feat: add offline training in learner

feat: add RLT algorithm
feat: add RLT policy RL-token encoder-decoder and actor
2026-05-12 07:09:43 +00:00 · 2026-03-22 23:00:07 +01:00 · 2026-03-22 22:59:35 +01:00 · 2026-03-22 22:57:43 +01:00 · 2026-03-22 22:52:56 +01:00 · 2026-03-19 13:42:46 +01:00
164 changed files with 9908 additions and 2625 deletions
@@ -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:
@@ -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:
@@ -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%**   |
@@ -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 +27,8 @@
    title: Porting Large Datasets
  - local: using_dataset_tools
    title: Using the Dataset Tools
+  - local: dataset_subtask
+    title: Using Subtasks in the Dataset
  title: "Datasets"
 - sections:
  - local: act
@@ -57,8 +57,6 @@
    title: Use Async Inference
  - local: rtc
    title: Real-Time Chunking (RTC)
-  - local: training_time_rtc
-    title: Training-Time RTC
  title: "Inference"
 - sections:
  - local: envhub
@@ -103,11 +101,17 @@
    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
@@ -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`.
@@ -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
@@ -185,7 +185,7 @@ 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 \
@@ -224,7 +224,7 @@ 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 \
@@ -241,7 +241,7 @@ 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 +249,13 @@ 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
@@ -270,7 +270,7 @@ 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 \
  --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
@@ -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 +40,7 @@ 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 🤗
+## Step 3: Install LeRobot 🤗

 ### From Source

@@ -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)
@@ -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 \
@@ -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 \
@@ -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 \
@@ -1,86 +0,0 @@
-# Training-Time RTC
-
-Training-Time RTC teaches the model to handle inference delay during training.
-It feeds the **ground-truth action prefix** to the model and trains only on the remaining postfix actions.
-This keeps chunk transitions smooth without doing any inference-time inpainting.
-
-Based on: [Training-Time Action Conditioning for Efficient Real-Time Chunking](https://arxiv.org/abs/2512.05964).
-
-LeRobot supports this for `pi0`, `pi05` and `smolvla` without changing model parameters.
-
---
-
-## How It Works
-
-### At Training Time
-
- Sample a delay `d` per batch element.
- Keep the first `d` action steps as **ground truth** (no noise).
- Add noise only to the postfix actions.
- Set the flow-matching timestep to **1.0** for prefix tokens and normal timesteps for postfix tokens.
- Mask the loss to only train on the postfix.
-
-### At Inference Time
-
-When `rtc_training_config.enabled=true`, the model uses training-time RTC inference:
-
- Replace prefix positions in `x_t` with previous chunk's leftover actions.
- Set timestep to **1.0** for prefix positions.
-
---
-
-## Quick Start (CLI)
-
-```bash
-lerobot-train \
-  --policy.type=pi0 \
-  --dataset.repo_id=your/dataset \
-  --policy.rtc_training_config.enabled=true \
-  --policy.rtc_training_config.min_delay=0 \
-  --policy.rtc_training_config.max_delay=6 \
-  --policy.rtc_training_config.delay_distribution=UNIFORM
-```
-
---
-
-## Inference with Training-Time RTC
-
-After training with `rtc_training_config`, use the same config at inference. The model will automatically use training-time RTC inference:
-
-```python
-policy = PI0Policy.from_pretrained("path/to/trained/model")
-# rtc_training_config is loaded from the saved config
-
-actions = policy.predict_action_chunk(
-    batch,
-    inference_delay=5,  # estimated delay in timesteps
-    prev_chunk_left_over=previous_actions,  # from previous chunk
-)
-```
-
---
-
-## Key Parameters
-
-`RTCTrainingConfig` is available on the policy config (`pi0`, `pi05`, `smolvla`, `xvla`):
-
- **`enabled`**: Toggle training-time RTC (both training and inference).
- **`min_delay` / `max_delay`**: Delay range (inclusive).
- **`delay_distribution`**:
-  - `UNIFORM`: uniform in `[min_delay, max_delay]`
-  - `EXP`: exponentially decayed distribution over delays
- **`exp_decay`**: Exponential decay factor for `EXP` sampling.
-
---
-
-## Notes and Recommendations
-
- Start with `min_delay=0` and `max_delay` around your expected worst-case inference delay.
- Use `EXP` if you want more supervision on smaller delays.
-
---
-
-## Related Docs
-
- [Real-Time Chunking (Inference-Time RTC)](./rtc)
- [Pi0](./pi0), [Pi0.5](./pi05), [SmolVLA](./smolvla)
@@ -188,7 +188,105 @@ 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
+```
+
+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
+```
+
+**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

@@ -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:
@@ -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 \
@@ -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" \
@@ -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__":
@@ -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__":
@@ -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__":
@@ -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__":
@@ -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__":
@@ -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__":
@@ -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__":
@@ -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 \
@@ -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 \
@@ -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__":
@@ -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__":
@@ -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__":
@@ -4,7 +4,6 @@ from pathlib import Path
 from queue import Empty, Full

 import torch
-import torch.optim as optim

 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
@@ -12,6 +11,7 @@ from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.modeling_sac import SACPolicy
 from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -40,8 +40,9 @@ def run_learner(
    policy_learner.train()
    policy_learner.to(device)

-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
+    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
+    algorithm.make_optimizers()

    print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
    print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -83,24 +84,26 @@ def run_learner(
                else:
                    batch[key] = online_batch[key]

-            loss, _ = policy_learner.forward(batch)
+            def batch_iter(b=batch):
+                while True:
+                    yield b

-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+            stats = algorithm.update(batch_iter())
            training_step += 1

            if training_step % LOG_EVERY == 0:
+                log_dict = stats.to_log_dict()
                print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"[LEARNER] Training step {training_step}, "
+                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
                    f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                )

            # Send updated parameters to actor every 10 training steps
            if training_step % SEND_EVERY == 0:
                try:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
+                    weights = algorithm.get_weights()
+                    parameters_queue.put_nowait(weights)
                    print("[LEARNER] Sent updated parameters to actor")
                except Full:
                    # Missing write due to queue not being consumed (should happen rarely)
@@ -144,15 +147,15 @@ def run_actor(

            while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
+                    new_weights = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_weights)
                    print("[ACTOR] Updated policy parameters from learner")
                except Empty:  # No new updated parameters available from learner, waiting
                    pass

-                # Get action from policy
+                # Get action from policy (returns full action: continuous + discrete)
                policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action_tensor = policy_actor.select_action(policy_obs)
                action = action_tensor.squeeze(0).cpu().numpy()

                # Step environment
@@ -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",
@@ -105,12 +105,17 @@ dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
 damiao = ["python-can>=4.2.0,<5.0.0"]

 # 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]"]
@@ -355,9 +360,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false

-# [[tool.mypy.overrides]]
-# module = "lerobot.motors.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.motors.*"
+ignore_errors = false

 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"
@@ -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
@@ -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."""
@@ -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
@@ -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.")
@@ -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(
@@ -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)
@@ -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()
@@ -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.")
@@ -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):
@@ -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)
@@ -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.")
@@ -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.")
@@ -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
@@ -211,3 +211,15 @@ class TrainRLServerPipelineConfig(TrainPipelineConfig):
    # NOTE: In RL, we don't need an offline dataset
    # TODO: Make `TrainPipelineConfig.dataset` optional
    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional
+
+    # Algorithm name registered in RLAlgorithmConfig registry
+    algorithm: str = "sac"
+
+    # Data mixer strategy name. Currently supports "online_offline"
+    mixer: str = "online_offline"
+    # Fraction sampled from online replay when using OnlineOfflineMixer
+    online_ratio: float = 0.5
+
+    # RL trainer iterator
+    async_prefetch: bool = True
+    queue_size: int = 2
@@ -50,8 +50,3 @@ class RTCAttentionSchedule(str, Enum):
    ONES = "ONES"
    LINEAR = "LINEAR"
    EXP = "EXP"
-
-
-class RTCTrainingDelayDistribution(str, Enum):
-    UNIFORM = "UNIFORM"
-    EXP = "EXP"
@@ -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):
@@ -1396,6 +1396,132 @@ 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,
@@ -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,
@@ -162,6 +163,7 @@ 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 +520,7 @@ class LeRobotDatasetMetadata:
        _validate_feature_names(features)

        obj.tasks = None
+        obj.subtasks = None
        obj.episodes = None
        obj.stats = None
        obj.info = create_empty_dataset_info(
@@ -653,7 +656,7 @@ 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.
@@ -1075,6 +1078,12 @@ 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):
@@ -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):
@@ -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.
@@ -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",
@@ -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


@@ -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,
        )
@@ -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")
@@ -23,17 +23,20 @@ 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:
-    can.Message = object
-    can.interface = None
+
+    class can:  # noqa: N801
+        Message = object
+        interface = None
+

 import numpy as np

-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import enter_pressed, move_cursor_up

@@ -152,6 +155,7 @@ class DamiaoMotorsBus(MotorsBusBase):
        """Check if the CAN bus is connected."""
        return self._is_connected and self.canbus is not None

+    @check_if_already_connected
    def connect(self, handshake: bool = True) -> None:
        """
        Open the CAN bus and initialize communication.
@@ -159,10 +163,6 @@ class DamiaoMotorsBus(MotorsBusBase):
        Args:
            handshake: If True, ping all motors to verify they're present
        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(
-                f"{self.__class__.__name__}('{self.port}') is already connected."
-            )

        try:
            # Auto-detect interface type based on port name
@@ -206,11 +206,34 @@ class DamiaoMotorsBus(MotorsBusBase):
        Raises ConnectionError if any motor fails to respond.
        """
        logger.info("Starting handshake with motors...")
-        missing_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:
-            msg = self._refresh_motor(motor_name)
-            if msg is None:
+            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)
@@ -223,6 +246,7 @@ class DamiaoMotorsBus(MotorsBusBase):
            )
        logger.info("Handshake successful. All motors ready.")

+    @check_if_not_connected
    def disconnect(self, disable_torque: bool = True) -> None:
        """
        Close the CAN bus connection.
@@ -230,8 +254,6 @@ class DamiaoMotorsBus(MotorsBusBase):
        Args:
            disable_torque: If True, disable torque on all motors before disconnecting
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self.__class__.__name__}('{self.port}') is not connected.")

        if disable_torque:
            try:
@@ -259,7 +281,11 @@ class DamiaoMotorsBus(MotorsBusBase):
        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)
+        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)
@@ -317,7 +343,11 @@ class DamiaoMotorsBus(MotorsBusBase):
        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)
+        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)

@@ -333,6 +363,10 @@ class DamiaoMotorsBus(MotorsBusBase):
        Returns:
            CAN message if received, None otherwise
        """
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        try:
            start_time = time.time()
            messages_seen = []
@@ -371,10 +405,13 @@ class DamiaoMotorsBus(MotorsBusBase):
        Returns:
            Dictionary mapping recv_id to CAN message
        """
-        responses = {}
+        responses: dict[int, can.Message] = {}
        expected_set = set(expected_recv_ids)
        start_time = time.time()

+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        try:
            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
                # 100us poll timeout
@@ -438,8 +475,11 @@ class DamiaoMotorsBus(MotorsBusBase):
        motor_name = self._get_motor_name(motor)
        motor_type = self._motor_types[motor_name]

+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
-        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        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)
@@ -465,6 +505,9 @@ class DamiaoMotorsBus(MotorsBusBase):

        recv_id_to_motor: dict[int, str] = {}

+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        # Step 1: Send all MIT control commands
        for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
            motor_id = self._get_motor_id(motor)
@@ -472,7 +515,7 @@ class DamiaoMotorsBus(MotorsBusBase):
            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)
+            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
@@ -539,10 +582,9 @@ class DamiaoMotorsBus(MotorsBusBase):
        except Exception as e:
            logger.warning(f"Failed to decode response from {motor}: {e}")

+    @check_if_not_connected
    def read(self, data_name: str, motor: str) -> Value:
        """Read a value from a single motor. Positions are always in degrees."""
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        # Refresh motor to get latest state
        msg = self._refresh_motor(motor)
@@ -572,6 +614,7 @@ class DamiaoMotorsBus(MotorsBusBase):
            raise ValueError(f"Unknown data_name: {data_name}")
        return mapping[data_name]

+    @check_if_not_connected
    def write(
        self,
        data_name: str,
@@ -582,8 +625,6 @@ class DamiaoMotorsBus(MotorsBusBase):
        Write a value to a single motor. Positions are always in degrees.
        Can write 'Goal_Position', 'Kp', or 'Kd'.
        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")

        if data_name in ("Kp", "Kd"):
            self._gains[motor][data_name.lower()] = float(value)
@@ -633,14 +674,18 @@ class DamiaoMotorsBus(MotorsBusBase):

    def _batch_refresh(self, motors: list[str]) -> None:
        """Internal helper to refresh a list of motors and update cache."""
+
+        if self.canbus is None:
+            raise RuntimeError("CAN bus is not initialized.")
+
        # Send refresh commands
        for motor in motors:
            motor_id = self._get_motor_id(motor)
            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)
+            msg = can.Message(
+                arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd
+            )
            self.canbus.send(msg)
-            # Small delay to reduce bus congestion if necessary, though removed in sync_read previously
-            # precise_sleep(PRECISE_SLEEP_SEC)

        # Collect responses
        expected_recv_ids = [self._get_motor_recv_id(m) for m in motors]
@@ -655,10 +700,12 @@ class DamiaoMotorsBus(MotorsBusBase):
            else:
                logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")

-    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+    @check_if_not_connected
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
        """
        Write values to multiple motors simultaneously. Positions are always in degrees.
        """
+
        if data_name in ("Kp", "Kd"):
            key = data_name.lower()
            for motor, val in values.items():
@@ -667,6 +714,8 @@ class DamiaoMotorsBus(MotorsBusBase):
        elif data_name == "Goal_Position":
            # Step 1: Send all MIT control commands
            recv_id_to_motor: dict[int, str] = {}
+            if self.canbus is None:
+                raise RuntimeError("CAN bus is not initialized.")
            for motor, value_degrees in values.items():
                motor_id = self._get_motor_id(motor)
                motor_name = self._get_motor_name(motor)
@@ -676,7 +725,9 @@ class DamiaoMotorsBus(MotorsBusBase):
                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)
+                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)

@@ -707,9 +758,9 @@ class DamiaoMotorsBus(MotorsBusBase):

    def record_ranges_of_motion(
        self,
-        motors: NameOrID | list[NameOrID] | None = None,
+        motors: str | list[str] | None = None,
        display_values: bool = True,
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
        """
        Interactively record the min/max values of each motor in degrees.

@@ -181,10 +181,10 @@ class DynamixelMotorsBus(SerialMotorsBus):
        for motor, m in self.motors.items():
            calibration[motor] = MotorCalibration(
                id=m.id,
-                drive_mode=drive_modes[motor],
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                drive_mode=int(drive_modes[motor]),
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
            )

        return calibration
@@ -198,7 +198,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
        if cache:
            self.calibration = calibration_dict

-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)

@@ -206,7 +206,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
        addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)

-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)

@@ -235,7 +235,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
        On Dynamixel Motors:
        Present_Position = Actual_Position + Homing_Offset
        """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
        for motor, pos in positions.items():
            model = self._get_motor_model(motor)
            max_res = self.model_resolution_table[model] - 1
@@ -258,6 +258,6 @@ class DynamixelMotorsBus(SerialMotorsBus):
            if raise_on_error:
                raise ConnectionError(self.packet_handler.getTxRxResult(comm))

-            return
+            return None

        return {id_: data[0] for id_, data in data_list.items()}
@@ -126,7 +126,7 @@ class FeetechMotorsBus(SerialMotorsBus):

        self.port_handler = scs.PortHandler(self.port)
        # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(  # type: ignore[method-assign]
            self.port_handler, scs.PortHandler
        )
        self.packet_handler = scs.PacketHandler(protocol_version)
@@ -262,9 +262,9 @@ class FeetechMotorsBus(SerialMotorsBus):
            calibration[motor] = MotorCalibration(
                id=m.id,
                drive_mode=0,
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
+                homing_offset=int(offsets[motor]),
+                range_min=int(mins[motor]),
+                range_max=int(maxes[motor]),
            )

        return calibration
@@ -284,7 +284,7 @@ class FeetechMotorsBus(SerialMotorsBus):
        On Feetech Motors:
        Present_Position = Actual_Position - Homing_Offset
        """
-        half_turn_homings = {}
+        half_turn_homings: dict[NameOrID, Value] = {}
        for motor, pos in positions.items():
            model = self._get_motor_model(motor)
            max_res = self.model_resolution_table[model] - 1
@@ -292,7 +292,7 @@ class FeetechMotorsBus(SerialMotorsBus):

        return half_turn_homings

-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
            self.write("Lock", motor, 0, num_retry=num_retry)
@@ -303,7 +303,7 @@ class FeetechMotorsBus(SerialMotorsBus):
        addr, length = get_address(self.model_ctrl_table, model, "Lock")
        self._write(addr, length, motor, 0, num_retry=num_retry)

-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        for motor in self._get_motors_list(motors):
            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
            self.write("Lock", motor, 1, num_retry=num_retry)
@@ -334,7 +334,7 @@ class FeetechMotorsBus(SerialMotorsBus):
    def _broadcast_ping(self) -> tuple[dict[int, int], int]:
        import scservo_sdk as scs

-        data_list = {}
+        data_list: dict[int, int] = {}

        status_length = 6

@@ -414,7 +414,7 @@ class FeetechMotorsBus(SerialMotorsBus):
        if not self._is_comm_success(comm):
            if raise_on_error:
                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return
+            return None

        ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
        if ids_errors:
@@ -23,6 +23,7 @@ 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
@@ -93,7 +94,7 @@ class MotorsBusBase(abc.ABC):
        pass

    @abc.abstractmethod
-    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
        """Write values to multiple motors."""
        pass

@@ -179,15 +180,16 @@ class Motor:


 class PortHandler(Protocol):
-    def __init__(self, port_name):
-        self.is_open: bool
-        self.baudrate: int
-        self.packet_start_time: float
-        self.packet_timeout: float
-        self.tx_time_per_byte: float
-        self.is_using: bool
-        self.port_name: str
-        self.ser: serial.Serial
+    is_open: bool
+    baudrate: int
+    packet_start_time: float
+    packet_timeout: float
+    tx_time_per_byte: float
+    is_using: bool
+    port_name: str
+    ser: serial.Serial
+
+    def __init__(self, port_name: str) -> None: ...

    def openPort(self): ...
    def closePort(self): ...
@@ -240,19 +242,22 @@ class PacketHandler(Protocol):
    def regWriteTxRx(self, port, id, address, length, data): ...
    def syncReadTx(self, port, start_address, data_length, param, param_length): ...
    def syncWriteTxOnly(self, port, start_address, data_length, param, param_length): ...
+    def broadcastPing(self, port): ...


 class GroupSyncRead(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.last_result: bool
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    last_result: bool
+    is_param_changed: bool
+    param: list
+    data_dict: dict

+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
    def makeParam(self): ...
    def addParam(self, id): ...
    def removeParam(self, id): ...
@@ -265,15 +270,17 @@ class GroupSyncRead(Protocol):


 class GroupSyncWrite(Protocol):
-    def __init__(self, port, ph, start_address, data_length):
-        self.port: str
-        self.ph: PortHandler
-        self.start_address: int
-        self.data_length: int
-        self.is_param_changed: bool
-        self.param: list
-        self.data_dict: dict
+    port: str
+    ph: PortHandler
+    start_address: int
+    data_length: int
+    is_param_changed: bool
+    param: list
+    data_dict: dict

+    def __init__(
+        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
+    ) -> None: ...
    def makeParam(self): ...
    def addParam(self, id, data): ...
    def removeParam(self, id): ...
@@ -400,7 +407,7 @@ class SerialMotorsBus(MotorsBusBase):
        else:
            raise TypeError(f"'{motor}' should be int, str.")

