add for rest of policies

Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

.github/workflows/full_tests.yml

@@ -101,9 +101,11 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-      github.event_name == 'push' ||
-      github.event_name == 'workflow_dispatch'
+      github.repository == 'huggingface/lerobot' && (
+        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+        github.event_name == 'push' ||
+        github.event_name == 'workflow_dispatch'
+      )
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:
@@ -171,6 +173,8 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Fix ptxas permissions
+        run: chmod +x /lerobot/.venv/lib/python3.10/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/unbound_deps_tests.yml

@@ -91,6 +91,7 @@ jobs:
     name: Build and Push Docker
     runs-on:
       group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
     env:

MANIFEST.in

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

benchmarks/video/README.md

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

docker/Dockerfile.internal

@@ -85,6 +85,8 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
 
 RUN uv pip install --no-cache ".[all]"
 
+RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
+
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .

docs/source/_toctree.yml

@@ -7,8 +7,6 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
-  - local: cameras
-    title: Cameras
   - local: bring_your_own_policies
     title: Bring Your Own Policies
   - local: integrate_hardware
@@ -29,6 +27,10 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     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
@@ -101,11 +103,17 @@
     title: Earth Rover Mini
   - local: omx
     title: OMX
+  - local: openarm
+    title: OpenArm
   title: "Robots"
 - sections:
   - local: phone_teleop
     title: Phone
   title: "Teleoperators"
+- sections:
+  - local: cameras
+    title: Cameras
+  title: "Sensors"
 - sections:
   - local: torch_accelerators
     title: PyTorch accelerators
@@ -115,6 +123,8 @@
     title: Notebooks
   - local: feetech
     title: Updating Feetech Firmware
+  - local: damiao
+    title: Damiao Motors and CAN Bus
   title: "Resources"
 - sections:
   - local: contributing

docs/source/act.mdx

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

docs/source/cameras.mdx

@@ -1,12 +1,22 @@
 # Cameras
 
-LeRobot offers multiple options for video capture, including phone cameras, built-in laptop cameras, external webcams, and Intel RealSense cameras. To efficiently record frames from most cameras, you can use either the `OpenCVCamera` or `RealSenseCamera` class. For additional compatibility details on the `OpenCVCamera` class, refer to the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
+LeRobot offers multiple options for video capture:
 
-### Finding your camera
+| Class             | Supported Cameras                   |
+| ----------------- | ----------------------------------- |
+| `OpenCVCamera`    | Phone, built-in laptop, USB webcams |
+| `ZMQCamera`       | Network-connected cameras           |
+| `RealSenseCamera` | Intel RealSense (with depth)        |
+| `Reachy2Camera`   | Reachy 2 robot cameras              |
 
-To instantiate a camera, you need a camera identifier. This identifier might change if you reboot your computer or re-plug your camera, a behavior mostly dependant on your operating system.
+> [!TIP]
+> For `OpenCVCamera` compatibility details, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
 
-To find the camera indices of the cameras plugged into your system, run the following script:
+### Find your camera
+
+Every camera requires a unique identifier to be instantiated, allowing you to distinguish between multiple connected devices.
+
+`OpenCVCamera` and `RealSenseCamera` support auto-discovery. Run the command below to list available devices and their identifiers. Note that these identifiers may change after rebooting your computer or re-plugging the camera, depending on your operating system.
 
 ```bash
 lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
@@ -14,7 +24,7 @@ lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 
 The output will look something like this if you have two cameras connected:
 
-```
+```bash
 --- Detected Cameras ---
 Camera #0:
   Name: OpenCV Camera @ 0
@@ -33,13 +43,37 @@ Camera #0:
 > [!WARNING]
 > When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.
 
-## Use Cameras
+`ZMQCamera` and `Reachy2Camera` do not support auto-discovery. They must be configured manually by providing their network address and port or robot SDK settings.
 
-Below are two examples, demonstrating how to work with the API.
+## Use cameras
 
-- **Asynchronous frame capture** using an OpenCV-based camera
+### Frame access modes
+
+All camera classes implement three access modes for capturing frames:
+
+| Method                    | Behavior                                                                                                                                                   | Blocks?        | Best For                                 |
+| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | ---------------------------------------- |
+| `read()`                  | Waits for the camera hardware to return a frame. May block for a long time depending on the camera and SDK.                                                | Yes            | Simple scripts, sequential capture       |
+| `async_read(timeout_ms)`  | Returns the latest unconsumed frame from background thread. Blocks only if buffer is empty, up to `timeout_ms`. Raises `TimeoutError` if no frame arrives. | With a timeout | Control loops synchronized to camera FPS |
+| `read_latest(max_age_ms)` | Peeks at the most recent frame in buffer (may be stale). Raises `TimeoutError` if frame is older than `max_age_ms`.                                        | No             | UI visualization, logging, monitoring    |
+
+### Usage examples
+
+The following examples show how to use the camera API to configure and capture frames from different camera types.
+
+- **Blocking and non-blocking frame capture** using an OpenCV-based camera
 - **Color and depth capture** using an Intel RealSense camera
 
+> [!WARNING]
+> Failing to cleanly disconnect cameras can cause resource leaks. Use the context manager protocol to ensure automatic cleanup:
+>
+> ```python
+> with OpenCVCamera(config) as camera:
+>     ...
+> ```
+>
+> You can also call `connect()` and `disconnect()` manually, but always use a `finally` block for the latter.
+
 <hfoptions id="shell_restart">
 <hfoption id="Open CV Camera">
 
@@ -60,16 +94,30 @@ config = OpenCVCameraConfig(
 )
 
 # Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
-camera = OpenCVCamera(config)
-camera.connect()
+with OpenCVCamera(config) as camera:
+
+    # Read a frame synchronously — blocks until hardware delivers a new frame
+    frame = camera.read()
+    print(f"read() call returned frame with shape:", frame.shape)
+
+    # Read a frame asynchronously with a timeout — returns the latest unconsumed frame or waits up to timeout_ms for a new one
+    try:
+        for i in range(10):
+            frame = camera.async_read(timeout_ms=200)
+            print(f"async_read call returned frame {i} with shape:", frame.shape)
+    except TimeoutError as e:
+        print(f"No frame received within timeout: {e}")
+
+    # Instantly return a frame - returns the most recent frame captured by the camera
+    try:
+        initial_frame = camera.read_latest(max_age_ms=1000)
+        for i in range(10):
+            frame = camera.read_latest(max_age_ms=1000)
+            print(f"read_latest call returned frame {i} with shape:", frame.shape)
+            print(f"Was a new frame received by the camera? {not (initial_frame == frame).any()}")
+    except TimeoutError as e:
+        print(f"Frame too old: {e}")
 
-# Read frames asynchronously in a loop via `async_read(timeout_ms)`
-try:
-    for i in range(10):
-        frame = camera.async_read(timeout_ms=200)
-        print(f"Async frame {i} shape:", frame.shape)
-finally:
-    camera.disconnect()
 ```
 <!-- prettier-ignore-end -->
 
@@ -111,10 +159,10 @@ finally:
 </hfoption>
 </hfoptions>
 
-## Use your phone
+## Use your phone's camera
 
 <hfoptions id="use phone">
-<hfoption id="Mac">
+<hfoption id="iPhone & macOS">
 
 To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
@@ -124,83 +172,49 @@ To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
 For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
 
-Your iPhone should be detected automatically when running the camera setup script in the next section.
-
 </hfoption>
-<hfoption id="Linux">
+<hfoption id="OBS virtual camera">
 
-If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
+If you want to use your phone as a camera using OBS, follow these steps to set up a virtual camera.
 
-1. _Install `v4l2loopback-dkms` and `v4l-utils`_. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
+1. _(Linux only) Install `v4l2loopback-dkms` and `v4l-utils`_. These packages create virtual camera devices and verify their settings. Install with:
 
-<!-- prettier-ignore-start -->
-```python
+```bash
 sudo apt install v4l2loopback-dkms v4l-utils
 ```
-<!-- prettier-ignore-end -->
 
-2. _Install [DroidCam](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
-3. _Install [OBS Studio](https://obsproject.com)_. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
+2. _Install the [DroidCam app](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
+3. _Download and install [OBS Studio](https://obsproject.com)_.
+4. _Download and install the [DroidCam OBS plugin](https://droidcam.app/obs)_.
+5. _Start OBS Studio_.
 
-<!-- prettier-ignore-start -->
-```python
-flatpak install flathub com.obsproject.Studio
-```
-<!-- prettier-ignore-end -->
-
-4. _Install the DroidCam OBS plugin_. This plugin integrates DroidCam with OBS Studio. Install it with:
-
-<!-- prettier-ignore-start -->
-```python
-flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
-```
-<!-- prettier-ignore-end -->
-
-5. _Start OBS Studio_. Launch with:
-
-<!-- prettier-ignore-start -->
-```python
-flatpak run com.obsproject.Studio
-```
-<!-- prettier-ignore-end -->
-
-6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
-7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
+6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480` to avoid the watermarks.
+7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video` or `OBS > Preferences... > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it.
 8. _Start virtual camera_. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. _Verify the virtual camera setup_. Use `v4l2-ctl` to list the devices:
+9. _Verify the virtual camera setup and resolution_.
+   - **Linux**: Use `v4l2-ctl` to list devices and check resolution:
+     ```bash
+     v4l2-ctl --list-devices  # find VirtualCam and note its /dev/videoX path
+     v4l2-ctl -d /dev/videoX --get-fmt-video  # replace with your VirtualCam path
+     ```
+     You should see `VirtualCam` listed and resolution `640x480`.
+   - **macOS**: Open Photo Booth or FaceTime and select "OBS Virtual Camera" as the input.
+   - **Windows**: The native Camera app doesn't support virtual cameras. Use a video conferencing app (Zoom, Teams) or run `lerobot-find-cameras opencv` directly to verify.
 
-<!-- prettier-ignore-start -->
-```python
-v4l2-ctl --list-devices
-```
-<!-- prettier-ignore-end -->
+<details>
+<summary><strong>Troubleshooting</strong></summary>
 
-You should see an entry like:
+> The virtual camera resolution is incorrect.
 
-```
-VirtualCam (platform:v4l2loopback-000):
-/dev/video1
-```
+Delete the virtual camera source and recreate it. The resolution cannot be changed after creation.
 
-10. _Check the camera resolution_. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
+> Error reading frame in background thread for OpenCVCamera(X): OpenCVCamera(X) frame width=640 or height=480 do not match configured width=1920 or height=1080.
 
-<!-- prettier-ignore-start -->
-```python
-v4l2-ctl -d /dev/video1 --get-fmt-video
-```
-<!-- prettier-ignore-end -->
+This error is caused by OBS Virtual Camera advertising a `1920x1080` resolution despite rescaling. The only fix for now is to comment out the width and height check in `_postprocess_image()`.
 
-You should see an entry like:
-
-```
->>> Format Video Capture:
->>>	Width/Height      : 640/480
->>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
-```
-
-Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
-
-If everything is set up correctly, you can proceed with the rest of the tutorial.
+</details>
 
 </hfoption>
 </hfoptions>
+
+If everything is set up correctly, your phone will appear as a standard OpenCV camera and can be used with `OpenCVCamera`.

docs/source/damiao.mdx

@@ -0,0 +1,165 @@
+# Damiao Motors and CAN Bus
+
+This guide covers setup and usage of Damiao motors with LeRobot via CAN bus communication.
+
+Currently, only Linux is supported, as the OpenArms CAN adapter only has drivers for Linux.
+
+## Linux CAN Setup
+
+Before using Damiao motors, you need to set up the CAN interface on your Linux system.
+
+### Install CAN Utilities
+
+```bash
+sudo apt-get install can-utils
+```
+
+### Configure CAN Interface (Manual)
+
+For standard CAN FD (recommended for OpenArms):
+
+```bash
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000 dbitrate 5000000 fd on
+sudo ip link set can0 up
+```
+
+For standard CAN (without FD):
+
+```bash
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000
+sudo ip link set can0 up
+```
+
+### Configure CAN Interface (Using LeRobot)
+
+LeRobot provides a utility script to setup and test CAN interfaces:
+
+```bash
+# Setup multiple interfaces (e.g., OpenArms Followers with 2 CAN buses)
+lerobot-setup-can --mode=setup --interfaces=can0,can1
+```
+
+## Debugging CAN Communication
+
+Use the built-in debug tools to test motor communication:
+
+```bash
+# Test motors on all interfaces
+lerobot-setup-can --mode=test --interfaces=can0,can1
+
+# Run speed/latency test
+lerobot-setup-can --mode=speed --interfaces=can0
+```
+
+The test mode will scan for motors (IDs 0x01-0x08) and report which ones respond. Example output:
+
+```
+can0: UP (CAN FD)
+  Motor 0x01 (joint_1): ✓ FOUND
+    → Response 0x11 [FD]: 00112233...
+  Motor 0x02 (joint_2): ✓ FOUND
+  Motor 0x03 (joint_3): ✗ No response
+  ...
+  Summary: 2/8 motors found
+```
+
+## Usage
+
+### Basic Setup
+
+```python
+from lerobot.motors import Motor
+from lerobot.motors.damiao import DamiaoMotorsBus
+
+# Define your motors with send/receive CAN IDs
+motors = {
+    "joint_1": Motor(id=0x01, motor_type_str="dm8009", recv_id=0x11),
+    "joint_2": Motor(id=0x02, motor_type_str="dm4340", recv_id=0x12),
+    "joint_3": Motor(id=0x03, motor_type_str="dm4310", recv_id=0x13),
+}
+
+# Create the bus
+bus = DamiaoMotorsBus(
+    port="can0",  # Linux socketcan interface
+    motors=motors,
+)
+
+# Connect
+bus.connect()
+```
+
+### Reading Motor States
+
+```python
+# Read single motor position (degrees)
+position = bus.read("Present_Position", "joint_1")
+
+# Read from multiple motors
+positions = bus.sync_read("Present_Position")  # All motors
+positions = bus.sync_read("Present_Position", ["joint_1", "joint_2"])
+
+# Read all states at once (position, velocity, torque)
+states = bus.sync_read_all_states()
+# Returns: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
+```
+
+### Writing Motor Commands
+
+```python
+# Enable torque
+bus.enable_torque()
+
+# Set goal position (degrees)
+bus.write("Goal_Position", "joint_1", 45.0)
+
+# Set positions for multiple motors
+bus.sync_write("Goal_Position", {
+    "joint_1": 45.0,
+    "joint_2": -30.0,
+    "joint_3": 90.0,
+})
+
+# Disable torque
+bus.disable_torque()
+```
+
+## Configuration Options
+
+| Parameter      | Default   | Description                                                 |
+| -------------- | --------- | ----------------------------------------------------------- |
+| `port`         | -         | CAN interface (`can0`) or serial port (`/dev/cu.usbmodem*`) |
+| `use_can_fd`   | `True`    | Enable CAN FD for higher data rates                         |
+| `bitrate`      | `1000000` | Nominal bitrate (1 Mbps)                                    |
+| `data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)                                |
+
+## Motor Configuration
+
+Each motor requires:
+
+- `id`: CAN ID for sending commands
+- `motor_type`: One of the supported motor types (e.g., `"dm8009"`, `"dm4340"`)
+- `recv_id`: CAN ID for receiving responses
+
+OpenArms default IDs follow the pattern: send ID `0x0N`, receive ID `0x1N` where N is the joint number.
+
+## Troubleshooting
+
+### No Response from Motors
+
+1. **Check power**
+2. **Verify CAN wiring**: Check CAN-H, CAN-L, and GND connections
+3. **Check motor IDs**: Use Damiao Debugging Tools to verify/configure IDs
+4. **Test CAN interface**: Run `candump can0` to see if messages are being received
+5. **Run diagnostics**: `lerobot-setup-can --mode=test --interfaces=can0`
+
+### Motor Timeout Parameter
+
+If motors were configured with timeout=0, they won't respond to commands. Use Damiao Debugging Tools to set a non-zero timeout value.
+
+### Verify CAN FD Status
+
+```bash
+ip -d link show can0 | grep fd
+```

docs/source/dataset_subtask.mdx

@@ -0,0 +1,278 @@
+# Using Subtasks in LeRobot Datasets
+
+Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
+
+- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
+- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
+- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
+
+LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
+
+## What are Subtasks?
+
+While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
+
+1. "Approach the apple"
+2. "Grasp the apple"
+3. "Lift the apple"
+4. "Move to basket"
+5. "Release the apple"
+
+Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
+
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
+  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
+  width="80%"
+/>
+
+<p>
+  <em>Figure: Overview of subtask annotation.</em>
+</p>
+
+**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
+
+## Dataset Structure
+
+Subtask information is stored in the dataset metadata:
+
+```
+my-dataset/
+├── data/
+│   └── ...
+├── meta/
+│   ├── info.json
+│   ├── stats.json
+│   ├── tasks.parquet
+│   ├── subtasks.parquet      # Subtask index → subtask string mapping
+│   └── episodes/
+│       └── ...
+└── videos/
+    └── ...
+```
+
+### Subtasks Parquet File
+
+The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
+
+| subtask_index | subtask (index column) |
+| ------------- | ---------------------- |
+| 0             | "Approach the apple"   |
+| 1             | "Grasp the apple"      |
+| 2             | "Lift the apple"       |
+| ...           | ...                    |
+
+### Frame-Level Annotations
+
+Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
+
+```python
+# Example frame data in the parquet file
+{
+    "index": 42,
+    "timestamp": 1.4,
+    "episode_index": 0,
+    "task_index": 0,
+    "subtask_index": 2,  # References "Lift the apple"
+    "observation.state": [...],
+    "action": [...],
+}
+```
+
+## Annotating Datasets with Subtasks
+
+We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
+
+**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
+
+After completing your annotation:
+
+1. Click "Push to Hub" to upload your annotated dataset
+2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
+
+## Loading Datasets with Subtasks
+
+When you load a dataset with subtask annotations, the subtask information is automatically available:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Load a dataset with subtask annotations
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+# Access a sample
+sample = dataset[100]
+
+# The sample includes both task and subtask information
+print(sample["task"])        # "Collect the fruit"
+print(sample["subtask"])     # "Grasp the apple"
+print(sample["task_index"])  # tensor(0)
+print(sample["subtask_index"])  # tensor(2)
+```
+
+### Checking for Subtask Support
+
+You can check if a dataset has subtask annotations:
+
+```python
+# Check if subtasks are available
+has_subtasks = (
+    "subtask_index" in dataset.features
+    and dataset.meta.subtasks is not None
+)
+
+if has_subtasks:
+    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
+    print("Subtasks:", list(dataset.meta.subtasks.index))
+```
+
+## Using Subtasks for Training
+
+### With the Tokenizer Processor
+
+The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
+
+```python
+from lerobot.processor.tokenizer_processor import TokenizerProcessor
+from lerobot.processor.pipeline import ProcessorPipeline
+
+# Create a tokenizer processor
+tokenizer_processor = TokenizerProcessor(
+    tokenizer_name_or_path="google/paligemma-3b-pt-224",
+    padding="max_length",
+    max_length=64,
+)
+
+# The processor will automatically tokenize subtasks if present in the batch
+# and add them to the observation under:
+# - "observation.subtask.tokens"
+# - "observation.subtask.attention_mask"
+```
+
+When subtasks are available in the batch, the tokenizer processor adds:
+
+- `observation.subtask.tokens`: Tokenized subtask text
+- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
+
+### DataLoader with Subtasks
+
+```python
+import torch
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    batch_size=16,
+    shuffle=True,
+)
+
+for batch in dataloader:
+    # Access subtask information in the batch
+    subtasks = batch["subtask"]  # List of subtask strings
+    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
+
+    # Use for training hierarchical policies or reward models
+    print(f"Batch subtasks: {set(subtasks)}")
+```
+
+## Example Datasets with Subtask Annotations
+
+Try loading a dataset with subtask annotations:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Example dataset with subtask annotations
+dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
+
+# Explore the subtasks
+print("Available subtasks:")
+for subtask_name in dataset.meta.subtasks.index:
+    print(f"  - {subtask_name}")
+
+# Get subtask distribution
+subtask_counts = {}
+for i in range(len(dataset)):
+    sample = dataset[i]
+    subtask = sample["subtask"]
+    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
+
+print("\nSubtask distribution:")
+for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
+    print(f"  {subtask}: {count} frames")
+```
+
+## Use Cases
+
+### 1. Hierarchical Policy Training
+
+Train policies that predict both actions and current subtask:
+
+```python
+class HierarchicalPolicy(nn.Module):
+    def __init__(self, num_subtasks):
+        super().__init__()
+        self.action_head = nn.Linear(hidden_dim, action_dim)
+        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
+
+    def forward(self, observations):
+        features = self.encoder(observations)
+        actions = self.action_head(features)
+        subtask_logits = self.subtask_head(features)
+        return actions, subtask_logits
+```
+
+### 2. Stage-Aware Reward Modeling (SARM)
+
+Build reward models that understand task progression:
+
+```python
+# SARM predicts:
+# - Stage: Which subtask is being executed (discrete)
+# - Progress: How far along the subtask (continuous 0-1)
+
+class SARMRewardModel(nn.Module):
+    def forward(self, observations):
+        features = self.encoder(observations)
+        stage_logits = self.stage_classifier(features)
+        progress = self.progress_regressor(features)
+        return stage_logits, progress
+```
+
+### 3. Progress Visualization
+
+Monitor robot execution by tracking subtask progression:
+
+```python
+def visualize_execution(model, observations):
+    for t, obs in enumerate(observations):
+        action, subtask_logits = model(obs)
+        predicted_subtask = subtask_names[subtask_logits.argmax()]
+        print(f"t={t}: Executing '{predicted_subtask}'")
+```
+
+## API Reference
+
+### LeRobotDataset Properties
+
+| Property                    | Type                   | Description                                |
+| --------------------------- | ---------------------- | ------------------------------------------ |
+| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
+| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
+
+### Sample Keys
+
+When subtasks are available, each sample includes:
+
+| Key             | Type           | Description                          |
+| --------------- | -------------- | ------------------------------------ |
+| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
+| `subtask`       | `str`          | Natural language subtask description |
+
+## Related Resources
+
+- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
+- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
+- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation

docs/source/earthrover_mini_plus.mdx

@@ -1,5 +1,11 @@
 # EarthRover Mini Plus
 
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
+  alt="EarthRover Mini Plus"
+  width="70%"
+/>
+
 The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
 
 ## What You Need
@@ -179,13 +185,16 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-python src/lerobot/scripts/lerobot_record.py \
+lerobot-record \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/groot.mdx

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

docs/source/hope_jr.mdx

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

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -185,7 +185,10 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube"
+    --dataset.single_task="Grab the black cube" \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -515,6 +518,9 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${HF_USER}/eval_so100 \
   --dataset.single_task="Put lego brick into the transparent box" \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2 \
+  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/installation.mdx

@@ -1,13 +1,15 @@
 # Installation
 
-## Install [`miniforge`](https://conda-forge.org/download/)
+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/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```
 
-## Environment Setup
+## Step 2: Environment Setup
 
 Create a virtual environment with Python 3.10, using conda:
 
@@ -38,7 +40,14 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-## Install LeRobot 🤗
+> [!NOTE]
+> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
+>
+> ```bash
+> conda install evdev -c conda-forge
+> ```
+
+## Step 3: Install LeRobot 🤗
 
 ### From Source
 

docs/source/lekiwi.mdx

@@ -1,5 +1,11 @@
 # LeKiwi
 
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/1740517739083.jpeg"
+  alt="LeKiwi"
+  width="70%"
+/>
+
 In the steps below, we explain how to assemble the LeKiwi mobile robot.
 
 ## Source the parts

docs/source/lerobot-dataset-v3.mdx

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

docs/source/libero.mdx

@@ -42,6 +42,7 @@ lerobot-eval \
 ```
 
