 11-bit IDs
+- **Update rate**: Typically 50-100 Hz depending on motor count
+- **Latency**: ~10-20ms per motor command
+
+## References
+
+- [OpenArm Official Documentation](https://docs.openarm.dev/)
+- [OpenArm Setup Guide](https://docs.openarm.dev/software/setup/)
+- [Motor ID Configuration](https://docs.openarm.dev/software/setup/motor-id)
+- [CAN Interface Setup](https://docs.openarm.dev/software/setup/can-setup)
+- [Motor Communication Test](https://docs.openarm.dev/software/setup/configure-test)
+- [Damiao Motor Documentation](https://wiki.seeedstudio.com/damiao_series/)
+- [Enactic GitHub](https://github.com/enactic/openarm_can)

docs/source/phone_teleop.mdx

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

docs/source/pi0.mdx

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

docs/source/pi05.mdx

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

docs/source/pi0fast.mdx

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

docs/source/policy_multi_task_dit_README.md

@@ -1,37 +0,0 @@
-# Multitask DiT Policy
-
-## Citation
-
-If you use this work, please cite the following works:
-
-```bibtex
-@misc{jones2025multitaskditpolicy,
-  author = {Bryson Jones},
-  title = {Dissecting and Open-Sourcing Multitask Diffusion Transformer Policy},
-  year = {2025},
-  url = {https://brysonkjones.substack.com/p/dissecting-and-open-sourcing-multitask-diffusion-transformer-policy},
-  note = {Blog post}
-}
-```
-
-```bibtex
-@misc{trilbmteam2025carefulexaminationlargebehaviormodels,
-  author       = {TRI LBM Team},
-  title        = {A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation},
-  year         = {2025},
-  eprint       = {arXiv:2507.05331},
-  archivePrefix = {arXiv},
-  primaryClass = {cs.RO},
-  url          = {https://arxiv.org/abs/2507.05331}
-}
-```
-
-```bibtex
-@misc{bostondynamics2025largebehaviormodelsatlas,
-  author       = {Boston Dynamics and TRI Research Team},
-  title        = {Large Behavior Models and Atlas Find New Footing},
-  year         = {2025},
-  url          = {https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/},
-  note         = {Blog post}
-}
-```

docs/source/rac.mdx

@@ -0,0 +1,291 @@
+# RaC: Recovery and Correction Training
+
+RaC (Recovery and Correction) is a human-in-the-loop data collection and training paradigm that improves robot policy performance on long-horizon tasks by explicitly teaching recovery and correction behaviors.
+
+**Key References:**
+- [RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction](https://arxiv.org/abs/2509.07953) (Hu et al., 2025)
+- [HG-DAgger: Interactive Imitation Learning with Human Experts](https://arxiv.org/abs/1810.02890) (Kelly et al., 2019)
+- [π∗0.6: a VLA That Learns From Experience](https://pi.website/blog/pistar06) (Physical Intelligence, 2025)
+- [SARM: Stage-Aware Reward Modeling](https://arxiv.org/abs/2509.25358) (Chen et al., 2025)
+
+---
+
+## Why RaC? The Problem with Standard Data Collection
+
+### Standard Behavioral Cloning Data Collection Limitations
+
+Standard behavior cloning trains policies on successful demonstrations. This approach can be sensitive to distribution shift and compounding errors. Because during deployment small errors can cascade and push the robot into states never seen during training.
+This is where RaC and methods like Dagger and HG-DAgger come in.
+
+### Prior Human-in-the-Loop Methods
+
+**DAgger** (Dataset Aggregation) addresses distribution shift by:
+- Running the novice policy to collect states
+- Querying expert for correct actions at those states
+- Aggregating new labels into training set
+
+**HG-DAgger** (Human-Gated DAgger) improves on DAgger by:
+- Giving human full control authority during interventions
+- Human takes over when unsafe, provides correction, returns control
+- Better action labels because human has uninterrupted control
+
+### RaC
+
+RaC explicitly collects **recovery + correction** data:
+
+```
+BC/DAgger:   policy → mistake → human corrects → continue
+RaC:         policy → mistake → human RECOVERS (teleop back) → CORRECTS → END
+```
+
+The critical insight is **Rule 1 (Recover then Correct)**:
+- Every intervention starts with human teleoperating back to an in-distribution state
+- Then human provides correction to complete the current subtask
+- Both segments are recorded as training data
+- This teaches the policy: "when things go wrong, go back and retry"
+
+**Rule 2 (Terminate after Intervention)**:
+- Episode ends after correction completes
+- Avoids mixed policy/human data on later subtasks
+- Keeps data distribution clean
+
+---
+
+## Comparison Table
+
+| Method | Data Type | Recovery Behavior | Correction Behavior |
+|--------|-----------|-------------------|---------------------|
+| BC | Success only | ✗ | ✗ |
+| DAgger | Success + corrections | ✗ | ✓ |
+| HG-DAgger | Success + corrections | Sometimes | ✓ |
+| RaC | Success + recovery + correction | ✓ Explicit | ✓ |
+
+---
+
+## The RaC Pipeline
+
+```
+┌─────────────────────────────────────────────────────────────────────────┐
+│                         RaC Training Pipeline                           │
+├─────────────────────────────────────────────────────────────────────────┤
+│                                                                         │
+│  1. PRE-TRAINING (Standard BC)                                          │
+│     └─> Train initial policy on clean demonstrations                    │
+│                                                                         │
+│  2. RAC DATA COLLECTION (Human-in-the-loop)                             │
+│     ├─> Policy runs autonomously                                        │
+│     ├─> Human monitors and intervenes when failure imminent             │
+│     │   ├─> RECOVERY: Human teleoperates robot back to good state       │
+│     │   └─> CORRECTION: Human completes the current subtask             │
+│     └─> Episode terminates after correction (Rule 2)                    │
+│                                                                         │
+│  3. REWARD LABELING (Optional: SARM)                                    │
+│     └─> Compute progress rewards for advantage-weighted training        │
+│                                                                         │
+│  4. FINE-TUNING                                                         │
+│     └─> Train on combined demos + RaC data (optionally with RA-BC)      │
+│                                                                         │
+└─────────────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## Step-by-Step Guide
+
+### Step 1: Pre-train a Base Policy
+
+First, train a policy on your demonstration dataset:
+
+```bash
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/demo-dataset \
+    --policy.type=pi0 \
+    --output_dir=outputs/pretrain \
+    --batch_size=32 \
+    --steps=50000
+```
+
+### Step 2: Collect RaC Data
+
+Run the RaC data collection script with your pre-trained policy:
+
+```bash
+python examples/rac/rac_data_collection.py \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --dataset.repo_id=your-username/rac-dataset \
+    --dataset.single_task="Pick up the cube and place it in the bowl" \
+    --dataset.num_episodes=50
+```
+
+**Controls (Keyboard + Foot Pedal):**
+
+| Key / Pedal | Action |
+|-------------|--------|
+| **SPACE** / Right pedal | Pause policy (teleop mirrors robot, no recording) |
+| **c** / Left pedal | Take control (start correction, recording resumes) |
+| **→** / Right pedal | End episode (save) - when in correction mode |
+| **←** | Re-record episode |
+| **ESC** | Stop session and push to hub |
+| Any key/pedal during reset | Start next episode |
+
+**The RaC Protocol:**
+
+1. Watch the policy run autonomously (teleop is idle/free)
+2. When you see imminent failure, press **SPACE** or **right pedal** to pause
+   - Policy stops
+   - Teleoperator moves to match robot position (torque enabled)
+   - No frames recorded during pause
+3. Press **c** or **left pedal** to take control
+   - Teleoperator torque disabled, free to move
+   - **RECOVERY**: Teleoperate back to a good state
+   - **CORRECTION**: Complete the subtask
+   - All movements are recorded
+4. Press **→** or **right pedal** to save and end episode
+5. **RESET**: Teleop moves to robot position, you can move robot to starting position
+6. Press any key/pedal to start next episode
+
+The recovery and correction segments teach the policy how to recover from errors.
+
+**Foot Pedal Setup (Linux):**
+
+If using a USB foot pedal (PCsensor FootSwitch), ensure access:
+```bash
+sudo setfacl -m u:$USER:rw /dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd
+```
+
+### Step 3: (Optional) Compute SARM Rewards
+
+For advantage-weighted training (RA-BC / Pi0.6-style), compute SARM progress values:
+
+```bash
+python src/lerobot/policies/sarm/compute_rabc_weights.py \
+    --dataset-repo-id your-username/rac-dataset \
+    --reward-model-path your-username/sarm-model \
+    --head-mode sparse \
+    --push-to-hub
+```
+
+### Step 4: Fine-tune Policy
+
+Fine-tune on the RaC data:
+
+```bash
+# Without RA-BC (standard fine-tuning)
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/rac-dataset \
+    --policy.type=pi0 \
+    --policy.pretrained_path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --output_dir=outputs/rac_finetune \
+    --steps=20000
+
+# With RA-BC (advantage-weighted, Pi0.6-style)
+python src/lerobot/scripts/lerobot_train.py \
+    --dataset.repo_id=your-username/rac-dataset \
+    --policy.type=pi0 \
+    --policy.pretrained_path=outputs/pretrain/checkpoints/last/pretrained_model \
+    --output_dir=outputs/rac_finetune_rabc \
+    --use_rabc=true \
+    --rabc_kappa=0.01 \
+    --steps=20000
+```
+
+---
+
+## Connection to Pi0.6 / RECAP
+
+Pi0.6's RECAP method shares similar principles:
+- Collect autonomous rollouts + expert interventions
+- Use value function to compute **advantages**: A(s,a) = V(s') - V(s)
+- **Advantage conditioning**: Weight training based on expected improvement
+
+In LeRobot, we can use **SARM** as the value function:
+- SARM progress φ(s) ∈ [0,1] measures task completion
+- Progress delta = φ(s') - φ(s) approximates advantage
+- RA-BC uses these to weight training samples (higher weight for good corrections)
+
+---
+
+## Tips for Effective RaC Collection
+
+### When to Intervene
+
+Intervene when you see:
+- Robot about to make an irreversible mistake
+- Robot hesitating or showing uncertain behavior
+- Robot deviating from expected trajectory
+
+### Recovery: Teleoperating Back to Good State
+
+During recovery, teleoperate the robot back to a state where:
+- The robot is in a familiar, in-distribution configuration
+- The current subtask can still be completed
+- The recovery trajectory itself is informative training data
+
+### Quality of Corrections
+
+During correction:
+- Provide **confident, clean** trajectories
+- Complete the current subtask fully
+- Don't overcorrect or add unnecessary movements
+
+---
+
+## Iterative Improvement
+
+RaC can be applied iteratively:
+
+```
+┌─────────────────────────────────────────────────────────────────────────┐
+│  Policy v0 (demos)                                                      │
+│       ↓                                                                 │
+│  RaC Collection (target current failure modes) → Policy v1              │
+│       ↓                                                                 │
+│  RaC Collection (target new failure modes) → Policy v2                  │
+│       ↓                                                                 │
+│  ... (repeat until satisfactory performance)                            │
+└─────────────────────────────────────────────────────────────────────────┘
+```
+
+Each iteration:
+1. Deploy current policy
+2. Collect RaC interventions on failure cases
+3. Fine-tune on accumulated data
+
+---
+
+## References
+
+```bibtex
+@article{hu2025rac,
+  title={RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction},
+  author={Hu, Zheyuan and Wu, Robyn and Enock, Naveen and Li, Jasmine and Kadakia, Riya and Erickson, Zackory and Kumar, Aviral},
+  journal={arXiv preprint arXiv:2509.07953},
+  year={2025}
+}
+
+@article{kelly2019hgdagger,
+  title={HG-DAgger: Interactive Imitation Learning with Human Experts},
+  author={Kelly, Michael and Sidrane, Chelsea and Driggs-Campbell, Katherine and Kochenderfer, Mykel J},
+  journal={arXiv preprint arXiv:1810.02890},
+  year={2019}
+}
+
+@article{pi2025recap,
+  title={π∗0.6: a VLA That Learns From Experience},
+  author={Physical Intelligence},
+  year={2025}
+}
+
+@article{chen2025sarm,
+  title={SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation},
+  author={Chen, Qianzhong and Yu, Justin and Schwager, Mac and Abbeel, Pieter and Shentu, Yide and Wu, Philipp},
+  journal={arXiv preprint arXiv:2509.25358},
+  year={2025}
+}
+```
+

docs/source/reachy2.mdx

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

docs/source/sarm.mdx

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

docs/source/smolvla.mdx

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

docs/source/streaming_video_encoding.mdx

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

docs/source/training_time_rtc.mdx

@@ -0,0 +1,86 @@
+# Training-Time RTC
+
+Training-Time RTC teaches the model to handle inference delay during training.
+It feeds the **ground-truth action prefix** to the model and trains only on the remaining postfix actions.
+This keeps chunk transitions smooth without doing any inference-time inpainting.
+
+Based on: [Training-Time Action Conditioning for Efficient Real-Time Chunking](https://arxiv.org/abs/2512.05964).
+
+LeRobot supports this for `pi0`, `pi05` and `smolvla` without changing model parameters.
+
+---
+
+## How It Works
+
+### At Training Time
+
+- Sample a delay `d` per batch element.
+- Keep the first `d` action steps as **ground truth** (no noise).
+- Add noise only to the postfix actions.
+- Set the flow-matching timestep to **1.0** for prefix tokens and normal timesteps for postfix tokens.
+- Mask the loss to only train on the postfix.
+
+### At Inference Time
+
+When `rtc_training_config.enabled=true`, the model uses training-time RTC inference:
+
+- Replace prefix positions in `x_t` with previous chunk's leftover actions.
+- Set timestep to **1.0** for prefix positions.
+
+---
+
+## Quick Start (CLI)
+
+```bash
+lerobot-train \
+  --policy.type=pi0 \
+  --dataset.repo_id=your/dataset \
+  --policy.rtc_training_config.enabled=true \
+  --policy.rtc_training_config.min_delay=0 \
+  --policy.rtc_training_config.max_delay=6 \
+  --policy.rtc_training_config.delay_distribution=UNIFORM
+```
+
+---
+
+## Inference with Training-Time RTC
+
+After training with `rtc_training_config`, use the same config at inference. The model will automatically use training-time RTC inference:
+
+```python
+policy = PI0Policy.from_pretrained("path/to/trained/model")
+# rtc_training_config is loaded from the saved config
+
+actions = policy.predict_action_chunk(
+    batch,
+    inference_delay=5,  # estimated delay in timesteps
+    prev_chunk_left_over=previous_actions,  # from previous chunk
+)
+```
+
+---
+
+## Key Parameters
+
+`RTCTrainingConfig` is available on the policy config (`pi0`, `pi05`, `smolvla`, `xvla`):
+
+- **`enabled`**: Toggle training-time RTC (both training and inference).
+- **`min_delay` / `max_delay`**: Delay range (inclusive).
+- **`delay_distribution`**:
+  - `UNIFORM`: uniform in `[min_delay, max_delay]`
+  - `EXP`: exponentially decayed distribution over delays
+- **`exp_decay`**: Exponential decay factor for `EXP` sampling.
+
+---
+
+## Notes and Recommendations
+
+- Start with `min_delay=0` and `max_delay` around your expected worst-case inference delay.
+- Use `EXP` if you want more supervision on smaller delays.
+
+---
+
+## Related Docs
+
+- [Real-Time Chunking (Inference-Time RTC)](./rtc)
+- [Pi0](./pi0), [Pi0.5](./pi05), [SmolVLA](./smolvla)

docs/source/unitree_g1.mdx

@@ -123,7 +123,7 @@ SSH into the robot and install LeRobot:
 ```bash
 ssh unitree@<YOUR_ROBOT_IP>
 
-conda create -y -n lerobot python=3.12
+conda create -y -n lerobot python=3.10
 conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
@@ -153,7 +153,7 @@ With the robot server running, you can now control the robot remotely. Let's lau
 ### Step 1: Install LeRobot on your machine
 
 ```bash
-conda create -y -n lerobot python=3.12
+conda create -y -n lerobot python=3.10
 conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git
 cd lerobot
@@ -216,7 +216,7 @@ lerobot-teleoperate \
 ### Record Dataset in Simulation
 
 ```bash
-lerobot-record \
+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}}' \
@@ -229,10 +229,7 @@ lerobot-record \
     --dataset.num_episodes=2 \
     --dataset.episode_time_s=5 \
     --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
+    --dataset.push_to_hub=true
 ```
 
 Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
@@ -269,7 +266,7 @@ lerobot-teleoperate \
 ### Record Dataset on Real Robot
 
 ```bash
-lerobot-record \
+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}}' \
@@ -282,10 +279,7 @@ lerobot-record \
     --dataset.num_episodes=2 \
     --dataset.episode_time_s=5 \
     --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
+    --dataset.push_to_hub=true
 ```
 
 **Note**: Update `server_address` to match your robot's camera server IP.

docs/source/using_dataset_tools.mdx

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

docs/source/walloss.mdx

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

docs/source/xvla.mdx

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

examples/backward_compatibility/replay.py

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

examples/dataset/slurm_compute_rabc.py

@@ -1,490 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-SLURM-distributed SARM RA-BC annotation pipeline.
-
-Computes SARM progress values for all frames in a dataset, distributed across
-SLURM workers, then merges the shards into a single sarm_progress.parquet.
-
-Two subcommands, each a separate SLURM submission:
-
-  compute    – N workers, each computes progress for a subset of episodes
-  aggregate  – 1 worker, merges N shards into sarm_progress.parquet, pushes to hub
-
-Usage:
-    python slurm_compute_rabc.py compute \\
-        --repo-id user/dataset --reward-model-path user/sarm_model \\
-        --stride 10 --device cpu --workers 50 --partition cpu
-
-    python slurm_compute_rabc.py aggregate \\
-        --repo-id user/dataset --reward-model-path user/sarm_model \\
-        --partition cpu --push-to-hub
-"""
-
-import argparse
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-
-
-class ComputeProgressShards(PipelineStep):
-    """Each worker computes SARM progress for its assigned episodes."""
-
-    def __init__(
-        self, repo_id, reward_model_path, stride=1, head_mode="sparse", device="cpu", shard_dir="rabc_shards"
-    ):
-        super().__init__()
-        if stride < 1:
-            raise ValueError(f"stride must be >= 1, got {stride}")
-        self.repo_id = repo_id
-        self.reward_model_path = reward_model_path
-        self.stride = stride
-        self.head_mode = head_mode
-        self.device = device
-        self.shard_dir = shard_dir
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import logging
-        from pathlib import Path
-
-        import numpy as np
-        import pyarrow as pa
-        import pyarrow.parquet as pq
-        import torch
-        from tqdm import tqdm
-
-        from lerobot.policies.sarm.compute_rabc_weights import (
-            generate_all_frame_indices,
-            interpolate_progress,
-            load_sarm_resources,
-        )
-        from lerobot.utils.utils import init_logging
-
-        init_logging()
-
-        dataset, reward_model, preprocess = load_sarm_resources(
-            self.repo_id,
-            self.reward_model_path,
-            self.device,
-        )
-
-        if hasattr(preprocess, "eval"):
-            preprocess.eval()
-        for step in preprocess.steps:
-            if hasattr(step, "eval"):
-                step.eval()
-
-        image_key = reward_model.config.image_key
-        state_key = reward_model.config.state_key
-        frame_gap = reward_model.config.frame_gap
-        center_idx = reward_model.config.n_obs_steps // 2
-
-        dual_mode = reward_model.config.uses_dual_heads
-        compute_sparse = self.head_mode in ("sparse", "both") or not dual_mode
-        compute_dense = self.head_mode in ("dense", "both") and dual_mode
-
-        my_episodes = list(range(dataset.num_episodes))[rank::world_size]
-        if not my_episodes:
-            logging.info(f"Rank {rank}: no episodes assigned")
-            return
-        logging.info(f"Rank {rank}: {len(my_episodes)} / {dataset.num_episodes} episodes")
-
-        all_rows = []
-
-        for ep_idx in tqdm(my_episodes, desc=f"Rank {rank}"):
-            ep = dataset.meta.episodes[ep_idx]
-            ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
-            task = dataset[ep_start].get("task", "perform the task")
-
-            all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
-            if self.stride > 1:
-                compute_indices = [i for i in all_ep_indices if (i - ep_start) % self.stride == 0]
-                if (ep_end - 1) not in compute_indices:
-                    compute_indices.append(ep_end - 1)
-                compute_indices = sorted(set(compute_indices))
-            else:
-                compute_indices = all_ep_indices
-
-            frame_results = {}
-            for qi in tqdm(compute_indices, desc=f"  Ep {ep_idx}", leave=False):
-                try:
-                    sample = dataset[qi]
-                    batch = {
-                        image_key: sample[image_key],
-                        "task": task,
-                        "index": qi,
-                        "episode_index": ep_idx,
-                    }
-                    if state_key in sample:
-                        batch[state_key] = sample[state_key]
-
-                    with torch.no_grad():
-                        processed = preprocess(batch)
-                        vf = processed["video_features"].to(self.device)
-                        tf = processed["text_features"].to(self.device)
-                        sf = processed.get("state_features")
-                        if sf is not None:
-                            sf = sf.to(self.device)
-                        lengths = processed.get("lengths")
-
-                        sparse_val = dense_val = np.nan
-                        if compute_sparse:
-                            r = reward_model.calculate_rewards(
-                                text_embeddings=tf,
-                                video_embeddings=vf,
-                                state_features=sf,
-                                lengths=lengths,
-                                return_all_frames=True,
-                                head_mode="sparse",
-                            )
-                            sparse_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
-                        if compute_dense:
-                            r = reward_model.calculate_rewards(
-                                text_embeddings=tf,
-                                video_embeddings=vf,
-                                state_features=sf,
-                                lengths=lengths,
-                                return_all_frames=True,
-                                head_mode="dense",
-                            )
-                            dense_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
-
-                        frame_results[qi] = (sparse_val, dense_val)
-                except Exception as e:
-                    logging.warning(f"Failed frame {qi}: {e}")
-
-            if not frame_results:
-                logging.warning(f"Episode {ep_idx}: all frames failed, skipping")
-                continue
-
-            # Interpolate to all frames in this episode
-            computed_idx = np.array(sorted(frame_results.keys()))
-            all_frame_arr = np.arange(ep_start, ep_end)
-
-            sparse_vals = np.array([frame_results[i][0] for i in computed_idx]) if compute_sparse else None
-            dense_vals = np.array([frame_results[i][1] for i in computed_idx]) if compute_dense else None
-
-            if self.stride > 1 and len(computed_idx) > 1:
-                if compute_sparse:
-                    sparse_vals = interpolate_progress(computed_idx, sparse_vals, all_frame_arr)
-                if compute_dense:
-                    dense_vals = interpolate_progress(computed_idx, dense_vals, all_frame_arr)
-                output_frames = all_frame_arr
-            else:
-                # Use only successfully computed frames to avoid indexing mismatch on failures
-                output_frames = computed_idx
-
-            for i, fi in enumerate(output_frames):
-                row = {"index": int(fi), "episode_index": ep_idx, "frame_index": int(fi - ep_start)}
-                if compute_sparse:
-                    row["progress_sparse"] = float(sparse_vals[i])
-                if compute_dense:
-                    row["progress_dense"] = float(dense_vals[i])
-                all_rows.append(row)
-
-        if all_rows:
-            import pandas as pd
-
-            df = pd.DataFrame(all_rows).sort_values("index").reset_index(drop=True)
-            table = pa.Table.from_pandas(df, preserve_index=False)
-            table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
-            shard_dir = Path(self.shard_dir)
-            shard_dir.mkdir(parents=True, exist_ok=True)
-            out = shard_dir / f"shard_{rank:05d}.parquet"
-            pq.write_table(table, out)
-            logging.info(f"Rank {rank}: saved {len(df)} rows to {out}")
-
-
-class AggregateProgress(PipelineStep):
-    """Merge all shard parquets into final sarm_progress.parquet."""
-
-    def __init__(self, repo_id, reward_model_path, shard_dir="rabc_shards", push_to_hub=False):
-        super().__init__()
-        self.repo_id = repo_id
-        self.reward_model_path = reward_model_path
-        self.shard_dir = shard_dir
-        self.push_to_hub = push_to_hub
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import datetime
-        import logging
-        import os
-        from pathlib import Path
-
-        import pandas as pd
-        import pyarrow as pa
-        import pyarrow.parquet as pq
-
-        from lerobot.datasets.lerobot_dataset import LeRobotDataset
-        from lerobot.utils.utils import init_logging
-
-        init_logging()
-        if rank != 0:
-            return
-
-        shard_dir = Path(self.shard_dir)
-        shards = sorted(shard_dir.glob("shard_*.parquet"))
-        if not shards:
-            raise FileNotFoundError(f"No shards found in {shard_dir}")
-
-        # Log shard modification time range to help detect stale files
-        mtimes = [os.path.getmtime(s) for s in shards]
-        oldest = datetime.datetime.fromtimestamp(min(mtimes)).isoformat(timespec="seconds")
-        newest = datetime.datetime.fromtimestamp(max(mtimes)).isoformat(timespec="seconds")
-        logging.info(f"Aggregating {len(shards)} shards (oldest: {oldest}, newest: {newest})")
-
-        df = pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True)
-        df = df.sort_values("index").reset_index(drop=True)
-
-        table = pa.Table.from_pandas(df, preserve_index=False)
-        table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
-
-        temp_ds = LeRobotDataset(self.repo_id, download_videos=False)
-        out_path = Path(temp_ds.root) / "sarm_progress.parquet"
-        out_path.parent.mkdir(parents=True, exist_ok=True)
-        pq.write_table(table, out_path)
-        logging.info(f"Saved {len(df)} rows to {out_path}")
-
-        for col in ["progress_sparse", "progress_dense"]:
-            if col in df.columns:
-                v = df[col].dropna()
-                logging.info(
-                    f"{col}: mean={v.mean():.4f} std={v.std():.4f} min={v.min():.4f} max={v.max():.4f}"
-                )
-
-        if self.push_to_hub:
-            from huggingface_hub import HfApi
-
-            api = HfApi()
-            hub_path = "sarm_progress.parquet"
-            logging.info(f"Uploading to {self.repo_id}/{hub_path}")
-            api.upload_file(
-                path_or_fileobj=str(out_path),
-                path_in_repo=hub_path,
-                repo_id=self.repo_id,
-                repo_type="dataset",
-            )
-            logging.info(f"Uploaded: https://huggingface.co/datasets/{self.repo_id}/blob/main/{hub_path}")
-
-
-def make_compute_executor(
-    repo_id,
-    reward_model_path,
-    stride,
-    head_mode,
-    device,
-    shard_dir,
-    logs_dir,
-    job_name,
-    slurm,
-    workers,
-    partition,
-    cpus_per_task,
-    mem_per_cpu,
-):
-    kwargs = {
-        "pipeline": [
-            ComputeProgressShards(repo_id, reward_model_path, stride, head_mode, device, str(shard_dir)),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": workers,
-                "workers": workers,
-                "time": "24:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        return SlurmPipelineExecutor(**kwargs)
-
-    kwargs.update({"tasks": workers, "workers": 1})
-    return LocalPipelineExecutor(**kwargs)
-
-
-def make_aggregate_executor(
-    repo_id,
-    reward_model_path,
-    shard_dir,
-    logs_dir,
-    job_name,
-    slurm,
-    partition,
-    cpus_per_task,
-    mem_per_cpu,
-    push_to_hub,
-):
-    kwargs = {
-        "pipeline": [
-            AggregateProgress(repo_id, reward_model_path, str(shard_dir), push_to_hub),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": 1,
-                "workers": 1,
-                "time": "02:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        return SlurmPipelineExecutor(**kwargs)
-
-    kwargs.update({"tasks": 1, "workers": 1})
-    return LocalPipelineExecutor(**kwargs)
-
-
-def _add_shared_args(p):
-    p.add_argument(
-        "--repo-id",
-        type=str,
-        required=True,
-        help="Hugging Face repository identifier, e.g. 'user/dataset'.",
-    )
-    p.add_argument(
-        "--shard-dir",
-        type=Path,
-        default=Path("rabc_shards"),
-        help="Directory to read/write per-rank parquet shards.",
-    )
-    p.add_argument(
-        "--logs-dir",
-        type=Path,
-        default=Path("logs"),
-        help="Directory for datatrove logs.",
-    )
-    p.add_argument(
-        "--job-name",
-        type=str,
-        default=None,
-        help="SLURM job name (defaults to rabc_<subcommand>).",
-    )
-    p.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="1 = submit via SLURM; 0 = run locally (useful for debugging).",
-    )
-    p.add_argument(
-        "--partition",
-        type=str,
-        default=None,
-        help="SLURM partition to submit to.",
-    )
-    p.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=4,
-        help="Number of CPUs per SLURM task.",
-    )
-    p.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="4G",
-        help="Memory per CPU, e.g. '4G' or '1950M'.",
-    )
-
-
-def main():
-    parser = argparse.ArgumentParser(
-        description="SLURM-distributed SARM RA-BC annotation pipeline",
-        formatter_class=argparse.RawDescriptionHelpFormatter,
-    )
-    sub = parser.add_subparsers(dest="command", required=True)
-
-    # compute subcommand
-    cp = sub.add_parser(
-        "compute",
-        help="Distribute progress computation across SLURM workers.",
-    )
-    _add_shared_args(cp)
-    cp.add_argument(
-        "--reward-model-path",
-        type=str,
-        required=True,
-        help="Path or HF repo id of the SARM reward model.",
-    )
-    cp.add_argument(
-        "--stride",
-        type=int,
-        default=1,
-        help="Compute every Nth frame; intermediate frames are interpolated (must be >= 1).",
-    )
-    cp.add_argument(
-        "--head-mode",
-        type=str,
-        default="sparse",
-        choices=["sparse", "dense", "both"],
-        help="Which reward head(s) to compute.",
-    )
-    cp.add_argument(
-        "--device",
-        type=str,
-        default="cpu",
-        help="Device for reward model inference, e.g. 'cpu' or 'cuda'.",
-    )
-    cp.add_argument(
-        "--workers",
-        type=int,
-        default=50,
-        help="Number of parallel SLURM tasks (one shard per worker).",
-    )
-
-    # aggregate subcommand
-    ap = sub.add_parser(
-        "aggregate",
-        help="Merge per-rank shards into a single sarm_progress.parquet.",
-    )
-    _add_shared_args(ap)
-    ap.add_argument(
-        "--reward-model-path",
-        type=str,
-        required=True,
-        help="Path or HF repo id of the SARM reward model (stored in parquet metadata).",
-    )
-    ap.add_argument(
-        "--push-to-hub",
-        action="store_true",
-        help="Upload sarm_progress.parquet to the Hugging Face Hub after aggregation.",
-    )
-
-    args = parser.parse_args()
-    job_name = args.job_name or f"rabc_{args.command}"
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    kwargs["job_name"] = job_name
-    command = kwargs.pop("command")
-
-    executor = make_compute_executor(**kwargs) if command == "compute" else make_aggregate_executor(**kwargs)
-
-    executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/openarms/debug_can_communication.py