-    def _get_motor_model(self, motor: NameOrID) -> int:
+    def _get_motor_model(self, motor: NameOrID) -> str:
        if isinstance(motor, str):
            return self.motors[motor].model
        elif isinstance(motor, int):
@@ -408,17 +415,19 @@ class SerialMotorsBus(MotorsBusBase):
        else:
            raise TypeError(f"'{motor}' should be int, str.")

-    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+    def _get_motors_list(self, motors: NameOrID | Sequence[NameOrID] | None) -> list[str]:
        if motors is None:
            return list(self.motors)
        elif isinstance(motors, str):
            return [motors]
-        elif isinstance(motors, list):
-            return motors.copy()
+        elif isinstance(motors, int):
+            return [self._id_to_name(motors)]
+        elif isinstance(motors, Sequence):
+            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
        else:
            raise TypeError(motors)

-    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
+    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> dict[int, Value]:
        if isinstance(values, (int | float)):
            return dict.fromkeys(self.ids, values)
        elif isinstance(values, dict):
@@ -640,18 +649,19 @@ class SerialMotorsBus(MotorsBusBase):
        pass

    @abc.abstractmethod
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
        """Enable torque on selected motors.

        Args:
-            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
+            motors (int | str | list[str] | None, optional): Same semantics as :pymeth:`disable_torque`.
+                Defaults to `None`.
            num_retry (int, optional): Number of additional retry attempts on communication failure.
                Defaults to 0.
        """
        pass

    @contextmanager
-    def torque_disabled(self, motors: int | str | list[str] | None = None):
+    def torque_disabled(self, motors: str | list[str] | None = None):
        """Context-manager that guarantees torque is re-enabled.

        This helper is useful to temporarily disable torque when configuring motors.
@@ -728,24 +738,19 @@ class SerialMotorsBus(MotorsBusBase):
        """
        pass

-    def reset_calibration(self, motors: NameOrID | list[NameOrID] | None = None) -> None:
+    def reset_calibration(self, motors: NameOrID | Sequence[NameOrID] | None = None) -> None:
        """Restore factory calibration for the selected motors.

        Homing offset is set to ``0`` and min/max position limits are set to the full usable range.
        The in-memory :pyattr:`calibration` is cleared.

        Args:
-            motors (NameOrID | list[NameOrID] | None, optional): Selection of motors. `None` (default)
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Selection of motors. `None` (default)
                resets every motor.
        """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)

-        for motor in motors:
+        for motor in motor_names:
            model = self._get_motor_model(motor)
            max_res = self.model_resolution_table[model] - 1
            self.write("Homing_Offset", motor, 0, normalize=False)
@@ -754,7 +759,9 @@ class SerialMotorsBus(MotorsBusBase):

        self.calibration = {}

-    def set_half_turn_homings(self, motors: NameOrID | list[NameOrID] | None = None) -> dict[NameOrID, Value]:
+    def set_half_turn_homings(
+        self, motors: NameOrID | Sequence[NameOrID] | None = None
+    ) -> dict[NameOrID, Value]:
        """Centre each motor range around its current position.

        The function computes and writes a homing offset such that the present position becomes exactly one
@@ -764,17 +771,12 @@ class SerialMotorsBus(MotorsBusBase):
            motors (NameOrID | list[NameOrID] | None, optional): Motors to adjust. Defaults to all motors (`None`).

        Returns:
-            dict[NameOrID, Value]: Mapping *motor → written homing offset*.
+            dict[str, Value]: Mapping *motor name → written homing offset*.
        """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)

-        self.reset_calibration(motors)
-        actual_positions = self.sync_read("Present_Position", motors, normalize=False)
+        self.reset_calibration(motor_names)
+        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
        homing_offsets = self._get_half_turn_homings(actual_positions)
        for motor, offset in homing_offsets.items():
            self.write("Homing_Offset", motor, offset)
@@ -786,8 +788,8 @@ class SerialMotorsBus(MotorsBusBase):
        pass

    def record_ranges_of_motion(
-        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
-    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        self, motors: NameOrID | Sequence[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[str, Value], dict[str, Value]]:
        """Interactively record the min/max encoder values of each motor.

        Move the joints by hand (with torque disabled) while the method streams live positions. Press
@@ -799,30 +801,25 @@ class SerialMotorsBus(MotorsBusBase):
            display_values (bool, optional): When `True` (default) a live table is printed to the console.

        Returns:
-            tuple[dict[NameOrID, Value], dict[NameOrID, Value]]: Two dictionaries *mins* and *maxes* with the
+            tuple[dict[str, Value], dict[str, Value]]: Two dictionaries *mins* and *maxes* with the
                extreme values observed for each motor.
        """
-        if motors is None:
-            motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
-            motors = [motors]
-        elif not isinstance(motors, list):
-            raise TypeError(motors)
+        motor_names = self._get_motors_list(motors)

-        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
        mins = start_positions.copy()
        maxes = start_positions.copy()

        user_pressed_enter = False
        while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motors, normalize=False)
+            positions = self.sync_read("Present_Position", motor_names, normalize=False)
            mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
            maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}

            if display_values:
                print("\n-------------------------------------------")
                print(f"{'NAME':<15} | {'MIN':>6} | {'POS':>6} | {'MAX':>6}")
-                for motor in motors:
+                for motor in motor_names:
                    print(f"{motor:<15} | {mins[motor]:>6} | {positions[motor]:>6} | {maxes[motor]:>6}")

            if enter_pressed():
@@ -830,9 +827,9 @@ class SerialMotorsBus(MotorsBusBase):

            if display_values and not user_pressed_enter:
                # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motors) + 3)
+                move_cursor_up(len(motor_names) + 3)

-        same_min_max = [motor for motor in motors if mins[motor] == maxes[motor]]
+        same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
        if same_min_max:
            raise ValueError(f"Some motors have the same min and max values:\n{pformat(same_min_max)}")

@@ -955,12 +952,12 @@ class SerialMotorsBus(MotorsBusBase):
            if raise_on_error:
                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
            else:
-                return
+                return None
        if self._is_error(error):
            if raise_on_error:
                raise RuntimeError(self.packet_handler.getRxPacketError(error))
            else:
-                return
+                return None

        return model_number

@@ -1007,12 +1004,13 @@ class SerialMotorsBus(MotorsBusBase):
        err_msg = f"Failed to read '{data_name}' on {id_=} after {num_retry + 1} tries."
        value, _, _ = self._read(addr, length, id_, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)

-        id_value = self._decode_sign(data_name, {id_: value})
+        decoded = self._decode_sign(data_name, {id_: value})

        if normalize and data_name in self.normalized_data:
-            id_value = self._normalize(id_value)
+            normalized = self._normalize(decoded)
+            return normalized[id_]

-        return id_value[id_]
+        return decoded[id_]

    def _read(
        self,
@@ -1023,7 +1021,7 @@ class SerialMotorsBus(MotorsBusBase):
        num_retry: int = 0,
        raise_on_error: bool = True,
        err_msg: str = "",
-    ) -> tuple[int, int]:
+    ) -> tuple[int, int, int]:
        if length == 1:
            read_fn = self.packet_handler.read1ByteTxRx
        elif length == 2:
@@ -1073,13 +1071,14 @@ class SerialMotorsBus(MotorsBusBase):
        model = self.motors[motor].model
        addr, length = get_address(self.model_ctrl_table, model, data_name)

+        int_value = int(value)
        if normalize and data_name in self.normalized_data:
-            value = self._unnormalize({id_: value})[id_]
+            int_value = self._unnormalize({id_: value})[id_]

-        value = self._encode_sign(data_name, {id_: value})[id_]
+        int_value = self._encode_sign(data_name, {id_: int_value})[id_]

-        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
-        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{int_value}' after {num_retry + 1} tries."
+        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)

    def _write(
        self,
@@ -1113,7 +1112,7 @@ class SerialMotorsBus(MotorsBusBase):
    def sync_read(
        self,
        data_name: str,
-        motors: str | list[str] | None = None,
+        motors: NameOrID | Sequence[NameOrID] | None = None,
        *,
        normalize: bool = True,
        num_retry: int = 0,
@@ -1122,7 +1121,7 @@ class SerialMotorsBus(MotorsBusBase):

        Args:
            data_name (str): Register name.
-            motors (str | list[str] | None, optional): Motors to query. `None` (default) reads every motor.
+            motors (NameOrID | Sequence[NameOrID] | None, optional): Motors to query. `None` (default) reads every motor.
            normalize (bool, optional): Normalisation flag.  Defaults to `True`.
            num_retry (int, optional): Retry attempts.  Defaults to `0`.

@@ -1143,16 +1142,17 @@ class SerialMotorsBus(MotorsBusBase):
        addr, length = get_address(self.model_ctrl_table, model, data_name)

        err_msg = f"Failed to sync read '{data_name}' on {ids=} after {num_retry + 1} tries."
-        ids_values, _ = self._sync_read(
+        raw_ids_values, _ = self._sync_read(
            addr, length, ids, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
        )

-        ids_values = self._decode_sign(data_name, ids_values)
+        decoded = self._decode_sign(data_name, raw_ids_values)

        if normalize and data_name in self.normalized_data:
-            ids_values = self._normalize(ids_values)
+            normalized = self._normalize(decoded)
+            return {self._id_to_name(id_): value for id_, value in normalized.items()}

-        return {self._id_to_name(id_): value for id_, value in ids_values.items()}
+        return {self._id_to_name(id_): value for id_, value in decoded.items()}

    def _sync_read(
        self,
@@ -1224,21 +1224,24 @@ class SerialMotorsBus(MotorsBusBase):
            num_retry (int, optional): Retry attempts.  Defaults to `0`.
        """

-        ids_values = self._get_ids_values_dict(values)
-        models = [self._id_to_model(id_) for id_ in ids_values]
+        raw_ids_values = self._get_ids_values_dict(values)
+        models = [self._id_to_model(id_) for id_ in raw_ids_values]
        if self._has_different_ctrl_tables:
            assert_same_address(self.model_ctrl_table, models, data_name)

        model = next(iter(models))
        addr, length = get_address(self.model_ctrl_table, model, data_name)

+        int_ids_values = {id_: int(val) for id_, val in raw_ids_values.items()}
        if normalize and data_name in self.normalized_data:
-            ids_values = self._unnormalize(ids_values)
+            int_ids_values = self._unnormalize(raw_ids_values)

-        ids_values = self._encode_sign(data_name, ids_values)
+        int_ids_values = self._encode_sign(data_name, int_ids_values)

-        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
-        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to sync write '{data_name}' with ids_values={int_ids_values} after {num_retry + 1} tries."
+        self._sync_write(
+            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
+        )

    def _sync_write(
        self,
@@ -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.
@@ -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.
@@ -20,7 +20,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
-from lerobot.policies.rtc.configuration_rtc import RTCConfig, RTCTrainingConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE

 DEFAULT_IMAGE_SIZE = 224
@@ -50,9 +50,8 @@ class PI0Config(PreTrainedConfig):
    min_period: float = 4e-3
    max_period: float = 4.0

-    # Real-Time Chunking (RTC) configurations
+    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None
-    rtc_training_config: RTCTrainingConfig | None = None

    image_resolution: tuple[int, int] = (
        DEFAULT_IMAGE_SIZE,
@@ -44,12 +44,6 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
-from lerobot.policies.rtc.training_time import (
-    apply_rtc_training_time,
-    apply_training_time_rtc_inference,
-    masked_mean,
-    sample_rtc_delay,
-)
 from lerobot.utils.constants import (
    ACTION,
    OBS_LANGUAGE_ATTENTION_MASK,
@@ -85,8 +79,8 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
    if dimension % 2 != 0:
        raise ValueError(f"dimension ({dimension}) must be divisible by 2")

-    if time.ndim not in (1, 2):
-        raise ValueError("The time tensor is expected to be of shape `(batch_size,)` or `(batch_size, T)`.")
+    if time.ndim != 1:
+        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")

    dtype = get_safe_dtype(torch.float64, device.type)
    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
@@ -94,14 +88,8 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd

    # Compute the outer product
    scaling_factor = 1.0 / period * 2 * math.pi
-    if time.ndim == 1:
-        sin_input = scaling_factor[None, :] * time[:, None]
-        return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-
-    time_flat = time.reshape(-1)
-    sin_input = scaling_factor[None, :] * time_flat[:, None]
-    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-    return pos_emb.reshape(*time.shape, dimension)
+    sin_input = scaling_factor[None, :] * time[:, None]
+    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)


 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
@@ -617,9 +605,6 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
    def _rtc_enabled(self):
        return self.config.rtc_config is not None and self.config.rtc_config.enabled

-    def _training_time_rtc_inference_enabled(self):
-        return self.config.rtc_training_config is not None and self.config.rtc_training_config.enabled
-
    def _apply_checkpoint(self, func, *args, **kwargs):
        """Helper method to apply gradient checkpointing if enabled."""
        if self.gradient_checkpointing_enabled and self.training:
@@ -729,10 +714,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`

        action_emb = self._apply_checkpoint(action_proj_func, noisy_actions)

-        if time_emb.dim() == 2:
-            time_emb = time_emb[:, None, :].expand_as(action_emb)
-        elif time_emb.shape[:2] != action_emb.shape[:2]:
-            raise ValueError(f"Expected time_emb shape {action_emb.shape[:2]}, got {time_emb.shape[:2]}")
+        time_emb = time_emb[:, None, :].expand_as(action_emb)
        action_time_emb = torch.cat([action_emb, time_emb], dim=2)

        def mlp_func(action_time_emb):
@@ -768,12 +750,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        if time is None:
            time = self.sample_time(actions.shape[0], actions.device)

-        if time.ndim == 1:
-            time_expanded = time[:, None, None]
-        elif time.ndim == 2:
-            time_expanded = time[:, :, None]
-        else:
-            raise ValueError(f"Expected time shape (B,) or (B, T), got {time.shape}")
+        time_expanded = time[:, None, None]
        x_t = time_expanded * noise + (1 - time_expanded) * actions
        u_t = noise - actions

@@ -869,37 +846,24 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`

        dt = -1.0 / num_steps

-        inference_delay = kwargs.get("inference_delay")
-        prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
-        execution_horizon = kwargs.get("execution_horizon")
-        use_training_time_rtc = self._training_time_rtc_inference_enabled()
-
        x_t = noise
        for step in range(num_steps):
            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)

-            if use_training_time_rtc:
-                x_t_cond, time_tensor = apply_training_time_rtc_inference(
-                    x_t, time, inference_delay, prev_chunk_left_over, self.config.chunk_size
-                )
-                v_t = self.denoise_step(
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
                    state=state,
                    prefix_pad_masks=prefix_pad_masks,
                    past_key_values=past_key_values,
-                    x_t=x_t_cond,
-                    timestep=time_tensor,
+                    x_t=input_x_t,
+                    timestep=current_timestep,
                )
-            elif self._rtc_enabled():
-                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)

-                def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
-                    return self.denoise_step(
-                        state=state,
-                        prefix_pad_masks=prefix_pad_masks,
-                        past_key_values=past_key_values,
-                        x_t=input_x_t,
-                        timestep=current_timestep,
-                    )
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")

                v_t = self.rtc_processor.denoise_step(
                    x_t=x_t,
@@ -910,14 +874,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
                    execution_horizon=execution_horizon,
                )
            else:
-                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
-                v_t = self.denoise_step(
-                    state=state,
-                    prefix_pad_masks=prefix_pad_masks,
-                    past_key_values=past_key_values,
-                    x_t=x_t,
-                    timestep=time_tensor,
-                )
+                v_t = denoise_step_partial_call(x_t)

            x_t = x_t + dt * v_t

@@ -1320,19 +1277,7 @@ class PI0Policy(PreTrainedPolicy):
        actions = self.prepare_action(batch)

        # Compute loss
-        postfix_mask = None
-        rtc_cfg = self.config.rtc_training_config
-        if rtc_cfg is not None and rtc_cfg.enabled and self.training:
-            batch_size = actions.shape[0]
-            time = self.model.sample_time(batch_size, actions.device)
-            noise = self.model.sample_noise(actions.shape, actions.device)
-            delay = sample_rtc_delay(rtc_cfg, batch_size, actions.device)
-            time, postfix_mask = apply_rtc_training_time(time, delay, actions.shape[1])
-            losses = self.model.forward(
-                images, img_masks, lang_tokens, lang_masks, state, actions, noise=noise, time=time
-            )
-        else:
-            losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions)
+        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions)

        # Truncate losses to actual action dimensions
        original_action_dim = self.config.output_features[ACTION].shape[0]
@@ -1344,12 +1289,12 @@ class PI0Policy(PreTrainedPolicy):

        if reduction == "none":
            # Return per-sample losses (B,) by averaging over time and action dims
-            per_sample_loss = masked_mean(losses, postfix_mask, reduce_dims=(1, 2))
+            per_sample_loss = losses.mean(dim=(1, 2))
            loss_dict["loss"] = per_sample_loss.mean().item()
            return per_sample_loss, loss_dict
        else:
            # Default: return scalar mean loss
-            loss = masked_mean(losses, postfix_mask, reduce_dims=(0, 1, 2))
+            loss = losses.mean()
            loss_dict["loss"] = loss.item()
            return loss, loss_dict

@@ -20,7 +20,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
-from lerobot.policies.rtc.configuration_rtc import RTCConfig, RTCTrainingConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE

 DEFAULT_IMAGE_SIZE = 224
@@ -52,7 +52,6 @@ class PI05Config(PreTrainedConfig):

    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None
-    rtc_training_config: RTCTrainingConfig | None = None

    image_resolution: tuple[int, int] = (
        DEFAULT_IMAGE_SIZE,
@@ -44,12 +44,6 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
-from lerobot.policies.rtc.training_time import (
-    apply_rtc_training_time,
-    apply_training_time_rtc_inference,
-    masked_mean,
-    sample_rtc_delay,
-)
 from lerobot.utils.constants import (
    ACTION,
    OBS_LANGUAGE_ATTENTION_MASK,
@@ -84,8 +78,8 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
    if dimension % 2 != 0:
        raise ValueError(f"dimension ({dimension}) must be divisible by 2")

-    if time.ndim not in (1, 2):
-        raise ValueError("The time tensor is expected to be of shape `(batch_size,)` or `(batch_size, T)`.")
+    if time.ndim != 1:
+        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")

    dtype = get_safe_dtype(torch.float64, device.type)
    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
@@ -93,14 +87,8 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd

    # Compute the outer product
    scaling_factor = 1.0 / period * 2 * math.pi
-    if time.ndim == 1:
-        sin_input = scaling_factor[None, :] * time[:, None]
-        return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-
-    time_flat = time.reshape(-1)
-    sin_input = scaling_factor[None, :] * time_flat[:, None]
-    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-    return pos_emb.reshape(*time.shape, dimension)
+    sin_input = scaling_factor[None, :] * time[:, None]
+    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)


 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
@@ -614,9 +602,6 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
    def _rtc_enabled(self):
        return self.config.rtc_config is not None and self.config.rtc_config.enabled

-    def _training_time_rtc_inference_enabled(self):
-        return self.config.rtc_training_config is not None and self.config.rtc_training_config.enabled
-
    def _apply_checkpoint(self, func, *args, **kwargs):
        """Helper method to apply gradient checkpointing if enabled."""
        if self.gradient_checkpointing_enabled and self.training:
@@ -744,12 +729,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
        if time is None:
            time = self.sample_time(actions.shape[0], actions.device)

-        if time.ndim == 1:
-            time_expanded = time[:, None, None]
-        elif time.ndim == 2:
-            time_expanded = time[:, :, None]
-        else:
-            raise ValueError(f"Expected time shape (B,) or (B, T), got {time.shape}")
+        time_expanded = time[:, None, None]
        x_t = time_expanded * noise + (1 - time_expanded) * actions
        u_t = noise - actions

@@ -840,35 +820,23 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`

        dt = -1.0 / num_steps

-        inference_delay = kwargs.get("inference_delay")
-        prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
-        execution_horizon = kwargs.get("execution_horizon")
-        use_training_time_rtc = self._training_time_rtc_inference_enabled()
-
        x_t = noise
        for step in range(num_steps):
            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)

-            if use_training_time_rtc:
-                x_t_cond, time_tensor = apply_training_time_rtc_inference(
-                    x_t, time, inference_delay, prev_chunk_left_over, self.config.chunk_size
-                )
-                v_t = self.denoise_step(
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
                    prefix_pad_masks=prefix_pad_masks,
                    past_key_values=past_key_values,
-                    x_t=x_t_cond,
-                    timestep=time_tensor,
+                    x_t=input_x_t,
+                    timestep=current_timestep,
                )
-            elif self._rtc_enabled():
-                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)

-                def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
-                    return self.denoise_step(
-                        prefix_pad_masks=prefix_pad_masks,
-                        past_key_values=past_key_values,
-                        x_t=input_x_t,
-                        timestep=current_timestep,
-                    )
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")

                v_t = self.rtc_processor.denoise_step(
                    x_t=x_t,
@@ -879,13 +847,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
                    execution_horizon=execution_horizon,
                )
            else:
-                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
-                v_t = self.denoise_step(
-                    prefix_pad_masks=prefix_pad_masks,
-                    past_key_values=past_key_values,
-                    x_t=x_t,
-                    timestep=time_tensor,
-                )
+                v_t = denoise_step_partial_call(x_t)

            x_t = x_t + dt * v_t

@@ -1288,17 +1250,7 @@ class PI05Policy(PreTrainedPolicy):
        actions = self.prepare_action(batch)

        # Compute loss (no separate state needed for PI05)
-        postfix_mask = None
-        rtc_cfg = self.config.rtc_training_config
-        if rtc_cfg is not None and rtc_cfg.enabled and self.training:
-            batch_size = actions.shape[0]
-            time = self.model.sample_time(batch_size, actions.device)
-            noise = self.model.sample_noise(actions.shape, actions.device)
-            delay = sample_rtc_delay(rtc_cfg, batch_size, actions.device)
-            time, postfix_mask = apply_rtc_training_time(time, delay, actions.shape[1])
-            losses = self.model.forward(images, img_masks, tokens, masks, actions, noise=noise, time=time)
-        else:
-            losses = self.model.forward(images, img_masks, tokens, masks, actions)
+        losses = self.model.forward(images, img_masks, tokens, masks, actions)

        # Truncate losses to actual action dimensions
        original_action_dim = self.config.output_features[ACTION].shape[0]
@@ -1310,12 +1262,12 @@ class PI05Policy(PreTrainedPolicy):

        if reduction == "none":
            # Return per-sample losses (B,) by averaging over time and action dims
-            per_sample_loss = masked_mean(losses, postfix_mask, reduce_dims=(1, 2))
+            per_sample_loss = losses.mean(dim=(1, 2))
            loss_dict["loss"] = per_sample_loss.mean().item()
            return per_sample_loss, loss_dict
        else:
            # Default: return scalar mean loss
-            loss = masked_mean(losses, postfix_mask, reduce_dims=(0, 1, 2))
+            loss = losses.mean()
            loss_dict["loss"] = loss.item()
            return loss, loss_dict

@@ -0,0 +1,18 @@
+# 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 lerobot.policies.rlt.configuration_rlt import RLTConfig
+from lerobot.policies.rlt.modeling_rlt import RLTPolicy
+
+__all__ = ["RLTConfig", "RLTPolicy"]
@@ -0,0 +1,156 @@
+# 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.
+"""RLT (RL Token) policy configuration.
+
+Reference: "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"
+(Xu et al., Physical Intelligence, 2026)
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+from lerobot.policies.sac.configuration_sac import ActorLearnerConfig, ConcurrencyConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGE, OBS_STATE
+
+
+@dataclass
+class RLTokenConfig:
+    """Configuration for the RL-token encoder/decoder transformer."""
+
+    input_dim: int = 2048
+    rl_token_dim: int = 2048
+    num_encoder_layers: int = 2
+    num_decoder_layers: int = 2
+    num_heads: int = 8
+    ff_dim: int = 2048
+    dropout: float = 0.0
+
+
+@dataclass
+class RLTActorConfig:
+    """Configuration for the lightweight RL actor MLP."""
+
+    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
+    std: float = 0.1
+
+
+@dataclass
+class RLTCriticConfig:
+    """Configuration for the RLT critic MLP."""
+
+    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
+
+
+@PreTrainedConfig.register_subclass("rlt")
+@dataclass
+class RLTConfig(PreTrainedConfig):
+    """Configuration for the RLT (RL Token) policy.
+
+    RLT adds an RL-token encoder/decoder to a frozen VLA backbone, then trains
+    a lightweight actor-critic head using the RL token as state representation.
+    The frozen VLA also provides reference action chunks that the actor refines.
+    """
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MIN_MAX,
+            "ACTION": NormalizationMode.MIN_MAX,
+        }
+    )
+
+    dataset_stats: dict[str, dict[str, list[float]]] | None = field(
+        default_factory=lambda: {
+            OBS_IMAGE: {
+                "mean": [0.485, 0.456, 0.406],
+                "std": [0.229, 0.224, 0.225],
+            },
+            OBS_STATE: {"min": [0.0], "max": [1.0]},
+            ACTION: {"min": [0.0], "max": [1.0]},
+        }
+    )
+
+    # ── Device ──
+    device: str = "cuda"
+    storage_device: str = "cpu"
+
+    # ── VLA backbone ──
+    vla_checkpoint: str | None = None
+
+    # ── RL-token ──
+    rl_token: RLTokenConfig = field(default_factory=RLTokenConfig)
+
+    # ── Actor / Critic heads ──
+    actor: RLTActorConfig = field(default_factory=RLTActorConfig)
+    critic: RLTCriticConfig = field(default_factory=RLTCriticConfig)
+
+    # ── Action chunks ──
+    chunk_size: int = 10
+    vla_chunk_size: int = 50
+
+    # ── Training parameters ──
+    online_steps: int = 50000
+    offline_steps: int = 5000
+    online_buffer_capacity: int = 100000
+    offline_buffer_capacity: int = 100000
+    online_step_before_learning: int = 500
+    warmup_steps: int = 500
+    async_prefetch: bool = False
+
+    # ── Algorithm hyperparameters ──
+    utd_ratio: int = 5
+    policy_update_freq: int = 2
+    discount: float = 0.99
+    critic_lr: float = 3e-4
+    actor_lr: float = 3e-4
+    rl_token_lr: float = 1e-4
+    tau: float = 0.005
+    clip_grad_norm: float = 10.0
+    num_critics: int = 2
+    bc_reg_coeff: float = 0.1
+    ref_dropout: float = 0.5
+    chunk_stride: int = 2
+    vla_finetune_weight: float = 0.0
+
+    # ── Distributed ──
+    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
+    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
+
+    def __post_init__(self):
+        super().__post_init__()
+
+    def get_optimizer_preset(self):
+        return None
+
+    def get_scheduler_preset(self):
+        return None
+
+    def validate_features(self) -> None:
+        if ACTION not in self.output_features:
+            raise ValueError("You must provide 'action' in the output features")
+
+    @property
+    def observation_delta_indices(self) -> list | None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list | None:
+        return None
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None
@@ -0,0 +1,318 @@
+# 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.
+"""RLT (RL Token) policy networks.
+
+Reference: "RL Token: Bootstrapping Online RL with Vision-Language-Action Models"
+(Xu et al., Physical Intelligence, 2026)
+
+Architecture:
+  - RLTokenEncoder: compresses VLA token embeddings into a single compact RL token
+  - RLTokenDecoder: reconstructs VLA embeddings from the RL token (Stage 1 training only)
+  - RLTActor: refines VLA reference action chunks conditioned on (z_rl, proprioception, ref_action)
+  - RLTCritic: Q(x, action_chunk) where x = (z_rl, proprioception)
+  - RLTPolicy: bundles RL-token modules + actor into a PreTrainedPolicy for inference
+"""
+
+from __future__ import annotations
+
+import math
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rlt.configuration_rlt import RLTConfig
+
+# ── Building blocks ──────────────────────────────────────────────────
+
+
+class MLP(nn.Module):
+    """Simple feedforward network with ReLU activations."""
+
+    def __init__(self, input_dim: int, hidden_dims: list[int], output_dim: int):
+        super().__init__()
+        layers: list[nn.Module] = []
+        prev = input_dim
+        for h in hidden_dims:
+            layers.append(nn.Linear(prev, h))
+            layers.append(nn.ReLU())
+            prev = h
+        layers.append(nn.Linear(prev, output_dim))
+        self.net = nn.Sequential(*layers)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.net(x)
+
+
+# ── RL Token Encoder ─────────────────────────────────────────────────
+
+
+class RLTokenEncoder(nn.Module):
+    """Compress VLA token embeddings into a single RL token via a small transformer.
+
+    Appends a learnable ``e_rl`` embedding to the VLA token sequence, processes
+    through transformer encoder layers, and returns the output at the ``e_rl``
+    position as the RL token ``z_rl``.
+
+    Paper Eq. 1: z_rl = g_phi([z_{1:M}, e_rl])_{M+1}
+    """
+
+    def __init__(
+        self,
+        input_dim: int,
+        rl_token_dim: int,
+        num_layers: int,
+        num_heads: int,
+        ff_dim: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.rl_token_dim = rl_token_dim
+
+        self.e_rl = nn.Parameter(torch.randn(1, 1, input_dim) * 0.02)
+
+        if input_dim != rl_token_dim:
+            self.input_proj = nn.Linear(input_dim, rl_token_dim)
+        else:
+            self.input_proj = nn.Identity()
+
+        encoder_layer = nn.TransformerEncoderLayer(
+            d_model=rl_token_dim,
+            nhead=num_heads,
+            dim_feedforward=ff_dim,
+            dropout=dropout,
+            batch_first=True,
+        )
+        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
+
+    def forward(self, z_vla: Tensor) -> Tensor:
+        """
+        Args:
+            z_vla: VLA token embeddings, shape ``(B, M, D)``.
+
+        Returns:
+            RL token ``z_rl``, shape ``(B, rl_token_dim)``.
+        """
+        batch_size = z_vla.shape[0]
+        e_rl = self.e_rl.expand(batch_size, -1, -1)
+        seq = torch.cat([z_vla, e_rl], dim=1)  # (B, M+1, D)
+        seq = self.input_proj(seq)
+        out = self.transformer(seq)
+        z_rl = out[:, -1, :]  # output at e_rl position
+        return z_rl
+
+
+# ── RL Token Decoder ─────────────────────────────────────────────────
+
+
+class RLTokenDecoder(nn.Module):
+    """Autoregressively reconstruct VLA embeddings from z_rl.
+
+    Used only during Stage 1 (offline RL-token training).
+
+    Paper Eq. 2: L_ro = E[sum_i || h(d([z_rl, z_bar_{1:i-1}]))_i - z_bar_i ||^2]
+    """
+
+    def __init__(
+        self,
+        rl_token_dim: int,
+        output_dim: int,
+        num_layers: int,
+        num_heads: int,
+        ff_dim: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.output_dim = output_dim
+
+        if rl_token_dim != output_dim:
+            self.rl_proj = nn.Linear(rl_token_dim, output_dim)
+        else:
+            self.rl_proj = nn.Identity()
+
+        decoder_layer = nn.TransformerDecoderLayer(
+            d_model=output_dim,
+            nhead=num_heads,
+            dim_feedforward=ff_dim,
+            dropout=dropout,
+            batch_first=True,
+        )
+        self.transformer = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
+        self.output_head = nn.Linear(output_dim, output_dim)
+
+    def forward(self, z_rl: Tensor, z_vla_stopped: Tensor) -> Tensor:
+        """
+        Args:
+            z_rl: RL token, shape ``(B, D_rl)``.
+            z_vla_stopped: Stop-gradient VLA embeddings, shape ``(B, M, D)``.
+
+        Returns:
+            Reconstructed embeddings, shape ``(B, M, D)``.
+        """
+        seq_len = z_vla_stopped.shape[1]
+        z_rl_proj = self.rl_proj(z_rl).unsqueeze(1)
+
+        target = torch.cat([z_rl_proj, z_vla_stopped[:, :-1, :]], dim=1)
+
+        causal_mask = nn.Transformer.generate_square_subsequent_mask(seq_len, device=z_rl.device)
+
+        decoded = self.transformer(
+            tgt=target,
+            memory=z_rl_proj,
+            tgt_mask=causal_mask,
+        )
+        return self.output_head(decoded)  # (B, M, D)
+
+
+# ── Actor ────────────────────────────────────────────────────────────
+
+
+class RLTActor(nn.Module):
+    """Lightweight actor that refines VLA reference action chunks.
+
+    Paper Eq. 4: pi_theta(a_{1:C} | x, a_tilde_{1:C}) = N(mu_theta(x, a_tilde), sigma^2 I)
+
+    The actor is conditioned on both the RL state and the VLA's proposed action
+    chunk, acting as a "VLA-guided action editor".
+    """
+
+    def __init__(self, state_dim: int, action_chunk_dim: int, hidden_dims: list[int], std: float = 0.1):
+        super().__init__()
+        input_dim = state_dim + action_chunk_dim
+        self.net = MLP(input_dim, hidden_dims, action_chunk_dim)
+        self.log_std = math.log(std)
+
+    def forward(self, state: Tensor, ref_action_chunk: Tensor) -> Tensor:
+        """Return the mean action chunk.
+
+        Args:
+            state: RL state ``x = (z_rl, proprioception)``, shape ``(B, state_dim)``.
+            ref_action_chunk: Flattened VLA reference chunk, shape ``(B, C*d)``.
+
+        Returns:
+            Refined action chunk (mean), shape ``(B, C*d)``.
+        """
+        x = torch.cat([state, ref_action_chunk], dim=-1)
+        return self.net(x)
+
+    def sample(self, state: Tensor, ref_action_chunk: Tensor) -> tuple[Tensor, Tensor]:
+        """Sample an action and return (action, log_prob)."""
+        mean = self.forward(state, ref_action_chunk)
+        std = math.exp(self.log_std)
+        noise = torch.randn_like(mean) * std
+        action = mean + noise
+        log_prob = -0.5 * (noise / std).pow(2).sum(dim=-1) - mean.shape[-1] * math.log(
+            std * math.sqrt(2 * math.pi)
+        )
+        return action, log_prob
+
+
+# ── Policy (inference bundle) ────────────────────────────────────────
+
+
+class RLTPolicy(PreTrainedPolicy):
+    """RLT policy — bundles the RL-token encoder and actor for inference.
+
+    The frozen VLA backbone is **not** part of this module; it is loaded
+    separately and its embeddings / reference actions are passed in via the
+    observation dict (populated by the actor process or a preprocessor).
+
+    During training, the :class:`RLTAlgorithm` holds the critic, target networks,
+    and optimizers. This class only contains what is needed for ``select_action``.
+    """
+
+    name = "rlt"
+    config_class = RLTConfig
+
+    def __init__(self, config: RLTConfig, dataset_stats=None):
+        super().__init__(config, dataset_stats)
+        action_dim = config.output_features["action"].shape[0]
+        action_chunk_dim = config.chunk_size * action_dim
+        prop_feature = config.input_features.get("observation.state", None)
+        proprioception_dim = prop_feature.shape[0] if prop_feature is not None else 0
+
+        state_dim = config.rl_token.rl_token_dim + proprioception_dim
+
+        # RL-token encoder (frozen after Stage 1)
+        self.rl_token_encoder = RLTokenEncoder(
+            input_dim=config.rl_token.input_dim,
+            rl_token_dim=config.rl_token.rl_token_dim,
+            num_layers=config.rl_token.num_encoder_layers,
+            num_heads=config.rl_token.num_heads,
+            ff_dim=config.rl_token.ff_dim,
+            dropout=config.rl_token.dropout,
+        )
+
+        # RL-token decoder (used only during Stage 1 training)
+        self.rl_token_decoder = RLTokenDecoder(
+            rl_token_dim=config.rl_token.rl_token_dim,
+            output_dim=config.rl_token.input_dim,
+            num_layers=config.rl_token.num_decoder_layers,
+            num_heads=config.rl_token.num_heads,
+            ff_dim=config.rl_token.ff_dim,
+            dropout=config.rl_token.dropout,
+        )
+
+        # Actor MLP
+        self.actor = RLTActor(
+            state_dim=state_dim,
+            action_chunk_dim=action_chunk_dim,
+            hidden_dims=config.actor.hidden_dims,
+            std=config.actor.std,
+        )
+
+        self._action_dim = action_dim
+        self._action_chunk_dim = action_chunk_dim
+        self._state_dim = state_dim
+        self._proprioception_dim = proprioception_dim
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a refined action chunk given an observation.
+
+        Expects the observation dict to contain:
+          - ``"observation.vla_embeddings"``: VLA internal token embeddings ``(M, D)``
+          - ``"observation.reference_action"``: VLA reference chunk ``(C*d,)``
+          - ``"observation.state"`` (optional): proprioceptive state ``(P,)``
+
+        Returns:
+            Action chunk tensor of shape ``(C*d,)``.
+        """
+        self.eval()
+
+        vla_emb = batch["observation.vla_embeddings"]
+        if vla_emb.dim() == 2:
+            vla_emb = vla_emb.unsqueeze(0)
+
+        z_rl = self.rl_token_encoder(vla_emb)  # (1, D_rl)
+
+        parts = [z_rl]
+        if "observation.state" in batch and self._proprioception_dim > 0:
+            prop = batch["observation.state"]
+            if prop.dim() == 1:
+                prop = prop.unsqueeze(0)
+            parts.append(prop)
+
+        state = torch.cat(parts, dim=-1)
+
+        ref = batch["observation.reference_action"]
+        if ref.dim() == 1:
+            ref = ref.unsqueeze(0)
+
+        action = self.actor(state, ref)
+        return action.squeeze(0)
+
+    def reset(self):
+        pass
@@ -23,7 +23,7 @@ Based on:

 from dataclasses import dataclass

-from lerobot.configs.types import RTCAttentionSchedule, RTCTrainingDelayDistribution
+from lerobot.configs.types import RTCAttentionSchedule