 - `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
+- `--env.task_ids` picks task ids to run (`[0]`, `[1,2,3]`, etc.). Omit this flag (or set it to `null`) to run all tasks in the suite.
 - `--eval.batch_size` controls how many environments run in parallel.
 - `--eval.n_episodes` sets how many episodes to run in total.
 

docs/source/openarm.mdx

@@ -0,0 +1,276 @@
+# OpenArm
+
+[OpenArm](https://openarm.dev) is an open-source 7DOF humanoid arm designed for physical AI research and deployment.
+
+To get your OpenArm, assembled or DIY, and join the global community, browse verified and certified manufacturers worldwide at [openarm.dev](https://openarm.dev).
+
+## What's Unique?
+
+- **Human-Scale Design**: OpenArm is designed with human-like proportions, scaled for a person around 160-165cm tall. This provides an optimal balance between practical reach and manageable inertia for safe, responsive operation.
+
+- **Safety-First Architecture**: Built with QDD backdrivable motors and high compliance, OpenArm prioritizes safe human-robot interaction while maintaining practical payload capabilities (6.0kg peak / 4.1kg nominal) for real-world tasks.
+
+- **Built for Durability**: Critical structural components use aluminum and stainless steel construction, ensuring robust performance for repetitive data collection and continuous research use.
+
+- **Fully Accessible & Buildable**: Every component, from CNC parts and 3D-printed casings to electrical wiring is designed to be purchasable and buildable by individual researchers and labs, with complete fabrication data provided.
+
+- **Practical & Affordable**: At $6,500 USD for a complete bimanual system, OpenArm delivers research-grade capabilities at a fraction of traditional humanoid robot costs.
+
+## Platform Requirements
+
+<Tip warning={true}>
+  **Linux Only**: OpenArm currently only works on Linux. The CAN bus USB adapter
+  does not have macOS drivers and has not been tested on Windows.
+</Tip>
+
+## Safety Guide
+
+Before operating OpenArm, please read the [official safety guide](https://docs.openarm.dev/getting-started/safety-guide). Key points:
+
+- **Secure installation**: Fasten the arm to a flat, stable surface with screws or clamps
+- **Safe distance**: Keep body parts and objects outside the range of motion during operation
+- **Protective equipment**: Always wear safety goggles; use additional PPE as needed
+- **Payload limits**: Do not exceed specified payload limits (6.0kg peak / 4.1kg nominal per arm)
+- **Emergency stop**: Know the location and operation of the emergency stop device
+- **Regular inspection**: Check for loose screws, damaged mechanical limits, unusual noises, and wiring damage
+
+## Hardware Setup
+
+Follow the official [OpenArm hardware documentation](https://docs.openarm.dev) for:
+
+- Bill of materials and sourcing
+- 3D printing instructions
+- Mechanical assembly
+- Electrical wiring
+
+The hardware repositories are available at [github.com/enactic/openarm](https://github.com/enactic/openarm).
+
+## CAN Bus Setup
+
+OpenArm uses CAN bus communication with Damiao motors. Once you have the CAN bus USB adapter plugged into your Linux PC, follow the [Damiao Motors and CAN Bus guide](./damiao) to configure the interface.
+
+Quick setup:
+
+```bash
+# Setup CAN interfaces
+lerobot-setup-can --mode=setup --interfaces=can0,can1
+
+# Test motor communication
+lerobot-setup-can --mode=test --interfaces=can0,can1
+```
+
+## Install LeRobot 🤗
+
+Follow our [Installation Guide](./installation), then install the Damiao motor support:
+
+```bash
+pip install -e ".[damiao]"
+```
+
+## Usage
+
+### Follower Arm (Robot)
+
+<hfoptions id="follower">
+<hfoption id="Command">
+
+```bash
+lerobot-calibrate \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_openarm_follower
+```
+
+</hfoption>
+<hfoption id="API example">
+
+```python
+from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
+
+config = OpenArmFollowerConfig(
+    port="can0",
+    side="right",  # or "left" for left arm
+    id="my_openarm_follower",
+)
+
+follower = OpenArmFollower(config)
+follower.connect()
+
+# Read current state
+obs = follower.get_observation()
+print(obs)
+
+# Send action (position in degrees)
+action = {
+    "joint_1.pos": 0.0,
+    "joint_2.pos": 0.0,
+    "joint_3.pos": 0.0,
+    "joint_4.pos": 45.0,
+    "joint_5.pos": 0.0,
+    "joint_6.pos": 0.0,
+    "joint_7.pos": 0.0,
+    "gripper.pos": 0.0,
+}
+follower.send_action(action)
+
+follower.disconnect()
+```
+
+</hfoption>
+</hfoptions>
+
+### Leader Arm (Teleoperator)
+
+The leader arm is used for teleoperation - manually moving it to control the follower arm.
+
+<hfoptions id="leader">
+<hfoption id="Command">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_openarm_leader
+```
+
+</hfoption>
+<hfoption id="API example">
+
+```python
+from lerobot.teleoperators.openarm_leader import OpenArmLeader, OpenArmLeaderConfig
+
+config = OpenArmLeaderConfig(
+    port="can1",
+    id="my_openarm_leader",
+    manual_control=True,  # Disable torque for manual movement
+)
+
+leader = OpenArmLeader(config)
+leader.connect()
+
+# Read current position (as action to send to follower)
+action = leader.get_action()
+print(action)
+
+leader.disconnect()
+```
+
+</hfoption>
+</hfoptions>
+
+### Teleoperation
+
+To teleoperate OpenArm with leader-follower control:
+
+```bash
+lerobot-teleoperate \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_follower \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_leader
+```
+
+### Bimanual Teleoperation
+
+To teleoperate a bimanual OpenArm setup with two leader and two follower arms:
+
+```bash
+lerobot-teleoperate \
+    --robot.type=bi_openarm_follower \
+    --robot.left_arm_config.port=can0 \
+    --robot.left_arm_config.side=left \
+    --robot.right_arm_config.port=can1 \
+    --robot.right_arm_config.side=right \
+    --robot.id=my_bimanual_follower \
+    --teleop.type=bi_openarm_leader \
+    --teleop.left_arm_config.port=can2 \
+    --teleop.right_arm_config.port=can3 \
+    --teleop.id=my_bimanual_leader
+```
+
+### Recording Data
+
+To record a dataset during teleoperation:
+
+```bash
+lerobot-record \
+    --robot.type=openarm_follower \
+    --robot.port=can0 \
+    --robot.side=right \
+    --robot.id=my_follower \
+    --teleop.type=openarm_leader \
+    --teleop.port=can1 \
+    --teleop.id=my_leader \
+    --repo-id=my_hf_username/my_openarm_dataset \
+    --fps=30 \
+    --num-episodes=10
+```
+
+## Configuration Options
+
+### Follower Configuration
+
+| Parameter             | Default   | Description                                                |
+| --------------------- | --------- | ---------------------------------------------------------- |
+| `port`                | -         | CAN interface (e.g., `can0`)                               |
+| `side`                | `None`    | Arm side: `"left"`, `"right"`, or `None` for custom limits |
+| `use_can_fd`          | `True`    | Enable CAN FD for higher data rates                        |
+| `can_bitrate`         | `1000000` | Nominal bitrate (1 Mbps)                                   |
+| `can_data_bitrate`    | `5000000` | CAN FD data bitrate (5 Mbps)                               |
+| `max_relative_target` | `None`    | Safety limit for relative target positions                 |
+| `position_kp`         | Per-joint | Position control proportional gains                        |
+| `position_kd`         | Per-joint | Position control derivative gains                          |
+
+### Leader Configuration
+
+| Parameter          | Default   | Description                         |
+| ------------------ | --------- | ----------------------------------- |
+| `port`             | -         | CAN interface (e.g., `can1`)        |
+| `manual_control`   | `True`    | Disable torque for manual movement  |
+| `use_can_fd`       | `True`    | Enable CAN FD for higher data rates |
+| `can_bitrate`      | `1000000` | Nominal bitrate (1 Mbps)            |
+| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)        |
+
+## Motor Configuration
+
+OpenArm uses Damiao motors with the following default configuration:
+
+| Joint                       | Motor Type | Send ID | Recv ID |
+| --------------------------- | ---------- | ------- | ------- |
+| joint_1 (Shoulder pan)      | DM8009     | 0x01    | 0x11    |
+| joint_2 (Shoulder lift)     | DM8009     | 0x02    | 0x12    |
+| joint_3 (Shoulder rotation) | DM4340     | 0x03    | 0x13    |
+| joint_4 (Elbow flex)        | DM4340     | 0x04    | 0x14    |
+| joint_5 (Wrist roll)        | DM4310     | 0x05    | 0x15    |
+| joint_6 (Wrist pitch)       | DM4310     | 0x06    | 0x16    |
+| joint_7 (Wrist rotation)    | DM4310     | 0x07    | 0x17    |
+| gripper                     | DM4310     | 0x08    | 0x18    |
+
+## Troubleshooting
+
+### No Response from Motors
+
+1. Check power supply connections
+2. Verify CAN wiring (CAN-H, CAN-L, GND)
+3. Run diagnostics: `lerobot-setup-can --mode=test --interfaces=can0`
+4. See the [Damiao troubleshooting guide](./damiao#troubleshooting) for more details
+
+### CAN Interface Not Found
+
+Ensure the CAN interface is configured:
+
+```bash
+ip link show can0
+```
+
+## Resources
+
+- [OpenArm Website](https://openarm.dev)
+- [OpenArm Documentation](https://docs.openarm.dev)
+- [OpenArm GitHub](https://github.com/enactic/openarm)
+- [Safety Guide](https://docs.openarm.dev/getting-started/safety-guide)
+- [Damiao Motors and CAN Bus](./damiao)

docs/source/phone_teleop.mdx

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

docs/source/pi0.mdx

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

docs/source/pi05.mdx

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

docs/source/reachy2.mdx

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

docs/source/sarm.mdx

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

docs/source/smolvla.mdx

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

docs/source/so101.mdx

@@ -1,5 +1,18 @@
 # SO-101
 
+<div style="display: flex; align-items: center; gap: 10px;">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Follower.webp"
+    alt="SO-101"
+    width="60%"
+  />
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Leader.webp"
+    alt="SO-101"
+    width="60%"
+  />
+</div>
+
 In the steps below, we explain how to assemble our flagship robot, the SO-101.
 
 ## Source the parts

docs/source/streaming_video_encoding.mdx

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

docs/source/unitree_g1.mdx

@@ -188,7 +188,111 @@ Press `Ctrl+C` to stop the policy.
 
 ## Running in Simulation Mode (MuJoCo)
 
-You can now test policies before unleashing them on the physical robot using MuJoCo. To do so simply set `is_simulation=True` in config.
+You can test policies before deploying on the physical robot using MuJoCo simulation. Set `is_simulation=True` in config or pass `--robot.is_simulation=true` via CLI.
+
+### Calibrate Exoskeleton Teleoperator
+
+```bash
+lerobot-calibrate \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo
+```
+
+### Teleoperate in Simulation
+
+```bash
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
+```
+
+### Record Dataset in Simulation
+
+```bash
+lerobot-record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
+```
+
+Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
+
+---
+
+## Running on Real Robot
+
+Once the robot server is running on the G1 (see Part 3), you can teleoperate and record on the real robot.
+
+### Start the Camera Server
+
+On the robot, start the ZMQ image server:
+
+```bash
+python src/lerobot/cameras/zmq/image_server.py
+```
+
+Keep this running in a separate terminal for camera streaming during recording.
+
+### Teleoperate Real Robot
+
+```bash
+lerobot-teleoperate \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --fps=100
+```
+
+### Record Dataset on Real Robot
+
+```bash
+lerobot-record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true \
+    --dataset.streaming_encoding=true \
+    # --dataset.vcodec=auto \
+    --dataset.encoder_threads=2
+```
+
+**Note**: Update `server_address` to match your robot's camera server IP.
+
+Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
+
+---
 