@@ -0,0 +1,416 @@
+#!/usr/bin/env python3
+"""
+Comprehensive debug script for OpenArms CAN FD communication.
+Tests all 4 CAN interfaces with CAN FD support.
+"""
+
+import can
+import time
+import sys
+import subprocess
+
+def check_can_interface(port):
+    """Check if CAN interface is UP and configured."""
+    try:
+        result = subprocess.run(['ip', 'link', 'show', port], 
+                              capture_output=True, text=True)
+        if result.returncode != 0:
+            return False, "Interface not found", None
+        
+        output = result.stdout
+        if 'UP' not in output:
+            return False, "Interface is DOWN", None
+        
+        # Check if CAN FD is enabled
+        is_fd = 'fd on' in output.lower() or 'canfd' in output.lower()
+        
+        return True, "Interface is UP", is_fd
+    except FileNotFoundError:
+        return None, "Cannot check (ip command not found)", None
+
+
+def test_motor_on_interface(bus, motor_id, timeout=2.0, use_fd=False):
+    """
+    Test a single motor and return all responses.
+    
+    Returns:
+        list of (arbitration_id, data) tuples for all responses received
+    """
+    # Send enable command
+    enable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+        is_extended_id=False,
+        is_fd=use_fd
+    )
+    
+    try:
+        bus.send(enable_msg)
+    except Exception as e:
+        return None, f"Send error: {e}"
+    
+    # Listen for responses
+    responses = []
+    start_time = time.time()
+    
+    while time.time() - start_time < timeout:
+        msg = bus.recv(timeout=0.1)
+        if msg:
+            responses.append((msg.arbitration_id, msg.data, msg.is_fd if hasattr(msg, 'is_fd') else False))
+    
+    # Send disable command
+    disable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFD],
+        is_extended_id=False,
+        is_fd=use_fd
+    )
+    try:
+        bus.send(disable_msg)
+    except:
+        pass
+    
+    return responses, None
+
+
+def test_interface(port, interface_type="socketcan", use_can_fd=True):
+    """Test all 8 motors on a single CAN interface."""
+    
+    results = {
+        'interface': port,
+        'status': None,
+        'is_fd': use_can_fd,
+        'motors': {}
+    }
+    
+    # Check interface status
+    status_ok, status_msg, interface_has_fd = check_can_interface(port)
+    
+    if interface_has_fd is not None:
+        results['interface_fd_enabled'] = interface_has_fd
+        if use_can_fd and not interface_has_fd:
+            status_msg += " (CAN FD NOT enabled on interface!)"
+        elif interface_has_fd:
+            status_msg += " (CAN FD enabled)"
+    
+    results['status'] = status_msg
+    
+    if status_ok is False:
+        return results
+    
+    # Try to connect
+    try:
+        if use_can_fd:
+            print(f"  Connecting to {port} with CAN FD (1 Mbps / 5 Mbps)...")
+            bus = can.interface.Bus(
+                channel=port,
+                interface=interface_type,
+                bitrate=1000000,
+                data_bitrate=5000000,
+                fd=True
+            )
+        else:
+            print(f"  Connecting to {port} with CAN 2.0 (1 Mbps)...")
+            bus = can.interface.Bus(
+                channel=port,
+                interface=interface_type,
+                bitrate=1000000
+            )
+    except Exception as e:
+        results['status'] = f"Connection failed: {e}"
+        return results
+    
+    try:
+        # Clear any pending messages
+        while bus.recv(timeout=0.01):
+            pass
+        
+        # Test each motor (0x01 to 0x08)
+        for motor_id in range(0x01, 0x09):
+            responses, error = test_motor_on_interface(bus, motor_id, timeout=1.0, use_fd=use_can_fd)
+            
+            if error:
+                results['motors'][motor_id] = {'error': error}
+            elif responses:
+                results['motors'][motor_id] = {
+                    'found': True,
+                    'responses': responses
+                }
+            else:
+                results['motors'][motor_id] = {
+                    'found': False,
+                    'responses': []
+                }
+            
+            time.sleep(0.05)  # Small delay between motors
+        
+    finally:
+        bus.shutdown()
+    
+    return results
+
+
+def print_results(all_results):
+    """Print formatted results for all interfaces."""
+    
+    print("SUMMARY - Motors Found on Each Interface")
+    
+    motor_names = {
+        0x01: "joint_1 (Shoulder pan)",
+        0x02: "joint_2 (Shoulder lift)",
+        0x03: "joint_3 (Shoulder rotation)",
+        0x04: "joint_4 (Elbow flex)",
+        0x05: "joint_5 (Wrist roll)",
+        0x06: "joint_6 (Wrist pitch)",
+        0x07: "joint_7 (Wrist rotation)",
+        0x08: "gripper",
+    }
+    
+    total_found = 0
+    
+    for result in all_results:
+        interface = result['interface']
+        status = result['status']
+        
+        print(f"{interface}: {status}")
+        if result.get('is_fd'):
+            print(f"  Mode: CAN FD")
+        else:
+            print(f"  Mode: CAN 2.0")
+        
+        if 'Connection failed' in status or 'DOWN' in status:
+            print(f"  ⚠ Cannot test {interface}")
+            continue
+        
+        motors_found = 0
+        
+        for motor_id in range(0x01, 0x09):
+            motor_data = result['motors'].get(motor_id, {})
+            motor_name = motor_names.get(motor_id, "Unknown")
+            
+            if motor_data.get('error'):
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ {motor_data['error']}")
+            elif motor_data.get('found'):
+                motors_found += 1
+                total_found += 1
+                responses = motor_data['responses']
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✓ FOUND")
+                
+                for resp_id, data, is_fd in responses:
+                    data_hex = data.hex()
+                    fd_flag = " [FD]" if is_fd else " [2.0]"
+                    print(f"    → Response from 0x{resp_id:02X}{fd_flag}: {data_hex}")
+            else:
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ No response")
+        
+        print(f"\n  Summary: {motors_found}/8 motors found on {interface}")
+    
+    # Overall summary
+    print("OVERALL SUMMARY")
+    print(f"Total motors found across all interfaces: {total_found}")
+    
+    # Analyze configuration
+    print("DIAGNOSIS")
+    
+    for result in all_results:
+        interface = result['interface']
+        motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+        
+        if motors_found == 0:
+            print(f"\n⚠ {interface}: NO MOTORS FOUND")
+            print("  Possible issues:")
+            print("    1. CAN FD mode mismatch (interface vs motor configuration)")
+            print("    2. Missing 120Ω termination resistors at BOTH cable ends")
+            print("    3. Motor timeout parameter set incorrectly (should NOT be 0)")
+            print("    4. CANH/CANL wiring issue")
+            print("    5. Cable too long (>40m for CAN FD at 5Mbps)")
+            
+            # Check FD mismatch
+            if result.get('is_fd') and not result.get('interface_fd_enabled'):
+                print("    ⚠️ CRITICAL: Trying CAN FD but interface NOT configured for FD!")
+                print(f"       Fix: sudo ip link set {interface} type can bitrate 1000000 dbitrate 5000000 fd on")
+                
+        elif motors_found < 8:
+            print(f"\n⚠ {interface}: Only {motors_found}/8 motors responding")
+            print("  Check power and connections for missing motors")
+        else:
+            print(f"\n✓ {interface}: All 8 motors responding correctly!")
+    
+    # Check for unexpected response IDs
+    print("RESPONSE ID ANALYSIS")
+    
+    for result in all_results:
+        interface = result['interface']
+        unexpected = []
+        
+        for motor_id, motor_data in result['motors'].items():
+            if motor_data.get('found'):
+                expected_id = motor_id + 0x10
+                actual_ids = [resp[0] for resp in motor_data['responses']]
+                
+                if expected_id not in actual_ids:
+                    unexpected.append((motor_id, actual_ids))
+        
+        if unexpected:
+            print(f"\n⚠ {interface}: Unexpected response IDs detected")
+            for motor_id, actual_ids in unexpected:
+                expected_id = motor_id + 0x10
+                print(f"  Motor 0x{motor_id:02X}: Expected 0x{expected_id:02X}, "
+                      f"got {[f'0x{id:02X}' for id in actual_ids]}")
+            print("  → Motor Master IDs need reconfiguration")
+        else:
+            motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+            if motors_found > 0:
+                print(f"\n✓ {interface}: All responding motors use correct IDs")
+
+
+def test_communication_speed(interface, motor_id, num_iterations=100):
+    """
+    Test communication speed with a motor.
+    
+    Returns:
+        tuple: (hz, avg_latency_ms) or (None, None) if test failed
+    """
+    try:
+        # Connect to interface
+        bus = can.interface.Bus(
+            channel=interface,
+            interface="socketcan",
+            bitrate=1000000,
+            data_bitrate=5000000,
+            fd=True
+        )
+        
+        # Send refresh commands and measure round-trip time
+        latencies = []
+        successful = 0
+        
+        for _ in range(num_iterations):
+            start = time.perf_counter()
+            
+            # Send enable command (lightweight operation)
+            enable_msg = can.Message(
+                arbitration_id=motor_id,
+                data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+                is_extended_id=False,
+                is_fd=True
+            )
+            bus.send(enable_msg)
+            
+            # Wait for response
+            msg = bus.recv(timeout=0.1)
+            
+            if msg:
+                latency = (time.perf_counter() - start) * 1000  # Convert to ms
+                latencies.append(latency)
+                successful += 1
+        
+        bus.shutdown()
+        
+        if successful > 0:
+            avg_latency = sum(latencies) / len(latencies)
+            hz = 1000.0 / avg_latency if avg_latency > 0 else 0
+            return hz, avg_latency
+        
+        return None, None
+        
+    except Exception as e:
+        print(f"    Speed test error: {e}")
+        return None, None
+
+
+def main():
+    """Main function to test all CAN interfaces with CAN FD."""
+    
+    print("\nThis will test all 4 CAN interfaces (can0-can3) with CAN FD")
+    print("Testing motors 0x01-0x08 on each interface")
+    print()
+    print("Make sure:")
+    print("  ✓ Motors are powered (24V)")
+    print("  ✓ CAN interfaces configured with FD mode:")
+    print("    ./examples/openarms/setup_can.sh")
+    print("  ✓ Motor 'timeout' parameter NOT set to 0 (use Damiao tools)")
+    print("  ✓ CAN wiring includes 120Ω termination at BOTH ends")
+    print()
+    
+    input("Press ENTER to start testing...")
+    
+    # Test all 4 interfaces with CAN FD
+    all_results = []
+    
+    for i in range(4):
+        interface = f"can{i}"
+        print(f"Testing {interface}...")
+        
+        result = test_interface(interface, use_can_fd=True)
+        all_results.append(result)
+        
+        # Quick status
+        if 'Connection failed' in result['status'] or 'DOWN' in result['status']:
+            print(f"  ⚠ {interface}: {result['status']}")
+        else:
+            motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+            print(f"  {interface}: {motors_found}/8 motors found")
+        
+        time.sleep(0.2)
+    
+    # Print detailed results
+    print_results(all_results)
+    
+    print("Testing Complete!")
+    
+    all_found = sum(sum(1 for m in r['motors'].values() if m.get('found')) for r in all_results)
+    
+    if all_found == 0:
+        print("\n⚠️ CRITICAL: No motors found on any interface!")
+        print("\nTop issues to check:")
+        print("  1. Motor 'timeout' parameter (use Damiao tools to set > 0)")
+        print("  2. CAN FD not enabled (run ./examples/openarms/setup_can.sh)")
+        print("  3. Missing termination resistors")
+        print("\nTry:")
+        print("  a) Check motor parameters with Damiao Debugging Tools")
+        print("  b) Verify CAN FD is enabled: ip -d link show can0 | grep fd")
+        print("  c) Run setup script: ./examples/openarms/setup_can.sh")
+    else:
+        # Run speed test on interfaces with motors
+        print("COMMUNICATION SPEED TEST")
+        print("\nTesting maximum communication frequency...")
+        
+        for result in all_results:
+            interface = result['interface']
+            
+            # Find first responding motor
+            responding_motor = None
+            for motor_id, motor_data in result['motors'].items():
+                if motor_data.get('found'):
+                    responding_motor = motor_id
+                    break
+            
+            if responding_motor:
+                print(f"\n{interface}: Testing with motor 0x{responding_motor:02X}...")
+                hz, latency = test_communication_speed(interface, responding_motor, num_iterations=100)
+                
+                if hz:
+                    print(f"  ✓ Max frequency: {hz:.1f} Hz")
+                    print(f"  ✓ Avg latency: {latency:.2f} ms")
+                    print(f"  ✓ Commands per second: ~{int(hz)}")
+                else:
+                    print(f"  ✗ Speed test failed")
+            else:
+                print(f"\n{interface}: No motors found, skipping speed test")
+
+        print()
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except KeyboardInterrupt:
+        print("\n\nTesting interrupted by user.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"\nUnexpected error: {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+

examples/openarms/evaluate.py

@@ -0,0 +1,360 @@
+#!/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.
+
+"""
+OpenArms Policy Evaluation
+
+Evaluates a trained policy on the OpenArms robot by running inference and recording
+the evaluation episodes to a dataset. Supports optional leader arm for manual resets.
+
+Example usage:
+    python examples/openarms/evaluate.py
+"""
+
+import time
+from pathlib import Path
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.processor import make_default_processors
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+
+HF_MODEL_ID = "lerobot-data-collection/level1_rac2_100k"  # TODO: Replace with your trained model
+HF_EVAL_DATASET_ID = "lerobot-data-collection/three-folds-pi0_eval_raccc3"  # TODO: Replace with your eval dataset name
+TASK_DESCRIPTION = "Fold the T-shirt properly" # TODO: Replace with your task, this should match!!
+
+NUM_EPISODES = 1
+FPS = 30
+EPISODE_TIME_SEC = 1000
+RESET_TIME_SEC = 60
+
+# Robot CAN interfaces
+FOLLOWER_LEFT_PORT = "can0"
+FOLLOWER_RIGHT_PORT = "can1"
+
+# If enabled, you can manually reset the environment between evaluation episodes
+USE_LEADER_FOR_RESETS = False  # Set to False if you don't want to use leader
+LEADER_LEFT_PORT = "can2"
+LEADER_RIGHT_PORT = "can3"
+
+# Camera configuration
+CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video0", width=1280, height=720, fps=FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video5", width=1280, height=720, fps=FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video2", width=640, height=480, fps=FPS),
+}
+
+def main():
+    """Main evaluation function."""
+    print("OpenArms Policy Evaluation")
+    print(f"\nModel: {HF_MODEL_ID}")
+    print(f"Evaluation Dataset: {HF_EVAL_DATASET_ID}")
+    print(f"Task: {TASK_DESCRIPTION}")
+    print(f"Episodes: {NUM_EPISODES}")
+    print(f"Episode Duration: {EPISODE_TIME_SEC}s")
+    print(f"Reset Duration: {RESET_TIME_SEC}s")
+    print(f"Use Leader for Resets: {USE_LEADER_FOR_RESETS}")
+    
+    follower_config = OpenArmsFollowerConfig(
+        port_left=FOLLOWER_LEFT_PORT,
+        port_right=FOLLOWER_RIGHT_PORT,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=CAMERA_CONFIG,
+    )
+    
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+    
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    
+    leader = None
+    if USE_LEADER_FOR_RESETS:
+        leader_config = OpenArmsLeaderConfig(
+            port_left=LEADER_LEFT_PORT,
+            port_right=LEADER_RIGHT_PORT,
+            can_interface="socketcan",
+            id="openarms_leader",
+            manual_control=False,  # Enable torque control for gravity compensation
+        )
+        
+        leader = OpenArmsLeader(leader_config)
+        leader.connect(calibrate=False)
+        
+        if not leader.is_connected:
+            raise RuntimeError("Leader robot failed to connect!")
+        
+        # Enable gravity compensation
+        if leader.pin_robot is not None:
+            leader.bus_right.enable_torque()
+            leader.bus_left.enable_torque()
+            time.sleep(0.1)
+            print(f"Leader connected with gravity compensation ({LEADER_LEFT_PORT}, {LEADER_RIGHT_PORT})")
+        else:
+            print(f"Leader connected but gravity compensation unavailable (no URDF)")
+
+    # Build default processors for action and observation
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+    
+    # Build dataset features from robot features and processors
+    # For actions, only include positions (no velocity or torque)
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+    
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+    
+    # Check if dataset already exists
+    dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / HF_EVAL_DATASET_ID
+    if dataset_path.exists():
+        print(f"Evaluation dataset already exists at: {dataset_path}")
+        print("This will append new episodes to the existing dataset.")
+        choice = input("  Continue? (y/n): ").strip().lower()
+        if choice != 'y':
+            print("  Aborting evaluation.")
+            follower.disconnect()
+            if leader:
+                leader.disconnect()
+            return
+    
+    # Create dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_EVAL_DATASET_ID,
+        fps=FPS,
+        features=dataset_features,
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_processes=0,
+        image_writer_threads=12, 
+    )
+    
+    # Load policy config from pretrained model and create policy using factory
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+    policy = make_policy(policy_config, ds_meta=dataset.meta)
+    
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy.config,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        preprocessor_overrides={
+            "device_processor": {"device": str(policy.config.device)}
+        },
+    )
+
+    print(f"\nRunning evaluation...")
+    # Initialize keyboard listener and visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="openarms_evaluation")
+    episode_idx = 0
+    
+    try:
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Evaluating episode {episode_idx + 1} of {NUM_EPISODES}")
+            print(f"\nRunning inference for episode {episode_idx + 1}...")
+            
+            # Run inference with policy
+            record_loop(
+                robot=follower,
+                events=events,
+                fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+                policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
+            
+            # Handle re-recording
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
+            
+            # Save episode
+            if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                print(f"Saving episode {episode_idx + 1} ({dataset.episode_buffer['size']} frames)...")
+                dataset.save_episode()
+                episode_idx += 1
+            
+            # Reset environment between episodes (if not last episode)
+            if not events["stop_recording"] and episode_idx < NUM_EPISODES:
+                if USE_LEADER_FOR_RESETS and leader:
+                    log_say("Reset the environment using leader arms")
+                    print(f"\nManual reset period ({RESET_TIME_SEC}s)...")
+                    
+                    # Use leader for manual reset with gravity compensation
+                    import numpy as np
+                    
+                    dt = 1 / FPS
+                    reset_start_time = time.perf_counter()
+                    
+                    while time.perf_counter() - reset_start_time < RESET_TIME_SEC:
+                        if events["exit_early"] or events["stop_recording"]:
+                            break
+                        
+                        loop_start = time.perf_counter()
+                        
+                        # Get leader state
+                        leader_action = leader.get_action()
+                        
+                        # Extract positions and velocities
+                        leader_positions_deg = {}
+                        leader_velocities_deg_per_sec = {}
+                        
+                        for motor in leader.bus_right.motors:
+                            pos_key = f"right_{motor}.pos"
+                            vel_key = f"right_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+                        
+                        for motor in leader.bus_left.motors:
+                            pos_key = f"left_{motor}.pos"
+                            vel_key = f"left_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+                        
+                        # Calculate gravity and friction torques
+                        leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+                        leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+                        
+                        leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+                        leader_friction_torques_nm = leader._friction_from_velocity(
+                            leader_velocities_rad_per_sec,
+                            friction_scale=1.0
+                        )
+                        
+                        # Combine torques
+                        leader_total_torques_nm = {}
+                        for motor_name in leader_gravity_torques_nm:
+                            gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+                            friction = leader_friction_torques_nm.get(motor_name, 0.0)
+                            leader_total_torques_nm[motor_name] = gravity + friction
+                        
+                        # Apply compensation
+                        for motor in leader.bus_right.motors:
+                            full_name = f"right_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_right._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        for motor in leader.bus_left.motors:
+                            full_name = f"left_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_left._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        # Send leader positions to follower
+                        follower_action = {}
+                        for joint in leader_positions_deg.keys():
+                            pos_key = f"{joint}.pos"
+                            if pos_key in leader_action:
+                                follower_action[pos_key] = leader_action[pos_key]
+                        
+                        if follower_action:
+                            follower.send_action(follower_action)
+                        
+                        # Maintain loop rate
+                        loop_duration = time.perf_counter() - loop_start
+                        sleep_time = dt - loop_duration
+                        if sleep_time > 0:
+                            time.sleep(sleep_time)
+                    
+                    print("Reset complete")
+                else:
+                    log_say("Waiting for manual reset")
+                    print(f"Manually reset the environment and press ENTER to continue")
+                    input("Press ENTER when ready...")
+        
+        print(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+    
+    except KeyboardInterrupt:
+        print("\n\nEvaluation interrupted by user")
+    
+    finally:
+        if leader:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+
+        follower.disconnect()
+        
+        if listener is not None:
+            listener.stop()
+        
+        dataset.finalize()
+        print("\nUploading to Hugging Face Hub...")
+        dataset.push_to_hub(private=True)
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/evaluate_with_rtc.py

@@ -0,0 +1,703 @@
+#!/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.
+
+"""
+OpenArms Policy Evaluation with Real-Time Chunking (RTC)
+
+Evaluates a trained policy on the OpenArms robot using RTC for smooth, continuous motion.
+RTC enables large flow-matching policies (Pi0, Pi0.5, SmolVLA) to produce reactive motion
+despite high inference latency by asynchronously generating action chunks.
+
+Features:
+- Thread-based asynchronous action generation and execution
+- RTC for smooth transitions between action chunks
+- Dataset recording for evaluation episodes
+
+Example usage:
+    python examples/openarms/evaluate_with_rtc.py
+
+    # With custom RTC parameters
+    python examples/openarms/evaluate_with_rtc.py \
+        --rtc.execution_horizon=12 \
+        --rtc.max_guidance_weight=10.0
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from pathlib import Path
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor import make_default_processors
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging, log_say
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+# ============================================================================
+# Default Configuration Constants
+# ============================================================================
+
+DEFAULT_HF_MODEL_ID = "lerobot-data-collection/level1_rac3_100k"
+DEFAULT_HF_EVAL_DATASET_ID = "lerobot-data-collection/test"
+DEFAULT_TASK_DESCRIPTION = "Fold the T-shirt properly"
+
+DEFAULT_NUM_EPISODES = 1
+DEFAULT_FPS = 30
+DEFAULT_EPISODE_TIME_SEC = 1000
+DEFAULT_RESET_TIME_SEC = 60
+
+DEFAULT_FOLLOWER_LEFT_PORT = "can0"
+DEFAULT_FOLLOWER_RIGHT_PORT = "can1"
+
+DEFAULT_CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video0", width=1280, height=720, fps=DEFAULT_FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video4", width=1280, height=720, fps=DEFAULT_FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video2", width=640, height=480, fps=DEFAULT_FPS),
+}
+
+
+# ============================================================================
+# Thread-Safe Robot Wrapper
+# ============================================================================
+
+
+class RobotWrapper:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: OpenArmsFollower):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self.lock:
+            self.robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        with self.lock:
+            return self.robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        with self.lock:
+            return self.robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self.robot.name
+
+
+# ============================================================================
+# Configuration
+# ============================================================================
+
+
+@dataclass
+class OpenArmsRTCEvalConfig(HubMixin):
+    """Configuration for OpenArms evaluation with RTC."""
+
+    policy: PreTrainedConfig | None = None
+
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=20,
+            max_guidance_weight=5.0,
+            prefix_attention_schedule=RTCAttentionSchedule.LINEAR,
+        )
+    )
+
+    model_id: str = DEFAULT_HF_MODEL_ID
+    eval_dataset_id: str = DEFAULT_HF_EVAL_DATASET_ID
+    task: str = DEFAULT_TASK_DESCRIPTION
+
+    num_episodes: int = DEFAULT_NUM_EPISODES
+    fps: float = DEFAULT_FPS
+    episode_time_sec: float = DEFAULT_EPISODE_TIME_SEC
+    reset_time_sec: float = DEFAULT_RESET_TIME_SEC
+
+    follower_left_port: str = DEFAULT_FOLLOWER_LEFT_PORT
+    follower_right_port: str = DEFAULT_FOLLOWER_RIGHT_PORT
+
+    device: str = "cuda"
+
+    # Should be higher than inference_delay + execution_horizon
+    action_queue_size_to_get_new_actions: int = 30
+
+    record_dataset: bool = True
+    push_to_hub: bool = True
+
+    interpolation: bool = True
+
+    use_torch_compile: bool = False
+    torch_compile_backend: str = "inductor"
+    torch_compile_mode: str = "default"
+    torch_compile_disable_cudagraphs: bool = True
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+            self.model_id = policy_path
+        elif self.model_id:
+            self.policy = PreTrainedConfig.from_pretrained(self.model_id)
+            self.policy.pretrained_path = self.model_id
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+# ============================================================================
+# Action Generation Thread
+# ============================================================================
+
+
+def get_actions_thread(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+):
+    """Thread function to asynchronously generate action chunks from the policy."""
+    try:
+        logger.info("[GET_ACTIONS] Starting action generation thread")
+
+        latency_tracker = LatencyTracker()
+        time_per_chunk = 1.0 / cfg.fps
+
+        hw_features = hw_to_dataset_features(robot.observation_features, "observation")
+        policy_device = policy.config.device
+
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                time.sleep(0.01)
+                continue
+
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk) if inference_latency else 0
+
+                obs = robot.get_observation()
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    hw_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]
+                obs_with_policy_features["robot_type"] = robot.name
+
+                preprocessed_obs = preprocessor(obs_with_policy_features)
+
+                actions = policy.predict_action_chunk(
+                    preprocessed_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                original_actions = actions.squeeze(0).clone()
+                postprocessed_actions = postprocessor(actions).squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] action_queue_size_to_get_new_actions too small. "
+                        "Should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+
+                logger.debug(
+                    f"[GET_ACTIONS] Generated chunk, latency={new_latency:.3f}s, "
+                    f"delay={new_delay}, queue_size={action_queue.qsize()}"
+                )
+            else:
+                time.sleep(0.01)
+
+        logger.info("[GET_ACTIONS] Action generation thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Action Execution Thread
+# ============================================================================
+
+
+def actor_thread(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+    dataset: LeRobotDataset | None,
+    dataset_lock: Lock,
+    teleop_action_processor,
+    robot_observation_processor,
+):
+    """Thread function to execute actions on the robot."""
+    try:
+        action_keys = [k for k in robot.action_features.keys() if k.endswith(".pos")]
+
+        if cfg.interpolation:
+            interp_factor = 2
+            robot_interval = 1.0 / (cfg.fps * interp_factor)
+            logger.info(f"[ACTOR] Interpolation ON: policy={cfg.fps}Hz -> robot={cfg.fps * interp_factor}Hz (2x)")
+        else:
+            interp_factor = 1
+            robot_interval = 1.0 / cfg.fps
+            logger.info(f"[ACTOR] Interpolation OFF: policy={cfg.fps}Hz, robot={cfg.fps}Hz")
+
+        prev_action: Tensor | None = None
+        interpolated_actions: list[Tensor] = []
+        interp_idx = 0
+
+        robot_send_count = 0
+        policy_consume_count = 0
+        last_hz_print = time.perf_counter()
+        last_dataset_time = 0.0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                prev_action = None
+                interpolated_actions = []
+                interp_idx = 0
+                robot_send_count = 0
+                policy_consume_count = 0
+                last_hz_print = time.perf_counter()
+                time.sleep(0.01)
+                continue
+
+            start_time = time.perf_counter()
+
+            if interp_idx >= len(interpolated_actions):
+                new_action = action_queue.get()
+                if new_action is not None:
+                    current_action = new_action.cpu()
+                    policy_consume_count += 1
+
+                    if cfg.interpolation and prev_action is not None:
+                        mid = prev_action + 0.5 * (current_action - prev_action)
+                        interpolated_actions = [mid, current_action]
+                    else:
+                        interpolated_actions = [current_action]
+
+                    prev_action = current_action
+                    interp_idx = 0
+
+            if interp_idx < len(interpolated_actions):
+                action_to_send = interpolated_actions[interp_idx]
+                interp_idx += 1
+
+                action_dict = {}
+                for i, key in enumerate(action_keys):
+                    if i < len(action_to_send):
+                        action_dict[key] = action_to_send[i].item()
+
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+                robot_send_count += 1
+
+                if cfg.record_dataset and dataset is not None:
+                    now = time.perf_counter()
+                    if now - last_dataset_time >= (1.0 / cfg.fps):
+                        last_dataset_time = now
+                        with dataset_lock:
+                            obs = robot.get_observation()
+                            obs_processed = robot_observation_processor(obs)
+                            action_for_dataset = teleop_action_processor((action_dict, None))
+                            frame = {}
+                            for key, value in obs_processed.items():
+                                frame[f"observation.{key}"] = value
+                            for key, value in action_for_dataset.items():
+                                frame[f"action.{key}"] = value
+                            frame["task"] = cfg.task
+                            dataset.add_frame(frame)
+
+            now = time.perf_counter()
+            if now - last_hz_print >= 5.0:
+                elapsed = now - last_hz_print
+                actual_robot_hz = robot_send_count / elapsed if elapsed > 0 else 0
+                actual_policy_hz = policy_consume_count / elapsed if elapsed > 0 else 0
+                logger.info(f"[ACTOR] Actual Hz - Robot: {actual_robot_hz:.1f}, Policy: {actual_policy_hz:.1f}")
+                robot_send_count = 0
+                policy_consume_count = 0
+                last_hz_print = now
+
+            dt_s = time.perf_counter() - start_time
+            sleep_time = max(0, robot_interval - dt_s - 0.001)
+            if sleep_time > 0:
+                time.sleep(sleep_time)
+
+        logger.info("[ACTOR] Shutting down")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Main Evaluation Function
+# ============================================================================
+
+
+def _apply_torch_compile(policy, cfg: OpenArmsRTCEvalConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method."""
+    if policy.name in ["pi05", "pi0"]:
+        return policy
+
+    try:
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def main(cfg: OpenArmsRTCEvalConfig):
+    """Main evaluation function with RTC."""
+    init_logging()
+
+    print("=" * 60)
+    print("OpenArms Policy Evaluation with RTC")
+    print("=" * 60)
+    print(f"\nModel: {cfg.model_id}")
+    print(f"Evaluation Dataset: {cfg.eval_dataset_id}")
+    print(f"Task: {cfg.task}")
+    print(f"Episodes: {cfg.num_episodes}")
+    print(f"Episode Duration: {cfg.episode_time_sec}s")
+    print(f"RTC Enabled: {cfg.rtc.enabled}")
+    print(f"RTC Execution Horizon: {cfg.rtc.execution_horizon}")
+    print(f"RTC Max Guidance Weight: {cfg.rtc.max_guidance_weight}")
+    print(f"Policy Hz: {cfg.fps}")
+    print(f"Robot Hz: {cfg.fps * 2 if cfg.interpolation else cfg.fps}")
+    print(f"Interpolation: {cfg.interpolation}")
+    print(f"Device: {cfg.device}")
+    print("=" * 60)
+
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+    episode_active = Event()
+
+    # Initialize Robot
+    follower_config = OpenArmsFollowerConfig(
+        port_left=cfg.follower_left_port,
+        port_right=cfg.follower_right_port,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=DEFAULT_CAMERA_CONFIG,
+    )
+
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    robot = RobotWrapper(follower)
+    logger.info("Follower robot connected")
+
+    # Build Processors and Dataset Features
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+
+    # Create or Load Dataset
+    dataset = None
+    dataset_lock = Lock()
+
+    if cfg.record_dataset:
+        dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / cfg.eval_dataset_id
+        if dataset_path.exists():
+            logger.info(f"Evaluation dataset exists at: {dataset_path}")
+            logger.info("New episodes will be appended.")
+            choice = input("Continue? (y/n): ").strip().lower()
+            if choice != "y":
+                logger.info("Aborting evaluation.")
+                follower.disconnect()
+                return
+
+        dataset = LeRobotDataset.create(
+            repo_id=cfg.eval_dataset_id,
+            fps=int(cfg.fps),
+            features=dataset_features,
+            robot_type=follower.name,
+            use_videos=True,
+            image_writer_processes=0,
+            image_writer_threads=12,
+        )
+        logger.info(f"Dataset created: {cfg.eval_dataset_id}")
+
+    # Load Policy
+    logger.info(f"Loading policy from: {cfg.model_id}")
+
+    policy_class = get_policy_class(cfg.policy.type)
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type in ["pi05", "pi0"]:
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    policy.config.rtc_config = cfg.rtc
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    logger.info(f"Policy loaded: {policy.name}")
+
+    # Create Action Queue and Start Threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    get_actions_t = Thread(
+        target=get_actions_thread,
+        args=(
+            policy,
+            robot,
+            robot_observation_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+        ),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_t.start()
+    logger.info("Started action generation thread")
+
+    actor_t = Thread(
+        target=actor_thread,
+        args=(
+            robot,
+            robot_action_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+            dataset,
+            dataset_lock,
+            teleop_action_processor,
+            robot_observation_processor,
+        ),
+        daemon=True,
+        name="Actor",
+    )
+    actor_t.start()
+    logger.info("Started actor thread")
+
+    # Run Evaluation Episodes
+    episode_idx = 0
+
+    try:
+        while episode_idx < cfg.num_episodes and not shutdown_event.is_set():
+            log_say(f"Evaluating episode {episode_idx + 1} of {cfg.num_episodes}")
+            logger.info(f"\n{'='*40}")
+            logger.info(f"Episode {episode_idx + 1} / {cfg.num_episodes}")
+            logger.info(f"{'='*40}")
+
+            action_queue = ActionQueue(cfg.rtc)
+            episode_active.set()
+            episode_start_time = time.time()
+
+            while (time.time() - episode_start_time) < cfg.episode_time_sec:
+                if shutdown_event.is_set():
+                    break
+
+                elapsed = time.time() - episode_start_time
+                if int(elapsed) % 10 == 0 and int(elapsed) > 0:
+                    logger.info(
+                        f"[MAIN] Episode progress: {elapsed:.0f}/{cfg.episode_time_sec}s, "
+                        f"queue_size={action_queue.qsize()}"
+                    )
+
+                time.sleep(0.5)
+
+            episode_active.clear()
+            logger.info(f"Episode {episode_idx + 1} completed")
+
+            if cfg.record_dataset and dataset is not None:
+                with dataset_lock:
+                    if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                        logger.info(
+                            f"Saving episode {episode_idx + 1} "
+                            f"({dataset.episode_buffer['size']} frames)"
+                        )
+                        dataset.save_episode()
+
+            episode_idx += 1
+
+            # Manual reset between episodes
+            if not shutdown_event.is_set() and episode_idx < cfg.num_episodes:
+                log_say("Waiting for manual reset")
+                logger.info("Manually reset the environment and press ENTER to continue")
+                input("Press ENTER when ready...")
+
+        logger.info(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+
+    except KeyboardInterrupt:
+        logger.info("\n\nEvaluation interrupted by user")
+
+    finally:
+        shutdown_event.set()
+        episode_active.clear()
+
+        if get_actions_t.is_alive():
+            logger.info("Waiting for action generation thread to finish...")
+            get_actions_t.join(timeout=5.0)
+
+        if actor_t.is_alive():
+            logger.info("Waiting for actor thread to finish...")
+            actor_t.join(timeout=5.0)
+
+        follower.disconnect()
+        logger.info("Follower disconnected")
+
+        if cfg.record_dataset and dataset is not None:
+            dataset.finalize()
+            if cfg.push_to_hub:
+                logger.info("Uploading to Hugging Face Hub...")
+                dataset.push_to_hub(private=True)
+
+        logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    main()