@dataclass
@@ -53,22 +53,3 @@ class RTCConfig:
            raise ValueError(f"max_guidance_weight must be positive, got {self.max_guidance_weight}")
        if self.debug_maxlen <= 0:
            raise ValueError(f"debug_maxlen must be positive, got {self.debug_maxlen}")
-
-
-@dataclass
-class RTCTrainingConfig:
-    """Configuration for training-time RTC action prefix conditioning."""
-
-    enabled: bool = False
-    min_delay: int = 0
-    max_delay: int = 0
-    delay_distribution: RTCTrainingDelayDistribution = RTCTrainingDelayDistribution.UNIFORM
-    exp_decay: float = 1.0
-
-    def __post_init__(self):
-        if self.min_delay < 0:
-            raise ValueError(f"min_delay must be >= 0, got {self.min_delay}")
-        if self.max_delay < self.min_delay:
-            raise ValueError(f"max_delay ({self.max_delay}) must be >= min_delay ({self.min_delay})")
-        if self.exp_decay <= 0:
-            raise ValueError(f"exp_decay must be positive, got {self.exp_decay}")
@@ -1,110 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import annotations
-
-import torch
-
-from lerobot.configs.types import RTCTrainingDelayDistribution
-from lerobot.policies.rtc.configuration_rtc import RTCTrainingConfig
-
-
-def sample_rtc_delay(cfg: RTCTrainingConfig, batch_size: int, device: torch.device) -> torch.Tensor:
-    if cfg.max_delay == cfg.min_delay:
-        return torch.full((batch_size,), cfg.min_delay, device=device, dtype=torch.long)
-
-    if cfg.delay_distribution == RTCTrainingDelayDistribution.UNIFORM:
-        return torch.randint(cfg.min_delay, cfg.max_delay + 1, (batch_size,), device=device, dtype=torch.long)
-
-    delay_values = torch.arange(cfg.min_delay, cfg.max_delay + 1, device=device, dtype=torch.long)
-    weights = torch.exp(-cfg.exp_decay * delay_values.to(dtype=torch.float32))
-    probs = weights / weights.sum()
-    samples = torch.multinomial(probs, batch_size, replacement=True)
-    return delay_values[samples]
-
-
-def apply_rtc_training_time(
-    time: torch.Tensor, delay: torch.Tensor, seq_len: int
-) -> tuple[torch.Tensor, torch.Tensor]:
-    device = time.device
-    delay = torch.clamp(delay, max=seq_len)
-    prefix_mask = torch.arange(seq_len, device=device)[None, :] < delay[:, None]
-    time_tokens = time[:, None].expand(-1, seq_len)
-    time_tokens = time_tokens.masked_fill(prefix_mask, 0.0)
-    postfix_mask = ~prefix_mask
-    return time_tokens, postfix_mask
-
-
-def masked_mean(
-    losses: torch.Tensor, mask: torch.Tensor | None, reduce_dims: tuple[int, ...], eps: float = 1e-8
-) -> torch.Tensor:
-    if mask is None:
-        return losses.mean(dim=reduce_dims)
-
-    mask = mask.to(dtype=losses.dtype)
-    while mask.dim() < losses.dim():
-        mask = mask.unsqueeze(-1)
-    masked = losses * mask
-    denom = mask.sum(dim=reduce_dims).clamp_min(eps)
-    return masked.sum(dim=reduce_dims) / denom
-
-
-def apply_training_time_rtc_inference(
-    x_t: torch.Tensor,
-    time: float,
-    inference_delay: int | None,
-    prev_chunk_left_over: torch.Tensor | None,
-    chunk_size: int,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Apply training-time RTC conditioning during inference.
-
-    Based on Algorithm 1 from "Training-Time Action Conditioning for Efficient Real-Time Chunking".
-
-    At each denoising step:
-    1. Replace prefix positions in x_t with ground truth from previous chunk
-    2. Create per-token timesteps with 1.0 for prefix positions
-
-    Args:
-        x_t: Current noisy actions (B, T, D)
-        time: Current flow matching timestep (scalar)
-        inference_delay: Number of prefix actions to condition on
-        prev_chunk_left_over: Previous chunk's leftover actions (B, T, D)
-        chunk_size: Total chunk size T
-
-    Returns:
-        x_t_conditioned: x_t with prefix replaced by previous actions
-        time_per_token: Per-token timesteps (B, T) with 1.0 for prefix
-    """
-    batch_size = x_t.shape[0]
-    device = x_t.device
-
-    if inference_delay is None or inference_delay <= 0 or prev_chunk_left_over is None:
-        time_scalar = torch.full((batch_size,), time, device=device, dtype=torch.float32)
-        return x_t, time_scalar
-
-    delay = min(inference_delay, chunk_size)
-    prefix_mask = torch.arange(chunk_size, device=device)[None, :] < delay
-
-    x_t_conditioned = torch.where(
-        prefix_mask[:, :, None].expand_as(x_t),
-        prev_chunk_left_over[:, :chunk_size, :],
-        x_t,
-    )
-
-    time_per_token = torch.full((batch_size, chunk_size), time, device=device, dtype=torch.float32)
-    time_per_token = time_per_token.masked_fill(prefix_mask, 1.0)
-
-    return x_t_conditioned, time_per_token
@@ -15,16 +15,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-import math
 from collections.abc import Callable
 from dataclasses import asdict
-from typing import Literal

-import einops
-import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution

@@ -52,20 +47,13 @@ class SACPolicy(

        # Determine action dimension and initialize all components
        continuous_action_dim = config.output_features[ACTION].shape[0]
-        self._init_encoders()
-        self._init_critics(continuous_action_dim)
+        self.encoder = SACObservationEncoder(config)
        self._init_actor(continuous_action_dim)
-        self._init_temperature()
+        self._init_discrete_critic()

    def get_optim_params(self) -> dict:
        optim_params = {
-            "actor": [
-                p
-                for n, p in self.actor.named_parameters()
-                if not n.startswith("encoder") or not self.shared_encoder
-            ],
-            "critic": self.critic_ensemble.parameters(),
-            "temperature": self.log_alpha,
+            "actor": [self.actor.parameters()],
        }
        if self.config.num_discrete_actions is not None:
            optim_params["discrete_critic"] = self.discrete_critic.parameters()
@@ -83,10 +71,9 @@ class SACPolicy(
    @torch.no_grad()
    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
        """Select action for inference/evaluation"""
-
        observations_features = None
-        if self.shared_encoder and self.actor.encoder.has_images:
-            observations_features = self.actor.encoder.get_cached_image_features(batch)
+        if self.encoder.has_images:
+            observations_features = self.encoder.get_cached_image_features(batch)

        actions, _, _ = self.actor(batch, observations_features)

@@ -97,371 +84,35 @@ class SACPolicy(

        return actions

-    def critic_forward(
-        self,
-        observations: dict[str, Tensor],
-        actions: Tensor,
-        use_target: bool = False,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        """Forward pass through a critic network ensemble
-
-        Args:
-            observations: Dictionary of observations
-            actions: Action tensor
-            use_target: If True, use target critics, otherwise use ensemble critics
-
-        Returns:
-            Tensor of Q-values from all critics
-        """
-
-        critics = self.critic_target if use_target else self.critic_ensemble
-        q_values = critics(observations, actions, observation_features)
-        return q_values
-
-    def discrete_critic_forward(
-        self, observations, use_target=False, observation_features=None
-    ) -> torch.Tensor:
-        """Forward pass through a discrete critic network
-
-        Args:
-            observations: Dictionary of observations
-            use_target: If True, use target critics, otherwise use ensemble critics
-            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
-
-        Returns:
-            Tensor of Q-values from the discrete critic network
-        """
-        discrete_critic = self.discrete_critic_target if use_target else self.discrete_critic
-        q_values = discrete_critic(observations, observation_features)
-        return q_values
-
    def forward(
        self,
        batch: dict[str, Tensor | dict[str, Tensor]],
-        model: Literal["actor", "critic", "temperature", "discrete_critic"] = "critic",
    ) -> dict[str, Tensor]:
-        """Compute the loss for the given model
+        """Actor forward pass."""
+        observations = batch.get("state", batch)
+        observation_features = batch.get("observation_feature") if isinstance(batch, dict) else None
+        actions, log_probs, means = self.actor(observations, observation_features)
+        return {"action": actions, "log_prob": log_probs, "action_mean": means}

-        Args:
-            batch: Dictionary containing:
-                - action: Action tensor
-                - reward: Reward tensor
-                - state: Observations tensor dict
-                - next_state: Next observations tensor dict
-                - done: Done mask tensor
-                - observation_feature: Optional pre-computed observation features
-                - next_observation_feature: Optional pre-computed next observation features
-            model: Which model to compute the loss for ("actor", "critic", "discrete_critic", or "temperature")
-
-        Returns:
-            The computed loss tensor
-        """
-        # Extract common components from batch
-        actions: Tensor = batch[ACTION]
-        observations: dict[str, Tensor] = batch["state"]
-        observation_features: Tensor = batch.get("observation_feature")
-
-        if model == "critic":
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-
-            loss_critic = self.compute_loss_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-            )
-
-            return {"loss_critic": loss_critic}
-
-        if model == "discrete_critic" and self.config.num_discrete_actions is not None:
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-            complementary_info = batch.get("complementary_info")
-            loss_discrete_critic = self.compute_loss_discrete_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-                complementary_info=complementary_info,
-            )
-            return {"loss_discrete_critic": loss_discrete_critic}
-        if model == "actor":
-            return {
-                "loss_actor": self.compute_loss_actor(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        if model == "temperature":
-            return {
-                "loss_temperature": self.compute_loss_temperature(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        raise ValueError(f"Unknown model type: {model}")
-
-    def update_target_networks(self):
-        """Update target networks with exponential moving average"""
-        for target_param, param in zip(
-            self.critic_target.parameters(),
-            self.critic_ensemble.parameters(),
-            strict=True,
-        ):
-            target_param.data.copy_(
-                param.data * self.config.critic_target_update_weight
-                + target_param.data * (1.0 - self.config.critic_target_update_weight)
-            )
-        if self.config.num_discrete_actions is not None:
-            for target_param, param in zip(
-                self.discrete_critic_target.parameters(),
-                self.discrete_critic.parameters(),
-                strict=True,
-            ):
-                target_param.data.copy_(
-                    param.data * self.config.critic_target_update_weight
-                    + target_param.data * (1.0 - self.config.critic_target_update_weight)
-                )
-
-    def update_temperature(self):
-        self.temperature = self.log_alpha.exp().item()
-
-    def compute_loss_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features: Tensor | None = None,
-        next_observation_features: Tensor | None = None,
-    ) -> Tensor:
-        with torch.no_grad():
-            next_action_preds, next_log_probs, _ = self.actor(next_observations, next_observation_features)
-
-            # 2- compute q targets
-            q_targets = self.critic_forward(
-                observations=next_observations,
-                actions=next_action_preds,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # subsample critics to prevent overfitting if use high UTD (update to date)
-            # TODO: Get indices before forward pass to avoid unnecessary computation
-            if self.config.num_subsample_critics is not None:
-                indices = torch.randperm(self.config.num_critics)
-                indices = indices[: self.config.num_subsample_critics]
-                q_targets = q_targets[indices]
-
-            # critics subsample size
-            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
-            if self.config.use_backup_entropy:
-                min_q = min_q - (self.temperature * next_log_probs)
-
-            td_target = rewards + (1 - done) * self.config.discount * min_q
-
-        # 3- compute predicted qs
-        if self.config.num_discrete_actions is not None:
-            # NOTE: We only want to keep the continuous action part
-            # In the buffer we have the full action space (continuous + discrete)
-            # We need to split them before concatenating them in the critic forward
-            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions,
-            use_target=False,
-            observation_features=observation_features,
-        )
-
-        # 4- Calculate loss
-        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
-        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
-        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
-        critics_loss = (
-            F.mse_loss(
-                input=q_preds,
-                target=td_target_duplicate,
-                reduction="none",
-            ).mean(dim=1)
-        ).sum()
-        return critics_loss
-
-    def compute_loss_discrete_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features=None,
-        next_observation_features=None,
-        complementary_info=None,
-    ):
-        # NOTE: We only want to keep the discrete action part
-        # In the buffer we have the full action space (continuous + discrete)
-        # We need to split them before concatenating them in the critic forward
-        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
-        actions_discrete = torch.round(actions_discrete)
-        actions_discrete = actions_discrete.long()
-
-        discrete_penalties: Tensor | None = None
-        if complementary_info is not None:
-            discrete_penalties: Tensor | None = complementary_info.get("discrete_penalty")
-
-        with torch.no_grad():
-            # For DQN, select actions using online network, evaluate with target network
-            next_discrete_qs = self.discrete_critic_forward(
-                next_observations, use_target=False, observation_features=next_observation_features
-            )
-            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
-
-            # Get target Q-values from target network
-            target_next_discrete_qs = self.discrete_critic_forward(
-                observations=next_observations,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # Use gather to select Q-values for best actions
-            target_next_discrete_q = torch.gather(
-                target_next_discrete_qs, dim=1, index=best_next_discrete_action
-            ).squeeze(-1)
-
-            # Compute target Q-value with Bellman equation
-            rewards_discrete = rewards
-            if discrete_penalties is not None:
-                rewards_discrete = rewards + discrete_penalties
-            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
-
-        # Get predicted Q-values for current observations
-        predicted_discrete_qs = self.discrete_critic_forward(
-            observations=observations, use_target=False, observation_features=observation_features
-        )
-
-        # Use gather to select Q-values for taken actions
-        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
-
-        # Compute MSE loss between predicted and target Q-values
-        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
-        return discrete_critic_loss
-
-    def compute_loss_temperature(self, observations, observation_features: Tensor | None = None) -> Tensor:
-        """Compute the temperature loss"""
-        # calculate temperature loss
-        with torch.no_grad():
-            _, log_probs, _ = self.actor(observations, observation_features)
-        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
-        return temperature_loss
-
-    def compute_loss_actor(
-        self,
-        observations,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        actions_pi, log_probs, _ = self.actor(observations, observation_features)
-
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions_pi,
-            use_target=False,
-            observation_features=observation_features,
-        )
-        min_q_preds = q_preds.min(dim=0)[0]
-
-        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
-        return actor_loss
-
-    def _init_encoders(self):
-        """Initialize shared or separate encoders for actor and critic."""
-        self.shared_encoder = self.config.shared_encoder
-        self.encoder_critic = SACObservationEncoder(self.config)
-        self.encoder_actor = (
-            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
-        )
-
-    def _init_critics(self, continuous_action_dim):
-        """Build critic ensemble, targets, and optional discrete critic."""
-        heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
-        target_heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_target = CriticEnsemble(encoder=self.encoder_critic, ensemble=target_heads)
-        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
-
-        if self.config.use_torch_compile:
-            self.critic_ensemble = torch.compile(self.critic_ensemble)
-            self.critic_target = torch.compile(self.critic_target)
-
-        if self.config.num_discrete_actions is not None:
-            self._init_discrete_critics()
-
-    def _init_discrete_critics(self):
-        """Build discrete discrete critic ensemble and target networks."""
-        self.discrete_critic = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-        self.discrete_critic_target = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-
-        # TODO: (maractingi, azouitine) Compile the discrete critic
-        self.discrete_critic_target.load_state_dict(self.discrete_critic.state_dict())
-
-    def _init_actor(self, continuous_action_dim):
-        """Initialize policy actor network and default target entropy."""
-        # NOTE: The actor select only the continuous action part
+    def _init_actor(self, continuous_action_dim: int) -> None:
        self.actor = Policy(
-            encoder=self.encoder_actor,
-            network=MLP(input_dim=self.encoder_actor.output_dim, **asdict(self.config.actor_network_kwargs)),
+            encoder=self.encoder,
+            network=MLP(input_dim=self.encoder.output_dim, **asdict(self.config.actor_network_kwargs)),
            action_dim=continuous_action_dim,
-            encoder_is_shared=self.shared_encoder,
+            encoder_is_shared=False,
            **asdict(self.config.policy_kwargs),
        )

-        self.target_entropy = self.config.target_entropy
-        if self.target_entropy is None:
-            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."""
-        temp_init = self.config.temperature_init
-        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
-        self.temperature = self.log_alpha.exp().item()
+    def _init_discrete_critic(self) -> None:
+        if self.config.num_discrete_actions is None:
+            self.discrete_critic = None
+            return
+        self.discrete_critic = DiscreteCritic(
+            encoder=self.encoder,
+            input_dim=self.encoder.output_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )


 class SACObservationEncoder(nn.Module):
@@ -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
        """,
@@ -20,7 +20,7 @@ from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import (
    CosineDecayWithWarmupSchedulerConfig,
 )
-from lerobot.policies.rtc.configuration_rtc import RTCConfig, RTCTrainingConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.utils.constants import OBS_IMAGES


@@ -103,9 +103,8 @@ class SmolVLAConfig(PreTrainedConfig):
    min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding
    max_period: float = 4.0