 ## Additional Resources
 

docs/source/using_dataset_tools.mdx

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

docs/source/walloss.mdx

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

docs/source/xvla.mdx

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

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
+    --dataset.repo_id=<USER>/record-test \
     --dataset.episode=2
 ```
 """
@@ -81,24 +81,25 @@ def replay(cfg: ReplayConfig):
     actions = dataset.hf_dataset.select_columns(ACTION)
     robot.connect()
 
-    log_say("Replaying episode", cfg.play_sounds, blocking=True)
-    for idx in range(dataset.num_frames):
-        start_episode_t = time.perf_counter()
+    try:
+        log_say("Replaying episode", cfg.play_sounds, blocking=True)
+        for idx in range(dataset.num_frames):
+            start_episode_t = time.perf_counter()
 
-        action_array = actions[idx][ACTION]
-        action = {}
-        for i, name in enumerate(dataset.features[ACTION]["names"]):
-            key = f"{name.removeprefix('main_')}.pos"
-            action[key] = action_array[i].item()
+            action_array = actions[idx][ACTION]
+            action = {}
+            for i, name in enumerate(dataset.features[ACTION]["names"]):
+                key = f"{name.removeprefix('main_')}.pos"
+                action[key] = action_array[i].item()
 
-        action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
-        action["elbow_flex.pos"] -= 90
-        robot.send_action(action)
+            action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+            action["elbow_flex.pos"] -= 90
+            robot.send_action(action)
 
-        dt_s = time.perf_counter() - start_episode_t
-        precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
-
-    robot.disconnect()
+            dt_s = time.perf_counter() - start_episode_t
+            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+    finally:
+        robot.disconnect()
 
 
 if __name__ == "__main__":

examples/lekiwi/evaluate.py

@@ -78,40 +78,24 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_evaluate")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting evaluate loop...")
-    recorded_episodes = 0
-    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        recorded_episodes = 0
+        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=teleop_action_processor,
-            robot_action_processor=robot_action_processor,
-            robot_observation_processor=robot_observation_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
@@ -120,24 +104,42 @@ def main():
                 robot_observation_processor=robot_observation_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    control_time_s=EPISODE_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=teleop_action_processor,
+                    robot_action_processor=robot_action_processor,
+                    robot_observation_processor=robot_observation_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            recorded_episodes += 1
 
-    dataset.finalize()
-    dataset.push_to_hub()
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
+
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/record.py

@@ -74,40 +74,23 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_record")
 
-    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+            raise ValueError("Robot or teleop is not connected!")
 
-    print("Starting record loop...")
-    recorded_episodes = 0
-    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {recorded_episodes}")
+        print("Starting record loop...")
+        recorded_episodes = 0
+        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {recorded_episodes}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            dataset=dataset,
-            teleop=[leader_arm, keyboard],
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=teleop_action_processor,
-            robot_action_processor=robot_action_processor,
-            robot_observation_processor=robot_observation_processor,
-        )
-
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                dataset=dataset,
                 teleop=[leader_arm, keyboard],
-                control_time_s=RESET_TIME_SEC,
+                control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=teleop_action_processor,
@@ -115,26 +98,44 @@ def main():
                 robot_observation_processor=robot_observation_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # Reset the environment if not stopping or re-recording
+            if not events["stop_recording"] and (
+                (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+            ):
+                log_say("Reset the environment")
+                record_loop(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    teleop=[leader_arm, keyboard],
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=teleop_action_processor,
+                    robot_action_processor=robot_action_processor,
+                    robot_observation_processor=robot_observation_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    leader_arm.disconnect()
-    keyboard.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            recorded_episodes += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        leader_arm.disconnect()
+        keyboard.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/replay.py

@@ -42,25 +42,27 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()
 
-        # Get recorded action from dataset
-        action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }
 
-        # Send action to robot
-        _ = robot.send_action(action)
+            # Send action to robot
+            _ = robot.send_action(action)
 
-        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-
-    robot.disconnect()
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+    finally:
+        robot.disconnect()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/evaluate.py

@@ -142,38 +142,24 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_evaluate")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting evaluate loop...")
-    episode_idx = 0
-    for episode_idx in range(NUM_EPISODES):
-        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        episode_idx = 0
+        for episode_idx in range(NUM_EPISODES):
+            log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=make_default_teleop_action_processor(),
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose_processor,
-        )
-
-        # 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")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
@@ -182,24 +168,41 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # 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(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    control_time_s=EPISODE_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=make_default_teleop_action_processor(),
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/record.py

@@ -149,38 +149,23 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_record")
 
-    if not robot.is_connected or not phone.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not robot.is_connected or not phone.is_connected:
+            raise ValueError("Robot or teleop is not connected!")
 
-    print("Starting record loop. Move your phone to teleoperate the robot...")
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting record loop. Move your phone to teleoperate the robot...")
+        episode_idx = 0
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            teleop=phone,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=phone_to_robot_ee_pose_processor,
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose,
-        )
-
-        # 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")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
                 teleop=phone,
-                control_time_s=RESET_TIME_SEC,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -188,25 +173,43 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # 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(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    teleop=phone,
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=phone_to_robot_ee_pose_processor,
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    phone.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        phone.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/replay.py

@@ -73,32 +73,34 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()
 
-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            ee_action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+            # Get robot observation
+            robot_obs = robot.get_observation()
 
-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+            # Dataset EE -> robot joints
+            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+            # Send action to robot
+            _ = robot.send_action(joint_action)
 
-        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-
-    # Clean up
-    robot.disconnect()
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+    finally:
+        # Clean up
+        robot.disconnect()
 
 
 if __name__ == "__main__":

examples/rtc/eval_dataset.py

@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
     # Basic usage with smolvla policy
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=lipsop/reuben_pi0 \
-        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --policy.path=<USER>/reuben_pi0 \
+        --dataset.repo_id=<USER>/so101_cube_in_cup \
         --rtc.execution_horizon=8 \
         --device=cuda
 
     # With torch.compile for faster inference (PyTorch 2.0+)
     # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=helper2424/check_rtc \
+        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --dataset.repo_id=<USER>/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \

examples/rtc/eval_with_real_robot.py

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

examples/so100_to_so100_EE/evaluate.py

@@ -142,38 +142,24 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="so100_so100_evaluate")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting evaluate loop...")
-    episode_idx = 0
-    for episode_idx in range(NUM_EPISODES):
-        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting evaluate loop...")
+        episode_idx = 0
+        for episode_idx in range(NUM_EPISODES):
+            log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=robot,
-            events=events,
-            fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=make_default_teleop_action_processor(),
-            robot_action_processor=robot_ee_to_joints_processor,
-            robot_observation_processor=robot_joints_to_ee_pose_processor,
-        )
-
-        # 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")
+            # Main record loop
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
+                policy=policy,
+                preprocessor=preprocessor,  # Pass the pre and post policy processors
+                postprocessor=postprocessor,
+                dataset=dataset,
                 control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
@@ -182,24 +168,41 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose_processor,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # 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(
+                    robot=robot,
+                    events=events,
+                    fps=FPS,
+                    control_time_s=EPISODE_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=make_default_teleop_action_processor(),
+                    robot_action_processor=robot_ee_to_joints_processor,
+                    robot_observation_processor=robot_joints_to_ee_pose_processor,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-record episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        robot.disconnect()
+        listener.stop()
 
-    dataset.finalize()
-    dataset.push_to_hub()
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/record.py

@@ -146,38 +146,23 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="recording_phone")
 
-    if not leader.is_connected or not follower.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+    try:
+        if not leader.is_connected or not follower.is_connected:
+            raise ValueError("Robot or teleop is not connected!")
 
-    print("Starting record loop...")
-    episode_idx = 0
-    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+        print("Starting record loop...")
+        episode_idx = 0
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
 
-        # Main record loop
-        record_loop(
-            robot=follower,
-            events=events,
-            fps=FPS,
-            teleop=leader,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
-            single_task=TASK_DESCRIPTION,
-            display_data=True,
-            teleop_action_processor=leader_joints_to_ee,
-            robot_action_processor=ee_to_follower_joints,
-            robot_observation_processor=follower_joints_to_ee,
-        )
-
-        # 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")
+            # Main record loop
             record_loop(
                 robot=follower,
                 events=events,
                 fps=FPS,
                 teleop=leader,
-                control_time_s=RESET_TIME_SEC,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=leader_joints_to_ee,
@@ -185,25 +170,44 @@ def main():
                 robot_observation_processor=follower_joints_to_ee,
             )
 
-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+            # 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(
+                    robot=follower,
+                    events=events,
+                    fps=FPS,
+                    teleop=leader,
+                    control_time_s=RESET_TIME_SEC,
+                    single_task=TASK_DESCRIPTION,
+                    display_data=True,
+                    teleop_action_processor=leader_joints_to_ee,
+                    robot_action_processor=ee_to_follower_joints,
+                    robot_observation_processor=follower_joints_to_ee,
+                )
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
 
-    # Clean up
-    log_say("Stop recording")
-    leader.disconnect()
-    follower.disconnect()
-    listener.stop()
+            # Save episode
+            dataset.save_episode()
+            episode_idx += 1
 
-    dataset.finalize()
-    dataset.push_to_hub()
+    finally:
+        # Clean up
+        log_say("Stop recording")
+        leader.disconnect()
+        follower.disconnect()
+        listener.stop()
+
+        dataset.finalize()
+        dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/replay.py

@@ -74,32 +74,35 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+    try:
+        if not robot.is_connected:
+            raise ValueError("Robot is not connected!")
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+        print("Starting replay loop...")
+        log_say(f"Replaying episode {EPISODE_IDX}")
+        for idx in range(len(episode_frames)):
+            t0 = time.perf_counter()
 
-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+            # Get recorded action from dataset
+            ee_action = {
+                name: float(actions[idx][ACTION][i])
+                for i, name in enumerate(dataset.features[ACTION]["names"])
+            }
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+            # Get robot observation
+            robot_obs = robot.get_observation()
 
-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+            # Dataset EE -> robot joints
+            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+            # Send action to robot
+            _ = robot.send_action(joint_action)
 
-        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
-    # Clean up
-    robot.disconnect()
+    finally:
+        # Clean up
+        robot.disconnect()
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -59,7 +59,7 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<4.2.0",
+    "datasets>=4.0.0,<5.0.0",
     "diffusers>=0.27.2,<0.36.0",
     "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
     "accelerate>=1.10.0,<2.0.0",
@@ -76,9 +76,9 @@ dependencies = [
     "pyserial>=3.5,<4.0",
     "wandb>=0.24.0,<0.25.0",
 
-    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
-    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
-    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
+    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
+    "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')", # TODO: Bump dependency
+    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
 
     "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
@@ -98,18 +98,26 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
+can-dep = ["python-can>=4.2.0,<5.0.0"]
 
 # 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]"]
 
 # Robots
+openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
     "pyzmq>=26.2.1,<28.0.0",
-    "onnxruntime>=1.16.0,<2.0.0"
+    "onnxruntime>=1.16.0,<2.0.0",
+    "pin>=3.0.0,<4.0.0",
+    "meshcat>=0.3.0,<0.4.0",
+    "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]"]
@@ -203,8 +211,12 @@ lerobot-info="lerobot.scripts.lerobot_info:main"
 lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 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"]
 
@@ -278,6 +290,7 @@ default.extend-ignore-identifiers-re = [
     "thw",
     "inpt",
     "ROBOTIS",
+    "OT_VALUE"
 ]
 
 # TODO: Uncomment when ready to use
@@ -352,9 +365,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.motors.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.motors.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"

src/lerobot/async_inference/robot_client.py

@@ -49,23 +49,18 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.robots import (  # noqa: F401
-    Robot,
-    RobotConfig,
-    bi_so_follower,
-    koch_follower,
+from lerobot.robots import (
+    RobotConfig,  # noqa: F401
     make_robot_from_config,
-    omx_follower,
-    so_follower,
 )
 from lerobot.transport import (
     services_pb2,  # type: ignore
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
+from lerobot.utils.import_utils import register_third_party_plugins
 
 from .configs import RobotClientConfig
-from .constants import SUPPORTED_ROBOTS
 from .helpers import (
     Action,
     FPSTracker,
@@ -485,8 +480,9 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    if cfg.robot.type not in SUPPORTED_ROBOTS:
-        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    # TODO: Assert if checking robot support is still needed with the plugin system
+    # if cfg.robot.type not in SUPPORTED_ROBOTS:
+    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -512,4 +508,5 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
+    register_third_party_plugins()
     async_client()  # run the client

src/lerobot/cameras/__init__.py

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

src/lerobot/cameras/camera.py

@@ -15,11 +15,12 @@
 # limitations under the License.
 
 import abc
+import warnings
 from typing import Any
 
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-from .configs import CameraConfig, ColorMode
+from .configs import CameraConfig
 
 
 class Camera(abc.ABC):
@@ -30,20 +31,12 @@ class Camera(abc.ABC):
 
     Manages basic camera properties (FPS, resolution) and core operations:
     - Connection/disconnection
-    - Frame capture (sync/async)
+    - Frame capture (sync/async/latest)
 
     Attributes:
         fps (int | None): Configured frames per second
         width (int | None): Frame width in pixels
         height (int | None): Frame height in pixels
-
-    Example:
-        class MyCamera(Camera):
-            def __init__(self, config): ...
-            @property
-            def is_connected(self) -> bool: ...
-            def connect(self, warmup=True): ...
-            # Plus other required methods
     """
 
     def __init__(self, config: CameraConfig):
@@ -56,6 +49,32 @@ class Camera(abc.ABC):
         self.width: int | None = config.width
         self.height: int | None = config.height
 
+    def __enter__(self):
+        """
+        Context manager entry.
+        Automatically connects to the camera.
+        """
+        self.connect()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback) -> None:
+        """
+        Context manager exit.
+        Automatically disconnects, ensuring resources are released even on error.
+        """
+        self.disconnect()
+
+    def __del__(self) -> None:
+        """
+        Destructor safety net.
+        Attempts to disconnect if the object is garbage collected without cleanup.
+        """
+        try:
+            if self.is_connected:
+                self.disconnect()
+        except Exception:  # nosec B110
+            pass
+
     @property
     @abc.abstractmethod
     def is_connected(self) -> bool:
@@ -89,12 +108,10 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """Capture and return a single frame from the camera.
+    def read(self) -> NDArray[Any]:
+        """Capture and return a single frame from the camera synchronously.
 
-        Args:
-            color_mode: Desired color mode for the output frame. If None,
-                        uses the camera's default color mode.
+        This is a blocking call that will wait for the hardware and its SDK.
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
@@ -103,17 +120,64 @@ class Camera(abc.ABC):
 
     @abc.abstractmethod
     def async_read(self, timeout_ms: float = ...) -> NDArray[Any]:
-        """Asynchronously capture and return a single frame from the camera.
+        """Return the most recent new frame.
+
+        This method retrieves the latest frame captured by the background thread.
+        If a new frame is already available in the buffer (captured since the last call),
+        it returns it immediately.
+
+        It blocks up to `timeout_ms` only if the buffer is empty or if the latest frame
+        was already consumed by a previous `async_read` call.
+
+        Essentially, this method return the latest unconsumed frame, waiting if necessary
+        for a new one to arrive within the specified timeout.
+
+        Usage:
+            - Ideal for control loops where you want to ensure every processed frame
+            is fresh, effectively synchronizing your loop to the camera's FPS.
+            - Causes of a timeout usually include: very low camera FPS, heavy processing load,
+            or if the camera is disconnected.
 
         Args:
-            timeout_ms: Maximum time to wait for a frame in milliseconds.
-                        Defaults to implementation-specific timeout.
+            timeout_ms: Maximum time to wait for a new frame in milliseconds.
+                        Defaults to 200ms (0.2s).
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
+
+        Raises:
+            TimeoutError: If no new frame arrives within `timeout_ms`.
         """
         pass
 
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Usage:
+            Ideal for scenarios requiring zero latency or decoupled frequencies & when
+            we want a guaranteed frame, such as UI visualization, logging, or
+            non-critical monitoring.
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            NotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+        warnings.warn(
+            f"{self.__class__.__name__}.read_latest() is not implemented. "
+            "Please override read_latest(); it will be required in future releases.",
+            FutureWarning,
+            stacklevel=2,
+        )
+        return self.async_read()
+
     @abc.abstractmethod
     def disconnect(self) -> None:
         """Disconnect from the camera and release resources."""

src/lerobot/cameras/configs.py

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

src/lerobot/cameras/opencv/camera_opencv.py

@@ -32,10 +32,11 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_backend, get_cv2_rotation
+from ..utils import get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -70,34 +71,24 @@ class OpenCVCamera(Camera):
     Example:
         ```python
         from lerobot.cameras.opencv import OpenCVCamera
-        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
+        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig
 
         # Basic usage with camera index 0
         config = OpenCVCameraConfig(index_or_path=0)
         camera = OpenCVCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously
+        # Read 1 frame synchronously (blocking)
         color_image = camera.read()
-        print(color_image.shape)
 
-        # Read 1 frame asynchronously
+        # Read 1 frame asynchronously (waits for new frame with a timeout)
         async_image = camera.async_read()
 
+        # Get the latest frame immediately (no wait, returns timestamp)
+        latest_image, timestamp = camera.read_latest()
+
         # When done, properly disconnect the camera using
         camera.disconnect()
-
-        # Example with custom settings
-        custom_config = OpenCVCameraConfig(
-            index_or_path='/dev/video0', # Or use an index
-            fps=30,
-            width=1280,
-            height=720,
-            color_mode=ColorMode.RGB,
-            rotation=Cv2Rotation.ROTATE_90
-        )
-        custom_camera = OpenCVCamera(custom_config)
-        # ... connect, read, disconnect ...
         ```
     """
 
@@ -123,10 +114,11 @@ class OpenCVCamera(Camera):
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = get_cv2_backend()
+        self.backend: int = config.backend
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -141,20 +133,23 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
+    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
 
         Initializes the OpenCV VideoCapture object, sets desired camera properties
-        (FPS, width, height), and performs initial checks.
+        (FPS, width, height), starts the background reading thread and performs initial checks.
+
+        Args:
+            warmup (bool): If True, waits at connect() time until at least one valid frame
+                           has been captured by the background thread. Defaults to True.
 
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
-            ConnectionError: If the specified camera index/path is not found or the camera is found but fails to open.
-            RuntimeError: If the camera opens but fails to apply requested FPS/resolution settings.
+            ConnectionError: If the specified camera index/path is not found or fails to open.
+            RuntimeError: If the camera opens but fails to apply requested settings.
         """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         # Use 1 thread for OpenCV operations to avoid potential conflicts or
         # blocking in multi-threaded applications, especially during data collection.
@@ -170,15 +165,20 @@ class OpenCVCamera(Camera):
             )
 
         self._configure_capture_settings()
+        self._start_read_thread()
 
-        if warmup:
+        if warmup and self.warmup_s > 0:
             start_time = time.time()
             while time.time() - start_time < self.warmup_s:
-                self.read()
+                self.async_read(timeout_ms=self.warmup_s * 1000)
                 time.sleep(0.1)
+            with self.frame_lock:
+                if self.latest_frame is None:
+                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
 
+    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """
         Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -196,11 +196,8 @@ class OpenCVCamera(Camera):
         Raises:
             RuntimeError: If the camera fails to set any of the specified properties
                           to the requested value.
-            DeviceNotConnectedError: If the camera is not connected when attempting
-                                     to configure settings.
+            DeviceNotConnectedError: If the camera is not connected.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
         # Set FOURCC first (if specified) as it can affect available FPS/resolution options
         if self.config.fourcc is not None:
@@ -339,6 +336,18 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
+    def _read_from_hardware(self) -> NDArray[Any]:
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        ret, frame = self.videocapture.read()
+
+        if not ret:
+            raise RuntimeError(f"{self} read failed (status={ret}).")
+
+        return frame
+
+    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -346,11 +355,6 @@ class OpenCVCamera(Camera):
         This is a blocking call. It waits for the next available frame from the
         camera hardware via OpenCV.
 
-        Args:
-            color_mode (Optional[ColorMode]): If specified, overrides the default
-                color mode (`self.color_mode`) for this read operation (e.g.,
-                request RGB even if default is BGR).
-
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
                        (height, width, channels), using the specified or default
@@ -362,34 +366,31 @@ class OpenCVCamera(Camera):
                           received frame dimensions don't match expectations before rotation.
             ValueError: If an invalid `color_mode` is requested.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         start_time = time.perf_counter()
 
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
 
-        ret, frame = self.videocapture.read()
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        processed_frame = self._postprocess_image(frame, color_mode)
+        self.new_frame_event.clear()
+        frame = self.async_read(timeout_ms=10000)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return processed_frame
+        return frame
 
-    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
+    def _postprocess_image(self, image: NDArray[Any]) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw frame.
 
         Args:
             image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
-            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
-                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame.
@@ -399,11 +400,10 @@ class OpenCVCamera(Camera):
             RuntimeError: If the raw frame dimensions do not match the configured
                           `width` and `height`.
         """
-        requested_color_mode = self.color_mode if color_mode is None else color_mode
 
-        if requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         h, w, c = image.shape
@@ -417,7 +417,7 @@ class OpenCVCamera(Camera):
             raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
 
         processed_image = image
-        if requested_color_mode == ColorMode.RGB:
+        if self.color_mode == ColorMode.RGB:
             processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
         if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE, cv2.ROTATE_180]:
@@ -431,7 +431,7 @@ class OpenCVCamera(Camera):
 
         On each iteration:
         1. Reads a color frame
-        2. Stores result in latest_frame (thread-safe)
+        2. Stores result in latest_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -439,30 +439,37 @@ class OpenCVCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
+        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                color_image = self.read()
+                raw_frame = self._read_from_hardware()
+                processed_frame = self._postprocess_image(raw_frame)
+                capture_time = time.perf_counter()
 
                 with self.frame_lock:
-                    self.latest_frame = color_image
+                    self.latest_frame = processed_frame
+                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
+                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                if failure_count <= 10:
+                    failure_count += 1
+                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                else:
+                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
 
     def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=0.1)
-        if self.stop_event is not None:
-            self.stop_event.set()
+        self._stop_read_thread()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
         self.thread.daemon = True
         self.thread.start()
+        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
@@ -475,6 +482,12 @@ class OpenCVCamera(Camera):
         self.thread = None
         self.stop_event = None
 
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -482,6 +495,7 @@ class OpenCVCamera(Camera):
         This method retrieves the most recent frame captured by the background
         read thread. It does not block waiting for the camera hardware directly,
         but may wait up to timeout_ms for the background thread to provide a frame.
+        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -496,17 +510,14 @@ class OpenCVCamera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
-            self._start_read_thread()
+            raise RuntimeError(f"{self} read thread is not running.")
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
             raise TimeoutError(
                 f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
+                f"Read thread alive: {self.thread.is_alive()}."
             )
 
         with self.frame_lock:
@@ -518,6 +529,41 @@ class OpenCVCamera(Camera):
 
         return frame
 
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
     def disconnect(self) -> None:
         """
         Disconnects from the camera and cleans up resources.
@@ -538,4 +584,9 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
 
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/opencv/configuration_opencv.py

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

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

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

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -32,6 +32,7 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
+from lerobot.utils.decorators import check_if_not_connected
 from lerobot.utils.import_utils import _reachy2_sdk_available
 
 if TYPE_CHECKING or _reachy2_sdk_available:
@@ -80,6 +81,8 @@ class Reachy2Camera(Camera):
         self.config = config
 
         self.color_mode = config.color_mode
+        self.latest_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
 
         self.cam_manager: CameraManager | None = None
 
@@ -121,16 +124,12 @@ class Reachy2Camera(Camera):
         """
         raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")
 
+    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
-        This is a blocking call.
-
-        Args:
-            color_mode (Optional[ColorMode]): If specified, overrides the default
-                color mode (`self.color_mode`) for this read operation (e.g.,
-                request RGB even if default is BGR).
+        This method retrieves the most recent frame available in Reachy 2's low-level software.
 
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
@@ -139,12 +138,14 @@ class Reachy2Camera(Camera):
         """
         start_time = time.perf_counter()
 
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
         if self.cam_manager is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
+
         frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
 
         if self.config.name == "teleop" and hasattr(self.cam_manager, "teleop"):
@@ -165,25 +166,27 @@ class Reachy2Camera(Camera):
             raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
 
         if frame is None:
-            return np.empty((0, 0, 3), dtype=np.uint8)
+            raise RuntimeError(f"Internal error: No frame available for {self}.")
 
-        if self.config.color_mode == "rgb":
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+            )
+        if self.color_mode == ColorMode.RGB:
             frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
+        self.latest_frame = frame
+        self.latest_timestamp = time.perf_counter()
+
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
         return frame
 
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
-        Reads the latest available frame.
-
-        This method retrieves the most recent frame available in Reachy 2's low-level software.
-
-        Args:
-            timeout_ms (float): Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 200ms (0.2 seconds).
+        Same as read()
 
         Returns:
             np.ndarray: The latest captured frame as a NumPy array in the format
@@ -194,16 +197,40 @@ class Reachy2Camera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        frame = self.read()
+        return self.read()
 
-        if frame is None:
-            raise RuntimeError(f"Internal error: No frame available for {self}.")
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
 
-        return frame
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
 
+        Returns:
+            tuple[NDArray, float]:
+                - The frame image (numpy array).
+                - The timestamp (time.perf_counter) when this frame was captured.
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.latest_frame is None or self.latest_timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - self.latest_timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return self.latest_frame
+
+    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -211,8 +238,6 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} not connected.")
 
         if self.cam_manager is not None:
             self.cam_manager.disconnect()

src/lerobot/cameras/realsense/camera_realsense.py

@@ -30,7 +30,8 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -72,15 +73,14 @@ class RealSenseCamera(Camera):
         camera = RealSenseCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously
+        # Read 1 frame synchronously (blocking)
         color_image = camera.read()
-        print(color_image.shape)
 
-        # Read 1 frame asynchronously
+        # Read 1 frame asynchronously (waits for new frame with a timeout)
         async_image = camera.async_read()
 
-        # When done, properly disconnect the camera using
-        camera.disconnect()
+        # Get the latest frame immediately (no wait, returns timestamp)
+        latest_image, timestamp = camera.read_latest()
 
         # Example with depth capture and custom settings
         custom_config = RealSenseCameraConfig(
@@ -133,7 +133,9 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: NDArray[Any] | None = None
+        self.latest_color_frame: NDArray[Any] | None = None
+        self.latest_depth_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -151,6 +153,7 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
+    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
@@ -158,14 +161,16 @@ class RealSenseCamera(Camera):
         Initializes the RealSense pipeline, configures the required streams (color
         and optionally depth), starts the pipeline, and validates the actual stream settings.
 
+        Args:
+            warmup (bool): If True, waits at connect() time until at least one valid frame
+                           has been captured by the background thread. Defaults to True.
+
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
             ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         self.rs_pipeline = rs.pipeline()
         rs_config = rs.config()
@@ -181,15 +186,18 @@ class RealSenseCamera(Camera):
             ) from e
 
         self._configure_capture_settings()
+        self._start_read_thread()
 
-        if warmup:
-            time.sleep(
-                1
-            )  # NOTE(Steven): RS cameras need a bit of time to warm up before the first read. If we don't wait, the first read from the warmup will raise.
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                self.read()
-                time.sleep(0.1)
+        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
+        self.warmup_s = max(self.warmup_s, 1)
+
+        start_time = time.time()
+        while time.time() - start_time < self.warmup_s:
+            self.async_read(timeout_ms=self.warmup_s * 1000)
+            time.sleep(0.1)
+        with self.frame_lock:
+            if self.latest_color_frame is None or self.use_depth and self.latest_depth_frame is None:
+                raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
 
@@ -282,6 +290,7 @@ class RealSenseCamera(Camera):
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
+    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
@@ -291,8 +300,6 @@ class RealSenseCamera(Camera):
         Raises:
             DeviceNotConnectedError: If device is not connected.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -312,6 +319,7 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
+    @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
@@ -319,9 +327,6 @@ class RealSenseCamera(Camera):
         This is a blocking call. It waits for a coherent set of frames (depth)
         from the camera hardware via the RealSense pipeline.
 
-        Args:
-            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
-
         Returns:
             np.ndarray: The depth map as a NumPy array (height, width)
                   of type `np.uint16` (raw depth values in millimeters) and rotation.
@@ -330,44 +335,50 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If reading frames from the pipeline fails or frames are invalid.
         """
+        if timeout_ms:
+            logger.warning(
+                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
+            )
 
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
         if not self.use_depth:
             raise RuntimeError(
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
-        start_time = time.perf_counter()
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
+        self.new_frame_event.clear()
+
+        _ = self.async_read(timeout_ms=10000)
+
+        with self.frame_lock:
+            depth_map = self.latest_depth_frame
+
+        if depth_map is None:
+            raise RuntimeError("No depth frame available. Ensure camera is streaming.")
+
+        return depth_map
+
+    def _read_from_hardware(self):
         if self.rs_pipeline is None:
             raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
 
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=10000)
 
         if not ret or frame is None:
-            raise RuntimeError(f"{self} read_depth failed (status={ret}).")
+            raise RuntimeError(f"{self} read failed (status={ret}).")
 
-        depth_frame = frame.get_depth_frame()
-        depth_map = np.asanyarray(depth_frame.get_data())
+        return frame
 
-        depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return depth_map_processed
-
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
+    @check_if_not_connected
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
 
         This is a blocking call. It waits for a coherent set of frames (color)
         from the camera hardware via the RealSense pipeline.
 
-        Args:
-            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
-
         Returns:
             np.ndarray: The captured color frame as a NumPy array
               (height, width, channels), processed according to `color_mode` and rotation.
@@ -378,39 +389,36 @@ class RealSenseCamera(Camera):
             ValueError: If an invalid `color_mode` is requested.
         """
 
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
         start_time = time.perf_counter()
 
-        if self.rs_pipeline is None:
-            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
 
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
+        if timeout_ms:
+            logger.warning(
+                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
+            )
 
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
-        color_frame = frame.get_color_frame()
-        color_image_raw = np.asanyarray(color_frame.get_data())
+        self.new_frame_event.clear()
 
-        color_image_processed = self._postprocess_image(color_image_raw, color_mode)
+        frame = self.async_read(timeout_ms=10000)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return color_image_processed
+        return frame
 
-    def _postprocess_image(
-        self, image: NDArray[Any], color_mode: ColorMode | None = None, depth_frame: bool = False
-    ) -> NDArray[Any]:
+    def _postprocess_image(self, image: NDArray[Any], depth_frame: bool = False) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw color frame.
 
         Args:
             image (np.ndarray): The raw image frame (expected RGB format from RealSense).
-            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
-                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
@@ -421,9 +429,9 @@ class RealSenseCamera(Camera):
                           `width` and `height`.
         """
 
-        if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if self.color_mode and self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid requested color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         if depth_frame:
@@ -454,7 +462,7 @@ class RealSenseCamera(Camera):
 
         On each iteration:
         1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame (thread-safe)
+        2. Stores result in latest_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -462,25 +470,41 @@ class RealSenseCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
+        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                color_image = self.read(timeout_ms=500)
+                frame = self._read_from_hardware()
+                color_frame_raw = frame.get_color_frame()
+                color_frame = np.asanyarray(color_frame_raw.get_data())
+                processed_color_frame = self._postprocess_image(color_frame)
+
+                if self.use_depth:
+                    depth_frame_raw = frame.get_depth_frame()
+                    depth_frame = np.asanyarray(depth_frame_raw.get_data())
+                    processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True)
+
+                capture_time = time.perf_counter()
 
                 with self.frame_lock:
-                    self.latest_frame = color_image
+                    self.latest_color_frame = processed_color_frame
+                    if self.use_depth:
+                        self.latest_depth_frame = processed_depth_frame
+                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
+                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                if failure_count <= 10:
+                    failure_count += 1
+                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
+                else:
+                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
 
     def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=0.1)
-        if self.stop_event is not None:
-            self.stop_event.set()
+        self._stop_read_thread()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
@@ -498,7 +522,14 @@ class RealSenseCamera(Camera):
         self.thread = None
         self.stop_event = None
 
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
     # NOTE(Steven): Missing implementation for depth for now
+    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
@@ -506,6 +537,7 @@ class RealSenseCamera(Camera):
         This method retrieves the most recent color frame captured by the background
         read thread. It does not block waiting for the camera hardware directly,
         but may wait up to timeout_ms for the background thread to provide a frame.
+        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -520,21 +552,18 @@ class RealSenseCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
-            self._start_read_thread()
+            raise RuntimeError(f"{self} read thread is not running.")
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
             raise TimeoutError(
                 f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
+                f"Read thread alive: {self.thread.is_alive()}."
             )
 
         with self.frame_lock:
-            frame = self.latest_frame
+            frame = self.latest_color_frame
             self.new_frame_event.clear()
 
         if frame is None:
@@ -542,6 +571,42 @@ class RealSenseCamera(Camera):
 
         return frame
 
+    # NOTE(Steven): Missing implementation for depth for now
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent (color) frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_color_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
     def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
@@ -565,4 +630,10 @@ class RealSenseCamera(Camera):
             self.rs_pipeline = None
             self.rs_profile = None
 
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/configuration_realsense.py

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

src/lerobot/cameras/utils.py

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

src/lerobot/cameras/zmq/camera_zmq.py

@@ -34,7 +34,8 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -45,6 +46,12 @@ logger = logging.getLogger(__name__)
 
 class ZMQCamera(Camera):
     """
+    Manages camera interactions via ZeroMQ for receiving frames from a remote server.
+
+    This class connects to a ZMQ Publisher, subscribes to frame topics, and decodes
+    incoming JSON messages containing Base64 encoded images. It supports both
+    synchronous and asynchronous frame reading patterns.
+
     Example usage:
         ```python
         from lerobot.cameras.zmq import ZMQCamera, ZMQCameraConfig
@@ -52,7 +59,16 @@ class ZMQCamera(Camera):
         config = ZMQCameraConfig(server_address="192.168.123.164", port=5555, camera_name="head_camera")
         camera = ZMQCamera(config)
         camera.connect()
-        frame = camera.read()
+
+        # Read 1 frame synchronously (blocking)
+        color_image = camera.read()
+
+        # Read 1 frame asynchronously (waits for new frame with a timeout)
+        async_image = camera.async_read()
+
+        # Get the latest frame immediately (no wait, returns timestamp)
+        latest_image, timestamp = camera.read_latest()
+
         camera.disconnect()
         ```
     """
@@ -68,14 +84,17 @@ class ZMQCamera(Camera):
         self.color_mode = config.color_mode
         self.timeout_ms = config.timeout_ms
 
+        # ZMQ Context and Socket
         self.context: zmq.Context | None = None
         self.socket: zmq.Socket | None = None
         self._connected = False
 
+        # Threading resources
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
+        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
     def __str__(self) -> str:
@@ -83,12 +102,17 @@ class ZMQCamera(Camera):
 
     @property
     def is_connected(self) -> bool:
+        """Checks if the ZMQ socket is initialized and connected."""
         return self._connected and self.context is not None and self.socket is not None
 
+    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
-        """Connect to ZMQ camera server."""
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
+        """Connect to ZMQ camera server.
+
+        Args:
+            warmup (bool): If True, waits for the camera to provide at least one
+                           valid frame before returning. Defaults to True.
+        """
 
         logger.info(f"Connecting to {self}...")
 
@@ -103,17 +127,28 @@ class ZMQCamera(Camera):
             self.socket.connect(f"tcp://{self.server_address}:{self.port}")
             self._connected = True
 
-            # Auto-detect resolution
+            # Auto-detect resolution if not provided
             if self.width is None or self.height is None:
-                h, w = self.read().shape[:2]
+                # Read directly from hardware because the thread isn't running yet
+                temp_frame = self._read_from_hardware()
+                h, w = temp_frame.shape[:2]
                 self.height = h
                 self.width = w
-                logger.info(f"{self} resolution: {w}x{h}")
+                logger.info(f"{self} resolution detected: {w}x{h}")
 
+            self._start_read_thread()
             logger.info(f"{self} connected.")
 
             if warmup:
-                time.sleep(0.1)
+                # Ensure we have captured at least one frame via the thread
+                start_time = time.time()
+                while time.time() - start_time < (self.config.warmup_s):  # Wait a bit more than timeout
+                    self.async_read(timeout_ms=self.config.warmup_s * 1000)
+                    time.sleep(0.1)
+
+                with self.frame_lock:
+                    if self.latest_frame is None:
+                        raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         except Exception as e:
             self._cleanup()
@@ -131,15 +166,14 @@ class ZMQCamera(Camera):
 
     @staticmethod
     def find_cameras() -> list[dict[str, Any]]:
-        """ZMQ cameras require manual configuration (server address/port)."""
-        return []
-
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
-        Read a single frame from the ZMQ camera.
+        Detection not implemented for ZMQ cameras. These cameras require manual configuration (server address/port).
+        """
+        raise NotImplementedError("Camera detection is not implemented for ZMQ cameras.")
 
-        Returns:
-            np.ndarray: Decoded frame (height, width, 3)
+    def _read_from_hardware(self) -> NDArray[Any]:
+        """
+        Reads a single frame directly from the ZMQ socket.
         """
         if not self.is_connected or self.socket is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
@@ -147,6 +181,7 @@ class ZMQCamera(Camera):
         try:
             message = self.socket.recv_string()
         except Exception as e:
+            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
             if type(e).__name__ == "Again":
                 raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
             raise
@@ -176,42 +211,114 @@ class ZMQCamera(Camera):
 
         return frame
 
+    @check_if_not_connected
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+        """
+        Reads a single frame synchronously from the camera.
+
+        This is a blocking call. It waits for the next available frame from the
+        camera background thread.
+
+        Returns:
+            np.ndarray: Decoded frame (height, width, 3)
+        """
+        start_time = time.perf_counter()
+
+        if color_mode is not None:
+            logger.warning(
+                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
+            )
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        self.new_frame_event.clear()
+        frame = self.async_read(timeout_ms=10000)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
+
+        return frame
+
     def _read_loop(self) -> None:
-        while self.stop_event and not self.stop_event.is_set():
+        """
+        Internal loop run by the background thread for asynchronous reading.
+        """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized.")
+
+        failure_count = 0
+        while not self.stop_event.is_set():
             try:
-                frame = self.read()
+                frame = self._read_from_hardware()
+                capture_time = time.perf_counter()
+
                 with self.frame_lock:
                     self.latest_frame = frame
+                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
+                failure_count = 0
+
             except DeviceNotConnectedError:
                 break
-            except TimeoutError:
-                pass
-            except Exception as e:
-                logger.warning(f"Read error: {e}")
+            except (TimeoutError, Exception) as e:
+                if failure_count <= 10:
+                    failure_count += 1
+                    logger.warning(f"Read error: {e}")
+                else:
+                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
 
     def _start_read_thread(self) -> None:
-        if self.thread and self.thread.is_alive():
-            return
+        if self.stop_event is not None:
+            self.stop_event.set()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, daemon=True)
+        self.thread = Thread(target=self._read_loop, daemon=True, name=f"{self}_read_loop")
         self.thread.start()
+        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
-        if self.stop_event:
+        if self.stop_event is not None:
             self.stop_event.set()
-        if self.thread and self.thread.is_alive():
+
+        if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+
         self.thread = None
         self.stop_event = None
 
-    def async_read(self, timeout_ms: float = 10000) -> NDArray[Any]:
-        """Read latest frame asynchronously (non-blocking)."""
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
 
-        if not self.thread or not self.thread.is_alive():
-            self._start_read_thread()
+    @check_if_not_connected
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
+        """
+        Reads the latest available frame asynchronously.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available. Defaults to 200ms.
+
+        Returns:
+            np.ndarray: The latest captured frame.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            TimeoutError: If no frame data becomes available within the specified timeout.
+            RuntimeError: If the background thread is not running.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
             raise TimeoutError(f"{self} async_read timeout after {timeout_ms}ms")
@@ -225,11 +332,54 @@ class ZMQCamera(Camera):
 
         return frame
 
+    @check_if_not_connected
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+        """Return the most recent frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        memory buffer. The frame may be stale,
+        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+
+        Returns:
+            NDArray[Any]: The frame image (numpy array).
+
+        Raises:
+            TimeoutError: If the latest frame is older than `max_age_ms`.
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If the camera is connected but has not captured any frames yet.
+        """
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
     def disconnect(self) -> None:
         """Disconnect from ZMQ camera."""
-        if not self.is_connected and not self.thread:
+        if not self.is_connected and self.thread is None:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        self._stop_read_thread()
+        if self.thread is not None:
+            self._stop_read_thread()
+
         self._cleanup()
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -29,12 +29,10 @@ class ZMQCameraConfig(CameraConfig):
     camera_name: str = "zmq_camera"
     color_mode: ColorMode = ColorMode.RGB
     timeout_ms: int = 5000
+    warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
+        self.color_mode = ColorMode(self.color_mode)
 
         if self.timeout_ms <= 0:
             raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")

src/lerobot/configs/policies.py

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

src/lerobot/datasets/aggregate.py

@@ -116,6 +116,9 @@ def update_meta_data(
     Adjusts all indices and timestamps to account for previously aggregated
     data and videos in the destination dataset.
 
+    For data file indices, uses the 'src_to_dst' mapping from aggregate_data()
+    to correctly map source file indices to their destination locations.
+
     Args:
         df: DataFrame containing the metadata to be updated.
         dst_meta: Destination dataset metadata.
@@ -129,8 +132,50 @@ def update_meta_data(
 
     df["meta/episodes/chunk_index"] = df["meta/episodes/chunk_index"] + meta_idx["chunk"]
     df["meta/episodes/file_index"] = df["meta/episodes/file_index"] + meta_idx["file"]
-    df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
-    df["data/file_index"] = df["data/file_index"] + data_idx["file"]
+
+    # Update data file indices using source-to-destination mapping
+    # This is critical for handling datasets that are already results of a merge
+    data_src_to_dst = data_idx.get("src_to_dst", {})
+    if data_src_to_dst:
+        # Store original indices for lookup
+        df["_orig_data_chunk"] = df["data/chunk_index"].copy()
+        df["_orig_data_file"] = df["data/file_index"].copy()
+
+        # Vectorized mapping from (src_chunk, src_file) to (dst_chunk, dst_file)
+        # This is much faster than per-row iteration for large metadata tables
+        mapping_index = pd.MultiIndex.from_tuples(
+            list(data_src_to_dst.keys()),
+            names=["chunk_index", "file_index"],
+        )
+        mapping_values = list(data_src_to_dst.values())
+        mapping_df = pd.DataFrame(
+            mapping_values,
+            index=mapping_index,
+            columns=["dst_chunk", "dst_file"],
+        )
+
+        # Construct a MultiIndex for each row based on original data indices
+        row_index = pd.MultiIndex.from_arrays(
+            [df["_orig_data_chunk"], df["_orig_data_file"]],
+            names=["chunk_index", "file_index"],
+        )
+
+        # Align mapping to rows; missing keys fall back to the default destination
+        reindexed = mapping_df.reindex(row_index)
+        reindexed[["dst_chunk", "dst_file"]] = reindexed[["dst_chunk", "dst_file"]].fillna(
+            {"dst_chunk": data_idx["chunk"], "dst_file": data_idx["file"]}
+        )
+
+        # Assign mapped destination indices back to the DataFrame
+        df["data/chunk_index"] = reindexed["dst_chunk"].to_numpy()
+        df["data/file_index"] = reindexed["dst_file"].to_numpy()
+
+        # Clean up temporary columns
+        df = df.drop(columns=["_orig_data_chunk", "_orig_data_file"])
+    else:
+        # Fallback to simple offset (backward compatibility for single-file sources)
+        df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
+        df["data/file_index"] = df["data/file_index"] + data_idx["file"]
     for key, video_idx in videos_idx.items():
         # Store original video file indices before updating
         orig_chunk_col = f"videos/{key}/chunk_index"
@@ -146,8 +191,7 @@ def update_meta_data(
         if src_to_dst:
             # Map each episode to its correct destination file and apply offset
             for idx in df.index:
-                # Convert to Python int to avoid numpy type mismatch in dict lookup
-                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
+                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
 
                 # Get destination chunk/file for this source file
                 dst_chunk, dst_file = src_to_dst.get(src_key, (video_idx["chunk"], video_idx["file"]))
@@ -163,8 +207,7 @@ def update_meta_data(
             df[orig_chunk_col] = video_idx["chunk"]
             df[orig_file_col] = video_idx["file"]
             for idx in df.index:
-                # Convert to Python int to avoid numpy type mismatch in dict lookup
-                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
+                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
                 offset = src_to_offset.get(src_key, 0)
                 df.at[idx, f"videos/{key}/from_timestamp"] += offset
                 df.at[idx, f"videos/{key}/to_timestamp"] += offset
@@ -262,6 +305,10 @@ def aggregate_datasets(
 
         meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx)
 
+        # Clear the src_to_dst mapping after processing each source dataset
+        # to avoid interference between different source datasets
+        data_idx.pop("src_to_dst", None)
+