examples/openarms/friction_compensation.py

@@ -0,0 +1,216 @@
+import time
+import numpy as np
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+
+# Friction model parameters from OpenArms config/follower.yaml
+# τ_fric(ω) = Fo + Fv·ω + Fc·tanh(k·ω)
+# For 8 motors: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
+FRICTION_PARAMS = {
+    "Fc": [0.306, 0.306, 0.40, 0.166, 0.050, 0.093, 0.172, 0.0512],  # Coulomb friction [Nm]
+    "k": [28.417, 28.417, 29.065, 130.038, 151.771, 242.287, 7.888, 4.000],  # tanh steepness
+    "Fv": [0.063, 0.0630, 0.604, 0.813, 0.029, 0.072, 0.084, 0.084],  # Viscous friction [Nm·s/rad]
+    "Fo": [0.088, 0.088, 0.008, -0.058, 0.005, 0.009, -0.059, -0.050],  # Offset torque [Nm]
+}
+
+# Constants from OpenArms C++ implementation
+AMP_TMP = 1.0
+COEF_TMP = 0.1
+
+FRICTION_SCALE = 1.0  # OpenArms C++ uses 0.3 factor in unilateral mode
+DAMPING_KD = [0.5, 0.5, 0.5, 0.5, 0.1, 0.1, 0.1, 0.1]  # Damping gains for stability
+
+def compute_friction_torque(velocity_rad_per_sec: float, motor_index: int) -> float:
+    """
+    Compute friction torque for a single motor using the tanh friction model.
+    
+    Args:
+        velocity_rad_per_sec: Angular velocity in rad/s
+        motor_index: Index of the motor (0-7)
+        
+    Returns:
+        Friction torque in N·m (scaled for stability)
+    """
+    
+    Fc = FRICTION_PARAMS["Fc"][motor_index]
+    k = FRICTION_PARAMS["k"][motor_index]
+    Fv = FRICTION_PARAMS["Fv"][motor_index]
+    Fo = FRICTION_PARAMS["Fo"][motor_index]
+    
+    # Friction model: τ_fric = amp * Fc * tanh(coef * k * ω) + Fv * ω + Fo
+    friction_torque = (
+        AMP_TMP * Fc * np.tanh(COEF_TMP * k * velocity_rad_per_sec) +
+        Fv * velocity_rad_per_sec +
+        Fo
+    )
+    
+    # Scale down friction compensation for stability at lower control rates
+    # (OpenArms C++ uses 0.3 factor in unilateral mode)!!
+    friction_torque *= FRICTION_SCALE
+    
+    return friction_torque
+
+
+def main() -> None:
+    config = OpenArmsFollowerConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=5.0,
+    )
+    
+    print("Initializing robot...")
+    follower = OpenArmsFollower(config)
+    follower.connect(calibrate=True)
+    
+    print(f"Applying friction compensation")
+    print("  1. Support the arm before starting")
+    print("  2. The arm will be held in place by friction compensation")
+    print("  3. You should be able to move it with gentle force")
+    print("\nPress ENTER when ready to start...")
+    input()
+    
+    print(f"✓ Motors enabled")
+    print("\nStarting friction compensation loop...")
+    print("Press Ctrl+C to stop\n")
+    
+    loop_times = []
+    last_print_time = time.perf_counter()
+    
+    # Motor name to index mapping
+    motor_name_to_index = {
+        "joint_1": 0,
+        "joint_2": 1,
+        "joint_3": 2,
+        "joint_4": 3,
+        "joint_5": 4,
+        "joint_6": 5,
+        "joint_7": 6,
+        "gripper": 7,
+    }
+    
+    try:
+        while True:
+            loop_start = time.perf_counter()
+            
+            # Get current joint positions and velocities from robot
+            obs = follower.get_observation()
+            
+            # Extract velocities in degrees per second
+            velocities_deg_per_sec = {}
+            positions_deg = {}
+            
+            for motor in follower.bus_right.motors:
+                vel_key = f"right_{motor}.vel"
+                pos_key = f"right_{motor}.pos"
+                if vel_key in obs:
+                    velocities_deg_per_sec[f"right_{motor}"] = obs[vel_key]
+                if pos_key in obs:
+                    positions_deg[f"right_{motor}"] = obs[pos_key]
+            
+            for motor in follower.bus_left.motors:
+                vel_key = f"left_{motor}.vel"
+                pos_key = f"left_{motor}.pos"
+                if vel_key in obs:
+                    velocities_deg_per_sec[f"left_{motor}"] = obs[vel_key]
+                if pos_key in obs:
+                    positions_deg[f"left_{motor}"] = obs[pos_key]
+            
+            # Convert velocities to rad/s and compute friction torques
+            friction_torques_nm = {}
+            for motor_full_name, velocity_deg_per_sec in velocities_deg_per_sec.items():
+                # Extract motor name without arm prefix
+                if motor_full_name.startswith("right_"):
+                    motor_name = motor_full_name.removeprefix("right_")
+                elif motor_full_name.startswith("left_"):
+                    motor_name = motor_full_name.removeprefix("left_")
+                else:
+                    continue
+                
+                # Get motor index for friction parameters
+                motor_index = motor_name_to_index.get(motor_name, 0)
+                
+                # Convert velocity to rad/s
+                velocity_rad_per_sec = np.deg2rad(velocity_deg_per_sec)
+                
+                # Compute friction torque
+                friction_torque = compute_friction_torque(velocity_rad_per_sec, motor_index)
+                friction_torques_nm[motor_full_name] = friction_torque
+            
+            # Apply friction compensation to right arm (all joints INCLUDING gripper)
+            for motor in follower.bus_right.motors:
+                full_name = f"right_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = friction_torques_nm.get(full_name, 0.0)
+                
+                # Get motor index for damping gain
+                motor_index = motor_name_to_index.get(motor, 0)
+                kd = DAMPING_KD[motor_index]
+                
+                # Send MIT control command with friction compensation + damping
+                follower.bus_right._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=kd,   # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+
+            # Apply friction compensation to left arm (all joints INCLUDING gripper)
+            for motor in follower.bus_left.motors:
+                full_name = f"left_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = friction_torques_nm.get(full_name, 0.0)
+                
+                # Get motor index for damping gain
+                motor_index = motor_name_to_index.get(motor, 0)
+                kd = DAMPING_KD[motor_index]
+                
+                # Send MIT control command with friction compensation + damping
+                follower.bus_left._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=kd,   # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+            
+            # Measure loop time
+            loop_end = time.perf_counter()
+            loop_time = loop_end - loop_start
+            loop_times.append(loop_time)
+            
+            # Print status every 2 seconds
+            if loop_end - last_print_time >= 2.0:
+                if loop_times:
+                    avg_time = sum(loop_times) / len(loop_times)
+                    current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                    
+                    print(f"{current_hz:.1f} Hz")
+
+                    loop_times = []
+                    last_print_time = loop_end
+            
+            time.sleep(0.001)
+            
+    except KeyboardInterrupt:
+        print("\n\nStopping friction compensation...")
+    
+    finally:
+        print("\nDisabling all motors and disconnecting...")
+        follower.bus_right.disable_torque()
+        follower.bus_left.disable_torque()
+        time.sleep(0.1)
+        follower.disconnect()
+        print("✓ Safe shutdown complete")
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/gravity_compensation.py

@@ -0,0 +1,142 @@
+import time
+import numpy as np
+import pinocchio as pin
+from os.path import join, dirname, exists, expanduser
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+
+def main() -> None:
+    config = OpenArmsFollowerConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=5.0,
+    )
+    
+    
+    print("Initializing robot...")
+    follower = OpenArmsFollower(config)
+    follower.connect(calibrate=True)
+    
+    # Load URDF for Pinocchio dynamics
+    urdf_path = "/home/croissant/Documents/openarm_description/openarm_bimanual_pybullet.urdf"
+    
+    pin_robot = pin.RobotWrapper.BuildFromURDF(urdf_path, dirname(urdf_path))
+    pin_robot.data = pin_robot.model.createData()
+    print(f"✓ Loaded Pinocchio model with {pin_robot.nq} DoFs")
+    
+    follower.pin_robot = pin_robot
+    
+    print(f"Applying gravity compensation")
+    print("  1. Support the arm before starting")
+    print("  2. The arm will be held in place by gravity compensation")
+    print("  3. You should be able to move it with gentle force")
+    print("\nPress ENTER when ready to start...")
+    input()
+    
+    print(f"✓ Motors enabled")
+    print("\nStarting gravity compensation loop...")
+    print("Press Ctrl+C to stop\n")
+    
+    loop_times = []
+    last_print_time = time.perf_counter()
+    
+    try:
+        while True:
+            loop_start = time.perf_counter()
+            
+            # Get current joint positions from robot
+            obs = follower.get_observation()
+            
+            # Extract positions in degrees
+            positions_deg = {}
+            for motor in follower.bus_right.motors:
+                key = f"right_{motor}.pos"
+                if key in obs:
+                    positions_deg[f"right_{motor}"] = obs[key]
+            
+            for motor in follower.bus_left.motors:
+                key = f"left_{motor}.pos"
+                if key in obs:
+                    positions_deg[f"left_{motor}"] = obs[key]
+            
+            # Convert to radians and calculate gravity torques
+            # Use the built-in method from OpenArmsFollower
+            positions_rad = {k: np.deg2rad(v) for k, v in positions_deg.items()}
+            torques_nm = follower._gravity_from_q(positions_rad)
+            
+            # Apply gravity compensation to right arm (all joints except gripper)
+            for motor in follower.bus_right.motors:
+                if motor == "gripper":
+                    continue  # Skip gripper
+                    
+                full_name = f"right_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = torques_nm.get(full_name, 0.0)
+                
+                # Send MIT control command with gravity compensation torque
+                follower.bus_right._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=0.0,  # No velocity damping
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+
+            # Apply gravity compensation to left arm (all joints except gripper)
+            for motor in follower.bus_left.motors:
+                if motor == "gripper":
+                    continue  # Skip gripper
+                    
+                full_name = f"left_{motor}"
+                position = positions_deg.get(full_name, 0.0)
+                torque = torques_nm.get(full_name, 0.0)
+                
+                # Send MIT control command with gravity compensation torque
+                follower.bus_left._mit_control(
+                    motor=motor,
+                    kp=0.0,  # No position control
+                    kd=0.0,  # No velocity damping
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque
+                )
+            
+            # Measure loop time
+            loop_end = time.perf_counter()
+            loop_time = loop_end - loop_start
+            loop_times.append(loop_time)
+            
+            # Print status every 2 seconds
+            if loop_end - last_print_time >= 2.0:
+                if loop_times:
+                    avg_time = sum(loop_times) / len(loop_times)
+                    current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                    
+                    print(f"{current_hz:.1f} Hz ({avg_time*1000:.1f} ms)")
+
+                    loop_times = []
+                    last_print_time = loop_end
+            
+            time.sleep(0.005)
+            
+    except KeyboardInterrupt:
+        print("\n\nStopping gravity compensation...")
+    
+    finally:
+        print("\nDisabling all motors and disconnecting...")
+        follower.bus_right.disable_torque()
+        follower.bus_left.disable_torque()
+        time.sleep(0.1)
+        follower.disconnect()
+        print("✓ Safe shutdown complete")
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/record_with_compensation.py

@@ -0,0 +1,395 @@
+"""
+OpenArms Dataset Recording with Gravity + Friction Compensation
+
+Records a dataset using OpenArms follower robot with leader teleoperator.
+Leader arms have gravity and friction compensation for weightless, easy movement.
+Includes 3 cameras: left wrist, right wrist, and base camera.
+
+Uses the same compensation approach as teleop_with_compensation.py
+"""
+
+import shutil
+import time
+from pathlib import Path
+
+import numpy as np
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+# Recording parameters
+NUM_EPISODES = 1
+FPS = 30
+EPISODE_TIME_SEC = 600
+RESET_TIME_SEC = 120
+TASK_DESCRIPTION = "OpenArms task description"
+
+# Friction compensation scale factor (1.0 = full, 0.3 = 30% for stability)
+FRICTION_SCALE = 1.0
+
+def record_loop_with_compensation(
+    robot,
+    leader,
+    events,
+    fps,
+    dataset,
+    dataset_features,
+    control_time_s,
+    single_task,
+    display_data=True,
+):
+    """
+    Custom record loop that applies gravity + friction compensation to leader.
+    Based on record_loop but with integrated compensation.
+    """
+    dt = 1 / fps
+    episode_start_time = time.perf_counter()
+    
+    # All joints (both arms)
+    all_joints = []
+    for motor in leader.bus_right.motors:
+        all_joints.append(f"right_{motor}")
+    for motor in leader.bus_left.motors:
+        all_joints.append(f"left_{motor}")
+    
+    while True:
+        loop_start = time.perf_counter()
+        elapsed = loop_start - episode_start_time
+        
+        # Check if we should exit
+        if elapsed >= control_time_s or events["exit_early"] or events["stop_recording"]:
+            break
+        
+        # Get leader state
+        leader_action = leader.get_action()
+        
+        # Extract positions and velocities in degrees
+        leader_positions_deg = {}
+        leader_velocities_deg_per_sec = {}
+        
+        for motor in leader.bus_right.motors:
+            pos_key = f"right_{motor}.pos"
+            vel_key = f"right_{motor}.vel"
+            if pos_key in leader_action:
+                leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+            if vel_key in leader_action:
+                leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+        
+        for motor in leader.bus_left.motors:
+            pos_key = f"left_{motor}.pos"
+            vel_key = f"left_{motor}.vel"
+            if pos_key in leader_action:
+                leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+            if vel_key in leader_action:
+                leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+        
+        # Calculate gravity torques for leader using built-in method
+        leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+        leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+        
+        # Calculate friction torques for leader using built-in method
+        leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+        leader_friction_torques_nm = leader._friction_from_velocity(
+            leader_velocities_rad_per_sec,
+            friction_scale=FRICTION_SCALE
+        )
+        
+        # Combine gravity + friction torques
+        leader_total_torques_nm = {}
+        for motor_name in leader_gravity_torques_nm:
+            gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+            friction = leader_friction_torques_nm.get(motor_name, 0.0)
+            leader_total_torques_nm[motor_name] = gravity + friction
+        
+        # Apply gravity + friction compensation to leader RIGHT arm (all joints including gripper)
+        for motor in leader.bus_right.motors:
+            full_name = f"right_{motor}"
+            position = leader_positions_deg.get(full_name, 0.0)
+            torque = leader_total_torques_nm.get(full_name, 0.0)
+            
+            # Get damping gain for stability
+            kd = leader.get_damping_kd(motor)
+            
+            leader.bus_right._mit_control(
+                motor=motor,
+                kp=0.0,
+                kd=kd,  # Add damping for stability
+                position_degrees=position,
+                velocity_deg_per_sec=0.0,
+                torque=torque,
+            )
+        
+        # Apply gravity + friction compensation to leader LEFT arm (all joints including gripper)
+        for motor in leader.bus_left.motors:
+            full_name = f"left_{motor}"
+            position = leader_positions_deg.get(full_name, 0.0)
+            torque = leader_total_torques_nm.get(full_name, 0.0)
+            
+            # Get damping gain for stability
+            kd = leader.get_damping_kd(motor)
+            
+            leader.bus_left._mit_control(
+                motor=motor,
+                kp=0.0,
+                kd=kd,  # Add damping for stability
+                position_degrees=position,
+                velocity_deg_per_sec=0.0,
+                torque=torque,
+            )
+        
+        # Send leader positions to follower (both arms)
+        follower_action = {}
+        for joint in all_joints:
+            pos_key = f"{joint}.pos"
+            if pos_key in leader_action:
+                follower_action[pos_key] = leader_action[pos_key]
+        
+        # Send action to robot
+        if follower_action:
+            robot.send_action(follower_action)
+        
+        # Get observation from robot (includes camera images)
+        observation = robot.get_observation()
+        
+        # Add to dataset if we have a dataset
+        if dataset is not None:
+            # Build properly formatted observation frame
+            obs_frame = build_dataset_frame(dataset_features, observation, prefix="observation")
+            
+            # Build properly formatted action frame (keep .pos suffix - it matches the feature names)
+            action_frame = build_dataset_frame(dataset_features, follower_action, prefix="action")
+            
+            # Combine into single frame
+            frame = {**obs_frame, **action_frame}
+            
+            # Add metadata (task is required, timestamp will be auto-calculated by add_frame)
+            frame["task"] = single_task
+            
+            dataset.add_frame(frame)
+        
+        # Display data if requested
+        if display_data:
+            log_rerun_data(observation=observation, action=follower_action)
+        
+        # Maintain loop rate
+        loop_duration = time.perf_counter() - loop_start
+        sleep_time = dt - loop_duration
+        if sleep_time > 0:
+            time.sleep(sleep_time)
+
+
+def main():
+    """Main recording loop with gravity compensation."""
+    
+    print("=" * 70)
+    print("OpenArms Dataset Recording with Compensation")
+    print("=" * 70)
+    
+    # Create camera configurations (3 cameras: left wrist, right wrist, base)
+    # Using actual device paths found by lerobot-find-cameras opencv
+    camera_config = {
+        "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video0", width=640, height=480, fps=FPS),
+        "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video1", width=640, height=480, fps=FPS),
+        "base": OpenCVCameraConfig(index_or_path="/dev/video7", width=640, height=480, fps=FPS),
+    }
+    
+    # Configure follower robot with cameras
+    follower_config = OpenArmsFollowerConfig(
+        port_left="can2",
+        port_right="can3",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=camera_config,
+    )
+    
+    # Configure leader teleoperator (no cameras needed)
+    leader_config = OpenArmsLeaderConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_leader",
+        manual_control=False,  # Enable torque control for gravity compensation
+    )
+    
+    # Initialize robot and teleoperator
+    print("\nInitializing devices...")
+    follower = OpenArmsFollower(follower_config)
+    leader = OpenArmsLeader(leader_config)
+    
+    # Connect devices
+    print("Connecting and calibrating...")
+    follower.connect(calibrate=True)
+    leader.connect(calibrate=True)
+    
+    # Verify URDF is loaded for gravity compensation
+    if leader.pin_robot is None:
+        raise RuntimeError("URDF model not loaded on leader. Gravity compensation not available.")
+    
+    # Configure the dataset features
+    # For actions, we only want to record positions (not velocity or torque)
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+    
+    action_features = hw_to_dataset_features(action_features_hw, "action")
+    obs_features = hw_to_dataset_features(follower.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
+    
+    # Create the dataset
+    print("\nCreating dataset...")
+    repo_id = "<hf_username>/<dataset_repo_id>"  # TODO: Replace with your Hugging Face repo
+    
+    # Check if dataset already exists and prompt user
+    dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / repo_id
+    while dataset_path.exists():
+        print(f"\nDataset already exists at: {dataset_path}")
+        print("\nOptions:")
+        print("  1. Overwrite existing dataset")
+        print("  2. Use a different name")
+        print("  3. Abort")
+        
+        choice = input("\nEnter your choice (1/2/3): ").strip()
+        
+        if choice == '1':
+            print(f"Removing existing dataset...")
+            shutil.rmtree(dataset_path)
+            print("✓ Existing dataset removed")
+            break
+        elif choice == '2':
+            print("\nCurrent repo_id:", repo_id)
+            new_repo_id = input("Enter new repo_id (format: <username>/<dataset_name>): ").strip()
+            if new_repo_id and '/' in new_repo_id:
+                repo_id = new_repo_id
+                dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / repo_id
+                print(f"✓ Using new repo_id: {repo_id}")
+                # Loop will continue if this new path also exists
+            else:
+                print("Invalid repo_id format. Please use format: <username>/<dataset_name>")
+        elif choice == '3':
+            print("Aborting. Please remove the existing dataset manually or restart with a different repo_id.")
+            follower.disconnect()
+            leader.disconnect()
+            return
+        else:
+            print("Invalid choice. Please enter 1, 2, or 3.")
+    
+    dataset = LeRobotDataset.create(
+        repo_id=repo_id,
+        fps=FPS,
+        features=dataset_features,
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+    
+    # Initialize keyboard listener and visualization
+    _, events = init_keyboard_listener()
+    init_rerun(session_name="openarms_recording")
+    
+    # Enable motors on both leader arms for gravity compensation
+    leader.bus_right.enable_torque()
+    leader.bus_left.enable_torque()
+    time.sleep(0.1)
+    
+    print("\n" + "=" * 70)
+    print(f"Recording {NUM_EPISODES} episodes")
+    print(f"Task: {TASK_DESCRIPTION}")
+    print("=" * 70)
+    print("\nLeader BOTH arms: Gravity + Friction comp | Follower BOTH arms: Teleop")
+    print("\nKeyboard controls:")
+    print("  - Press 'q' to stop recording")
+    print("  - Press 'r' to re-record current episode")
+    print("=" * 70)
+    
+    episode_idx = 0
+    
+    try:
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+            
+            # Record episode with compensation active
+            record_loop_with_compensation(
+                robot=follower,
+                leader=leader,
+                events=events,
+                fps=FPS,
+                dataset=dataset,
+                dataset_features=dataset_features,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
+            
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+                log_say("Reset the environment")
+                record_loop_with_compensation(
+                    robot=follower,
+                    leader=leader,
+                    events=events,
+                    fps=FPS,
+                    dataset=None,  # Don't save reset period
+                    dataset_features=dataset_features,
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                )
+            
+            # Handle re-recording
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
+            
+            # Only save episode if frames were recorded
+            if dataset.episode_buffer is not None and dataset.episode_buffer["size"] > 0:
+                dataset.save_episode()
+                episode_idx += 1
+            else:
+                log_say("No frames recorded, skipping episode save")
+                # Clear the empty buffer
+                dataset.episode_buffer = None
+    
+    except KeyboardInterrupt:
+        print("\n\nStopping recording...")
+    
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        try:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+            follower.disconnect()
+            print("✓ Shutdown complete")
+        except Exception as e:
+            print(f"Shutdown error: {e}")
+        
+        # Upload dataset
+        print("\nUploading dataset to Hugging Face Hub...")
+        try:
+            dataset.push_to_hub()
+            print("✓ Dataset uploaded successfully")
+        except Exception as e:
+            print(f"Warning: Failed to upload dataset: {e}")
+            print("You can manually upload later using: dataset.push_to_hub()")
+        
+        print("✓ Recording complete!")
+
+
+if __name__ == "__main__":
+    main()

examples/openarms/replay.py

@@ -0,0 +1,166 @@
+#!/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.
+
+"""
+OpenArms Dataset Replay Example
+
+Replays position actions from a recorded dataset on an OpenArms follower robot.
+Only position commands (ending with .pos) are replayed, not velocity or torque.
+
+Example usage:
+    python examples/openarms/replay.py
+"""
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.constants import ACTION
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+# Configuration
+EPISODE_IDX = 0
+DATASET_REPO_ID = "lerobot-data-collection/replay-this-2025-11-02-17-58"  # TODO: Replace with your dataset
+DATASET_ROOT = None  # Use default cache location, or specify custom path
+
+# Robot configuration - adjust these to match your setup
+ROBOT_CONFIG = OpenArmsFollowerConfig(
+    port_left="can2",      # CAN interface for left arm
+    port_right="can3",     # CAN interface for right arm
+    can_interface="socketcan", 
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=10.0,  # Safety limit: max degrees to move per step
+)
+
+
+def main():
+    """Main replay function."""
+    print("=" * 70)
+    print("OpenArms Dataset Replay")
+    print("=" * 70)
+    print(f"\nDataset: {DATASET_REPO_ID}")
+    print(f"Episode: {EPISODE_IDX}")
+    print(f"Robot: {ROBOT_CONFIG.id}")
+    print(f"  Left arm: {ROBOT_CONFIG.port_left}")
+    print(f"  Right arm: {ROBOT_CONFIG.port_right}")
+    print("\n" + "=" * 70)
+    
+    # Initialize the robot
+    print("\n[1/3] Initializing robot...")
+    robot = OpenArmsFollower(ROBOT_CONFIG)
+    
+    # Load the dataset
+    print(f"\n[2/3] Loading dataset '{DATASET_REPO_ID}'...")
+    dataset = LeRobotDataset(
+        DATASET_REPO_ID,
+        root=DATASET_ROOT,
+        episodes=[EPISODE_IDX]
+    )
+    
+    # Filter dataset to only include frames from the specified episode
+    # (required for dataset V3.0 where episodes are chunked)
+    episode_frames = dataset.hf_dataset.filter(
+        lambda x: x["episode_index"] == EPISODE_IDX
+    )
+    
+    if len(episode_frames) == 0:
+        raise ValueError(
+            f"No frames found for episode {EPISODE_IDX} in dataset {DATASET_REPO_ID}"
+        )
+    
+    print(f"  Found {len(episode_frames)} frames in episode {EPISODE_IDX}")
+    
+    # Extract action features from dataset
+    action_features = dataset.features.get(ACTION, {})
+    action_names = action_features.get("names", [])
+    
+    # Filter to only position actions (ending with .pos)
+    position_action_names = [name for name in action_names if name.endswith(".pos")]
+    
+    if not position_action_names:
+        raise ValueError(
+            f"No position actions found in dataset. Action names: {action_names}"
+        )
+    
+    print(f"  Found {len(position_action_names)} position actions to replay")
+    print(f"  Actions: {', '.join(position_action_names[:5])}{'...' if len(position_action_names) > 5 else ''}")
+    
+    # Select only action columns from dataset
+    actions = episode_frames.select_columns(ACTION)
+    
+    # Connect to the robot
+    print(f"\n[3/3] Connecting to robot...")
+    robot.connect(calibrate=False)  # Skip calibration for replay
+    
+    if not robot.is_connected:
+        raise RuntimeError("Robot failed to connect!")
+    
+    print("\n" + "=" * 70)
+    print("Ready to replay!")
+    print("=" * 70)
+    print("\nThe robot will replay the recorded positions.")
+    print("Press Ctrl+C to stop at any time.\n")
+    
+    input("Press ENTER to start replaying...")
+    
+    # Replay loop
+    log_say(f"Replaying episode {EPISODE_IDX}", blocking=True)
+    
+    try:
+        for idx in range(len(episode_frames)):
+            loop_start = time.perf_counter()
+            
+            # Extract action array from dataset
+            action_array = actions[idx][ACTION]
+            
+            # Build action dictionary, but only include position actions
+            action = {}
+            for i, name in enumerate(action_names):
+                # Only include position actions (ending with .pos)
+                if name.endswith(".pos"):
+                    action[name] = float(action_array[i])
+            
+            # Send action to robot
+            robot.send_action(action)
+            
+            # Maintain replay rate (use dataset fps)
+            loop_duration = time.perf_counter() - loop_start
+            dt_s = 1.0 / dataset.fps - loop_duration
+            busy_wait(dt_s)
+            
+            # Progress indicator every 100 frames
+            if (idx + 1) % 100 == 0:
+                progress = (idx + 1) / len(episode_frames) * 100
+                print(f"Progress: {idx + 1}/{len(episode_frames)} frames ({progress:.1f}%)")
+        
+        print(f"\n✓ Successfully replayed {len(episode_frames)} frames")
+        log_say("Replay complete", blocking=True)
+        
+    except KeyboardInterrupt:
+        print("\n\nReplay interrupted by user")
+    finally:
+        # Disconnect robot
+        print("\nDisconnecting robot...")
+        robot.disconnect()
+        print("✓ Replay complete!")
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/setup_can.sh

@@ -0,0 +1,73 @@
+#!/bin/bash
+# Setup all OpenArms CAN interfaces with CAN FD
+
+set -e
+
+echo "=========================================="
+echo "OpenArms CAN FD Interface Setup"
+echo "=========================================="
+echo ""
+echo "Mode: CAN FD"
+echo "  - Nominal bitrate: 1 Mbps"
+echo "  - Data bitrate: 5 Mbps"
+echo ""
+echo "Configuring interfaces can0, can1, can2, can3..."
+echo ""
+
+# Configure each CAN interface with CAN FD
+for i in 0 1 2 3; do
+    interface="can$i"
+    
+    # Check if interface exists
+    if ! ip link show "$interface" &> /dev/null; then
+        echo "⚠ $interface: Not found, skipping"
+        continue
+    fi
+    
+    # Bring down interface
+    sudo ip link set "$interface" down 2>/dev/null
+    
+    # Configure CAN FD mode
+    sudo ip link set "$interface" type can \
+        bitrate 1000000 \
+        dbitrate 5000000 \
+        fd on
+    
+    # Bring up interface
+    sudo ip link set "$interface" up
+    
+    # Verify configuration
+    if ip link show "$interface" | grep -q "UP"; then
+        echo "✓ $interface: Configured and UP"
+    else
+        echo "✗ $interface: Failed to bring UP"
+    fi
+done
+
+echo ""
+echo "=========================================="
+echo "Verification"
+echo "=========================================="
+echo ""
+
+# Show detailed status for each interface
+for i in 0 1 2 3; do
+    interface="can$i"
+    if ip link show "$interface" &> /dev/null; then
+        echo "$interface:"
+        # Show key parameters
+        ip -d link show "$interface" | grep -E "can|state|bitrate|dbitrate" | head -3
+        echo ""
+    fi
+done
+
+echo "=========================================="
+echo "Setup Complete!"
+echo "=========================================="
+echo ""
+echo "All interfaces configured for CAN FD mode"
+echo ""
+echo "Next steps:"
+echo "  1. Test motors: python debug_can_communication.py"
+echo "  2. Run teleoperation: python examples/openarms/teleop.py"
+echo ""

examples/openarms/teleop.py

@@ -0,0 +1,148 @@
+"""
+OpenArms Teleoperation Example - Full Dual Arms
+
+This script demonstrates teleoperation of OpenArms follower robot using an OpenArms leader arm.
+It first calibrates both devices, then enters a teleoperation loop for both arms.
+"""
+
+import time
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+
+
+follower_config = OpenArmsFollowerConfig(
+    port_left="can2",   # CAN interface for follower left arm
+    port_right="can3",  # CAN interface for follower right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=5.0,  # Safety limit
+)
+
+
+leader_config = OpenArmsLeaderConfig(
+    port_left="can0",   # CAN interface for leader left arm
+    port_right="can1",  # CAN interface for leader right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_leader",
+    manual_control=True,  # Enable manual control (torque disabled)
+)
+
+print("=" * 60)
+print("OpenArms Teleoperation - Full Dual Arms")
+print("=" * 60)
+
+# Initialize devices
+print("\n[1/4] Initializing devices...")
+follower = OpenArmsFollower(follower_config)
+leader = OpenArmsLeader(leader_config)
+
+# Connect and calibrate follower
+print("\n[2/4] Connecting and calibrating follower robot...")
+print("Note: If you have existing calibration, just press ENTER to use it.")
+follower.connect(calibrate=True)
+
+# Connect and calibrate leader
+print("\n[3/4] Connecting and calibrating leader arm...")
+print("Note: The leader arm will have torque disabled for manual control.")
+leader.connect(calibrate=True)
+
+# Wait for user to be ready
+print("\n[4/4] Ready for teleoperation!")
+print("\nBoth arms will be controlled (16 motors total):")
+print("  RIGHT ARM: joints 1-7 + gripper")
+print("  LEFT ARM: joints 1-7 + gripper")
+
+print("\nPress ENTER to start teleoperation...")
+input()
+
+print("\nTeleoperation started! Move both leader arms.")
+print("Press Ctrl+C to stop.\n")
+
+# All joints for both arms (16 motors total)
+all_joints = [
+    # Right arm
+    "right_joint_1",
+    "right_joint_2",
+    "right_joint_3",
+    "right_joint_4",
+    "right_joint_5",
+    "right_joint_6",
+    "right_joint_7",
+    "right_gripper",
+    # Left arm
+    "left_joint_1",
+    "left_joint_2",
+    "left_joint_3",
+    "left_joint_4",
+    "left_joint_5",
+    "left_joint_6",
+    "left_joint_7",
+    "left_gripper",
+]
+
+# Performance monitoring
+loop_times = []
+start_time = time.perf_counter()
+last_print_time = start_time
+
+try:
+    while True:
+        loop_start = time.perf_counter()
+        
+        # Get action from leader
+        leader_action = leader.get_action()
+        
+        # Filter to only position data for all joints (both arms)
+        joint_action = {}
+        for joint in all_joints:
+            pos_key = f"{joint}.pos"
+            if pos_key in leader_action:
+                joint_action[pos_key] = leader_action[pos_key]
+        
+        # Send action to follower (both arms)
+        if joint_action:
+            follower.send_action(joint_action)
+        
+        # Measure loop time
+        loop_end = time.perf_counter()
+        loop_time = loop_end - loop_start
+        loop_times.append(loop_time)
+        
+        # Print stats every 2 seconds
+        if loop_end - last_print_time >= 2.0:
+            if loop_times:
+                avg_time = sum(loop_times) / len(loop_times)
+                current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                min_time = min(loop_times)
+                max_time = max(loop_times)
+                max_hz = 1.0 / min_time if min_time > 0 else 0
+                min_hz = 1.0 / max_time if max_time > 0 else 0
+                
+                print(f"[Hz Stats] Avg: {current_hz:.1f} Hz | "
+                      f"Range: {min_hz:.1f}-{max_hz:.1f} Hz | "
+                      f"Avg loop time: {avg_time*1000:.1f} ms")
+                
+                # Reset for next measurement window
+                loop_times = []
+                last_print_time = loop_end
+            
+except KeyboardInterrupt:
+    print("\n\nStopping teleoperation...")
+finally:
+    # Disconnect devices
+    print("Disconnecting devices...")
+    try:
+        follower.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting follower: {e}")
+    
+    try:
+        leader.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting leader: {e}")
+    
+    print("Done!")