-    # Real-Time Chunking (RTC) configurations
+    # Real-Time Chunking (RTC) configuration
    rtc_config: RTCConfig | None = None
-    rtc_training_config: RTCTrainingConfig | None = None

    def __post_init__(self):
        super().__post_init__()
@@ -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
 ```
@@ -63,12 +63,6 @@ from typing_extensions import Unpack

 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
-from lerobot.policies.rtc.training_time import (
-    apply_rtc_training_time,
-    apply_training_time_rtc_inference,
-    masked_mean,
-    sample_rtc_delay,
-)
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.smolvla.smolvlm_with_expert import SmolVLMWithExpertModel
 from lerobot.policies.utils import (
@@ -91,8 +85,8 @@ def create_sinusoidal_pos_embedding(
    if dimension % 2 != 0:
        raise ValueError(f"dimension ({dimension}) must be divisible by 2")

-    if time.ndim not in (1, 2):
-        raise ValueError("The time tensor is expected to be of shape `(batch_size,)` or `(batch_size, T)`.")
+    if time.ndim != 1:
+        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")

    dtype = get_safe_dtype(torch.float64, device.type)
    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
@@ -100,14 +94,9 @@ def create_sinusoidal_pos_embedding(

    # Compute the outer product
    scaling_factor = 1.0 / period * 2 * math.pi
-    if time.ndim == 1:
-        sin_input = scaling_factor[None, :] * time[:, None]
-        return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-
-    time_flat = time.reshape(-1)
-    sin_input = scaling_factor[None, :] * time_flat[:, None]
+    sin_input = scaling_factor[None, :] * time[:, None]
    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-    return pos_emb.reshape(*time.shape, dimension)
+    return pos_emb


 def make_att_2d_masks(pad_masks, att_masks):
@@ -386,39 +375,28 @@ class SmolVLAPolicy(PreTrainedPolicy):
        lang_tokens = batch[f"{OBS_LANGUAGE_TOKENS}"]
        lang_masks = batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
        actions = self.prepare_action(batch)
-        postfix_mask = None
-        rtc_cfg = self.config.rtc_training_config
-        if rtc_cfg is not None and rtc_cfg.enabled and self.training:
-            batch_size = actions.shape[0]
-            if time is None:
-                time = self.model.sample_time(batch_size, actions.device)
-            if noise is None:
-                noise = self.model.sample_noise(actions.shape, actions.device)
-            delay = sample_rtc_delay(rtc_cfg, batch_size, actions.device)
-            time, postfix_mask = apply_rtc_training_time(time, delay, actions.shape[1])
        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()
-            postfix_mask = in_episode_bound if postfix_mask is None else (postfix_mask & in_episode_bound)
+            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
-            per_sample_loss = masked_mean(losses, postfix_mask, reduce_dims=(1, 2))
+            per_sample_loss = losses.mean(dim=(1, 2))
            loss_dict["loss"] = per_sample_loss.mean().item()
            return per_sample_loss, loss_dict
        else:
            # Default: return scalar mean loss
-            loss = masked_mean(losses, postfix_mask, reduce_dims=(0, 1, 2))
+            loss = losses.mean()
            loss_dict["loss"] = loss.item()
            return loss, loss_dict

@@ -618,9 +596,6 @@ class VLAFlowMatching(nn.Module):
    def _rtc_enabled(self):
        return self.config.rtc_config is not None and self.config.rtc_config.enabled

-    def _training_time_rtc_inference_enabled(self):
-        return self.config.rtc_training_config is not None and self.config.rtc_training_config.enabled
-
    def set_requires_grad(self):
        for params in self.state_proj.parameters():
            params.requires_grad = self.config.train_state_proj
@@ -756,10 +731,7 @@ class VLAFlowMatching(nn.Module):
        )
        time_emb = time_emb.type(dtype=dtype)

-        if time_emb.dim() == 2:
-            time_emb = time_emb[:, None, :].expand_as(action_emb)
-        elif time_emb.shape[:2] != action_emb.shape[:2]:
-            raise ValueError(f"Expected time_emb shape {action_emb.shape[:2]}, got {time_emb.shape[:2]}")
+        time_emb = time_emb[:, None, :].expand_as(action_emb)
        action_time_emb = torch.cat([action_emb, time_emb], dim=2)

        action_time_emb = self.action_time_mlp_in(action_time_emb)
@@ -791,12 +763,7 @@ class VLAFlowMatching(nn.Module):
        if time is None:
            time = self.sample_time(actions.shape[0], actions.device)

-        if time.ndim == 1:
-            time_expanded = time[:, None, None]
-        elif time.ndim == 2:
-            time_expanded = time[:, :, None]
-        else:
-            raise ValueError(f"Expected time shape (B,) or (B, T), got {time.shape}")
+        time_expanded = time[:, None, None]
        x_t = time_expanded * noise + (1 - time_expanded) * actions
        u_t = noise - actions
        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
@@ -859,35 +826,23 @@ class VLAFlowMatching(nn.Module):
        num_steps = self.config.num_steps
        dt = -1.0 / num_steps

-        inference_delay = kwargs.get("inference_delay")
-        prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
-        execution_horizon = kwargs.get("execution_horizon")
-        use_training_time_rtc = self._training_time_rtc_inference_enabled()
-
        x_t = noise
        for step in range(num_steps):
            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)

-            if use_training_time_rtc:
-                x_t_cond, time_tensor = apply_training_time_rtc_inference(
-                    x_t, time, inference_delay, prev_chunk_left_over, self.config.chunk_size
-                )
-                v_t = self.denoise_step(
-                    x_t=x_t_cond,
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
+                    x_t=input_x_t,
                    prefix_pad_masks=prefix_pad_masks,
                    past_key_values=past_key_values,
-                    timestep=time_tensor,
+                    timestep=current_timestep,
                )
-            elif self._rtc_enabled():
-                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)

-                def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
-                    return self.denoise_step(
-                        x_t=input_x_t,
-                        prefix_pad_masks=prefix_pad_masks,
-                        past_key_values=past_key_values,
-                        timestep=current_timestep,
-                    )
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")

                v_t = self.rtc_processor.denoise_step(
                    x_t=x_t,
@@ -898,13 +853,7 @@ class VLAFlowMatching(nn.Module):
                    execution_horizon=execution_horizon,
                )
            else:
-                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
-                v_t = self.denoise_step(
-                    x_t=x_t,
-                    prefix_pad_masks=prefix_pad_masks,
-                    past_key_values=past_key_values,
-                    timestep=time_tensor,
-                )
+                v_t = denoise_step_partial_call(x_t)

            x_t = x_t + dt * v_t

@@ -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.
@@ -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.
@@ -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",
@@ -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(
@@ -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

@@ -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(
@@ -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]:
        """
@@ -131,6 +131,15 @@ class _NormalizationMixin:
        if self.dtype is None:
            self.dtype = torch.float32
        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        self._reshape_visual_stats()
+
+    def _reshape_visual_stats(self) -> None:
+        """Reshape visual stats from ``[C]`` to ``[C, 1, 1]`` for image broadcasting."""
+        for key, feature in self.features.items():
+            if feature.type == FeatureType.VISUAL and key in self._tensor_stats:
+                for stat_name, stat_tensor in self._tensor_stats[key].items():
+                    if isinstance(stat_tensor, Tensor) and stat_tensor.ndim == 1:
+                        self._tensor_stats[key][stat_name] = stat_tensor.reshape(-1, 1, 1)

    def to(
        self, device: torch.device | str | None = None, dtype: torch.dtype | None = None
@@ -149,6 +158,7 @@ class _NormalizationMixin:
        if dtype is not None:
            self.dtype = dtype
        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        self._reshape_visual_stats()
        return self

    def state_dict(self) -> dict[str, Tensor]:
@@ -198,6 +208,7 @@ class _NormalizationMixin:
            # Don't load from state_dict, keep the explicitly provided stats
            # But ensure _tensor_stats is properly initialized
            self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)  # type: ignore[assignment]
+            self._reshape_visual_stats()
            return

        # Normal behavior: load stats from state_dict
@@ -208,6 +219,7 @@ class _NormalizationMixin:
            self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
                dtype=torch.float32, device=self.device
            )
+        self._reshape_visual_stats()

        # Reconstruct the original stats dict from tensor stats for compatibility with to() method
        # and other functions that rely on self.stats
@@ -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
@@ -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:
@@ -0,0 +1,13 @@
+# 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.
@@ -61,7 +61,7 @@ from lerobot.cameras import opencv  # noqa: F401
 from lerobot.configs import parser
 from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.policies.factory import make_policy
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.processor import TransitionKey
 from lerobot.rl.process import ProcessSignalHandler
 from lerobot.rl.queue import get_last_item_from_queue
@@ -248,16 +248,16 @@ def act_with_policy(

    logging.info("make_policy")

-    ### Instantiate the policy in both the actor and learner processes
-    ### To avoid sending a SACPolicy object through the port, we create a policy instance
-    ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
-    policy: SACPolicy = make_policy(
+    policy = make_policy(
        cfg=cfg.policy,
        env_cfg=cfg.env,
    )
    policy = policy.eval()
    assert isinstance(policy, nn.Module)

+    # TODO: Re-enable processor pipeline once refactoring is validated against main
+    # preprocessor, postprocessor = None, None
+
    obs, info = online_env.reset()
    env_processor.reset()
    action_processor.reset()
@@ -288,7 +288,6 @@ def act_with_policy(

        # Time policy inference and check if it meets FPS requirement
        with policy_timer:
-            # Extract observation from transition for policy
            action = policy.select_action(batch=observation)
        policy_fps = policy_timer.fps_last

@@ -649,12 +648,12 @@ def interactions_stream(
 #  Policy functions


-def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
+def update_policy_parameters(policy: PreTrainedPolicy, parameters_queue: Queue, device):
+    """Load the latest policy weights from the learner."""
    bytes_state_dict = get_last_item_from_queue(parameters_queue, block=False)
    if bytes_state_dict is not None:
        logging.info("[ACTOR] Load new parameters from Learner.")
        state_dicts = bytes_to_state_dict(bytes_state_dict)
-
        # TODO: check encoder parameter synchronization possible issues:
        # 1. When shared_encoder=True, we're loading stale encoder params from actor's state_dict
        #    instead of the updated encoder params from critic (which is optimized separately)
@@ -664,18 +663,9 @@ def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device)
        # - Send critic's encoder state when shared_encoder=True
        # - Skip encoder params entirely when freeze_vision_encoder=True
        # - Ensure discrete_critic gets correct encoder state (currently uses encoder_critic)
-
        # Load actor state dict
-        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
-        policy.actor.load_state_dict(actor_state_dict)
-
-        # Load discrete critic if present
-        if hasattr(policy, "discrete_critic") and "discrete_critic" in state_dicts:
-            discrete_critic_state_dict = move_state_dict_to_device(
-                state_dicts["discrete_critic"], device=device
-            )
-            policy.discrete_critic.load_state_dict(discrete_critic_state_dict)
-            logging.info("[ACTOR] Loaded discrete critic parameters from Learner.")
+        state_dicts = move_state_dict_to_device(state_dicts, device=device)
+        policy.load_state_dict(state_dicts)


 #  Utilities functions
@@ -0,0 +1,70 @@
+# 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 __future__ import annotations
+
+import torch
+
+from lerobot.rl.algorithms.base import (
+    RLAlgorithm,
+    RLAlgorithmConfig,
+    TrainingStats,
+)
+from lerobot.rl.algorithms.rlt import RLTAlgorithm, RLTAlgorithmConfig
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
+
+
+def make_algorithm(
+    policy: torch.nn.Module,
+    policy_cfg,
+    *,
+    algorithm_name: str,
+) -> RLAlgorithm:
+    """Construct an :class:`RLAlgorithm` from a policy and its config.
+
+    Algorithm selection is explicit via ``algorithm_name`` (from
+    ``cfg.algorithm``).
+
+    This is fully registry-driven — adding a new algorithm only requires
+    registering an ``RLAlgorithmConfig`` subclass; no changes here.
+
+    The returned algorithm has **no optimizers** yet.  On the learner side,
+    call ``algorithm.make_optimizers()`` afterwards to create them.  On the
+    actor side (inference-only), leave them empty.
+
+    Args:
+        policy: Instantiated policy (e.g. ``SACPolicy``).
+        policy_cfg: The policy's ``PreTrainedConfig`` with the hyper-parameters
+            expected by the algorithm config's ``from_policy_config`` class-method.
+        algorithm_name: Algorithm registry key to instantiate.
+    """
+    known = RLAlgorithmConfig.get_known_choices()
+    if algorithm_name not in known:
+        raise ValueError(f"No RLAlgorithmConfig registered for '{algorithm_name}'. Known: {list(known)}")
+
+    config_cls = RLAlgorithmConfig.get_choice_class(algorithm_name)
+    algo_config = config_cls.from_policy_config(policy_cfg)
+    return algo_config.build_algorithm(policy)
+
+
+__all__ = [
+    "RLAlgorithm",
+    "RLAlgorithmConfig",
+    "TrainingStats",
+    "SACAlgorithm",
+    "SACAlgorithmConfig",
+    "RLTAlgorithm",
+    "RLTAlgorithmConfig",
+    "make_algorithm",
+]
@@ -0,0 +1,183 @@
+# 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.
+"""Base classes for RL algorithms.
+
+Defines the abstract interface that every algorithm must implement, a registry
+for algorithm configs, and a dataclass for training statistics.
+"""
+
+from __future__ import annotations
+
+import abc
+from collections.abc import Iterator
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import draccus
+import torch
+from torch import Tensor
+from torch.optim import Optimizer
+
+if TYPE_CHECKING:
+    from lerobot.rl.data_sources.data_mixer import DataMixer
+
+BatchType = dict[str, Any]
+
+
+@dataclass
+class TrainingStats:
+    """Returned by ``algorithm.update()`` for logging and checkpointing."""
+
+    # Generic containers for all algorithms
+    losses: dict[str, float] = field(default_factory=dict)
+    grad_norms: dict[str, float] = field(default_factory=dict)
+    extra: dict[str, float] = field(default_factory=dict)
+
+    def to_log_dict(self) -> dict[str, float]:
+        """Flatten all stats into a single dict for logging."""
+
+        d: dict[str, float] = {}
+        for name, val in self.losses.items():
+            d[name] = val
+        for name, val in self.grad_norms.items():
+            d[f"{name}_grad_norm"] = val
+        for name, val in self.extra.items():
+            d[name] = val
+        return d
+
+
+@dataclass
+class RLAlgorithmConfig(draccus.ChoiceRegistry):
+    """Registry for algorithm configs."""
+
+    def build_algorithm(self, policy: torch.nn.Module) -> RLAlgorithm:
+        """Construct the :class:`RLAlgorithm` for this config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{type(self).__name__} must implement build_algorithm()")
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg: Any) -> RLAlgorithmConfig:
+        """Build an algorithm config from a policy config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{cls.__name__} must implement from_policy_config()")
+
+
+class RLAlgorithm(abc.ABC):
+    """Base for all RL algorithms."""
+
+    @abc.abstractmethod
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One complete training step.
+
+        The algorithm calls ``next(batch_iterator)`` as many times as it
+        needs (e.g. ``utd_ratio`` times for SAC) to obtain fresh batches.
+        The iterator is owned by the trainer; the algorithm just consumes
+        from it.
+        """
+        ...
+
+    def supports_offline_phase(self) -> bool:
+        """Whether this algorithm has an offline pretraining phase.
+
+        Algorithms like RLT (RL-token training) or ConRFT (Cal-QL pretraining)
+        return ``True`` here. The learner checks this before the main online
+        loop and routes to :meth:`offline_update` accordingly.
+        """
+        return False
+
+    def offline_update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One offline training step (called before any online collection).
+
+        Only called when :meth:`supports_offline_phase` returns ``True``.
+        Uses the same iterator protocol as :meth:`update`.
+        """
+        raise NotImplementedError(
+            f"{type(self).__name__} does not implement offline_update(). "
+            "Either override this method or return False from supports_offline_phase()."
+        )
+
+    def transition_to_online(self) -> None:  # noqa: B027
+        """Called once when switching from offline to online phase.
+
+        Use this to freeze modules trained offline, rebuild optimizers for the
+        online phase, reset step counters, etc.
+
+        Default is a no-op; subclasses override when they have an offline phase.
+        """
+
+    def configure_data_iterator(
+        self,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ) -> Iterator[BatchType]:
+        """Create the data iterator this algorithm needs.
+
+        The default implementation uses the standard ``data_mixer.get_iterator()``.
+        Algorithms that need specialised sampling should override this method.
+        """
+        return data_mixer.get_iterator(
+            batch_size=batch_size,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+
+    def make_optimizers(self) -> dict[str, Optimizer]:
+        """Create, store, and return the optimizers needed for training.
+
+        Called on the **learner** side after construction.  Subclasses must
+        override this with algorithm-specific optimizer setup.
+        """
+        return {}
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        """Return optimizers for checkpointing / external scheduling."""
+        return {}
+
+    @property
+    def optimization_step(self) -> int:
+        """Current learner optimization step.
+
+        Part of the stable contract for checkpoint/resume. Algorithms can
+        either use this default storage or override for custom behavior.
+        """
+        return getattr(self, "_optimization_step", 0)
+
+    @optimization_step.setter
+    def optimization_step(self, value: int) -> None:
+        self._optimization_step = int(value)
+
+    def get_weights(self) -> dict[str, Any]:
+        """Policy state-dict to push to actors."""
+        return {}
+
+    @abc.abstractmethod
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load policy state-dict received from the learner (inverse of ``get_weights``)."""
+
+    @torch.no_grad()
+    def get_observation_features(
+        self, observations: Tensor, next_observations: Tensor
+    ) -> tuple[Tensor | None, Tensor | None]:
+        """Pre-compute observation features (e.g. frozen encoder cache).
+
+        Returns ``(None, None)`` when caching is not applicable.
+        """
+        return None, None
@@ -0,0 +1,18 @@
+# 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 lerobot.rl.algorithms.rlt.configuration_rlt import RLTAlgorithmConfig
+from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
+
+__all__ = ["RLTAlgorithm", "RLTAlgorithmConfig"]
@@ -0,0 +1,83 @@
+# 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.
+"""RLT algorithm configuration."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+import torch
+
+from lerobot.rl.algorithms.base import RLAlgorithmConfig
+
+if TYPE_CHECKING:
+    from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
+
+
+@RLAlgorithmConfig.register_subclass("rlt")
+@dataclass
+class RLTAlgorithmConfig(RLAlgorithmConfig):
+    """RLT-specific hyper-parameters that control the update loop."""
+
+    # ── Action chunks ──
+    chunk_size: int = 10