         dst_meta.info["total_episodes"] += src_meta.total_episodes
         dst_meta.info["total_frames"] += src_meta.total_frames
 
@@ -312,10 +359,6 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
         dst_file_durations = video_idx["dst_file_durations"]
 
         for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
-            # Convert to Python int to ensure consistent dict keys
-            src_chunk_idx = int(src_chunk_idx)
-            src_file_idx = int(src_file_idx)
-
             src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
                 video_key=key,
                 chunk_index=src_chunk_idx,
@@ -388,10 +431,16 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     Reads source data files, updates indices to match the aggregated dataset,
     and writes them to the destination with proper file rotation.
 
+    Tracks a `src_to_dst` mapping from source (chunk, file) to destination (chunk, file)
+    which is critical for correctly updating episode metadata when source datasets
+    have multiple data files (e.g., from a previous merge operation).
+
     Args:
         src_meta: Source dataset metadata.
         dst_meta: Destination dataset metadata.
         data_idx: Dictionary tracking data chunk and file indices.
+        data_files_size_in_mb: Maximum size for data files in MB.
+        chunk_size: Maximum number of files per chunk.
 
     Returns:
         dict: Updated data_idx with current chunk and file indices.
@@ -409,6 +458,10 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     # retrieve features schema for proper image typing in parquet
     hf_features = get_hf_features_from_features(dst_meta.features) if contains_images else None
 
+    # Track source to destination file mapping for metadata update
+    # This is critical for handling datasets that are already results of a merge
+    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}
+
     for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
         src_path = src_meta.root / DEFAULT_DATA_PATH.format(
             chunk_index=src_chunk_idx, file_index=src_file_idx
@@ -421,7 +474,9 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             df = pd.read_parquet(src_path)
         df = update_data_df(df, src_meta, dst_meta)
 
-        data_idx = append_or_create_parquet_file(
+        # Write data and get the actual destination file it was written to
+        # This avoids duplicating the rotation logic here
+        data_idx, (dst_chunk, dst_file) = append_or_create_parquet_file(
             df,
             src_path,
             data_idx,
@@ -433,6 +488,12 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             hf_features=hf_features,
         )
 
+        # Record the mapping from source to actual destination
+        src_to_dst[(src_chunk_idx, src_file_idx)] = (dst_chunk, dst_file)
+
+    # Add the mapping to data_idx for use in metadata update
+    data_idx["src_to_dst"] = src_to_dst
+
     return data_idx
 
 
@@ -473,7 +534,7 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
             videos_idx,
         )
 
-        meta_idx = append_or_create_parquet_file(
+        meta_idx, _ = append_or_create_parquet_file(
             df,
             src_path,
             meta_idx,
@@ -501,7 +562,7 @@ def append_or_create_parquet_file(
     contains_images: bool = False,
     aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
-):
+) -> tuple[dict[str, int], tuple[int, int]]:
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
     Manages file rotation when size limits are exceeded to prevent individual files
@@ -519,9 +580,11 @@ def append_or_create_parquet_file(
         hf_features: Optional HuggingFace Features schema for proper image typing.
 
     Returns:
-        dict: Updated index dictionary with current chunk and file indices.
+        tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
+               and (dst_chunk, dst_file) is the actual destination file the data was written to.
     """
-    dst_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
+    dst_chunk, dst_file = idx["chunk"], idx["file"]
+    dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
 
     if not dst_path.exists():
         dst_path.parent.mkdir(parents=True, exist_ok=True)
@@ -529,14 +592,15 @@ def append_or_create_parquet_file(
             to_parquet_with_hf_images(df, dst_path, features=hf_features)
         else:
             df.to_parquet(dst_path)
-        return idx
+        return idx, (dst_chunk, dst_file)
 
     src_size = get_parquet_file_size_in_mb(src_path)
     dst_size = get_parquet_file_size_in_mb(dst_path)
 
     if dst_size + src_size >= max_mb:
         idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
-        new_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
+        dst_chunk, dst_file = idx["chunk"], idx["file"]
+        new_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
         new_path.parent.mkdir(parents=True, exist_ok=True)
         final_df = df
         target_path = new_path
@@ -555,7 +619,7 @@ def append_or_create_parquet_file(
     else:
         final_df.to_parquet(target_path)
 
-    return idx
+    return idx, (dst_chunk, dst_file)
 
 
 def finalize_aggregation(aggr_meta, all_metadata):

src/lerobot/datasets/card_template.md

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

src/lerobot/datasets/dataset_tools.py

@@ -567,20 +567,22 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
     input_path: Path,
     output_path: Path,
-    episodes_to_keep: list[tuple[float, float]],
+    episodes_to_keep: list[tuple[int, int]],
     fps: float,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
-    This function decodes frames from specified time ranges and re-encodes them with
+    This function decodes frames from specified frame ranges and re-encodes them with
     properly reset timestamps to ensure monotonic progression.
 
     Args:
         input_path: Source video file path.
         output_path: Destination video file path.
-        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
+        episodes_to_keep: List of (start_frame, end_frame) tuples for episodes to keep.
+            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
+            is inclusive and end_frame is exclusive.
         fps: Frame rate of the video.
         vcodec: Video codec to use for encoding.
         pix_fmt: Pixel format for output video.
@@ -622,9 +624,10 @@ def _keep_episodes_from_video_with_av(
 
     # Create set of (start, end) ranges for fast lookup.
     # Convert to a sorted list for efficient checking.
-    time_ranges = sorted(episodes_to_keep)
+    frame_ranges = sorted(episodes_to_keep)
 
     # Track frame index for setting PTS and current range being processed.
+    src_frame_count = 0
     frame_count = 0
     range_idx = 0
 
@@ -634,21 +637,20 @@ def _keep_episodes_from_video_with_av(
             if frame is None:
                 continue
 
-            # Get frame timestamp.
-            frame_time = float(frame.pts * frame.time_base) if frame.pts is not None else 0.0
-
-            # Check if frame is in any of our desired time ranges.
+            # Check if frame is in any of our desired frame ranges.
             # Skip ranges that have already passed.
-            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
+            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
                 range_idx += 1
 
             # If we've passed all ranges, stop processing.
-            if range_idx >= len(time_ranges):
+            if range_idx >= len(frame_ranges):
                 break
 
             # Check if frame is in current range.
-            start_ts, end_ts = time_ranges[range_idx]
-            if frame_time < start_ts:
+            start_frame = frame_ranges[range_idx][0]
+
+            if src_frame_count < start_frame:
+                src_frame_count += 1
                 continue
 
             # Frame is in range - create a new frame with reset timestamps.
@@ -661,6 +663,7 @@ def _keep_episodes_from_video_with_av(
             for pkt in v_out.encode(new_frame):
                 out.mux(pkt)
 
+            src_frame_count += 1
             frame_count += 1
 
     # Flush encoder.
@@ -749,15 +752,17 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of time ranges to keep, in sorted order.
+                # Build list of frame ranges to keep, in sorted order.
                 sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
-                episodes_to_keep_ranges: list[tuple[float, float]] = []
-
+                episodes_to_keep_ranges: list[tuple[int, int]] = []
                 for old_idx in sorted_keep_episodes:
                     src_ep = src_dataset.meta.episodes[old_idx]
-                    from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
-                    to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
-                    episodes_to_keep_ranges.append((from_ts, to_ts))
+                    from_frame = round(src_ep[f"videos/{video_key}/from_timestamp"] * src_dataset.meta.fps)
+                    to_frame = round(src_ep[f"videos/{video_key}/to_timestamp"] * src_dataset.meta.fps)
+                    assert src_ep["length"] == to_frame - from_frame, (
+                        f"Episode length mismatch: {src_ep['length']} vs {to_frame - from_frame}"
+                    )
+                    episodes_to_keep_ranges.append((from_frame, to_frame))
 
                 # Use PyAV filters to efficiently re-encode only the desired segments.
                 assert src_dataset.meta.video_path is not None
@@ -1396,6 +1401,132 @@ BYTES_PER_KIB = 1024
 BYTES_PER_MIB = BYTES_PER_KIB * BYTES_PER_KIB
 
 
+def modify_tasks(
+    dataset: LeRobotDataset,
+    new_task: str | None = None,
+    episode_tasks: dict[int, str] | None = None,
+) -> LeRobotDataset:
+    """Modify tasks in a LeRobotDataset.
+
+    This function allows you to either:
+    1. Set a single task for the entire dataset (using `new_task`)
+    2. Set specific tasks for specific episodes (using `episode_tasks`)
+
+    You can combine both: `new_task` sets the default, and `episode_tasks` overrides
+    specific episodes.
+
+    The dataset is modified in-place, updating only the task-related files:
+    - meta/tasks.parquet
+    - data/**/*.parquet (task_index column)
+    - meta/episodes/**/*.parquet (tasks column)
+    - meta/info.json (total_tasks)
+
+    Args:
+        dataset: The source LeRobotDataset to modify.
+        new_task: A single task string to apply to all episodes. If None and episode_tasks
+            is also None, raises an error.
+        episode_tasks: Optional dict mapping episode indices to their task strings.
+            Overrides `new_task` for specific episodes.
+
+
+    Examples:
+        Set a single task for all episodes:
+            dataset = modify_tasks(dataset, new_task="Pick up the cube")
+
+        Set different tasks for specific episodes:
+            dataset = modify_tasks(
+                dataset,
+                episode_tasks={0: "Task A", 1: "Task B", 2: "Task A"}
+            )
+
+        Set a default task with overrides:
+            dataset = modify_tasks(
+                dataset,
+                new_task="Default task",
+                episode_tasks={5: "Special task for episode 5"}
+            )
+    """
+    if new_task is None and episode_tasks is None:
+        raise ValueError("Must specify at least one of new_task or episode_tasks")
+
+    if episode_tasks is not None:
+        valid_indices = set(range(dataset.meta.total_episodes))
+        invalid = set(episode_tasks.keys()) - valid_indices
+        if invalid:
+            raise ValueError(f"Invalid episode indices: {invalid}")
+
+    # Ensure episodes metadata is loaded
+    if dataset.meta.episodes is None:
+        dataset.meta.episodes = load_episodes(dataset.root)
+
+    # Build the mapping from episode index to task string
+    episode_to_task: dict[int, str] = {}
+    for ep_idx in range(dataset.meta.total_episodes):
+        if episode_tasks and ep_idx in episode_tasks:
+            episode_to_task[ep_idx] = episode_tasks[ep_idx]
+        elif new_task is not None:
+            episode_to_task[ep_idx] = new_task
+        else:
+            # Keep original task if not overridden and no default provided
+            original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
+            if not original_tasks:
+                raise ValueError(f"Episode {ep_idx} has no tasks and no default task was provided")
+            episode_to_task[ep_idx] = original_tasks[0]
+
+    # Collect all unique tasks and create new task mapping
+    unique_tasks = sorted(set(episode_to_task.values()))
+    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
+    task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
+
+    logging.info(f"Modifying tasks in {dataset.repo_id}")
+    logging.info(f"New tasks: {unique_tasks}")
+
+    root = dataset.root
+
+    # Update data files - modify task_index column
+    logging.info("Updating data files...")
+    data_dir = root / DATA_DIR
+
+    for parquet_path in tqdm(sorted(data_dir.rglob("*.parquet")), desc="Updating data"):
+        df = pd.read_parquet(parquet_path)
+
+        # Build a mapping from episode_index to new task_index for rows in this file
+        episode_indices_in_file = df["episode_index"].unique()
+        ep_to_new_task_idx = {
+            ep_idx: task_to_index[episode_to_task[ep_idx]] for ep_idx in episode_indices_in_file
+        }
+
+        # Update task_index column
+        df["task_index"] = df["episode_index"].map(ep_to_new_task_idx)
+        df.to_parquet(parquet_path, index=False)
+
+    # Update episodes metadata - modify tasks column
+    logging.info("Updating episodes metadata...")
+    episodes_dir = root / "meta" / "episodes"
+
+    for parquet_path in tqdm(sorted(episodes_dir.rglob("*.parquet")), desc="Updating episodes"):
+        df = pd.read_parquet(parquet_path)
+
+        # Update tasks column
+        df["tasks"] = df["episode_index"].apply(lambda ep_idx: [episode_to_task[ep_idx]])
+        df.to_parquet(parquet_path, index=False)
+
+    # Write new tasks.parquet
+    write_tasks(new_task_df, root)
+
+    # Update info.json
+    dataset.meta.info["total_tasks"] = len(unique_tasks)
+    write_info(dataset.meta.info, root)
+
+    # Reload metadata to reflect changes
+    dataset.meta.tasks = new_task_df
+    dataset.meta.episodes = load_episodes(root)
+
+    logging.info(f"Tasks: {unique_tasks}")
+
+    return dataset
+
+
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path,

src/lerobot/datasets/lerobot_dataset.py

@@ -57,6 +57,7 @@ from lerobot.datasets.utils import (
     load_info,
     load_nested_dataset,
     load_stats,
+    load_subtasks,
     load_tasks,
     update_chunk_file_indices,
     validate_episode_buffer,
@@ -67,6 +68,7 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
+    StreamingVideoEncoder,
     VideoFrame,
     concatenate_video_files,
     decode_video_frames,
@@ -74,11 +76,11 @@ from lerobot.datasets.video_utils import (
     get_safe_default_codec,
     get_video_duration_in_s,
     get_video_info,
+    resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -162,6 +164,7 @@ class LeRobotDatasetMetadata:
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
         self.tasks = load_tasks(self.root)
+        self.subtasks = load_subtasks(self.root)
         self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
 
@@ -518,6 +521,7 @@ class LeRobotDatasetMetadata:
         _validate_feature_names(features)
 
         obj.tasks = None
+        obj.subtasks = None
         obj.episodes = None
         obj.stats = None
         obj.info = create_empty_dataset_info(
@@ -542,12 +546,19 @@ class LeRobotDatasetMetadata:
 
 
 def _encode_video_worker(
-    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
+    video_key: str,
+    episode_index: int,
+    root: Path,
+    fps: int,
+    vcodec: str = "libsvtav1",
+    encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
+    encode_video_frames(
+        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
+    )
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -567,6 +578,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
+        streaming_encoding: bool = False,
+        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -653,7 +667,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             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
+                set the HF_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.
@@ -680,12 +694,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
             vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
-                encoding is CPU-heavy.
+                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
+                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
+            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
+                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
+            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
+                streaming encoding. Defaults to 30 (~1s at 30fps).
+            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
+                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
+                libsvtav1 and 'threads' for h264/hevc.
         """
         super().__init__()
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -697,7 +716,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = vcodec
+        self.vcodec = resolve_vcodec(vcodec)
+        self._encoder_threads = encoder_threads
 
         # Unused attributes
         self.image_writer = None
@@ -705,6 +725,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
+        self._streaming_encoder = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -726,7 +747,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Check if cached dataset contains all requested episodes
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
-        except (AssertionError, FileNotFoundError, NotADirectoryError):
+        except (FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
                 self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
@@ -746,6 +767,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
+        # Initialize streaming encoder for resumed recording
+        if streaming_encoding and len(self.meta.video_keys) > 0:
+            self._streaming_encoder = StreamingVideoEncoder(
+                fps=self.meta.fps,
+                vcodec=self.vcodec,
+                pix_fmt="yuv420p",
+                g=2,
+                crf=30,
+                preset=None,
+                queue_maxsize=encoder_queue_maxsize,
+                encoder_threads=encoder_threads,
+            )
+
     def _close_writer(self) -> None:
         """Close and cleanup the parquet writer if it exists."""
         writer = getattr(self, "writer", None)
@@ -805,7 +839,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             hub_api.upload_folder(**upload_kwargs)
 
         card = create_lerobot_dataset_card(
-            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, repo_id=self.repo_id, **card_kwargs
         )
         card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
 
@@ -1075,6 +1109,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Add task as a string
         task_idx = item["task_index"].item()
         item["task"] = self.meta.tasks.iloc[task_idx].name
+
+        # add subtask information if available
+        if "subtask_index" in self.features and self.meta.subtasks is not None:
+            subtask_idx = item["subtask_index"].item()
+            item["subtask"] = self.meta.subtasks.iloc[subtask_idx].name
+
         return item
 
     def __repr__(self):
@@ -1095,6 +1135,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         self._close_writer()
         self.meta._close_writer()
+        if self._streaming_encoder is not None:
+            self._streaming_encoder.close()
 
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
@@ -1149,6 +1191,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing
 
+        # Start streaming encoder on first frame of episode (once, before iterating keys)
+        if frame_index == 0 and self._streaming_encoder is not None:
+            self._streaming_encoder.start_episode(
+                video_keys=list(self.meta.video_keys),
+                temp_dir=self.root,
+            )
+
         # Add frame features to episode_buffer
         for key in frame:
             if key not in self.features:
@@ -1156,7 +1205,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                 )
 
-            if self.features[key]["dtype"] in ["image", "video"]:
+            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
+                self._streaming_encoder.feed_frame(key, frame[key])
+                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
+            elif self.features[key]["dtype"] in ["image", "video"]:
                 img_path = self._get_image_file_path(
                     episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                 )
@@ -1217,13 +1269,38 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-        ep_stats = compute_episode_stats(episode_buffer, self.features)
 
-        ep_metadata = self._save_episode_data(episode_buffer)
         has_video_keys = len(self.meta.video_keys) > 0
+        use_streaming = self._streaming_encoder is not None and has_video_keys
         use_batched_encoding = self.batch_encoding_size > 1
 
-        if has_video_keys and not use_batched_encoding:
+        if use_streaming:
+            # Compute stats for non-video features only (video stats come from encoder)
+            non_video_buffer = {
+                k: v
+                for k, v in episode_buffer.items()
+                if self.features.get(k, {}).get("dtype") not in ("video",)
+            }
+            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
+            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
+        else:
+            ep_stats = compute_episode_stats(episode_buffer, self.features)
+
+        ep_metadata = self._save_episode_data(episode_buffer)
+
+        if use_streaming:
+            # Finish streaming encoding and collect results
+            streaming_results = self._streaming_encoder.finish_episode()
+            for video_key in self.meta.video_keys:
+                temp_path, video_stats = streaming_results[video_key]
+                if video_stats is not None:
+                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
+                    ep_stats[video_key] = {
+                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
+                        for k, v in video_stats.items()
+                    }
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
+        elif has_video_keys and not use_batched_encoding:
             num_cameras = len(self.meta.video_keys)
             if parallel_encoding and num_cameras > 1:
                 # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1237,6 +1314,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             self.root,
                             self.fps,
                             self.vcodec,
+                            self._encoder_threads,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1505,6 +1583,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
+        # Cancel streaming encoder if active
+        if self._streaming_encoder is not None:
+            self._streaming_encoder.cancel_episode()
+
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1552,7 +1634,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
+        return _encode_video_worker(
+            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
+        )
 
     @classmethod
     def create(
@@ -1569,10 +1653,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
+        metadata_buffer_size: int = 10,
+        streaming_encoding: bool = False,
+        encoder_queue_maxsize: int = 30,
+        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+        vcodec = resolve_vcodec(vcodec)
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1581,6 +1668,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             features=features,
             root=root,
             use_videos=use_videos,
+            metadata_buffer_size=metadata_buffer_size,
         )
         obj.repo_id = obj.meta.repo_id
         obj.root = obj.meta.root
@@ -1590,6 +1678,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
         obj.vcodec = vcodec
+        obj._encoder_threads = encoder_threads
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1611,6 +1700,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj._lazy_loading = False
         obj._recorded_frames = 0
         obj._writer_closed_for_reading = False
+
+        # Initialize streaming encoder
+        if streaming_encoding and len(obj.meta.video_keys) > 0:
+            obj._streaming_encoder = StreamingVideoEncoder(
+                fps=fps,
+                vcodec=vcodec,
+                pix_fmt="yuv420p",
+                g=2,
+                crf=30,
+                preset=None,
+                queue_maxsize=encoder_queue_maxsize,
+                encoder_threads=encoder_threads,
+            )
+        else:
+            obj._streaming_encoder = None
+
         return obj
 
 
@@ -1666,11 +1771,12 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
             )
         for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
             extra_keys = set(ds.features).difference(intersection_features)
-            logging.warning(
-                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                "other datasets."
-            )
-            self.disabled_features.update(extra_keys)
+            if extra_keys:
+                logging.warning(
+                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                    "other datasets."
+                )
+                self.disabled_features.update(extra_keys)
 
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps

src/lerobot/datasets/transforms.py

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

src/lerobot/datasets/utils.py

@@ -60,6 +60,7 @@ VIDEO_DIR = "videos"
 
 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
+DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
@@ -121,19 +122,9 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
-        # PyArrow loads the entire dataset into memory
-        if episodes is None:
-            return Dataset.from_parquet([str(path) for path in paths], features=features)
-
-        arrow_dataset = pa_ds.dataset(paths, format="parquet")
-        filter_expr = pa_ds.field("episode_index").isin(episodes)
-        table = arrow_dataset.to_table(filter=filter_expr)
-
-        if features is not None:
-            table = table.cast(features.arrow_schema)
-
-        return Dataset(table)
+        # We use .from_parquet() memory-mapped loading for efficiency
+        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -353,6 +344,14 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
     return tasks
 
 
+def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
+    """Load subtasks from subtasks.parquet if it exists."""
+    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
+    if subtasks_path.exists():
+        return pd.read_parquet(subtasks_path)
+    return None
+
+
 def write_episodes(episodes: Dataset, local_dir: Path) -> None:
     """Write episode metadata to a parquet file in the LeRobot v3.0 format.
     This function writes episode-level metadata to a single parquet file.

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -529,7 +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`, `cadene/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
     )
     parser.add_argument(
         "--branch",

src/lerobot/datasets/video_utils.py

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

src/lerobot/envs/configs.py

@@ -205,6 +205,7 @@ class ObservationConfig:
 
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
+    add_ee_pose_to_observation: bool = False
     display_cameras: bool = False
 
 
@@ -260,6 +261,7 @@ class HILSerlRobotEnvConfig(EnvConfig):
 @dataclass
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
+    task_ids: list[int] | None = None
     fps: int = 30
     episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
@@ -338,10 +340,10 @@ class LiberoEnv(EnvConfig):
 
     @property
     def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-        }
+        kwargs: dict[str, Any] = {"obs_type": self.obs_type, "render_mode": self.render_mode}
+        if self.task_ids is not None:
+            kwargs["task_ids"] = self.task_ids
+        return kwargs
 
 
 @EnvConfig.register_subclass("metaworld")

src/lerobot/envs/libero.py

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

src/lerobot/motors/__init__.py

@@ -14,4 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus
+from .motors_bus import (
+    Motor,
+    MotorCalibration,
+    MotorNormMode,
+)

src/lerobot/motors/calibration_gui.py

@@ -18,7 +18,7 @@ from dataclasses import dataclass
 
 os.environ["PYGAME_HIDE_SUPPORT_PROMPT"] = "1"
 
-from lerobot.motors import MotorCalibration, MotorsBus
+from .motors_bus import MotorCalibration, MotorsBus
 
 BAR_LEN, BAR_THICKNESS = 450, 8
 HANDLE_R = 10
@@ -221,7 +221,7 @@ class RangeFinderGUI:
 
         self.bus = bus
         self.groups = groups if groups is not None else {"all": list(bus.motors)}
-        self.group_names = list(groups)
+        self.group_names = list(self.groups)
         self.current_group = self.group_names[0]
 
         if not bus.is_connected:
@@ -230,18 +230,20 @@ class RangeFinderGUI:
         self.calibration = bus.read_calibration()
         self.res_table = bus.model_resolution_table
         self.present_cache = {
-            m: bus.read("Present_Position", m, normalize=False) for motors in groups.values() for m in motors
+            m: bus.read("Present_Position", m, normalize=False)
+            for motors in self.groups.values()
+            for m in motors
         }
 
         pygame.init()
         self.font = pygame.font.Font(None, FONT_SIZE)
 
-        label_pad = max(self.font.size(m)[0] for ms in groups.values() for m in ms)
+        label_pad = max(self.font.size(m)[0] for ms in self.groups.values() for m in ms)
         self.label_pad = label_pad
         width = 40 + label_pad + BAR_LEN + 6 + BTN_W + 10 + SAVE_W + 10
         self.controls_bottom = 10 + SAVE_H
         self.base_y = self.controls_bottom + TOP_GAP
-        height = self.base_y + PADDING_Y * len(groups[self.current_group]) + 40
+        height = self.base_y + PADDING_Y * len(self.groups[self.current_group]) + 40
 
         self.screen = pygame.display.set_mode((width, height))
         pygame.display.set_caption("Motors range finder")

src/lerobot/motors/damiao/__init__.py

@@ -0,0 +1,18 @@
+#!/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.
+
+from .damiao import DamiaoMotorsBus
+from .tables import *

src/lerobot/motors/damiao/damiao.py

@@ -0,0 +1,859 @@
+# 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.
+
+# Portions of this file are derived from DM_Control_Python by cmjang.
+# Licensed under the MIT License; see `LICENSE` for the full text:
+# https://github.com/cmjang/DM_Control_Python
+
+import logging
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from functools import cached_property
+from typing import TYPE_CHECKING, Any, TypedDict
+
+from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.import_utils import _can_available
+
+if TYPE_CHECKING or _can_available:
+    import can
+else:
+
+    class can:  # noqa: N801
+        Message = object
+        interface = None
+
+
+import numpy as np
+
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import enter_pressed, move_cursor_up
+
+from ..motors_bus import Motor, MotorCalibration, MotorsBusBase, NameOrID, Value
+from .tables import (
+    AVAILABLE_BAUDRATES,
+    CAN_CMD_DISABLE,
+    CAN_CMD_ENABLE,
+    CAN_CMD_REFRESH,
+    CAN_CMD_SET_ZERO,
+    CAN_PARAM_ID,
+    DEFAULT_BAUDRATE,
+    DEFAULT_TIMEOUT_MS,
+    MIT_KD_RANGE,
+    MIT_KP_RANGE,
+    MOTOR_LIMIT_PARAMS,