examples/openarms/teleop_openarms_mini.py

@@ -0,0 +1,197 @@
+"""
+OpenArms Mini Teleoperation Example
+
+This script demonstrates teleoperation of an OpenArms follower robot using 
+an OpenArms Mini leader (Feetech-based) with dual arms (16 motors total).
+
+The OpenArms Mini has:
+- Right arm: 8 motors (joint_1 to joint_7 + gripper)
+- Left arm: 8 motors (joint_1 to joint_7 + gripper)
+
+Note on gripper normalization:
+- OpenArms Mini gripper: 0-100 scale (0=closed, 100=open)
+- OpenArms follower gripper: degrees (0=closed, -65=open)
+- This script automatically converts between the two ranges
+"""
+
+import time
+import os
+import sys
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.teleoperators.openarms_mini.openarms_mini import OpenArmsMini
+from lerobot.teleoperators.openarms_mini.config_openarms_mini import OpenArmsMiniConfig
+from lerobot.utils.robot_utils import busy_wait
+
+# Target control frequency
+TARGET_FPS = 30
+
+# Configure the OpenArms follower (Damiao motors on CAN bus)
+follower_config = OpenArmsFollowerConfig(
+    port_left="can0",   # CAN interface for follower left arm
+    port_right="can1",  # CAN interface for follower right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=10.0,  # Safety limit (degrees per step)
+)
+
+# Configure the OpenArms Mini leader (Feetech motors on serial)
+leader_config = OpenArmsMiniConfig(
+    port_right="/dev/ttyACM0",  # Serial port for right arm
+    port_left="/dev/ttyACM1",   # Serial port for left arm
+    id="openarms_mini",
+    use_degrees=True,
+)
+
+print("OpenArms Mini → OpenArms Follower Teleoperation")
+
+# Initialize devices
+follower = OpenArmsFollower(follower_config)
+leader = OpenArmsMini(leader_config)
+
+# Connect and calibrate follower
+print("Note: If you have existing calibration, just press ENTER to use it.")
+follower.connect(calibrate=True)
+
+# Connect and calibrate leader
+print("Note: The leader arms will have torque disabled for manual control.")
+leader.connect(calibrate=True)
+
+print("\nPress ENTER to start teleoperation...")
+input()
+
+print("Press Ctrl+C to stop.\n")
+
+# All joints for both arms (16 motors total)
+all_joints = [
+    # Right arm
+    "right_joint_1",
+    "right_joint_2",
+    "right_joint_3",
+    "right_joint_4",
+    "right_joint_5",
+    "right_joint_6",
+    "right_joint_7",
+    "right_gripper",
+    # Left arm
+    "left_joint_1",
+    "left_joint_2",
+    "left_joint_3",
+    "left_joint_4",
+    "left_joint_5",
+    "left_joint_6",
+    "left_joint_7",
+    "left_gripper",
+]
+
+# Performance monitoring
+loop_times = []
+avg_loop_time = 0.0
+min_loop_time = float('inf')
+max_loop_time = 0.0
+stats_update_interval = 1.0  # Update stats every 1 second
+last_stats_update = time.perf_counter()
+
+
+SWAPPED_JOINTS = {
+    "right_joint_6": "right_joint_7",
+    "right_joint_7": "right_joint_6",
+    "left_joint_6": "left_joint_7",
+    "left_joint_7": "left_joint_6",
+}
+
+try:
+    while True:
+        loop_start = time.perf_counter()
+
+        # Get actions and observations
+        leader_action = leader.get_action()
+        follower_obs = follower.get_observation()
+
+        joint_action = {}
+        for joint in all_joints:
+            leader_key = f"{joint}.pos"
+
+            # Determine which follower joint this leader joint controls
+            follower_joint = SWAPPED_JOINTS.get(joint, joint)
+            follower_key = f"{follower_joint}.pos"
+
+            # Get leader position (default 0 if missing)
+            pos = leader_action.get(leader_key, 0.0)
+
+            # Convert gripper values: Mini uses 0-100, OpenArms uses 0 to -65 degrees
+            if "gripper" in joint:
+                # Map 0-100 (Mini) to 0 to -65 (OpenArms)
+                # 0 (closed) -> 0°, 100 (open) -> -65°
+                pos = (pos / 100.0) * -65.0
+
+            # Store in action dict for follower
+            joint_action[follower_key] = pos
+
+        follower.send_action(joint_action)
+
+        # Loop timing
+        loop_end = time.perf_counter()
+        loop_time = loop_end - loop_start
+        loop_times.append(loop_time)
+
+        # Update stats periodically
+        current_time = time.perf_counter()
+        if current_time - last_stats_update >= stats_update_interval:
+            if loop_times:
+                avg_loop_time = sum(loop_times) / len(loop_times)
+                min_loop_time = min(loop_times)
+                max_loop_time = max(loop_times)
+                loop_times = []
+                last_stats_update = current_time
+
+        # Display everything
+        sys.stdout.write("\033[H\033[J")  # Clear screen
+        
+        # Show timing stats at the top
+        if avg_loop_time > 0:
+            avg_hz = 1.0 / avg_loop_time
+            min_hz = 1.0 / max_loop_time if max_loop_time > 0 else 0
+            max_hz = 1.0 / min_loop_time if min_loop_time > 0 and min_loop_time < float('inf') else 0
+            print(f"[Performance] Target: {TARGET_FPS} Hz | Avg: {avg_hz:.1f} Hz | Range: {min_hz:.1f}-{max_hz:.1f} Hz | Loop: {avg_loop_time*1000:.1f} ms\n")
+        else:
+            print(f"[Performance] Target: {TARGET_FPS} Hz | Measuring...\n")
+
+        # Show joint positions
+        print(f"{'Joint':<20} {'Leader':>15} {'Follower':>15}")
+        print(f"{'':20} {'(0-100/deg)':>15} {'(deg)':>15}")
+        print("-" * 52)
+
+        for joint in all_joints:
+            leader_key = f"{joint}.pos"
+            follower_joint = SWAPPED_JOINTS.get(joint, joint)
+            follower_key = f"{follower_joint}.pos"
+
+            leader_pos = leader_action.get(leader_key, 0.0)
+            follower_pos = follower_obs.get(follower_key, 0.0)
+
+            print(f"{joint:<20} {leader_pos:>15.2f} {follower_pos:>15.2f}")
+
+        # Smart sleep to maintain target FPS
+        dt_s = time.perf_counter() - loop_start
+        busy_wait(max(0, 1.0 / TARGET_FPS - dt_s))
+            
+except KeyboardInterrupt:
+    print("\n\nStopping teleoperation...")
+finally:
+    # Disconnect devices
+    print("Disconnecting devices...")
+    try:
+        follower.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting follower: {e}")
+    
+    try:
+        leader.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting leader: {e}")
+    
+    print("Done!")
+

examples/openarms/teleop_with_compensation.py

@@ -0,0 +1,202 @@
+"""
+OpenArms Teleoperation with Gravity + Friction Compensation
+
+Leader arms (both LEFT and RIGHT): Gravity + Friction compensation (weightless, easy to move)
+Follower arms (both LEFT and RIGHT): Mirror leader movements
+
+Uses the URDF file from the lerobot repository.
+"""
+
+import time
+
+import numpy as np
+
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+
+# Friction compensation scale factor (1.0 = full, 0.3 = 30% for stability)
+FRICTION_SCALE = 1.0
+
+def main():
+    """Main teleoperation loop with gravity compensation"""
+
+    print("=" * 70)
+    print("OpenArms Teleoperation with Gravity Compensation")
+    print("=" * 70)
+
+    # Configuration
+    follower_config = OpenArmsFollowerConfig(
+        port_left="can2",
+        port_right="can3",
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+    )
+
+    leader_config = OpenArmsLeaderConfig(
+        port_left="can0",
+        port_right="can1",
+        can_interface="socketcan",
+        id="openarms_leader",
+        manual_control=False,  # Enable torque control for gravity compensation
+    )
+
+    # Initialize and connect
+    print("\nInitializing devices...")
+    follower = OpenArmsFollower(follower_config)
+    leader = OpenArmsLeader(leader_config)
+
+    follower.connect()
+    leader.connect()
+
+    # URDF is automatically loaded in the leader constructor
+    if leader.pin_robot is None:
+        raise RuntimeError("URDF model not loaded on leader. Gravity compensation not available.")
+
+    print("\nLeader BOTH arms: Gravity + Friction comp | Follower BOTH arms: Teleop")
+    print("Press ENTER to start...")
+    input()
+
+    # Enable motors on both leader arms for gravity compensation
+    leader.bus_right.enable_torque()
+    leader.bus_left.enable_torque()
+    time.sleep(0.1)
+
+    print("Press Ctrl+C to stop\n")
+
+    # Main control loop
+    loop_times = []
+    last_print_time = time.perf_counter()
+
+    # All joints (both arms)
+    all_joints = []
+    for motor in leader.bus_right.motors:
+        all_joints.append(f"right_{motor}")
+    for motor in leader.bus_left.motors:
+        all_joints.append(f"left_{motor}")
+
+    try:
+        while True:
+            loop_start = time.perf_counter()
+
+            # Get leader state
+            leader_action = leader.get_action()
+
+            # Extract positions and velocities in degrees
+            leader_positions_deg = {}
+            leader_velocities_deg_per_sec = {}
+            
+            for motor in leader.bus_right.motors:
+                pos_key = f"right_{motor}.pos"
+                vel_key = f"right_{motor}.vel"
+                if pos_key in leader_action:
+                    leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+                if vel_key in leader_action:
+                    leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+
+            for motor in leader.bus_left.motors:
+                pos_key = f"left_{motor}.pos"
+                vel_key = f"left_{motor}.vel"
+                if pos_key in leader_action:
+                    leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+                if vel_key in leader_action:
+                    leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+
+            # Calculate gravity torques for leader using built-in method
+            leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+            leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+            
+            # Calculate friction torques for leader using built-in method
+            leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+            leader_friction_torques_nm = leader._friction_from_velocity(
+                leader_velocities_rad_per_sec,
+                friction_scale=FRICTION_SCALE
+            )
+            
+            # Combine gravity + friction torques 
+            leader_total_torques_nm = {}
+            for motor_name in leader_gravity_torques_nm:
+                gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+                friction = leader_friction_torques_nm.get(motor_name, 0.0)
+                leader_total_torques_nm[motor_name] = gravity + friction
+
+            # Apply gravity + friction compensation to leader RIGHT arm (all joints including gripper)
+            for motor in leader.bus_right.motors:
+                full_name = f"right_{motor}"
+                position = leader_positions_deg.get(full_name, 0.0)
+                torque = leader_total_torques_nm.get(full_name, 0.0)
+                
+                # Get damping gain for stability
+                kd = leader.get_damping_kd(motor)
+
+                leader.bus_right._mit_control(
+                    motor=motor,
+                    kp=0.0,
+                    kd=kd,  # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque,
+                )
+
+            # Apply gravity + friction compensation to leader LEFT arm (all joints including gripper)
+            for motor in leader.bus_left.motors:
+                full_name = f"left_{motor}"
+                position = leader_positions_deg.get(full_name, 0.0)
+                torque = leader_total_torques_nm.get(full_name, 0.0)
+                
+                # Get damping gain for stability
+                kd = leader.get_damping_kd(motor)
+
+                leader.bus_left._mit_control(                    
+                    motor=motor,
+                    kp=0.0,
+                    kd=kd,  # Add damping for stability
+                    position_degrees=position,
+                    velocity_deg_per_sec=0.0,
+                    torque=torque,
+                )
+
+            # Send leader positions to follower (both arms)
+            follower_action = {}
+            for joint in all_joints:
+                pos_key = f"{joint}.pos"
+                if pos_key in leader_action:
+                    follower_action[pos_key] = leader_action[pos_key]
+
+            if follower_action:
+                follower.send_action(follower_action)
+
+            # Performance monitoring
+            loop_end = time.perf_counter()
+            loop_time = loop_end - loop_start
+            loop_times.append(loop_time)
+
+            if loop_end - last_print_time >= 2.0:
+                if loop_times:
+                    avg_time = sum(loop_times) / len(loop_times)
+                    current_hz = 1.0 / avg_time if avg_time > 0 else 0
+
+                    print(f"{current_hz:.1f} Hz ({avg_time*1000:.1f} ms)")
+
+                    loop_times = []
+                    last_print_time = loop_end
+
+    except KeyboardInterrupt:
+        print("\n\nStopping...")
+    finally:
+        try:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+            follower.disconnect()
+            print("✓ Shutdown complete")
+        except Exception as e:
+            print(f"Shutdown error: {e}")
+
+
+if __name__ == "__main__":
+    main()

examples/openarms/unify_task.py

@@ -0,0 +1,152 @@
+#!/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.
+
+"""
+Unify all tasks in a dataset to a single task (modifies in-place).
+
+This script:
+1. Loads a dataset
+2. Sets all task_index to 0 and task description to "fold"
+3. Updates tasks.parquet and task_index in data files (in-place, no copying)
+
+Usage:
+    python examples/openarms/unify_task.py --repo-id lerobot-data-collection/level1_rac1
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+from pathlib import Path
+
+import pandas as pd
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.utils import (
+    DATA_DIR,
+    write_info,
+    write_tasks,
+)
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+
+# Single unified task
+UNIFIED_TASK = "fold"
+
+
+def unify_dataset_tasks(
+    repo_id: str,
+    root: Path | None = None,
+    push_to_hub: bool = False,
+) -> None:
+    """Unify all tasks in a dataset to a single task (modifies in-place).
+
+    Args:
+        repo_id: Dataset repository ID.
+        root: Optional root path for dataset.
+        push_to_hub: Whether to push the result to HuggingFace Hub.
+    """
+    input_root = root if root else HF_LEROBOT_HOME / repo_id
+    input_repo_id = repo_id
+
+    logging.info(f"Loading metadata from {repo_id}")
+
+    # Load source metadata
+    src_meta = LeRobotDatasetMetadata(repo_id, root=input_root)
+
+    logging.info(f"Source dataset: {src_meta.total_episodes} episodes, {src_meta.total_frames} frames")
+    logging.info(f"Original tasks: {len(src_meta.tasks)}")
+
+    # Modify in-place (input_root == output_root supported)
+    data_dir = input_root / DATA_DIR
+
+    # Process data files - set all task_index to 0
+    logging.info("Processing data files (in-place)...")
+    for parquet_file in tqdm(sorted(data_dir.rglob("*.parquet")), desc="Processing data"):
+        df = pd.read_parquet(parquet_file)
+        df["task_index"] = 0  # All tasks unified to index 0
+        df.to_parquet(parquet_file)
+
+    # Process episodes metadata - set all tasks to unified task
+    logging.info("Processing episodes metadata (in-place)...")
+    episodes_dir = input_root / "meta" / "episodes"
+    if episodes_dir.exists():
+        for parquet_file in tqdm(sorted(episodes_dir.rglob("*.parquet")), desc="Processing episodes"):
+            df = pd.read_parquet(parquet_file)
+            df["tasks"] = [[UNIFIED_TASK]] * len(df)  # All episodes get the unified task
+            df.to_parquet(parquet_file)
+    else:
+        logging.warning(f"No episodes directory found at {episodes_dir}, skipping")
+
+    # Update tasks.parquet with single task
+    logging.info(f"Creating single task: {UNIFIED_TASK}")
+    new_tasks = pd.DataFrame({"task_index": [0]}, index=[UNIFIED_TASK])
+    write_tasks(new_tasks, input_root)
+
+    # Update info.json
+    new_info = src_meta.info.copy()
+    new_info["total_tasks"] = 1
+    write_info(new_info, input_root)
+
+    logging.info(f"Dataset modified in-place at {input_root}")
+    logging.info(f"Task: {UNIFIED_TASK}")
+
+    if push_to_hub:
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        logging.info(f"Pushing {input_repo_id} to hub")
+        dataset = LeRobotDataset(input_repo_id, root=input_root)
+        dataset.push_to_hub(private=True)
+        logging.info("Push complete!")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Unify all tasks in a dataset to a single task 'fold' (modifies in-place)."
+    )
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Dataset repository ID",
+    )
+    parser.add_argument(
+        "--root",
+        type=Path,
+        default=None,
+        help="Optional root path (defaults to HF_LEROBOT_HOME/repo_id)",
+    )
+    parser.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Push result to HuggingFace Hub",
+    )
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
+
+    unify_dataset_tasks(
+        repo_id=args.repo_id,
+        root=args.root,
+        push_to_hub=args.push_to_hub,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/openarms_web_interface/App.css

@@ -0,0 +1,745 @@
+body {
+  margin: 0;
+  padding: 0;
+  font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, sans-serif;
+  background: #f5f5f5;
+}
+
+main {
+  min-height: 100vh;
+  padding: 2rem;
+}
+
+header {
+  text-align: center;
+  margin-bottom: 2rem;
+}
+
+h1 {
+  font-size: 2rem;
+  font-weight: 600;
+  color: #333;
+  margin: 0;
+}
+
+h2 {
+  font-size: 1.25rem;
+  font-weight: 600;
+  color: #333;
+  margin: 0 0 1rem 0;
+}
+
+h3 {
+  font-size: 0.875rem;
+  font-weight: 600;
+  color: #666;
+  margin: 0 0 0.5rem 0;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+}
+
+.container {
+  max-width: 1920px;
+  margin: 0 auto;
+  display: grid;
+  grid-template-columns: minmax(500px, 600px) 1fr;
+  gap: 2rem;
+  align-items: start;
+}
+
+/* Left column container */
+.left-column {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+/* Right column container */
+.right-column {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+/* Responsive: Stack on smaller screens */
+@media (max-width: 1200px) {
+  .container {
+    grid-template-columns: 1fr;
+  }
+}
+
+.panel {
+  background: white;
+  border-radius: 8px;
+  padding: 1.5rem;
+  box-shadow: 0 1px 3px rgba(0,0,0,0.1);
+}
+
+.config-panel {
+  border: 2px solid #e5e7eb;
+}
+
+.config-header {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  cursor: pointer;
+  user-select: none;
+  padding: 0.5rem 0;
+}
+
+.config-header:hover {
+  opacity: 0.7;
+}
+
+.toggle-icon {
+  font-size: 1rem;
+  color: #6b7280;
+  transition: transform 0.2s;
+}
+
+.config-content {
+  margin-top: 1rem;
+  padding-top: 1rem;
+  border-top: 1px solid #e5e7eb;
+}
+
+.robot-setup {
+  margin-bottom: 0.5rem;
+}
+
+.robot-status {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 1rem;
+  border-radius: 6px;
+  font-weight: 500;
+  gap: 1rem;
+}
+
+.robot-status.ready {
+  background: linear-gradient(135deg, #d1fae5 0%, #a7f3d0 100%);
+  color: #065f46;
+  border: 1px solid #10b981;
+}
+
+.robot-status.not-ready {
+  background: linear-gradient(135deg, #fef3c7 0%, #fde68a 100%);
+  color: #92400e;
+  border: 1px solid #f59e0b;
+}
+
+.btn-setup {
+  background: #10b981;
+  color: white;
+  border: none;
+  padding: 0.5rem 1rem;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-setup:hover:not(:disabled) {
+  background: #059669;
+}
+
+.btn-setup:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.btn-zero {
+  background: #8b5cf6;
+  color: white;
+  border: none;
+  padding: 0.5rem 1rem;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-zero:hover:not(:disabled) {
+  background: #7c3aed;
+}
+
+.btn-zero:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.zero-position-section {
+  margin-top: 1rem;
+  padding-top: 1rem;
+  border-top: 1px solid #e5e7eb;
+}
+
+.btn-zero-large {
+  width: 100%;
+  background: #8b5cf6;
+  color: white;
+  border: none;
+  padding: 0.875rem 1.5rem;
+  border-radius: 8px;
+  font-size: 1rem;
+  font-weight: 600;
+  cursor: pointer;
+  transition: all 0.2s;
+  box-shadow: 0 2px 4px rgba(139, 92, 246, 0.2);
+}
+
+.btn-zero-large:hover:not(:disabled) {
+  background: #7c3aed;
+  box-shadow: 0 4px 8px rgba(139, 92, 246, 0.3);
+  transform: translateY(-1px);
+}
+
+.btn-zero-large:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+  box-shadow: none;
+  transform: none;
+}
+
+.delete-episode-section {
+  margin-top: 1rem;
+  padding-top: 1rem;
+  border-top: 1px solid #e5e7eb;
+}
+
+.btn-delete {
+  width: 100%;
+  background: #ef4444;
+  color: white;
+  border: none;
+  padding: 0.875rem 1.5rem;
+  border-radius: 8px;
+  font-size: 1rem;
+  font-weight: 600;
+  cursor: pointer;
+  transition: all 0.2s;
+  box-shadow: 0 2px 4px rgba(239, 68, 68, 0.2);
+}
+
+.btn-delete:hover:not(:disabled) {
+  background: #dc2626;
+  box-shadow: 0 4px 8px rgba(239, 68, 68, 0.3);
+  transform: translateY(-1px);
+}
+
+.btn-delete:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+  box-shadow: none;
+  transform: none;
+}
+
+.delete-info {
+  margin-top: 0.5rem;
+  font-size: 0.875rem;
+  color: #666;
+  text-align: center;
+  font-style: italic;
+}
+
+.btn-disconnect {
+  background: #ef4444;
+  color: white;
+  border: none;
+  padding: 0.5rem 1rem;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-disconnect:hover {
+  background: #dc2626;
+}
+
+.btn-refresh {
+  background: #3b82f6;
+  color: white;
+  border: none;
+  padding: 0.4rem 0.8rem;
+  border-radius: 4px;
+  font-size: 0.75rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.btn-refresh:hover:not(:disabled) {
+  background: #2563eb;
+}
+
+.btn-refresh:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.control-panel {
+  border: 2px solid #10b981;
+}
+
+.status-banner {
+  display: flex;
+  align-items: center;
+  gap: 1rem;
+  padding: 1rem 1.5rem;
+  border-radius: 6px;
+  margin-bottom: 1.5rem;
+  font-weight: 500;
+  font-size: 0.95rem;
+}
+
+.status-banner.initializing {
+  background: linear-gradient(135deg, #dbeafe 0%, #bfdbfe 100%);
+  color: #1e40af;
+  border-left: 4px solid #3b82f6;
+}
+
+.status-banner.encoding {
+  background: linear-gradient(135deg, #fef3c7 0%, #fde68a 100%);
+  color: #92400e;
+  border-left: 4px solid #f59e0b;
+}
+
+.status-banner.uploading {
+  background: linear-gradient(135deg, #e0e7ff 0%, #c7d2fe 100%);
+  color: #3730a3;
+  border-left: 4px solid #6366f1;
+}
+
+.status-banner.success {
+  background: linear-gradient(135deg, #d1fae5 0%, #a7f3d0 100%);
+  color: #065f46;
+  border-left: 4px solid #10b981;
+}
+
+.status-banner.warning {
+  background: linear-gradient(135deg, #fee2e2 0%, #fecaca 100%);
+  color: #991b1b;
+  border-left: 4px solid #ef4444;
+}
+
+.spinner {
+  width: 20px;
+  height: 20px;
+  border: 3px solid rgba(0, 0, 0, 0.1);
+  border-top-color: currentColor;
+  border-radius: 50%;
+  animation: spin 0.8s linear infinite;
+}
+
+@keyframes spin {
+  to { transform: rotate(360deg); }
+}
+
+.control-horizontal {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+.control-left {
+  display: flex;
+  flex-direction: column;
+  gap: 1rem;
+}
+
+.control-right {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+}
+
+.input-group {
+  display: flex;
+  gap: 0.5rem;
+  margin-bottom: 0;
+}
+
+input[type="text"] {
+  flex: 1;
+  padding: 0.75rem;
+  border: 1px solid #ddd;
+  border-radius: 4px;
+  font-size: 1rem;
+}
+
+input[type="text"]:disabled {
+  background: #f5f5f5;
+  cursor: not-allowed;
+}
+
+input[type="text"]:focus {
+  outline: none;
+  border-color: #10b981;
+}
+
+button {
+  padding: 0.75rem 1.5rem;
+  border: none;
+  border-radius: 4px;
+  font-size: 1rem;
+  font-weight: 500;
+  cursor: pointer;
+  transition: all 0.2s;
+}
+
+.btn-set-task {
+  background: #3b82f6;
+  color: white;
+  min-width: 120px;
+}
+
+.btn-set-task:hover:not(:disabled) {
+  background: #2563eb;
+}
+
+.btn-set-task:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.btn-start {
+  background: #10b981;
+  color: white;
+}
+
+.btn-start:hover:not(:disabled) {
+  background: #059669;
+}
+
+.btn-start:disabled {
+  background: #d1d5db;
+  cursor: not-allowed;
+}
+
+.btn-stop {
+  background: #ef4444;
+  color: white;
+}
+
+.btn-stop:hover {
+  background: #dc2626;
+}
+
+.btn-reset {
+  padding: 0.5rem 1rem;
+  background: #6b7280;
+  color: white;
+  font-size: 0.875rem;
+}
+
+.btn-reset:hover {
+  background: #4b5563;
+}
+
+.status {
+  display: flex;
+  align-items: center;
+  gap: 0.75rem;
+  padding: 1rem;
+  border-radius: 4px;
+  margin-bottom: 1rem;
+}
+
+.status.recording {
+  background: #fee2e2;
+  color: #991b1b;
+}
+
+.status.recording.recording-active {
+  display: flex;
+  flex-direction: column;
+  gap: 1rem;
+  background: #dc2626;
+  color: white;
+  padding: 1.5rem;
+  border: 4px solid #991b1b;
+  box-shadow: 0 4px 12px rgba(220, 38, 38, 0.4);
+  font-weight: 700;
+  font-size: 1rem;
+}
+
+.status.recording.recording-active .indicator {
+  width: 20px;
+  height: 20px;
+  background: #fef2f2;
+  animation: pulse-strong 1s ease-in-out infinite;
+}
+
+@keyframes pulse-strong {
+  0%, 100% { 
+    opacity: 1;
+    transform: scale(1);
+  }
+  50% { 
+    opacity: 0.7;
+    transform: scale(1.1);
+  }
+}
+
+.status.recording.recording-active .time-display {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+  font-size: 1.5rem;
+  font-weight: 700;
+  color: white;
+}
+
+.fps-display {
+  font-size: 1rem;
+  font-weight: 500;
+  opacity: 0.95;
+}
+
+.fps-warning {
+  color: #fef2f2;
+  animation: pulse-warning 1s ease-in-out infinite;
+}
+
+@keyframes pulse-warning {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0.5; }
+}
+
+.status.recording.recording-active .btn-stop {
+  align-self: stretch;
+}
+
+.ramp-up-countdown {
+  display: flex;
+  justify-content: center;
+  margin-bottom: 1rem;
+}
+
+.countdown-box {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  padding: 2rem 3rem;
+  background: linear-gradient(135deg, #fef3c7 0%, #fde68a 100%);
+  border: 4px solid #f59e0b;
+  border-radius: 16px;
+  box-shadow: 0 6px 20px rgba(245, 158, 11, 0.4);
+  min-width: 280px;
+  animation: pulse-warm 1.5s ease-in-out infinite;
+}
+
+@keyframes pulse-warm {
+  0%, 100% {
+    box-shadow: 0 6px 20px rgba(245, 158, 11, 0.4);
+  }
+  50% {
+    box-shadow: 0 6px 25px rgba(245, 158, 11, 0.6);
+  }
+}
+
+.countdown-label {
+  font-size: 1rem;
+  color: #92400e;
+  text-transform: uppercase;
+  letter-spacing: 1.5px;
+  font-weight: 800;
+  margin-bottom: 1rem;
+  text-align: center;
+}
+
+.countdown-value {
+  font-size: 4.5rem;
+  font-weight: 900;
+  color: #d97706;
+  font-family: 'Courier New', monospace;
+  line-height: 1;
+  text-shadow: 2px 2px 6px rgba(0, 0, 0, 0.15);
+  margin-bottom: 0.5rem;
+}
+
+.countdown-subtitle {
+  font-size: 0.875rem;
+  color: #78350f;
+  font-weight: 600;
+  font-style: italic;
+  text-align: center;
+  margin-top: 0.5rem;
+}
+
+.status.idle {
+  background: #f3f4f6;
+  color: #374151;
+}
+
+.indicator {
+  width: 12px;
+  height: 12px;
+  border-radius: 50%;
+  background: #ef4444;
+  animation: pulse 1.5s ease-in-out infinite;
+}
+
+@keyframes pulse {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0.5; }
+}
+
+.counter {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  gap: 0.75rem;
+  padding: 1.5rem;
+  background: linear-gradient(135deg, #f9fafb 0%, #f3f4f6 100%);
+  border-radius: 8px;
+  border: 2px solid #e5e7eb;
+  min-width: 200px;
+}
+
+.counter-label {
+  font-size: 0.75rem;
+  color: #6b7280;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  font-weight: 600;
+}
+
+.counter-value {
+  font-size: 3rem;
+  font-weight: 700;
+  color: #10b981;
+  line-height: 1;
+}
+
+.time-display {
+  font-size: 1.5rem;
+  font-weight: 600;
+  font-family: 'Courier New', monospace;
+}
+
+.error-box {
+  padding: 1rem;
+  background: #fee2e2;
+  color: #991b1b;
+  border-radius: 4px;
+  border-left: 4px solid #ef4444;
+  font-size: 0.875rem;
+}
+
+.config-section {
+  margin-bottom: 1.5rem;
+}
+
+.config-section:last-child {
+  margin-bottom: 0;
+}
+
+.config-grid {
+  display: grid;
+  grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+  gap: 1rem;
+}
+
+label {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+  font-size: 0.875rem;
+  color: #374151;
+  font-weight: 500;
+}
+
+select {
+  padding: 0.5rem;
+  border: 1px solid #ddd;
+  border-radius: 4px;
+  font-size: 0.875rem;
+  background: white;
+}
+
+select:disabled {
+  background: #f5f5f5;
+  cursor: not-allowed;
+}
+
+/* Camera Layout */
+.camera-layout {
+  display: flex;
+  flex-direction: column;
+  gap: 1.5rem;
+}
+
+.camera-base {
+  width: 100%;
+}
+
+.camera-wrist-container {
+  display: grid;
+  grid-template-columns: repeat(2, 1fr);
+  gap: 1.5rem;
+}
+
+.camera-wrist {
+  width: 100%;
+}
+
+.camera {
+  border: 1px solid #e5e7eb;
+  border-radius: 4px;
+  overflow: hidden;
+}
+
+.camera h3 {
+  padding: 0.75rem;
+  background: #f9fafb;
+  border-bottom: 1px solid #e5e7eb;
+  margin: 0;
+}
+
+.camera img {
+  width: 100%;
+  height: auto;
+  display: block;
+  background: #000;
+  min-height: 300px;
+  object-fit: cover;
+}
+
+.camera-placeholder {
+  text-align: center;
+  padding: 4rem 2rem;
+  background: #f9fafb;
+  border-radius: 4px;
+  border: 2px dashed #d1d5db;
+}
+
+.camera-placeholder p {
+  margin: 0.5rem 0;
+  font-size: 1rem;
+  color: #6b7280;
+}
+
+.camera-placeholder p:first-child {
+  font-size: 1.25rem;
+  font-weight: 500;
+  color: #374151;
+}
+
+.hint {
+  margin-top: 0.5rem;
+  font-size: 0.75rem;
+  color: #6b7280;
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  flex-wrap: wrap;
+}
+