+    chunk_stride: int = 2
+
+    # ── Update cadence ──
+    utd_ratio: int = 5
+    policy_update_freq: int = 2
+    clip_grad_norm: float = 10.0
+
+    # ── Learning rates ──
+    actor_lr: float = 3e-4
+    critic_lr: float = 3e-4
+    rl_token_lr: float = 1e-4
+
+    # ── TD learning ──
+    discount: float = 0.99
+    tau: float = 0.005
+    num_critics: int = 2
+
+    # ── Policy constraint (paper Eq. 5) ──
+    bc_reg_coeff: float = 0.1
+    ref_dropout: float = 0.5
+
+    # ── Offline RL-token training ──
+    vla_finetune_weight: float = 0.0
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg) -> RLTAlgorithmConfig:
+        """Build from an existing ``RLTConfig`` (cfg.policy)."""
+        return cls(
+            chunk_size=policy_cfg.chunk_size,
+            chunk_stride=policy_cfg.chunk_stride,
+            utd_ratio=policy_cfg.utd_ratio,
+            policy_update_freq=policy_cfg.policy_update_freq,
+            clip_grad_norm=policy_cfg.clip_grad_norm,
+            actor_lr=policy_cfg.actor_lr,
+            critic_lr=policy_cfg.critic_lr,
+            rl_token_lr=policy_cfg.rl_token_lr,
+            discount=policy_cfg.discount,
+            tau=policy_cfg.tau,
+            num_critics=policy_cfg.num_critics,
+            bc_reg_coeff=policy_cfg.bc_reg_coeff,
+            ref_dropout=policy_cfg.ref_dropout,
+            vla_finetune_weight=policy_cfg.vla_finetune_weight,
+        )
+
+    def build_algorithm(self, policy: torch.nn.Module) -> RLTAlgorithm:
+        from lerobot.rl.algorithms.rlt.rlt_algorithm import RLTAlgorithm
+
+        return RLTAlgorithm(policy=policy, config=self)
@@ -0,0 +1,319 @@
+# 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.
+"""RLT (RL Token) algorithm.
+
+Implements the two-stage training from "RL Token: Bootstrapping Online RL
+with Vision-Language-Action Models" (Xu et al., Physical Intelligence, 2026).
+
+Stage 1 (offline): Train RL-token encoder/decoder via reconstruction loss.
+Stage 2 (online):  Train actor-critic with chunked TD, BC regularization,
+                   reference-action pass-through, and reference-action dropout.
+"""
+
+from __future__ import annotations
+
+import copy
+from collections.abc import Iterator
+from typing import Any
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+from torch.optim import Optimizer
+
+from lerobot.policies.rlt.modeling_rlt import MLP, RLTPolicy
+from lerobot.policies.utils import get_device_from_parameters
+from lerobot.rl.algorithms.base import (
+    BatchType,
+    RLAlgorithm,
+    TrainingStats,
+)
+from lerobot.rl.algorithms.rlt.configuration_rlt import RLTAlgorithmConfig
+from lerobot.utils.constants import ACTION
+
+
+class RLTCritic(nn.Module):
+    """Q-function over (state, action_chunk) pairs.
+
+    Paper Eq. 3: Q_psi(x, a_{1:C})
+
+    Training-only component — lives on the algorithm side, not in the policy.
+    """
+
+    def __init__(self, state_dim: int, action_chunk_dim: int, hidden_dims: list[int]):
+        super().__init__()
+        self.net = MLP(state_dim + action_chunk_dim, hidden_dims, output_dim=1)
+
+    def forward(self, state: Tensor, action_chunk: Tensor) -> Tensor:
+        x = torch.cat([state, action_chunk], dim=-1)
+        return self.net(x)
+
+
+class RLTAlgorithm(RLAlgorithm):
+    """RL Token: lightweight actor-critic on frozen VLA features.
+
+    Owns the ``RLTPolicy`` (RL-token encoder/decoder + actor), a critic
+    ensemble, and target networks.  All VLA-specific logic (embedding
+    extraction, reference actions) lives in ``_prepare_forward_batch``.
+    """
+
+    def __init__(self, policy: RLTPolicy, config: RLTAlgorithmConfig):
+        self.policy = policy
+        self.config = config
+        self.optimizers: dict[str, Optimizer] = {}
+        self._optimization_step: int = 0
+        self._device = get_device_from_parameters(self.policy)
+        self._is_online = False
+
+        self._init_critics()
+        self._move_to_device()
+
+    # ── Initialization ───────────────────────────────────────────────
+
+    def _init_critics(self) -> None:
+        state_dim = self.policy._state_dim
+        action_chunk_dim = self.policy._action_chunk_dim
+        hidden_dims = self.policy.config.critic.hidden_dims
+
+        self.critics = torch.nn.ModuleList(
+            [RLTCritic(state_dim, action_chunk_dim, hidden_dims) for _ in range(self.config.num_critics)]
+        )
+        self.critic_targets = torch.nn.ModuleList([copy.deepcopy(c) for c in self.critics])
+        for ct in self.critic_targets:
+            ct.requires_grad_(False)
+
+    def _move_to_device(self) -> None:
+        self.critics.to(self._device)
+        self.critic_targets.to(self._device)
+
+    # ── Offline phase (Stage 1): RL-token training ───────────────────
+
+    def supports_offline_phase(self) -> bool:
+        return True
+
+    def offline_update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """Train RL-token encoder/decoder on demonstration data.
+
+        Paper Eq. 2: L_ro = E[ sum_i || h(d([z_rl, z_bar_{1:i-1}]))_i - z_bar_i ||^2 ]
+        """
+        batch = next(batch_iterator)
+
+        vla_embeddings = batch["state"]["observation.vla_embeddings"].to(self._device)
+        z_vla = vla_embeddings.detach()  # stop-gradient on VLA embeddings
+
+        z_rl = self.policy.rl_token_encoder(z_vla)
+        z_reconstructed = self.policy.rl_token_decoder(z_rl, z_vla)
+
+        loss_ro = F.mse_loss(z_reconstructed, z_vla)
+
+        self.optimizers["rl_token"].zero_grad()
+        loss_ro.backward()
+        torch.nn.utils.clip_grad_norm_(
+            list(self.policy.rl_token_encoder.parameters()) + list(self.policy.rl_token_decoder.parameters()),
+            max_norm=self.config.clip_grad_norm,
+        )
+        self.optimizers["rl_token"].step()
+
+        self._optimization_step += 1
+        return TrainingStats(losses={"loss_rl_token": loss_ro.item()})
+
+    def transition_to_online(self) -> None:
+        """Freeze RL-token modules; rebuild optimizers for actor-critic only."""
+        self.policy.rl_token_encoder.requires_grad_(False)
+        self.policy.rl_token_decoder.requires_grad_(False)
+        self._is_online = True
+
+        self.optimizers = {
+            "actor": torch.optim.Adam(self.policy.actor.parameters(), lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critics.parameters(), lr=self.config.critic_lr),
+        }
+        self._optimization_step = 0
+
+    # ── Online phase (Stage 2): Actor-Critic ─────────────────────────
+
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One full RLT update step with UTD critic warm-up.
+
+        Pulls ``utd_ratio`` batches. First ``utd_ratio - 1`` are critic-only;
+        the last batch also updates the actor (every ``policy_update_freq`` steps).
+        """
+        for _ in range(self.config.utd_ratio - 1):
+            batch = next(batch_iterator)
+            fb = self._prepare_forward_batch(batch)
+            self._critic_step(fb)
+            self._update_target_networks()
+
+        batch = next(batch_iterator)
+        fb = self._prepare_forward_batch(batch)
+        critic_loss = self._critic_step(fb)
+
+        stats = TrainingStats(losses={"loss_critic": critic_loss})
+
+        if self._optimization_step % self.config.policy_update_freq == 0:
+            actor_loss, bc_loss, q_val = self._actor_step(fb)
+            stats.losses["loss_actor"] = actor_loss
+            stats.extra["bc_loss"] = bc_loss
+            stats.extra["q_value_mean"] = q_val
+
+        self._update_target_networks()
+        self._optimization_step += 1
+        return stats
+
+    def _prepare_forward_batch(self, batch: BatchType) -> dict[str, Any]:
+        """Convert a replay batch into algorithm-ready tensors.
+
+        Extracts RL-token from VLA embeddings, builds RL state, reads
+        reference action from complementary_info.
+        """
+        obs = batch["state"]
+        next_obs = batch["next_state"]
+        device = self._device
+
+        vla_emb = obs["observation.vla_embeddings"].to(device)
+        next_vla_emb = next_obs["observation.vla_embeddings"].to(device)
+
+        with torch.no_grad():
+            z_rl = self.policy.rl_token_encoder(vla_emb)
+            z_rl_next = self.policy.rl_token_encoder(next_vla_emb)
+
+        parts = [z_rl]
+        next_parts = [z_rl_next]
+        if "observation.state" in obs and self.policy._proprioception_dim > 0:
+            prop = obs["observation.state"].to(device)
+            next_prop = next_obs["observation.state"].to(device)
+            parts.append(prop)
+            next_parts.append(next_prop)
+
+        state = torch.cat(parts, dim=-1)
+        next_state = torch.cat(next_parts, dim=-1)
+
+        action = batch[ACTION].to(device)
+        reward = batch["reward"].to(device)
+        done = batch["done"].to(device)
+
+        ref_action = None
+        comp_info = batch.get("complementary_info")
+        if comp_info is not None and "reference_action" in comp_info:
+            ref_action = comp_info["reference_action"].to(device)
+
+        return {
+            "state": state,
+            "next_state": next_state,
+            "action": action,
+            "reward": reward,
+            "done": done,
+            "reference_action": ref_action,
+        }
+
+    def _critic_step(self, fb: dict[str, Any]) -> float:
+        """Paper Eq. 3: chunked TD with clipped double-Q target."""
+        state = fb["state"]
+        next_state = fb["next_state"]
+        action = fb["action"]
+        reward = fb["reward"]
+        done = fb["done"]
+
+        with torch.no_grad():
+            ref = fb.get("reference_action")
+            if ref is None:
+                ref = torch.zeros_like(action)
+            next_action = self.policy.actor(next_state, ref)
+
+            target_qs = [ct(next_state, next_action) for ct in self.critic_targets]
+            min_target_q = torch.min(torch.cat(target_qs, dim=-1), dim=-1, keepdim=True).values
+
+            discount_chunk = self.config.discount**self.config.chunk_size
+            td_target = reward.unsqueeze(-1) + (1 - done.unsqueeze(-1)) * discount_chunk * min_target_q
+
+        q_preds = [c(state, action) for c in self.critics]
+        loss = sum(F.mse_loss(q, td_target) for q in q_preds)
+
+        self.optimizers["critic"].zero_grad()
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_(self.critics.parameters(), max_norm=self.config.clip_grad_norm)
+        self.optimizers["critic"].step()
+        return loss.item()
+
+    def _actor_step(self, fb: dict[str, Any]) -> tuple[float, float, float]:
+        """Paper Eq. 5: maximize Q while staying near VLA reference.
+
+        L_pi(theta) = E[ -Q(x, a) + beta * ||a - a_tilde||^2 ]
+        With reference-action dropout applied to the actor's ref input.
+        """
+        state = fb["state"]
+        ref = fb.get("reference_action")
+        if ref is None:
+            ref = torch.zeros(state.shape[0], self.policy._action_chunk_dim, device=self._device)
+
+        # Reference-action dropout (paper Section IV-B)
+        mask = (torch.rand(ref.shape[0], 1, device=self._device) > self.config.ref_dropout).float()
+        ref_input = ref * mask
+
+        action = self.policy.actor(state, ref_input)
+
+        q_value = self.critics[0](state, action)
+
+        bc_loss = F.mse_loss(action, ref)
+
+        loss = -q_value.mean() + self.config.bc_reg_coeff * bc_loss
+
+        self.optimizers["actor"].zero_grad()
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_(self.policy.actor.parameters(), max_norm=self.config.clip_grad_norm)
+        self.optimizers["actor"].step()
+
+        return loss.item(), bc_loss.item(), q_value.mean().item()
+
+    def _update_target_networks(self) -> None:
+        tau = self.config.tau
+        for critic, target in zip(self.critics, self.critic_targets, strict=True):
+            for p, tp in zip(critic.parameters(), target.parameters(), strict=True):
+                tp.data.copy_(tau * p.data + (1 - tau) * tp.data)
+
+    # ── Optimizer management ─────────────────────────────────────────
+
+    def make_optimizers(self) -> dict[str, Optimizer]:
+        """Create optimizers. Initially for RL-token (Stage 1)."""
+        self.optimizers = {
+            "rl_token": torch.optim.Adam(
+                list(self.policy.rl_token_encoder.parameters())
+                + list(self.policy.rl_token_decoder.parameters()),
+                lr=self.config.rl_token_lr,
+            ),
+            "actor": torch.optim.Adam(self.policy.actor.parameters(), lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critics.parameters(), lr=self.config.critic_lr),
+        }
+        return self.optimizers
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        return self.optimizers
+
+    # ── Weight sync ──────────────────────────────────────────────────
+
+    def get_weights(self) -> dict[str, Any]:
+        """Push actor + RL-token encoder to actors (small footprint)."""
+        weights = {
+            "actor": self.policy.actor.state_dict(),
+            "rl_token_encoder": self.policy.rl_token_encoder.state_dict(),
+        }
+        return {k: {kk: vv.cpu() for kk, vv in v.items()} for k, v in weights.items()}
+
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        if "actor" in weights:
+            self.policy.actor.load_state_dict({k: v.to(device) for k, v in weights["actor"].items()})
+        if "rl_token_encoder" in weights:
+            self.policy.rl_token_encoder.load_state_dict(
+                {k: v.to(device) for k, v in weights["rl_token_encoder"].items()}
+            )
@@ -0,0 +1,18 @@
+# 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 lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
+from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+__all__ = ["SACAlgorithm", "SACAlgorithmConfig"]
@@ -0,0 +1,81 @@
+# 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.
+"""SAC algorithm configuration."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
+
+import torch
+
+from lerobot.policies.sac.configuration_sac import CriticNetworkConfig
+from lerobot.rl.algorithms.base import RLAlgorithmConfig
+
+if TYPE_CHECKING:
+    from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+
+@RLAlgorithmConfig.register_subclass("sac")
+@dataclass
+class SACAlgorithmConfig(RLAlgorithmConfig):
+    """SAC-specific hyper-parameters that control the update loop."""
+
+    utd_ratio: int = 1
+    policy_update_freq: int = 1
+    clip_grad_norm: float = 40.0
+    actor_lr: float = 3e-4
+    critic_lr: float = 3e-4
+    temperature_lr: float = 3e-4
+    discount: float = 0.99
+    temperature_init: float = 1.0
+    target_entropy: float | None = None
+    use_backup_entropy: bool = True
+    critic_target_update_weight: float = 0.005
+    num_critics: int = 2
+    num_subsample_critics: int | None = None
+    num_discrete_actions: int | None = None
+    shared_encoder: bool = True
+    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    use_torch_compile: bool = True
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg) -> SACAlgorithmConfig:
+        """Build from an existing ``SACConfig`` (cfg.policy) for backwards compat."""
+        return cls(
+            utd_ratio=policy_cfg.utd_ratio,
+            policy_update_freq=policy_cfg.policy_update_freq,
+            clip_grad_norm=policy_cfg.grad_clip_norm,
+            actor_lr=policy_cfg.actor_lr,
+            critic_lr=policy_cfg.critic_lr,
+            temperature_lr=policy_cfg.temperature_lr,
+            discount=policy_cfg.discount,
+            temperature_init=policy_cfg.temperature_init,
+            target_entropy=policy_cfg.target_entropy,
+            use_backup_entropy=policy_cfg.use_backup_entropy,
+            critic_target_update_weight=policy_cfg.critic_target_update_weight,
+            num_critics=policy_cfg.num_critics,
+            num_subsample_critics=policy_cfg.num_subsample_critics,
+            num_discrete_actions=policy_cfg.num_discrete_actions,
+            shared_encoder=policy_cfg.shared_encoder,
+            critic_network_kwargs=policy_cfg.critic_network_kwargs,
+            discrete_critic_network_kwargs=policy_cfg.discrete_critic_network_kwargs,
+            use_torch_compile=policy_cfg.use_torch_compile,
+        )
+
+    def build_algorithm(self, policy: torch.nn.Module) -> SACAlgorithm:
+        from lerobot.rl.algorithms.sac.sac_algorithm import SACAlgorithm
+
+        return SACAlgorithm(policy=policy, config=self)
@@ -0,0 +1,409 @@
+# 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.
+"""SAC (Soft Actor-Critic) algorithm.
+
+This module encapsulates all SAC-specific training logic (critic, actor,
+temperature, and discrete-critic updates) behind the ``RLAlgorithm`` interface.
+"""
+
+from __future__ import annotations
+
+import math
+from collections.abc import Iterator
+from dataclasses import asdict
+from typing import Any
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+from torch.optim import Optimizer
+
+from lerobot.policies.sac.modeling_sac import (
+    DISCRETE_DIMENSION_INDEX,
+    CriticEnsemble,
+    CriticHead,
+    DiscreteCritic,
+    SACObservationEncoder,
+    SACPolicy,
+)
+from lerobot.policies.utils import get_device_from_parameters
+from lerobot.rl.algorithms.base import (
+    BatchType,
+    RLAlgorithm,
+    TrainingStats,
+)
+from lerobot.rl.algorithms.sac.configuration_sac import SACAlgorithmConfig
+from lerobot.utils.constants import ACTION
+from lerobot.utils.transition import move_state_dict_to_device
+
+
+class SACAlgorithm(RLAlgorithm):
+    """Soft Actor-Critic with optional discrete-critic head.
+
+    Owns the ``SACPolicy`` and its optimizers.  All loss methods call
+    ``self.policy(batch_dict)`` rather than reaching into ``self.policy.actor``
+    directly, so any policy that returns ``{"action", "log_prob"}`` from its
+    ``forward()`` is compatible.
+    """
+
+    def __init__(
+        self,
+        policy: SACPolicy,
+        config: SACAlgorithmConfig,
+    ):
+        self.policy = policy
+        self.config = config
+        self.optimizers: dict[str, Optimizer] = {}
+        self._optimization_step: int = 0
+
+        self._device = get_device_from_parameters(self.policy)
+        self._init_critic_encoder()
+        self._init_critics()
+        self._init_temperature()
+        self._move_to_device()
+
+    def _init_critic_encoder(self) -> None:
+        """Build or share the encoder used by critics."""
+        if self.config.shared_encoder:
+            self.critic_encoder = self.policy.encoder
+            self.policy.actor.encoder_is_shared = True
+        else:
+            self.critic_encoder = SACObservationEncoder(self.policy.config)
+
+    def _init_critics(self) -> None:
+        """Build critic ensemble, targets, and optional discrete critic."""
+        action_dim = self.policy.config.output_features[ACTION].shape[0]
+        input_dim = self.critic_encoder.output_dim + action_dim
+
+        heads = [
+            CriticHead(input_dim=input_dim, **asdict(self.config.critic_network_kwargs))
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_ensemble = CriticEnsemble(encoder=self.critic_encoder, ensemble=heads)
+
+        target_heads = [
+            CriticHead(input_dim=input_dim, **asdict(self.config.critic_network_kwargs))
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_target = CriticEnsemble(encoder=self.critic_encoder, ensemble=target_heads)
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
+
+        if self.config.use_torch_compile:
+            self.critic_ensemble = torch.compile(self.critic_ensemble)
+            self.critic_target = torch.compile(self.critic_target)
+
+        if self.config.num_discrete_actions is not None:
+            self._init_discrete_critic_target()
+
+    def _init_discrete_critic_target(self) -> None:
+        """Build only the target discrete critic."""
+        input_dim = self.critic_encoder.output_dim
+        self.discrete_critic_target = DiscreteCritic(
+            encoder=self.critic_encoder,
+            input_dim=input_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
+        # TODO: (kmeftah) Compile the discrete critic
+        self.discrete_critic_target.load_state_dict(self.policy.discrete_critic.state_dict())
+
+    def _init_temperature(self) -> None:
+        """Set up temperature parameter (log_alpha) and default target entropy."""
+        temp_init = self.config.temperature_init
+        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+
+        action_dim = self.policy.config.output_features[ACTION].shape[0]
+        self.target_entropy = self.config.target_entropy
+        if self.target_entropy is None:
+            dim = action_dim + (1 if self.config.num_discrete_actions is not None else 0)
+            self.target_entropy = -np.prod(dim) / 2
+
+    def _move_to_device(self) -> None:
+        """Move algorithm-owned modules to the policy device."""
+        self.critic_ensemble.to(self._device)
+        self.critic_target.to(self._device)
+        self.log_alpha = nn.Parameter(self.log_alpha.data.to(self._device))
+        if hasattr(self, "discrete_critic_target"):
+            self.discrete_critic_target.to(self._device)
+
+    @property
+    def temperature(self) -> float:
+        return self.log_alpha.exp().item()
+
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """Run one full SAC update with UTD critic warm-up.
+