+    MotorType,
+)
+
+logger = logging.getLogger(__name__)
+
+
+LONG_TIMEOUT_SEC = 0.1
+MEDIUM_TIMEOUT_SEC = 0.01
+SHORT_TIMEOUT_SEC = 0.001
+PRECISE_TIMEOUT_SEC = 0.0001
+
+
+class MotorState(TypedDict):
+    position: float
+    velocity: float
+    torque: float
+    temp_mos: float
+    temp_rotor: float
+
+
+class DamiaoMotorsBus(MotorsBusBase):
+    """
+    The Damiao implementation for a MotorsBus using CAN bus communication.
+
+    This class uses python-can for CAN bus communication with Damiao motors.
+    For more info, see:
+    - python-can documentation: https://python-can.readthedocs.io/en/stable/
+    - Seedstudio documentation: https://wiki.seeedstudio.com/damiao_series/
+    - DM_Control_Python repo: https://github.com/cmjang/DM_Control_Python
+    """
+
+    # CAN-specific settings
+    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
+    default_baudrate = DEFAULT_BAUDRATE
+    default_timeout = DEFAULT_TIMEOUT_MS
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+        can_interface: str = "auto",
+        use_can_fd: bool = True,
+        bitrate: int = 1000000,
+        data_bitrate: int | None = 5000000,
+    ):
+        """
+        Initialize the Damiao motors bus.
+
+        Args:
+            port: CAN interface name (e.g., "can0" for Linux, "/dev/cu.usbmodem*" for macOS)
+            motors: Dictionary mapping motor names to Motor objects
+            calibration: Optional calibration data
+            can_interface: CAN interface type - "auto" (default), "socketcan" (Linux), or "slcan" (macOS/serial)
+            use_can_fd: Whether to use CAN FD mode (default: True for OpenArms)
+            bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
+            data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
+        """
+        super().__init__(port, motors, calibration)
+        self.port = port
+        self.can_interface = can_interface
+        self.use_can_fd = use_can_fd
+        self.bitrate = bitrate
+        self.data_bitrate = data_bitrate
+        self.canbus: can.interface.Bus | None = None
+        self._is_connected = False
+
+        # Map motor names to CAN IDs
+        self._motor_can_ids: dict[str, int] = {}
+        self._recv_id_to_motor: dict[int, str] = {}
+        self._motor_types: dict[str, MotorType] = {}
+
+        for name, motor in self.motors.items():
+            if motor.motor_type_str is None:
+                raise ValueError(f"Motor '{name}' is missing required 'motor_type'")
+            self._motor_types[name] = getattr(MotorType, motor.motor_type_str.upper().replace("-", "_"))
+
+            # Map recv_id to motor name for filtering responses
+            if motor.recv_id is not None:
+                self._recv_id_to_motor[motor.recv_id] = name
+
+        # State cache for handling packet drops safely
+        self._last_known_states: dict[str, MotorState] = {
+            name: {
+                "position": 0.0,
+                "velocity": 0.0,
+                "torque": 0.0,
+                "temp_mos": 0.0,
+                "temp_rotor": 0.0,
+            }
+            for name in self.motors
+        }
+
+        # Dynamic gains storage
+        # Defaults: Kp=10.0 (Stiffness), Kd=0.5 (Damping)
+        self._gains: dict[str, dict[str, float]] = {name: {"kp": 10.0, "kd": 0.5} for name in self.motors}
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if the CAN bus is connected."""
+        return self._is_connected and self.canbus is not None
+
+    @check_if_already_connected
+    def connect(self, handshake: bool = True) -> None:
+        """
+        Open the CAN bus and initialize communication.
+
+        Args:
+            handshake: If True, ping all motors to verify they're present
+        """
+
+        try:
+            # Auto-detect interface type based on port name
+            if self.can_interface == "auto":
+                if self.port.startswith("/dev/"):
+                    self.can_interface = "slcan"
+                    logger.info(f"Auto-detected slcan interface for port {self.port}")
+                else:
+                    self.can_interface = "socketcan"
+                    logger.info(f"Auto-detected socketcan interface for port {self.port}")
+
+            # Connect to CAN bus
+            kwargs = {
+                "channel": self.port,
+                "bitrate": self.bitrate,
+                "interface": self.can_interface,
+            }
+
+            if self.can_interface == "socketcan" and self.use_can_fd and self.data_bitrate is not None:
+                kwargs.update({"data_bitrate": self.data_bitrate, "fd": True})
+                logger.info(
+                    f"Connected to {self.port} with CAN FD (bitrate={self.bitrate}, data_bitrate={self.data_bitrate})"
+                )
+            else:
+                logger.info(f"Connected to {self.port} with {self.can_interface} (bitrate={self.bitrate})")
+
+            self.canbus = can.interface.Bus(**kwargs)
+            self._is_connected = True
+
+            if handshake:
+                self._handshake()
+
+            logger.debug(f"{self.__class__.__name__} connected via {self.can_interface}.")
+        except Exception as e:
+            self._is_connected = False
+            raise ConnectionError(f"Failed to connect to CAN bus: {e}") from e
+
+    def _handshake(self) -> None:
+        """
+        Verify all motors are present and populate initial state cache.
+        Raises ConnectionError if any motor fails to respond.
+        """
+        logger.info("Starting handshake with motors...")
+
+        # Drain any pending messages
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        while self.canbus.recv(timeout=0.01):
+            pass
+
+        missing_motors = []
+        for motor_name in self.motors:
+            motor_id = self._get_motor_id(motor_name)
+            recv_id = self._get_motor_recv_id(motor_name)
+
+            # Send enable command
+            data = [0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, CAN_CMD_ENABLE]
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+            self.canbus.send(msg)
+
+            # Wait for response with longer timeout
+            response = None
+            start_time = time.time()
+            while time.time() - start_time < 0.1:
+                response = self.canbus.recv(timeout=0.1)
+                if response and response.arbitration_id == recv_id:
+                    break
+                response = None
+
+            if response is None:
+                missing_motors.append(motor_name)
+            else:
+                self._process_response(motor_name, msg)
+            time.sleep(MEDIUM_TIMEOUT_SEC)
+
+        if missing_motors:
+            raise ConnectionError(
+                f"Handshake failed. The following motors did not respond: {missing_motors}. "
+                "Check power (24V) and CAN wiring."
+            )
+        logger.info("Handshake successful. All motors ready.")
+
+    @check_if_not_connected
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """
+        Close the CAN bus connection.
+
+        Args:
+            disable_torque: If True, disable torque on all motors before disconnecting
+        """
+
+        if disable_torque:
+            try:
+                self.disable_torque()
+            except Exception as e:
+                logger.warning(f"Failed to disable torque during disconnect: {e}")
+
+        if self.canbus:
+            self.canbus.shutdown()
+            self.canbus = None
+        self._is_connected = False
+        logger.debug(f"{self.__class__.__name__} disconnected.")
+
+    def configure_motors(self) -> None:
+        """Configure all motors with default settings."""
+        # Damiao motors don't require much configuration in MIT mode
+        # Just ensure they're enabled
+        for motor in self.motors:
+            self._send_simple_command(motor, CAN_CMD_ENABLE)
+            time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    def _send_simple_command(self, motor: NameOrID, command_byte: int) -> None:
+        """Helper to send simple 8-byte commands (Enable, Disable, Zero)."""
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [command_byte]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        self.canbus.send(msg)
+        if msg := self._recv_motor_response(expected_recv_id=recv_id):
+            self._process_response(motor_name, msg)
+        else:
+            logger.debug(f"No response from {motor_name} after command 0x{command_byte:02X}")
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        target_motors = self._get_motors_list(motors)
+        for motor in target_motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._send_simple_command(motor, CAN_CMD_ENABLE)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        target_motors = self._get_motors_list(motors)
+        for motor in target_motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._send_simple_command(motor, CAN_CMD_DISABLE)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    @contextmanager
+    def torque_disabled(self, motors: str | list[str] | None = None):
+        """
+        Context manager that guarantees torque is re-enabled.
+
+        This helper is useful to temporarily disable torque when configuring motors.
+        """
+        self.disable_torque(motors)
+        try:
+            yield
+        finally:
+            self.enable_torque(motors)
+
+    def set_zero_position(self, motors: str | list[str] | None = None) -> None:
+        """Set current position as zero for selected motors."""
+        target_motors = self._get_motors_list(motors)
+        for motor in target_motors:
+            self._send_simple_command(motor, CAN_CMD_SET_ZERO)
+            time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    def _refresh_motor(self, motor: NameOrID) -> can.Message | None:
+        """Refresh motor status and return the response."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+        msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        self.canbus.send(msg)
+        return self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _recv_motor_response(
+        self, expected_recv_id: int | None = None, timeout: float = 0.001
+    ) -> can.Message | None:
+        """
+        Receive a response from a motor.
+
+        Args:
+            expected_recv_id: If provided, only return messages from this CAN ID
+            timeout: Timeout in seconds (default: 1ms for high-speed operation)
+        Returns:
+            CAN message if received, None otherwise
+        """
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        try:
+            start_time = time.time()
+            messages_seen = []
+            while time.time() - start_time < timeout:
+                msg = self.canbus.recv(timeout=PRECISE_TIMEOUT_SEC)
+                if msg:
+                    messages_seen.append(f"0x{msg.arbitration_id:02X}")
+                    if expected_recv_id is None or msg.arbitration_id == expected_recv_id:
+                        return msg
+                    logger.debug(
+                        f"Ignoring message from 0x{msg.arbitration_id:02X}, expected 0x{expected_recv_id:02X}"
+                    )
+
+            if logger.isEnabledFor(logging.DEBUG):
+                if messages_seen:
+                    logger.debug(
+                        f"Received {len(messages_seen)} msgs from {set(messages_seen)}, expected 0x{expected_recv_id:02X}"
+                    )
+                else:
+                    logger.debug(f"No CAN messages received (expected 0x{expected_recv_id:02X})")
+        except Exception as e:
+            logger.debug(f"Failed to receive CAN message: {e}")
+        return None
+
+    def _recv_all_responses(
+        self, expected_recv_ids: list[int], timeout: float = 0.002
+    ) -> dict[int, can.Message]:
+        """
+        Efficiently receive responses from multiple motors at once.
+        Uses the OpenArms pattern: collect all available messages within timeout.
+
+        Args:
+            expected_recv_ids: List of CAN IDs we expect responses from
+            timeout: Total timeout in seconds (default: 2ms)
+
+        Returns:
+            Dictionary mapping recv_id to CAN message
+        """
+        responses: dict[int, can.Message] = {}
+        expected_set = set(expected_recv_ids)
+        start_time = time.time()
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        try:
+            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
+                # 100us poll timeout
+                msg = self.canbus.recv(timeout=PRECISE_TIMEOUT_SEC)
+                if msg and msg.arbitration_id in expected_set:
+                    responses[msg.arbitration_id] = msg
+                    if len(responses) == len(expected_recv_ids):
+                        break
+        except Exception as e:
+            logger.debug(f"Error receiving responses: {e}")
+
+        return responses
+
+    def _encode_mit_packet(
+        self,
+        motor_type: MotorType,
+        kp: float,
+        kd: float,
+        position_degrees: float,
+        velocity_deg_per_sec: float,
+        torque: float,
+    ) -> list[int]:
+        """Helper to encode control parameters into 8 bytes for MIT mode."""
+        # Convert degrees to radians
+        position_rad = np.radians(position_degrees)
+        velocity_rad_per_sec = np.radians(velocity_deg_per_sec)
+
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+
+        # Encode parameters
+        kp_uint = self._float_to_uint(kp, *MIT_KP_RANGE, 12)
+        kd_uint = self._float_to_uint(kd, *MIT_KD_RANGE, 12)
+        q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+        dq_uint = self._float_to_uint(velocity_rad_per_sec, -vmax, vmax, 12)
+        tau_uint = self._float_to_uint(torque, -tmax, tmax, 12)
+
+        # Pack data
+        data = [0] * 8
+        data[0] = (q_uint >> 8) & 0xFF
+        data[1] = q_uint & 0xFF
+        data[2] = dq_uint >> 4
+        data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+        data[4] = kp_uint & 0xFF
+        data[5] = kd_uint >> 4
+        data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+        data[7] = tau_uint & 0xFF
+        return data
+
+    def _mit_control(
+        self,
+        motor: NameOrID,
+        kp: float,
+        kd: float,
+        position_degrees: float,
+        velocity_deg_per_sec: float,
+        torque: float,
+    ) -> None:
+        """Send MIT control command to a motor."""
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        motor_type = self._motor_types[motor_name]
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+        self.canbus.send(msg)
+
+        recv_id = self._get_motor_recv_id(motor)
+        if msg := self._recv_motor_response(expected_recv_id=recv_id):
+            self._process_response(motor_name, msg)
+        else:
+            logger.debug(f"No response from {motor_name} after MIT control command")
+
+    def _mit_control_batch(
+        self,
+        commands: dict[NameOrID, tuple[float, float, float, float, float]],
+    ) -> None:
+        """
+        Send MIT control commands to multiple motors in batch.
+        Sends all commands first, then collects responses.
+
+        Args:
+            commands: Dict mapping motor name/ID to (kp, kd, position_deg, velocity_deg/s, torque)
+                     Example: {'joint_1': (10.0, 0.5, 45.0, 0.0, 0.0), ...}
+        """
+        if not commands:
+            return
+
+        recv_id_to_motor: dict[int, str] = {}
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        # Step 1: Send all MIT control commands
+        for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
+            motor_id = self._get_motor_id(motor)
+            motor_name = self._get_motor_name(motor)
+            motor_type = self._motor_types[motor_name]
+
+            data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+            self.canbus.send(msg)
+
+            recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
+
+        # Step 2: Collect responses and update state cache
+        responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=SHORT_TIMEOUT_SEC)
+        for recv_id, motor_name in recv_id_to_motor.items():
+            if msg := responses.get(recv_id):
+                self._process_response(motor_name, msg)
+
+    def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
+        """Convert float to unsigned integer for CAN transmission."""
+        x = max(x_min, min(x_max, x))  # Clamp to range
+        span = x_max - x_min
+        data_norm = (x - x_min) / span
+        return int(data_norm * ((1 << bits) - 1))
+
+    def _uint_to_float(self, x: int, x_min: float, x_max: float, bits: int) -> float:
+        """Convert unsigned integer from CAN to float."""
+        span = x_max - x_min
+        data_norm = float(x) / ((1 << bits) - 1)
+        return data_norm * span + x_min
+
+    def _decode_motor_state(
+        self, data: bytearray | bytes, motor_type: MotorType
+    ) -> tuple[float, float, float, int, int]:
+        """
+        Decode motor state from CAN data.
+        Returns: (position_deg, velocity_deg_s, torque, temp_mos, temp_rotor)
+        """
+        if len(data) < 8:
+            raise ValueError("Invalid motor state data")
+
+        # Extract encoded values
+        q_uint = (data[1] << 8) | data[2]
+        dq_uint = (data[3] << 4) | (data[4] >> 4)
+        tau_uint = ((data[4] & 0x0F) << 8) | data[5]
+        t_mos = data[6]
+        t_rotor = data[7]
+
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+
+        # Decode to physical values
+        position_rad = self._uint_to_float(q_uint, -pmax, pmax, 16)
+        velocity_rad_per_sec = self._uint_to_float(dq_uint, -vmax, vmax, 12)
+        torque = self._uint_to_float(tau_uint, -tmax, tmax, 12)
+
+        return np.degrees(position_rad), np.degrees(velocity_rad_per_sec), torque, t_mos, t_rotor
+
+    def _process_response(self, motor: str, msg: can.Message) -> None:
+        """Decode a message and update the motor state cache."""
+        try:
+            motor_type = self._motor_types[motor]
+            pos, vel, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+
+            self._last_known_states[motor] = {
+                "position": pos,
+                "velocity": vel,
+                "torque": torque,
+                "temp_mos": float(t_mos),
+                "temp_rotor": float(t_rotor),
+            }
+        except Exception as e:
+            logger.warning(f"Failed to decode response from {motor}: {e}")
+
+    @check_if_not_connected
+    def read(self, data_name: str, motor: str) -> Value:
+        """Read a value from a single motor. Positions are always in degrees."""
+
+        # Refresh motor to get latest state
+        msg = self._refresh_motor(motor)
+        if msg is None:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            raise ConnectionError(
+                f"No response from motor '{motor}' (send ID: 0x{motor_id:02X}, recv ID: 0x{recv_id:02X}). "
+                f"Check that: 1) Motor is powered (24V), 2) CAN wiring is correct, "
+                f"3) Motor IDs are configured correctly using Damiao Debugging Tools"
+            )
+
+        self._process_response(motor, msg)
+        return self._get_cached_value(motor, data_name)
+
+    def _get_cached_value(self, motor: str, data_name: str) -> Value:
+        """Retrieve a specific value from the cache."""
+        state = self._last_known_states[motor]
+        mapping: dict[str, Any] = {
+            "Present_Position": state["position"],
+            "Present_Velocity": state["velocity"],
+            "Present_Torque": state["torque"],
+            "Temperature_MOS": state["temp_mos"],
+            "Temperature_Rotor": state["temp_rotor"],
+        }
+        if data_name not in mapping:
+            raise ValueError(f"Unknown data_name: {data_name}")
+        return mapping[data_name]
+
+    @check_if_not_connected
+    def write(
+        self,
+        data_name: str,
+        motor: str,
+        value: Value,
+    ) -> None:
+        """
+        Write a value to a single motor. Positions are always in degrees.
+        Can write 'Goal_Position', 'Kp', or 'Kd'.
+        """
+
+        if data_name in ("Kp", "Kd"):
+            self._gains[motor][data_name.lower()] = float(value)
+        elif data_name == "Goal_Position":
+            kp = self._gains[motor]["kp"]
+            kd = self._gains[motor]["kd"]
+            self._mit_control(motor, kp, kd, float(value), 0.0, 0.0)
+        else:
+            raise ValueError(f"Writing {data_name} not supported in MIT mode")
+
+    def sync_read(
+        self,
+        data_name: str,
+        motors: str | list[str] | None = None,
+    ) -> dict[str, Value]:
+        """
+        Read the same value from multiple motors simultaneously.
+        """
+        target_motors = self._get_motors_list(motors)
+        self._batch_refresh(target_motors)
+
+        result = {}
+        for motor in target_motors:
+            result[motor] = self._get_cached_value(motor, data_name)
+        return result
+
+    def sync_read_all_states(
+        self,
+        motors: str | list[str] | None = None,
+        *,
+        num_retry: int = 0,
+    ) -> dict[str, MotorState]:
+        """
+        Read ALL motor states (position, velocity, torque) from multiple motors in ONE refresh cycle.
+
+        Returns:
+            Dictionary mapping motor names to state dicts with keys: 'position', 'velocity', 'torque'
+            Example: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
+        """
+        target_motors = self._get_motors_list(motors)
+        self._batch_refresh(target_motors)
+
+        result = {}
+        for motor in target_motors:
+            result[motor] = self._last_known_states[motor].copy()
+        return result
+
+    def _batch_refresh(self, motors: list[str]) -> None:
+        """Internal helper to refresh a list of motors and update cache."""
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
+        # Send refresh commands
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+            msg = can.Message(
+                arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd
+            )
+            self.canbus.send(msg)
+
+        # Collect responses
+        expected_recv_ids = [self._get_motor_recv_id(m) for m in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=MEDIUM_TIMEOUT_SEC)
+
+        # Update cache
+        for motor in motors:
+            recv_id = self._get_motor_recv_id(motor)
+            msg = responses.get(recv_id)
+            if msg:
+                self._process_response(motor, msg)
+            else:
+                logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
+
+    @check_if_not_connected
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
+        """
+        Write values to multiple motors simultaneously. Positions are always in degrees.
+        """
+
+        if data_name in ("Kp", "Kd"):
+            key = data_name.lower()
+            for motor, val in values.items():
+                self._gains[motor][key] = float(val)
+
+        elif data_name == "Goal_Position":
+            # Step 1: Send all MIT control commands
+            recv_id_to_motor: dict[int, str] = {}
+            if self.canbus is None:
+                raise RuntimeError("CAN bus is not initialized.")
+            for motor, value_degrees in values.items():
+                motor_id = self._get_motor_id(motor)
+                motor_name = self._get_motor_name(motor)
+                motor_type = self._motor_types[motor_name]
+
+                kp = self._gains[motor]["kp"]
+                kd = self._gains[motor]["kd"]
+
+                data = self._encode_mit_packet(motor_type, kp, kd, float(value_degrees), 0.0, 0.0)
+                msg = can.Message(
+                    arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd
+                )
+                self.canbus.send(msg)
+                precise_sleep(PRECISE_TIMEOUT_SEC)
+
+                recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
+
+            # Step 2: Collect responses and update state cache
+            responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=MEDIUM_TIMEOUT_SEC)
+            for recv_id, motor_name in recv_id_to_motor.items():
+                if msg := responses.get(recv_id):
+                    self._process_response(motor_name, msg)
+        else:
+            # Fall back to individual writes
+            for motor, value in values.items():
+                self.write(data_name, motor, value)
+
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration data from motors."""
+        # Damiao motors don't store calibration internally
+        # Return existing calibration or empty dict
+        return self.calibration if self.calibration else {}
+
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration data to motors."""
+        # Damiao motors don't store calibration internally
+        # Just cache it in memory
+        if cache:
+            self.calibration = calibration_dict
+
+    def record_ranges_of_motion(
+        self,
+        motors: str | list[str] | None = None,
+        display_values: bool = True,
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
+        """
+        Interactively record the min/max values of each motor in degrees.
+
+        Move the joints by hand (with torque disabled) while the method streams live positions.
+        Press Enter to finish.
+        """
+        target_motors = self._get_motors_list(motors)
+
+        self.disable_torque(target_motors)
+        time.sleep(LONG_TIMEOUT_SEC)
+
+        start_positions = self.sync_read("Present_Position", target_motors)
+        mins = start_positions.copy()
+        maxes = start_positions.copy()
+
+        print("\nMove joints through their full range of motion. Press ENTER when done.")
+        user_pressed_enter = False
+
+        while not user_pressed_enter:
+            positions = self.sync_read("Present_Position", target_motors)
+
+            for motor in target_motors:
+                if motor in positions:
+                    mins[motor] = min(positions[motor], mins.get(motor, positions[motor]))
+                    maxes[motor] = max(positions[motor], maxes.get(motor, positions[motor]))
+
+            if display_values:
+                print("\n" + "=" * 50)
+                print(f"{'MOTOR':<20} | {'MIN (deg)':>12} | {'POS (deg)':>12} | {'MAX (deg)':>12}")
+                print("-" * 50)
+                for motor in target_motors:
+                    if motor in positions:
+                        print(
+                            f"{motor:<20} | {mins[motor]:>12.1f} | {positions[motor]:>12.1f} | {maxes[motor]:>12.1f}"
+                        )
+
+            if enter_pressed():
+                user_pressed_enter = True
+
+            if display_values and not user_pressed_enter:
+                move_cursor_up(len(target_motors) + 4)
+
+            time.sleep(LONG_TIMEOUT_SEC)
+
+        self.enable_torque(target_motors)
+
+        for motor in target_motors:
+            if (motor in mins) and (motor in maxes) and (int(abs(maxes[motor] - mins[motor])) < 5):
+                raise ValueError(f"Motor {motor} has insufficient range of motion (< 5 degrees)")
+
+        return mins, maxes
+
+    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+        """Convert motor specification to list of motor names."""
+        if motors is None:
+            return list(self.motors.keys())
+        elif isinstance(motors, str):
+            return [motors]
+        elif isinstance(motors, list):
+            return motors
+        else:
+            raise TypeError(f"Invalid motors type: {type(motors)}")
+
+    def _get_motor_id(self, motor: NameOrID) -> int:
+        """Get CAN ID for a motor."""
+        if isinstance(motor, str):
+            if motor in self.motors:
+                return self.motors[motor].id
+            else:
+                raise ValueError(f"Unknown motor: {motor}")
+        else:
+            return motor
+
+    def _get_motor_name(self, motor: NameOrID) -> str:
+        """Get motor name from name or ID."""
+        if isinstance(motor, str):
+            return motor
+        else:
+            for name, m in self.motors.items():
+                if m.id == motor:
+                    return name
+            raise ValueError(f"Unknown motor ID: {motor}")
+
+    def _get_motor_recv_id(self, motor: NameOrID) -> int:
+        """Get motor recv_id from name or ID."""
+        motor_name = self._get_motor_name(motor)
+        motor_obj = self.motors.get(motor_name)
+        if motor_obj and motor_obj.recv_id is not None:
+            return motor_obj.recv_id
+        else:
+            raise ValueError(f"Motor {motor_obj} doesn't have a valid recv_id (None).")
+
+    @cached_property
+    def is_calibrated(self) -> bool:
+        """Check if motors are calibrated."""
+        return bool(self.calibration)

src/lerobot/motors/damiao/tables.py

@@ -0,0 +1,209 @@
+# 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):
+    DM3507 = 0
+    DM4310 = 1
+    DM4310_48V = 2
+    DM4340 = 3
+    DM4340_48V = 4
+    DM6006 = 5
+    DM8006 = 6
+    DM8009 = 7
+    DM10010L = 8
+    DM10010 = 9
+    DMH3510 = 10
+    DMH6215 = 11
+    DMG6220 = 12
+
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 1
+    POS_VEL = 2
+    VEL = 3
+    TORQUE_POS = 4
+
+
+# Motor variable IDs (RID)
+class MotorVariable(IntEnum):
+    UV_VALUE = 0
+    KT_VALUE = 1
+    OT_VALUE = 2
+    OC_VALUE = 3
+    ACC = 4
+    DEC = 5
+    MAX_SPD = 6
+    MST_ID = 7
+    ESC_ID = 8
+    TIMEOUT = 9
+    CTRL_MODE = 10
+    DAMP = 11
+    INERTIA = 12
+    HW_VER = 13
+    SW_VER = 14
+    SN = 15
+    NPP = 16
+    RS = 17
+    LS = 18
+    FLUX = 19
+    GR = 20
+    PMAX = 21
+    VMAX = 22
+    TMAX = 23
+    I_BW = 24
+    KP_ASR = 25
+    KI_ASR = 26
+    KP_APR = 27
+    KI_APR = 28
+    OV_VALUE = 29
+    GREF = 30
+    DETA = 31
+    V_BW = 32
+    IQ_C1 = 33
+    VL_C1 = 34
+    CAN_BR = 35
+    SUB_VER = 36
+    U_OFF = 50
+    V_OFF = 51
+    K1 = 52
+    K2 = 53
+    M_OFF = 54
+    DIR = 55
+    P_M = 80
+    XOUT = 81
+
+
+# Motor limit parameters [PMAX, VMAX, TMAX]
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS = {
+    MotorType.DM3507: (12.5, 30, 10),
+    MotorType.DM4310: (12.5, 30, 10),
+    MotorType.DM4310_48V: (12.5, 50, 10),
+    MotorType.DM4340: (12.5, 8, 28),
+    MotorType.DM4340_48V: (12.5, 10, 28),
+    MotorType.DM6006: (12.5, 45, 20),
+    MotorType.DM8006: (12.5, 45, 40),
+    MotorType.DM8009: (12.5, 45, 54),
+    MotorType.DM10010L: (12.5, 25, 200),
+    MotorType.DM10010: (12.5, 20, 200),
+    MotorType.DMH3510: (12.5, 280, 1),
+    MotorType.DMH6215: (12.5, 45, 10),
+    MotorType.DMG6220: (12.5, 45, 10),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.DM3507: "dm3507",
+    MotorType.DM4310: "dm4310",
+    MotorType.DM4310_48V: "dm4310_48v",
+    MotorType.DM4340: "dm4340",
+    MotorType.DM4340_48V: "dm4340_48v",
+    MotorType.DM6006: "dm6006",
+    MotorType.DM8006: "dm8006",
+    MotorType.DM8009: "dm8009",
+    MotorType.DM10010L: "dm10010l",
+    MotorType.DM10010: "dm10010",
+    MotorType.DMH3510: "dmh3510",
+    MotorType.DMH6215: "dmh6215",
+    MotorType.DMG6220: "dmg6220",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "dm3507": 65536,
+    "dm4310": 65536,
+    "dm4310_48v": 65536,
+    "dm4340": 65536,
+    "dm4340_48v": 65536,
+    "dm6006": 65536,
+    "dm8006": 65536,
+    "dm8009": 65536,
+    "dm10010l": 65536,
+    "dm10010": 65536,
+    "dmh3510": 65536,
+    "dmh6215": 65536,
+    "dmg6220": 65536,
+}
+
+# CAN baudrates supported by Damiao motors
+AVAILABLE_BAUDRATES = [
+    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
+    3200000,  # 7: 3.2 mbps
+    4000000,  # 8: 4 mbps
+    5000000,  # 9: 5 mbps
+]
+DEFAULT_BAUDRATE = 1000000  # 1 Mbps is standard for OpenArms
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 1000
+
+# OpenArms specific configurations
+# Based on: https://docs.openarm.dev/software/setup/configure-test
+# OpenArms has 7 DOF per arm (14 total for dual arm)
+OPENARMS_ARM_MOTOR_IDS = {