examples/openarms_web_interface/App.jsx

@@ -0,0 +1,857 @@
+import { useState, useEffect, useCallback, useRef } from 'react';
+import './App.css';
+
+const API_BASE = 'http://localhost:8000/api';
+
+function App() {
+  // State
+  const [task, setTask] = useState('');
+  const [isRecording, setIsRecording] = useState(false);
+  const [isInitializing, setIsInitializing] = useState(false);
+  const [isEncoding, setIsEncoding] = useState(false);
+  const [isUploading, setIsUploading] = useState(false);
+  const [robotsReady, setRobotsReady] = useState(false);
+  const [elapsedTime, setElapsedTime] = useState(0);
+  const [currentFps, setCurrentFps] = useState(0);
+  const [loopFps, setLoopFps] = useState(0);
+  const [episodeCount, setEpisodeCount] = useState(0);
+  const [error, setError] = useState(null);
+  const [statusMessage, setStatusMessage] = useState('Ready');
+  const [uploadStatus, setUploadStatus] = useState(null);
+  const [rampUpRemaining, setRampUpRemaining] = useState(0);
+  const [movingToZero, setMovingToZero] = useState(false);
+  const [configExpanded, setConfigExpanded] = useState(false);
+  const [latestRepoId, setLatestRepoId] = useState(null);
+
+  // Configuration
+  const [config, setConfig] = useState({
+    leader_type: 'openarms',  // 'openarms' or 'openarms_mini'
+    leader_left: 'can0',
+    leader_right: 'can1',
+    follower_left: 'can2',
+    follower_right: 'can3',
+    left_wrist: '/dev/video0',
+    right_wrist: '/dev/video1',
+    base: '/dev/video4'
+  });
+
+  // Available options
+  const [availableCameras, setAvailableCameras] = useState([]);
+  const [availableUsbPorts, setAvailableUsbPorts] = useState([]);
+  const canInterfaces = ['can0', 'can1', 'can2', 'can3'];
+
+  const statusIntervalRef = useRef(null);
+  const hasInitializedRef = useRef(false);
+
+  const loadConfig = () => {
+    try {
+      const saved = localStorage.getItem('openarms_config');
+      if (saved) {
+        const loadedConfig = JSON.parse(saved);
+        setConfig(prev => ({ ...prev, ...loadedConfig }));
+      }
+    } catch (e) {
+      console.error('Load config error:', e);
+    }
+  };
+
+  const saveConfig = (newConfig) => {
+    try {
+      localStorage.setItem('openarms_config', JSON.stringify(newConfig || config));
+    } catch (e) {
+      console.error('Save config error:', e);
+    }
+  };
+
+  // Fetch status periodically
+  const fetchStatus = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/status`);
+      const data = await response.json();
+
+      setIsRecording(data.is_recording);
+      setIsInitializing(data.is_initializing);
+      setIsEncoding(data.is_encoding);
+      setIsUploading(data.is_uploading);
+      setRobotsReady(data.robots_ready);
+      setElapsedTime(data.elapsed_time);
+      setCurrentFps(data.current_fps || 0);
+      setLoopFps(data.loop_fps || 0);
+      setEpisodeCount(data.episode_count);
+      setError(data.error);
+      setStatusMessage(data.status_message || 'Ready');
+      setUploadStatus(data.upload_status);
+      setRampUpRemaining(data.ramp_up_remaining || 0);
+      setMovingToZero(data.moving_to_zero || false);
+      
+      // Track the latest repo_id from the backend
+      if (data.latest_repo_id) {
+        setLatestRepoId(data.latest_repo_id);
+      }
+
+      if (data.config) {
+        // Only merge server config if we don't have a saved config (first load)
+        if (!localStorage.getItem('openarms_config')) {
+          setConfig(prev => {
+            const merged = { ...data.config, ...prev };
+            localStorage.setItem('openarms_config', JSON.stringify(merged));
+            return merged;
+          });
+        }
+      }
+    } catch (e) {
+      console.error('Failed to fetch status:', e);
+    }
+  };
+
+  const setupRobots = async () => {
+    // Show warning to verify camera positions
+    const confirmed = window.confirm(
+      '⚠️ IMPORTANT: Before connecting robots, please verify:\n\n' +
+      '📹 Check that cameras are correctly positioned:\n' +
+      '   • LEFT wrist camera is actually on the LEFT arm\n' +
+      '   • RIGHT wrist camera is actually on the RIGHT arm\n' +
+      '   • BASE camera is actually the BASE/overhead camera\n\n' +
+      'Incorrect camera positioning will result in invalid training data!\n\n' +
+      'Click OK to continue with robot setup, or Cancel to review configuration.'
+    );
+    
+    if (!confirmed) {
+      return; // User cancelled, don't proceed
+    }
+    
+    setError(null);
+    try {
+      const response = await fetch(`${API_BASE}/robots/setup`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(config)
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to setup robots');
+      }
+
+      await response.json();
+      saveConfig(config);
+    } catch (e) {
+      setError(`Robot setup failed: ${e.message}`);
+    }
+  };
+
+  // Disconnect robots
+  const disconnectRobots = async () => {
+    try {
+      await fetch(`${API_BASE}/robots/disconnect`, { method: 'POST' });
+      setRobotsReady(false);
+    } catch (e) {
+      console.error('Failed to disconnect robots:', e);
+    }
+  };
+
+  // Discover cameras
+  const discoverCameras = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/cameras/discover`);
+      const data = await response.json();
+      const cameras = data.cameras || [];
+      setAvailableCameras(cameras);
+
+      // Get list of valid camera IDs
+      const validCameraIds = cameras.map(cam => String(cam.id));
+
+      // Auto-fix config if current values are invalid or not set
+      const updated = { ...config };
+      let changed = false;
+
+      // Auto-fix invalid camera config
+      if (!config.left_wrist || !validCameraIds.includes(config.left_wrist)) {
+        if (cameras.length >= 1) {
+          updated.left_wrist = String(cameras[0].id);
+          changed = true;
+        }
+      }
+
+      if (!config.right_wrist || !validCameraIds.includes(config.right_wrist)) {
+        if (cameras.length >= 2) {
+          updated.right_wrist = String(cameras[1].id);
+          changed = true;
+        }
+      }
+
+      if (!config.base || !validCameraIds.includes(config.base)) {
+        if (cameras.length >= 3) {
+          updated.base = String(cameras[2].id);
+          changed = true;
+        }
+      }
+
+      if (changed) {
+        setConfig(updated);
+        saveConfig(updated);
+      }
+
+      if (cameras.length === 0) {
+        setError('No cameras detected! Please connect cameras and refresh.');
+      }
+    } catch (e) {
+      console.error('Failed to discover cameras:', e);
+      setError(`Camera discovery failed: ${e.message}`);
+    }
+  };
+
+  // Discover USB ports
+  const discoverUsbPorts = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/usb/discover`);
+      const data = await response.json();
+      const ports = data.ports || [];
+      setAvailableUsbPorts(ports);
+
+      // Auto-fix config if OpenArms Mini is selected and ports are invalid
+      if (config.leader_type === 'openarms_mini') {
+        const updated = { ...config };
+        let changed = false;
+
+        if (ports.length >= 1 && !ports.includes(config.leader_left)) {
+          updated.leader_left = ports[0];
+          changed = true;
+        }
+
+        if (ports.length >= 2 && !ports.includes(config.leader_right)) {
+          updated.leader_right = ports[1];
+          changed = true;
+        }
+
+        if (changed) {
+          setConfig(updated);
+          saveConfig(updated);
+        }
+      }
+
+      if (ports.length === 0) {
+        console.warn('No USB ports detected for OpenArms Mini');
+      }
+    } catch (e) {
+      console.error('Failed to discover USB ports:', e);
+    }
+  };
+
+  // Set task only (for pedal use)
+  const setTaskOnly = async () => {
+    if (!task.trim()) {
+      setError('Please enter a task description');
+      return;
+    }
+
+    setError(null);
+
+    try {
+      const response = await fetch(`${API_BASE}/recording/set-task`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ task, ...config })
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to set task');
+      }
+
+      const result = await response.json();
+      setStatusMessage(result.message || `Task set: ${task}`);
+      saveConfig(config);
+      
+      // Clear success message after 3 seconds
+      setTimeout(() => {
+        if (!isRecording && !isInitializing) {
+          setStatusMessage('Ready');
+        }
+      }, 3000);
+    } catch (e) {
+      setError(e.message);
+    }
+  };
+
+  // Start recording
+  const startRecording = async () => {
+    if (!task.trim()) {
+      setError('Please enter a task description');
+      return;
+    }
+
+    setError(null);
+
+    try {
+      const response = await fetch(`${API_BASE}/recording/start`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ task, ...config })
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to start recording');
+      }
+
+      await response.json();
+      saveConfig(config);
+    } catch (e) {
+      setError(e.message);
+    }
+  };
+
+  // Stop recording
+  const stopRecording = async () => {
+    try {
+      const response = await fetch(`${API_BASE}/recording/stop`, {
+        method: 'POST'
+      });
+
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to stop recording');
+      }
+
+      const data = await response.json();
+      setError(null);
+      // Update latest repo_id after recording
+      if (data.dataset_name) {
+        setLatestRepoId(`lerobot-data-collection/${data.dataset_name}`);
+      }
+    } catch (e) {
+      setError(e.message);
+    }
+  };
+
+  const deleteLatestEpisode = async () => {
+    if (!latestRepoId) {
+      setError('No episode to delete');
+      return;
+    }
+    
+    const confirmed = window.confirm(
+      `WARNING: This will permanently delete the repository:\n\n${latestRepoId}\n\nThis action cannot be undone. Continue?`
+    );
+    
+    if (!confirmed) {
+      return;
+    }
+    
+    try {
+      const response = await fetch(`${API_BASE}/recording/delete-latest`, { method: 'POST' });
+      
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to delete episode');
+      }
+
+      const data = await response.json();
+      setLatestRepoId(null);
+      setEpisodeCount(Math.max(0, episodeCount - 1));
+      setStatusMessage(`Deleted: ${data.deleted_repo}`);
+      
+      setTimeout(() => {
+        if (!isRecording && !isInitializing) {
+          setStatusMessage('Ready');
+        }
+      }, 3000);
+    } catch (e) {
+      setError(`Delete failed: ${e.message}`);
+    }
+  };
+
+  // Reset counter
+  const resetCounter = async () => {
+    try {
+      await fetch(`${API_BASE}/counter/reset`, { method: 'POST' });
+      setEpisodeCount(0);
+    } catch (e) {
+      console.error('Failed to reset counter:', e);
+    }
+  };
+
+  // Move robot to zero position
+  const moveToZero = async () => {
+    setError(null);
+    try {
+      const response = await fetch(`${API_BASE}/robots/move-to-zero`, { method: 'POST' });
+      if (!response.ok) {
+        const data = await response.json();
+        throw new Error(data.detail || 'Failed to move to zero position');
+      }
+      await response.json();
+    } catch (e) {
+      setError(`Move to zero failed: ${e.message}`);
+    }
+  };
+
+  // Format time as MM:SS
+  const formatTime = (seconds) => {
+    const mins = Math.floor(seconds / 60);
+    const secs = Math.floor(seconds % 60);
+    return `${mins.toString().padStart(2, '0')}:${secs.toString().padStart(2, '0')}`;
+  };
+
+  // Update config and save
+  const updateConfig = (key, value) => {
+    const updated = { ...config, [key]: value };
+    setConfig(updated);
+    saveConfig(updated);
+  };
+
+  // Initialize on mount only
+  useEffect(() => {
+    // Prevent double-initialization in development
+    if (hasInitializedRef.current) {
+      return;
+    }
+    hasInitializedRef.current = true;
+
+    loadConfig();
+    discoverCameras();
+    discoverUsbPorts();
+    fetchStatus();
+    statusIntervalRef.current = setInterval(fetchStatus, 1000);
+
+    return () => {
+      if (statusIntervalRef.current) {
+        clearInterval(statusIntervalRef.current);
+      }
+    };
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []); // Run only once on mount
+
+  // Discover USB ports when leader type changes to Mini
+  useEffect(() => {
+    if (config.leader_type === 'openarms_mini') {
+      discoverUsbPorts();
+    }
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [config.leader_type]);
+
+  return (
+    <main>
+      <header>
+        <h1>OpenArms Recording</h1>
+      </header>
+
+      <div className="container">
+        {/* Left Column: Configuration and Recording Control */}
+        <div className="left-column">
+          {/* Configuration Panel */}
+          <section className="panel config-panel">
+          <div
+            className="config-header"
+            onClick={() => setConfigExpanded(!configExpanded)}
+            role="button"
+            tabIndex={0}
+            onKeyDown={(e) => e.key === 'Enter' && setConfigExpanded(!configExpanded)}
+          >
+            <h2>⚙️ Configuration</h2>
+            <span className="toggle-icon">{configExpanded ? '▼' : '▶'}</span>
+          </div>
+
+          {configExpanded && (
+            <div className="config-content">
+              {/* Robot Setup */}
+              <div className="config-section">
+                <h3>🤖 Robot Setup</h3>
+                <div className="robot-setup">
+                  {robotsReady ? (
+                    <div className="robot-status ready">
+                      <span>✅ Robots Ready - Recording will start instantly</span>
+                      <button onClick={disconnectRobots} className="btn-disconnect">
+                        Disconnect Robots
+                      </button>
+                    </div>
+                  ) : (
+                    <div className="robot-status not-ready">
+                      <span>⚠️ Robots not initialized - Recording will take ~10 seconds</span>
+                      <button
+                        onClick={setupRobots}
+                        disabled={isRecording || isInitializing}
+                        className="btn-setup"
+                      >
+                        🚀 Setup Robots
+                      </button>
+                    </div>
+                  )}
+                </div>
+              </div>
+
+              {/* Leader Type Selection */}
+              <div className="config-section">
+                <h3>🎮 Leader Type</h3>
+                <div className="config-grid">
+                  <label style={{gridColumn: '1 / -1'}}>
+                    Leader Arm Type
+                    <select
+                      value={config.leader_type}
+                      onChange={(e) => updateConfig('leader_type', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      <option value="openarms">OpenArms (CAN Bus - Damiao Motors)</option>
+                      <option value="openarms_mini">OpenArms Mini (USB - Feetech Motors)</option>
+                    </select>
+                  </label>
+                </div>
+              </div>
+
+              {/* Leader Interfaces (CAN or USB based on type) */}
+              <div className="config-section">
+                <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'center', marginBottom: '0.5rem' }}>
+                  <h3>
+                    {config.leader_type === 'openarms_mini' 
+                      ? `Leader Ports (USB/Serial) ${availableUsbPorts.length > 0 ? `(${availableUsbPorts.length} detected)` : ''}` 
+                      : 'Leader Interfaces (CAN)'}
+                  </h3>
+                  {config.leader_type === 'openarms_mini' && (
+                    <button
+                      onClick={discoverUsbPorts}
+                      className="btn-refresh"
+                      disabled={isRecording || robotsReady}
+                    >
+                      🔄 Refresh
+                    </button>
+                  )}
+                </div>
+                
+                <div className="config-grid">
+                  <label>
+                    Leader Left
+                    <select
+                      value={config.leader_left}
+                      onChange={(e) => updateConfig('leader_left', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {config.leader_type === 'openarms_mini' ? (
+                        availableUsbPorts.length > 0 ? (
+                          availableUsbPorts.map((port) => (
+                            <option key={port} value={port}>{port}</option>
+                          ))
+                        ) : (
+                          <option value="">No USB ports detected</option>
+                        )
+                      ) : (
+                        canInterfaces.map((iface) => (
+                          <option key={iface} value={iface}>{iface}</option>
+                        ))
+                      )}
+                    </select>
+                  </label>
+
+                  <label>
+                    Leader Right
+                    <select
+                      value={config.leader_right}
+                      onChange={(e) => updateConfig('leader_right', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {config.leader_type === 'openarms_mini' ? (
+                        availableUsbPorts.length > 0 ? (
+                          availableUsbPorts.map((port) => (
+                            <option key={port} value={port}>{port}</option>
+                          ))
+                        ) : (
+                          <option value="">No USB ports detected</option>
+                        )
+                      ) : (
+                        canInterfaces.map((iface) => (
+                          <option key={iface} value={iface}>{iface}</option>
+                        ))
+                      )}
+                    </select>
+                  </label>
+                </div>
+              </div>
+
+              {/* Follower CAN Interfaces */}
+              <div className="config-section">
+                <h3>Follower Interfaces (CAN)</h3>
+                
+                <div className="config-grid">
+                  <label>
+                    Follower Left
+                    <select
+                      value={config.follower_left}
+                      onChange={(e) => updateConfig('follower_left', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {canInterfaces.map((iface) => (
+                        <option key={iface} value={iface}>{iface}</option>
+                      ))}
+                    </select>
+                  </label>
+
+                  <label>
+                    Follower Right
+                    <select
+                      value={config.follower_right}
+                      onChange={(e) => updateConfig('follower_right', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {canInterfaces.map((iface) => (
+                        <option key={iface} value={iface}>{iface}</option>
+                      ))}
+                    </select>
+                  </label>
+                </div>
+              </div>
+
+              {/* Camera Configuration */}
+              <div className="config-section">
+                <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'center', marginBottom: '0.5rem' }}>
+                  <h3>Cameras {availableCameras.length > 0 && `(${availableCameras.length} detected)`}</h3>
+                  <button
+                    onClick={discoverCameras}
+                    className="btn-refresh"
+                    disabled={isRecording || robotsReady}
+                  >
+                    🔄 Refresh
+                  </button>
+                </div>
+                <div className="config-grid">
+                  <label>
+                    Left Wrist
+                    <select
+                      value={config.left_wrist}
+                      onChange={(e) => updateConfig('left_wrist', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {availableCameras.map((cam) => (
+                        <option key={cam.id} value={String(cam.id)}>
+                          {cam.name || `Camera @ ${cam.id}`}
+                        </option>
+                      ))}
+                    </select>
+                  </label>
+
+                  <label>
+                    Right Wrist
+                    <select
+                      value={config.right_wrist}
+                      onChange={(e) => updateConfig('right_wrist', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {availableCameras.map((cam) => (
+                        <option key={cam.id} value={String(cam.id)}>
+                          {cam.name || `Camera @ ${cam.id}`}
+                        </option>
+                      ))}
+                    </select>
+                  </label>
+
+                  <label>
+                    Base Camera
+                    <select
+                      value={config.base}
+                      onChange={(e) => updateConfig('base', e.target.value)}
+                      disabled={isRecording || robotsReady}
+                    >
+                      {availableCameras.map((cam) => (
+                        <option key={cam.id} value={String(cam.id)}>
+                          {cam.name || `Camera @ ${cam.id}`}
+                        </option>
+                      ))}
+                    </select>
+                  </label>
+                </div>
+              </div>
+            </div>
+          )}
+        </section>
+
+        {/* Control Panel */}
+        <section className="panel control-panel">
+          <h2>🎬 Recording Control</h2>
+
+          {/* Status Banner - Always show important statuses */}
+          {isInitializing && (
+            <div className="status-banner initializing">
+              <div className="spinner"></div>
+              <span>{statusMessage}</span>
+            </div>
+          )}
+
+          {isEncoding && (
+            <div className="status-banner encoding">
+              <div className="spinner"></div>
+              <span>📹 {statusMessage}</span>
+            </div>
+          )}
+
+          {isUploading && (
+            <div className="status-banner uploading">
+              <div className="spinner"></div>
+              <span>☁️ {statusMessage}</span>
+            </div>
+          )}
+
+          {uploadStatus && !isRecording && !isEncoding && !isUploading && (
+            <div className={`status-banner ${uploadStatus.startsWith('✓') ? 'success' : 'warning'}`}>
+              <span>{uploadStatus}</span>
+            </div>
+          )}
+
+          <div className="control-horizontal">
+            {/* Task Input and Status */}
+            <div className="control-left">
+              <div className="input-group">
+                <input
+                  type="text"
+                  value={task}
+                  onChange={(e) => setTask(e.target.value)}
+                  placeholder="Task description (e.g., 'pick and place')"
+                  disabled={isRecording || isInitializing || isEncoding || isUploading}
+                  onKeyPress={(e) => {
+                    if (e.key === 'Enter' && robotsReady) {
+                      setTaskOnly();
+                    }
+                  }}
+                />
+                <button
+                  onClick={setTaskOnly}
+                  disabled={isRecording || isInitializing || isEncoding || isUploading || !robotsReady}
+                  className="btn-set-task"
+                  title={!robotsReady ? 'Please setup robots first' : 'Store task for pedal use (Enter key)'}
+                >
+                  💾 Set Task
+                </button>
+                <button
+                  onClick={startRecording}
+                  disabled={isRecording || isInitializing || isEncoding || isUploading || !robotsReady}
+                  className="btn-start"
+                  title={!robotsReady ? 'Please setup robots first' : ''}
+                >
+                  {isInitializing
+                    ? '⏳ Initializing...'
+                    : isRecording
+                    ? '⏺ Recording...'
+                    : robotsReady
+                    ? '⏺ Start Recording'
+                    : '⏺ Setup Robots First'}
+                </button>
+              </div>
+
+              {/* Ramp-up Countdown */}
+              {isRecording && rampUpRemaining > 0 && (
+                <div className="ramp-up-countdown">
+                  <div className="countdown-box">
+                    <div className="countdown-label">⚡ WARMING UP - PID RAMP-UP</div>
+                    <div className="countdown-value">{rampUpRemaining.toFixed(1)}s</div>
+                    <div className="countdown-subtitle">Recording will start automatically...</div>
+                  </div>
+                </div>
+              )}
+
+              {/* Recording Status - Only show after ramp-up */}
+              {isRecording && rampUpRemaining <= 0 && (
+                <div className="status recording recording-active">
+                  <div className="indicator"></div>
+                  <div className="time-display">
+                    <span>{formatTime(elapsedTime)}</span>
+                    <span className="fps-display">
+                      Loop: {loopFps.toFixed(1)} Hz
+                      {loopFps > 0 && loopFps < 29 && <span className="fps-warning"> ⚠️</span>}
+                    </span>
+                    <span className="fps-display">Recording: {currentFps.toFixed(1)} FPS</span>
+                  </div>
+                  <button onClick={stopRecording} className="btn-stop">
+                    ⏹ Stop
+                  </button>
+                </div>
+              )}
+            </div>
+
+            {/* Episode Counter */}
+            <div className="control-right">
+              <div className="counter">
+                <div className="counter-label">Episodes Recorded</div>
+                <div className="counter-value">{episodeCount}</div>
+                <button onClick={resetCounter} className="btn-reset">
+                  Reset
+                </button>
+              </div>
+            </div>
+          </div>
+
+          {/* Delete Latest Episode Button */}
+          {!isRecording && !isInitializing && latestRepoId && (
+            <div className="delete-episode-section">
+              <button 
+                onClick={deleteLatestEpisode}
+                className="btn-delete"
+                title="Delete the latest recorded episode from HuggingFace Hub"
+              >
+                Delete Latest Episode
+              </button>
+              <div className="delete-info">Will delete: {latestRepoId}</div>
+            </div>
+          )}
+
+          {/* Move to Zero Button */}
+          {robotsReady && !isRecording && !isInitializing && (
+            <div className="zero-position-section">
+              <button 
+                onClick={moveToZero} 
+                disabled={movingToZero}
+                className="btn-zero-large"
+                title="Move both leader and follower robots to zero position (2s)"
+              >
+                {movingToZero ? '⏳ Moving to Zero Position...' : '🎯 Move to Zero Position (Leader + Follower)'}
+              </button>
+            </div>
+          )}
+
+          {/* Error Display */}
+          {error && (
+            <div className="error-box">
+              ⚠️ {error}
+            </div>
+          )}
+        </section>
+        </div>
+
+        {/* Right Column: Camera Feeds */}
+        <div className="right-column">
+          <section className="panel cameras">
+          <h2>📹 Camera Views</h2>
+          {robotsReady || isRecording || isInitializing ? (
+            <div className="camera-layout">
+              {/* Base camera - full width */}
+              <div className="camera camera-base">
+                <h3>Base Camera</h3>
+                <img src={`${API_BASE}/camera/stream/base`} alt="Base Camera" />
+              </div>
+
+              {/* Wrist cameras - side by side */}
+              <div className="camera-wrist-container">
+                <div className="camera camera-wrist">
+                  <h3>Left Wrist</h3>
+                  <img src={`${API_BASE}/camera/stream/left_wrist`} alt="Left Wrist Camera" />
+                </div>
+
+                <div className="camera camera-wrist">
+                  <h3>Right Wrist</h3>
+                  <img src={`${API_BASE}/camera/stream/right_wrist`} alt="Right Wrist Camera" />
+                </div>
+              </div>
+            </div>
+          ) : (
+            <div className="camera-placeholder">
+              <p>📷 Camera feeds will appear when robots are set up</p>
+              <p className="hint">Click "Setup Robots" above to preview camera feeds</p>
+            </div>
+          )}
+        </section>
+        </div>
+
+      </div>
+    </main>
+  );
+}
+
+export default App;
+

examples/openarms_web_interface/README.md

@@ -0,0 +1,41 @@
+# OpenArms Web Recording Interface
+
+A web interface for recording OpenArms datasets.
+
+## Installation
+
+```bash
+cd examples/openarms_web_interface
+npm install
+```
+
+## Usage
+
+**Start everything with one command:**
+
+```bash
+./launch.sh
+```
+
+This will:
+- Start the FastAPI backend on port 8000
+- Start the React frontend on port 5173
+- Show live logs from both services
+
+Then open your browser to: **http://localhost:5173**
+
+**Stop with:** `Ctrl+C`
+
+---
+
+## Workflow
+
+1. **Configure CAN interfaces** and **camera paths** in the dropdowns
+2. Click **"Setup Robots"** to initialize (once at start)
+3. Enter a **task description**
+4. Click **"Start Recording"** to begin an episode
+5. Click **"Stop Recording"** when done
+6. Dataset is automatically encoded and uploaded to HuggingFace Hub as **private**
+7. Repeat steps 3-6 for more episodes (no need to re-setup robots!)
+
+---

examples/openarms_web_interface/index.html

@@ -0,0 +1,12 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>OpenArms Recording Interface</title>
+  </head>
+  <body>
+    <div id="root"></div>
+    <script type="module" src="/main.jsx"></script>
+  </body>
+</html>

examples/openarms_web_interface/launch.sh

@@ -0,0 +1,142 @@
+#!/bin/bash
+
+# OpenArms Web Interface Launcher
+# Starts Rerun viewer, FastAPI backend, and React frontend
+
+set -e
+
+# Colors for output
+GREEN='\033[0;32m'
+BLUE='\033[0;34m'
+YELLOW='\033[1;33m'
+RED='\033[0;31m'
+NC='\033[0m' # No Color
+
+# Get script directory
+SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+cd "$SCRIPT_DIR"
+
+echo -e "${BLUE}╔════════════════════════════════════════╗${NC}"
+echo -e "${BLUE}║   OpenArms Web Recording Interface    ║${NC}"
+echo -e "${BLUE}╚════════════════════════════════════════╝${NC}"
+echo ""
+
+# Function to cleanup on exit
+cleanup() {
+    echo ""
+    echo -e "${YELLOW}Shutting down services...${NC}"
+    
+    # Kill all child processes
+    pkill -P $$ 2>/dev/null || true
+    
+    # Kill specific services by port
+    lsof -ti:8000 | xargs kill -9 2>/dev/null || true  # Backend
+    lsof -ti:5173 | xargs kill -9 2>/dev/null || true  # Frontend
+    lsof -ti:9876 | xargs kill -9 2>/dev/null || true  # Rerun (if spawned)
+    
+    echo -e "${GREEN}✓ Services stopped${NC}"
+    exit 0
+}
+
+# Register cleanup on script exit
+trap cleanup EXIT INT TERM
+
+# Check if required commands exist
+command -v rerun >/dev/null 2>&1 || { 
+    echo -e "${RED}✗ Error: 'rerun' not found. Please install: pip install rerun-sdk${NC}"
+    exit 1
+}
+
+command -v python >/dev/null 2>&1 || { 
+    echo -e "${RED}✗ Error: 'python' not found${NC}"
+    exit 1
+}
+
+command -v npm >/dev/null 2>&1 || { 
+    echo -e "${RED}✗ Error: 'npm' not found${NC}"
+    exit 1
+}
+
+# Check if node_modules exists
+if [ ! -d "node_modules" ]; then
+    echo -e "${YELLOW}⚠ node_modules not found. Running npm install...${NC}"
+    npm install
+    echo -e "${GREEN}✓ Dependencies installed${NC}"
+    echo ""
+fi
+
+echo -e "${GREEN}Starting services...${NC}"
+echo ""
+
+# 1. Start FastAPI backend (Rerun will start when recording begins)
+echo -e "${BLUE}[1/2]${NC} Starting FastAPI backend on port 8000..."
+cd "$SCRIPT_DIR"
+
+# Use Python from current environment (if lerobot env is active, it will use that)
+# Otherwise, check if we need to use conda run
+if [[ "$CONDA_DEFAULT_ENV" == "lerobot" ]]; then
+    # Already in lerobot environment
+    echo -e "${GREEN}✓ Using active lerobot environment${NC}"
+    PYTHON_CMD="python"
+elif command -v conda >/dev/null 2>&1 && conda env list | grep -q "^lerobot "; then
+    # lerobot env exists but not active - use conda run
+    echo -e "${YELLOW}Using conda run with lerobot environment...${NC}"
+    PYTHON_CMD="conda run -n lerobot --no-capture-output python"
+else
+    # Fall back to system python
+    echo -e "${YELLOW}⚠ Warning: lerobot environment not found, using system python${NC}"
+    PYTHON_CMD="python"
+fi
+
+$PYTHON_CMD web_record_server.py > /tmp/openarms_backend.log 2>&1 &
+BACKEND_PID=$!
+sleep 3
+
+if ps -p $BACKEND_PID > /dev/null; then
+    echo -e "${GREEN}✓ Backend started${NC} (PID: $BACKEND_PID)"
+    echo -e "      URL: ${BLUE}http://localhost:8000${NC}"
+else
+    echo -e "${RED}✗ Failed to start backend${NC}"
+    echo -e "${YELLOW}Check logs: tail -f /tmp/openarms_backend.log${NC}"
+    exit 1
+fi
+echo ""
+
+# 2. Start React frontend
+echo -e "${BLUE}[2/2]${NC} Starting React frontend on port 5173..."
+cd "$SCRIPT_DIR"
+npm run dev > /tmp/openarms_frontend.log 2>&1 &
+FRONTEND_PID=$!
+sleep 3
+
+if ps -p $FRONTEND_PID > /dev/null; then
+    echo -e "${GREEN}✓ Frontend started${NC} (PID: $FRONTEND_PID)"
+    echo -e "      URL: ${BLUE}http://localhost:5173${NC}"
+else
+    echo -e "${RED}✗ Failed to start frontend${NC}"
+    echo -e "${YELLOW}Check logs: tail -f /tmp/openarms_frontend.log${NC}"
+    exit 1
+fi
+echo ""
+
+# Display status
+echo -e "${GREEN}╔════════════════════════════════════════╗${NC}"
+echo -e "${GREEN}║     All services running! 🚀           ║${NC}"
+echo -e "${GREEN}╚════════════════════════════════════════╝${NC}"
+echo ""
+echo -e "🔧 ${BLUE}Backend:${NC}  http://localhost:8000"
+echo -e "🌐 ${BLUE}Frontend:${NC} http://localhost:5173"
+echo -e "📊 ${BLUE}Rerun:${NC}    Will spawn automatically when recording starts"
+echo ""
+echo -e "${YELLOW}Open your browser to:${NC} ${BLUE}http://localhost:5173${NC}"
+echo ""
+echo -e "${YELLOW}Logs:${NC}"
+echo -e "  • Backend:  tail -f /tmp/openarms_backend.log"
+echo -e "  • Frontend: tail -f /tmp/openarms_frontend.log"
+echo ""
+echo -e "${RED}Press Ctrl+C to stop all services${NC}"
+echo ""
+
+# Keep script running and wait for any service to exit
+wait
+

examples/openarms_web_interface/main.jsx

@@ -0,0 +1,7 @@
+import { createRoot } from 'react-dom/client'
+import App from './App.jsx'
+
+createRoot(document.getElementById('root')).render(
+  <App />
+)
+

examples/openarms_web_interface/package.json

@@ -0,0 +1,21 @@
+{
+  "name": "openarms-web-interface",
+  "private": true,
+  "version": "0.0.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "vite build",
+    "preview": "vite preview"
+  },
+  "dependencies": {
+    "react": "^18.3.1",
+    "react-dom": "^18.3.1"
+  },
+  "devDependencies": {
+    "@types/react": "^18.3.12",
+    "@types/react-dom": "^18.3.1",
+    "@vitejs/plugin-react": "^4.3.4",
+    "vite": "^6.0.1"
+  }
+}

examples/openarms_web_interface/vite.config.js

@@ -0,0 +1,17 @@
+import { defineConfig } from 'vite'
+import react from '@vitejs/plugin-react'
+
+// https://vite.dev/config/
+export default defineConfig({
+  plugins: [react()],
+  server: {
+    port: 5173,
+    strictPort: false,
+    host: true,
+    open: false
+  },
+  build: {
+    outDir: 'dist',
+    sourcemap: true
+  }
+})

examples/rac/cmd.sh

@@ -0,0 +1 @@
+ python examples/rac/rac_data_collection_openarms_rtc.py     --robot.type=openarms_follower     --robot.port_right=can1     --robot.port_left=can0      --robot.cameras="{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}"      --teleop.type=openarms_mini     --teleop.port_right=/dev/ttyACM0     --teleop.port_left=/dev/ttyACM1         --policy.path=lerobot-data-collection/level1_rac3_100k        --dataset.repo_id=lerobot-data-collection/level1_rac3_rtc_s5_2  --dataset.single_task="Fold the T-shirt properly"   --dataset.num_episodes=5

examples/rac/rac_data_collection.py

@@ -0,0 +1,638 @@
+#!/usr/bin/env python
+"""
+RaC (Recovery and Correction) Data Collection with Policy Rollout + Human Intervention.
+
+This implements the RaC paradigm from "RaC: Robot Learning for Long-Horizon Tasks
+by Scaling Recovery and Correction" (Hu et al., 2025) for LeRobot.
+
+RaC improves upon standard data collection (BC) and prior human-in-the-loop methods
+(DAgger, HG-DAgger) by explicitly collecting recovery and correction behaviors:
+
+The workflow:
+1. Policy runs autonomously
+2. Press SPACE to pause - robot holds position
+3. Press 'c' to take control - human provides RECOVERY + CORRECTION
+4. Press → to end episode (save and continue to next)
+5. Reset, then do next rollout
+
+Key RaC Rules:
+- Rule 1 (Recover then Correct): Every intervention = recovery + correction (both human)
+- Rule 2 (Terminate after Intervention): Episode ends after correction
+
+The recovery segment (teleoperating back to good state) is recorded as training data -
+this teaches the policy how to recover from errors.
+
+Keyboard Controls:
+    SPACE  - Pause policy (robot holds position, no recording)
+    c      - Take control (start correction, recording resumes)
+    →      - End episode (save and continue to next)
+    ←      - Re-record episode
+    ESC    - Stop recording and push dataset to hub
+
+Usage:
+    python examples/rac/rac_data_collection.py \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58760431541 \
+        --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --teleop.type=so100_leader \
+        --teleop.port=/dev/tty.usbmodem58760431551 \
+        --policy.path=outputs/train/my_policy/checkpoints/last/pretrained_model \
+        --dataset.repo_id=my_user/rac_dataset \
+        --dataset.single_task="Pick up the cube"
+"""
+
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from pprint import pformat
+from typing import Any
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.image_writer import safe_stop_image_writer
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts
+from lerobot.datasets.video_utils import VideoEncodingManager
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.utils import make_robot_action
+from lerobot.processor import (
+    IdentityProcessor,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.processor.rename_processor import rename_stats
+from lerobot.robots import Robot, RobotConfig, make_robot_from_config
+from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.control_utils import is_headless, predict_action
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import get_safe_torch_device, init_logging, log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+
+@dataclass
+class RaCDatasetConfig:
+    repo_id: str
+    single_task: str
+    root: str | Path | None = None
+    fps: int = 30
+    episode_time_s: float = 120
+    reset_time_s: float = 30
+    num_episodes: int = 50
+    video: bool = True
+    push_to_hub: bool = True
+    private: bool = False
+    tags: list[str] | None = None
+    num_image_writer_processes: int = 0
+    num_image_writer_threads_per_camera: int = 4
+    video_encoding_batch_size: int = 1
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+
+@dataclass
+class RaCConfig:
+    robot: RobotConfig
+    dataset: RaCDatasetConfig
+    policy: PreTrainedConfig
+    teleop: TeleoperatorConfig
+    display_data: bool = True
+    play_sounds: bool = True
+    resume: bool = False
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+def init_rac_keyboard_listener():
+    """Initialize keyboard listener with RaC-specific controls."""
+    events = {
+        "exit_early": False,
+        "rerecord_episode": False,
+        "stop_recording": False,
+        "policy_paused": False,      # SPACE pressed - policy paused, teleop tracking robot
+        "correction_active": False,  # 'c' pressed - human controlling, recording correction
+        "in_reset": False,           # True during reset period
+        "start_next_episode": False, # Signal to start next episode
+    }
+
+    if is_headless():
+        logging.warning("Headless environment - keyboard controls unavailable")
+        return None, events
+
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if events["in_reset"]:
+                # During reset: any action key starts next episode
+                if key == keyboard.Key.space or key == keyboard.Key.right:
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["start_next_episode"] = True
+            else:
+                # During episode
+                if key == keyboard.Key.space:
+                    if not events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ⏸ PAUSED - Policy stopped, teleop moving to robot position")
+                        print("      Press 'c' or START to take control")
+                        events["policy_paused"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    if events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ▶ START pressed - taking control")
+                        events["start_next_episode"] = True
+                elif key == keyboard.Key.right:
+                    print("[RaC] → End episode")
+                    events["exit_early"] = True
+                elif key == keyboard.Key.left:
+                    print("[RaC] ← Re-record episode")
+                    events["rerecord_episode"] = True
+                    events["exit_early"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["exit_early"] = True
+        except Exception as e:
+            print(f"Key error: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+    
+    start_pedal_listener(events)
+    
+    return listener, events
+
+
+def start_pedal_listener(events: dict):
+    """Start foot pedal listener thread if evdev is available."""
+    import threading
+    
+    try:
+        from evdev import InputDevice, ecodes
+    except ImportError:
+        logging.info("[Pedal] evdev not installed - pedal support disabled")
+        return
+    
+    PEDAL_DEVICE = "/dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd"
+    KEY_LEFT = "KEY_A"   # Left pedal
+    KEY_RIGHT = "KEY_C"  # Right pedal
+    
+    def pedal_reader():
+        try:
+            dev = InputDevice(PEDAL_DEVICE)
+            print(f"[Pedal] Connected: {dev.name}")
+            print(f"[Pedal] Right=pause/next, Left=take control/start")
+            
+            for ev in dev.read_loop():
+                if ev.type != ecodes.EV_KEY:
+                    continue
+                
+                from evdev import categorize
+                key = categorize(ev)
+                code = key.keycode
+                if isinstance(code, (list, tuple)):
+                    code = code[0]
+                
+                # Only trigger on key down
+                if key.keystate != 1:
+                    continue
+                
+                if events["in_reset"]:
+                    # During reset: either pedal starts next episode
+                    if code in [KEY_LEFT, KEY_RIGHT]:
+                        print("\n[Pedal] Starting next episode...")
+                        events["start_next_episode"] = True
+                else:
+                    # During episode
+                    if code == KEY_RIGHT:
+                        # Right pedal: SPACE (pause) when running, → (next) when in correction
+                        if events["correction_active"]:
+                            print("\n[Pedal] → End episode")
+                            events["exit_early"] = True
+                        elif not events["policy_paused"]:
+                            print("\n[Pedal] ⏸ PAUSED - Policy stopped, teleop moving to robot")
+                            print("        Press left pedal to take control")
+                            events["policy_paused"] = True
+                    
+                    elif code == KEY_LEFT:
+                        # Left pedal: START (take control) when paused
+                        if events["policy_paused"] and not events["correction_active"]:
+                            print("\n[Pedal] ▶ START pressed - taking control")
+                            events["start_next_episode"] = True
+                        
+        except FileNotFoundError:
+            logging.info(f"[Pedal] Device not found: {PEDAL_DEVICE}")
+        except PermissionError:
+            logging.warning(f"[Pedal] Permission denied. Run: sudo setfacl -m u:$USER:rw {PEDAL_DEVICE}")
+        except Exception as e:
+            logging.debug(f"[Pedal] Error: {e}")
+    
+    thread = threading.Thread(target=pedal_reader, daemon=True)
+    thread.start()
+
+
+def make_identity_processors():
+    """Create identity processors for RaC recording."""
+    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessor()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    robot_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessor()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[IdentityProcessor()],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+    return teleop_proc, robot_proc, obs_proc
+
+
+def move_robot_to_zero(robot: Robot, duration_s: float = 2.0, fps: int = 50):
+    """Smoothly move all robot joints to zero position."""
+    obs = robot.get_observation()
+    current_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
+    target_pos = {k: 0.0 for k in current_pos}
+    
+    print(f"[RaC] Moving robot to zero position ({duration_s}s)...")
+    steps = int(duration_s * fps)
+    for step in range(steps + 1):
+        t = step / steps
+        interp_pos = {k: current_pos[k] * (1 - t) + target_pos[k] * t for k in current_pos}
+        robot.send_action(interp_pos)
+        time.sleep(1 / fps)
+    print("[RaC] Robot at zero position.")
+
+@safe_stop_image_writer
+def rac_rollout_loop(
+    robot: Robot,
+    teleop: Teleoperator,
+    policy: PreTrainedPolicy,
+    preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction],
+    dataset: LeRobotDataset,
+    events: dict,
+    fps: int,
+    control_time_s: float,
+    single_task: str,
+    display_data: bool = True,
+) -> dict:
+    """
+    RaC rollout loop with two-stage intervention:
+    
+    1. Policy runs autonomously (recording)
+    2. SPACE: Policy pauses (NOT recording) - robot holds position
+    3. 'c': Human takes control (recording correction)
+    4. →: End episode
+    """
+    policy.reset()
+    preprocessor.reset()
+    postprocessor.reset()
+
+    device = get_safe_torch_device(policy.config.device)
+    frame_buffer = []
+
+    stats = {
+        "total_frames": 0,
+        "autonomous_frames": 0,
+        "paused_frames": 0,
+        "correction_frames": 0,
+    }
+
+    last_robot_action = None
+    was_paused = False
+    was_correction_active = False
+    waiting_for_takeover = False
+    timestamp = 0
+    start_t = time.perf_counter()
+
+    while timestamp < control_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["policy_paused"] = False
+            events["correction_active"] = False
+            break
+
+        # Detect transition to paused state
+        if events["policy_paused"] and not was_paused:
+            obs = robot.get_observation()
+            robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
+            print("[RaC] Moving teleop to robot position (2s smooth transition)...")
+            teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+            print("[RaC] Teleop aligned. Press START to take control.")
+            events["start_next_episode"] = False
+            waiting_for_takeover = True
+            was_paused = True
+
+        # Wait for start button before enabling correction mode
+        if waiting_for_takeover and events["start_next_episode"]:
+            print("[RaC] Start pressed - enabling teleop control...")
+            events["start_next_episode"] = False
+            events["correction_active"] = True
+            waiting_for_takeover = False
+            was_correction_active = True
+
+        obs = robot.get_observation()
+        obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
+
+        if events["correction_active"]:
+            # Human controlling - record correction data
+            robot_action = teleop.get_action()
+            robot.send_action(robot_action)
+            stats["correction_frames"] += 1
+            
+            # Record this frame
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame = {**obs_frame, **action_frame, "task": single_task}
+            frame_buffer.append(frame)
+            stats["total_frames"] += 1
+            
+        elif waiting_for_takeover:
+            # Waiting for START - policy stopped, no recording, robot holds position
+            if last_robot_action is not None:
+                robot.send_action(last_robot_action)
+            stats["paused_frames"] += 1
+            
+        elif events["policy_paused"]:
+            # Paused and user acknowledged - hold last position, don't record
+            if last_robot_action is not None:
+                robot.send_action(last_robot_action)
+            stats["paused_frames"] += 1
+            robot_action = last_robot_action
+            
+        else:
+            # Normal policy execution - record
+            action_values = predict_action(
+                observation=obs_frame,
+                policy=policy,
+                device=device,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                use_amp=policy.config.use_amp,
+                task=single_task,
+                robot_type=robot.robot_type,
+            )
+            robot_action: RobotAction = make_robot_action(action_values, dataset.features)
+            robot.send_action(robot_action)
+            last_robot_action = robot_action
+            stats["autonomous_frames"] += 1
+            
+            # Record this frame
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame = {**obs_frame, **action_frame, "task": single_task}
+            frame_buffer.append(frame)
+            stats["total_frames"] += 1
+
+        if display_data and robot_action is not None:
+            log_rerun_data(observation=obs, action=robot_action)
+
+        dt = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt)
+        timestamp = time.perf_counter() - start_t
+
+    for frame in frame_buffer:
+        dataset.add_frame(frame)
+
+    return stats
+
+
+def reset_loop(
+    robot: Robot,
+    teleop: Teleoperator,
+    events: dict,
+    fps: int,
+):
+    """Reset period where human repositions environment. Two-stage: enable teleop, then start episode."""
+    print("\n" + "=" * 65)
+    print("  [RaC] RESET - Moving teleop to robot position...")
+    print("=" * 65)
+    
+    # Enter reset mode
+    events["in_reset"] = True
+    events["start_next_episode"] = False
+    
+    # Move teleop to match robot position to avoid sudden jumps
+    obs = robot.get_observation()
+    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
+    teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+    
+    # Stage 1: Wait for user to press start to enable teleoperation
+    print("  Teleop aligned. Press any key/pedal to enable teleoperation")
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        precise_sleep(0.05)
+    
+    if events["stop_recording"]:
+        return
+    
+    # Stage 2: Enable teleop and let user move robot to starting position
+    events["start_next_episode"] = False
+    teleop.disable_torque()
+    print("  Teleop enabled - move robot to starting position")
+    print("  Press any key/pedal to start next episode")
+
+    # Wait for user to signal ready for next episode
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        loop_start = time.perf_counter()
+
+        action = teleop.get_action()
+        robot.send_action(action)
+
+        dt = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt)
+    
+    # Exit reset mode and clear flags for next episode
+    events["in_reset"] = False
+    events["start_next_episode"] = False
+    events["exit_early"] = False
+    events["policy_paused"] = False
+    events["correction_active"] = False
+
+
+@parser.wrap()
+def rac_collect(cfg: RaCConfig) -> LeRobotDataset:
+    """Main RaC data collection function."""
+    init_logging()
+    logging.info(pformat(cfg.__dict__))
+
+    if cfg.display_data:
+        init_rerun(session_name="rac_collection")
+
+    robot = make_robot_from_config(cfg.robot)
+    teleop = make_teleoperator_from_config(cfg.teleop)
+
+    teleop_proc, robot_proc, obs_proc = make_identity_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    dataset = None
+    listener = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+            if hasattr(robot, "cameras") and robot.cameras:
+                dataset.start_image_writer(
+                    num_processes=cfg.dataset.num_image_writer_processes,
+                    num_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
+                )
+        else:
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot.cameras if hasattr(robot, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+
+        policy = make_policy(cfg.policy, ds_meta=dataset.meta)
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.policy.device},
+                "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
+
+        robot.connect()
+        teleop.connect()
+        listener, events = init_rac_keyboard_listener()
+
+        print("\n" + "=" * 65)
+        print("  RaC (Recovery and Correction) Data Collection")
+        print("=" * 65)
+        print("  Policy runs autonomously until you intervene.")
+        print()
+        print("  Controls:")
+        print("    SPACE  - Pause policy (robot holds position, no recording)")
+        print("    c      - Take control (start correction, recording)")
+        print("    →      - End episode (save)")
+        print("    ←      - Re-record episode")
+        print("    ESC    - Stop session and push to hub")
+        print("=" * 65 + "\n")
+
+        with VideoEncodingManager(dataset):
+            recorded = 0
+            while recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                log_say(f"RaC episode {dataset.num_episodes}", cfg.play_sounds)
+                
+                move_robot_to_zero(robot, duration_s=2.0, fps=cfg.dataset.fps)
+
+                stats = rac_rollout_loop(
+                    robot=robot,
+                    teleop=teleop,
+                    policy=policy,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    dataset=dataset,
+                    events=events,
+                    fps=cfg.dataset.fps,
+                    control_time_s=cfg.dataset.episode_time_s,
+                    single_task=cfg.dataset.single_task,
+                    display_data=cfg.display_data,
+                )
+
+                logging.info(f"Episode stats: {stats}")
+
+                if events["rerecord_episode"]:
+                    log_say("Re-recording", cfg.play_sounds)
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                dataset.save_episode()
+                recorded += 1
+
+                # Reset between episodes
+                if recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                    reset_loop(
+                        robot=robot,
+                        teleop=teleop,
+                        events=events,
+                        fps=cfg.dataset.fps,
+                    )
+
+    finally:
+        log_say("Stop recording", cfg.play_sounds, blocking=True)
+
+        if dataset:
+            dataset.finalize()
+
+        if robot.is_connected:
+            robot.disconnect()
+        if teleop.is_connected:
+            teleop.disconnect()
+
+        if not is_headless() and listener:
+            listener.stop()
+
+        if cfg.dataset.push_to_hub:
+            dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+def main():
+    from lerobot.utils.import_utils import register_third_party_plugins
+
+    register_third_party_plugins()
+    rac_collect()
+
+
+if __name__ == "__main__":
+    main()
+

examples/rac/rac_data_collection_openarms.py

@@ -0,0 +1,659 @@
+#!/usr/bin/env python
+"""
+RaC (Recovery and Correction) Data Collection for OpenArms Robot.
+
+This implements the RaC paradigm from "RaC: Robot Learning for Long-Horizon Tasks
+by Scaling Recovery and Correction" (Hu et al., 2025) for LeRobot with OpenArms.
+
+RaC improves upon standard data collection (BC) and prior human-in-the-loop methods
+(DAgger, HG-DAgger) by explicitly collecting recovery and correction behaviors:
+
+The workflow:
+1. Policy runs autonomously (teleop is idle/free)
+2. Press SPACE to pause - teleop moves to match robot position
+3. Press 'c' to take control - teleop is free, human provides RECOVERY + CORRECTION
+4. Press → to end episode (save and continue to next)
+5. Reset, then do next rollout
+
+Key RaC Rules:
+- Rule 1 (Recover then Correct): Every intervention = recovery + correction (both human)
+- Rule 2 (Terminate after Intervention): Episode ends after correction
+
+The recovery segment (teleoperating back to good state) is recorded as training data -
+this teaches the policy how to recover from errors.
+
+Keyboard Controls:
+    SPACE  - Pause policy (teleop mirrors robot, no recording)
+    c      - Take control (teleop free, recording correction)
+    →      - End episode (save and continue to next)
+    ←      - Re-record episode
+    ESC    - Stop recording and push dataset to hub
+
+Usage:
+    python examples/rac/rac_data_collection_openarms.py \
+        --robot.type=openarms_follower \
+        --robot.port_right=can0 \
+        --robot.port_left=can1 \
+        --robot.cameras="{ left_wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, right_wrist: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30}}" \
+        --teleop.type=openarms_mini \
+        --teleop.port_right=/dev/ttyUSB0 \
+        --teleop.port_left=/dev/ttyUSB1 \
+        --policy.path=outputs/train/my_policy/checkpoints/last/pretrained_model \
+        --dataset.repo_id=my_user/rac_openarms_dataset \
+        --dataset.single_task="Pick up the cube"
+"""
+
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from pprint import pformat
+from typing import Any
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.image_writer import safe_stop_image_writer
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts
+from lerobot.datasets.video_utils import VideoEncodingManager
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.utils import make_robot_action
+from lerobot.processor import (
+    IdentityProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.processor.rename_processor import rename_stats
+from lerobot.robots import Robot, RobotConfig, make_robot_from_config
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig  # noqa: F401
+from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
+from lerobot.teleoperators.openarms_mini.config_openarms_mini import OpenArmsMiniConfig  # noqa: F401
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.control_utils import is_headless, predict_action
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import get_safe_torch_device, init_logging, log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+
+@dataclass
+class RaCDatasetConfig:
+    repo_id: str
+    single_task: str
+    root: str | Path | None = None
+    fps: int = 30
+    episode_time_s: float = 120
+    reset_time_s: float = 30
+    num_episodes: int = 50
+    video: bool = True
+    push_to_hub: bool = True
+    private: bool = False
+    tags: list[str] | None = None
+    num_image_writer_processes: int = 0
+    num_image_writer_threads_per_camera: int = 4
+    video_encoding_batch_size: int = 1
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+
+@dataclass
+class RaCConfig:
+    robot: RobotConfig
+    dataset: RaCDatasetConfig
+    teleop: TeleoperatorConfig
+    policy: PreTrainedConfig | None = None
+    display_data: bool = True
+    play_sounds: bool = True
+    resume: bool = False
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        if self.policy is None:
+            raise ValueError("policy.path is required")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:  
+        return ["policy"]
+
+
+def init_rac_keyboard_listener():
+    """Initialize keyboard listener with RaC-specific controls."""
+    events = {
+        "exit_early": False,
+        "rerecord_episode": False,
+        "stop_recording": False,
+        "policy_paused": False,      # SPACE pressed - policy paused, teleop tracking robot
+        "correction_active": False,  # 'c' pressed - human controlling, recording correction
+        "in_reset": False,           # True during reset period
+        "start_next_episode": False, # Signal to start next episode
+    }
+
+    if is_headless():
+        logging.warning("Headless environment - keyboard controls unavailable")
+        return None, events
+
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if events["in_reset"]:
+                # During reset: any action key starts next episode
+                if key == keyboard.Key.space or key == keyboard.Key.right:
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["start_next_episode"] = True
+            else:
+                # During episode
+                if key == keyboard.Key.space:
+                    if not events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ⏸ PAUSED - Policy stopped, teleop moving to robot position")
+                        print("      Press 'c' or START to take control")
+                        events["policy_paused"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    if events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ▶ START pressed - taking control")
+                        events["start_next_episode"] = True
+                elif key == keyboard.Key.right:
+                    print("[RaC] → End episode")
+                    events["exit_early"] = True
+                elif key == keyboard.Key.left:
+                    print("[RaC] ← Re-record episode")
+                    events["rerecord_episode"] = True
+                    events["exit_early"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["exit_early"] = True
+        except Exception as e:
+            print(f"Key error: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+    
+    start_pedal_listener(events)
+    
+    return listener, events
+
+
+def start_pedal_listener(events: dict):
+    """Start foot pedal listener thread if evdev is available."""
+    import threading
+    
+    try:
+        from evdev import InputDevice, ecodes
+    except ImportError:
+        logging.info("[Pedal] evdev not installed - pedal support disabled")
+        return
+    
+    PEDAL_DEVICE = "/dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd"
+    KEY_LEFT = "KEY_A"   # Left pedal
+    KEY_RIGHT = "KEY_C"  # Right pedal
+    
+    def pedal_reader():
+        try:
+            dev = InputDevice(PEDAL_DEVICE)
+            print(f"[Pedal] Connected: {dev.name}")
+            print(f"[Pedal] Right=pause/next, Left=take control/start")
+            
+            for ev in dev.read_loop():
+                if ev.type != ecodes.EV_KEY:
+                    continue
+                
+                from evdev import categorize
+                key = categorize(ev)
+                code = key.keycode
+                if isinstance(code, (list, tuple)):
+                    code = code[0]
+                
+                # Only trigger on key down
+                if key.keystate != 1:
+                    continue
+                
+                if events["in_reset"]:
+                    # During reset: either pedal starts next episode
+                    if code in [KEY_LEFT, KEY_RIGHT]:
+                        print("\n[Pedal] Starting next episode...")
+                        events["start_next_episode"] = True
+                else:
+                    # During episode
+                    if code == KEY_RIGHT:
+                        # Right pedal: SPACE (pause) when running, → (next) when in correction
+                        if events["correction_active"]:
+                            print("\n[Pedal] → End episode")
+                            events["exit_early"] = True
+                        elif not events["policy_paused"]:
+                            print("\n[Pedal] ⏸ PAUSED - Policy stopped, teleop moving to robot")
+                            print("        Press left pedal to take control")
+                            events["policy_paused"] = True
+                    
+                    elif code == KEY_LEFT:
+                        # Left pedal: START (take control) when paused
+                        if events["policy_paused"] and not events["correction_active"]:
+                            print("\n[Pedal] ▶ START pressed - taking control")
+                            events["start_next_episode"] = True
+                        
+        except FileNotFoundError:
+            logging.info(f"[Pedal] Device not found: {PEDAL_DEVICE}")
+        except PermissionError:
+            logging.warning(f"[Pedal] Permission denied. Run: sudo setfacl -m u:$USER:rw {PEDAL_DEVICE}")
+        except Exception as e:
+            logging.debug(f"[Pedal] Error: {e}")
+    
+    thread = threading.Thread(target=pedal_reader, daemon=True)
+    thread.start()
+
+
+def make_identity_processors():
+    """Create identity processors for RaC recording."""
+    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    robot_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+    return teleop_proc, robot_proc, obs_proc
+
+
+def move_robot_to_zero(robot: Robot, duration_s: float = 2.0, fps: int = 50):
+    """Smoothly move all robot joints to zero position."""
+    obs = robot.get_observation()
+    current_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
+    target_pos = {k: 0.0 for k in current_pos}
+    
+    print(f"[RaC] Moving robot to zero position ({duration_s}s)...")
+    steps = int(duration_s * fps)
+    for step in range(steps + 1):
+        t = step / steps
+        interp_pos = {k: current_pos[k] * (1 - t) + target_pos[k] * t for k in current_pos}
+        robot.send_action(interp_pos)
+        time.sleep(1 / fps)
+    print("[RaC] Robot at zero position.")
+
+
+@safe_stop_image_writer
+def rac_rollout_loop(
+    robot: Robot,
+    teleop: Teleoperator,
+    policy: PreTrainedPolicy,
+    preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction],
+    dataset: LeRobotDataset,
+    events: dict,
+    fps: int,
+    control_time_s: float,
+    single_task: str,
+    display_data: bool = True,
+) -> dict:
+    """
+    RaC rollout loop with two-stage intervention:
+    
+    1. Policy runs autonomously (recording) - teleop free/idle
+    2. SPACE: Policy pauses, teleop mirrors robot position (NOT recording)
+    3. 'c': Human takes control, teleop torque disabled (recording correction)
+    4. →: End episode
+    
+    This allows smooth handoff - teleop tracks robot only when paused.
+    """
+    policy.reset()
+    preprocessor.reset()
+    postprocessor.reset()
+
+    device = get_safe_torch_device(policy.config.device)
+    frame_buffer = []
+
+    stats = {
+        "total_frames": 0,
+        "autonomous_frames": 0,
+        "paused_frames": 0,
+        "correction_frames": 0,
+    }
+
+    # Start with teleop torque disabled - only enable when paused to track robot
+    teleop.disable_torque()
+    was_paused = False
+    was_correction_active = False
+    waiting_for_takeover = False
+
+    timestamp = 0
+    start_t = time.perf_counter()
+
+    while timestamp < control_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["policy_paused"] = False
+            events["correction_active"] = False
+            break
+
+        # Detect transition to paused state - smooth move teleop to robot position
+        if events["policy_paused"] and not was_paused:
+            obs = robot.get_observation()
+            obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+            robot_pos = {k: v for k, v in obs_filtered.items() if k.endswith(".pos")}
+            print("[RaC] Moving teleop to robot position (2s smooth transition)...")
+            teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+            print("[RaC] Teleop aligned. Press START to take control.")
+            events["start_next_episode"] = False
+            waiting_for_takeover = True
+            was_paused = True
+
+        # Wait for start button before enabling correction mode
+        if waiting_for_takeover and events["start_next_episode"]:
+            print("[RaC] Start pressed - enabling teleop control...")
+            teleop.disable_torque()
+            events["start_next_episode"] = False
+            events["correction_active"] = True
+            waiting_for_takeover = False
+            was_correction_active = True
+
+        obs = robot.get_observation()
+        obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+        obs_frame = build_dataset_frame(dataset.features, obs_filtered, prefix=OBS_STR)
+
+        if events["correction_active"]:
+            # Human controlling - record correction data
+            robot_action = teleop.get_action()
+            # Convert gripper from teleop range (0-100) to robot degrees (-65 to 0)
+            for key in robot_action:
+                if "gripper" in key:
+                    robot_action[key] = -0.65 * robot_action[key]
+            robot.send_action(robot_action)
+            stats["correction_frames"] += 1
+            
+            # Record this frame
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame = {**obs_frame, **action_frame, "task": single_task}
+            frame_buffer.append(frame)
+            stats["total_frames"] += 1
+            
+        elif waiting_for_takeover:
+            # Waiting for START - policy stopped, no recording, robot holds position
+            stats["paused_frames"] += 1
+            
+        elif events["policy_paused"]:
+            # Paused and user acknowledged - teleop tracks robot position, don't record
+            robot_action = {k: v for k, v in obs_filtered.items() if k.endswith(".pos")}
+            teleop.send_feedback(robot_action)
+            stats["paused_frames"] += 1
+            
+        else:
+            # Normal policy execution - record (teleop is free/idle)
+            action_values = predict_action(
+                observation=obs_frame,
+                policy=policy,
+                device=device,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                use_amp=policy.config.use_amp,
+                task=single_task,
+                robot_type=robot.robot_type,
+            )
+            robot_action: RobotAction = make_robot_action(action_values, dataset.features)
+            robot.send_action(robot_action)
+            stats["autonomous_frames"] += 1
+            
+            # Record this frame
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame = {**obs_frame, **action_frame, "task": single_task}
+            frame_buffer.append(frame)
+            stats["total_frames"] += 1
+
+        if display_data:
+            log_rerun_data(observation=obs_filtered, action=robot_action)
+
+        dt = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt)
+        timestamp = time.perf_counter() - start_t
+
+    # Ensure teleoperator torque is disabled at end
+    teleop.disable_torque()
+
+    for frame in frame_buffer:
+        dataset.add_frame(frame)
+
+    return stats
+
+
+def reset_loop(
+    robot: Robot,
+    teleop: Teleoperator,
+    events: dict,
+    fps: int,
+):
+    """Reset period where human repositions environment. Two-stage: enable teleop, then start episode."""
+    print("\n" + "=" * 65)
+    print("  [RaC] RESET - Moving teleop to robot position...")
+    print("=" * 65)
+    
+    # Enter reset mode
+    events["in_reset"] = True
+    events["start_next_episode"] = False
+    
+    # First move teleop to match robot position to avoid sudden jumps
+    obs = robot.get_observation()
+    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
+    teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+    
+    # Stage 1: Wait for user to press start to enable teleoperation
+    print("  Teleop aligned. Press any key/pedal to enable teleoperation")
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        precise_sleep(0.05)
+    
+    if events["stop_recording"]:
+        return
+    
+    # Stage 2: Enable teleop and let user move robot to starting position
+    events["start_next_episode"] = False
+    teleop.disable_torque()
+    print("  Teleop enabled - move robot to starting position")
+    print("  Press any key/pedal to start next episode")
+
+    # Wait for user to signal ready for next episode
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        loop_start = time.perf_counter()
+
+        action = teleop.get_action()
+        # Convert gripper from teleop range (0-100) to robot degrees (-65 to 0)
+        for key in action:
+            if "gripper" in key:
+                action[key] = -0.65 * action[key]
+        robot.send_action(action)
+
+        dt = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt)
+    
+    # Exit reset mode and clear flags for next episode
+    events["in_reset"] = False
+    events["start_next_episode"] = False
+    events["exit_early"] = False
+    events["policy_paused"] = False
+    events["correction_active"] = False
+
+
+@parser.wrap()
+def rac_collect(cfg: RaCConfig) -> LeRobotDataset:
+    """Main RaC data collection function."""
+    init_logging()
+    logging.info(pformat(cfg.__dict__))
+
+    if cfg.display_data:
+        init_rerun(session_name="rac_collection_openarms")
+
+    robot = make_robot_from_config(cfg.robot)
+    teleop = make_teleoperator_from_config(cfg.teleop)
+
+    teleop_proc, robot_proc, obs_proc = make_identity_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    dataset = None
+    listener = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+            if hasattr(robot, "cameras") and robot.cameras:
+                dataset.start_image_writer(
+                    num_processes=cfg.dataset.num_image_writer_processes,
+                    num_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
+                )
+        else:
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot.cameras if hasattr(robot, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+
+        policy = make_policy(cfg.policy, ds_meta=dataset.meta)
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.policy.device},
+                "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
+
+        robot.connect()
+        teleop.connect()
+        listener, events = init_rac_keyboard_listener()
+
+        print("\n" + "=" * 65)
+        print("  RaC (Recovery and Correction) Data Collection - OpenArms")
+        print("=" * 65)
+        print("  Policy runs autonomously until you intervene.")
+        print()
+        print("  Controls:")
+        print("    SPACE  - Pause policy (teleop tracks robot, no recording)")
+        print("    c      - Take control (start correction, recording)")
+        print("    →      - End episode (save)")
+        print("    ←      - Re-record episode")
+        print("    ESC    - Stop session and push to hub")
+        print("=" * 65 + "\n")
+
+        with VideoEncodingManager(dataset):
+            recorded = 0
+            while recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                log_say(f"RaC episode {dataset.num_episodes}", cfg.play_sounds)
+                
+                move_robot_to_zero(robot, duration_s=2.0, fps=cfg.dataset.fps)
+
+                stats = rac_rollout_loop(
+                    robot=robot,
+                    teleop=teleop,
+                    policy=policy,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    dataset=dataset,
+                    events=events,
+                    fps=cfg.dataset.fps,
+                    control_time_s=cfg.dataset.episode_time_s,
+                    single_task=cfg.dataset.single_task,
+                    display_data=cfg.display_data,
+                )
+
+                logging.info(f"Episode stats: {stats}")
+
+                if events["rerecord_episode"]:
+                    log_say("Re-recording", cfg.play_sounds)
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                dataset.save_episode()
+                recorded += 1
+
+                # Reset between episodes
+                if recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                    reset_loop(
+                        robot=robot,
+                        teleop=teleop,
+                        events=events,
+                        fps=cfg.dataset.fps,
+                    )
+
+    finally:
+        log_say("Stop recording", cfg.play_sounds, blocking=True)
+
+        if dataset:
+            dataset.finalize()
+
+        if robot.is_connected:
+            robot.disconnect()
+        if teleop.is_connected:
+            teleop.disconnect()
+
+        if not is_headless() and listener:
+            listener.stop()
+
+        if cfg.dataset.push_to_hub:
+            dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+def main():
+    from lerobot.utils.import_utils import register_third_party_plugins
+
+    register_third_party_plugins()
+    rac_collect()
+
+
+if __name__ == "__main__":
+    main()
+

examples/rac/rac_data_collection_openarms_rtc.py

@@ -0,0 +1,902 @@
+#!/usr/bin/env python
+"""
+RaC (Recovery and Correction) Data Collection for OpenArms Robot with RTC.
+
+This combines RaC data collection with Real-Time Chunking (RTC) for smooth policy execution.
+RTC enables large flow-matching policies (Pi0, Pi0.5, SmolVLA) to produce reactive motion
+despite high inference latency by asynchronously generating action chunks.
+
+The workflow:
+1. Policy runs autonomously with RTC (teleop is idle/free)
+2. Press SPACE to pause - teleop moves to match robot position
+3. Press 'c' to take control - teleop is free, human provides RECOVERY + CORRECTION
+4. Press → to end episode (save and continue to next)
+5. Reset, then do next rollout
+
+Usage:
+    python examples/rac/rac_data_collection_openarms_rtc.py \
+        --robot.port_right=can0 \
+        --robot.port_left=can1 \
+        --teleop.port_right=/dev/ttyUSB0 \
+        --teleop.port_left=/dev/ttyUSB1 \
+        --policy.path=outputs/train/my_policy/checkpoints/last/pretrained_model \
+        --dataset.repo_id=my_user/rac_openarms_dataset \
+        --dataset.single_task="Pick up the cube"
+"""
+
+import logging
+import math
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from pprint import pformat
+from threading import Event, Lock, Thread
+from typing import Any
+
+import torch
+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.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.image_writer import safe_stop_image_writer
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.datasets.video_utils import VideoEncodingManager
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.policies.utils import make_robot_action
+from lerobot.processor import (
+    IdentityProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.processor.rename_processor import rename_stats
+from lerobot.robots import Robot, RobotConfig, make_robot_from_config
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig  # noqa: F401
+from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
+from lerobot.teleoperators.openarms_mini.config_openarms_mini import OpenArmsMiniConfig  # noqa: F401
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.control_utils import is_headless, predict_action
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import get_safe_torch_device, init_logging, log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+# ============================================================================
+# Configuration
+# ============================================================================
+
+@dataclass
+class RaCRTCDatasetConfig:
+    repo_id: str = "lerobot/rac_openarms_rtc"
+    single_task: str = "default task"
+    root: str | Path | None = None
+    fps: int = 30
+    episode_time_s: float = 500
+    reset_time_s: float = 30
+    num_episodes: int = 50
+    video: bool = True
+    push_to_hub: bool = True
+    private: bool = False
+    tags: list[str] | None = None
+    num_image_writer_processes: int = 0
+    num_image_writer_threads_per_camera: int = 4
+    video_encoding_batch_size: int = 1
+    streaming_encoding: bool = True
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+
+@dataclass
+class RaCRTCConfig:
+    robot: RobotConfig = field(default_factory=lambda: OpenArmsFollowerConfig(
+        port_left="can0",
+        port_right="can1",
+    ))
+    teleop: TeleoperatorConfig = field(default_factory=lambda: OpenArmsMiniConfig(
+        port_left="/dev/ttyUSB1",
+        port_right="/dev/ttyUSB0",
+    ))
+    dataset: RaCRTCDatasetConfig = field(default_factory=RaCRTCDatasetConfig)
+    policy: PreTrainedConfig | None = None
+    
+    rtc: RTCConfig = field(default_factory=lambda: RTCConfig(
+        enabled=True, 
+        execution_horizon=20,
+        max_guidance_weight=5.0,
+        prefix_attention_schedule=RTCAttentionSchedule.LINEAR,
+    ))
+    
+    interpolation: bool = True
+    display_data: bool = True
+    play_sounds: bool = True
+    resume: bool = False
+    device: str = "cuda"
+    action_queue_size_to_get_new_actions: int = 30
+    
+    # Torch compile is disabled by default for real-time inference
+    # First inference with compile takes minutes to compile kernels
+    use_torch_compile: bool = False
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        if self.policy is None:
+            raise ValueError("policy.path is required")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+# ============================================================================
+# Thread-Safe Robot Wrapper (from evaluate_with_rtc.py)
+# ============================================================================
+
+class RobotWrapper:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: Robot):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self.lock:
+            self.robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        return self.robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        return self.robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self.robot.name
+    
+    @property
+    def robot_type(self) -> str:
+        return self.robot.robot_type
+
+
+# ============================================================================
+# Keyboard/Pedal Listeners
+# ============================================================================
+
+def init_rac_keyboard_listener():
+    """Initialize keyboard listener with RaC-specific controls."""
+    events = {
+        "exit_early": False,
+        "rerecord_episode": False,
+        "stop_recording": False,
+        "policy_paused": False,
+        "correction_active": False,
+        "in_reset": False,
+        "start_next_episode": False,
+    }
+
+    if is_headless():
+        logging.warning("Headless environment - keyboard controls unavailable")
+        return None, events
+
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if events["in_reset"]:
+                if key == keyboard.Key.space or key == keyboard.Key.right:
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["start_next_episode"] = True
+            else:
+                if key == keyboard.Key.space:
+                    if not events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ⏸ PAUSED - Policy stopped, teleop moving to robot position")
+                        print("      Press 'c' or START to take control")
+                        events["policy_paused"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    if events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ▶ START pressed - taking control")
+                        events["start_next_episode"] = True
+                elif key == keyboard.Key.right:
+                    print("[RaC] → End episode")
+                    events["exit_early"] = True
+                elif key == keyboard.Key.left:
+                    print("[RaC] ← Re-record episode")
+                    events["rerecord_episode"] = True
+                    events["exit_early"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["exit_early"] = True
+        except Exception as e:
+            print(f"Key error: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+    
+    start_pedal_listener(events)
+    
+    return listener, events
+
+
+def start_pedal_listener(events: dict):
+    """Start foot pedal listener thread if evdev is available."""
+    import threading
+    
+    try:
+        from evdev import InputDevice, ecodes  # noqa: F401
+    except ImportError:
+        logging.info("[Pedal] evdev not installed - pedal support disabled")
+        return
+    
+    PEDAL_DEVICE = "/dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd"
+    KEY_LEFT = "KEY_A"
+    KEY_RIGHT = "KEY_C"
+    
+    def pedal_reader():
+        try:
+            dev = InputDevice(PEDAL_DEVICE)
+            print(f"[Pedal] Connected: {dev.name}")
+            
+            for ev in dev.read_loop():
+                if ev.type != ecodes.EV_KEY:
+                    continue
+                
+                from evdev import categorize  # noqa: F401
+                key = categorize(ev)
+                code = key.keycode
+                if isinstance(code, (list, tuple)):
+                    code = code[0]
+                
+                if key.keystate != 1:
+                    continue
+                
+                if events["in_reset"]:
+                    if code in [KEY_LEFT, KEY_RIGHT]:
+                        events["start_next_episode"] = True
+                else:
+                    if code == KEY_RIGHT:
+                        if events["correction_active"]:
+                            events["exit_early"] = True
+                        elif not events["policy_paused"]:
+                            events["policy_paused"] = True
+                    elif code == KEY_LEFT:
+                        if events["policy_paused"] and not events["correction_active"]:
+                            events["start_next_episode"] = True
+                        
+        except FileNotFoundError:
+            logging.info(f"[Pedal] Device not found: {PEDAL_DEVICE}")
+        except PermissionError:
+            logging.warning(f"[Pedal] Permission denied for {PEDAL_DEVICE}")
+        except Exception as e:
+            logging.debug(f"[Pedal] Error: {e}")
+    
+    thread = threading.Thread(target=pedal_reader, daemon=True)
+    thread.start()
+
+
+def make_identity_processors():
+    """Create identity processors for RaC recording."""
+    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    robot_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+    return teleop_proc, robot_proc, obs_proc
+
+
+# ============================================================================
+# RTC Inference Thread (from evaluate_with_rtc.py)
+# ============================================================================
+
+def rtc_inference_thread(
+    policy,
+    obs_holder: dict,
+    hw_features: dict,
+    preprocessor,
+    postprocessor,
+    queue_holder: dict,
+    shutdown_event: Event,
+    policy_active: Event,
+    cfg: RaCRTCConfig,
+):
+    """Background thread that generates action chunks using RTC."""
+    try:
+        logger.info("[RTC] ========== INFERENCE THREAD STARTED ==========")
+        logger.info(f"[RTC] policy={policy.name}, hw_features has {len(hw_features)} keys")
+        
+        latency_tracker = LatencyTracker()
+        time_per_chunk = 1.0 / cfg.dataset.fps
+        policy_device = policy.config.device
+        
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+        
+        inference_count = 0
+        wait_logged = False
+        
+        while not shutdown_event.is_set():
+            if not policy_active.is_set():
+                if not wait_logged:
+                    logger.info("[RTC] Waiting for policy_active...")
+                    wait_logged = True
+                time.sleep(0.01)
+                continue
+            
+            wait_logged = False
+            
+            action_queue = queue_holder["queue"]
+            if action_queue is None:
+                logger.warning("[RTC] queue_holder['queue'] is None!")
+                time.sleep(0.01)
+                continue
+            
+            obs_filtered = obs_holder.get("obs")
+            if obs_filtered is None:
+                logger.warning("[RTC] obs_holder['obs'] is None!")
+                time.sleep(0.01)
+                continue
+            
+            qsize = action_queue.qsize()
+            if qsize <= get_actions_threshold:
+                try:
+                    if inference_count == 0:
+                        logger.info(f"[RTC] Starting first inference, obs keys={len(obs_filtered)}, qsize={qsize}")
+                    
+                    current_time = time.perf_counter()
+                    action_index_before_inference = action_queue.get_action_index()
+                    prev_actions = action_queue.get_left_over()
+                    
+                    inference_latency = latency_tracker.max()
+                    inference_delay = math.ceil(inference_latency / time_per_chunk) if inference_latency else 0
+                    
+                    obs_with_policy_features = build_dataset_frame(hw_features, obs_filtered, prefix="observation")
+                    
+                    for name in obs_with_policy_features:
+                        obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                        if "image" in name:
+                            obs_with_policy_features[name] = obs_with_policy_features[name].float() / 255
+                            obs_with_policy_features[name] = obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0).to(policy_device)
+                    
+                    obs_with_policy_features["task"] = [cfg.dataset.single_task]
+                    obs_with_policy_features["robot_type"] = obs_holder.get("robot_type", "openarms_follower")
+                    
+                    preprocessed_obs = preprocessor(obs_with_policy_features)
+                    
+                    actions = policy.predict_action_chunk(
+                        preprocessed_obs,
+                        inference_delay=inference_delay,
+                        prev_chunk_left_over=prev_actions,
+                    )
+                    
+                    original_actions = actions.squeeze(0).clone()
+                    postprocessed_actions = postprocessor(actions).squeeze(0)
+                    
+                    new_latency = time.perf_counter() - current_time
+                    new_delay = math.ceil(new_latency / time_per_chunk)
+                    latency_tracker.add(new_latency)
+                    
+                    action_queue.merge(original_actions, postprocessed_actions, new_delay, action_index_before_inference)
+                    
+                    inference_count += 1
+                    logger.info(f"[RTC] Inference #{inference_count}, latency={new_latency:.2f}s, queue={action_queue.qsize()}")
+                except Exception as e:
+                    logger.error(f"[RTC] Inference error: {e}")
+                    import traceback
+                    traceback.print_exc()
+                    time.sleep(1.0)
+            else:
+                time.sleep(0.01)
+        
+        logger.info("[RTC] Inference thread shutting down")
+    except Exception as e:
+        logger.error(f"[RTC] THREAD CRASHED: {e}")
+        import traceback
+        traceback.print_exc()
+
+
+# ============================================================================
+# Main Rollout Loop
+# ============================================================================
+
+@safe_stop_image_writer
+def rac_rtc_rollout_loop(
+    robot: RobotWrapper,
+    teleop: Teleoperator,
+    policy: PreTrainedPolicy,
+    preprocessor,
+    postprocessor,
+    dataset: LeRobotDataset,
+    events: dict,
+    cfg: RaCRTCConfig,
+    queue_holder: dict,
+    obs_holder: dict,  # Main loop writes obs here for RTC thread to read
+    policy_active: Event,
+    hw_features: dict,
+) -> dict:
+    """RaC rollout loop with RTC for smooth policy execution."""
+    fps = cfg.dataset.fps
+    single_task = cfg.dataset.single_task
+    control_time_s = cfg.dataset.episode_time_s
+    device = get_safe_torch_device(cfg.device)
+    
+    # Reset policy state
+    policy.reset()
+    preprocessor.reset()
+    postprocessor.reset()
+    
+    streaming = dataset._streaming_encoder is not None
+    frame_buffer = [] if not streaming else None
+    stats = {
+        "total_frames": 0,
+        "autonomous_frames": 0,
+        "paused_frames": 0,
+        "correction_frames": 0,
+    }
+
+    teleop.disable_torque()
+    was_paused = False
+    waiting_for_takeover = False
+    
+    # Action keys for converting tensor to dict
+    action_keys = [k for k in robot.action_features.keys() if k.endswith(".pos")]
+    
+    # Interpolation state
+    prev_action: Tensor | None = None
+    interpolated_actions: list[Tensor] = []
+    interp_idx = 0
+    
+    if cfg.interpolation:
+        control_interval = 1.0 / (fps * 2)  # 2x rate
+    else:
+        control_interval = 1.0 / fps
+    
+    robot_action = {}
+    timestamp = 0
+    start_t = time.perf_counter()
+
+    while timestamp < control_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["policy_paused"] = False
+            events["correction_active"] = False
+            break
+
+        # State transition: entering paused state
+        if events["policy_paused"] and not was_paused:
+            policy_active.clear()  # Stop RTC inference
+            obs = robot.get_observation()
+            obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+            robot_pos = {k: v for k, v in obs_filtered.items() if k.endswith(".pos")}
+            print("[RaC] Moving teleop to robot position...")
+            teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+            print("[RaC] Teleop aligned. Press 'c' to take control.")
+            events["start_next_episode"] = False
+            waiting_for_takeover = True
+            was_paused = True
+            # Reset interpolation
+            prev_action = None
+            interpolated_actions = []
+            interp_idx = 0
+
+        # Wait for takeover
+        if waiting_for_takeover and events["start_next_episode"]:
+            print("[RaC] Taking control...")
+            teleop.disable_torque()
+            events["start_next_episode"] = False
+            events["correction_active"] = True
+            waiting_for_takeover = False
+
+        # Get observation (ONLY the main loop reads from robot!)
+        obs = robot.get_observation()
+        obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+        obs_frame = build_dataset_frame(dataset.features, obs_filtered, prefix=OBS_STR)
+        
+        # Share observation with RTC thread (thread reads, main loop writes)
+        obs_holder["obs"] = obs_filtered
+
+        if events["correction_active"]:
+            # Human controlling
+            robot_action = teleop.get_action()
+            for key in robot_action:
+                if "gripper" in key:
+                    robot_action[key] = -0.65 * robot_action[key]
+            robot.send_action(robot_action)
+            stats["correction_frames"] += 1
+            
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame = {**obs_frame, **action_frame, "task": single_task}
+            if streaming:
+                dataset.add_frame(frame)
+            else:
+                frame_buffer.append(frame)
+            stats["total_frames"] += 1
+            
+        elif waiting_for_takeover:
+            stats["paused_frames"] += 1
+            
+        elif events["policy_paused"]:
+            robot_pos = {k: v for k, v in obs_filtered.items() if k.endswith(".pos")}
+            teleop.send_feedback(robot_pos)
+            stats["paused_frames"] += 1
+            
+        else:
+            # Policy execution with RTC
+            if not policy_active.is_set():
+                policy_active.set()
+                logger.info("[ROLLOUT] Policy activated, waiting for first actions...")
+            
+            action_queue = queue_holder["queue"]
+            
+            # Get action from queue (with interpolation)
+            if interp_idx >= len(interpolated_actions):
+                new_action = action_queue.get() if action_queue else None
+                
+                # Log queue status periodically
+                if stats["autonomous_frames"] == 0 and new_action is None:
+                    qsize = action_queue.qsize() if action_queue else -1
+                    if timestamp < 0.5 or int(timestamp * 10) % 10 == 0:
+                        logger.info(f"[ROLLOUT] Waiting for actions... queue_size={qsize}, obs_set={obs_holder.get('obs') is not None}")
+                
+                if new_action is not None:
+                    current_action = new_action.cpu()
+                    
+                    if cfg.interpolation and prev_action is not None:
+                        mid = prev_action + 0.5 * (current_action - prev_action)
+                        interpolated_actions = [mid, current_action]
+                    else:
+                        interpolated_actions = [current_action]
+                    
+                    prev_action = current_action
+                    interp_idx = 0
+                    
+                    if stats["autonomous_frames"] == 0:
+                        logger.info(f"[ROLLOUT] Got first action! Starting robot motion.")
+            
+            if interp_idx < len(interpolated_actions):
+                action_to_send = interpolated_actions[interp_idx]
+                interp_idx += 1
+                
+                robot_action = {}
+                for i, key in enumerate(action_keys):
+                    if i < len(action_to_send):
+                        robot_action[key] = action_to_send[i].item()
+                
+                robot.send_action(robot_action)
+                stats["autonomous_frames"] += 1
+                
+                # Record at original fps
+                action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+                frame = {**obs_frame, **action_frame, "task": single_task}
+                if streaming:
+                    dataset.add_frame(frame)
+                else:
+                    frame_buffer.append(frame)
+                stats["total_frames"] += 1
+
+        if cfg.display_data:
+            log_rerun_data(observation=obs_filtered, action=robot_action)
+
+        dt = time.perf_counter() - loop_start
+        sleep_time = control_interval - dt
+        if sleep_time > 0:
+            precise_sleep(sleep_time)
+        timestamp = time.perf_counter() - start_t
+
+    policy_active.clear()
+    teleop.disable_torque()
+
+    if not streaming:
+        for frame in frame_buffer:
+            dataset.add_frame(frame)
+
+    return stats
+
+
+def reset_loop(robot: RobotWrapper, teleop: Teleoperator, events: dict, fps: int):
+    """Reset period where human repositions environment."""
+    print("\n" + "=" * 65)
+    print("  [RaC] RESET")
+    print("=" * 65)
+    
+    events["in_reset"] = True
+    events["start_next_episode"] = False
+    
+    obs = robot.get_observation()
+    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
+    teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+    
+    print("  Press any key/pedal to enable teleoperation")
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        precise_sleep(0.05)
+    
+    if events["stop_recording"]:
+        return
+    
+    events["start_next_episode"] = False
+    teleop.disable_torque()
+    print("  Teleop enabled - press any key/pedal to start episode")
+
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        loop_start = time.perf_counter()
+        action = teleop.get_action()
+        for key in action:
+            if "gripper" in key:
+                action[key] = -0.65 * action[key]
+        robot.send_action(action)
+        dt = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt)
+    
+    events["in_reset"] = False
+    events["start_next_episode"] = False
+    events["exit_early"] = False
+    events["policy_paused"] = False
+    events["correction_active"] = False
+
+
+# ============================================================================
+# Main Entry Point
+# ============================================================================
+
+@parser.wrap()
+def rac_rtc_collect(cfg: RaCRTCConfig) -> LeRobotDataset:
+    """Main RaC data collection function with RTC."""
+    init_logging()
+    logging.info(pformat(cfg.__dict__))
+
+    if cfg.display_data:
+        init_rerun(session_name="rac_rtc_collection_openarms")
+
+    robot_raw = make_robot_from_config(cfg.robot)
+    teleop = make_teleoperator_from_config(cfg.teleop)
+    
+    teleop_proc, robot_proc, obs_proc = make_identity_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot_raw.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot_raw.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    dataset = None
+    listener = None
+    shutdown_event = Event()
+    policy_active = Event()
+    rtc_thread = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+            if cfg.dataset.streaming_encoding:
+                dataset.start_streaming_encoder()
+            if hasattr(robot_raw, "cameras") and robot_raw.cameras:
+                dataset.start_image_writer(
+                    num_processes=cfg.dataset.num_image_writer_processes,
+                    num_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot_raw.cameras),
+                )
+        else:
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot_raw.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot_raw.cameras if hasattr(robot_raw, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+            )
+
+        # Load policy
+        logger.info(f"Loading policy from: {cfg.policy.pretrained_path}")
+        policy_class = get_policy_class(cfg.policy.type)
+        
+        # Override compile_model for real-time inference (first compile takes minutes)
+        policy_config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+        if cfg.policy.type in ["pi05", "pi0"]:
+            policy_config.compile_model = cfg.use_torch_compile
+            logger.info(f"Set compile_model={cfg.use_torch_compile} for real-time inference")
+        
+        policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=policy_config)
+        policy.config.rtc_config = cfg.rtc
+        policy.init_rtc_processor()
+        policy = policy.to(cfg.device)
+        policy.eval()
+        logger.info(f"Policy loaded: {policy.name}")
+
+        # Setup preprocessor/postprocessor
+        hw_features = hw_to_dataset_features(robot_raw.observation_features, "observation")
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+                "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
+
+        # Connect robot and wrap for thread safety
+        robot_raw.connect()
+        robot = RobotWrapper(robot_raw)
+        
+        teleop.connect()
+        listener, events = init_rac_keyboard_listener()
+
+        # Shared state holders (main loop writes, RTC thread reads)
+        queue_holder = {"queue": ActionQueue(cfg.rtc)}
+        obs_holder = {"obs": None, "robot_type": robot.robot_type}  # Main loop updates obs
+
+        # Start RTC inference thread
+        # NOTE: Thread does NOT access robot directly - reads from obs_holder
+        rtc_thread = Thread(
+            target=rtc_inference_thread,
+            args=(
+                policy,
+                obs_holder,  # Thread reads obs from here (set by main loop)
+                hw_features,
+                preprocessor,
+                postprocessor,
+                queue_holder,
+                shutdown_event,
+                policy_active,
+                cfg,
+            ),
+            daemon=True,
+            name="RTCInference",
+        )
+        rtc_thread.start()
+        logger.info("Started RTC inference thread")
+
+        print("\n" + "=" * 65)
+        print("  RaC Data Collection with RTC")
+        print("=" * 65)
+        print(f"  Policy: {cfg.policy.pretrained_path}")
+        print(f"  Task: {cfg.dataset.single_task}")
+        print(f"  FPS: {cfg.dataset.fps}")
+        print(f"  Interpolation: {cfg.interpolation}")
+        print()
+        print("  Controls:")
+        print("    SPACE  - Pause policy")
+        print("    c      - Take control")
+        print("    →      - End episode")
+        print("    ESC    - Stop and push to hub")
+        print("=" * 65 + "\n")
+
+        with VideoEncodingManager(dataset):
+            recorded = 0
+            while recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                log_say(f"RaC episode {dataset.num_episodes}", cfg.play_sounds)
+                
+                # Fresh action queue per episode (update holder so thread sees it)
+                queue_holder["queue"] = ActionQueue(cfg.rtc)
+                
+                logger.info(f"Episode {recorded + 1} / {cfg.dataset.num_episodes}")
+
+                stats = rac_rtc_rollout_loop(
+                    robot=robot,
+                    teleop=teleop,
+                    policy=policy,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    dataset=dataset,
+                    events=events,
+                    cfg=cfg,
+                    queue_holder=queue_holder,
+                    obs_holder=obs_holder,
+                    policy_active=policy_active,
+                    hw_features=hw_features,
+                )
+
+                logging.info(f"Episode stats: {stats}")
+
+                if events["rerecord_episode"]:
+                    log_say("Re-recording", cfg.play_sounds)
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                t_save_start = time.perf_counter()
+                dataset.save_episode()
+                logging.info(f"[RaC] save_episode total: {time.perf_counter() - t_save_start:.2f}s")
+                recorded += 1
+
+                if recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                    reset_loop(robot, teleop, events, cfg.dataset.fps)
+
+    finally:
+        log_say("Stop recording", cfg.play_sounds, blocking=True)
+        
+        shutdown_event.set()
+        policy_active.clear()
+        
+        if rtc_thread and rtc_thread.is_alive():
+            rtc_thread.join(timeout=2.0)
+
+        if dataset:
+            dataset.finalize()
+
+        if robot_raw.is_connected:
+            robot_raw.disconnect()
+        if teleop.is_connected:
+            teleop.disconnect()
+
+        if not is_headless() and listener:
+            listener.stop()
+
+        if cfg.dataset.push_to_hub:
+            dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+def main():
+    from lerobot.utils.import_utils import register_third_party_plugins
+    register_third_party_plugins()
+    rac_rtc_collect()
+
+
+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=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/reuben_pi0 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --policy.path=lipsop/reuben_pi0 \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
         --rtc.execution_horizon=8 \
         --device=cuda
 
     # With torch.compile for faster inference (PyTorch 2.0+)
     # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
     # Run RTC with Real robot with RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
+        --policy.path=helper2424/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \

loop_datasets.py

@@ -0,0 +1,10 @@
+from huggingface_hub import HfApi, list_datasets
+
+api = HfApi()
+datasets = list_datasets(author="lerobot-data-collection")
+print('"[', end="")
+i=0
+for dataset in datasets:
+    if "three-folds-dataset" in dataset.id:
+        print("'" + dataset.id + "',", end="")
+print(']"',)

pyproject.toml

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.5"
+version = "0.4.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.12"
+requires-python = ">=3.10"
 authors = [
     { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
     { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,8 +50,7 @@ classifiers = [
     "Intended Audience :: Education",
     "Intended Audience :: Science/Research",
     "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.12",
-    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.10",
     "Topic :: Software Development :: Build Tools",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -60,30 +59,28 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.0.0,<4.2.0",
     "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub>=1.0.0,<2.0.0",
+    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
     "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
-    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "setuptools>=71.0.0,<81.0.0",
     "cmake>=3.29.0.1,<4.2.0",
-    "packaging>=24.2,<26.0",
-
-    "torch>=2.2.1,<2.11.0",
-    "torchcodec>=0.2.1,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')",
-    "torchvision>=0.21.0,<0.26.0",
-
     "einops>=0.8.0,<0.9.0",
     "opencv-python-headless>=4.9.0,<4.13.0",
     "av>=15.0.0,<16.0.0",
     "jsonlines>=4.0.0,<5.0.0",
-    "pynput>=1.7.8,<1.9.0",
+    "packaging>=24.2,<26.0",
+    "pynput>=1.7.7,<1.9.0",
     "pyserial>=3.5,<4.0",
-
     "wandb>=0.24.0,<0.25.0",
-    "draccus==0.10.0", # TODO: Relax version constraint
+
+    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
+    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
+
+    "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -98,20 +95,14 @@ dependencies = [
 
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.9.17"]
-transformers-dep = ["transformers>=5.3.0,<6.0.0"]
+placo-dep = ["placo>=0.9.6,<0.10.0"]
+transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
-can-dep = ["python-can>=4.2.0,<5.0.0"]
-peft-dep = ["peft>=0.18.0,<1.0.0"]
-scipy-dep = ["scipy>=1.14.0,<2.0.0"]
-qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
-matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["lerobot[can-dep]"]
-robstride = ["lerobot[can-dep]"]
+damiao = ["python-can>=4.2.0,<5.0.0"]
 
 # Robots
 openarms = ["lerobot[damiao]"]
@@ -123,31 +114,30 @@ unitree_g1 = [
     "onnxruntime>=1.16.0,<2.0.0",
     "pin>=3.0.0,<4.0.0",
     "meshcat>=0.3.0,<0.4.0",
-    "lerobot[matplotlib-dep]",
+    "matplotlib>=3.9.0,<4.0.0",
     "casadi>=3.6.0,<4.0.0",
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
-    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
 
 # Policies
 wallx = [
-    "lerobot[transformers-dep]",
-    "lerobot[peft]",
-    "lerobot[scipy-dep]",
-    "torchdiffeq>=0.2.4,<0.3.0",
-    "lerobot[qwen-vl-utils-dep]",
+    "transformers==4.49.0",
+    "peft==0.17.1",
+    "scipy==1.15.3",
+    "torchdiffeq==0.2.5",
+    "qwen_vl_utils==0.0.11"
 ]
-pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
+pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
-multi_task_dit = ["lerobot[transformers-dep]"]
 groot = [
     "lerobot[transformers-dep]",
-    "lerobot[peft]",
+    "peft>=0.13.0,<1.0.0",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
     "safetensors>=0.4.3,<1.0.0",
@@ -156,13 +146,13 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
 xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
-async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
-peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
+async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
+peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -170,19 +160,13 @@ test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
-# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+aloha = ["gym-aloha>=0.1.2,<0.2.0"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
-metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
+metaworld = ["metaworld==3.0.0"]
 
 # All
 all = [
-    # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
-    # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
-    # helps pip's resolver converge by constraining scipy early, before it encounters
-    # the loose scipy requirements from transitive deps like dm-control and metaworld.
-    "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
@@ -190,8 +174,8 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    "lerobot[wallx]",
-    "lerobot[pi]",
+    # "lerobot[wallx]",
+    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
@@ -203,7 +187,7 @@ all = [
     "lerobot[aloha]",
     "lerobot[pusht]",
     "lerobot[phone]",
-    "lerobot[libero]; sys_platform == 'linux'",
+    "lerobot[libero]",
     "lerobot[metaworld]",
     "lerobot[sarm]",
     "lerobot[peft]",
@@ -228,14 +212,11 @@ lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
-[tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
-
 [tool.setuptools.packages.find]
 where = ["src"]
 
 [tool.ruff]
-target-version = "py312"
+target-version = "py310"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -327,7 +308,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
 
 [tool.mypy]
-python_version = "3.12"
+python_version = "3.10"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -411,3 +392,85 @@ ignore_errors = false
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
+
+[tool.uv]
+# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
+conflicts = [
+    [
+        { extra = "wallx" },
+        { extra = "transformers-dep" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "pi" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "smolvla" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "groot" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "xvla" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "sarm" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "hilserl" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "libero" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "peft" },
+    ],
+    [
+        { extra = "wallx" },
+        { extra = "all" },
+    ],
+    # pi uses custom branch which conflicts with transformers-dep
+    [
+        { extra = "pi" },
+        { extra = "transformers-dep" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "smolvla" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "groot" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "xvla" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "sarm" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "hilserl" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "libero" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "peft" },
+    ],
+    [
+        { extra = "pi" },
+        { extra = "all" },
+    ],
+]

src/lerobot/async_inference/robot_client.py

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

src/lerobot/cameras/camera.py

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

src/lerobot/cameras/opencv/camera_opencv.py

@@ -530,7 +530,7 @@ class OpenCVCamera(Camera):
         return frame
 
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -201,7 +201,7 @@ class Reachy2Camera(Camera):
         return self.read()
 
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/realsense/camera_realsense.py

@@ -573,7 +573,7 @@ class RealSenseCamera(Camera):
 
     # NOTE(Steven): Missing implementation for depth for now
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/configs/default.py

@@ -27,7 +27,7 @@ class DatasetConfig:
     # "dataset_index" into the returned item. The index mapping is made according to the order in which the
     # datasets are provided.
     repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/configs/types.py

@@ -50,3 +50,8 @@ class RTCAttentionSchedule(str, Enum):
     ONES = "ONES"
     LINEAR = "LINEAR"
     EXP = "EXP"
+
+
+class RTCTrainingDelayDistribution(str, Enum):
+    UNIFORM = "UNIFORM"
+    EXP = "EXP"

src/lerobot/datasets/aggregate.py

@@ -289,9 +289,7 @@ def aggregate_datasets(
 
     logging.info("Find all tasks")
     unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame(
-        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
-    )
+    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
 
     meta_idx = {"chunk": 0, "file": 0}
     data_idx = {"chunk": 0, "file": 0}

src/lerobot/datasets/card_template.md

@@ -7,13 +7,6 @@
 
 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/compute_stats.py

@@ -13,6 +13,10 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import logging
+import time
+from concurrent.futures import ThreadPoolExecutor
+
 import numpy as np
 
 from lerobot.datasets.utils import load_image_as_numpy
@@ -227,19 +231,20 @@ def auto_downsample_height_width(img: np.ndarray, target_size: int = 150, max_si
     return img[:, ::downsample_factor, ::downsample_factor]
 
 
+def _load_single_image(path: str) -> np.ndarray:
+    img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True)
+    return auto_downsample_height_width(img)
+
+
 def sample_images(image_paths: list[str]) -> np.ndarray:
     sampled_indices = sample_indices(len(image_paths))
+    paths = [image_paths[idx] for idx in sampled_indices]
 
-    images = None
-    for i, idx in enumerate(sampled_indices):
-        path = image_paths[idx]
-        # we load as uint8 to reduce memory usage
-        img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True)
-        img = auto_downsample_height_width(img)
-
-        if images is None:
-            images = np.empty((len(sampled_indices), *img.shape), dtype=np.uint8)
+    with ThreadPoolExecutor(max_workers=min(8, len(paths))) as pool:
+        loaded = list(pool.map(_load_single_image, paths))
 
+    images = np.empty((len(loaded), *loaded[0].shape), dtype=np.uint8)
+    for i, img in enumerate(loaded):
         images[i] = img
 
     return images
@@ -504,27 +509,46 @@ def compute_episode_stats(
         quantile_list = DEFAULT_QUANTILES
 
     ep_stats = {}
-    for key, data in episode_data.items():
-        if features[key]["dtype"] == "string":
-            continue
 
+    def _compute_single_feature_stats(key, data):
+        t0 = time.perf_counter()
         if features[key]["dtype"] in ["image", "video"]:
             ep_ft_array = sample_images(data)
             axes_to_reduce = (0, 2, 3)
-            keepdims = True
+            kd = True
         else:
             ep_ft_array = data
             axes_to_reduce = 0
-            keepdims = data.ndim == 1
+            kd = data.ndim == 1
 
-        ep_stats[key] = get_feature_stats(
-            ep_ft_array, axis=axes_to_reduce, keepdims=keepdims, quantile_list=quantile_list
-        )
+        stats = get_feature_stats(ep_ft_array, axis=axes_to_reduce, keepdims=kd, quantile_list=quantile_list)
 
         if features[key]["dtype"] in ["image", "video"]:
-            ep_stats[key] = {
-                k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items()
-            }
+            stats = {k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in stats.items()}
+
+        dt = time.perf_counter() - t0
+        if dt > 0.1:
+            logging.info(f"[compute_episode_stats] {key} ({features[key]['dtype']}): {dt:.2f}s")
+        return key, stats
+
+    # Split into image/video features (heavy I/O) and numeric features (fast)
+    image_keys = [(k, d) for k, d in episode_data.items()
+                  if k in features and features[k]["dtype"] in ["image", "video"]]
+    numeric_keys = [(k, d) for k, d in episode_data.items()
+                    if k in features and features[k]["dtype"] not in ["image", "video", "string"]]
+
+    # Run image features in parallel (I/O bound)
+    if image_keys:
+        with ThreadPoolExecutor(max_workers=len(image_keys)) as pool:
+            futures = [pool.submit(_compute_single_feature_stats, k, d) for k, d in image_keys]
+            for f in futures:
+                key, stats = f.result()
+                ep_stats[key] = stats
+
+    # Numeric features are fast — run sequentially
+    for k, d in numeric_keys:
+        _, stats = _compute_single_feature_stats(k, d)
+        ep_stats[k] = stats
 
     return ep_stats
 

src/lerobot/datasets/dataset_tools.py

@@ -37,7 +37,7 @@ import torch
 from tqdm import tqdm
 
 from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.datasets.compute_stats import aggregate_stats
+from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats, get_feature_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
     DATA_DIR,
@@ -89,8 +89,8 @@ def delete_episodes(
     Args:
         dataset: The source LeRobotDataset.
         episode_indices: List of episode indices to delete.
-        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.
+        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.
     """
     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: Root directory where the split datasets will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
+        output_dir: Base directory for output datasets. If None, uses default location.
 
     Examples:
       Split by specific episodes
@@ -243,8 +243,8 @@ def merge_datasets(
 
     Args:
         datasets: List of LeRobotDatasets to merge.
-        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.
+        output_repo_id: Repository ID for the merged dataset.
+        output_dir: Directory to save the merged dataset. If None, uses default location.
     """
     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: 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.
+        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.
 
     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: 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.
+        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.
 
     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: 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.
+        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.
 
     Returns:
         New dataset with features removed.
@@ -567,22 +567,20 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
     input_path: Path,
     output_path: Path,
-    episodes_to_keep: list[tuple[int, int]],
+    episodes_to_keep: list[tuple[float, float]],
     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 frame ranges and re-encodes them with
+    This function decodes frames from specified time 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_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.
+        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
         fps: Frame rate of the video.
         vcodec: Video codec to use for encoding.
         pix_fmt: Pixel format for output video.
@@ -624,10 +622,9 @@ def _keep_episodes_from_video_with_av(
 
     # Create set of (start, end) ranges for fast lookup.
     # Convert to a sorted list for efficient checking.
-    frame_ranges = sorted(episodes_to_keep)
+    time_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
 
@@ -637,20 +634,21 @@ def _keep_episodes_from_video_with_av(
             if frame is None:
                 continue
 
-            # Check if frame is in any of our desired frame ranges.
+            # 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.
             # Skip ranges that have already passed.
-            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
+            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
                 range_idx += 1
 
             # If we've passed all ranges, stop processing.
-            if range_idx >= len(frame_ranges):
+            if range_idx >= len(time_ranges):
                 break
 
             # Check if frame is in current range.
-            start_frame = frame_ranges[range_idx][0]
-
-            if src_frame_count < start_frame:
-                src_frame_count += 1
+            start_ts, end_ts = time_ranges[range_idx]
+            if frame_time < start_ts:
                 continue
 
             # Frame is in range - create a new frame with reset timestamps.
@@ -663,7 +661,6 @@ 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.
@@ -752,17 +749,15 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of frame ranges to keep, in sorted order.
+                # Build list of time 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[int, int]] = []
+                episodes_to_keep_ranges: list[tuple[float, float]] = []
+
                 for old_idx in sorted_keep_episodes:
                     src_ep = src_dataset.meta.episodes[old_idx]
-                    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))
+                    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))
 
                 # Use PyAV filters to efficiently re-encode only the desired segments.
                 assert src_dataset.meta.video_path is not None
@@ -1475,9 +1470,7 @@ 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=pd.Index(unique_tasks, name="task")
-    )
+    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
     task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
 
     logging.info(f"Modifying tasks in {dataset.repo_id}")
@@ -1529,9 +1522,125 @@ def modify_tasks(
     return dataset
 
 
+def recompute_stats(
+    dataset: LeRobotDataset,
+    skip_image_video: bool = True,
+    delta_action: bool = False,
+    delta_exclude_joints: list[str] | None = None,
+) -> LeRobotDataset:
+    """Recompute stats.json from scratch by iterating all episodes.
+
+    Args:
+        dataset: The LeRobotDataset to recompute stats for.
+        skip_image_video: If True (default), only recompute stats for numeric features
+            (action, state, etc.) and keep existing image/video stats unchanged.
+        delta_action: If True, compute action stats as delta (action - state).
+            Useful when training with use_delta_actions=True so normalization matches.
+        delta_exclude_joints: Joint names to exclude from delta conversion when
+            delta_action=True. These dims keep absolute stats. Uses dataset's
+            action feature names to build the mask. Default: ["gripper"].
+
+    Returns:
+        The same dataset with updated stats.
+    """
+    features = dataset.meta.features
+    numeric_features = {
+        k: v for k, v in features.items()
+        if v["dtype"] not in ["image", "video", "string"]
+        and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+    }
+
+    if skip_image_video:
+        features_to_compute = numeric_features
+    else:
+        features_to_compute = {
+            k: v for k, v in features.items()
+            if v["dtype"] != "string"
+            and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+        }
+
+    # Build delta mask if delta_action is enabled
+    delta_mask = None
+    if delta_action and "action" in features and "observation.state" in features:
+        if delta_exclude_joints is None:
+            delta_exclude_joints = ["gripper"]
+        action_names = features["action"].get("names")
+        if action_names is not None:
+            exclude = set(delta_exclude_joints)
+            delta_mask = [n not in exclude for n in action_names]
+        else:
+            action_dim = features["action"]["shape"][0]
+            delta_mask = [True] * action_dim
+        # Only recompute action stats when delta is enabled — state stays unchanged
+        features_to_compute = {"action": features["action"]}
+        logging.info(f"Recomputing action stats as delta (exclude: {delta_exclude_joints})")
+    else:
+        logging.info(f"Recomputing stats for features: {list(features_to_compute.keys())}")
+
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    all_episode_stats = []
+    numeric_keys = [k for k, v in features_to_compute.items() if v["dtype"] not in ["image", "video"]]
+    # Also need state for delta computation even though we don't recompute state stats
+    needs_state = delta_mask is not None
+
+    for parquet_path in tqdm(parquet_files, desc="Computing stats from data files"):
+        df = pd.read_parquet(parquet_path)
+
+        for ep_idx in sorted(df["episode_index"].unique()):
+            ep_df = df[df["episode_index"] == ep_idx]
+            episode_data = {}
+            for key in numeric_keys:
+                if key in ep_df.columns:
+                    values = ep_df[key].values
+                    if hasattr(values[0], "__len__"):
+                        episode_data[key] = np.stack(values)
+                    else:
+                        episode_data[key] = np.array(values)
+
+            # Apply delta conversion to actions before computing stats
+            if delta_mask is not None and "action" in episode_data:
+                from lerobot.processor.delta_action_processor import to_delta_actions
+
+                # Load state for delta even if we're not computing state stats
+                if needs_state and "observation.state" in ep_df.columns:
+                    state_values = ep_df["observation.state"].values
+                    if hasattr(state_values[0], "__len__"):
+                        states = np.stack(state_values)
+                    else:
+                        states = np.array(state_values)
+                    actions_t = torch.from_numpy(episode_data["action"]).float()
+                    states_t = torch.from_numpy(states).float()
+                    episode_data["action"] = to_delta_actions(actions_t, states_t, delta_mask).numpy()
+
+            ep_stats = compute_episode_stats(episode_data, features_to_compute)
+            all_episode_stats.append(ep_stats)
+
+    if not all_episode_stats:
+        logging.warning("No episode stats computed")
+        return dataset
+
+    new_stats = aggregate_stats(all_episode_stats)
+
+    # Merge: keep existing stats for features we didn't recompute
+    if dataset.meta.stats:
+        for key, value in dataset.meta.stats.items():
+            if key not in new_stats:
+                new_stats[key] = value
+
+    write_stats(new_stats, dataset.root)
+    dataset.meta.stats = new_stats
+
+    logging.info(f"Stats recomputed for {len(all_episode_stats)} episodes")
+    return dataset
+
+
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
-    output_dir: Path | None = None,
+    output_dir: Path,
     repo_id: str | None = None,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -1550,8 +1659,8 @@ def convert_image_to_video_dataset(
 
     Args:
         dataset: The source LeRobot dataset with images
-        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.
+        output_dir: Directory to save the new video dataset
+        repo_id: Repository ID for the new dataset (default: original_id + "_video")
         vcodec: Video codec (default: libsvtav1)
         pix_fmt: Pixel format (default: yuv420p)
         g: Group of pictures size (default: 2)
@@ -1602,7 +1711,6 @@ 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

@@ -18,22 +18,30 @@ import contextlib
 import logging
 import shutil
 import tempfile
+import time
 from collections.abc import Callable
 from pathlib import Path
 
 import datasets
 import numpy as np
+import os
 import packaging.version
 import pandas as pd
 import PIL.Image
 import pyarrow as pa
 import pyarrow.parquet as pq
+from concurrent.futures import ProcessPoolExecutor
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
+from lerobot.datasets.compute_stats import (
+    RunningQuantileStats,
+    aggregate_stats,
+    auto_downsample_height_width,
+    compute_episode_stats,
+)
 from lerobot.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
@@ -76,7 +84,6 @@ 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
 
@@ -314,7 +321,7 @@ class LeRobotDatasetMetadata:
         if self.tasks is None:
             new_tasks = tasks
             task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(tasks, name="task"))
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
         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))
@@ -421,8 +428,10 @@ class LeRobotDatasetMetadata:
 
         write_info(self.info, self.root)
 
+        t0 = time.perf_counter()
         self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
         write_stats(self.stats, self.root)
+        logging.info(f"[meta.save_episode] aggregate+write_stats: {time.perf_counter() - t0:.2f}s")
 
     def update_video_info(self, video_key: str | None = None) -> None:
         """
@@ -545,20 +554,11 @@ class LeRobotDatasetMetadata:
         return obj
 
 
-def _encode_video_worker(
-    video_key: str,
-    episode_index: int,
-    root: Path,
-    fps: int,
-    vcodec: str = "libsvtav1",
-    encoder_threads: int | None = None,
-) -> Path:
+def _encode_video_worker(video_key: str, episode_index: int, root: Path, fps: int) -> 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, encoder_threads=encoder_threads
-    )
+    encode_video_frames(img_dir, temp_path, fps, overwrite=True)
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -577,10 +577,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         download_videos: bool = True,
         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:
@@ -664,11 +660,11 @@ 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.
-            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).
+            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'.
             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
@@ -693,16 +689,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
             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', '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__()
         self.repo_id = repo_id
@@ -716,8 +702,6 @@ 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 = resolve_vcodec(vcodec)
-        self._encoder_threads = encoder_threads
 
         # Unused attributes
         self.image_writer = None
@@ -726,6 +710,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         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._running_video_stats = {}
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -747,7 +732,7 @@ 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 (FileNotFoundError, NotADirectoryError):
+        except (AssertionError, FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
                 self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
@@ -767,19 +752,6 @@ 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)
@@ -839,7 +811,7 @@ 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, repo_id=self.repo_id, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
         )
         card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
 
@@ -969,30 +941,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
         else:
             return get_hf_features_from_features(self.features)
 
-    def _get_query_indices(
-        self, abs_idx: int, ep_idx: int
-    ) -> tuple[dict[str, list[int]], dict[str, torch.Tensor]]:
-        """Compute query indices for delta timestamps.
-
-        Args:
-            abs_idx: The absolute index in the full dataset (not the relative index in filtered episodes).
-            ep_idx: The episode index.
-
-        Returns:
-            A tuple of (query_indices, padding) where:
-            - query_indices: Dict mapping keys to lists of absolute indices to query
-            - padding: Dict mapping "{key}_is_pad" to boolean tensors indicating padded positions
-        """
+    def _get_query_indices(self, idx: int, ep_idx: int) -> tuple[dict[str, list[int | bool]]]:
         ep = self.meta.episodes[ep_idx]
         ep_start = ep["dataset_from_index"]
         ep_end = ep["dataset_to_index"]
         query_indices = {
-            key: [max(ep_start, min(ep_end - 1, abs_idx + delta)) for delta in delta_idx]
+            key: [max(ep_start, min(ep_end - 1, idx + delta)) for delta in delta_idx]
             for key, delta_idx in self.delta_indices.items()
         }
         padding = {  # Pad values outside of current episode range
             f"{key}_is_pad": torch.BoolTensor(
-                [(abs_idx + delta < ep_start) | (abs_idx + delta >= ep_end) for delta in delta_idx]
+                [(idx + delta < ep_start) | (idx + delta >= ep_end) for delta in delta_idx]
             )
             for key, delta_idx in self.delta_indices.items()
         }
@@ -1084,12 +1043,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self._ensure_hf_dataset_loaded()
         item = self.hf_dataset[idx]
         ep_idx = item["episode_index"].item()
-        # Use the absolute index from the dataset for delta timestamp calculations
-        abs_idx = item["index"].item()
 
         query_indices = None
         if self.delta_indices is not None:
-            query_indices, padding = self._get_query_indices(abs_idx, ep_idx)
+            query_indices, padding = self._get_query_indices(idx, ep_idx)
             query_result = self._query_hf_dataset(query_indices)
             item = {**item, **padding}
             for key, val in query_result.items():
@@ -1133,10 +1090,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Close the parquet writers. This function needs to be called after data collection/conversion, else footer metadata won't be written to the parquet files.
         The dataset won't be valid and can't be loaded as ds = LeRobotDataset(repo_id=repo, root=HF_LEROBOT_HOME.joinpath(repo))
         """
+        if self._streaming_encoder:
+            self._streaming_encoder.close()
         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
@@ -1186,18 +1143,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Automatically add frame_index and timestamp to episode buffer
         frame_index = self.episode_buffer["size"]
+        if frame_index == 0 and self._streaming_encoder:
+            self._streaming_encoder.start_episode(self.meta.video_keys, self.root)
+            self._init_running_video_stats()
         timestamp = frame.pop("timestamp") if "timestamp" in frame else frame_index / self.fps
         self.episode_buffer["frame_index"].append(frame_index)
         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:
@@ -1205,18 +1158,19 @@ 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"] == "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
-                )
-                if frame_index == 0:
-                    img_path.parent.mkdir(parents=True, exist_ok=True)
-                compress_level = 1 if self.features[key]["dtype"] == "video" else 6
-                self._save_image(frame[key], img_path, compress_level)
-                self.episode_buffer[key].append(str(img_path))
+            if self.features[key]["dtype"] in ["image", "video"]:
+                if self._streaming_encoder and self.features[key]["dtype"] == "video":
+                    self._feed_streaming_frame(key, frame[key])
+                    self.episode_buffer[key].append(None)
+                else:
+                    img_path = self._get_image_file_path(
+                        episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
+                    )
+                    if frame_index == 0:
+                        img_path.parent.mkdir(parents=True, exist_ok=True)
+                    compress_level = 1 if self.features[key]["dtype"] == "video" else 6
+                    self._save_image(frame[key], img_path, compress_level)
+                    self.episode_buffer[key].append(str(img_path))
             else:
                 self.episode_buffer[key].append(frame[key])
 
@@ -1267,79 +1221,50 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 continue
             episode_buffer[key] = np.stack(episode_buffer[key])
 
-        # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
 
-        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 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)
+        t0 = time.perf_counter()
+        if self._streaming_encoder:
+            filtered = {k: v for k, v in episode_buffer.items() if k not in self.meta.video_keys}
+            ep_stats = compute_episode_stats(filtered, self.features)
+            for key in self.meta.video_keys:
+                stats = self._running_video_stats[key].get_statistics()
+                ep_stats[key] = {
+                    k: v if k == "count" else (v.reshape(-1, 1, 1) / 255.0)
+                    for k, v in stats.items()
+                }
         else:
             ep_stats = compute_episode_stats(episode_buffer, self.features)
+        t_stats = time.perf_counter() - t0
 
+        t0 = time.perf_counter()
         ep_metadata = self._save_episode_data(episode_buffer)
+        t_save_data = time.perf_counter() - t0
 
-        if use_streaming:
-            # Finish streaming encoding and collect results
-            streaming_results = self._streaming_encoder.finish_episode()
+        has_video_keys = len(self.meta.video_keys) > 0
+        use_batched_encoding = self.batch_encoding_size > 1
+
+        t0 = time.perf_counter()
+        if has_video_keys and self._streaming_encoder:
+            video_paths = 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))
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, video_paths[video_key]))
         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:
-                # num_cameras * num_threads = (total_cpu -1)
-                with concurrent.futures.ProcessPoolExecutor(max_workers=num_cameras) as executor:
-                    future_to_key = {
-                        executor.submit(
-                            _encode_video_worker,
-                            video_key,
-                            episode_index,
-                            self.root,
-                            self.fps,
-                            self.vcodec,
-                            self._encoder_threads,
-                        ): video_key
-                        for video_key in self.meta.video_keys
-                    }
-
-                    results = {}
-                    for future in concurrent.futures.as_completed(future_to_key):
-                        video_key = future_to_key[future]
-                        try:
-                            temp_path = future.result()
-                            results[video_key] = temp_path
-                        except Exception as exc:
-                            logging.error(f"Video encoding failed for {video_key}: {exc}")
-                            raise exc
-
-                for video_key in self.meta.video_keys:
-                    temp_path = results[video_key]
-                    ep_metadata.update(
-                        self._save_episode_video(video_key, episode_index, temp_path=temp_path)
-                    )
-            else:
-                for video_key in self.meta.video_keys:
-                    ep_metadata.update(self._save_episode_video(video_key, episode_index))
+            video_paths = self._encode_multiple_temporary_episode_videos(self.meta.video_keys, episode_index)
+            for video_key, video_path in zip(self.meta.video_keys, video_paths):
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, video_path))
+        t_video = time.perf_counter() - t0
 
         # `meta.save_episode` need to be executed after encoding the videos
+        t0 = time.perf_counter()
         self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
+        t_meta = time.perf_counter() - t0
+
+        logging.info(
+            f"[save_episode] ep={episode_index} frames={episode_length} | "
+            f"stats={t_stats:.2f}s data={t_save_data:.2f}s video={t_video:.2f}s meta={t_meta:.2f}s "
+            f"total={t_stats + t_save_data + t_video + t_meta:.2f}s"
+        )
 
         if has_video_keys and use_batched_encoding:
             # Check if we should trigger batch encoding
@@ -1507,6 +1432,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episode_index: int,
         temp_path: Path | None = None,
     ) -> dict:
+        t0 = time.perf_counter()
         # Encode episode frames into a temporary video
         if temp_path is None:
             ep_path = self._encode_temporary_episode_video(video_key, episode_index)
@@ -1580,13 +1506,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
             f"videos/{video_key}/from_timestamp": latest_duration_in_s,
             f"videos/{video_key}/to_timestamp": latest_duration_in_s + ep_duration_in_s,
         }
+        save_time = time.perf_counter() - t0
+        rate = ep_duration_in_s / save_time if save_time > 0 else float("inf")
+        logging.info(
+            f"[save_episode_video] {video_key} ep={episode_index} "
+            f"save={save_time:.2f}s video_dur={ep_duration_in_s:.1f}s "
+            f"size={ep_size_in_mb:.1f}MB rate={rate:.2f}x realtime"
+        )
         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()
-
+        if self._streaming_encoder:
+            self._streaming_encoder.stop_episode()
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1595,7 +1526,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             episode_index = self.episode_buffer["episode_index"]
             if isinstance(episode_index, np.ndarray):
                 episode_index = episode_index.item() if episode_index.size == 1 else episode_index[0]
-            for cam_key in self.meta.image_keys:
+            for cam_key in self.meta.camera_keys:
                 img_dir = self._get_image_file_dir(episode_index, cam_key)
                 if img_dir.is_dir():
                     shutil.rmtree(img_dir)
@@ -1628,16 +1559,67 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
+    def start_streaming_encoder(self):
+        """Enable streaming video encoding for recording."""
+        if len(self.meta.video_keys) > 0:
+            self._streaming_encoder = StreamingVideoEncoder(fps=self.fps)
+            self._running_video_stats = {}
+
+    def _init_running_video_stats(self):
+        self._running_video_stats = {key: RunningQuantileStats() for key in self.meta.video_keys}
+
+    def _feed_streaming_frame(self, key: str, image) -> None:
+        """Feed image to streaming encoder and accumulate running stats."""
+        if isinstance(image, np.ndarray):
+            if image.ndim == 3 and image.shape[0] in (1, 3, 4):
+                img_chw = image
+            else:
+                img_chw = image.transpose(2, 0, 1)
+        else:
+            img_chw = np.array(image).transpose(2, 0, 1)
+
+        self._streaming_encoder.feed_frame(key, image)
+        img_ds = auto_downsample_height_width(img_chw)
+        c, h, w = img_ds.shape
+        self._running_video_stats[key].update(
+            img_ds.transpose(1, 2, 0).reshape(-1, c).astype(np.float64)
+        )
+
     def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         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, self._encoder_threads
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps)
+
+    def _encode_multiple_temporary_episode_videos(self, video_keys, episode_index):
+        temp_paths = []
+        img_dirs = []
+        for video_key in video_keys:
+            temp_paths.append(Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4")
+            img_dirs.append(self._get_image_file_dir(episode_index, video_key))
+        fps = [self.fps]*len(video_keys)
+
+        t0 = time.perf_counter()
+        with ProcessPoolExecutor(max_workers=len(video_keys)) as executor:
+            executor.map(encode_video_frames,img_dirs,temp_paths,fps)
+        encode_time = time.perf_counter() - t0
+
+        n_frames = len(list(img_dirs[0].glob("*"))) if img_dirs and img_dirs[0].exists() else 0
+        video_duration_s = n_frames / self.fps if n_frames > 0 else 0
+        rate = video_duration_s / encode_time if encode_time > 0 else float("inf")
+        logging.info(
+            f"[encode_videos] ep={episode_index} keys={len(video_keys)} "
+            f"encode={encode_time:.2f}s video_dur={video_duration_s:.1f}s "
+            f"rate={rate:.2f}x realtime"
         )
 
+        for img_dir in img_dirs:
+            shutil.rmtree(img_dir)
+
+        return temp_paths
+
     @classmethod
     def create(
         cls,
@@ -1652,14 +1634,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_threads: int = 0,
         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."""
-        vcodec = resolve_vcodec(vcodec)
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1668,7 +1645,6 @@ 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
@@ -1677,8 +1653,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_writer = None
         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)
@@ -1696,26 +1670,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.writer = None
         obj.latest_episode = None
         obj._current_file_start_frame = None
+        obj._streaming_encoder = None
+        obj._running_video_stats = {}
+        if streaming_encoding and len(obj.meta.video_keys) > 0:
+            obj._streaming_encoder = StreamingVideoEncoder(fps=fps)
         # Initialize tracking for incremental recording
         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
 
 
@@ -1771,12 +1733,11 @@ 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)
-            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)
+            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/utils.py

@@ -21,7 +21,7 @@ from collections import deque
 from collections.abc import Iterable, Iterator
 from pathlib import Path
 from pprint import pformat
-from typing import Any
+from typing import Any, Generic, TypeVar
 
 import datasets
 import numpy as np
@@ -78,6 +78,8 @@ 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)
@@ -120,9 +122,19 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # 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)
+        # 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)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -339,7 +351,6 @@ 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
 
 
@@ -1232,7 +1243,7 @@ class LookAheadError(Exception):
     pass
 
 
-class Backtrackable[T]:
+class Backtrackable(Generic[T]):
     """
     Wrap any iterator/iterable so you can step back up to `history` items
     and look ahead up to `lookahead` items.

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -36,11 +36,8 @@ 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}.
 ```
 
 """
@@ -108,7 +105,7 @@ episodes.jsonl
 {"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
 
 NEW
-meta/episodes/chunk-000/file_000.parquet
+meta/episodes/chunk-000/episodes_000.parquet
 episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
 -------------------------
 OLD
@@ -116,16 +113,15 @@ tasks.jsonl
 {"task_index": 1, "task": "Put the blue block in the green bowl"}
 
 NEW
-meta/tasks.parquet
+meta/tasks/chunk-000/file_000.parquet
 task_index | task
 -------------------------
 OLD
 episodes_stats.jsonl
-{"episode_index": 1, "stats": {"feature_name": {"min": ..., "max": ..., "mean": ..., "std": ..., "count": ...}}}
 
 NEW
-meta/episodes/chunk-000/file_000.parquet
-episode_index | feature_name/min | feature_name/max | feature_name/mean | feature_name/std | feature_name/count
+meta/episodes_stats/chunk-000/file_000.parquet
+episode_index | mean | std | min | max
 -------------------------
 UPDATE
 meta/info.json
@@ -174,7 +170,7 @@ 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=pd.Index(task_strings, name="task"))
+    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
     write_tasks(df_tasks, new_root)
 
 
@@ -205,6 +201,7 @@ 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
@@ -214,23 +211,9 @@ 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_idx, ep_path in enumerate(tqdm.tqdm(ep_paths, desc="convert data files")):
+    for ep_path in 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,
@@ -241,7 +224,20 @@ 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)
-        paths_to_cat.append(ep_path)
+        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)
 
     # Write remaining data if any
     if paths_to_cat:
@@ -473,7 +469,7 @@ 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)
+    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root) / repo_id
     if root.exists():
         validate_local_dataset_version(root)
         use_local_dataset = True
@@ -533,7 +529,7 @@ 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`, `<USER>/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
     )
     parser.add_argument(
         "--branch",
@@ -557,7 +553,7 @@ if __name__ == "__main__":
         "--root",
         type=str,
         default=None,
-        help="Local directory to use for downloading/writing the dataset. Defaults to $HF_LEROBOT_HOME/repo_id.",
+        help="Local directory to use for downloading/writing the dataset.",
     )
     parser.add_argument(
         "--push-to-hub",

src/lerobot/datasets/video_utils.py

@@ -13,19 +13,16 @@
 # 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 threading import Lock, Thread
 from typing import Any, ClassVar
 
 import av
@@ -37,82 +34,6 @@ 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"):
@@ -227,17 +148,16 @@ def decode_video_frames_torchvision(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    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}"
-        )
+    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}"
+    )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -249,11 +169,7 @@ 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
 
-    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)})"
-        )
+    assert len(timestamps) == len(closest_frames)
     return closest_frames
 
 
@@ -358,16 +274,15 @@ def decode_video_frames_torchcodec(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    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}"
-        )
+    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}"
+    )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -396,13 +311,14 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.WARNING,
-    overwrite: bool = False,
+    log_level: int | None = av.logging.ERROR,
+    overwrite: bool = True,
     preset: int | None = None,
-    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    vcodec = resolve_vcodec(vcodec)
+    # Check encoder availability
+    if vcodec not in ["h264", "hevc", "libsvtav1"]:
+        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -433,22 +349,24 @@ def encode_video_frames(
         width, height = dummy_image.size
 
     # Define video codec options
-    video_options = _get_codec_options(vcodec, g, crf, preset)
+    video_options = {}
+
+    if g is not None:
+        video_options["g"] = str(g)
+
+    if crf is not None:
+        video_options["crf"] = str(crf)
+
+    #TEMPORARY FIX
+    video_options["preset"] = "12"
 
     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 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)
+    if vcodec == "libsvtav1":
+        video_options["preset"] = str(preset) if preset is not None else "12"
 
     # Set logging level
     if log_level is not None:
@@ -484,6 +402,141 @@ def encode_video_frames(
         raise OSError(f"Video encoding did not work. File not found: {video_path}.")
 
 
+_DONE = object()
+
+
+class _CameraEncoder:
+    """Encodes frames for one camera in a daemon thread."""
+
+    def __init__(self, video_path, fps, vcodec, pix_fmt, g, crf):
+        self.video_path = Path(video_path)
+        self.fps = fps
+        self.vcodec = vcodec
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.queue = queue.Queue()
+        self._thread = None
+        self._cancelled = False
+
+    def start(self):
+        self.video_path.parent.mkdir(parents=True, exist_ok=True)
+        self._thread = Thread(target=self._run, daemon=True)
+        self._thread.start()
+
+    def finish(self) -> Path:
+        self.queue.put(_DONE)
+        self._thread.join(timeout=120)
+        return self.video_path
+
+    def cancel(self):
+        self._cancelled = True
+        while not self.queue.empty():
+            try:
+                self.queue.get_nowait()
+            except queue.Empty:
+                break
+        self.queue.put(_DONE)
+        if self._thread:
+            self._thread.join(timeout=5)
+        if self.video_path.parent.exists():
+            shutil.rmtree(self.video_path.parent, ignore_errors=True)
+
+    def _run(self):
+        options = {}
+        if self.g is not None:
+            options["g"] = str(self.g)
+        if self.crf is not None:
+            options["crf"] = str(self.crf)
+        if self.vcodec == "libsvtav1":
+            options["preset"] = "12"
+
+        output = None
+        output_stream = None
+        try:
+            while True:
+                data = self.queue.get()
+                if data is _DONE or self._cancelled:
+                    break
+
+                if isinstance(data, np.ndarray):
+                    if data.ndim == 3 and data.shape[0] in (1, 3, 4):
+                        data = data.transpose(1, 2, 0)
+                    pil = Image.fromarray(data.astype(np.uint8)).convert("RGB")
+                else:
+                    pil = data.convert("RGB")
+
+                if output is None:
+                    w, h = pil.size
+                    output = av.open(str(self.video_path), "w")
+                    output_stream = output.add_stream(self.vcodec, self.fps, options=options)
+                    output_stream.pix_fmt = self.pix_fmt
+                    output_stream.width = w
+                    output_stream.height = h
+
+                pkt = output_stream.encode(av.VideoFrame.from_image(pil))
+                if pkt:
+                    output.mux(pkt)
+
+            if output_stream and not self._cancelled:
+                pkt = output_stream.encode()
+                if pkt:
+                    output.mux(pkt)
+        except Exception as e:
+            logging.error(f"[StreamingEncoder] {e}")
+        finally:
+            if output:
+                output.close()
+
+
+class StreamingVideoEncoder:
+    """Encodes video on-the-fly using one background thread per camera.
+
+    PyAV releases the GIL during encoding, so Python threads give true
+    parallelism for the CPU-intensive codec work. The queue is unbounded
+    so feed_frame never blocks the caller (teleop thread always has priority).
+    """
+
+    def __init__(self, fps, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30):
+        self.fps = fps
+        self._vcodec = vcodec
+        self._pix_fmt = pix_fmt
+        self._g = g
+        self._crf = crf
+        self._encoders: dict[str, _CameraEncoder] = {}
+
+    def start_episode(self, video_keys, temp_dir):
+        self.stop_episode()
+        for key in video_keys:
+            path = Path(tempfile.mkdtemp(dir=temp_dir)) / f"{key}_stream.mp4"
+            enc = _CameraEncoder(path, self.fps, self._vcodec, self._pix_fmt, self._g, self._crf)
+            enc.start()
+            self._encoders[key] = enc
+
+    def feed_frame(self, video_key, image):
+        """Non-blocking: put frame on unbounded queue (never blocks caller)."""
+        enc = self._encoders.get(video_key)
+        if enc:
+            enc.queue.put(image)
+
+    def finish_episode(self) -> dict[str, Path]:
+        """Flush all encoders, wait for completion, return {key: video_path}."""
+        paths = {}
+        for key, enc in self._encoders.items():
+            paths[key] = enc.finish()
+        self._encoders.clear()
+        return paths
+
+    def stop_episode(self):
+        """Cancel current episode encoding (for re-record)."""
+        for enc in self._encoders.values():
+            enc.cancel()
+        self._encoders.clear()
+
+    def close(self):
+        self.stop_episode()
+
+
 def concatenate_video_files(
     input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
 ):
@@ -567,348 +620,6 @@ 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
@@ -943,7 +654,7 @@ with warnings.catch_warnings():
 
 def get_audio_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting audio stream information
     audio_info = {}
@@ -975,7 +686,7 @@ 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.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting video stream information
     video_info = {}
@@ -1061,15 +772,8 @@ class VideoEncodingManager:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        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
+        # Handle any remaining episodes that haven't been batch encoded
+        if self.dataset.episodes_since_last_encoding > 0:
             if exc_type is not None:
                 logging.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
@@ -1086,8 +790,8 @@ class VideoEncodingManager:
         # Finalize the dataset to properly close all writers
         self.dataset.finalize()
 
-        # Clean up episode images if recording was interrupted (only for non-streaming mode)
-        if exc_type is not None and streaming_encoder is None:
+        # Clean up episode images if recording was interrupted
+        if exc_type is not None:
             interrupted_episode_index = self.dataset.num_episodes
             for key in self.dataset.meta.video_keys:
                 img_dir = self.dataset._get_image_file_path(
@@ -1101,12 +805,14 @@ class VideoEncodingManager:
 
         # Clean up any remaining images directory if it's empty
         img_dir = self.dataset.root / "images"
-        if img_dir.exists():
-            png_files = list(img_dir.rglob("*.png"))
-            if len(png_files) == 0:
+        # 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():
                 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/motors/__init__.py

@@ -18,4 +18,7 @@ from .motors_bus import (
     Motor,
     MotorCalibration,
     MotorNormMode,
+    MotorsBus,  # Backward compatibility (alias for SerialMotorsBus)
+    MotorsBusBase,
+    SerialMotorsBus,
 )

src/lerobot/motors/damiao/tables.py

@@ -15,7 +15,7 @@
 """Configuration tables for Damiao motors."""
 
 from enum import IntEnum
-
+from typing import Dict, List, Tuple
 
 # Motor type definitions
 class MotorType(IntEnum):
@@ -33,7 +33,6 @@ class MotorType(IntEnum):
     DMH6215 = 11
     DMG6220 = 12
 
-
 # Control modes
 class ControlMode(IntEnum):
     MIT = 1
@@ -41,7 +40,6 @@ class ControlMode(IntEnum):
     VEL = 3
     TORQUE_POS = 4
 
-
 # Motor variable IDs (RID)
 class MotorVariable(IntEnum):
     UV_VALUE = 0
@@ -90,8 +88,7 @@ class MotorVariable(IntEnum):
     P_M = 80
     XOUT = 81
 
-
-# Motor limit parameters [PMAX, VMAX, TMAX]
+# Motor limit parameters [PMAX, VMAX, TMAX] 
 # PMAX: Maximum position (rad)
 # VMAX: Maximum velocity (rad/s)
 # TMAX: Maximum torque (N·m)
@@ -147,10 +144,10 @@ MODEL_RESOLUTION = {
 
 # CAN baudrates supported by Damiao motors
 AVAILABLE_BAUDRATES = [
-    125000,  # 0: 125 kbps
-    200000,  # 1: 200 kbps
-    250000,  # 2: 250 kbps
-    500000,  # 3: 500 kbps
+    125000,   # 0: 125 kbps
+    200000,   # 1: 200 kbps
+    250000,   # 2: 250 kbps
+    500000,   # 3: 500 kbps
     1000000,  # 4: 1 mbps (default for OpenArms)
     2000000,  # 5: 2 mbps
     2500000,  # 6: 2.5 mbps
@@ -163,6 +160,9 @@ DEFAULT_BAUDRATE = 1000000  # 1 Mbps is standard for OpenArms
 # Default timeout in milliseconds
 DEFAULT_TIMEOUT_MS = 1000
 
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
 # OpenArms specific configurations
 # Based on: https://docs.openarm.dev/software/setup/configure-test
 # OpenArms has 7 DOF per arm (14 total for dual arm)
@@ -182,14 +182,14 @@ OPENARMS_GRIPPER_MOTOR_IDS = {
 
 # Default motor types for OpenArms
 OPENARMS_DEFAULT_MOTOR_TYPES = {
-    "joint_1": MotorType.DM8009,  # Shoulder pan - high torque
-    "joint_2": MotorType.DM8009,  # Shoulder lift - high torque
-    "joint_3": MotorType.DM4340,  # Shoulder rotation
-    "joint_4": MotorType.DM4340,  # Elbow flex
-    "joint_5": MotorType.DM4310,  # Wrist roll
-    "joint_6": MotorType.DM4310,  # Wrist pitch
-    "joint_7": MotorType.DM4310,  # Wrist rotation
-    "gripper": MotorType.DM4310,  # Gripper
+    "joint_1": MotorType.DM8009,   # Shoulder pan - high torque
+    "joint_2": MotorType.DM8009,   # Shoulder lift - high torque
+    "joint_3": MotorType.DM4340,   # Shoulder rotation
+    "joint_4": MotorType.DM4340,   # Elbow flex
+    "joint_5": MotorType.DM4310,   # Wrist roll
+    "joint_6": MotorType.DM4310,   # Wrist pitch
+    "joint_7": MotorType.DM4310,   # Wrist rotation
+    "gripper": MotorType.DM4310,   # Gripper
 }
 
 # MIT control parameter ranges

src/lerobot/motors/dynamixel/dynamixel.py

@@ -22,7 +22,8 @@ import logging
 from copy import deepcopy
 from enum import Enum
 
-from ..encoding_utils import decode_twos_complement, encode_twos_complement
+from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
+
 from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     AVAILABLE_BAUDRATES,
@@ -202,9 +203,9 @@ class DynamixelMotorsBus(SerialMotorsBus):
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
 
     def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):

src/lerobot/motors/feetech/feetech.py

@@ -17,7 +17,8 @@ from copy import deepcopy
 from enum import Enum
 from pprint import pformat
 
-from ..encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+
 from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     FIRMWARE_MAJOR_VERSION,
@@ -164,7 +165,7 @@ class FeetechMotorsBus(SerialMotorsBus):
 
     def _handshake(self) -> None:
         self._assert_motors_exist()
-        self._assert_same_firmware()
+        #self._assert_same_firmware()
 
     def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
         if self.protocol_version == 0:
@@ -297,11 +298,11 @@ class FeetechMotorsBus(SerialMotorsBus):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 0, num_retry=num_retry)
 
-    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
         addr, length = get_address(self.model_ctrl_table, model, "Lock")
-        self._write(addr, length, motor, 0, num_retry=num_retry)
+        self._write(addr, length, motor_id, 0, num_retry=num_retry)
 
     def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):

src/lerobot/motors/motors_bus.py

@@ -29,7 +29,7 @@ from dataclasses import dataclass
 from enum import Enum
 from functools import cached_property
 from pprint import pformat
-from typing import Protocol
+from typing import Protocol, TypeAlias
 
 import serial
 from deepdiff import DeepDiff
@@ -38,8 +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
 
-type NameOrID = str | int
-type Value = int | float
+NameOrID: TypeAlias = str | int
+Value: TypeAlias = int | float
 
 logger = logging.getLogger(__name__)
 
@@ -1277,4 +1277,4 @@ class SerialMotorsBus(MotorsBusBase):
 
 
 # Backward compatibility alias
-MotorsBus = SerialMotorsBus
+MotorsBus: TypeAlias = SerialMotorsBus

src/lerobot/motors/robstride/tables.py

@@ -1,120 +0,0 @@
-# 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/__init__.py

@@ -15,7 +15,6 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .groot.configuration_groot import GrootConfig as GrootConfig
-from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
@@ -29,7 +28,6 @@ from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
 __all__ = [
     "ACTConfig",
     "DiffusionConfig",
-    "MultiTaskDiTConfig",
     "PI0Config",
     "PI05Config",
     "PI0FastConfig",

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -55,16 +55,10 @@ class DiffusionConfig(PreTrainedConfig):
         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.
-        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).
+        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).
         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.
@@ -120,9 +114,7 @@ class DiffusionConfig(PreTrainedConfig):
     # Architecture / modeling.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    resize_shape: tuple[int, int] | None = None
-    crop_ratio: float = 1.0
-    crop_shape: tuple[int, int] | None = None
+    crop_shape: tuple[int, int] | None = (84, 84)
     crop_is_random: bool = True
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
@@ -147,10 +139,6 @@ 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
 
@@ -183,25 +171,6 @@ 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)
@@ -229,12 +198,13 @@ 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.resize_shape is None and self.crop_shape is not None:
+        if 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 image shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
+                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for "
+                        f"`{key}`."
                     )
 
         # Check that all input images have the same shape.