+        Pulls ``utd_ratio`` batches from ``batch_iterator``.  The first
+        ``utd_ratio - 1`` batches are used for critic-only warm-up steps;
+        the last batch drives the full update (critic + actor + temperature).
+        """
+        for _ in range(self.config.utd_ratio - 1):
+            batch = next(batch_iterator)
+            forward_batch = self._prepare_forward_batch(batch)
+
+            loss_critic = self._compute_loss_critic(forward_batch)
+            self.optimizers["critic"].zero_grad()
+            loss_critic.backward()
+            torch.nn.utils.clip_grad_norm_(
+                self.critic_ensemble.parameters(),
+                max_norm=self.config.clip_grad_norm,
+            ).item()
+            self.optimizers["critic"].step()
+
+            if self.config.num_discrete_actions is not None:
+                loss_discrete = self._compute_loss_discrete_critic(forward_batch)
+                self.optimizers["discrete_critic"].zero_grad()
+                loss_discrete.backward()
+                torch.nn.utils.clip_grad_norm_(
+                    self.policy.discrete_critic.parameters(),
+                    max_norm=self.config.clip_grad_norm,
+                ).item()
+                self.optimizers["discrete_critic"].step()
+            self._update_target_networks()
+
+        batch = next(batch_iterator)
+        forward_batch = self._prepare_forward_batch(batch)
+
+        loss_critic = self._compute_loss_critic(forward_batch)
+        self.optimizers["critic"].zero_grad()
+        loss_critic.backward()
+        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
+            self.critic_ensemble.parameters(),
+            max_norm=self.config.clip_grad_norm,
+        ).item()
+        self.optimizers["critic"].step()
+
+        critic_loss_val = loss_critic.item()
+        stats = TrainingStats(
+            losses={"loss_critic": critic_loss_val},
+            grad_norms={"critic": critic_grad_norm},
+        )
+
+        if self.config.num_discrete_actions is not None:
+            loss_discrete = self._compute_loss_discrete_critic(forward_batch)
+            self.optimizers["discrete_critic"].zero_grad()
+            loss_discrete.backward()
+            dc_grad = torch.nn.utils.clip_grad_norm_(
+                self.policy.discrete_critic.parameters(),
+                max_norm=self.config.clip_grad_norm,
+            ).item()
+            self.optimizers["discrete_critic"].step()
+            stats.losses["loss_discrete_critic"] = loss_discrete.item()
+            stats.grad_norms["discrete_critic"] = dc_grad
+
+        if self._optimization_step % self.config.policy_update_freq == 0:
+            for _ in range(self.config.policy_update_freq):
+                actor_loss = self._compute_loss_actor(forward_batch)
+                self.optimizers["actor"].zero_grad()
+                actor_loss.backward()
+                actor_grad = torch.nn.utils.clip_grad_norm_(
+                    self.policy.actor.parameters(),
+                    max_norm=self.config.clip_grad_norm,
+                ).item()
+                self.optimizers["actor"].step()
+
+                temp_loss = self._compute_loss_temperature(forward_batch)
+                self.optimizers["temperature"].zero_grad()
+                temp_loss.backward()
+                temp_grad = torch.nn.utils.clip_grad_norm_(
+                    [self.log_alpha],
+                    max_norm=self.config.clip_grad_norm,
+                ).item()
+                self.optimizers["temperature"].step()
+
+            stats.losses["loss_actor"] = actor_loss.item()
+            stats.losses["loss_temperature"] = temp_loss.item()
+            stats.grad_norms["actor"] = actor_grad
+            stats.grad_norms["temperature"] = temp_grad
+            stats.extra["temperature"] = self.temperature
+
+        self._update_target_networks()
+
+        self._optimization_step += 1
+        return stats
+
+    def _compute_loss_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        obs_features = batch.get("observation_feature")
+        next_obs_features = batch.get("next_observation_feature")
+
+        with torch.no_grad():
+            next_output = self.policy({"state": next_observations, "observation_feature": next_obs_features})
+            next_actions = next_output["action"]
+            next_log_probs = next_output["log_prob"]
+
+            q_targets = self.critic_target(next_observations, next_actions, next_obs_features)
+
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            min_q, _ = q_targets.min(dim=0)
+            if self.config.use_backup_entropy:
+                min_q = min_q - (self.temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        if self.config.num_discrete_actions is not None:
+            actions = actions[:, :DISCRETE_DIMENSION_INDEX]
+
+        q_preds = self.critic_ensemble(observations, actions, obs_features)
+
+        td_target_dup = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        critics_loss = (F.mse_loss(input=q_preds, target=td_target_dup, reduction="none").mean(dim=1)).sum()
+        return critics_loss
+
+    def _compute_loss_discrete_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        obs_features = batch.get("observation_feature")
+        next_obs_features = batch.get("next_observation_feature")
+        complementary_info = batch.get("complementary_info")
+
+        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
+        actions_discrete = torch.round(actions_discrete).long()
+
+        discrete_penalties: Tensor | None = None
+        if complementary_info is not None:
+            discrete_penalties = complementary_info.get("discrete_penalty")
+
+        with torch.no_grad():
+            next_discrete_qs = self.policy.discrete_critic(next_observations, next_obs_features)
+            best_next_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
+
+            target_next_qs = self.discrete_critic_target(next_observations, next_obs_features)
+            target_next_q = torch.gather(target_next_qs, dim=1, index=best_next_action).squeeze(-1)
+
+            rewards_disc = rewards
+            if discrete_penalties is not None:
+                rewards_disc = rewards + discrete_penalties
+            target_q = rewards_disc + (1 - done) * self.config.discount * target_next_q
+
+        predicted_qs = self.policy.discrete_critic(observations, obs_features)
+        predicted_q = torch.gather(predicted_qs, dim=1, index=actions_discrete).squeeze(-1)
+
+        return F.mse_loss(input=predicted_q, target=target_q)
+
+    def _compute_loss_actor(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        obs_features = batch.get("observation_feature")
+
+        output = self.policy({"state": observations, "observation_feature": obs_features})
+        actions_pi = output["action"]
+        log_probs = output["log_prob"]
+
+        q_preds = self.critic_ensemble(observations, actions_pi, obs_features)
+        min_q = q_preds.min(dim=0)[0]
+
+        return ((self.temperature * log_probs) - min_q).mean()
+
+    def _compute_loss_temperature(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        obs_features = batch.get("observation_feature")
+
+        with torch.no_grad():
+            output = self.policy({"state": observations, "observation_feature": obs_features})
+            log_probs = output["log_prob"]
+
+        return (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
+
+    def _update_target_networks(self) -> None:
+        tau = self.config.critic_target_update_weight
+        for target_p, p in zip(
+            self.critic_target.parameters(), self.critic_ensemble.parameters(), strict=True
+        ):
+            target_p.data.copy_(p.data * tau + target_p.data * (1.0 - tau))
+        if self.config.num_discrete_actions is not None:
+            for target_p, p in zip(
+                self.discrete_critic_target.parameters(),
+                self.policy.discrete_critic.parameters(),
+                strict=True,
+            ):
+                target_p.data.copy_(p.data * tau + target_p.data * (1.0 - tau))
+
+    def _prepare_forward_batch(self, batch: BatchType) -> dict[str, Any]:
+        """Build the dict expected by loss computation from a sampled batch."""
+        observations = batch["state"]
+        next_observations = batch["next_state"]
+
+        observation_features, next_observation_features = self.get_observation_features(
+            observations, next_observations
+        )
+        forward_batch: dict[str, Any] = {
+            ACTION: batch[ACTION],
+            "reward": batch["reward"],
+            "state": observations,
+            "next_state": next_observations,
+            "done": batch["done"],
+            "observation_feature": observation_features,
+            "next_observation_feature": next_observation_features,
+        }
+        if "complementary_info" in batch:
+            forward_batch["complementary_info"] = batch["complementary_info"]
+        return forward_batch
+
+    def make_optimizers(self) -> dict[str, Optimizer]:
+        """Create Adam optimizers for the SAC components and store them."""
+        actor_params = [
+            p
+            for n, p in self.policy.actor.named_parameters()
+            if not self.config.shared_encoder or not n.startswith("encoder")
+        ]
+        self.optimizers = {
+            "actor": torch.optim.Adam(actor_params, lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critic_ensemble.parameters(), lr=self.config.critic_lr),
+            "temperature": torch.optim.Adam([self.log_alpha], lr=self.config.temperature_lr),
+        }
+        if self.config.num_discrete_actions is not None:
+            self.optimizers["discrete_critic"] = torch.optim.Adam(
+                self.policy.discrete_critic.parameters(), lr=self.config.critic_lr
+            )
+        return self.optimizers
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        return self.optimizers
+
+    def get_weights(self) -> dict[str, Any]:
+        """Policy state-dict to push to actors (includes actor + discrete critic)."""
+        return move_state_dict_to_device(self.policy.state_dict(), device="cpu")
+
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load policy state-dict received from the learner."""
+        state = move_state_dict_to_device(weights, device=device)
+        self.policy.load_state_dict(state)
+
+    @torch.no_grad()
+    def get_observation_features(
+        self, observations: Tensor, next_observations: Tensor
+    ) -> tuple[Tensor | None, Tensor | None]:
+        if not self.config.shared_encoder:
+            return None, None
+        if self.policy.config.vision_encoder_name is None or not self.policy.config.freeze_vision_encoder:
+            return None, None
+        if not self.policy.encoder.has_images:
+            return None, None
+        observation_features = self.policy.encoder.get_cached_image_features(observations)
+        next_observation_features = self.policy.encoder.get_cached_image_features(next_observations)
+        return observation_features, next_observation_features
@@ -0,0 +1,17 @@
+# 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 lerobot.rl.data_sources.data_mixer import BatchType, DataMixer, OnlineOfflineMixer
+
+__all__ = ["BatchType", "DataMixer", "OnlineOfflineMixer"]
@@ -0,0 +1,94 @@
+# 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 __future__ import annotations
+
+import abc
+from typing import Any
+
+from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
+
+BatchType = dict[str, Any]
+
+
+class DataMixer(abc.ABC):
+    """Abstract interface for all data mixing strategies.
+
+    Subclasses must implement ``sample(batch_size)`` and may override
+    ``get_iterator`` for specialised iteration.
+    """
+
+    @abc.abstractmethod
+    def sample(self, batch_size: int) -> BatchType:
+        """Draw one batch of ``batch_size`` transitions."""
+        ...
+
+    def get_iterator(
+        self,
+        batch_size: int,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        """Infinite iterator that yields batches.
+
+        The default implementation repeatedly calls ``self.sample()``.
+        Subclasses with underlying buffer iterators (async prefetch)
+        should override this for better throughput.
+        """
+        while True:
+            yield self.sample(batch_size)
+
+
+class OnlineOfflineMixer(DataMixer):
+    """Mixes transitions from an online and an optional offline replay buffer.
+
+    When both buffers are present, each batch is constructed by sampling
+    ``ceil(batch_size * online_ratio)`` from the online buffer and the
+    remainder from the offline buffer, then concatenating.
+
+    This mixer assumes both online and offline buffers are present.
+    """
+
+    def __init__(
+        self,
+        online_buffer: ReplayBuffer,
+        offline_buffer: ReplayBuffer | None = None,
+        online_ratio: float = 1.0,
+    ):
+        if not 0.0 <= online_ratio <= 1.0:
+            raise ValueError(f"online_ratio must be in [0, 1], got {online_ratio}")
+        self.online_buffer = online_buffer
+        self.offline_buffer = offline_buffer
+        self.online_ratio = online_ratio
+
+    def sample(self, batch_size: int) -> BatchType:
+        if self.offline_buffer is None:
+            return self.online_buffer.sample(batch_size)
+
+        n_online = max(1, int(batch_size * self.online_ratio))
+        n_offline = batch_size - n_online
+
+        online_batch = self.online_buffer.sample(n_online)
+        offline_batch = self.offline_buffer.sample(n_offline)
+        return concatenate_batch_transitions(online_batch, offline_batch)
+
+    def get_iterator(
+        self,
+        batch_size: int,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        """Yield batches from online/offline mixed sampling."""
+        while True:
+            yield self.sample(batch_size)
@@ -36,6 +36,7 @@ from lerobot.processor import (
    DeviceProcessorStep,
    EnvTransition,
    GripperPenaltyProcessorStep,
+    GymHILAdapterProcessorStep,
    ImageCropResizeProcessorStep,
    InterventionActionProcessorStep,
    MapDeltaActionToRobotActionStep,
@@ -379,6 +380,7 @@ def make_processors(
        ]

        env_pipeline_steps = [
+            GymHILAdapterProcessorStep(),
            Numpy2TorchActionProcessorStep(),
            VanillaObservationProcessorStep(),
            AddBatchDimensionProcessorStep(),
@@ -412,7 +414,10 @@ def make_processors(
        if cfg.processor.observation.add_current_to_observation:
            env_pipeline_steps.append(MotorCurrentProcessorStep(robot=env.robot))

-    if kinematics_solver is not None:
+    add_ee_pose = (
+        cfg.processor.observation is not None and cfg.processor.observation.add_ee_pose_to_observation
+    )
+    if kinematics_solver is not None and add_ee_pose:
        env_pipeline_steps.append(
            ForwardKinematicsJointsToEEObservation(
                kinematics=kinematics_solver,
@@ -435,7 +440,12 @@ def make_processors(
        )

    # Add gripper penalty processor if gripper config exists and enabled
-    if cfg.processor.gripper is not None and cfg.processor.gripper.use_gripper:
+    # Only add if max_gripper_pos is explicitly configured (required for normalization)
+    if (
+        cfg.processor.gripper is not None
+        and cfg.processor.gripper.use_gripper
+        and cfg.processor.max_gripper_pos is not None
+    ):
        env_pipeline_steps.append(
            GripperPenaltyProcessorStep(
                penalty=cfg.processor.gripper.gripper_penalty,
@@ -600,7 +610,14 @@ def control_loop(

    dataset = None
    if cfg.mode == "record":
-        action_features = teleop_device.action_features
+        if teleop_device:
+            action_features = teleop_device.action_features
+        else:
+            action_features = {
+                "dtype": "float32",
+                "shape": (4,),
+                "names": ["delta_x", "delta_y", "delta_z", "gripper"],
+            }
        features = {
            ACTION: action_features,
            REWARD: {"dtype": "float32", "shape": (1,), "names": None},
@@ -648,7 +665,7 @@ def control_loop(
        # Create a neutral action (no movement)
        neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
        if use_gripper:
-            neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
+            neutral_action = torch.cat([neutral_action, torch.tensor([0.0])])  # Gripper stay

        # Use the new step function
        transition = step_env_and_process_transition(
@@ -717,6 +734,8 @@ def control_loop(
        precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))

    if dataset is not None and cfg.dataset.push_to_hub:
+        logging.info("Finalizing dataset before pushing to hub")
+        dataset.finalize()
        logging.info("Pushing dataset to hub")
        dataset.push_to_hub()

@@ -65,9 +65,11 @@ from lerobot.configs.train import TrainRLServerPipelineConfig
 from lerobot.datasets.factory import make_dataset
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.policies.factory import make_policy
-from lerobot.policies.sac.modeling_sac import SACPolicy
-from lerobot.rl.buffer import ReplayBuffer, concatenate_batch_transitions
+from lerobot.rl.algorithms import make_algorithm
+from lerobot.rl.buffer import ReplayBuffer
+from lerobot.rl.data_sources import OnlineOfflineMixer
 from lerobot.rl.process import ProcessSignalHandler
+from lerobot.rl.trainer import RLTrainer
 from lerobot.rl.wandb_utils import WandBLogger
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
@@ -93,7 +95,7 @@ from lerobot.utils.train_utils import (
    save_checkpoint,
    update_last_checkpoint,
 )
-from lerobot.utils.transition import move_state_dict_to_device, move_transition_to_device
+from lerobot.utils.transition import move_transition_to_device
 from lerobot.utils.utils import (
    format_big_number,
    get_safe_torch_device,
@@ -264,8 +266,8 @@ def add_actor_information_and_train(
    - Transfers transitions from the actor to the replay buffer.
    - Logs received interaction messages.
    - Ensures training begins only when the replay buffer has a sufficient number of transitions.
-    - Samples batches from the replay buffer and performs multiple critic updates.
-    - Periodically updates the actor, critic, and temperature optimizers.
+    - Delegates training updates to an ``RLAlgorithm`` (currently ``SACAlgorithm``).
+    - Periodically pushes updated weights to actors.
    - Logs training statistics, including loss values and optimization frequency.

    NOTE: This function doesn't have a single responsibility, it should be split into multiple functions
@@ -284,17 +286,15 @@ def add_actor_information_and_train(
    # of 7%
    device = get_safe_torch_device(try_device=cfg.policy.device, log=True)
    storage_device = get_safe_torch_device(try_device=cfg.policy.storage_device)
-    clip_grad_norm_value = cfg.policy.grad_clip_norm
    online_step_before_learning = cfg.policy.online_step_before_learning
-    utd_ratio = cfg.policy.utd_ratio
    fps = cfg.env.fps
    log_freq = cfg.log_freq
    save_freq = cfg.save_freq
-    policy_update_freq = cfg.policy.policy_update_freq
    policy_parameters_push_frequency = cfg.policy.actor_learner_config.policy_parameters_push_frequency
    saving_checkpoint = cfg.save_checkpoint
    online_steps = cfg.policy.online_steps
-    async_prefetch = cfg.policy.async_prefetch
+    async_prefetch = cfg.async_prefetch
+    queue_size = cfg.queue_size

    # Initialize logging for multiprocessing
    if not use_threads(cfg):
@@ -306,7 +306,7 @@ def add_actor_information_and_train(

    logging.info("Initializing policy")

-    policy: SACPolicy = make_policy(
+    policy = make_policy(
        cfg=cfg.policy,
        env_cfg=cfg.env,
    )
@@ -315,19 +315,24 @@ def add_actor_information_and_train(

    policy.train()

-    push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
+    algorithm = make_algorithm(
+        policy=policy,
+        policy_cfg=cfg.policy,
+        algorithm_name=cfg.algorithm,
+    )

+    # TODO: Re-enable processor pipeline once refactoring is validated against main
+    preprocessor, postprocessor = None, None
+
+    # Push initial policy weights to actors (same path as periodic push)
+    state_bytes = state_to_bytes(algorithm.get_weights())
+    parameters_queue.put(state_bytes)
    last_time_policy_pushed = time.time()

-    optimizers, lr_scheduler = make_optimizers_and_scheduler(cfg=cfg, policy=policy)
-
-    # If we are resuming, we need to load the training state
-    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)
-
    log_training_info(cfg=cfg, policy=policy)

    replay_buffer = initialize_replay_buffer(cfg, device, storage_device)
-    batch_size = cfg.batch_size
+    total_batch_size = cfg.batch_size
    offline_replay_buffer = None

    if cfg.dataset is not None:
@@ -336,20 +341,70 @@ def add_actor_information_and_train(
            device=device,
            storage_device=storage_device,
        )
-        batch_size: int = batch_size // 2  # We will sample from both replay buffer
+
+    # DataMixer: online-only or online/offline 50-50 mix
+    data_mixer = OnlineOfflineMixer(
+        online_buffer=replay_buffer,
+        offline_buffer=offline_replay_buffer,
+        online_ratio=cfg.online_ratio,
+    )