+    "joint_1": {"send": 0x01, "recv": 0x11},  # J1 - Shoulder pan
+    "joint_2": {"send": 0x02, "recv": 0x12},  # J2 - Shoulder lift
+    "joint_3": {"send": 0x03, "recv": 0x13},  # J3 - Elbow flex
+    "joint_4": {"send": 0x04, "recv": 0x14},  # J4 - Wrist flex
+    "joint_5": {"send": 0x05, "recv": 0x15},  # J5 - Wrist roll
+    "joint_6": {"send": 0x06, "recv": 0x16},  # J6 - Wrist pitch
+    "joint_7": {"send": 0x07, "recv": 0x17},  # J7 - Wrist rotation
+}
+
+OPENARMS_GRIPPER_MOTOR_IDS = {
+    "gripper": {"send": 0x08, "recv": 0x18},  # J8 - Gripper
+}
+
+# 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
+}
+
+# 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_REFRESH = 0xCC
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF

src/lerobot/motors/dynamixel/dynamixel.py

@@ -22,9 +22,8 @@ import logging
 from copy import deepcopy
 from enum import Enum
 
-from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..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,
     MODEL_BAUDRATE_TABLE,
@@ -100,7 +99,7 @@ def _split_into_byte_chunks(value: int, length: int) -> list[int]:
     return data
 
 
-class DynamixelMotorsBus(MotorsBus):
+class DynamixelMotorsBus(SerialMotorsBus):
     """
     The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
     the motors. For more info, see the Dynamixel SDK Documentation:
@@ -182,10 +181,10 @@ class DynamixelMotorsBus(MotorsBus):
         for motor, m in self.motors.items():
             calibration[motor] = MotorCalibration(
                 id=m.id,
-                drive_mode=drive_modes[motor],
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                drive_mode=int(drive_modes[motor]),
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
             )
 
         return calibration
@@ -199,15 +198,15 @@ class DynamixelMotorsBus(MotorsBus):
         if cache:
             self.calibration = calibration_dict
 
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
 
@@ -236,7 +235,7 @@ class DynamixelMotorsBus(MotorsBus):
         On Dynamixel Motors:
         Present_Position = Actual_Position + Homing_Offset
         """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -259,6 +258,6 @@ class DynamixelMotorsBus(MotorsBus):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
 
-            return
+            return None
 
         return {id_: data[0] for id_, data in data_list.items()}

src/lerobot/motors/feetech/feetech.py

@@ -17,9 +17,8 @@ from copy import deepcopy
 from enum import Enum
 from pprint import pformat
 
-from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..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,
     FIRMWARE_MINOR_VERSION,
@@ -96,7 +95,7 @@ def patch_setPacketTimeout(self, packet_length):  # noqa: N802
     self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50
 
 
-class FeetechMotorsBus(MotorsBus):
+class FeetechMotorsBus(SerialMotorsBus):
     """
     The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
     python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.
@@ -127,7 +126,7 @@ class FeetechMotorsBus(MotorsBus):
 
         self.port_handler = scs.PortHandler(self.port)
         # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(  # type: ignore[method-assign]
             self.port_handler, scs.PortHandler
         )
         self.packet_handler = scs.PacketHandler(protocol_version)
@@ -263,9 +262,9 @@ class FeetechMotorsBus(MotorsBus):
             calibration[motor] = MotorCalibration(
                 id=m.id,
                 drive_mode=0,
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
             )
 
         return calibration
@@ -285,7 +284,7 @@ class FeetechMotorsBus(MotorsBus):
         On Feetech Motors:
         Present_Position = Actual_Position - Homing_Offset
         """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -293,18 +292,18 @@ class FeetechMotorsBus(MotorsBus):
 
         return half_turn_homings
 
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 0, num_retry=num_retry)
 
-    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
         addr, length = get_address(self.model_ctrl_table, model, "Lock")
-        self._write(addr, length, motor_id, 0, num_retry=num_retry)
+        self._write(addr, length, motor, 0, num_retry=num_retry)
 
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 1, num_retry=num_retry)
@@ -335,7 +334,7 @@ class FeetechMotorsBus(MotorsBus):
     def _broadcast_ping(self) -> tuple[dict[int, int], int]:
         import scservo_sdk as scs
 
-        data_list = {}
+        data_list: dict[int, int] = {}
 
         status_length = 6
 
@@ -415,7 +414,7 @@ class FeetechMotorsBus(MotorsBus):
         if not self._is_comm_success(comm):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return
+            return None
 
         ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
         if ids_errors:

src/lerobot/motors/motors_bus.py

@@ -19,8 +19,11 @@
 # TODO(aliberts): Add block noqa when feature below is available
 # https://github.com/astral-sh/ruff/issues/3711
 
+from __future__ import annotations
+
 import abc
 import logging
+from collections.abc import Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
 from enum import Enum
@@ -41,6 +44,81 @@ Value: TypeAlias = int | float
 logger = logging.getLogger(__name__)
 
 
+class MotorsBusBase(abc.ABC):
+    """
+    Base class for all motor bus implementations.
+
+    This is a minimal interface that all motor buses must implement, regardless of their
+    communication protocol (serial, CAN, etc.).
+    """
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+    ):
+        self.port = port
+        self.motors = motors
+        self.calibration = calibration if calibration else {}
+
+    @abc.abstractmethod
+    def connect(self, handshake: bool = True) -> None:
+        """Establish connection to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """Disconnect from the motors."""
+        pass
+
+    @property
+    @abc.abstractmethod
+    def is_connected(self) -> bool:
+        """Check if connected to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def read(self, data_name: str, motor: str) -> Value:
+        """Read a value from a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def write(self, data_name: str, motor: str, value: Value) -> None:
+        """Write a value to a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def sync_read(self, data_name: str, motors: str | list[str] | None = None) -> dict[str, Value]:
+        """Read a value from multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
+        """Write values to multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration parameters from the motors."""
+        pass
+
+    @abc.abstractmethod
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration parameters to the motors."""
+        pass
+
+
 def get_ctrl_table(model_ctrl_table: dict[str, dict], model: str) -> dict[str, tuple[int, int]]:
     ctrl_table = model_ctrl_table.get(model)
     if ctrl_table is None:
@@ -97,18 +175,21 @@ class Motor:
     id: int
     model: str
     norm_mode: MotorNormMode
+    motor_type_str: str | None = None
+    recv_id: int | None = None
 
 
 class PortHandler(Protocol):
-    def __init__(self, port_name):
-        self.is_open: bool
-        self.baudrate: int
-        self.packet_start_time: float
-        self.packet_timeout: float
-        self.tx_time_per_byte: float
-        self.is_using: bool
-        self.port_name: str
-        self.ser: serial.Serial
+    is_open: bool
+    baudrate: int
+    packet_start_time: float
+    packet_timeout: float
+    tx_time_per_byte: float
+    is_using: bool
+    port_name: str
+    ser: serial.Serial
+
+    def __init__(self, port_name: str) -> None: ...
 
     def openPort(self): ...
     def closePort(self): ...
@@ -161,19 +242,22 @@ class PacketHandler(Protocol):
     def regWriteTxRx(self, port, id, address, length, data): ...
     def syncReadTx(self, port, start_address, data_length, param, param_length): ...
     def syncWriteTxOnly(self, port, start_address, data_length, param, param_length): ...
+    def broadcastPing(self, port): ...
 
 
 class GroupSyncRead(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.last_result: bool
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    last_result: bool
+    is_param_changed: bool
+    param: list
+    data_dict: dict
 
+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id): ...
     def removeParam(self, id): ...
@@ -186,15 +270,17 @@ class GroupSyncRead(Protocol):
 
 
 class GroupSyncWrite(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    is_param_changed: bool
+    param: list
+    data_dict: dict
 
+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id, data): ...
     def removeParam(self, id): ...
@@ -203,15 +289,15 @@ class GroupSyncWrite(Protocol):
     def txPacket(self): ...
 
 
-class MotorsBus(abc.ABC):
+class SerialMotorsBus(MotorsBusBase):
     """
-    A MotorsBus allows to efficiently read and write to the attached motors.
+    A SerialMotorsBus allows to efficiently read and write to motors connected via serial communication.
     It represents several motors daisy-chained together and connected through a serial port.
-    There are currently two implementations of this abstract class:
+    There are currently two implementations of this class:
         - DynamixelMotorsBus
         - FeetechMotorsBus
 
-    Note: This class may evolve in the future should we add support for other types of bus.
+    This class is specifically for serial-based motor protocols (Dynamixel, Feetech, etc.).
 
     A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
     To find the port, you can run our utility script:
@@ -260,9 +346,7 @@ class MotorsBus(abc.ABC):
         motors: dict[str, Motor],
         calibration: dict[str, MotorCalibration] | None = None,
     ):
-        self.port = port
-        self.motors = motors
-        self.calibration = calibration if calibration else {}
+        super().__init__(port, motors, calibration)
 
         self.port_handler: PortHandler
         self.packet_handler: PacketHandler
@@ -323,7 +407,7 @@ class MotorsBus(abc.ABC):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motor_model(self, motor: NameOrID) -> int:
+    def _get_motor_model(self, motor: NameOrID) -> str:
         if isinstance(motor, str):
             return self.motors[motor].model
         elif isinstance(motor, int):
@@ -331,17 +415,19 @@ class MotorsBus(abc.ABC):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+    def _get_motors_list(self, motors: NameOrID | Sequence[NameOrID] | None) -> list[str]:
         if motors is None:
             return list(self.motors)
         elif isinstance(motors, str):
             return [motors]
-        elif isinstance(motors, list):
-            return motors.copy()
+        elif isinstance(motors, int):
+            return [self._id_to_name(motors)]
+        elif isinstance(motors, Sequence):
+            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
         else:
             raise TypeError(motors)
 
-    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
+    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> dict[int, Value]:
         if isinstance(values, (int | float)):
             return dict.fromkeys(self.ids, values)
         elif isinstance(values, dict):
@@ -532,7 +618,7 @@ class MotorsBus(abc.ABC):
         self.set_baudrate(self.default_baudrate)
 
     @abc.abstractmethod
-    def _find_single_motor(self, motor: str, initial_baudrate: int | None) -> tuple[int, int]:
+    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
         pass
 
     @abc.abstractmethod
@@ -545,13 +631,13 @@ class MotorsBus(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         """Disable torque on selected motors.
 
         Disabling Torque allows to write to the motors' permanent memory area (EPROM/EEPROM).
 
         Args:
-            motors (int | str | list[str] | None, optional): Target motors.  Accepts a motor name, an ID, a
+            motors ( str | list[str] | None, optional): Target motors.  Accepts a motor name, an ID, a
                 list of names or `None` to affect every registered motor.  Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
@@ -563,18 +649,19 @@ class MotorsBus(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         """Enable torque on selected motors.
 
         Args:
-            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
+            motors (int | str | list[str] | None, optional): Same semantics as :pymeth:`disable_torque`.
+                Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
         """
         pass
 
     @contextmanager
-    def torque_disabled(self, motors: int | str | list[str] | None = None):
+    def torque_disabled(self, motors: str | list[str] | None = None):
         """Context-manager that guarantees torque is re-enabled.
 
         This helper is useful to temporarily disable torque when configuring motors.
@@ -651,24 +738,19 @@ class MotorsBus(abc.ABC):
         """
         pass
 
-    def reset_calibration(self, motors: NameOrID | list[NameOrID] | None = None) -> None:
+    def reset_calibration(self, motors: NameOrID | Sequence[NameOrID] | None = None) -> None:
         """Restore factory calibration for the selected motors.
 
         Homing offset is set to ``0`` and min/max position limits are set to the full usable range.
         The in-memory :pyattr:`calibration` is cleared.
 
         Args:
-            motors (NameOrID | list[NameOrID] | None, optional): Selection of motors. `None` (default)
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Selection of motors. `None` (default)
                 resets every motor.
         """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)
 
-        for motor in motors:
+        for motor in motor_names:
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
             self.write("Homing_Offset", motor, 0, normalize=False)
@@ -677,7 +759,9 @@ class MotorsBus(abc.ABC):
 
         self.calibration = {}
 
-    def set_half_turn_homings(self, motors: NameOrID | list[NameOrID] | None = None) -> dict[NameOrID, Value]:
+    def set_half_turn_homings(
+        self, motors: NameOrID | Sequence[NameOrID] | None = None
+    ) -> dict[NameOrID, Value]:
         """Centre each motor range around its current position.
 
         The function computes and writes a homing offset such that the present position becomes exactly one
@@ -687,17 +771,12 @@ class MotorsBus(abc.ABC):
             motors (NameOrID | list[NameOrID] | None, optional): Motors to adjust. Defaults to all motors (`None`).
 
         Returns:
-            dict[NameOrID, Value]: Mapping *motor → written homing offset*.
+            dict[str, Value]: Mapping *motor name → written homing offset*.
         """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)
 
-        self.reset_calibration(motors)
-        actual_positions = self.sync_read("Present_Position", motors, normalize=False)
+        self.reset_calibration(motor_names)
+        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
         homing_offsets = self._get_half_turn_homings(actual_positions)
         for motor, offset in homing_offsets.items():
             self.write("Homing_Offset", motor, offset)
@@ -709,8 +788,8 @@ class MotorsBus(abc.ABC):
         pass
 
     def record_ranges_of_motion(
-        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        self, motors: NameOrID | Sequence[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
         """Interactively record the min/max encoder values of each motor.
 
         Move the joints by hand (with torque disabled) while the method streams live positions. Press
@@ -722,30 +801,25 @@ class MotorsBus(abc.ABC):
             display_values (bool, optional): When `True` (default) a live table is printed to the console.
 
         Returns:
-            tuple[dict[NameOrID, Value], dict[NameOrID, Value]]: Two dictionaries *mins* and *maxes* with the
+            tuple[dict[str, Value], dict[str, Value]]: Two dictionaries *mins* and *maxes* with the
                 extreme values observed for each motor.
         """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)
 
-        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
         mins = start_positions.copy()
         maxes = start_positions.copy()
 
         user_pressed_enter = False
         while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motors, normalize=False)
+            positions = self.sync_read("Present_Position", motor_names, normalize=False)
             mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
             maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}
 
             if display_values:
                 print("\n-------------------------------------------")
                 print(f"{'NAME':<15} | {'MIN':>6} | {'POS':>6} | {'MAX':>6}")
-                for motor in motors:
+                for motor in motor_names:
                     print(f"{motor:<15} | {mins[motor]:>6} | {positions[motor]:>6} | {maxes[motor]:>6}")
 
             if enter_pressed():
@@ -753,9 +827,9 @@ class MotorsBus(abc.ABC):
 
             if display_values and not user_pressed_enter:
                 # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motors) + 3)
+                move_cursor_up(len(motor_names) + 3)
 
-        same_min_max = [motor for motor in motors if mins[motor] == maxes[motor]]
+        same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
         if same_min_max:
             raise ValueError(f"Some motors have the same min and max values:\n{pformat(same_min_max)}")
 
@@ -878,12 +952,12 @@ class MotorsBus(abc.ABC):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
             else:
-                return
+                return None
         if self._is_error(error):
             if raise_on_error:
                 raise RuntimeError(self.packet_handler.getRxPacketError(error))
             else:
-                return
+                return None
 
         return model_number
 
@@ -930,12 +1004,13 @@ class MotorsBus(abc.ABC):
         err_msg = f"Failed to read '{data_name}' on {id_=} after {num_retry + 1} tries."
         value, _, _ = self._read(addr, length, id_, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
-        id_value = self._decode_sign(data_name, {id_: value})
+        decoded = self._decode_sign(data_name, {id_: value})
 
         if normalize and data_name in self.normalized_data:
-            id_value = self._normalize(id_value)
+            normalized = self._normalize(decoded)
+            return normalized[id_]
 
-        return id_value[id_]
+        return decoded[id_]
 
     def _read(
         self,
@@ -946,7 +1021,7 @@ class MotorsBus(abc.ABC):
         num_retry: int = 0,
         raise_on_error: bool = True,
         err_msg: str = "",
-    ) -> tuple[int, int]:
+    ) -> tuple[int, int, int]:
         if length == 1:
             read_fn = self.packet_handler.read1ByteTxRx
         elif length == 2:
@@ -996,13 +1071,14 @@ class MotorsBus(abc.ABC):
         model = self.motors[motor].model
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
+        int_value = int(value)
         if normalize and data_name in self.normalized_data:
-            value = self._unnormalize({id_: value})[id_]
+            int_value = self._unnormalize({id_: value})[id_]
 
-        value = self._encode_sign(data_name, {id_: value})[id_]
+        int_value = self._encode_sign(data_name, {id_: int_value})[id_]
 
-        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
-        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{int_value}' after {num_retry + 1} tries."
+        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     def _write(
         self,
@@ -1036,7 +1112,7 @@ class MotorsBus(abc.ABC):
     def sync_read(
         self,
         data_name: str,
-        motors: str | list[str] | None = None,
+        motors: NameOrID | Sequence[NameOrID] | None = None,
         *,
         normalize: bool = True,
         num_retry: int = 0,
@@ -1045,7 +1121,7 @@ class MotorsBus(abc.ABC):
 
         Args:
             data_name (str): Register name.
-            motors (str | list[str] | None, optional): Motors to query. `None` (default) reads every motor.
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Motors to query. `None` (default) reads every motor.
             normalize (bool, optional): Normalisation flag.  Defaults to `True`.
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
 
@@ -1066,16 +1142,17 @@ class MotorsBus(abc.ABC):
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
         err_msg = f"Failed to sync read '{data_name}' on {ids=} after {num_retry + 1} tries."
-        ids_values, _ = self._sync_read(
+        raw_ids_values, _ = self._sync_read(
             addr, length, ids, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
         )
 
-        ids_values = self._decode_sign(data_name, ids_values)
+        decoded = self._decode_sign(data_name, raw_ids_values)
 
         if normalize and data_name in self.normalized_data:
-            ids_values = self._normalize(ids_values)
+            normalized = self._normalize(decoded)
+            return {self._id_to_name(id_): value for id_, value in normalized.items()}
 
-        return {self._id_to_name(id_): value for id_, value in ids_values.items()}
+        return {self._id_to_name(id_): value for id_, value in decoded.items()}
 
     def _sync_read(
         self,
@@ -1147,21 +1224,24 @@ class MotorsBus(abc.ABC):
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
         """
 
-        ids_values = self._get_ids_values_dict(values)
-        models = [self._id_to_model(id_) for id_ in ids_values]
+        raw_ids_values = self._get_ids_values_dict(values)
+        models = [self._id_to_model(id_) for id_ in raw_ids_values]
         if self._has_different_ctrl_tables:
             assert_same_address(self.model_ctrl_table, models, data_name)
 
         model = next(iter(models))
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
+        int_ids_values = {id_: int(val) for id_, val in raw_ids_values.items()}
         if normalize and data_name in self.normalized_data:
-            ids_values = self._unnormalize(ids_values)
+            int_ids_values = self._unnormalize(raw_ids_values)
 
-        ids_values = self._encode_sign(data_name, ids_values)
+        int_ids_values = self._encode_sign(data_name, int_ids_values)
 
-        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
-        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to sync write '{data_name}' with ids_values={int_ids_values} after {num_retry + 1} tries."
+        self._sync_write(
+            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
+        )
 
     def _sync_write(
         self,
@@ -1194,3 +1274,7 @@ class MotorsBus(abc.ABC):
         for id_, value in ids_values.items():
             data = self._serialize_data(value, length)
             self.sync_writer.addParam(id_, data)
+
+
+# Backward compatibility alias
+MotorsBus: TypeAlias = SerialMotorsBus

src/lerobot/motors/robstride/__init__.py

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

src/lerobot/motors/robstride/tables.py

@@ -0,0 +1,120 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Configuration tables for Damiao motors."""
+
+from enum import IntEnum
+
+
+# Motor type definitions
+class MotorType(IntEnum):
+    O0 = 0
+    O1 = 1
+    O2 = 2
+    O3 = 3
+    O4 = 4
+    O5 = 5
+    ELO5 = 6
+    O6 = 7
+
+
+class CommMode(IntEnum):
+    PrivateProtocole = 0
+    CANopen = 1
+    MIT = 2
+
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 0
+    POS_VEL = 1
+    VEL = 2
+
+
+# Motor limit parameters [PMAX, VMAX, TMAX]
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS: dict[MotorType, tuple[float, float, float]] = {
+    MotorType.O0: (12.57, 33, 14),
+    MotorType.O1: (12.57, 44, 17),
+    MotorType.O2: (12.57, 33, 20),
+    MotorType.O3: (12.57, 33, 60),
+    MotorType.O4: (12.57, 33, 120),
+    MotorType.O5: (12.57, 50, 5.5),
+    MotorType.ELO5: (12.57, 50, 6),
+    MotorType.O6: (112.5, 50, 36),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.O0: "O0",
+    MotorType.O1: "O1",
+    MotorType.O2: "O2",
+    MotorType.O3: "O3",
+    MotorType.O4: "O4",
+    MotorType.O5: "O5",
+    MotorType.ELO5: "ELO5",
+    MotorType.O6: "O6",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "O0": 65536,
+    "O1": 65536,
+    "O2": 65536,
+    "O3": 65536,
+    "O4": 65536,
+    "O5": 65536,
+    "ELO5": 65536,
+    "O6": 65536,
+}
+
+# CAN baudrates supported by Robstride motors
+AVAILABLE_BAUDRATES = [
+    1000000,  # 4: 1 mbps (default)
+]
+DEFAULT_BAUDRATE = 1000000
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 0  # disabled by default, otherwise 20000 is 1s
+
+
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
+
+# MIT control parameter ranges
+MIT_KP_RANGE = (0.0, 500.0)
+MIT_KD_RANGE = (0.0, 5.0)
+
+# CAN frame command IDs
+CAN_CMD_ENABLE = 0xFC
+CAN_CMD_DISABLE = 0xFD
+CAN_CMD_SET_ZERO = 0xFE
+CAN_CMD_CLEAR_FAULT = 0xFB
+
+
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF
+
+
+RUNNING_TIMEOUT = 0.001
+PARAM_TIMEOUT = 0.01
+
+STATE_CACHE_TTL_S = 0.02

src/lerobot/policies/act/configuration_act.py

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

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -30,7 +30,7 @@ class DiffusionConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and `output_shapes`.
+    Those are: `input_features` and `output_features`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -48,21 +48,12 @@ class DiffusionConfig(PreTrainedConfig):
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.
@@ -73,7 +64,7 @@ class DiffusionConfig(PreTrainedConfig):
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
         spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
-        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
+        use_separate_rgb_encoder_per_camera: Whether to use a separate RGB encoder for each camera view.
         down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
             You may provide a variable number of dimensions, therefore also controlling the degree of
             downsampling.
@@ -148,6 +139,10 @@ class DiffusionConfig(PreTrainedConfig):
     # Inference
     num_inference_steps: int | None = None
 
+    # Optimization
+    compile_model: bool = False
+    compile_mode: str = "reduce-overhead"
+
     # Loss computation
     do_mask_loss_for_padding: bool = False
 

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -142,6 +142,9 @@ class DiffusionPolicy(PreTrainedPolicy):
         """Run the batch through the model and compute the loss for training or validation."""
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            for key in self.config.image_features:
+                if self.config.n_obs_steps == 1 and batch[key].ndim == 4:
+                    batch[key] = batch[key].unsqueeze(1)
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
         loss = self.diffusion.compute_loss(batch)
         # no output_dict so returning None
@@ -182,6 +185,11 @@ class DiffusionModel(nn.Module):
 
         self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
 
+        if config.compile_model:
+            # Compile the U-Net. "reduce-overhead" is preferred for the small-batch repetitive loops
+            # common in diffusion inference.
+            self.unet = torch.compile(self.unet, mode=config.compile_mode)
+
         self.noise_scheduler = _make_noise_scheduler(
             config.noise_scheduler_type,
             num_train_timesteps=config.num_train_timesteps,

src/lerobot/policies/pi0/modeling_pi0.py

@@ -995,7 +995,14 @@ class PI0Policy(PreTrainedPolicy):
 
         # Initialize model without loading weights
         # Check if dataset_stats were provided in kwargs
-        model = cls(config, **kwargs)
+        if _transformers_available:
+            from transformers.modeling_utils import no_init_weights
+
+            with no_init_weights():
+                model = cls(config, **kwargs)
+            model.model.paligemma_with_expert.paligemma.tie_weights()
+        else:
+            model = cls(config, **kwargs)
 
         # Now manually load and remap the state dict
         try:

src/lerobot/policies/pi05/modeling_pi05.py

@@ -967,7 +967,14 @@ class PI05Policy(PreTrainedPolicy):
 
         # Initialize model without loading weights
         # Check if dataset_stats were provided in kwargs
-        model = cls(config, **kwargs)
+        if _transformers_available:
+            from transformers.modeling_utils import no_init_weights
+
+            with no_init_weights():
+                model = cls(config, **kwargs)
+            model.model.paligemma_with_expert.paligemma.tie_weights()
+        else:
+            model = cls(config, **kwargs)
 
         # Now manually load and remap the state dict
         try:

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -895,7 +895,14 @@ class PI0FastPolicy(PreTrainedPolicy):
 
         # Initialize model without loading weights
         # Check if dataset_stats were provided in kwargs
-        model = cls(config, **kwargs)
+        if _transformers_available:
+            from transformers.modeling_utils import no_init_weights
+
+            with no_init_weights():
+                model = cls(config, **kwargs)
+            model.model.paligemma_with_expert.paligemma.tie_weights()
+        else:
+            model = cls(config, **kwargs)
 
         # Now manually load and remap the state dict
         try:

src/lerobot/policies/sac/modeling_sac.py

@@ -239,8 +239,10 @@ class SACPolicy(
                     + target_param.data * (1.0 - self.config.critic_target_update_weight)
                 )
 
-    def update_temperature(self):
-        self.temperature = self.log_alpha.exp().item()
+    @property
+    def temperature(self) -> float:
+        """Return the current temperature value, always in sync with log_alpha."""
+        return self.log_alpha.exp().item()
 
     def compute_loss_critic(
         self,
@@ -457,11 +459,10 @@ class SACPolicy(
             dim = continuous_action_dim + (1 if self.config.num_discrete_actions is not None else 0)
             self.target_entropy = -np.prod(dim) / 2
 
-    def _init_temperature(self):
-        """Set up temperature parameter and initial log_alpha."""
+    def _init_temperature(self) -> None:
+        """Set up temperature parameter (log_alpha)."""
         temp_init = self.config.temperature_init
         self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
-        self.temperature = self.log_alpha.exp().item()
 
 
 class SACObservationEncoder(nn.Module):

src/lerobot/policies/sarm/compute_rabc_weights.py

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

src/lerobot/policies/sarm/processor_sarm.py

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

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -85,7 +85,7 @@ class SmolVLAConfig(PreTrainedConfig):
     scheduler_decay_lr: float = 2.5e-6
 
     vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
-    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+    load_vlm_weights: bool = False  # Set to False in case of training the expert from scratch. True when init from pretrained SmolVLA weights
 
     add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.
 

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -30,7 +30,7 @@ Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
 lerobot-train \
 --policy.path=lerobot/smolvla_base \
---dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
+--dataset.repo_id=<USER>/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -40,7 +40,7 @@ and an action expert.
 ```bash
 lerobot-train \
 --policy.type=smolvla \
---dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
+--dataset.repo_id=<USER>/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -378,16 +378,16 @@ class SmolVLAPolicy(PreTrainedPolicy):
         actions_is_pad = batch.get("actions_id_pad")
         loss_dict = {}
         losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
-        loss_dict["losses_after_forward"] = losses.clone()
+        loss_dict["losses_after_forward"] = losses.clone().mean().item()
 