+    # RLTrainer owns the iterator, preprocessor, and creates optimizers.
+    trainer = RLTrainer(
+        algorithm=algorithm,
+        data_mixer=data_mixer,
+        batch_size=total_batch_size,
+        preprocessor=preprocessor,
+        action_dim=cfg.policy.output_features["action"].shape[0],
+        async_prefetch=async_prefetch,
+        queue_size=queue_size,
+    )
+
+    # If we are resuming, we need to load the training state
+    optimizers = algorithm.get_optimizers()
+    resume_optimization_step, resume_interaction_step = load_training_state(cfg=cfg, optimizers=optimizers)

    logging.info("Starting learner thread")
    interaction_message = None
    optimization_step = resume_optimization_step if resume_optimization_step is not None else 0
+    algorithm.optimization_step = optimization_step
    interaction_step_shift = resume_interaction_step if resume_interaction_step is not None else 0

    dataset_repo_id = None
    if cfg.dataset is not None:
        dataset_repo_id = cfg.dataset.repo_id

-    # Initialize iterators
-    online_iterator = None
-    offline_iterator = None
+    # ── Offline phase (e.g. RLT RL-token training, ConRFT Cal-QL pretraining) ──
+    offline_steps = getattr(cfg.policy, "offline_steps", 0)
+    if algorithm.supports_offline_phase() and offline_steps > 0 and offline_replay_buffer is not None:
+        logging.info(f"[LEARNER] Starting offline phase ({offline_steps} steps)")
+        offline_mixer = OnlineOfflineMixer(
+            online_buffer=offline_replay_buffer,
+            offline_buffer=None,
+            online_ratio=1.0,
+        )
+        offline_iterator = algorithm.configure_data_iterator(
+            data_mixer=offline_mixer,
+            batch_size=total_batch_size,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+        for step in range(offline_steps):
+            if shutdown_event is not None and shutdown_event.is_set():
+                logging.info("[LEARNER] Shutdown during offline phase. Exiting...")
+                return
+
+            stats = algorithm.offline_update(offline_iterator)
+
+            if step % log_freq == 0:
+                logging.info(f"[LEARNER] Offline step {step}/{offline_steps}: {stats.to_log_dict()}")
+                if wandb_logger:
+                    log_dict = stats.to_log_dict()
+                    log_dict["offline_step"] = step
+                    wandb_logger.log_dict(d=log_dict, mode="train", custom_step_key="offline_step")
+
+        algorithm.transition_to_online()
+        optimizers = algorithm.get_optimizers()
+        logging.info("[LEARNER] Offline phase complete, transitioned to online")

    # NOTE: THIS IS THE MAIN LOOP OF THE LEARNER
    while True:
@@ -380,183 +435,22 @@ def add_actor_information_and_train(
        if len(replay_buffer) < online_step_before_learning:
            continue

-        if online_iterator is None:
-            online_iterator = replay_buffer.get_iterator(
-                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
-            )
-
-        if offline_replay_buffer is not None and offline_iterator is None:
-            offline_iterator = offline_replay_buffer.get_iterator(
-                batch_size=batch_size, async_prefetch=async_prefetch, queue_size=2
-            )
-
        time_for_one_optimization_step = time.time()
-        for _ in range(utd_ratio - 1):
-            # Sample from the iterators
-            batch = next(online_iterator)

-            if dataset_repo_id is not None:
-                batch_offline = next(offline_iterator)
-                batch = concatenate_batch_transitions(
-                    left_batch_transitions=batch, right_batch_transition=batch_offline
-                )
-
-            actions = batch[ACTION]
-            rewards = batch["reward"]
-            observations = batch["state"]
-            next_observations = batch["next_state"]
-            done = batch["done"]
-            check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
-
-            observation_features, next_observation_features = get_observation_features(
-                policy=policy, observations=observations, next_observations=next_observations
-            )
-
-            # Create a batch dictionary with all required elements for the forward method
-            forward_batch = {
-                ACTION: actions,
-                "reward": rewards,
-                "state": observations,
-                "next_state": next_observations,
-                "done": done,
-                "observation_feature": observation_features,
-                "next_observation_feature": next_observation_features,
-                "complementary_info": batch["complementary_info"],
-            }
-
-            # Use the forward method for critic loss
-            critic_output = policy.forward(forward_batch, model="critic")
-
-            # Main critic optimization
-            loss_critic = critic_output["loss_critic"]
-            optimizers["critic"].zero_grad()
-            loss_critic.backward()
-            critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
-            )
-            optimizers["critic"].step()
-
-            # Discrete critic optimization (if available)
-            if policy.config.num_discrete_actions is not None:
-                discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
-                loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
-                optimizers["discrete_critic"].zero_grad()
-                loss_discrete_critic.backward()
-                discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
-                )
-                optimizers["discrete_critic"].step()
-
-            # Update target networks (main and discrete)
-            policy.update_target_networks()
-
-        # Sample for the last update in the UTD ratio
-        batch = next(online_iterator)
-
-        if dataset_repo_id is not None:
-            batch_offline = next(offline_iterator)
-            batch = concatenate_batch_transitions(
-                left_batch_transitions=batch, right_batch_transition=batch_offline
-            )
-
-        actions = batch[ACTION]
-        rewards = batch["reward"]
-        observations = batch["state"]
-        next_observations = batch["next_state"]
-        done = batch["done"]
-
-        check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
-
-        observation_features, next_observation_features = get_observation_features(
-            policy=policy, observations=observations, next_observations=next_observations
-        )
-
-        # Create a batch dictionary with all required elements for the forward method
-        forward_batch = {
-            ACTION: actions,
-            "reward": rewards,
-            "state": observations,
-            "next_state": next_observations,
-            "done": done,
-            "observation_feature": observation_features,
-            "next_observation_feature": next_observation_features,
-        }
-
-        critic_output = policy.forward(forward_batch, model="critic")
-
-        loss_critic = critic_output["loss_critic"]
-        optimizers["critic"].zero_grad()
-        loss_critic.backward()
-        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-            parameters=policy.critic_ensemble.parameters(), max_norm=clip_grad_norm_value
-        ).item()
-        optimizers["critic"].step()
-
-        # Initialize training info dictionary
-        training_infos = {
-            "loss_critic": loss_critic.item(),
-            "critic_grad_norm": critic_grad_norm,
-        }
-
-        # Discrete critic optimization (if available)
-        if policy.config.num_discrete_actions is not None:
-            discrete_critic_output = policy.forward(forward_batch, model="discrete_critic")
-            loss_discrete_critic = discrete_critic_output["loss_discrete_critic"]
-            optimizers["discrete_critic"].zero_grad()
-            loss_discrete_critic.backward()
-            discrete_critic_grad_norm = torch.nn.utils.clip_grad_norm_(
-                parameters=policy.discrete_critic.parameters(), max_norm=clip_grad_norm_value
-            ).item()
-            optimizers["discrete_critic"].step()
-
-            # Add discrete critic info to training info
-            training_infos["loss_discrete_critic"] = loss_discrete_critic.item()
-            training_infos["discrete_critic_grad_norm"] = discrete_critic_grad_norm
-
-        # Actor and temperature optimization (at specified frequency)
-        if optimization_step % policy_update_freq == 0:
-            for _ in range(policy_update_freq):
-                # Actor optimization
-                actor_output = policy.forward(forward_batch, model="actor")
-                loss_actor = actor_output["loss_actor"]
-                optimizers["actor"].zero_grad()
-                loss_actor.backward()
-                actor_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=policy.actor.parameters(), max_norm=clip_grad_norm_value
-                ).item()
-                optimizers["actor"].step()
-
-                # Add actor info to training info
-                training_infos["loss_actor"] = loss_actor.item()
-                training_infos["actor_grad_norm"] = actor_grad_norm
-
-                # Temperature optimization
-                temperature_output = policy.forward(forward_batch, model="temperature")
-                loss_temperature = temperature_output["loss_temperature"]
-                optimizers["temperature"].zero_grad()
-                loss_temperature.backward()
-                temp_grad_norm = torch.nn.utils.clip_grad_norm_(
-                    parameters=[policy.log_alpha], max_norm=clip_grad_norm_value
-                ).item()
-                optimizers["temperature"].step()
-
-                # Add temperature info to training info
-                training_infos["loss_temperature"] = loss_temperature.item()
-                training_infos["temperature_grad_norm"] = temp_grad_norm
-                training_infos["temperature"] = policy.temperature
-
-                # Update temperature
-                policy.update_temperature()
+        # One training step (trainer owns data_mixer iterator; algorithm owns UTD loop)
+        stats = trainer.training_step()

        # Push policy to actors if needed
        if time.time() - last_time_policy_pushed > policy_parameters_push_frequency:
-            push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)
+            state_dicts = algorithm.get_weights()
+            state_bytes = state_to_bytes(state_dicts)
+            parameters_queue.put(state_bytes)
            last_time_policy_pushed = time.time()

-        # Update target networks (main and discrete)
-        policy.update_target_networks()
+        training_infos = stats.to_log_dict()

        # Log training metrics at specified intervals
+        optimization_step = algorithm.optimization_step
        if optimization_step % log_freq == 0:
            training_infos["replay_buffer_size"] = len(replay_buffer)
            if offline_replay_buffer is not None:
@@ -584,7 +478,6 @@ def add_actor_information_and_train(
                custom_step_key="Optimization step",
            )

-        optimization_step += 1
        if optimization_step % log_freq == 0:
            logging.info(f"[LEARNER] Number of optimization step: {optimization_step}")

@@ -601,6 +494,8 @@ def add_actor_information_and_train(
                offline_replay_buffer=offline_replay_buffer,
                dataset_repo_id=dataset_repo_id,
                fps=fps,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
            )


@@ -685,6 +580,8 @@ def save_training_checkpoint(
    offline_replay_buffer: ReplayBuffer | None = None,
    dataset_repo_id: str | None = None,
    fps: int = 30,
+    preprocessor=None,
+    postprocessor=None,
 ) -> None:
    """
    Save training checkpoint and associated data.
@@ -708,6 +605,8 @@ def save_training_checkpoint(
        offline_replay_buffer: Optional offline replay buffer to save
        dataset_repo_id: Repository ID for dataset
        fps: Frames per second for dataset
+        preprocessor: Optional preprocessor pipeline to save
+        postprocessor: Optional postprocessor pipeline to save
    """
    logging.info(f"Checkpoint policy after step {optimization_step}")
    _num_digits = max(6, len(str(online_steps)))
@@ -724,6 +623,8 @@ def save_training_checkpoint(
        policy=policy,
        optimizer=optimizers,
        scheduler=None,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
    )

    # Save interaction step manually
@@ -761,58 +662,6 @@ def save_training_checkpoint(
    logging.info("Resume training")


-def make_optimizers_and_scheduler(cfg: TrainRLServerPipelineConfig, policy: nn.Module):
-    """
-    Creates and returns optimizers for the actor, critic, and temperature components of a reinforcement learning policy.
-
-    This function sets up Adam optimizers for:
-    - The **actor network**, ensuring that only relevant parameters are optimized.
-    - The **critic ensemble**, which evaluates the value function.
-    - The **temperature parameter**, which controls the entropy in soft actor-critic (SAC)-like methods.
-
-    It also initializes a learning rate scheduler, though currently, it is set to `None`.
-
-    NOTE:
-    - If the encoder is shared, its parameters are excluded from the actor's optimization process.
-    - The policy's log temperature (`log_alpha`) is wrapped in a list to ensure proper optimization as a standalone tensor.
-
-    Args:
-        cfg: Configuration object containing hyperparameters.
-        policy (nn.Module): The policy model containing the actor, critic, and temperature components.
-
-    Returns:
-        Tuple[Dict[str, torch.optim.Optimizer], Optional[torch.optim.lr_scheduler._LRScheduler]]:
-        A tuple containing:
-        - `optimizers`: A dictionary mapping component names ("actor", "critic", "temperature") to their respective Adam optimizers.
-        - `lr_scheduler`: Currently set to `None` but can be extended to support learning rate scheduling.
-
-    """
-    optimizer_actor = torch.optim.Adam(
-        params=[
-            p
-            for n, p in policy.actor.named_parameters()
-            if not policy.config.shared_encoder or not n.startswith("encoder")
-        ],
-        lr=cfg.policy.actor_lr,
-    )
-    optimizer_critic = torch.optim.Adam(params=policy.critic_ensemble.parameters(), lr=cfg.policy.critic_lr)
-
-    if cfg.policy.num_discrete_actions is not None:
-        optimizer_discrete_critic = torch.optim.Adam(
-            params=policy.discrete_critic.parameters(), lr=cfg.policy.critic_lr
-        )
-    optimizer_temperature = torch.optim.Adam(params=[policy.log_alpha], lr=cfg.policy.critic_lr)
-    lr_scheduler = None
-    optimizers = {
-        "actor": optimizer_actor,
-        "critic": optimizer_critic,
-        "temperature": optimizer_temperature,
-    }
-    if cfg.policy.num_discrete_actions is not None:
-        optimizers["discrete_critic"] = optimizer_discrete_critic
-    return optimizers, lr_scheduler
-
-
 # Training setup functions


@@ -1017,33 +866,6 @@ def initialize_offline_replay_buffer(
 # Utilities/Helpers functions


-def get_observation_features(
-    policy: SACPolicy, observations: torch.Tensor, next_observations: torch.Tensor
-) -> tuple[torch.Tensor | None, torch.Tensor | None]:
-    """
-    Get observation features from the policy encoder. It act as cache for the observation features.
-    when the encoder is frozen, the observation features are not updated.
-    We can save compute by caching the observation features.
-
-    Args:
-        policy: The policy model
-        observations: The current observations
-        next_observations: The next observations
-
-    Returns:
-        tuple: observation_features, next_observation_features
-    """
-
-    if policy.config.vision_encoder_name is None or not policy.config.freeze_vision_encoder:
-        return None, None
-
-    with torch.no_grad():
-        observation_features = policy.actor.encoder.get_cached_image_features(observations)
-        next_observation_features = policy.actor.encoder.get_cached_image_features(next_observations)
-
-    return observation_features, next_observation_features
-
-
 def use_threads(cfg: TrainRLServerPipelineConfig) -> bool:
    return cfg.policy.concurrency.learner == "threads"

@@ -1094,23 +916,6 @@ def check_nan_in_transition(
    return nan_detected


-def push_actor_policy_to_queue(parameters_queue: Queue, policy: nn.Module):
-    logging.debug("[LEARNER] Pushing actor policy to the queue")
-
-    # Create a dictionary to hold all the state dicts
-    state_dicts = {"policy": move_state_dict_to_device(policy.actor.state_dict(), device="cpu")}
-
-    # Add discrete critic if it exists
-    if hasattr(policy, "discrete_critic") and policy.discrete_critic is not None:
-        state_dicts["discrete_critic"] = move_state_dict_to_device(
-            policy.discrete_critic.state_dict(), device="cpu"
-        )
-        logging.debug("[LEARNER] Including discrete critic in state dict push")
-
-    state_bytes = state_to_bytes(state_dicts)
-    parameters_queue.put(state_bytes)
-
-
 def process_interaction_message(
    message, interaction_step_shift: int, wandb_logger: WandBLogger | None = None
 ):
@@ -0,0 +1,132 @@
+# 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 __future__ import annotations
+
+from collections.abc import Iterator
+from typing import Any
+
+import torch
+
+from lerobot.rl.algorithms.base import (
+    BatchType,
+    RLAlgorithm,
+    TrainingStats,
+)
+from lerobot.rl.data_sources.data_mixer import DataMixer
+from lerobot.utils.constants import ACTION
+
+
+def preprocess_rl_batch(preprocessor: Any, batch: BatchType, *, action_dim: int | None = None) -> BatchType:
+    """Apply a policy preprocessor to an RL batch."""
+    observations = batch["state"]
+    next_observations = batch["next_state"]
+    actions = batch[ACTION]
+
+    extra_action = None
+    if action_dim is not None and actions.shape[-1] > action_dim:
+        extra_action = actions[..., action_dim:]
+        actions = actions[..., :action_dim]
+
+    obs_action = {**observations, ACTION: actions}
+    obs_action = preprocessor(obs_action)
+    batch["state"] = {k: v for k, v in obs_action.items() if k.startswith("observation.")}
+    batch[ACTION] = obs_action[ACTION]
+
+    if extra_action is not None:
+        batch[ACTION] = torch.cat([batch[ACTION], extra_action], dim=-1)
+
+    next_obs = {**next_observations}
+    next_obs = preprocessor(next_obs)
+    batch["next_state"] = {k: v for k, v in next_obs.items() if k.startswith("observation.")}
+
+    return batch
+
+
+class _PreprocessedIterator:
+    """Iterator wrapper that preprocesses each sampled RL batch."""
+
+    __slots__ = ("_raw", "_preprocessor", "_action_dim")
+
+    def __init__(
+        self, raw_iterator: Iterator[BatchType], preprocessor: Any, action_dim: int | None = None
+    ) -> None:
+        self._raw = raw_iterator
+        self._preprocessor = preprocessor
+        self._action_dim = action_dim
+
+    def __iter__(self) -> _PreprocessedIterator:
+        return self
+
+    def __next__(self) -> BatchType:
+        batch = next(self._raw)
+        return preprocess_rl_batch(self._preprocessor, batch, action_dim=self._action_dim)
+
+
+class RLTrainer:
+    """Unified training step orchestrator.
+
+    Holds the algorithm, a DataMixer, and an optional preprocessor.
+    """
+
+    def __init__(
+        self,
+        algorithm: RLAlgorithm,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        preprocessor: Any | None = None,
+        action_dim: int | None = None,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ):
+        self.algorithm = algorithm
+        self.data_mixer = data_mixer
+        self.batch_size = batch_size
+        self._preprocessor = preprocessor
+        self._action_dim = action_dim
+        self.async_prefetch = async_prefetch
+        self.queue_size = queue_size
+
+        self._iterator: Iterator[BatchType] | None = None
+
+        self.algorithm.make_optimizers()
+
+    def _build_data_iterator(self) -> Iterator[BatchType]:
+        """Create a fresh algorithm-configured iterator (optionally preprocessed)."""
+        raw = self.algorithm.configure_data_iterator(
+            data_mixer=self.data_mixer,
+            batch_size=self.batch_size,
+            async_prefetch=self.async_prefetch,
+            queue_size=self.queue_size,
+        )
+        if self._preprocessor is not None:
+            return _PreprocessedIterator(raw, self._preprocessor, self._action_dim)
+        return raw
+
+    def reset_data_iterator(self) -> None:
+        """Discard the current iterator so it will be rebuilt lazily next step."""
+        self._iterator = None
+
+    def set_data_mixer(self, data_mixer: DataMixer, *, reset: bool = True) -> None:
+        """Swap the active data mixer, optionally resetting the iterator."""
+        self.data_mixer = data_mixer
+        if reset:
+            self.reset_data_iterator()
+
+    def training_step(self) -> TrainingStats:
+        """Run one training step (algorithm-agnostic)."""
+        if self._iterator is None:
+            self._iterator = self._build_data_iterator()
+        return self.algorithm.update(self._iterator)
--- a/Show More
+++ b/Show More