         if actions_is_pad is not None:
             in_episode_bound = ~actions_is_pad
             losses = losses * in_episode_bound.unsqueeze(-1)
-            loss_dict["losses_after_in_ep_bound"] = losses.clone()
+            loss_dict["losses_after_in_ep_bound"] = losses.clone().mean().item()
 
         # Remove padding
         losses = losses[:, :, : self.config.max_action_dim]
-        loss_dict["losses_after_rm_padding"] = losses.clone()
+        loss_dict["losses_after_rm_padding"] = losses.clone().mean().item()
 
         if reduction == "none":
             # Return per-sample losses (B,) by averaging over time and action dims

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -77,7 +77,6 @@ class SmolVLMWithExpertModel(nn.Module):
             print(f"Loading  {model_id} weights ...")
             self.vlm = AutoModelForImageTextToText.from_pretrained(
                 model_id,
-                device_map=device,
                 torch_dtype="bfloat16",
                 low_cpu_mem_usage=True,
             )

src/lerobot/policies/tdmpc/configuration_tdmpc.py

@@ -30,7 +30,7 @@ class TDMPCConfig(PreTrainedConfig):
     camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
+    Those are: `input_features`, `output_features`, and perhaps `max_random_shift_ratio`.
 
     Args:
         n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
@@ -40,24 +40,12 @@ class TDMPCConfig(PreTrainedConfig):
             is an alternative to using action repeats. If this is set to more than 1, then we require
             `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
             approach of using multiple steps from the plan is not in the original implementation.
-        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
-            the input data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
-            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
-            include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
-            the output data name, and the value is a list indicating the dimensions of the corresponding data.
-            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
-            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
-            match the original implementation.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
-            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
-            normalization mode here.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
         state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
         latent_dim: Observation's latent embedding dimension.

src/lerobot/policies/vqbet/configuration_vqbet.py

@@ -32,7 +32,7 @@ class VQBeTConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes` and `output_shapes`.
+    Those are: `input_features` and `output_features`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -46,21 +46,12 @@ class VQBeTConfig(PreTrainedConfig):
             current step and additional steps going back).
         n_action_pred_token: Total number of current token and future tokens that VQ-BeT predicts.
         action_chunk_size: Action chunk size of each action prediction token.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            The key represents the input data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "observation.image" refers to an input from
-            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, shapes doesnt include batch dimension or temporal dimension.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-            The key represents the output data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "action" refers to an output shape of [14], indicating
-            14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
-        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
-            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
-            [-1, 1] range.
-        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
-            original scale. Note that this is also used for normalizing the training targets.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.

src/lerobot/processor/__init__.py

@@ -44,6 +44,7 @@ from .hil_processor import (
     AddTeleopActionAsComplimentaryDataStep,
     AddTeleopEventsAsInfoStep,
     GripperPenaltyProcessorStep,
+    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     RewardClassifierProcessorStep,
@@ -87,6 +88,7 @@ __all__ = [
     "DoneProcessorStep",
     "EnvAction",
     "EnvTransition",
+    "GymHILAdapterProcessorStep",
     "GripperPenaltyProcessorStep",
     "hotswap_stats",
     "IdentityProcessorStep",

src/lerobot/processor/converters.py

@@ -168,11 +168,12 @@ def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
     """
     pad_keys = {k: v for k, v in batch.items() if "_is_pad" in k}
     task_key = {"task": batch["task"]} if "task" in batch else {}
+    subtask_key = {"subtask": batch["subtask"]} if "subtask" in batch else {}
     index_key = {"index": batch["index"]} if "index" in batch else {}
     task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
     episode_index_key = {"episode_index": batch["episode_index"]} if "episode_index" in batch else {}
 
-    return {**pad_keys, **task_key, **index_key, **task_index_key, **episode_index_key}
+    return {**pad_keys, **task_key, **subtask_key, **index_key, **task_index_key, **episode_index_key}
 
 
 def create_transition(

src/lerobot/processor/env_processor.py

@@ -17,7 +17,7 @@ from dataclasses import dataclass
 
 import torch
 
-from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.utils.constants import OBS_IMAGES, OBS_PREFIX, OBS_STATE, OBS_STR
 
 from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
@@ -92,7 +92,7 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # copy over non-STATE features
         for ft, feats in features.items():
-            if ft != PipelineFeatureType.STATE:
+            if ft != FeatureType.STATE:
                 new_features[ft] = feats.copy()
 
         # rebuild STATE features
@@ -100,13 +100,11 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # add our new flattened state
         state_feats[OBS_STATE] = PolicyFeature(
-            key=OBS_STATE,
+            type=FeatureType.STATE,
             shape=(8,),  # [eef_pos(3), axis_angle(3), gripper(2)]
-            dtype="float32",
-            description=("Concatenated end-effector position (3), axis-angle (3), and gripper qpos (2)."),
         )
 
-        new_features[PipelineFeatureType.STATE] = state_feats
+        new_features[FeatureType.STATE] = state_feats
 
         return new_features
 

src/lerobot/processor/gym_action_processor.py

@@ -20,6 +20,7 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 
 from .converters import to_tensor
 from .core import EnvAction, EnvTransition, PolicyAction
+from .hil_processor import TELEOP_ACTION_KEY
 from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry
 
 
@@ -89,6 +90,13 @@ class Numpy2TorchActionProcessorStep(ProcessorStep):
             torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
             new_transition[TransitionKey.ACTION] = torch_action
 
+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        if TELEOP_ACTION_KEY in complementary_data:
+            teleop_action = complementary_data[TELEOP_ACTION_KEY]
+            if isinstance(teleop_action, EnvAction):
+                complementary_data[TELEOP_ACTION_KEY] = to_tensor(teleop_action)
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
         return new_transition
 
     def transform_features(

src/lerobot/processor/hil_processor.py

@@ -18,16 +18,18 @@
 import math
 import time
 from dataclasses import dataclass
-from typing import Any, Protocol, TypeVar, runtime_checkable
+from typing import TYPE_CHECKING, Any, Protocol, TypeVar, runtime_checkable
 
 import numpy as np
 import torch
 import torchvision.transforms.functional as F  # noqa: N812
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
-from lerobot.teleoperators.teleoperator import Teleoperator
 from lerobot.teleoperators.utils import TeleopEvents
 
+if TYPE_CHECKING:
+    from lerobot.teleoperators.teleoperator import Teleoperator
+
 from .core import EnvTransition, PolicyAction, TransitionKey
 from .pipeline import (
     ComplementaryDataProcessorStep,
@@ -69,10 +71,10 @@ class HasTeleopEvents(Protocol):
 
 
 # Type variable constrained to Teleoperator subclasses that also implement events
-TeleopWithEvents = TypeVar("TeleopWithEvents", bound=Teleoperator)
+TeleopWithEvents = TypeVar("TeleopWithEvents", bound="Teleoperator")
 
 
-def _check_teleop_with_events(teleop: Teleoperator) -> None:
+def _check_teleop_with_events(teleop: "Teleoperator") -> None:
     """
     Runtime check that a teleoperator implements the `HasTeleopEvents` protocol.
 
@@ -103,7 +105,7 @@ class AddTeleopActionAsComplimentaryDataStep(ComplementaryDataProcessorStep):
         teleop_device: The teleoperator instance to get the action from.
     """
 
-    teleop_device: Teleoperator
+    teleop_device: "Teleoperator"
 
     def complementary_data(self, complementary_data: dict) -> dict:
         """
@@ -310,9 +312,40 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
         return features
 
 
+@ProcessorStepRegistry.register("gym_hil_adapter_processor")
+class GymHILAdapterProcessorStep(ProcessorStep):
+    """
+    Adapts the output of the `gym-hil` environment to the format expected by `lerobot` processors.
+
+    This step normalizes the `transition` object by:
+    1. Copying `teleop_action` from `info` to `complementary_data`.
+    2. Copying `is_intervention` from `info` (using the string key) to `info` (using the enum key).
+    """
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        info = transition.get(TransitionKey.INFO, {})
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+
+        if TELEOP_ACTION_KEY in info:
+            complementary_data[TELEOP_ACTION_KEY] = info[TELEOP_ACTION_KEY]
+
+        if "is_intervention" in info:
+            info[TeleopEvents.IS_INTERVENTION] = info["is_intervention"]
+
+        transition[TransitionKey.INFO] = info
+        transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
-class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
+class GripperPenaltyProcessorStep(ProcessorStep):
     """
     Applies a penalty for inefficient gripper usage.
 
@@ -327,26 +360,27 @@ class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
     penalty: float = -0.01
     max_gripper_pos: float = 30.0
 
-    def complementary_data(self, complementary_data: dict) -> dict:
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
         """
         Calculates the gripper penalty and adds it to the complementary data.
 
         Args:
-            complementary_data: The incoming complementary data, which should contain
-                                raw joint positions.
+            transition: The incoming environment transition.
 
         Returns:
-            A new complementary data dictionary with the `discrete_penalty` key added.
+            The modified transition with the penalty added to complementary data.
         """
-        action = self.transition.get(TransitionKey.ACTION)
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
 
         raw_joint_positions = complementary_data.get("raw_joint_positions")
         if raw_joint_positions is None:
-            return complementary_data
+            return new_transition
 
         current_gripper_pos = raw_joint_positions.get(GRIPPER_KEY, None)
         if current_gripper_pos is None:
-            return complementary_data
+            return new_transition
 
         # Gripper action is a PolicyAction at this stage
         gripper_action = action[-1].item()
@@ -362,11 +396,12 @@ class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
 
         gripper_penalty = self.penalty * int(gripper_penalty_bool)
 
-        # Create new complementary data with penalty info
+        # Update complementary data with penalty info
         new_complementary_data = dict(complementary_data)
         new_complementary_data[DISCRETE_PENALTY_KEY] = gripper_penalty
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
 
-        return new_complementary_data
+        return new_transition
 
     def get_config(self) -> dict[str, Any]:
         """

src/lerobot/processor/pipeline.py

@@ -413,7 +413,7 @@ class DataProcessorPipeline(HubMixin, Generic[TInput, TOutput]):
         Args:
             save_directory: The directory where the pipeline will be saved. If None, saves to
                 HF_LEROBOT_HOME/processors/{sanitized_pipeline_name}.
-            repo_id: ID of your repository on the Hub. Used only if `push_to_hub=True`.
+            repo_id: ID of your repository on the Hub. Used only if `push_to_hub=true`.
             push_to_hub: Whether or not to push your object to the Hugging Face Hub after saving it.
             card_kwargs: Additional arguments passed to the card template to customize the card.
             config_filename: The name of the JSON configuration file. If None, a name is

src/lerobot/processor/tokenizer_processor.py

@@ -34,6 +34,8 @@ from lerobot.utils.constants import (
     ACTION_TOKEN_MASK,
     ACTION_TOKENS,
     OBS_LANGUAGE_ATTENTION_MASK,
+    OBS_LANGUAGE_SUBTASK_ATTENTION_MASK,
+    OBS_LANGUAGE_SUBTASK_TOKENS,
     OBS_LANGUAGE_TOKENS,
 )
 from lerobot.utils.import_utils import _transformers_available
@@ -139,6 +141,32 @@ class TokenizerProcessorStep(ObservationProcessorStep):
 
         return None
 
+    def get_subtask(self, transition: EnvTransition) -> list[str] | None:
+        """
+        Extracts the subtask from the transition's complementary data.
+
+        Args:
+            transition: The environment transition.
+
+        Returns:
+            A list of subtask strings, or None if the subtask key is not found or the value is None.
+        """
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
+        if complementary_data is None:
+            return None
+
+        subtask = complementary_data.get("subtask")
+        if subtask is None:
+            return None
+
+        # Standardize to a list of strings for the tokenizer
+        if isinstance(subtask, str):
+            return [subtask]
+        elif isinstance(subtask, list) and all(isinstance(t, str) for t in subtask):
+            return subtask
+
+        return None
+
     def observation(self, observation: RobotObservation) -> RobotObservation:
         """
         Tokenizes the task description and adds it to the observation dictionary.
@@ -176,6 +204,24 @@ class TokenizerProcessorStep(ObservationProcessorStep):
         new_observation[OBS_LANGUAGE_TOKENS] = tokenized_prompt["input_ids"]
         new_observation[OBS_LANGUAGE_ATTENTION_MASK] = tokenized_prompt["attention_mask"].to(dtype=torch.bool)
 
+        # Tokenize subtask if available
+        subtask = self.get_subtask(self.transition)
+        if subtask is not None:
+            tokenized_subtask = self._tokenize_text(subtask)
+
+            # Move new tokenized tensors to the detected device
+            if target_device is not None:
+                tokenized_subtask = {
+                    k: v.to(target_device) if isinstance(v, torch.Tensor) else v
+                    for k, v in tokenized_subtask.items()
+                }
+
+            # Add tokenized subtask to the observation
+            new_observation[OBS_LANGUAGE_SUBTASK_TOKENS] = tokenized_subtask["input_ids"]
+            new_observation[OBS_LANGUAGE_SUBTASK_ATTENTION_MASK] = tokenized_subtask["attention_mask"].to(
+                dtype=torch.bool
+            )
+
         return new_observation
 
     def _detect_device(self, transition: EnvTransition) -> torch.device | None: