Potential noise injection for data recording: DART

* fix(RL) add missing config arguments

* respond to copilot review

* fix(revert penalty in reward): reverting gripper penalty addition in reward. This is already done in compute_loss_discrete_critic.

---------

Co-authored-by: CarolinePascal <caroline8.pascal@gmail.com>

docs/source/_toctree.yml

@@ -7,8 +7,6 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
-  - local: cameras
-    title: Cameras
   - local: bring_your_own_policies
     title: Bring Your Own Policies
   - local: integrate_hardware
@@ -29,6 +27,8 @@
     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
@@ -99,11 +99,19 @@
     title: Unitree G1
   - local: earthrover_mini_plus
     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
@@ -113,6 +121,8 @@
     title: Notebooks
   - local: feetech
     title: Updating Feetech Firmware
+  - local: damiao
+    title: Damiao Motors and CAN Bus
   title: "Resources"
 - sections:
   - local: contributing

docs/source/cameras.mdx

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

docs/source/damiao.mdx

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

docs/source/dataset_subtask.mdx

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

docs/source/earthrover_mini_plus.mdx

@@ -1,5 +1,11 @@
 # EarthRover Mini Plus
 
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
+  alt="EarthRover Mini Plus"
+  width="70%"
+/>
+
 The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
 
 ## What You Need

docs/source/lekiwi.mdx

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

docs/source/libero.mdx

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

docs/source/omx.mdx

@@ -0,0 +1,197 @@
+## Order and Assemble the parts
+
+First, assemble the OMX hardware following the official assembly guide.
+
+OMX Assembly Guide: https://ai.robotis.com/omx/assembly_guide_omx.html
+
+OMX robots are shipped preconfigured from the factory. Motor IDs, communication parameters, and joint offsets are already set, so no additional motor setup or calibration is required before using LeRobot.
+
+## Install LeRobot 🤗
+
+To install LeRobot, follow our [Installation Guide](./installation)
+
+In addition to these instructions, you need to install the Dynamixel SDK:
+
+```bash
+pip install -e ".[dynamixel]"
+```
+
+## Connect the robot
+
+To find the port for each bus servo adapter, run this script:
+
+```bash
+lerobot-find-port
+```
+
+This command runs and when prompted, disconnect the USB cable from either the leader or follower arm and press Enter. The output will show 'The port of this MotorsBus is [port]'. This identifies the port for the disconnected arm. Repeat for the other arm to identify both ports.
+
+<hfoptions id="find_port">
+<hfoption id="Mac">
+
+Example output on macOS:
+
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the USB cable from your MotorsBus and press Enter when done.
+
+[...Disconnect corresponding leader or follower arm and press Enter...]
+
+The port of this MotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the USB cable.
+```
+
+Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
+
+</hfoption>
+<hfoption id="Linux">
+
+On Linux, we strongly recommend using udev rules to assign persistent and human-readable device names to the OMX leader and follower arms. This avoids issues where device names such as ttyACM0 and ttyACM1 change when the robot is unplugged, replugged, or when the system is rebooted.
+
+#### 1. Find your device serial numbers
+
+You should have obtained the port numbers like ../../ttyACM? for the leader and follower using `lerobot-find-port`. You can match those results with the serial numbers using the `ls -l /dev/serial/by-id/` command.
+To create udev rules, you need the unique serial number for each OMX device. The easiest way is to list devices under:
+
+```bash
+ls -l /dev/serial/by-id/
+```
+
+You will see output similar to:
+
+```bash
+usb-ROBOTIS_OpenRB-150_228BDD7B503059384C2E3120FF0A2B19-if00 -> ../../ttyACM0
+usb-ROBOTIS_OpenRB-150_67E1ED68503059384C2E3120FF092234-if00 -> ../../ttyACM1
+```
+
+In each line, the serial number is the long string after `usb-ROBOTIS_OpenRB-150_` and before `-if00`.
+
+Follower serial: `228BDD7B503059384C2E3120FF0A2B19`
+
+Leader serial: `67E1ED68503059384C2E3120FF092234`
+
+#### 2. Create the udev rule
+
+Create a new udev rule file:
+
+```bash
+sudo nano /etc/udev/rules.d/99-omx.rules
+```
+
+Paste the following lines, replacing the serial numbers with the values you found above:
+
+```bash
+SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="228BDD7B503059384C2E3120FF0A2B19", SYMLINK+="omx_follower"
+SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="67E1ED68503059384C2E3120FF092234", SYMLINK+="omx_leader"
+```
+
+Save the file and reload udev rules:
+
+```bash
+sudo udevadm control --reload-rules
+sudo udevadm trigger
+```
+
+Now unplug and replug both devices once.
+
+#### 3. Verify the symlinks
+
+Check that the persistent device names exist:
+
+```bash
+ls -l /dev/omx_follower /dev/omx_leader
+```
+
+You should see them pointing to ttyACM\* devices:
+
+```bash
+/dev/omx_follower -> ttyACM*
+/dev/omx_leader   -> ttyACM*
+```
+
+These names remain stable across reboots and reconnections.
+
+</hfoption>
+</hfoptions>
+
+## Teleoperate
+
+After identifying the correct ports, you can directly teleoperate the follower arm using the leader arm.
+
+<hfoptions id="teleoperate">
+<hfoption id="Mac">
+
+### Teleoperate without camera
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=<your_follower_port> \
+  --robot.id=omx_follower_arm \
+  --teleop.type=omx_leader \
+  --teleop.port=<your_leader_port> \
+  --teleop.id=omx_leader_arm
+```
+
+During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
+
+### Teleoperate with camera
+
+You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=<your_follower_port> \
+  --robot.id=omx_follower_arm \
+  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
+  --teleop.type=omx_leader \
+  --teleop.port=<your_leader_port> \
+  --teleop.id=omx_leader_arm \
+  --display_data=true
+```
+
+When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
+
+</hfoption>
+<hfoption id="Linux">
+
+### Teleoperate without camera
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=/dev/omx_follower \
+  --robot.id=omx_follower_arm \
+  --teleop.type=omx_leader \
+  --teleop.port=/dev/omx_leader \
+  --teleop.id=omx_leader_arm
+```
+
+During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
+
+### Teleoperate with camera
+
+You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
+
+```bash
+lerobot-teleoperate \
+  --robot.type=omx_follower \
+  --robot.port=/dev/omx_follower \
+  --robot.id=omx_follower_arm \
+  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
+  --teleop.type=omx_leader \
+  --teleop.port=/dev/omx_leader \
+  --teleop.id=omx_leader_arm \
+  --display_data=true
+```
+
+When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
+
+</hfoption>
+</hfoptions>
+
+Congrats 🎉, your robot is all set to learn a task on its own.
+
+> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/robotis).

docs/source/openarm.mdx

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

docs/source/so101.mdx

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

docs/source/unitree_g1.mdx

@@ -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
+python -m lerobot.scripts.lerobot_record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=true \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true
+```
+
+Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
+
+---
+
+## 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
+python -m lerobot.scripts.lerobot_record \
+    --robot.type=unitree_g1 \
+    --robot.is_simulation=false \
+    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+    --teleop.type=unitree_g1 \
+    --teleop.left_arm_config.port=/dev/ttyACM1 \
+    --teleop.right_arm_config.port=/dev/ttyACM0 \
+    --teleop.id=exo \
+    --dataset.repo_id=your-username/dataset-name \
+    --dataset.single_task="Test" \
+    --dataset.num_episodes=2 \
+    --dataset.episode_time_s=5 \
+    --dataset.reset_time_s=5 \
+    --dataset.push_to_hub=true
+```
+
+**Note**: Update `server_address` to match your robot's camera server IP.
+
+Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
+
+---
 
 ## Additional Resources
 

examples/backward_compatibility/replay.py

@@ -81,24 +81,25 @@ def replay(cfg: ReplayConfig):
     actions = dataset.hf_dataset.select_columns(ACTION)
     robot.connect()
 
-    log_say("Replaying episode", cfg.play_sounds, blocking=True)
-    for idx in range(dataset.num_frames):
-        start_episode_t = time.perf_counter()
+    try:
+        log_say("Replaying episode", cfg.play_sounds, blocking=True)
+        for idx in range(dataset.num_frames):
+            start_episode_t = time.perf_counter()
 
-        action_array = actions[idx][ACTION]
-        action = {}
-        for i, name in enumerate(dataset.features[ACTION]["names"]):
-            key = f"{name.removeprefix('main_')}.pos"
-            action[key] = action_array[i].item()
+            action_array = actions[idx][ACTION]
+            action = {}
+            for i, name in enumerate(dataset.features[ACTION]["names"]):
+                key = f"{name.removeprefix('main_')}.pos"
+                action[key] = action_array[i].item()
 
-        action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
-        action["elbow_flex.pos"] -= 90
-        robot.send_action(action)
+            action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+            action["elbow_flex.pos"] -= 90
+            robot.send_action(action)
 
-        dt_s = time.perf_counter() - start_episode_t
-        precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
-
-    robot.disconnect()
+            dt_s = time.perf_counter() - start_episode_t
+            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+    finally:
+        robot.disconnect()
 
 
 if __name__ == "__main__":

examples/lekiwi/evaluate.py

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

examples/lekiwi/record.py

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

examples/lekiwi/replay.py

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

examples/phone_to_so100/evaluate.py

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

examples/phone_to_so100/record.py

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

examples/phone_to_so100/replay.py

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

examples/so100_to_so100_EE/evaluate.py

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

examples/so100_to_so100_EE/record.py

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

examples/so100_to_so100_EE/replay.py

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

pyproject.toml

@@ -102,14 +102,20 @@ grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+damiao = ["python-can>=4.2.0,<5.0.0"]
 
 # Robots
+openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
     "pyzmq>=26.2.1,<28.0.0",
-    "onnxruntime>=1.16.0,<2.0.0"
+    "onnxruntime>=1.16.0,<2.0.0",
+    "pin>=3.0.0,<4.0.0",
+    "meshcat>=0.3.0,<0.4.0",
+    "matplotlib>=3.9.0,<4.0.0",
+    "casadi>=3.6.0,<4.0.0",
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
@@ -203,6 +209,7 @@ lerobot-info="lerobot.scripts.lerobot_info:main"
 lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
+lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.packages.find]
@@ -278,6 +285,7 @@ default.extend-ignore-identifiers-re = [
     "thw",
     "inpt",
     "ROBOTIS",
+    "OT_VALUE"
 ]
 
 # TODO: Uncomment when ready to use

src/lerobot/cameras/camera.py

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

src/lerobot/cameras/opencv/camera_opencv.py

@@ -70,34 +70,24 @@ 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 ...
         ```
     """
 
@@ -123,6 +113,7 @@ 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)
@@ -146,12 +137,16 @@ class OpenCVCamera(Camera):
         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.")
@@ -170,12 +165,16 @@ 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.")
 
@@ -196,8 +195,7 @@ 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.")
@@ -339,6 +337,17 @@ 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
+
     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 +366,34 @@ 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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        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.")
 
-        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 +403,10 @@ 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 +420,7 @@ 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 +434,7 @@ 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 +442,37 @@ 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 +485,11 @@ 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()
+
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -482,6 +497,7 @@ 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
@@ -500,13 +516,12 @@ class OpenCVCamera(Camera):
             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 +533,42 @@ class OpenCVCamera(Camera):
 
         return frame
 
+    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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        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 +589,9 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
 
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -80,6 +80,8 @@ 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
 
@@ -125,12 +127,7 @@ class Reachy2Camera(Camera):
         """
         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
@@ -145,6 +142,11 @@ class Reachy2Camera(Camera):
         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"):
@@ -165,11 +167,18 @@ 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")
 
@@ -177,13 +186,7 @@ class Reachy2Camera(Camera):
 
     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
@@ -197,12 +200,38 @@ class Reachy2Camera(Camera):
         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}.")
+    def read_latest(self, max_age_ms: int = 1000) -> 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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.latest_frame is None or self.latest_timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        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
 
     def disconnect(self) -> None:
         """

src/lerobot/cameras/realsense/camera_realsense.py

@@ -72,15 +72,14 @@ 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 +132,9 @@ 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)
@@ -158,6 +159,10 @@ 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).
@@ -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.")
 
@@ -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 +335,52 @@ 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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        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]:
+    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 +391,39 @@ class RealSenseCamera(Camera):
             ValueError: If an invalid `color_mode` is requested.
         """
 
+        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 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.")
 
-        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.")
 
-        if self.rs_pipeline is None:
-            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+        self.new_frame_event.clear()
 
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
-
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        color_frame = frame.get_color_frame()
-        color_image_raw = np.asanyarray(color_frame.get_data())
-
-        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 +434,9 @@ 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 +467,7 @@ 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 +475,41 @@ 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")
@@ -498,6 +527,12 @@ 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
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
@@ -506,6 +541,7 @@ 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
@@ -524,17 +560,16 @@ class RealSenseCamera(Camera):
             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 +577,43 @@ class RealSenseCamera(Camera):
 
         return frame
 
+    # NOTE(Steven): Missing implementation for depth for now
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
+        """Return the most recent (color) frame captured immediately (Peeking).
+
+        This method is non-blocking and returns whatever is currently in the
+        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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        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 +637,10 @@ class RealSenseCamera(Camera):
             self.rs_pipeline = None
             self.rs_profile = None
 
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/zmq/camera_zmq.py

@@ -45,6 +45,12 @@ 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 +58,16 @@ 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 +83,17 @@ 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:
@@ -83,10 +101,16 @@ 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
 
     def connect(self, warmup: bool = True) -> None:
-        """Connect to ZMQ camera server."""
+        """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.
+        """
         if self.is_connected:
             raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
@@ -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 +211,117 @@ class ZMQCamera(Camera):
 
         return frame
 
-    def _read_loop(self) -> None:
-        while self.stop_event and not self.stop_event.is_set():
-            try:
-                frame = self.read()
-                with self.frame_lock:
-                    self.latest_frame = frame
-                self.new_frame_event.set()
-            except DeviceNotConnectedError:
-                break
-            except TimeoutError:
-                pass
-            except Exception as e:
-                logger.warning(f"Read error: {e}")
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+        """
+        Reads a single frame synchronously from the camera.
 
-    def _start_read_thread(self) -> None:
-        if self.thread and self.thread.is_alive():
-            return
-        self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, daemon=True)
-        self.thread.start()
+        This is a blocking call. It waits for the next available frame from the
+        camera background thread.
 
-    def _stop_read_thread(self) -> None:
-        if self.stop_event:
-            self.stop_event.set()
-        if self.thread and self.thread.is_alive():
-            self.thread.join(timeout=2.0)
-        self.thread = None
-        self.stop_event = None
+        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."
+            )
 
-    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.")
 
-        if not self.thread or not self.thread.is_alive():
-            self._start_read_thread()
+        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:
+        """
+        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_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, 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.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, name=f"{self}_read_loop")
+        self.thread.start()
+        time.sleep(0.1)
+
+    def _stop_read_thread(self) -> None:
+        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)
+
+        self.thread = None
+        self.stop_event = None
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
+    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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
 
         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 +335,55 @@ class ZMQCamera(Camera):
 
         return frame
 
+    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 not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        with self.frame_lock:
+            frame = self.latest_frame
+            timestamp = self.latest_timestamp
+
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
+
+        return frame
+
     def disconnect(self) -> None:
         """Disconnect from ZMQ camera."""
-        if not self.is_connected and not self.thread:
+        if not self.is_connected and self.thread is None:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        self._stop_read_thread()
+        if self.thread is not None:
+            self._stop_read_thread()
+
         self._cleanup()
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -29,6 +29,7 @@ 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):

src/lerobot/configs/policies.py

@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-        output_shapes: A dictionary defining the shapes of the output data for the policy.
-        input_normalization_modes: A dictionary with key representing the modality and the value specifies the
-            normalization mode to apply.
-        output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
-            the original scale.
+        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
+            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
+            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
+        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
+            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
     """
 
     n_obs_steps: int = 1
@@ -105,16 +105,6 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     def observation_delta_indices(self) -> list | None:  # type: ignore[type-arg] #TODO: No implementation
         raise NotImplementedError
 
-    @property
-    def image_observation_delta_indices(self) -> list | None:  # type: ignore[type-arg]
-        """Return indices for delta image observations only.
-
-        Unlike observation_delta_indices which applies to ALL observations,
-        this only applies to image observations (keys starting with observation.images).
-        Default returns None. Override in subclass to enable.
-        """
-        return None
-
     @property
     @abc.abstractmethod
     def action_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation

src/lerobot/datasets/aggregate.py

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

src/lerobot/datasets/dataset_tools.py

@@ -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,

src/lerobot/datasets/factory.py

@@ -27,7 +27,7 @@ from lerobot.datasets.lerobot_dataset import (
 )
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.transforms import ImageTransforms
-from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_PREFIX, REWARD
+from lerobot.utils.constants import ACTION, OBS_PREFIX, REWARD
 
 IMAGENET_STATS = {
     "mean": [[[0.485]], [[0.456]], [[0.406]]],  # (c,1,1)
@@ -59,12 +59,7 @@ def resolve_delta_timestamps(
             delta_timestamps[key] = [i / ds_meta.fps for i in cfg.reward_delta_indices]
         if key == ACTION and cfg.action_delta_indices is not None:
             delta_timestamps[key] = [i / ds_meta.fps for i in cfg.action_delta_indices]
-
-        # Check for image-specific delta indices first (e.g., for video encoding)
-        if key.startswith(OBS_IMAGES) and cfg.image_observation_delta_indices is not None:
-            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.image_observation_delta_indices]
-        # Fall back to generic observation delta indices for all observations
-        elif key.startswith(OBS_PREFIX) and cfg.observation_delta_indices is not None:
+        if key.startswith(OBS_PREFIX) and cfg.observation_delta_indices is not None:
             delta_timestamps[key] = [i / ds_meta.fps for i in cfg.observation_delta_indices]
 
     if len(delta_timestamps) == 0:

src/lerobot/datasets/lerobot_dataset.py

@@ -57,6 +57,7 @@ from lerobot.datasets.utils import (
     load_info,
     load_nested_dataset,
     load_stats,
+    load_subtasks,
     load_tasks,
     update_chunk_file_indices,
     validate_episode_buffer,
@@ -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(
@@ -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):

src/lerobot/datasets/utils.py

@@ -60,6 +60,7 @@ VIDEO_DIR = "videos"
 
 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
+DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
@@ -353,6 +354,14 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
     return tasks
 
 
+def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
+    """Load subtasks from subtasks.parquet if it exists."""
+    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
+    if subtasks_path.exists():
+        return pd.read_parquet(subtasks_path)
+    return None
+
+
 def write_episodes(episodes: Dataset, local_dir: Path) -> None:
     """Write episode metadata to a parquet file in the LeRobot v3.0 format.
     This function writes episode-level metadata to a single parquet file.

src/lerobot/envs/configs.py

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

src/lerobot/motors/__init__.py

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

src/lerobot/motors/calibration_gui.py

@@ -18,7 +18,7 @@ from dataclasses import dataclass
 
 os.environ["PYGAME_HIDE_SUPPORT_PROMPT"] = "1"
 
-from lerobot.motors import MotorCalibration, MotorsBus
+from .motors_bus import MotorCalibration, MotorsBus
 
 BAR_LEN, BAR_THICKNESS = 450, 8
 HANDLE_R = 10

src/lerobot/motors/damiao/__init__.py

@@ -0,0 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .damiao import DamiaoMotorsBus
+from .tables import *

src/lerobot/motors/damiao/damiao.py

@@ -0,0 +1,833 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Portions of this file are derived from DM_Control_Python by cmjang.
+# Licensed under the MIT License; see `LICENSE` for the full text:
+# https://github.com/cmjang/DM_Control_Python
+
+import logging
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from functools import cached_property
+from typing import TYPE_CHECKING, Any, TypedDict
+
+from lerobot.utils.import_utils import _can_available
+
+if TYPE_CHECKING or _can_available:
+    import can
+else:
+
+    class can:  # noqa: N801
+        Message = object
+        interface = None
+
+
+import numpy as np
+
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import enter_pressed, move_cursor_up
+
+from ..motors_bus import Motor, MotorCalibration, MotorsBusBase, NameOrID, Value
+from .tables import (
+    AVAILABLE_BAUDRATES,
+    CAN_CMD_DISABLE,
+    CAN_CMD_ENABLE,
+    CAN_CMD_REFRESH,
+    CAN_CMD_SET_ZERO,
+    CAN_PARAM_ID,
+    DEFAULT_BAUDRATE,
+    DEFAULT_TIMEOUT_MS,
+    MIT_KD_RANGE,
+    MIT_KP_RANGE,
+    MOTOR_LIMIT_PARAMS,
+    MotorType,
+)
+
+logger = logging.getLogger(__name__)
+
+
+LONG_TIMEOUT_SEC = 0.1
+MEDIUM_TIMEOUT_SEC = 0.01
+SHORT_TIMEOUT_SEC = 0.001
+PRECISE_TIMEOUT_SEC = 0.0001
+
+
+class MotorState(TypedDict):
+    position: float
+    velocity: float
+    torque: float
+    temp_mos: float
+    temp_rotor: float
+
+
+class DamiaoMotorsBus(MotorsBusBase):
+    """
+    The Damiao implementation for a MotorsBus using CAN bus communication.
+
+    This class uses python-can for CAN bus communication with Damiao motors.
+    For more info, see:
+    - python-can documentation: https://python-can.readthedocs.io/en/stable/
+    - Seedstudio documentation: https://wiki.seeedstudio.com/damiao_series/
+    - DM_Control_Python repo: https://github.com/cmjang/DM_Control_Python
+    """
+
+    # CAN-specific settings
+    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
+    default_baudrate = DEFAULT_BAUDRATE
+    default_timeout = DEFAULT_TIMEOUT_MS
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+        can_interface: str = "auto",
+        use_can_fd: bool = True,
+        bitrate: int = 1000000,
+        data_bitrate: int | None = 5000000,
+    ):
+        """
+        Initialize the Damiao motors bus.
+
+        Args:
+            port: CAN interface name (e.g., "can0" for Linux, "/dev/cu.usbmodem*" for macOS)
+            motors: Dictionary mapping motor names to Motor objects
+            calibration: Optional calibration data
+            can_interface: CAN interface type - "auto" (default), "socketcan" (Linux), or "slcan" (macOS/serial)
+            use_can_fd: Whether to use CAN FD mode (default: True for OpenArms)
+            bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
+            data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
+        """
+        super().__init__(port, motors, calibration)
+        self.port = port
+        self.can_interface = can_interface
+        self.use_can_fd = use_can_fd
+        self.bitrate = bitrate
+        self.data_bitrate = data_bitrate
+        self.canbus: can.interface.Bus | None = None
+        self._is_connected = False
+
+        # Map motor names to CAN IDs
+        self._motor_can_ids: dict[str, int] = {}
+        self._recv_id_to_motor: dict[int, str] = {}
+        self._motor_types: dict[str, MotorType] = {}
+
+        for name, motor in self.motors.items():
+            if motor.motor_type_str is None:
+                raise ValueError(f"Motor '{name}' is missing required 'motor_type'")
+            self._motor_types[name] = getattr(MotorType, motor.motor_type_str.upper().replace("-", "_"))
+
+            # Map recv_id to motor name for filtering responses
+            if motor.recv_id is not None:
+                self._recv_id_to_motor[motor.recv_id] = name
+
+        # State cache for handling packet drops safely
+        self._last_known_states: dict[str, MotorState] = {
+            name: {
+                "position": 0.0,
+                "velocity": 0.0,
+                "torque": 0.0,
+                "temp_mos": 0.0,
+                "temp_rotor": 0.0,
+            }
+            for name in self.motors
+        }
+
+        # Dynamic gains storage
+        # Defaults: Kp=10.0 (Stiffness), Kd=0.5 (Damping)
+        self._gains: dict[str, dict[str, float]] = {name: {"kp": 10.0, "kd": 0.5} for name in self.motors}
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if the CAN bus is connected."""
+        return self._is_connected and self.canbus is not None
+
+    def connect(self, handshake: bool = True) -> None:
+        """
+        Open the CAN bus and initialize communication.
+
+        Args:
+            handshake: If True, ping all motors to verify they're present
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(
+                f"{self.__class__.__name__}('{self.port}') is already connected."
+            )
+
+        try:
+            # Auto-detect interface type based on port name
+            if self.can_interface == "auto":
+                if self.port.startswith("/dev/"):
+                    self.can_interface = "slcan"
+                    logger.info(f"Auto-detected slcan interface for port {self.port}")
+                else:
+                    self.can_interface = "socketcan"
+                    logger.info(f"Auto-detected socketcan interface for port {self.port}")
+
+            # Connect to CAN bus
+            kwargs = {
+                "channel": self.port,
+                "bitrate": self.bitrate,
+                "interface": self.can_interface,
+            }
+
+            if self.can_interface == "socketcan" and self.use_can_fd and self.data_bitrate is not None:
+                kwargs.update({"data_bitrate": self.data_bitrate, "fd": True})
+                logger.info(
+                    f"Connected to {self.port} with CAN FD (bitrate={self.bitrate}, data_bitrate={self.data_bitrate})"
+                )
+            else:
+                logger.info(f"Connected to {self.port} with {self.can_interface} (bitrate={self.bitrate})")
+
+            self.canbus = can.interface.Bus(**kwargs)
+            self._is_connected = True
+
+            if handshake:
+                self._handshake()
+
+            logger.debug(f"{self.__class__.__name__} connected via {self.can_interface}.")
+        except Exception as e:
+            self._is_connected = False
+            raise ConnectionError(f"Failed to connect to CAN bus: {e}") from e
+
+    def _handshake(self) -> None:
+        """
+        Verify all motors are present and populate initial state cache.
+        Raises ConnectionError if any motor fails to respond.
+        """
+        logger.info("Starting handshake with motors...")
+
+        # Drain any pending messages
+        while self.canbus.recv(timeout=0.01):
+            pass
+
+        missing_motors = []
+        for motor_name in self.motors:
+            motor_id = self._get_motor_id(motor_name)
+            recv_id = self._get_motor_recv_id(motor_name)
+
+            # Send enable command
+            data = [0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, CAN_CMD_ENABLE]
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+            self.canbus.send(msg)
+
+            # Wait for response with longer timeout
+            response = None
+            start_time = time.time()
+            while time.time() - start_time < 0.1:
+                response = self.canbus.recv(timeout=0.1)
+                if response and response.arbitration_id == recv_id:
+                    break
+                response = None
+
+            if response is None:
+                missing_motors.append(motor_name)
+            else:
+                self._process_response(motor_name, msg)
+            time.sleep(MEDIUM_TIMEOUT_SEC)
+
+        if missing_motors:
+            raise ConnectionError(
+                f"Handshake failed. The following motors did not respond: {missing_motors}. "
+                "Check power (24V) and CAN wiring."
+            )
+        logger.info("Handshake successful. All motors ready.")
+
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """
+        Close the CAN bus connection.
+
+        Args:
+            disable_torque: If True, disable torque on all motors before disconnecting
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self.__class__.__name__}('{self.port}') is not connected.")
+
+        if disable_torque:
+            try:
+                self.disable_torque()
+            except Exception as e:
+                logger.warning(f"Failed to disable torque during disconnect: {e}")
+
+        if self.canbus:
+            self.canbus.shutdown()
+            self.canbus = None
+        self._is_connected = False
+        logger.debug(f"{self.__class__.__name__} disconnected.")
+
+    def configure_motors(self) -> None:
+        """Configure all motors with default settings."""
+        # Damiao motors don't require much configuration in MIT mode
+        # Just ensure they're enabled
+        for motor in self.motors:
+            self._send_simple_command(motor, CAN_CMD_ENABLE)
+            time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    def _send_simple_command(self, motor: NameOrID, command_byte: int) -> None:
+        """Helper to send simple 8-byte commands (Enable, Disable, Zero)."""
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [command_byte]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+        self.canbus.send(msg)
+        if msg := self._recv_motor_response(expected_recv_id=recv_id):
+            self._process_response(motor_name, msg)
+        else:
+            logger.debug(f"No response from {motor_name} after command 0x{command_byte:02X}")
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        target_motors = self._get_motors_list(motors)
+        for motor in target_motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._send_simple_command(motor, CAN_CMD_ENABLE)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        target_motors = self._get_motors_list(motors)
+        for motor in target_motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._send_simple_command(motor, CAN_CMD_DISABLE)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    @contextmanager
+    def torque_disabled(self, motors: str | list[str] | None = None):
+        """
+        Context manager that guarantees torque is re-enabled.
+
+        This helper is useful to temporarily disable torque when configuring motors.
+        """
+        self.disable_torque(motors)
+        try:
+            yield
+        finally:
+            self.enable_torque(motors)
+
+    def set_zero_position(self, motors: str | list[str] | None = None) -> None:
+        """Set current position as zero for selected motors."""
+        target_motors = self._get_motors_list(motors)
+        for motor in target_motors:
+            self._send_simple_command(motor, CAN_CMD_SET_ZERO)
+            time.sleep(MEDIUM_TIMEOUT_SEC)
+
+    def _refresh_motor(self, motor: NameOrID) -> can.Message | None:
+        """Refresh motor status and return the response."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+        msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+        self.canbus.send(msg)
+        return self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _recv_motor_response(
+        self, expected_recv_id: int | None = None, timeout: float = 0.001
+    ) -> can.Message | None:
+        """
+        Receive a response from a motor.
+
+        Args:
+            expected_recv_id: If provided, only return messages from this CAN ID
+            timeout: Timeout in seconds (default: 1ms for high-speed operation)
+        Returns:
+            CAN message if received, None otherwise
+        """
+        try:
+            start_time = time.time()
+            messages_seen = []
+            while time.time() - start_time < timeout:
+                msg = self.canbus.recv(timeout=PRECISE_TIMEOUT_SEC)
+                if msg:
+                    messages_seen.append(f"0x{msg.arbitration_id:02X}")
+                    if expected_recv_id is None or msg.arbitration_id == expected_recv_id:
+                        return msg
+                    logger.debug(
+                        f"Ignoring message from 0x{msg.arbitration_id:02X}, expected 0x{expected_recv_id:02X}"
+                    )
+
+            if logger.isEnabledFor(logging.DEBUG):
+                if messages_seen:
+                    logger.debug(
+                        f"Received {len(messages_seen)} msgs from {set(messages_seen)}, expected 0x{expected_recv_id:02X}"
+                    )
+                else:
+                    logger.debug(f"No CAN messages received (expected 0x{expected_recv_id:02X})")
+        except Exception as e:
+            logger.debug(f"Failed to receive CAN message: {e}")
+        return None
+
+    def _recv_all_responses(
+        self, expected_recv_ids: list[int], timeout: float = 0.002
+    ) -> dict[int, can.Message]:
+        """
+        Efficiently receive responses from multiple motors at once.
+        Uses the OpenArms pattern: collect all available messages within timeout.
+
+        Args:
+            expected_recv_ids: List of CAN IDs we expect responses from
+            timeout: Total timeout in seconds (default: 2ms)
+
+        Returns:
+            Dictionary mapping recv_id to CAN message
+        """
+        responses = {}
+        expected_set = set(expected_recv_ids)
+        start_time = time.time()
+
+        try:
+            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
+                # 100us poll timeout
+                msg = self.canbus.recv(timeout=PRECISE_TIMEOUT_SEC)
+                if msg and msg.arbitration_id in expected_set:
+                    responses[msg.arbitration_id] = msg
+                    if len(responses) == len(expected_recv_ids):
+                        break
+        except Exception as e:
+            logger.debug(f"Error receiving responses: {e}")
+
+        return responses
+
+    def _encode_mit_packet(
+        self,
+        motor_type: MotorType,
+        kp: float,
+        kd: float,
+        position_degrees: float,
+        velocity_deg_per_sec: float,
+        torque: float,
+    ) -> list[int]:
+        """Helper to encode control parameters into 8 bytes for MIT mode."""
+        # Convert degrees to radians
+        position_rad = np.radians(position_degrees)
+        velocity_rad_per_sec = np.radians(velocity_deg_per_sec)
+
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+
+        # Encode parameters
+        kp_uint = self._float_to_uint(kp, *MIT_KP_RANGE, 12)
+        kd_uint = self._float_to_uint(kd, *MIT_KD_RANGE, 12)
+        q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+        dq_uint = self._float_to_uint(velocity_rad_per_sec, -vmax, vmax, 12)
+        tau_uint = self._float_to_uint(torque, -tmax, tmax, 12)
+
+        # Pack data
+        data = [0] * 8
+        data[0] = (q_uint >> 8) & 0xFF
+        data[1] = q_uint & 0xFF
+        data[2] = dq_uint >> 4
+        data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+        data[4] = kp_uint & 0xFF
+        data[5] = kd_uint >> 4
+        data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+        data[7] = tau_uint & 0xFF
+        return data
+
+    def _mit_control(
+        self,
+        motor: NameOrID,
+        kp: float,
+        kd: float,
+        position_degrees: float,
+        velocity_deg_per_sec: float,
+        torque: float,
+    ) -> None:
+        """Send MIT control command to a motor."""
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        motor_type = self._motor_types[motor_name]
+
+        data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+        self.canbus.send(msg)
+
+        recv_id = self._get_motor_recv_id(motor)
+        if msg := self._recv_motor_response(expected_recv_id=recv_id):
+            self._process_response(motor_name, msg)
+        else:
+            logger.debug(f"No response from {motor_name} after MIT control command")
+
+    def _mit_control_batch(
+        self,
+        commands: dict[NameOrID, tuple[float, float, float, float, float]],
+    ) -> None:
+        """
+        Send MIT control commands to multiple motors in batch.
+        Sends all commands first, then collects responses.
+
+        Args:
+            commands: Dict mapping motor name/ID to (kp, kd, position_deg, velocity_deg/s, torque)
+                     Example: {'joint_1': (10.0, 0.5, 45.0, 0.0, 0.0), ...}
+        """
+        if not commands:
+            return
+
+        recv_id_to_motor: dict[int, str] = {}
+
+        # Step 1: Send all MIT control commands
+        for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
+            motor_id = self._get_motor_id(motor)
+            motor_name = self._get_motor_name(motor)
+            motor_type = self._motor_types[motor_name]
+
+            data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
+            self.canbus.send(msg)
+
+            recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
+
+        # Step 2: Collect responses and update state cache
+        responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=SHORT_TIMEOUT_SEC)
+        for recv_id, motor_name in recv_id_to_motor.items():
+            if msg := responses.get(recv_id):
+                self._process_response(motor_name, msg)
+
+    def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
+        """Convert float to unsigned integer for CAN transmission."""
+        x = max(x_min, min(x_max, x))  # Clamp to range
+        span = x_max - x_min
+        data_norm = (x - x_min) / span
+        return int(data_norm * ((1 << bits) - 1))
+
+    def _uint_to_float(self, x: int, x_min: float, x_max: float, bits: int) -> float:
+        """Convert unsigned integer from CAN to float."""
+        span = x_max - x_min
+        data_norm = float(x) / ((1 << bits) - 1)
+        return data_norm * span + x_min
+
+    def _decode_motor_state(
+        self, data: bytearray | bytes, motor_type: MotorType
+    ) -> tuple[float, float, float, int, int]:
+        """
+        Decode motor state from CAN data.
+        Returns: (position_deg, velocity_deg_s, torque, temp_mos, temp_rotor)
+        """
+        if len(data) < 8:
+            raise ValueError("Invalid motor state data")
+
+        # Extract encoded values
+        q_uint = (data[1] << 8) | data[2]
+        dq_uint = (data[3] << 4) | (data[4] >> 4)
+        tau_uint = ((data[4] & 0x0F) << 8) | data[5]
+        t_mos = data[6]
+        t_rotor = data[7]
+
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+
+        # Decode to physical values
+        position_rad = self._uint_to_float(q_uint, -pmax, pmax, 16)
+        velocity_rad_per_sec = self._uint_to_float(dq_uint, -vmax, vmax, 12)
+        torque = self._uint_to_float(tau_uint, -tmax, tmax, 12)
+
+        return np.degrees(position_rad), np.degrees(velocity_rad_per_sec), torque, t_mos, t_rotor
+
+    def _process_response(self, motor: str, msg: can.Message) -> None:
+        """Decode a message and update the motor state cache."""
+        try:
+            motor_type = self._motor_types[motor]
+            pos, vel, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+
+            self._last_known_states[motor] = {
+                "position": pos,
+                "velocity": vel,
+                "torque": torque,
+                "temp_mos": float(t_mos),
+                "temp_rotor": float(t_rotor),
+            }
+        except Exception as e:
+            logger.warning(f"Failed to decode response from {motor}: {e}")
+
+    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)
+        if msg is None:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            raise ConnectionError(
+                f"No response from motor '{motor}' (send ID: 0x{motor_id:02X}, recv ID: 0x{recv_id:02X}). "
+                f"Check that: 1) Motor is powered (24V), 2) CAN wiring is correct, "
+                f"3) Motor IDs are configured correctly using Damiao Debugging Tools"
+            )
+
+        self._process_response(motor, msg)
+        return self._get_cached_value(motor, data_name)
+
+    def _get_cached_value(self, motor: str, data_name: str) -> Value:
+        """Retrieve a specific value from the cache."""
+        state = self._last_known_states[motor]
+        mapping: dict[str, Any] = {
+            "Present_Position": state["position"],
+            "Present_Velocity": state["velocity"],
+            "Present_Torque": state["torque"],
+            "Temperature_MOS": state["temp_mos"],
+            "Temperature_Rotor": state["temp_rotor"],
+        }
+        if data_name not in mapping:
+            raise ValueError(f"Unknown data_name: {data_name}")
+        return mapping[data_name]
+
+    def write(
+        self,
+        data_name: str,
+        motor: str,
+        value: Value,
+    ) -> None:
+        """
+        Write a value to a single motor. Positions are always in degrees.
+        Can write 'Goal_Position', 'Kp', or 'Kd'.
+        """
+        if 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)
+        elif data_name == "Goal_Position":
+            kp = self._gains[motor]["kp"]
+            kd = self._gains[motor]["kd"]
+            self._mit_control(motor, kp, kd, float(value), 0.0, 0.0)
+        else:
+            raise ValueError(f"Writing {data_name} not supported in MIT mode")
+
+    def sync_read(
+        self,
+        data_name: str,
+        motors: str | list[str] | None = None,
+    ) -> dict[str, Value]:
+        """
+        Read the same value from multiple motors simultaneously.
+        """
+        target_motors = self._get_motors_list(motors)
+        self._batch_refresh(target_motors)
+
+        result = {}
+        for motor in target_motors:
+            result[motor] = self._get_cached_value(motor, data_name)
+        return result
+
+    def sync_read_all_states(
+        self,
+        motors: str | list[str] | None = None,
+        *,
+        num_retry: int = 0,
+    ) -> dict[str, MotorState]:
+        """
+        Read ALL motor states (position, velocity, torque) from multiple motors in ONE refresh cycle.
+
+        Returns:
+            Dictionary mapping motor names to state dicts with keys: 'position', 'velocity', 'torque'
+            Example: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
+        """
+        target_motors = self._get_motors_list(motors)
+        self._batch_refresh(target_motors)
+
+        result = {}
+        for motor in target_motors:
+            result[motor] = self._last_known_states[motor].copy()
+        return result
+
+    def _batch_refresh(self, motors: list[str]) -> None:
+        """Internal helper to refresh a list of motors and update cache."""
+        # Send refresh commands
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+            msg = can.Message(
+                arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd
+            )
+            self.canbus.send(msg)
+
+        # Collect responses
+        expected_recv_ids = [self._get_motor_recv_id(m) for m in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=MEDIUM_TIMEOUT_SEC)
+
+        # Update cache
+        for motor in motors:
+            recv_id = self._get_motor_recv_id(motor)
+            msg = responses.get(recv_id)
+            if msg:
+                self._process_response(motor, msg)
+            else:
+                logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
+
+    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+        """
+        Write values to multiple motors simultaneously. Positions are always in degrees.
+        """
+        if data_name in ("Kp", "Kd"):
+            key = data_name.lower()
+            for motor, val in values.items():
+                self._gains[motor][key] = float(val)
+
+        elif data_name == "Goal_Position":
+            # Step 1: Send all MIT control commands
+            recv_id_to_motor: dict[int, str] = {}
+            for motor, value_degrees in values.items():
+                motor_id = self._get_motor_id(motor)
+                motor_name = self._get_motor_name(motor)
+                motor_type = self._motor_types[motor_name]
+
+                kp = self._gains[motor]["kp"]
+                kd = self._gains[motor]["kd"]
+
+                data = self._encode_mit_packet(motor_type, kp, kd, float(value_degrees), 0.0, 0.0)
+                msg = can.Message(
+                    arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd
+                )
+                self.canbus.send(msg)
+                precise_sleep(PRECISE_TIMEOUT_SEC)
+
+                recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
+
+            # Step 2: Collect responses and update state cache
+            responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=MEDIUM_TIMEOUT_SEC)
+            for recv_id, motor_name in recv_id_to_motor.items():
+                if msg := responses.get(recv_id):
+                    self._process_response(motor_name, msg)
+        else:
+            # Fall back to individual writes
+            for motor, value in values.items():
+                self.write(data_name, motor, value)
+
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration data from motors."""
+        # Damiao motors don't store calibration internally
+        # Return existing calibration or empty dict
+        return self.calibration if self.calibration else {}
+
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration data to motors."""
+        # Damiao motors don't store calibration internally
+        # Just cache it in memory
+        if cache:
+            self.calibration = calibration_dict
+
+    def record_ranges_of_motion(
+        self,
+        motors: NameOrID | list[NameOrID] | None = None,
+        display_values: bool = True,
+    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        """
+        Interactively record the min/max values of each motor in degrees.
+
+        Move the joints by hand (with torque disabled) while the method streams live positions.
+        Press Enter to finish.
+        """
+        target_motors = self._get_motors_list(motors)
+
+        self.disable_torque(target_motors)
+        time.sleep(LONG_TIMEOUT_SEC)
+
+        start_positions = self.sync_read("Present_Position", target_motors)
+        mins = start_positions.copy()
+        maxes = start_positions.copy()
+
+        print("\nMove joints through their full range of motion. Press ENTER when done.")
+        user_pressed_enter = False
+
+        while not user_pressed_enter:
+            positions = self.sync_read("Present_Position", target_motors)
+
+            for motor in target_motors:
+                if motor in positions:
+                    mins[motor] = min(positions[motor], mins.get(motor, positions[motor]))
+                    maxes[motor] = max(positions[motor], maxes.get(motor, positions[motor]))
+
+            if display_values:
+                print("\n" + "=" * 50)
+                print(f"{'MOTOR':<20} | {'MIN (deg)':>12} | {'POS (deg)':>12} | {'MAX (deg)':>12}")
+                print("-" * 50)
+                for motor in target_motors:
+                    if motor in positions:
+                        print(
+                            f"{motor:<20} | {mins[motor]:>12.1f} | {positions[motor]:>12.1f} | {maxes[motor]:>12.1f}"
+                        )
+
+            if enter_pressed():
+                user_pressed_enter = True
+
+            if display_values and not user_pressed_enter:
+                move_cursor_up(len(target_motors) + 4)
+
+            time.sleep(LONG_TIMEOUT_SEC)
+
+        self.enable_torque(target_motors)
+
+        for motor in target_motors:
+            if (motor in mins) and (motor in maxes) and (int(abs(maxes[motor] - mins[motor])) < 5):
+                raise ValueError(f"Motor {motor} has insufficient range of motion (< 5 degrees)")
+
+        return mins, maxes
+
+    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+        """Convert motor specification to list of motor names."""
+        if motors is None:
+            return list(self.motors.keys())
+        elif isinstance(motors, str):
+            return [motors]
+        elif isinstance(motors, list):
+            return motors
+        else:
+            raise TypeError(f"Invalid motors type: {type(motors)}")
+
+    def _get_motor_id(self, motor: NameOrID) -> int:
+        """Get CAN ID for a motor."""
+        if isinstance(motor, str):
+            if motor in self.motors:
+                return self.motors[motor].id
+            else:
+                raise ValueError(f"Unknown motor: {motor}")
+        else:
+            return motor
+
+    def _get_motor_name(self, motor: NameOrID) -> str:
+        """Get motor name from name or ID."""
+        if isinstance(motor, str):
+            return motor
+        else:
+            for name, m in self.motors.items():
+                if m.id == motor:
+                    return name
+            raise ValueError(f"Unknown motor ID: {motor}")
+
+    def _get_motor_recv_id(self, motor: NameOrID) -> int:
+        """Get motor recv_id from name or ID."""
+        motor_name = self._get_motor_name(motor)
+        motor_obj = self.motors.get(motor_name)
+        if motor_obj and motor_obj.recv_id is not None:
+            return motor_obj.recv_id
+        else:
+            raise ValueError(f"Motor {motor_obj} doesn't have a valid recv_id (None).")
+
+    @cached_property
+    def is_calibrated(self) -> bool:
+        """Check if motors are calibrated."""
+        return bool(self.calibration)

src/lerobot/motors/damiao/tables.py

@@ -0,0 +1,209 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Configuration tables for Damiao motors."""
+
+from enum import IntEnum
+
+
+# Motor type definitions
+class MotorType(IntEnum):
+    DM3507 = 0
+    DM4310 = 1
+    DM4310_48V = 2
+    DM4340 = 3
+    DM4340_48V = 4
+    DM6006 = 5
+    DM8006 = 6
+    DM8009 = 7
+    DM10010L = 8
+    DM10010 = 9
+    DMH3510 = 10
+    DMH6215 = 11
+    DMG6220 = 12
+
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 1
+    POS_VEL = 2
+    VEL = 3
+    TORQUE_POS = 4
+
+
+# Motor variable IDs (RID)
+class MotorVariable(IntEnum):
+    UV_VALUE = 0
+    KT_VALUE = 1
+    OT_VALUE = 2
+    OC_VALUE = 3
+    ACC = 4
+    DEC = 5
+    MAX_SPD = 6
+    MST_ID = 7
+    ESC_ID = 8
+    TIMEOUT = 9
+    CTRL_MODE = 10
+    DAMP = 11
+    INERTIA = 12
+    HW_VER = 13
+    SW_VER = 14
+    SN = 15
+    NPP = 16
+    RS = 17
+    LS = 18
+    FLUX = 19
+    GR = 20
+    PMAX = 21
+    VMAX = 22
+    TMAX = 23
+    I_BW = 24
+    KP_ASR = 25
+    KI_ASR = 26
+    KP_APR = 27
+    KI_APR = 28
+    OV_VALUE = 29
+    GREF = 30
+    DETA = 31
+    V_BW = 32
+    IQ_C1 = 33
+    VL_C1 = 34
+    CAN_BR = 35
+    SUB_VER = 36
+    U_OFF = 50
+    V_OFF = 51
+    K1 = 52
+    K2 = 53
+    M_OFF = 54
+    DIR = 55
+    P_M = 80
+    XOUT = 81
+
+
+# Motor limit parameters [PMAX, VMAX, TMAX]
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS = {
+    MotorType.DM3507: (12.5, 30, 10),
+    MotorType.DM4310: (12.5, 30, 10),
+    MotorType.DM4310_48V: (12.5, 50, 10),
+    MotorType.DM4340: (12.5, 8, 28),
+    MotorType.DM4340_48V: (12.5, 10, 28),
+    MotorType.DM6006: (12.5, 45, 20),
+    MotorType.DM8006: (12.5, 45, 40),
+    MotorType.DM8009: (12.5, 45, 54),
+    MotorType.DM10010L: (12.5, 25, 200),
+    MotorType.DM10010: (12.5, 20, 200),
+    MotorType.DMH3510: (12.5, 280, 1),
+    MotorType.DMH6215: (12.5, 45, 10),
+    MotorType.DMG6220: (12.5, 45, 10),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.DM3507: "dm3507",
+    MotorType.DM4310: "dm4310",
+    MotorType.DM4310_48V: "dm4310_48v",
+    MotorType.DM4340: "dm4340",
+    MotorType.DM4340_48V: "dm4340_48v",
+    MotorType.DM6006: "dm6006",
+    MotorType.DM8006: "dm8006",
+    MotorType.DM8009: "dm8009",
+    MotorType.DM10010L: "dm10010l",
+    MotorType.DM10010: "dm10010",
+    MotorType.DMH3510: "dmh3510",
+    MotorType.DMH6215: "dmh6215",
+    MotorType.DMG6220: "dmg6220",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "dm3507": 65536,
+    "dm4310": 65536,
+    "dm4310_48v": 65536,
+    "dm4340": 65536,
+    "dm4340_48v": 65536,
+    "dm6006": 65536,
+    "dm8006": 65536,
+    "dm8009": 65536,
+    "dm10010l": 65536,
+    "dm10010": 65536,
+    "dmh3510": 65536,
+    "dmh6215": 65536,
+    "dmg6220": 65536,
+}
+
+# CAN baudrates supported by Damiao motors
+AVAILABLE_BAUDRATES = [
+    125000,  # 0: 125 kbps
+    200000,  # 1: 200 kbps
+    250000,  # 2: 250 kbps
+    500000,  # 3: 500 kbps
+    1000000,  # 4: 1 mbps (default for OpenArms)
+    2000000,  # 5: 2 mbps
+    2500000,  # 6: 2.5 mbps
+    3200000,  # 7: 3.2 mbps
+    4000000,  # 8: 4 mbps
+    5000000,  # 9: 5 mbps
+]
+DEFAULT_BAUDRATE = 1000000  # 1 Mbps is standard for OpenArms
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 1000
+
+# OpenArms specific configurations
+# Based on: https://docs.openarm.dev/software/setup/configure-test
+# OpenArms has 7 DOF per arm (14 total for dual arm)
+OPENARMS_ARM_MOTOR_IDS = {
+    "joint_1": {"send": 0x01, "recv": 0x11},  # J1 - Shoulder pan
+    "joint_2": {"send": 0x02, "recv": 0x12},  # J2 - Shoulder lift
+    "joint_3": {"send": 0x03, "recv": 0x13},  # J3 - Elbow flex
+    "joint_4": {"send": 0x04, "recv": 0x14},  # J4 - Wrist flex
+    "joint_5": {"send": 0x05, "recv": 0x15},  # J5 - Wrist roll
+    "joint_6": {"send": 0x06, "recv": 0x16},  # J6 - Wrist pitch
+    "joint_7": {"send": 0x07, "recv": 0x17},  # J7 - Wrist rotation
+}
+
+OPENARMS_GRIPPER_MOTOR_IDS = {
+    "gripper": {"send": 0x08, "recv": 0x18},  # J8 - Gripper
+}
+
+# Default motor types for OpenArms
+OPENARMS_DEFAULT_MOTOR_TYPES = {
+    "joint_1": MotorType.DM8009,  # Shoulder pan - high torque
+    "joint_2": MotorType.DM8009,  # Shoulder lift - high torque
+    "joint_3": MotorType.DM4340,  # Shoulder rotation
+    "joint_4": MotorType.DM4340,  # Elbow flex
+    "joint_5": MotorType.DM4310,  # Wrist roll
+    "joint_6": MotorType.DM4310,  # Wrist pitch
+    "joint_7": MotorType.DM4310,  # Wrist rotation
+    "gripper": MotorType.DM4310,  # Gripper
+}
+
+# MIT control parameter ranges
+MIT_KP_RANGE = (0.0, 500.0)
+MIT_KD_RANGE = (0.0, 5.0)
+
+# CAN frame command IDs
+CAN_CMD_ENABLE = 0xFC
+CAN_CMD_DISABLE = 0xFD
+CAN_CMD_SET_ZERO = 0xFE
+CAN_CMD_REFRESH = 0xCC
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF

src/lerobot/motors/dynamixel/dynamixel.py

@@ -22,9 +22,8 @@ import logging
 from copy import deepcopy
 from enum import Enum
 
-from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..encoding_utils import decode_twos_complement, encode_twos_complement
+from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     AVAILABLE_BAUDRATES,
     MODEL_BAUDRATE_TABLE,
@@ -100,7 +99,7 @@ def _split_into_byte_chunks(value: int, length: int) -> list[int]:
     return data
 
 
-class DynamixelMotorsBus(MotorsBus):
+class DynamixelMotorsBus(SerialMotorsBus):
     """
     The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
     the motors. For more info, see the Dynamixel SDK Documentation:
@@ -203,9 +202,9 @@ class DynamixelMotorsBus(MotorsBus):
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
     def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):

src/lerobot/motors/feetech/feetech.py

@@ -17,9 +17,8 @@ from copy import deepcopy
 from enum import Enum
 from pprint import pformat
 
-from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     FIRMWARE_MAJOR_VERSION,
     FIRMWARE_MINOR_VERSION,
@@ -96,7 +95,7 @@ def patch_setPacketTimeout(self, packet_length):  # noqa: N802
     self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50
 
 
-class FeetechMotorsBus(MotorsBus):
+class FeetechMotorsBus(SerialMotorsBus):
     """
     The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
     python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.
@@ -298,11 +297,11 @@ class FeetechMotorsBus(MotorsBus):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 0, num_retry=num_retry)
 
-    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
         addr, length = get_address(self.model_ctrl_table, model, "Lock")
-        self._write(addr, length, motor_id, 0, num_retry=num_retry)
+        self._write(addr, length, motor, 0, num_retry=num_retry)
 
     def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):

src/lerobot/motors/motors_bus.py

@@ -19,6 +19,8 @@
 # TODO(aliberts): Add block noqa when feature below is available
 # https://github.com/astral-sh/ruff/issues/3711
 
+from __future__ import annotations
+
 import abc
 import logging
 from contextlib import contextmanager
@@ -41,6 +43,81 @@ Value: TypeAlias = int | float
 logger = logging.getLogger(__name__)
 
 
+class MotorsBusBase(abc.ABC):
+    """
+    Base class for all motor bus implementations.
+
+    This is a minimal interface that all motor buses must implement, regardless of their
+    communication protocol (serial, CAN, etc.).
+    """
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+    ):
+        self.port = port
+        self.motors = motors
+        self.calibration = calibration if calibration else {}
+
+    @abc.abstractmethod
+    def connect(self, handshake: bool = True) -> None:
+        """Establish connection to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """Disconnect from the motors."""
+        pass
+
+    @property
+    @abc.abstractmethod
+    def is_connected(self) -> bool:
+        """Check if connected to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def read(self, data_name: str, motor: str) -> Value:
+        """Read a value from a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def write(self, data_name: str, motor: str, value: Value) -> None:
+        """Write a value to a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def sync_read(self, data_name: str, motors: str | list[str] | None = None) -> dict[str, Value]:
+        """Read a value from multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def sync_write(self, data_name: str, values: Value | dict[str, Value]) -> None:
+        """Write values to multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration parameters from the motors."""
+        pass
+
+    @abc.abstractmethod
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration parameters to the motors."""
+        pass
+
+
 def get_ctrl_table(model_ctrl_table: dict[str, dict], model: str) -> dict[str, tuple[int, int]]:
     ctrl_table = model_ctrl_table.get(model)
     if ctrl_table is None:
@@ -97,6 +174,8 @@ class Motor:
     id: int
     model: str
     norm_mode: MotorNormMode
+    motor_type_str: str | None = None
+    recv_id: int | None = None
 
 
 class PortHandler(Protocol):
@@ -203,15 +282,15 @@ class GroupSyncWrite(Protocol):
     def txPacket(self): ...
 
 
-class MotorsBus(abc.ABC):
+class SerialMotorsBus(MotorsBusBase):
     """
-    A MotorsBus allows to efficiently read and write to the attached motors.
+    A SerialMotorsBus allows to efficiently read and write to motors connected via serial communication.
     It represents several motors daisy-chained together and connected through a serial port.
-    There are currently two implementations of this abstract class:
+    There are currently two implementations of this class:
         - DynamixelMotorsBus
         - FeetechMotorsBus
 
-    Note: This class may evolve in the future should we add support for other types of bus.
+    This class is specifically for serial-based motor protocols (Dynamixel, Feetech, etc.).
 
     A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
     To find the port, you can run our utility script:
@@ -260,9 +339,7 @@ class MotorsBus(abc.ABC):
         motors: dict[str, Motor],
         calibration: dict[str, MotorCalibration] | None = None,
     ):
-        self.port = port
-        self.motors = motors
-        self.calibration = calibration if calibration else {}
+        super().__init__(port, motors, calibration)
 
         self.port_handler: PortHandler
         self.packet_handler: PacketHandler
@@ -532,7 +609,7 @@ class MotorsBus(abc.ABC):
         self.set_baudrate(self.default_baudrate)
 
     @abc.abstractmethod
-    def _find_single_motor(self, motor: str, initial_baudrate: int | None) -> tuple[int, int]:
+    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
         pass
 
     @abc.abstractmethod
@@ -545,13 +622,13 @@ class MotorsBus(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         """Disable torque on selected motors.
 
         Disabling Torque allows to write to the motors' permanent memory area (EPROM/EEPROM).
 
         Args:
-            motors (int | str | list[str] | None, optional): Target motors.  Accepts a motor name, an ID, a
+            motors ( str | list[str] | None, optional): Target motors.  Accepts a motor name, an ID, a
                 list of names or `None` to affect every registered motor.  Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
@@ -1194,3 +1271,7 @@ class MotorsBus(abc.ABC):
         for id_, value in ids_values.items():
             data = self._serialize_data(value, length)
             self.sync_writer.addParam(id_, data)
+
+
+# Backward compatibility alias
+MotorsBus: TypeAlias = SerialMotorsBus

src/lerobot/policies/act/configuration_act.py

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

src/lerobot/policies/diffusion/configuration_diffusion.py

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

src/lerobot/policies/factory.py

@@ -35,7 +35,6 @@ from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.policies.videovla.configuration_pi05 import PI05VideoConfig
 from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 from lerobot.policies.sarm.configuration_sarm import SARMConfig
@@ -68,7 +67,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-              "vqbet", "pi0", "pi05", "pi05_video", "sac", "reward_classifier", "smolvla", "wall_x".
+              "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla", "wall_x".
 
     Returns:
         The policy class corresponding to the given name.
@@ -104,10 +103,6 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.pi05.modeling_pi05 import PI05Policy
 
         return PI05Policy
-    elif name == "pi05_video":
-        from lerobot.policies.videovla.modeling_pi05 import PI05VideoPolicy
-
-        return PI05VideoPolicy
     elif name == "sac":
         from lerobot.policies.sac.modeling_sac import SACPolicy
 
@@ -152,7 +147,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
-                     "diffusion", "act", "vqbet", "pi0", "pi05", "pi05_video", "sac", "smolvla",
+                     "diffusion", "act", "vqbet", "pi0", "pi05", "sac", "smolvla",
                      "reward_classifier", "wall_x".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
@@ -174,8 +169,6 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi05":
         return PI05Config(**kwargs)
-    elif policy_type == "pi05_video":
-        return PI05VideoConfig(**kwargs)
     elif policy_type == "sac":
         return SACConfig(**kwargs)
     elif policy_type == "smolvla":
@@ -340,14 +333,6 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
-    elif isinstance(policy_cfg, PI05VideoConfig):
-        from lerobot.policies.videovla.processor_pi05 import make_pi05_video_pre_post_processors
-
-        processors = make_pi05_video_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
     elif isinstance(policy_cfg, SACConfig):
         from lerobot.policies.sac.processor_sac import make_sac_pre_post_processors
 

src/lerobot/policies/pi05/modeling_pi05.py

@@ -460,8 +460,8 @@ class PaliGemmaWithExpertModel(
                 inputs_embeds=inputs_embeds[1],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
-                use_cache=False,
-                past_key_values=None, #jadechoghari
+                past_key_values=past_key_values,
+                use_cache=use_cache,
                 adarms_cond=adarms_cond[1] if adarms_cond is not None else None,
             )
             suffix_output = suffix_output.last_hidden_state
@@ -575,13 +575,13 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
 
-        # try:
-        #     from transformers.models.siglip import check
+        try:
+            from transformers.models.siglip import check
 
-        #     if not check.check_whether_transformers_replace_is_installed_correctly():
-        #         raise ValueError(msg)
-        # except ImportError:
-        #     raise ValueError(msg) from None
+            if not check.check_whether_transformers_replace_is_installed_correctly():
+                raise ValueError(msg)
+        except ImportError:
+            raise ValueError(msg) from None
 
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""

src/lerobot/policies/sac/modeling_sac.py

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

src/lerobot/policies/tdmpc/configuration_tdmpc.py

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

src/lerobot/policies/videovla/README.md

@@ -1,49 +0,0 @@
-# π₀.₅ (pi05)
-
-This repository contains the Hugging Face port of **π₀.₅**, adapted from [OpenPI](https://github.com/Physical-Intelligence/openpi) by the Physical Intelligence.
-It is designed as a **Vision-Language-Action model with open-world generalization**.
-
----
-
-## Model Overview
-
-| Feature              | π₀                                                     | π₀.₅                                      |
-| -------------------- | ------------------------------------------------------ | ----------------------------------------- |
-| Time Conditioning    | Concatenates time with actions via `action_time_mlp_*` | Uses `time_mlp_*` for AdaRMS conditioning |
-| AdaRMS               | Not used                                               | Used in action expert                     |
-| Tokenizer Length     | 48 tokens                                              | 200 tokens                                |
-| Discrete State Input | False (Uses `state_proj` layer)                        | True                                      |
-| Parameter Count      | Higher (includes state embedding)                      | Lower (no state embedding)                |
-
----
-
-## Citation
-
-If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
-
-```bibtex
-@misc{openpi2024,
-  author       = {Physical Intelligence Lab},
-  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
-  year         = {2024},
-  publisher    = {GitHub},
-  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
-  license      = {Apache-2.0}
-}
-
-@misc{intelligence2025pi05visionlanguageactionmodelopenworld,
-  title        = {π₀.₅: a Vision-Language-Action Model with Open-World Generalization},
-  author       = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
-  year         = {2025},
-  eprint       = {2504.16054},
-  archivePrefix= {arXiv},
-  primaryClass = {cs.LG},
-  url          = {https://arxiv.org/abs/2504.16054},
-}
-```
-
----
-
-## License
-
-This port follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

src/lerobot/policies/videovla/__init__.py

@@ -1,31 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# Lazy imports to avoid conflicts with lerobot.policies.pi05.PI05Config
-# when only importing subpackages like videoprism
-def __getattr__(name):
-    if name == "PI05VideoConfig":
-        from .configuration_pi05 import PI05VideoConfig
-        return PI05VideoConfig
-    elif name == "PI05VideoPolicy":
-        from .modeling_pi05 import PI05VideoPolicy
-        return PI05VideoPolicy
-    elif name == "make_pi05_video_pre_post_processors":
-        from .processor_pi05 import make_pi05_video_pre_post_processors
-        return make_pi05_video_pre_post_processors
-    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
-
-__all__ = ["PI05VideoConfig", "PI05VideoPolicy", "make_pi05_video_pre_post_processors"]

src/lerobot/policies/videovla/configuration_pi05.py

@@ -1,212 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass, field
-
-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
-from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
-
-DEFAULT_IMAGE_SIZE = 224
-
-
-@PreTrainedConfig.register_subclass("pi05_video")
-@dataclass
-class PI05VideoConfig(PreTrainedConfig):
-    paligemma_variant: str = "gemma_2b"
-    action_expert_variant: str = "gemma_300m"
-    dtype: str = "float32"  # Options: "bfloat16", "float32"
-
-    n_obs_steps: int = 1
-    chunk_size: int = 50  # Number of action steps to predict, in openpi called "action_horizon"
-    n_action_steps: int = 50  # Number of action steps to execute
-
-    # Video encoder settings (VideoPrism)
-    use_video_encoder: bool = False  # Enable video encoding with VideoPrism
-    video_num_frames: int = 16  # Number of frames for video encoding (VideoPrism default is 16)
-    videoprism_model_name: str = "MHRDYN7/videoprism-base-f16r288"  # VideoPrism model to use
-    videoprism_image_size: int = 288  # VideoPrism expects 288x288 images
-    freeze_video_encoder: bool = True  # Whether to freeze the video encoder weights
-    video_padding_mode: str = "repeat"  # How to pad frames at episode start: "repeat" or "zero"
-    # Which camera to use for video encoding (None = first camera, or specify key like "observation.images.top")
-    video_encoder_camera_key: str | None = None
-    # Perceiver Resampler settings to reduce video tokens (4096 -> video_num_latents)
-    video_num_latents: int = 256  # Number of latent tokens for video resampler
-    video_resampler_num_heads: int = 8  # Number of attention heads in resampler
-
-    # Shorter state and action vectors will be padded to these dimensions
-    max_state_dim: int = 32
-    max_action_dim: int = 32
-
-    # Flow matching parameters: see openpi `PI0Pytorch`
-    num_inference_steps: int = 10
-    time_sampling_beta_alpha: float = 1.5
-    time_sampling_beta_beta: float = 1.0
-    time_sampling_scale: float = 0.999
-    time_sampling_offset: float = 0.001
-    min_period: float = 4e-3
-    max_period: float = 4.0
-
-    # Real-Time Chunking (RTC) configuration
-    rtc_config: RTCConfig | None = None
-
-    image_resolution: tuple[int, int] = (
-        DEFAULT_IMAGE_SIZE,
-        DEFAULT_IMAGE_SIZE,
-    )  # see openpi `preprocessing_pytorch.py`
-
-    # Add empty images. Used to add empty cameras when no image features are present.
-    empty_cameras: int = 0
-
-    tokenizer_max_length: int = 200  # see openpi `__post_init__`
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.QUANTILES,  # Pi0.5 uses quantiles for state
-            "ACTION": NormalizationMode.QUANTILES,  # Pi0.5 uses quantiles for action
-        }
-    )
-
-    # Training settings
-    gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
-    compile_model: bool = False  # Whether to use torch.compile for model optimization
-    compile_mode: str = "max-autotune"  # Torch compile mode
-    device: str | None = None  # Device to use for the model (None = auto-detect)
-
-    # Finetuning settings
-    freeze_vision_encoder: bool = False  # Freeze only the vision encoder
-    train_expert_only: bool = False  # Freeze entire VLM, train only action expert and projections
-
-    # Optimizer settings: see openpi `AdamW`
-    optimizer_lr: float = 2.5e-5  # see openpi `CosineDecaySchedule: peak_lr`
-    optimizer_betas: tuple[float, float] = (0.9, 0.95)
-    optimizer_eps: float = 1e-8
-    optimizer_weight_decay: float = 0.01
-    optimizer_grad_clip_norm: float = 1.0
-
-    # Scheduler settings: see openpi `CosineDecaySchedule`
-    # Note: These will auto-scale if --steps < scheduler_decay_steps
-    # For example, --steps=3000 will scale warmup to 100 and decay to 3000
-    scheduler_warmup_steps: int = 1_000
-    scheduler_decay_steps: int = 30_000
-    scheduler_decay_lr: float = 2.5e-6
-
-    tokenizer_max_length: int = 200  # see openpi `__post_init__`
-
-    def __post_init__(self):
-        super().__post_init__()
-
-        # Validate configuration
-        if self.n_action_steps > self.chunk_size:
-            raise ValueError(
-                f"n_action_steps ({self.n_action_steps}) cannot be greater than chunk_size ({self.chunk_size})"
-            )
-
-        if self.paligemma_variant not in ["gemma_300m", "gemma_2b"]:
-            raise ValueError(f"Invalid paligemma_variant: {self.paligemma_variant}")
-
-        if self.action_expert_variant not in ["gemma_300m", "gemma_2b"]:
-            raise ValueError(f"Invalid action_expert_variant: {self.action_expert_variant}")
-
-        if self.dtype not in ["bfloat16", "float32"]:
-            raise ValueError(f"Invalid dtype: {self.dtype}")
-
-        # Validate video encoder settings
-        if self.use_video_encoder:
-            if self.video_num_frames < 1:
-                raise ValueError(f"video_num_frames must be >= 1, got {self.video_num_frames}")
-            if self.videoprism_image_size < 1:
-                raise ValueError(f"videoprism_image_size must be >= 1, got {self.videoprism_image_size}")
-            if self.video_padding_mode not in ["repeat", "zero"]:
-                raise ValueError(
-                    f"video_padding_mode must be 'repeat' or 'zero', got {self.video_padding_mode}"
-                )
-
-    def validate_features(self) -> None:
-        """Validate and set up input/output features."""
-        for i in range(self.empty_cameras):
-            key = OBS_IMAGES + f".empty_camera_{i}"
-            empty_camera = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(3, *self.image_resolution),  # Use configured image resolution
-            )
-            self.input_features[key] = empty_camera
-
-        if OBS_STATE not in self.input_features:
-            state_feature = PolicyFeature(
-                type=FeatureType.STATE,
-                shape=(self.max_state_dim,),  # Padded to max_state_dim
-            )
-            self.input_features[OBS_STATE] = state_feature
-
-        if ACTION not in self.output_features:
-            action_feature = PolicyFeature(
-                type=FeatureType.ACTION,
-                shape=(self.max_action_dim,),  # Padded to max_action_dim
-            )
-            self.output_features[ACTION] = action_feature
-
-    def get_optimizer_preset(self) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            betas=self.optimizer_betas,
-            eps=self.optimizer_eps,
-            weight_decay=self.optimizer_weight_decay,
-            grad_clip_norm=self.optimizer_grad_clip_norm,
-        )
-
-    def get_scheduler_preset(self):
-        return CosineDecayWithWarmupSchedulerConfig(
-            peak_lr=self.optimizer_lr,
-            decay_lr=self.scheduler_decay_lr,
-            num_warmup_steps=self.scheduler_warmup_steps,
-            num_decay_steps=self.scheduler_decay_steps,
-        )
-
-    @property
-    def observation_delta_indices(self) -> list[int] | None:
-        """Return indices for delta observations.
-
-        For PI05, we don't use generic observation_delta_indices because it would
-        apply to both images AND state. Instead, we use image_observation_delta_indices
-        which only applies to image observations.
-        """
-        return None
-
-    @property
-    def image_observation_delta_indices(self) -> list[int] | None:
-        """Return indices for delta image observations only.
-
-        When video encoding is enabled, returns indices for the past frames
-        needed by VideoPrism (e.g., -15, -14, ..., -1, 0 for 16 frames).
-        This only applies to image observations, not state.
-        """
-        if self.use_video_encoder:
-            # Return indices for past frames: [-15, -14, ..., -1, 0] for 16 frames
-            return list(range(-(self.video_num_frames - 1), 1))
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return list(range(self.chunk_size))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

src/lerobot/policies/videovla/processor_pi05.py

@@ -1,171 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from copy import deepcopy
-from dataclasses import dataclass
-from typing import Any
-
-import numpy as np
-import torch
-
-from lerobot.configs.types import PipelineFeatureType, PolicyFeature
-from lerobot.policies.videovla.configuration_pi05 import PI05VideoConfig
-from lerobot.policies.pi05.modeling_pi05 import pad_vector
-from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
-    PolicyAction,
-    PolicyProcessorPipeline,
-    ProcessorStep,
-    ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-)
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.processor.core import EnvTransition, TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
-
-
-@ProcessorStepRegistry.register(name="pi05_prepare_state_tokenizer_processor_step")
-@dataclass
-class Pi05PrepareStateTokenizerProcessorStep(ProcessorStep):
-    """
-    Processor step to prepare the state and tokenize the language input.
-    """
-
-    max_state_dim: int = 32
-    task_key: str = "task"
-
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
-        transition = transition.copy()
-
-        state = transition.get(TransitionKey.OBSERVATION, {}).get(OBS_STATE)
-        if state is None:
-            raise ValueError("State is required for PI05")
-        tasks = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}).get(self.task_key)
-        if tasks is None:
-            raise ValueError("No task found in complementary data")
-
-        # TODO: check if this necessary
-        state = deepcopy(state)
-
-        # Prepare state (pad to max_state_dim)
-        state = pad_vector(state, self.max_state_dim)
-
-        # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
-        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
-        state_np = state.cpu().numpy()
-        discretized_states = np.digitize(state_np, bins=np.linspace(-1, 1, 256 + 1)[:-1]) - 1
-
-        full_prompts = []
-        for i, task in enumerate(tasks):
-            cleaned_text = task.strip().replace("_", " ").replace("\n", " ")
-            state_str = " ".join(map(str, discretized_states[i]))
-            full_prompt = f"Task: {cleaned_text}, State: {state_str};\nAction: "
-            full_prompts.append(full_prompt)
-
-        transition[TransitionKey.COMPLEMENTARY_DATA][self.task_key] = full_prompts
-        # Normalize state to [-1, 1] range if needed (assuming it's already normalized by normalizer processor step!!)
-        # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
-        return transition
-
-    def transform_features(
-        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-        """
-        This step does not alter the feature definitions.
-        """
-        return features
-
-
-def make_pi05_video_pre_post_processors(
-    config: PI05VideoConfig,
-    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
-) -> tuple[
-    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    PolicyProcessorPipeline[PolicyAction, PolicyAction],
-]:
-    """
-    Constructs pre-processor and post-processor pipelines for the PI05Video policy.
-
-    The pre-processing pipeline prepares input data for the model by:
-    1. Renaming features to match pretrained configurations.
-    2. Normalizing input and output features based on dataset statistics.
-    3. Adding a batch dimension.
-    4. Appending a newline character to the task description for tokenizer compatibility.
-    5. Tokenizing the text prompt using the PaliGemma tokenizer.
-    6. Moving all data to the specified device.
-
-    The post-processing pipeline handles the model's output by:
-    1. Moving data to the CPU.
-    2. Unnormalizing the output features to their original scale.
-
-    Args:
-        config: The configuration object for the PI0 policy.
-        dataset_stats: A dictionary of statistics for normalization.
-        preprocessor_kwargs: Additional arguments for the pre-processor pipeline.
-        postprocessor_kwargs: Additional arguments for the post-processor pipeline.
-
-    Returns:
-        A tuple containing the configured pre-processor and post-processor pipelines.
-    """
-
-    # Add remaining processors
-    input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
-        # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
-        # because the tokenizer step expects normalized state in [-1, 1] range for discretization
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        Pi05PrepareStateTokenizerProcessorStep(max_state_dim=config.max_state_dim),
-        TokenizerProcessorStep(
-            tokenizer_name="google/paligemma-3b-pt-224",
-            max_length=config.tokenizer_max_length,
-            padding_side="right",
-            padding="max_length",
-        ),
-        DeviceProcessorStep(device=config.device),
-    ]
-
-    output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )

src/lerobot/policies/videovla/test_video_encoder.py

@@ -1,214 +0,0 @@
-#!/usr/bin/env python
-"""
-Test script for PI05 with video encoder (VideoPrism).
-
-This script creates a dummy example to test the model with video encoding enabled.
-"""
-
-import torch
-
-from lerobot.configs.types import FeatureType, PolicyFeature
-from lerobot.policies.videovla.configuration_pi05 import PI05VideoConfig
-from lerobot.policies.videovla.modeling_pi05 import PI05VideoPolicy
-from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
-
-
-def create_dummy_batch(
-    batch_size: int = 2,
-    num_frames: int = 16,
-    image_size: int = 224,
-    num_cameras: int = 2,
-    state_dim: int = 14,
-    action_dim: int = 14,
-    chunk_size: int = 50,
-    seq_len: int = 10,
-    device: str = "cuda",
-) -> dict[str, torch.Tensor]:
-    """Create a dummy batch for testing."""
-    batch = {}
-
-    # Create image observations with temporal dimension [B, T, C, H, W]
-    for i in range(num_cameras):
-        key = f"{OBS_IMAGES}.camera_{i}"
-        # Images in [0, 1] range
-        batch[key] = torch.rand(batch_size, num_frames, 3, image_size, image_size, device=device)
-
-    # Create state observation [B, state_dim]
-    batch[OBS_STATE] = torch.rand(batch_size, state_dim, device=device)
-
-    # Create language tokens and attention mask [B, seq_len]
-    batch["observation.language.tokens"] = torch.randint(0, 1000, (batch_size, seq_len), device=device)
-    batch["observation.language.attention_mask"] = torch.ones(batch_size, seq_len, dtype=torch.bool, device=device)
-
-    # Create action targets [B, chunk_size, action_dim]
-    batch[ACTION] = torch.rand(batch_size, chunk_size, action_dim, device=device)
-
-    return batch
-
-
-def test_video_encoder():
-    """Test the PI05 model with video encoding enabled."""
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    print(f"Using device: {device}")
-
-    # Configuration
-    batch_size = 2
-    num_frames = 16
-    image_size = 224
-    num_cameras = 2
-    state_dim = 14
-    action_dim = 14
-    chunk_size = 50
-
-    # Create config with video encoder enabled
-    print("Creating PI05VideoConfig with video encoder...")
-    config = PI05VideoConfig(
-        use_video_encoder=True,
-        video_num_frames=num_frames,
-        videoprism_model_name="MHRDYN7/videoprism-base-f16r288",
-        videoprism_image_size=288,
-        freeze_video_encoder=True,
-        video_padding_mode="repeat",
-        video_encoder_camera_key=f"{OBS_IMAGES}.camera_0",  # Use first camera for video
-        chunk_size=chunk_size,
-        max_action_dim=32,
-        max_state_dim=32,
-        dtype="float32",  # Use float32 for testing
-        device=device,
-    )
-
-    # Set up input/output features
-    for i in range(num_cameras):
-        key = f"{OBS_IMAGES}.camera_{i}"
-        config.input_features[key] = PolicyFeature(
-            type=FeatureType.VISUAL,
-            shape=(3, image_size, image_size),
-        )
-
-    config.input_features[OBS_STATE] = PolicyFeature(
-        type=FeatureType.STATE,
-        shape=(state_dim,),
-    )
-
-    config.output_features[ACTION] = PolicyFeature(
-        type=FeatureType.ACTION,
-        shape=(action_dim,),
-    )
-
-    print(f"use_video_encoder: {config.use_video_encoder}")
-    print(f"video_num_frames: {config.video_num_frames}")
-    print(f"video_padding_mode: {config.video_padding_mode}")
-    print(f"video_encoder_camera_key: {config.video_encoder_camera_key}")
-    print(f"image_observation_delta_indices: {config.image_observation_delta_indices}")
-
-    # Create model
-    model = PI05VideoPolicy(config)
-    model.to(device)
-
-    # Create dummy batch
-    batch = create_dummy_batch(
-        batch_size=batch_size,
-        num_frames=num_frames,
-        image_size=image_size,
-        num_cameras=num_cameras,
-        state_dim=state_dim,
-        action_dim=action_dim,
-        chunk_size=chunk_size,
-        device=device,
-    )
-
-    print(f"Batch keys: {list(batch.keys())}"  )
-    for key, value in batch.items():
-        print(f"{key}: {value.shape}")
-
-    # Test forward pass
-    model.train()
-    try:
-        loss, loss_dict = model.forward(batch)
-        print(f"Forward pass successful!")
-        print(f"Loss: {loss.item():.4f}")
-        print(f"Loss dict: {loss_dict}")
-    except Exception as e:
-        print(f"Forward pass failed: {e}")
-        raise
-
-    # Test inference
-    model.eval()
-    with torch.no_grad():
-        try:
-            actions = model.predict_action_chunk(batch)
-            print(f"Test pass, inference pass!")
-            print(f"Predicted actions shape: {actions.shape}")
-        except Exception as e:
-            print(f"Inference failed: {e}")
-            raise
-
-    print("All tests passed!")
-
-
-def test_frame_padding():
-    """Test frame padding at episode start."""
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-
-    # Create config
-    config = PI05VideoConfig(
-        use_video_encoder=True,
-        video_num_frames=16,
-        videoprism_model_name="MHRDYN7/videoprism-base-f16r288",
-        freeze_video_encoder=True,
-        video_padding_mode="repeat",
-        chunk_size=50,
-        dtype="float32",
-        device=device,
-    )
-
-    # Set up minimal features
-    config.input_features[f"{OBS_IMAGES}.camera_0"] = PolicyFeature(
-        type=FeatureType.VISUAL,
-        shape=(3, 224, 224),
-    )
-    config.output_features[ACTION] = PolicyFeature(
-        type=FeatureType.ACTION,
-        shape=(14,),
-    )
-
-    # Create model
-    model = PI05VideoPolicy(config)
-    model.to(device)
-
-    # Test with fewer frames than expected (simulating episode start)
-    batch = {
-        f"{OBS_IMAGES}.camera_0": torch.rand(2, 5, 3, 224, 224, device=device),
-        "observation.language.tokens": torch.randint(0, 1000, (2, 10), device=device),
-        "observation.language.attention_mask": torch.ones(2, 10, dtype=torch.bool, device=device),
-        ACTION: torch.rand(2, 50, 14, device=device),
-    }
-
-    video_frames = model._preprocess_video(batch)
-    if video_frames is not None:
-        print(f"Input frames: 5")
-        print(f"Output video_frames shape: {video_frames.shape}")
-        print(f"Expected: [2, 16, 3, 224, 224]")
-        assert video_frames.shape == (2, 16, 3, 224, 224), f"Unexpected shape: {video_frames.shape}"
-        print("Frame padding test PASSED!")
-    else:
-        print("video_frames is None (unexpected)")
-
-    # Test with single frame
-    batch[f"{OBS_IMAGES}.camera_0"] = torch.rand(2, 3, 224, 224, device=device)  # [B, C, H, W]
-
-    video_frames = model._preprocess_video(batch)
-    if video_frames is not None:
-        print(f"Input: single frame [B, C, H, W]")
-        print(f"Output video_frames shape: {video_frames.shape}")
-        print(f"Expected: [2, 16, 3, 224, 224]")
-        assert video_frames.shape == (2, 16, 3, 224, 224), f"Unexpected shape: {video_frames.shape}"
-        print("Single frame expansion test PASSED!")
-    else:
-        print("video_frames is None (unexpected)")
-
-    print("All tests passed!")
-if __name__ == "__main__":
-    # Run tests
-    test_frame_padding()
-    test_video_encoder()

src/lerobot/policies/videovla/videoprism/__init__.py

@@ -1,37 +0,0 @@
-# Copyright 2025 The HuggingFace 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 .configuration_videoprism import VideoPrismConfig, VideoPrismTextConfig, VideoPrismVisionConfig
-from .modeling_videoprism import (
-    VideoPrismClipModel,
-    VideoPrismForVideoClassification,
-    VideoPrismPreTrainedModel,
-    VideoPrismTextModel,
-    VideoPrismVideoModel,
-    VideoPrismVisionModel,
-)
-from .video_processing_videoprism import VideoPrismVideoProcessor
-
-__all__ = [
-    "VideoPrismConfig",
-    "VideoPrismTextConfig",
-    "VideoPrismVisionConfig",
-    "VideoPrismClipModel",
-    "VideoPrismForVideoClassification",
-    "VideoPrismPreTrainedModel",
-    "VideoPrismTextModel",
-    "VideoPrismVideoModel",
-    "VideoPrismVisionModel",
-    "VideoPrismVideoProcessor",
-]

src/lerobot/policies/videovla/videoprism/configuration_videoprism.py

@@ -1,269 +0,0 @@
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-#           This file was automatically generated from src/transformers/models/videoprism/modular_videoprism.py.
-#               Do NOT edit this file manually as any edits will be overwritten by the generation of
-#             the file from the modular. If any change should be done, please apply the change to the
-#                          modular_videoprism.py file directly. One of our CI enforces this.
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-from transformers import PretrainedConfig
-from transformers.utils import logging
-
-
-logger = logging.get_logger(__name__)
-
-
-class VideoPrismVisionConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`VideoPrismVisionModel`]. It is used to instantiate a
-    VideoPrism vision encoder according to the specified arguments, defining the model architecture. Instantiating a
-    configuration with the defaults will yield a similar configuration to that of the VideoPrism
-    [google/videoprism](https://huggingface.co/google/videoprism) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        image_size (`int`, *optional*, defaults to 288):
-            The size of the input image.
-        num_frames (`int`, *optional*, defaults to 16):
-            The number of frames in the input video.
-        tubelet_size (`List[int]`, *optional*, defaults to `[1, 18, 18]`):
-            The size of the tubelet patch.
-        num_channels (`int`, *optional*, defaults to 3):
-            The number of input channels.
-        hidden_size (`int`, *optional*, defaults to 768):
-            Dimensionality of the encoder layers and the pooler layer.
-        num_spatial_layers (`int`, *optional*, defaults to 12):
-            Number of spatial transformer blocks.
-        num_temporal_layers (`int`, *optional*, defaults to 4):
-            Number of temporal transformer blocks.
-        num_attention_heads (`int`, *optional*, defaults to 12):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        intermediate_size (`int`, *optional*, defaults to 3072):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_python"`):
-            The non-linear activation function (function or string).
-        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
-            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
-        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-        initializer_range (`float`, *optional*, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the layer normalization layers.
-        qkv_bias (`bool`, *optional*, defaults to `True`):
-            Whether to add a bias to the qkv projections in attention layers.
-        attn_logit_softcapping (`float`, *optional*, defaults to 50.0):
-            Softcapping constant for attention logits.
-        num_auxiliary_layers (`int`, *optional*, defaults to 2):
-            Number of auxiliary layers. This is used in the VideoPrismVideoModel that is a part of VideoPrismClipModel.
-        apply_l2_norm (`bool`, *optional*, defaults to `True`):
-            Whether to apply L2 normalization to the output. This is used in the VideoPrismVideoModel that is a part of VideoPrismClipModel.
-
-    Example:
-
-    ```python
-    >>> from transformers import VideoPrismVisionConfig, VideoPrismVisionModel
-
-    >>> # Initializing a VideoPrismVisionConfig with default values
-    >>> configuration = VideoPrismVisionConfig()
-
-    >>> # Initializing a VideoPrismVisionModel with the configuration
-    >>> model = VideoPrismVisionModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "videoprism_vision_model"
-    base_config_key = "vision_config"
-
-    def __init__(
-        self,
-        image_size=288,
-        num_frames=16,
-        tubelet_size=[1, 18, 18],
-        num_channels=3,
-        hidden_size=768,
-        num_spatial_layers=12,
-        num_temporal_layers=4,
-        num_attention_heads=12,
-        intermediate_size=3072,
-        hidden_act="gelu_python",
-        hidden_dropout_prob=0.0,
-        attention_probs_dropout_prob=0.0,
-        initializer_range=0.02,
-        layer_norm_eps=1e-06,
-        qkv_bias=True,
-        attn_logit_softcapping=50.0,
-        num_auxiliary_layers=2,
-        apply_l2_norm=True,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-        self.hidden_size = hidden_size
-        self.num_attention_heads = num_attention_heads
-        self.intermediate_size = intermediate_size
-        self.hidden_act = hidden_act
-        self.hidden_dropout_prob = hidden_dropout_prob
-        self.attention_probs_dropout_prob = attention_probs_dropout_prob
-        self.initializer_range = initializer_range
-        self.layer_norm_eps = layer_norm_eps
-
-        self.image_size = image_size
-        self.num_frames = num_frames
-        self.tubelet_size = tubelet_size
-        self.num_channels = num_channels
-        self.qkv_bias = qkv_bias
-        self.num_spatial_layers = num_spatial_layers
-        self.num_temporal_layers = num_temporal_layers
-        self.attn_logit_softcapping = attn_logit_softcapping
-        self.num_auxiliary_layers = num_auxiliary_layers
-        self.apply_l2_norm = apply_l2_norm
-
-
-class VideoPrismTextConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`VideoPrismTextModel`]. It is used to instantiate a
-    VideoPrism text encoder according to the specified arguments, defining the model architecture. Instantiating a
-    configuration with the defaults will yield a similar configuration to that of the VideoPrism
-    [google/videoprism](https://huggingface.co/google/videoprism) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        hidden_size (`int`, *optional*, defaults to 768):
-            Dimensionality of the encoder layers and the pooler layer.
-        intermediate_size (`int`, *optional*, defaults to 3072):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 12):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        num_text_layers (`int`, *optional*, defaults to 12):
-            Number of hidden layers in the text Transformer encoder.
-        vocab_size (`int`, *optional*, defaults to 32000):
-            Vocabulary size of the text model. Defines the number of different tokens that can be represented by the
-            `input_ids` passed when calling [`VideoPrismTextModel`].
-        apply_l2_norm (`bool`, *optional*, defaults to `True`):
-            Whether to apply L2 normalization to the output text embeddings.
-        hidden_act (`str` or `function`, *optional*, defaults to `"relu"`):
-            The non-linear activation function (function or string) in the encoder and pooler.
-        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-        qkv_bias (`bool`, *optional*, defaults to `True`):
-            Whether to add a bias to the query, key, and value projections in the attention layers.
-        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
-            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
-        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the layer normalization layers.
-        initializer_range (`float`, *optional*, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        attn_logit_softcapping (`float`, *optional*, defaults to 50.0):
-            Softcapping constant for attention logits.
-
-    Example:
-
-    ```python
-    >>> from transformers import VideoPrismTextConfig, VideoPrismTextModel
-
-    >>> # Initializing a VideoPrismTextConfig with default values
-    >>> configuration = VideoPrismTextConfig()
-
-    >>> # Initializing a VideoPrismTextModel (with random weights) from the configuration
-    >>> model = VideoPrismTextModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "videoprism_text_model"
-    base_config_key = "text_config"
-
-    def __init__(
-        self,
-        hidden_size=768,
-        intermediate_size=3072,
-        num_attention_heads=12,
-        num_text_layers=12,
-        vocab_size=32000,
-        apply_l2_norm=True,
-        hidden_act="relu",
-        attention_probs_dropout_prob=0.0,
-        qkv_bias=True,
-        hidden_dropout_prob=0.0,
-        layer_norm_eps=1e-06,
-        initializer_range=0.02,
-        attn_logit_softcapping=50.0,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_attention_heads = num_attention_heads
-        self.num_text_layers = num_text_layers
-        self.vocab_size = vocab_size
-        self.apply_l2_norm = apply_l2_norm
-        self.hidden_act = hidden_act
-        self.attention_probs_dropout_prob = attention_probs_dropout_prob
-        self.qkv_bias = qkv_bias
-        self.hidden_dropout_prob = hidden_dropout_prob
-        self.layer_norm_eps = layer_norm_eps
-        self.initializer_range = initializer_range
-        self.attn_logit_softcapping = attn_logit_softcapping
-
-
-class VideoPrismConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`VideoPrismModel`]. It is used to instantiate a
-    VideoPrism model according to the specified arguments, defining the model architecture. Instantiating a
-    configuration with the defaults will yield a similar configuration to that of the VideoPrism
-    [google/videoprism](https://huggingface.co/google/videoprism) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        text_config (`VideoPrismTextConfig`, *optional*):
-            Configuration for the text model.
-        vision_config (`VideoPrismVisionConfig`, *optional*):
-            Configuration for the vision model.
-        kwargs (*optional*):
-            Dictionary of keyword arguments.
-
-    Example:
-
-    ```python
-    >>> from transformers import VideoPrismConfig, VideoPrismModel
-
-    >>> # Initializing a VideoPrismConfig with default values
-    >>> configuration = VideoPrismConfig()
-
-    >>> # Initializing a VideoPrismClipModel with the configuration
-    >>> model = VideoPrismClipModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "videoprism"
-    sub_configs = {"text_config": VideoPrismTextConfig, "vision_config": VideoPrismVisionConfig}
-
-    def __init__(self, text_config=None, vision_config=None, **kwargs):
-        if text_config is None:
-            text_config = VideoPrismTextConfig()
-            logger.info("`text_config` is `None`. Initializing the `VideoPrismTextConfig` with default values.")
-        elif isinstance(text_config, dict):
-            text_config = VideoPrismTextConfig(**text_config)
-
-        if vision_config is None:
-            vision_config = VideoPrismVisionConfig()
-            logger.info("`vision_config` is `None`. initializing the `VideoPrismVisionConfig` with default values.")
-        elif isinstance(vision_config, dict):
-            vision_config = VideoPrismVisionConfig(**vision_config)
-
-        self.text_config = text_config
-        self.vision_config = vision_config
-
-        super().__init__(**kwargs)
-
-
-__all__ = ["VideoPrismVisionConfig", "VideoPrismTextConfig", "VideoPrismConfig"]

src/lerobot/policies/videovla/videoprism/initialization.py

@@ -1,245 +0,0 @@
-# Copyright 2025 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import sys
-from collections import defaultdict
-from contextlib import contextmanager
-
-import torch
-
-
-# Record all the torch primitives in advance, so that we can use them without them being modified when we patch torch
-# in context managers
-TORCH_INIT_FUNCTIONS = {
-    "uniform_": torch.nn.init.uniform_,
-    "normal_": torch.nn.init.normal_,
-    "constant_": torch.nn.init.constant_,
-    "ones_": torch.nn.init.ones_,
-    "zeros_": torch.nn.init.zeros_,
-    "eye_": torch.nn.init.eye_,
-    "dirac_": torch.nn.init.dirac_,
-    "xavier_uniform_": torch.nn.init.xavier_uniform_,
-    "xavier_normal_": torch.nn.init.xavier_normal_,
-    "kaiming_uniform_": torch.nn.init.kaiming_uniform_,
-    "kaiming_normal_": torch.nn.init.kaiming_normal_,
-    "trunc_normal_": torch.nn.init.trunc_normal_,
-    "orthogonal_": torch.nn.init.orthogonal_,
-    "sparse_": torch.nn.init.sparse_,
-}
-
-
-def uniform_(
-    tensor: torch.Tensor, a: float = 0.0, b: float = 1.0, generator: torch.Generator | None = None
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["uniform_"](tensor, a=a, b=b, generator=generator)
-    return tensor
-
-
-def normal_(
-    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, generator: torch.Generator | None = None
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["normal_"](tensor, mean=mean, std=std, generator=generator)
-    return tensor
-
-
-def constant_(tensor: torch.Tensor, val: float) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["constant_"](tensor, val=val)
-    return tensor
-
-
-def ones_(tensor: torch.Tensor) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["ones_"](tensor)
-    return tensor
-
-
-def zeros_(tensor: torch.Tensor) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["zeros_"](tensor)
-    return tensor
-
-
-def eye_(tensor: torch.Tensor) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["eye_"](tensor)
-    return tensor
-
-
-def dirac_(tensor: torch.Tensor, groups: int = 1) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["dirac_"](tensor, groups=groups)
-    return tensor
-
-
-def xavier_uniform_(tensor: torch.Tensor, gain: float = 1.0, generator: torch.Generator | None = None) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["xavier_uniform_"](tensor, gain=gain, generator=generator)
-    return tensor
-
-
-def xavier_normal_(tensor: torch.Tensor, gain: float = 1.0, generator: torch.Generator | None = None) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["xavier_normal_"](tensor, gain=gain, generator=generator)
-    return tensor
-
-
-def kaiming_uniform_(
-    tensor: torch.Tensor,
-    a: float = 0,
-    mode: str = "fan_in",
-    nonlinearity: str = "leaky_relu",
-    generator: torch.Generator | None = None,
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["kaiming_uniform_"](
-            tensor, a=a, mode=mode, nonlinearity=nonlinearity, generator=generator
-        )
-    return tensor
-
-
-def kaiming_normal_(
-    tensor: torch.Tensor,
-    a: float = 0,
-    mode: str = "fan_in",
-    nonlinearity: str = "leaky_relu",
-    generator: torch.Generator | None = None,
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["kaiming_normal_"](
-            tensor, a=a, mode=mode, nonlinearity=nonlinearity, generator=generator
-        )
-    return tensor
-
-
-def trunc_normal_(
-    tensor: torch.Tensor,
-    mean: float = 0.0,
-    std: float = 1.0,
-    a: float = -2.0,
-    b: float = 2.0,
-    generator: torch.Generator | None = None,
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["trunc_normal_"](tensor, mean=mean, std=std, a=a, b=b, generator=generator)
-    return tensor
-
-
-def orthogonal_(
-    tensor: torch.Tensor,
-    gain: float = 1,
-    generator: torch.Generator | None = None,
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["orthogonal_"](tensor, gain=gain, generator=generator)
-    return tensor
-
-
-def sparse_(
-    tensor: torch.Tensor, sparsity: float, std: float = 0.01, generator: torch.Generator | None = None
-) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        return TORCH_INIT_FUNCTIONS["sparse_"](tensor, sparsity=sparsity, std=std, generator=generator)
-    return tensor
-
-
-def copy_(tensor: torch.Tensor, other: torch.Tensor) -> torch.Tensor:
-    if not getattr(tensor, "_is_hf_initialized", False):
-        with torch.no_grad():
-            return tensor.copy_(other)
-    return tensor
-
-
-# Here, we need to check several modules imported, and hot patch all of them, as sometimes torch does
-# something like `from torch.nn.init import xavier_uniform_` in their internals (e.g in torch.nn.modules.activations,
-# where MultiHeadAttention lives), so the function name is binded at import time and just doing
-# `setattr(torch.nn.init, name, globals()[name])` is thus not enough
-# The following list should be enough for all torch versions we work with
-TORCH_MODULES_TO_PATCH = (
-    "torch.nn.init",
-    "torch.nn.modules.activation",
-    "torch.nn.modules.transformer",
-    "torch.nn.modules.linear",
-    "torch.nn.modules.loss",
-    "torch.nn.modules.batchnorm",
-    "torch.nn.modules.conv",
-    "torch.nn.modules.normalization",
-    "torch.nn.modules.rnn",
-    "torch.nn.modules.sparse",
-)
-
-
-@contextmanager
-def guard_torch_init_functions():
-    """
-    Guard the `torch.nn.init` primitive functions to behave exactly like the functions in this file, i.e. be
-    protected against the `_is_hf_initialized` flag to avoid re-init if the param was already loaded.
-
-    Usually, all models are using the init from `transformers` which are already guarded, but just to make extra sure
-    and for remote code, we also use this context manager.
-    """
-    originals = defaultdict(dict)
-    try:
-        # Replace all torch funcs by the ones in this file
-        for module_name in TORCH_MODULES_TO_PATCH:
-            if module_name in sys.modules:
-                module = sys.modules[module_name]
-                for func_name in TORCH_INIT_FUNCTIONS.keys():
-                    if hasattr(module, func_name):
-                        originals[module][func_name] = getattr(module, func_name)
-                        setattr(module, func_name, globals()[func_name])
-        yield
-    finally:
-        # Set back the original functions on all modules
-        for module, functions in originals.items():
-            for func_name, func in functions.items():
-                setattr(module, func_name, func)
-
-
-@contextmanager
-def no_init_weights():
-    """
-    Disable weight initialization both at the torch-level, and at the transformers-level (`init_weights`).
-    This is used to speed-up initializing an empty model with deepspeed, as we do not initialize the model on meta device
-    with deepspeed, but we still don't need to run expensive weight initializations as we are loading params afterwards.
-    """
-    from .modeling_utils import PreTrainedModel
-
-    def empty_func(*args, **kwargs):
-        pass
-
-    originals = defaultdict(dict)
-    try:
-        # Replace all torch funcs by empty ones
-        for module_name in TORCH_MODULES_TO_PATCH:
-            if module_name in sys.modules:
-                module = sys.modules[module_name]
-                for func_name in TORCH_INIT_FUNCTIONS.keys():
-                    if hasattr(module, func_name):
-                        originals[module][func_name] = getattr(module, func_name)
-                        setattr(module, func_name, empty_func)
-
-        # Also patch our own `init_weights`
-        original_init_weights = PreTrainedModel.init_weights
-        PreTrainedModel.init_weights = empty_func
-
-        yield
-    finally:
-        # Set back the original torch functions on all modules
-        for module, functions in originals.items():
-            for func_name, func in functions.items():
-                setattr(module, func_name, func)
-        # Set back `init_weights`
-        PreTrainedModel.init_weights = original_init_weights

src/lerobot/policies/videovla/videoprism/modeling_videoprism.py

@@ -1,994 +0,0 @@
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-#           This file was automatically generated from src/transformers/models/videoprism/modular_videoprism.py.
-#               Do NOT edit this file manually as any edits will be overwritten by the generation of
-#             the file from the modular. If any change should be done, please apply the change to the
-#                          modular_videoprism.py file directly. One of our CI enforces this.
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-import math
-from collections.abc import Callable
-from dataclasses import dataclass
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from torch.nn.init import _calculate_fan_in_and_fan_out
-
-from . import initialization as init
-from transformers.activations import ACT2FN
-from transformers.masking_utils import create_causal_mask
-from transformers.modeling_layers import GradientCheckpointingLayer
-from transformers.modeling_outputs import BaseModelOutput, ImageClassifierOutput
-from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
-from transformers.file_utils import ModelOutput
-
-from .configuration_videoprism import VideoPrismConfig, VideoPrismTextConfig, VideoPrismVisionConfig
-
-def torch_int(x):
-    """
-    Casts an input to a torch int64 tensor if we are in a tracing context, otherwise to a Python int.
-    """
-    if not torch.is_available():
-        return int(x)
-
-    return x.to(torch.int64) if torch.jit.is_tracing() and isinstance(x, torch.Tensor) else int(x)
-
-@dataclass
-class BaseModelOutputWithSpatialAndTemporalStates(ModelOutput):
-    """
-    Base class for model outputs that include spatial and temporal states.
-
-    Args:
-        last_hidden_state (Optional[torch.FloatTensor]):
-            The last hidden state of the model, typically of shape
-            (batch_size, num_patches * num_frames, hidden_size).
-
-        temporal_hidden_state (Optional[torch.FloatTensor]):
-            The last hidden_state of the temporal encoder, typically of shape
-            (batch_size * num_patches, num_frames, hidden_size).
-
-        spatial_hidden_state (Optional[torch.FloatTensor]):
-            The last hidden_state of the spatial encoder, typically of shape
-            (batch_size * num_frames, num_patches, hidden_size).
-    """
-
-    last_hidden_state: torch.FloatTensor | None = None
-    temporal_hidden_state: torch.FloatTensor | None = None
-    spatial_hidden_state: torch.FloatTensor | None = None
-
-
-@dataclass
-class VideoPrismClipOutput(ModelOutput):
-    """
-    Base class for VideoPrismClip model outputs.
-    """
-
-    logits_per_video: torch.FloatTensor | None = None
-    logits_per_text: torch.FloatTensor | None = None
-    video_embeds: torch.FloatTensor | None = None
-    text_embeds: torch.FloatTensor | None = None
-
-
-@dataclass
-class VideoPrismVideoOutput(ModelOutput):
-    """
-    Base class for VideoPrismVideo model outputs.
-    """
-
-    video_last_hidden_state: torch.FloatTensor | None = None
-    auxiliary_output: torch.FloatTensor | None = None
-    attention_pooling_output: torch.FloatTensor | None = None
-
-
-class VideoPrismTubeletEmbeddings(nn.Module):
-    """
-    Construct VideoPrism Tubelet embeddings.
-
-    This module turns a batch of videos of shape (batch_size, num_frames, num_channels, height, width) into a tensor of
-    shape (batch_size, seq_len, hidden_size) to be consumed by a Transformer encoder.
-
-    The seq_len (the number of patches) equals (number of frames // tubelet_size[0]) * (height // tubelet_size[1]) *
-    (width // tubelet_size[2]).
-    """
-
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-        self.num_frames = config.num_frames
-        self.image_size = (
-            config.image_size
-            if isinstance(self.config.image_size, tuple)
-            else (self.config.image_size, self.config.image_size)
-        )
-        self.patch_size = config.tubelet_size
-        self.embed_dim = config.hidden_size
-
-        self.projection = nn.Conv3d(
-            config.num_channels, config.hidden_size, kernel_size=config.tubelet_size, stride=config.tubelet_size
-        )
-        self.pos_emb_shape = [self.image_size[0] // self.patch_size[1], self.image_size[1] // self.patch_size[2]]
-        self.num_patches = self.pos_emb_shape[0] * self.pos_emb_shape[1]
-
-    def forward(self, pixel_values_videos: torch.Tensor, interpolate_pos_encoding: bool = False) -> torch.Tensor:
-        batch_size, num_frames, num_channels, height, width = pixel_values_videos.shape
-        if not interpolate_pos_encoding and (height != self.image_size[0] or width != self.image_size[1]):
-            raise ValueError(
-                f"Image size ({height}*{width}) doesn't match model ({self.image_size[0]}*{self.image_size[1]}). Set interpolate_pos_encoding=True to automatically resize the model position embeddings."
-            )
-        # permute to (batch_size, num_channels, num_frames, height, width)
-        pixel_values_videos = pixel_values_videos.permute(0, 2, 1, 3, 4)
-
-        hidden_states = self.projection(pixel_values_videos)
-        # flatten the spatial part and permute to (B, T, num_patches, dim)
-        hidden_states = hidden_states.flatten(3).permute(0, 2, 3, 1)
-        # combine batch and time dimension
-        batch_size, num_frames, num_patches, hidden_size = hidden_states.shape
-        hidden_states = hidden_states.reshape(batch_size * num_frames, num_patches, hidden_size)
-
-        return hidden_states
-
-
-class VideoPrismSpatialEmbeddings(nn.Module):
-    """
-    VideoPrism Spatial Embeddings.
-
-    Creates embeddings from a video using VideoPrismSpatialTubeletEmbeddings and adds positional embeddings.
-    """
-
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-        self.patch_embeddings = VideoPrismTubeletEmbeddings(config)
-        self.position_embeddings = nn.Parameter(torch.zeros(1, self.patch_embeddings.num_patches, config.hidden_size))
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.patch_size = config.tubelet_size[1:]
-        self.tubelet_size = config.tubelet_size
-
-    # Adapted from transformers.models.vit.modeling_vit.ViTEmbeddings.interpolate_pos_encoding
-    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
-        """
-        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
-        images. This method is also adapted to support torch.jit tracing.
-
-        Adapted from:
-        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
-        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
-        """
-
-        num_patches = embeddings.shape[1]
-        num_positions = self.position_embeddings.shape[1]
-
-        # always interpolate when tracing to ensure the exported model works for dynamic input shapes
-        if not torch.jit.is_tracing() and num_patches == num_positions and height == width:
-            return self.position_embeddings
-
-        dim = embeddings.shape[-1]
-
-        num_row_patches = height // self.patch_size[0]
-        num_col_patches = width // self.patch_size[1]
-
-        sqrt_num_positions = torch_int(num_positions**0.5)
-        patch_pos_embed = self.position_embeddings.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
-        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
-
-        patch_pos_embed = nn.functional.interpolate(
-            patch_pos_embed,
-            size=(num_row_patches, num_col_patches),
-            mode="bilinear",
-            antialias=True,
-        )
-
-        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
-        return patch_pos_embed
-
-    def forward(
-        self, pixel_values_videos: torch.Tensor, interpolate_pos_encoding: bool | None = False
-    ) -> torch.Tensor:
-        b, t, c, h, w = pixel_values_videos.shape
-        assert h == w, "Input image height and width must be the same"
-        embeddings = self.patch_embeddings(pixel_values_videos, interpolate_pos_encoding)
-
-        # add positional encoding to each token
-        if interpolate_pos_encoding:
-            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, h, w)
-        else:
-            embeddings = embeddings + self.position_embeddings
-
-        embeddings = self.dropout(embeddings)
-
-        return embeddings
-
-
-class VideoPrismTemporalEmbeddings(nn.Module):
-    """
-    VideoPrism Temporal Embeddings.
-
-    Receives embeddings from spatial encoder, reshapes the hidden state to
-    (batch_size * num_patches, num_frames, hidden_size) and adds positional embeddings.
-    """
-
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-
-        self.position_embeddings = nn.Parameter(torch.zeros(1, self.config.num_frames, config.hidden_size))
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-
-    # Adapted from transformers.models.vit.modeling_vit.ViTEmbeddings.interpolate_pos_encoding
-    def interpolate_pos_encoding(self, embeddings: torch.Tensor) -> torch.Tensor:
-        """
-        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
-        images. This method is also adapted to support torch.jit tracing.
-
-        Adapted from:
-        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
-        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
-        """
-        target_emb_length = embeddings.shape[1]
-        source_emb_length = self.position_embeddings.shape[1]
-
-        # always interpolate when tracing to ensure the exported model works for dynamic input shapes
-        if not torch.jit.is_tracing() and target_emb_length == source_emb_length:
-            return self.position_embeddings
-
-        source_emb = self.position_embeddings
-        dim = embeddings.shape[-1]
-        source_emb = source_emb.unsqueeze(1)
-        source_emb = nn.functional.interpolate(
-            source_emb,
-            size=(target_emb_length, dim),
-            mode="bilinear",
-            antialias=True,
-        )
-
-        return source_emb.squeeze(1)
-
-    def forward(
-        self,
-        pixel_values_videos: torch.Tensor,
-        input_shape: torch.Size,
-        interpolate_pos_encoding: bool | None = False,
-    ) -> torch.Tensor:
-        if input_shape is not None:
-            b, t, c, h, w = input_shape
-        _, features, dim = pixel_values_videos.shape
-        hidden_states = pixel_values_videos.view(b, t, features, dim)
-        hidden_states = hidden_states.permute(0, 2, 1, 3)
-        embeddings = hidden_states.reshape(b * features, t, dim)
-
-        # add positional encoding to each token
-        if interpolate_pos_encoding:
-            embeddings = embeddings + self.interpolate_pos_encoding(embeddings)
-        else:
-            embeddings = embeddings + self.position_embeddings
-        embeddings = self.dropout(embeddings)
-        return embeddings
-
-
-def eager_attention_forward(
-    module: nn.Module,
-    query: torch.Tensor,
-    key: torch.Tensor,
-    value: torch.Tensor,
-    attention_mask: torch.Tensor | None,
-    scaling: float,
-    dropout: float = 0.0,
-    softcap: float | None = None,
-    **kwargs,
-):
-    # Take the dot product between "query" and "key" to get the raw attention scores.
-    attn_weights = torch.matmul(query, key.transpose(-1, -2)) * scaling
-
-    if softcap is not None:
-        attn_weights = attn_weights / softcap
-        attn_weights = torch.tanh(attn_weights)
-        attn_weights = attn_weights * softcap
-    if attention_mask is not None:
-        attn_weights = attn_weights + attention_mask.expand(*attn_weights.shape)
-
-    # Normalize the attention scores to probabilities.
-    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
-    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
-    attn_output = torch.matmul(attn_weights, value)
-    attn_output = attn_output.transpose(1, 2).contiguous()
-    return attn_output, attn_weights
-
-
-class VideoPrismSelfAttention(nn.Module):
-    def __init__(self, config: VideoPrismVisionConfig | VideoPrismTextConfig):
-        super().__init__()
-        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
-            raise ValueError(
-                f"The hidden size {config.hidden_size} is not a multiple of the number of attention "
-                f"heads {config.num_attention_heads}."
-            )
-
-        self.config = config
-        self.num_attention_heads = config.num_attention_heads
-        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
-        self.all_head_size = self.num_attention_heads * self.attention_head_size
-        self.dropout_prob = config.attention_probs_dropout_prob
-        self.scale = self.attention_head_size**-0.5
-        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
-        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
-        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None,
-        **kwargs,
-    ) -> tuple[torch.Tensor, torch.Tensor]:
-        batch_size = hidden_states.shape[0]
-        new_shape = batch_size, -1, self.num_attention_heads, self.attention_head_size
-        query = self.query(hidden_states).view(*new_shape).transpose(1, 2)
-        key = self.key(hidden_states).view(*new_shape).transpose(1, 2)
-        value = self.value(hidden_states).view(*new_shape).transpose(1, 2)
-
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
-
-        context_layer, attention_probs = attention_interface(
-            self,
-            query,
-            key,
-            value,
-            attention_mask,
-            scaling=self.scale,
-            dropout=0.0 if not self.training else self.dropout_prob,
-            softcap=self.config.attn_logit_softcapping,
-            **kwargs,
-        )
-
-        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
-        context_layer = context_layer.reshape(new_context_layer_shape)
-
-        return (context_layer, attention_probs)
-
-
-class VideoPrismSelfOutput(nn.Module):
-    """
-    The residual connection is defined in VideoPrismLayer instead of here (as is the case with other models), due to the
-    layernorm applied before each block.
-    """
-
-    def __init__(self, config: VideoPrismConfig):
-        super().__init__()
-        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-
-    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
-        hidden_states = self.dense(hidden_states)
-        hidden_states = self.dropout(hidden_states)
-        return hidden_states
-
-
-class VideoPrismAttention(nn.Module):
-    def __init__(self, config: VideoPrismConfig):
-        super().__init__()
-        self.attention = VideoPrismSelfAttention(config)
-        self.output = VideoPrismSelfOutput(config)
-
-    def forward(
-        self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, **kwargs
-    ) -> torch.Tensor:
-        self_attn_output, _ = self.attention(hidden_states, attention_mask, **kwargs)
-        output = self.output(self_attn_output, hidden_states)
-        return output
-
-
-class VideoPrismLayerNorm(nn.LayerNorm):
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        return F.layer_norm(hidden_states, self.normalized_shape, self.weight + 1, self.bias, self.eps)
-
-
-class VideoPrismIntermediate(nn.Module):
-    def __init__(self, config: VideoPrismConfig):
-        super().__init__()
-        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        if isinstance(config.hidden_act, str):
-            self.intermediate_act_fn = ACT2FN[config.hidden_act]
-        else:
-            self.intermediate_act_fn = config.hidden_act
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        hidden_states = self.dense(hidden_states)
-        hidden_states = self.intermediate_act_fn(hidden_states)
-        hidden_states = self.dropout(hidden_states)
-
-        return hidden_states
-
-
-class VideoPrismOutput(nn.Module):
-    def __init__(self, config: VideoPrismConfig):
-        super().__init__()
-        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-
-    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
-        hidden_states = self.dense(hidden_states)
-        hidden_states = self.dropout(hidden_states)
-        hidden_states = hidden_states + input_tensor
-        return hidden_states
-
-
-class VideoPrismLayer(GradientCheckpointingLayer):
-    """This corresponds to the EncoderBlock class in the scenic/videoprism implementation."""
-
-    def __init__(self, config: VideoPrismVisionConfig | VideoPrismTextConfig):
-        super().__init__()
-        self.config = config
-        self.attention = VideoPrismAttention(config)
-        self.intermediate = VideoPrismIntermediate(config)
-        self.output = VideoPrismOutput(config)
-        self.layernorm_before = VideoPrismLayerNorm(self.config.hidden_size, eps=self.config.layer_norm_eps)
-        self.layernorm_after = VideoPrismLayerNorm(self.config.hidden_size, eps=self.config.layer_norm_eps)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        **kwargs,
-    ) -> torch.Tensor:
-        hidden_states_norm = self.layernorm_before(hidden_states)
-        attention_output = self.attention(hidden_states_norm, attention_mask, **kwargs)
-
-        # first residual connection
-        hidden_states = attention_output + hidden_states
-
-        # in VideoPrism, layernorm is also applied after self-attention
-        layer_output = self.layernorm_after(hidden_states)
-        layer_output = self.intermediate(layer_output)
-
-        # second residual connection is done here
-        layer_output = self.output(layer_output, hidden_states)
-
-        return layer_output
-
-
-class VideoPrismSpatialEncoder(nn.Module):
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-        self.layer = nn.ModuleList([VideoPrismLayer(config) for _ in range(config.num_spatial_layers)])
-        self.gradient_checkpointing = False
-
-    def forward(self, hidden_states: torch.Tensor) -> BaseModelOutput:
-        for i, layer_module in enumerate(self.layer):
-            hidden_states = layer_module(hidden_states)
-
-        return BaseModelOutput(last_hidden_state=hidden_states)
-
-
-class VideoPrismTemporalEncoder(nn.Module):
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-        self.layer = nn.ModuleList([VideoPrismLayer(config) for _ in range(config.num_temporal_layers)])
-        self.gradient_checkpointing = False
-
-    def forward(self, hidden_states: torch.Tensor) -> BaseModelOutput:
-        for i, layer_module in enumerate(self.layer):
-            hidden_states = layer_module(hidden_states)
-
-        return BaseModelOutput(last_hidden_state=hidden_states)
-
-
-class VideoPrismAuxiliaryEncoder(nn.Module):
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-        self.layer = nn.ModuleList([VideoPrismLayer(self.config) for _ in range(config.num_auxiliary_layers)])
-        self.gradient_checkpointing = False
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        **kwargs,
-    ) -> BaseModelOutput:
-        for i, layer_module in enumerate(self.layer):
-            hidden_states = layer_module(hidden_states, attention_mask, **kwargs)
-
-        return BaseModelOutput(last_hidden_state=hidden_states)
-
-
-class VideoPrismTextEncoder(nn.Module):
-    def __init__(self, config: VideoPrismTextConfig):
-        super().__init__()
-        self.config = config
-        self.layer = nn.ModuleList([VideoPrismLayer(config) for _ in range(config.num_text_layers)])
-        self.gradient_checkpointing = False
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        **kwargs,
-    ) -> BaseModelOutput:
-        for i, layer_module in enumerate(self.layer):
-            hidden_states = layer_module(hidden_states, attention_mask, **kwargs)
-
-        return BaseModelOutput(last_hidden_state=hidden_states)
-
-
-def variance_scaling_(tensor, mode="fan_in", distribution="normal"):
-    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
-    if mode == "fan_in":
-        denom = fan_in
-    elif mode == "fan_out":
-        denom = fan_out
-    elif mode == "fan_avg":
-        denom = (fan_in + fan_out) / 2
-
-    variance = 1.0 / denom
-
-    if distribution == "truncated_normal":
-        init.trunc_normal_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
-    elif distribution == "normal":
-        init.normal_(tensor, std=math.sqrt(variance))
-    elif distribution == "uniform":
-        bound = math.sqrt(3 * variance)
-        init.uniform_(tensor, -bound, bound)
-    else:
-        raise ValueError(f"invalid distribution {distribution}")
-
-
-def lecun_normal_(tensor):
-    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
-
-
-class VideoPrismPreTrainedModel(PreTrainedModel):
-    config_class = VideoPrismConfig
-    config: VideoPrismConfig
-    base_model_prefix = "videoprism"
-    main_input_name = "pixel_values_videos"
-    input_modalities = ("video", "text")
-    supports_gradient_checkpointing = True
-    _no_split_modules = [
-        "VideoPrismSpatialEmbeddings",
-        "VideoPrismTemporalEmbeddings",
-        "VideoPrismSpatialEncoder",
-        "VideoPrismTemporalEncoder",
-        "VideoPrismAuxiliaryEncoder",
-        "VideoPrismTextEncoder",
-        "VideoPrismMultiheadAttentionPoolingHead",
-    ]
-    _supports_sdpa = True
-    _supports_flash_attn = True
-    _supports_attention_backend = True
-    _supports_flex_attention = True
-
-    def _init_weights(self, module):
-        if isinstance(module, (nn.Linear, nn.Conv3d)):
-            lecun_normal_(module.weight)
-            init.zeros_(module.bias)
-
-        elif isinstance(module, nn.LayerNorm):
-            init.zeros_(module.bias)
-            init.ones_(module.weight)
-
-
-class VideoPrismVisionModel(VideoPrismPreTrainedModel):
-    config_class = VideoPrismVisionConfig
-    config: VideoPrismVisionConfig
-
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__(config)
-        self.config = config
-        self.layernorm1 = VideoPrismLayerNorm(self.config.hidden_size, eps=self.config.layer_norm_eps)
-        self.layernorm2 = VideoPrismLayerNorm(self.config.hidden_size, eps=self.config.layer_norm_eps)
-        self.spatial_embeddings = VideoPrismSpatialEmbeddings(self.config)
-        self.temporal_embeddings = VideoPrismTemporalEmbeddings(self.config)
-        self.spatial_encoder = VideoPrismSpatialEncoder(self.config)
-        self.temporal_encoder = VideoPrismTemporalEncoder(self.config)
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.spatial_embeddings.patch_embeddings
-
-    def forward(
-        self,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        interpolate_pos_encoding: bool | None = False,
-        **kwargs,
-    ) -> BaseModelOutputWithSpatialAndTemporalStates:
-        r"""
-        Args:
-            pixel_values_videos (`torch.FloatTensor`):
-                Pixel values of the video frames of shape (batch_size, num_frames, num_channels, height, width).
-            interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
-                Whether to interpolate positional encodings to match input size.
-
-        Example:
-
-        ```python
-        >>> from transformers import VideoPrismVideoProcessor, VideoPrismVisionModel
-        >>> import torch
-
-        >>> processor = VideoPrismVideoProcessor.from_pretrained("google/videoprism")
-        >>> model = VideoPrismVisionModel.from_pretrained("google/videoprism")
-
-        >>> video = "sample_video.mp4"
-        >>> inputs = processor(videos=video)
-        >>> with torch.no_grad():
-        ...     outputs = model(**inputs)
-        ...     features = outputs.last_hidden_state
-        ```
-        """
-        if pixel_values_videos is None:
-            raise ValueError("You have to specify pixel_values_videos")
-
-        input_shape = pixel_values_videos.shape
-        spatial_embeds = self.spatial_embeddings(pixel_values_videos, interpolate_pos_encoding)
-        spatial_encoder_outputs: BaseModelOutput = self.spatial_encoder(hidden_states=spatial_embeds, **kwargs)
-        # shape of spatial_sequence_output is (B * num_frames, num_patches, dim)
-        spatial_sequence_output = spatial_encoder_outputs.last_hidden_state
-        features = self.layernorm1(spatial_sequence_output)
-
-        temporal_embeds = self.temporal_embeddings(features, input_shape, interpolate_pos_encoding)
-        temporal_encoder_outputs: BaseModelOutput = self.temporal_encoder(hidden_states=temporal_embeds, **kwargs)
-        # shape of temporal_sequence_output is (B * num_patches, num_frames, dim)
-        temporal_sequence_output = temporal_encoder_outputs.last_hidden_state
-        features = self.layernorm2(temporal_sequence_output)
-        _, num_frames, dim = features.shape
-        features = features.view(input_shape[0], -1, num_frames, dim).permute(0, 2, 1, 3).contiguous()
-        _, num_frames, num_patches, dim = features.shape
-        features = features.view(input_shape[0], num_frames * num_patches, -1)
-
-        return BaseModelOutputWithSpatialAndTemporalStates(
-            last_hidden_state=features,
-            temporal_hidden_state=temporal_sequence_output,
-            spatial_hidden_state=spatial_sequence_output,
-        )
-
-
-class VideoPrismMultiheadAttentionPoolingHead(nn.Module):
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__()
-        self.config = config
-        self.num_attention_heads = self.config.num_attention_heads
-        self.attention_head_size = int(self.config.intermediate_size / self.config.num_attention_heads)
-        self.all_head_size = self.num_attention_heads * self.attention_head_size
-        self.dropout_prob = self.config.attention_probs_dropout_prob
-        # PerDimScale
-        self.dim = int(self.config.intermediate_size / self.config.num_attention_heads)
-        self.per_dim_scale = nn.Parameter(torch.zeros(self.dim))
-        r_softplus_0 = 1.442695041
-        scale = torch.tensor(r_softplus_0 / (self.dim**0.5))
-        softplus = nn.functional.softplus(self.per_dim_scale)
-        scale = scale * softplus
-        self.register_buffer("scale", scale)
-
-        self.pooling_attention_query = nn.Parameter(torch.zeros(1, 1, self.config.hidden_size))
-        self.query = nn.Linear(self.config.hidden_size, self.config.intermediate_size, bias=self.config.qkv_bias)
-        self.key = nn.Linear(self.config.hidden_size, self.config.intermediate_size, bias=self.config.qkv_bias)
-        self.value = nn.Linear(self.config.hidden_size, self.config.intermediate_size, bias=self.config.qkv_bias)
-        self.projection = nn.Linear(self.config.intermediate_size, self.config.hidden_size, bias=self.config.qkv_bias)
-        self.layernorm = VideoPrismLayerNorm(self.config.hidden_size, eps=self.config.layer_norm_eps)
-        self.dim = int(self.config.intermediate_size / self.config.num_attention_heads)
-
-    def forward(
-        self,
-        hidden_states: torch.FloatTensor,
-        attention_mask: torch.LongTensor | None = None,
-        **kwargs,
-    ) -> tuple[torch.FloatTensor, torch.FloatTensor]:
-        batch_size, seq_length, hidden_size = hidden_states.shape
-        query = self.pooling_attention_query.expand(batch_size, -1, -1)
-        query_layer = (
-            self.query(query).view(batch_size, -1, self.num_attention_heads, self.attention_head_size).transpose(1, 2)
-        )
-        query_layer = query_layer * self.scale.expand(*query_layer.shape)
-
-        key_layer = (
-            self.key(hidden_states)
-            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
-            .transpose(1, 2)
-        )
-        value_layer = (
-            self.value(hidden_states)
-            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
-            .transpose(1, 2)
-        )
-
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
-
-        context_layer, attention_probs = attention_interface(
-            self,
-            query_layer,
-            key_layer,
-            value_layer,
-            attention_mask,
-            scaling=1.0,
-            dropout=0.0 if not self.training else self.dropout_prob,
-            softcap=None,
-            **kwargs,
-        )
-
-        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
-        context_layer = context_layer.reshape(new_context_layer_shape)
-        outputs = self.projection(context_layer)
-        outputs = self.layernorm(outputs)
-        return (outputs, attention_probs)
-
-
-def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
-    """This function is intended to align with the l2norm implementation in the FLA library."""
-    inv_norm = torch.rsqrt((x * x).sum(dim=dim, keepdim=True) + eps)
-    return x * inv_norm
-
-
-class VideoPrismTextModel(VideoPrismPreTrainedModel):
-    config_class = VideoPrismTextConfig
-    config: VideoPrismTextConfig
-
-    def __init__(self, config: VideoPrismTextConfig):
-        super().__init__(config)
-        self.config = config
-        self.text_encoder = VideoPrismTextEncoder(self.config)
-        self.token_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
-        self.cls_emb = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
-        self.layernorm = VideoPrismLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.normalize = config.apply_l2_norm
-        self.post_init()
-
-    def create_sinusoidal_positions(self, num_pos: int, dim: int) -> torch.Tensor:
-        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / (dim - 2)))
-        sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=torch.int64).float(), inv_freq).float()
-        return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        **kwargs,
-    ) -> BaseModelOutput:
-        r"""
-        Args:
-            input_ids (`torch.Tensor`):
-                Input token IDs.
-            attention_mask (`torch.Tensor`, *optional*):
-                Attention mask to avoid performing attention on padding token indices.
-        """
-        batch_size, seq_length = input_ids.shape
-        hidden_states = self.token_embeddings(input_ids)
-        hidden_states = hidden_states * (self.config.hidden_size**0.5)
-
-        cls_padding = torch.ones(batch_size, 1)
-        input_ids = torch.cat((input_ids, cls_padding), dim=1)
-        attention_mask = torch.cat((attention_mask, cls_padding), dim=1) if attention_mask is not None else None
-
-        if attention_mask is not None:
-            attention_mask = create_causal_mask(
-                config=self.config,
-                input_embeds=hidden_states,
-                attention_mask=attention_mask,
-                cache_position=torch.arange(hidden_states.shape[1] + 1, device=hidden_states.device),
-                past_key_values=None,
-            )
-
-        features = hidden_states + self.create_sinusoidal_positions(seq_length, self.config.hidden_size)
-        cls_emb = self.cls_emb * (self.config.hidden_size**0.5)
-        cls_emb = cls_emb.expand(features.shape[0], -1, -1)
-        features = torch.cat((features, cls_emb), dim=1)
-        text_encoder_output = self.text_encoder(features, attention_mask)
-        features = text_encoder_output.last_hidden_state
-        features = self.layernorm(features)
-        text_embeddings = features[:, -1]
-
-        if self.normalize:
-            text_embeddings = l2norm(text_embeddings, dim=-1)
-
-        return BaseModelOutput(
-            last_hidden_state=text_embeddings,
-        )
-
-
-
-class VideoPrismVideoModel(VideoPrismPreTrainedModel):
-    config_class = VideoPrismVisionConfig
-    config: VideoPrismVisionConfig
-
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__(config)
-        self.config = config
-        self.backbone = VideoPrismVisionModel(self.config)
-        self.auxiliary_encoder = VideoPrismAuxiliaryEncoder(self.config)
-        self.contrastive_vision_pooler = VideoPrismMultiheadAttentionPoolingHead(self.config)
-        self.normalize = self.config.apply_l2_norm
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.backbone.spatial_embeddings.patch_embeddings
-
-    def forward(
-        self,
-        pixel_values_videos: torch.FloatTensor,
-        interpolate_pos_encoding: bool | None = False,
-        **kwargs,
-    ) -> VideoPrismVideoOutput:
-        r"""
-        Args:
-            pixel_values_videos (`torch.FloatTensor`):
-                Pixel values of the video frames.
-            interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
-                Whether to interpolate positional encodings to match input size.
-        """
-        backbone_outputs = self.backbone(
-            pixel_values_videos=pixel_values_videos, interpolate_pos_encoding=interpolate_pos_encoding, **kwargs
-        )
-        video_features = backbone_outputs.last_hidden_state
-        auxiliary_output = self.auxiliary_encoder(video_features)
-        auxiliary_output_features = auxiliary_output.last_hidden_state
-        contrastive_vision_pooler_output = self.contrastive_vision_pooler(auxiliary_output_features, **kwargs)
-        video_embeddings = contrastive_vision_pooler_output[0]
-        if self.normalize:
-            video_embeddings = l2norm(video_embeddings, dim=-1)
-
-        return VideoPrismVideoOutput(
-            video_last_hidden_state=video_embeddings,
-            auxiliary_output=auxiliary_output,
-            attention_pooling_output=contrastive_vision_pooler_output,
-        )
-
-
-class VideoPrismClipModel(VideoPrismPreTrainedModel):
-    config_class = VideoPrismConfig
-
-    def __init__(self, config: VideoPrismConfig):
-        super().__init__(config)
-        self.config = config
-        self.vision_config = config.vision_config
-        self.text_config = config.text_config
-        self.video_model = VideoPrismVideoModel(self.vision_config)
-        self.text_model = VideoPrismTextModel(self.text_config)
-        self.post_init()
-
-    def forward(
-        self,
-        pixel_values_videos: torch.FloatTensor,
-        input_ids: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        interpolate_pos_encoding: bool | None = False,
-        temperature: float | None = None,
-        **kwargs,
-    ) -> VideoPrismClipOutput:
-        r"""
-        Args:
-            pixel_values_videos (`torch.FloatTensor`):
-                Pixel values of the video frames.
-            input_ids (`torch.Tensor`):
-                Input token IDs for text.
-            attention_mask (`torch.Tensor`, *optional*):
-                Attention mask for text inputs.
-            interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
-                Whether to interpolate positional encodings.
-            temperature (`float`, *optional*):
-                Temperature parameter for scaling similarity scores.
-
-        Example:
-
-        ```python
-        >>> from transformers import VideoPrismProcessor, VideoPrismClipModel
-        >>> import torch
-
-        >>> processor = VideoPrismProcessor.from_pretrained("google/videoprism")
-        >>> model = VideoPrismClipModel.from_pretrained("google/videoprism")
-
-        >>> video = "sample_video.mp4"
-        >>> texts = ["a dog", "a cat"]
-        >>> inputs = processor(videos=video, texts=texts, return_tensors="pt", padding=True)
-
-        >>> with torch.no_grad():
-        ...     outputs = model(**inputs)
-        ...     logits_per_video = outputs.logits_per_video
-        ```
-        """
-        video_model_outputs = self.video_model(
-            pixel_values_videos=pixel_values_videos, interpolate_pos_encoding=interpolate_pos_encoding, **kwargs
-        )
-        text_model_outputs = self.text_model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)
-
-        video_embeddings = video_model_outputs.video_last_hidden_state
-        text_embeddings = text_model_outputs.last_hidden_state
-        emb_dim = video_embeddings[0].shape[-1]
-        assert emb_dim == text_embeddings[0].shape[-1]
-
-        video_embeds = video_embeddings.reshape(-1, emb_dim)
-        text_embeds = text_embeddings.reshape(-1, emb_dim)
-        similarity_matrix = torch.matmul(video_embeds, text_embeds.T)
-
-        if temperature is not None:
-            similarity_matrix /= temperature
-
-        logits_per_video = torch.exp(similarity_matrix)
-        logits_per_text = logits_per_video.T
-        logits_per_video = logits_per_video / torch.sum(logits_per_video, dim=0, keepdims=True)
-        logits_per_text = logits_per_text / torch.sum(logits_per_text, dim=0, keepdims=True)
-
-        return VideoPrismClipOutput(
-            logits_per_video=logits_per_video,
-            logits_per_text=logits_per_text,
-            video_embeds=video_embeds,
-            text_embeds=text_embeds,
-        )
-
-
-
-class VideoPrismForVideoClassification(VideoPrismPreTrainedModel):
-    config_class = VideoPrismVisionConfig
-    config: VideoPrismVisionConfig
-
-    def __init__(self, config: VideoPrismVisionConfig):
-        super().__init__(config)
-        self.config = config
-        self.encoder = VideoPrismVisionModel(self.config)
-        self.contrastive_vision_pooler = VideoPrismMultiheadAttentionPoolingHead(self.config)
-        self.classifier = nn.Linear(self.config.hidden_size, self.config.num_labels)
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.encoder.spatial_embeddings.patch_embeddings
-
-    def forward(
-        self,
-        pixel_values_videos: torch.FloatTensor,
-        labels: torch.LongTensor | None = None,
-        interpolate_pos_encoding: bool | None = False,
-        **kwargs,
-    ) -> ImageClassifierOutput:
-        r"""
-        Args:
-            pixel_values_videos (`torch.FloatTensor`):
-                Pixel values of the video frames.
-            labels (`torch.LongTensor`, *optional*):
-                Video classification labels.
-            interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
-                Whether to interpolate positional encodings.
-
-        Example:
-
-        ```python
-        >>> from transformers import VideoPrismVideoProcessor, VideoPrismForVideoClassification
-        >>> import torch
-
-        >>> processor = VideoPrismVideoProcessor("google/videoprism")
-        >>> model = VideoPrismForVideoClassification.from_pretrained("google/videoprism", num_labels=1000)
-
-        >>> video = "sample_video.mp4"
-        >>> inputs = processor(videos=video, return_tensors="pt")
-
-        >>> with torch.no_grad():
-        ...     outputs = model(**inputs)
-        ...     logits = outputs.logits
-        ```
-        """
-        encoder_outputs = self.encoder(
-            pixel_values_videos=pixel_values_videos, interpolate_pos_encoding=interpolate_pos_encoding, **kwargs
-        )
-        sequence_output = encoder_outputs.last_hidden_state
-        pooled_output = self.contrastive_vision_pooler(sequence_output, **kwargs).pooled_output
-        logits = self.classifier(pooled_output)
-        loss = None
-        if labels is not None:
-            loss = self.loss_function(labels, logits, self.config, **kwargs)
-
-        return ImageClassifierOutput(
-            loss=loss,
-            logits=logits,
-            hidden_states=encoder_outputs.last_hidden_state,
-        )
-
-
-__all__ = [
-    "VideoPrismVisionModel",
-    "VideoPrismPreTrainedModel",
-    "VideoPrismVideoModel",
-    "VideoPrismTextModel",
-    "VideoPrismClipModel",
-    "VideoPrismForVideoClassification",
-]

src/lerobot/policies/videovla/videoprism/test.py

@@ -1,50 +0,0 @@
-import torch
-import numpy as np
-from torchcodec.decoders import VideoDecoder
-
-from lerobot.policies.videovla.videoprism import VideoPrismVideoProcessor
-from lerobot.policies.videovla.videoprism import VideoPrismVisionModel
-processor = VideoPrismVideoProcessor.from_pretrained(
-    "MHRDYN7/videoprism-base-f16r288"
-)
-
-model = VideoPrismVisionModel.from_pretrained(
-    "MHRDYN7/videoprism-base-f16r288",
-    torch_dtype=torch.float16,
-    device_map="auto",
-    attn_implementation="sdpa",
-)
-
-video_url = "https://huggingface.co/datasets/nateraw/kinetics-mini/resolve/main/val/archery/-Qz25rXdMjE_000014_000024.mp4"
-
-vr = VideoDecoder(video_url)
-frame_idx = np.arange(0, 64)
-video = vr.get_frames_at(indices=frame_idx).data  # T x C x H x W
-
-video = processor(video, return_tensors="pt")
-video = {k: v.to(model.device, model.dtype) for k, v in video.items()}
-outputs = model(**video)
-encoder_outputs = outputs.last_hidden_state
-print(encoder_outputs.shape) # 
-
-import time
-import torch
-
-# warmup
-for _ in range(10):
-    _ = model(**video)
-
-times = []
-for _ in range(50):
-    torch.cuda.synchronize()
-    t0 = time.perf_counter()
-
-    _ = model(**video)
-
-    torch.cuda.synchronize()
-    t1 = time.perf_counter()
-    times.append(t1 - t0)
-
-print(f"Mean: {1000*sum(times)/len(times):.2f} ms")
-print(f"Min : {1000*min(times):.2f} ms")
-print(f"Max : {1000*max(times):.2f} ms")

src/lerobot/policies/videovla/videoprism/video_processing_videoprism.py

@@ -1,44 +0,0 @@
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-#           This file was automatically generated from src/transformers/models/videoprism/modular_videoprism.py.
-#               Do NOT edit this file manually as any edits will be overwritten by the generation of
-#             the file from the modular. If any change should be done, please apply the change to the
-#                          modular_videoprism.py file directly. One of our CI enforces this.
-#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-
-from transformers.image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling
-from transformers.video_processing_utils import BaseVideoProcessor
-
-
-class VideoPrismVideoProcessor(BaseVideoProcessor):
-    r"""
-    Constructs a VideoPrism video processor.
-
-    This processor inherits from [`LlavaOnevisionVideoProcessor`] and sets default parameters for VideoPrism models.
-    Video frames are resized to 288x288 using bicubic resampling without normalization.
-
-    Args:
-        size (`Dict[str, int]`, *optional*, defaults to `{"height": 288, "width": 288}`):
-            The size to resize the video frames to.
-        resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
-            The resampling filter to use when resizing images.
-        do_normalize (`bool`, *optional*, defaults to `False`):
-            Whether to normalize the video frames.
-    """
-
-    resample = PILImageResampling.BICUBIC
-    image_mean = OPENAI_CLIP_MEAN
-    image_std = OPENAI_CLIP_STD
-
-    size = {"height": 288, "width": 288}
-    rescale_factor = 1 / 255
-    default_to_square = False
-    crop_size = None
-    do_resize = True
-    do_center_crop = None
-    do_rescale = True
-    do_normalize = False
-    do_convert_rgb = True
-    do_sample_frames = False  # Set to False for BC, recommended to set `True` in new models
-
-
-__all__ = ["VideoPrismVideoProcessor"]

src/lerobot/policies/vqbet/configuration_vqbet.py

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

src/lerobot/processor/converters.py

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

src/lerobot/processor/hil_processor.py

@@ -18,16 +18,18 @@
 import math
 import time
 from dataclasses import dataclass
-from typing import Any, Protocol, TypeVar, runtime_checkable
+from typing import TYPE_CHECKING, Any, Protocol, TypeVar, runtime_checkable
 
 import numpy as np
 import torch
 import torchvision.transforms.functional as F  # noqa: N812
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
-from lerobot.teleoperators.teleoperator import Teleoperator
 from lerobot.teleoperators.utils import TeleopEvents
 
+if TYPE_CHECKING:
+    from lerobot.teleoperators.teleoperator import Teleoperator
+
 from .core import EnvTransition, PolicyAction, TransitionKey
 from .pipeline import (
     ComplementaryDataProcessorStep,
@@ -69,10 +71,10 @@ class HasTeleopEvents(Protocol):
 
 
 # Type variable constrained to Teleoperator subclasses that also implement events
-TeleopWithEvents = TypeVar("TeleopWithEvents", bound=Teleoperator)
+TeleopWithEvents = TypeVar("TeleopWithEvents", bound="Teleoperator")
 
 
-def _check_teleop_with_events(teleop: Teleoperator) -> None:
+def _check_teleop_with_events(teleop: "Teleoperator") -> None:
     """
     Runtime check that a teleoperator implements the `HasTeleopEvents` protocol.
 
@@ -103,7 +105,7 @@ class AddTeleopActionAsComplimentaryDataStep(ComplementaryDataProcessorStep):
         teleop_device: The teleoperator instance to get the action from.
     """
 
-    teleop_device: Teleoperator
+    teleop_device: "Teleoperator"
 
     def complementary_data(self, complementary_data: dict) -> dict:
         """
@@ -312,7 +314,7 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
 
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
-class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
+class GripperPenaltyProcessorStep(ProcessorStep):
     """
     Applies a penalty for inefficient gripper usage.
 
@@ -327,26 +329,27 @@ 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 +365,12 @@ class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
 
         gripper_penalty = self.penalty * int(gripper_penalty_bool)
 
-        # Create new complementary data with penalty info
+        # Update complementary data with penalty info
         new_complementary_data = dict(complementary_data)
         new_complementary_data[DISCRETE_PENALTY_KEY] = gripper_penalty
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
 
-        return new_complementary_data
+        return new_transition
 
     def get_config(self) -> dict[str, Any]:
         """

src/lerobot/processor/tokenizer_processor.py

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

src/lerobot/rl/gym_manipulator.py

@@ -412,7 +412,10 @@ 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 +438,12 @@ 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,

src/lerobot/rl/learner.py

@@ -545,9 +545,6 @@ def add_actor_information_and_train(
                 training_infos["temperature_grad_norm"] = temp_grad_norm
                 training_infos["temperature"] = policy.temperature
 
-                # Update temperature
-                policy.update_temperature()
-
         # 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)

src/lerobot/rl/wandb_utils.py

@@ -26,8 +26,21 @@ from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.constants import PRETRAINED_MODEL_DIR
 
 
-def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[str] | str:
+def cfg_to_group(
+    cfg: TrainPipelineConfig, return_list: bool = False, truncate_tags: bool = False, max_tag_length: int = 64
+) -> list[str] | str:
     """Return a group name for logging. Optionally returns group name as list."""
+
+    def _maybe_truncate(tag: str) -> str:
+        """Truncate tag to max_tag_length characters if required.
+
+        wandb rejects tags longer than 64 characters.
+        See: https://github.com/wandb/wandb/blob/main/wandb/sdk/wandb_settings.py
+        """
+        if len(tag) <= max_tag_length:
+            return tag
+        return tag[:max_tag_length]
+
     lst = [
         f"policy:{cfg.policy.type}",
         f"seed:{cfg.seed}",
@@ -36,6 +49,8 @@ def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[st
         lst.append(f"dataset:{cfg.dataset.repo_id}")
     if cfg.env is not None:
         lst.append(f"env:{cfg.env.type}")
+    if truncate_tags:
+        lst = [_maybe_truncate(tag) for tag in lst]
     return lst if return_list else "-".join(lst)
 
 
@@ -83,7 +98,7 @@ class WandBLogger:
             entity=self.cfg.entity,
             name=self.job_name,
             notes=self.cfg.notes,
-            tags=cfg_to_group(cfg, return_list=True),
+            tags=cfg_to_group(cfg, return_list=True, truncate_tags=True),
             dir=self.log_dir,
             config=cfg.to_dict(),
             # TODO(rcadene): try set to True

src/lerobot/robots/bi_openarm_follower/__init__.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .bi_openarm_follower import BiOpenArmFollower
+from .config_bi_openarm_follower import BiOpenArmFollowerConfig
+
+__all__ = ["BiOpenArmFollower", "BiOpenArmFollowerConfig"]

src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py

@@ -0,0 +1,175 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from functools import cached_property
+
+from lerobot.processor import RobotAction, RobotObservation
+from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
+
+from ..robot import Robot
+from .config_bi_openarm_follower import BiOpenArmFollowerConfig
+
+logger = logging.getLogger(__name__)
+
+
+class BiOpenArmFollower(Robot):
+    """
+    Bimanual OpenArm Follower Arms
+    """
+
+    config_class = BiOpenArmFollowerConfig
+    name = "bi_openarm_follower"
+
+    def __init__(self, config: BiOpenArmFollowerConfig):
+        super().__init__(config)
+        self.config = config
+
+        left_arm_config = OpenArmFollowerConfig(
+            id=f"{config.id}_left" if config.id else None,
+            calibration_dir=config.calibration_dir,
+            port=config.left_arm_config.port,
+            disable_torque_on_disconnect=config.left_arm_config.disable_torque_on_disconnect,
+            max_relative_target=config.left_arm_config.max_relative_target,
+            cameras=config.left_arm_config.cameras,
+            side=config.left_arm_config.side,
+            can_interface=config.left_arm_config.can_interface,
+            use_can_fd=config.left_arm_config.use_can_fd,
+            can_bitrate=config.left_arm_config.can_bitrate,
+            can_data_bitrate=config.left_arm_config.can_data_bitrate,
+            motor_config=config.left_arm_config.motor_config,
+            position_kd=config.left_arm_config.position_kd,
+            position_kp=config.left_arm_config.position_kp,
+            joint_limits=config.left_arm_config.joint_limits,
+        )
+
+        right_arm_config = OpenArmFollowerConfig(
+            id=f"{config.id}_right" if config.id else None,
+            calibration_dir=config.calibration_dir,
+            port=config.right_arm_config.port,
+            disable_torque_on_disconnect=config.right_arm_config.disable_torque_on_disconnect,
+            max_relative_target=config.right_arm_config.max_relative_target,
+            cameras=config.right_arm_config.cameras,
+            side=config.right_arm_config.side,
+            can_interface=config.right_arm_config.can_interface,
+            use_can_fd=config.right_arm_config.use_can_fd,
+            can_bitrate=config.right_arm_config.can_bitrate,
+            can_data_bitrate=config.right_arm_config.can_data_bitrate,
+            motor_config=config.right_arm_config.motor_config,
+            position_kd=config.right_arm_config.position_kd,
+            position_kp=config.right_arm_config.position_kp,
+            joint_limits=config.right_arm_config.joint_limits,
+        )
+
+        self.left_arm = OpenArmFollower(left_arm_config)
+        self.right_arm = OpenArmFollower(right_arm_config)
+
+        # Only for compatibility with other parts of the codebase that expect a `robot.cameras` attribute
+        self.cameras = {**self.left_arm.cameras, **self.right_arm.cameras}
+
+    @property
+    def _motors_ft(self) -> dict[str, type]:
+        left_arm_motors_ft = self.left_arm._motors_ft
+        right_arm_motors_ft = self.right_arm._motors_ft
+
+        return {
+            **{f"left_{k}": v for k, v in left_arm_motors_ft.items()},
+            **{f"right_{k}": v for k, v in right_arm_motors_ft.items()},
+        }
+
+    @property
+    def _cameras_ft(self) -> dict[str, tuple]:
+        left_arm_cameras_ft = self.left_arm._cameras_ft
+        right_arm_cameras_ft = self.right_arm._cameras_ft
+
+        return {
+            **{f"left_{k}": v for k, v in left_arm_cameras_ft.items()},
+            **{f"right_{k}": v for k, v in right_arm_cameras_ft.items()},
+        }
+
+    @cached_property
+    def observation_features(self) -> dict[str, type | tuple]:
+        return {**self._motors_ft, **self._cameras_ft}
+
+    @cached_property
+    def action_features(self) -> dict[str, type]:
+        return self._motors_ft
+
+    @property
+    def is_connected(self) -> bool:
+        return self.left_arm.is_connected and self.right_arm.is_connected
+
+    def connect(self, calibrate: bool = True) -> None:
+        self.left_arm.connect(calibrate)
+        self.right_arm.connect(calibrate)
+
+    @property
+    def is_calibrated(self) -> bool:
+        return self.left_arm.is_calibrated and self.right_arm.is_calibrated
+
+    def calibrate(self) -> None:
+        self.left_arm.calibrate()
+        self.right_arm.calibrate()
+
+    def configure(self) -> None:
+        self.left_arm.configure()
+        self.right_arm.configure()
+
+    def setup_motors(self) -> None:
+        raise NotImplementedError(
+            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
+        )
+
+    def get_observation(self) -> RobotObservation:
+        obs_dict = {}
+
+        # Add "left_" prefix
+        left_obs = self.left_arm.get_observation()
+        obs_dict.update({f"left_{key}": value for key, value in left_obs.items()})
+
+        # Add "right_" prefix
+        right_obs = self.right_arm.get_observation()
+        obs_dict.update({f"right_{key}": value for key, value in right_obs.items()})
+
+        return obs_dict
+
+    def send_action(
+        self,
+        action: RobotAction,
+        custom_kp: dict[str, float] | None = None,
+        custom_kd: dict[str, float] | None = None,
+    ) -> RobotAction:
+        # Remove "left_" prefix
+        left_action = {
+            key.removeprefix("left_"): value for key, value in action.items() if key.startswith("left_")
+        }
+        # Remove "right_" prefix
+        right_action = {
+            key.removeprefix("right_"): value for key, value in action.items() if key.startswith("right_")
+        }
+
+        sent_action_left = self.left_arm.send_action(left_action, custom_kp, custom_kd)
+        sent_action_right = self.right_arm.send_action(right_action, custom_kp, custom_kd)
+
+        # Add prefixes back
+        prefixed_sent_action_left = {f"left_{key}": value for key, value in sent_action_left.items()}
+        prefixed_sent_action_right = {f"right_{key}": value for key, value in sent_action_right.items()}
+
+        return {**prefixed_sent_action_left, **prefixed_sent_action_right}
+
+    def disconnect(self):
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()

src/lerobot/robots/bi_openarm_follower/config_bi_openarm_follower.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from lerobot.robots.openarm_follower import OpenArmFollowerConfigBase
+
+from ..config import RobotConfig
+
+
+@RobotConfig.register_subclass("bi_openarm_follower")
+@dataclass
+class BiOpenArmFollowerConfig(RobotConfig):
+    """Configuration class for Bi OpenArm Follower robots."""
+
+    left_arm_config: OpenArmFollowerConfigBase
+    right_arm_config: OpenArmFollowerConfigBase

src/lerobot/robots/openarm_follower/__init__.py

@@ -0,0 +1,20 @@
+#!/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 .config_openarm_follower import OpenArmFollowerConfig, OpenArmFollowerConfigBase
+from .openarm_follower import OpenArmFollower
+
+__all__ = ["OpenArmFollower", "OpenArmFollowerConfig", "OpenArmFollowerConfigBase"]

src/lerobot/robots/openarm_follower/config_openarm_follower.py

@@ -0,0 +1,122 @@
+#!/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 dataclasses import dataclass, field
+
+from lerobot.cameras import CameraConfig
+
+from ..config import RobotConfig
+
+LEFT_DEFAULT_JOINTS_LIMITS: dict[str, tuple[float, float]] = {
+    "joint_1": (-75.0, 75.0),
+    "joint_2": (-90.0, 9.0),
+    "joint_3": (-85.0, 85.0),
+    "joint_4": (0.0, 135.0),
+    "joint_5": (-85.0, 85.0),
+    "joint_6": (-40.0, 40.0),
+    "joint_7": (-80.0, 80.0),
+    "gripper": (-65.0, 0.0),
+}
+
+RIGHT_DEFAULT_JOINTS_LIMITS: dict[str, tuple[float, float]] = {
+    "joint_1": (-75.0, 75.0),
+    "joint_2": (-9.0, 90.0),
+    "joint_3": (-85.0, 85.0),
+    "joint_4": (0.0, 135.0),
+    "joint_5": (-85.0, 85.0),
+    "joint_6": (-40.0, 40.0),
+    "joint_7": (-80.0, 80.0),
+    "gripper": (-65.0, 0.0),
+}
+
+
+@dataclass
+class OpenArmFollowerConfigBase:
+    """Base configuration for the OpenArms follower robot with Damiao motors."""
+
+    # CAN interfaces - one per arm
+    # arm CAN interface (e.g., "can1")
+    # Linux: "can0", "can1", etc.
+    port: str
+
+    # side of the arm: "left" or "right". If "None" default values will be used
+    side: str | None = None
+
+    # CAN interface type: "socketcan" (Linux), "slcan" (serial), or "auto" (auto-detect)
+    can_interface: str = "socketcan"
+
+    # CAN FD settings (OpenArms uses CAN FD by default)
+    use_can_fd: bool = True
+    can_bitrate: int = 1000000  # Nominal bitrate (1 Mbps)
+    can_data_bitrate: int = 5000000  # Data bitrate for CAN FD (5 Mbps)
+
+    # Whether to disable torque when disconnecting
+    disable_torque_on_disconnect: bool = True
+
+    # Safety limit for relative target positions
+    # Set to a positive scalar for all motors, or a dict mapping motor names to limits
+    max_relative_target: float | dict[str, float] | None = None
+
+    # Camera configurations
+    cameras: dict[str, CameraConfig] = field(default_factory=dict)
+
+    # Motor configuration for OpenArms (7 DOF per arm)
+    # Maps motor names to (send_can_id, recv_can_id, motor_type)
+    # Based on: https://docs.openarm.dev/software/setup/configure-test
+    # OpenArms uses 4 types of motors:
+    # - DM8009 (DM-J8009P-2EC) for shoulders (high torque)
+    # - DM4340P and DM4340 for shoulder rotation and elbow
+    # - DM4310 (DM-J4310-2EC V1.1) for wrist and gripper
+    motor_config: dict[str, tuple[int, int, str]] = field(
+        default_factory=lambda: {
+            "joint_1": (0x01, 0x11, "dm8009"),  # J1 - Shoulder pan (DM8009)
+            "joint_2": (0x02, 0x12, "dm8009"),  # J2 - Shoulder lift (DM8009)
+            "joint_3": (0x03, 0x13, "dm4340"),  # J3 - Shoulder rotation (DM4340)
+            "joint_4": (0x04, 0x14, "dm4340"),  # J4 - Elbow flex (DM4340)
+            "joint_5": (0x05, 0x15, "dm4310"),  # J5 - Wrist roll (DM4310)
+            "joint_6": (0x06, 0x16, "dm4310"),  # J6 - Wrist pitch (DM4310)
+            "joint_7": (0x07, 0x17, "dm4310"),  # J7 - Wrist rotation (DM4310)
+            "gripper": (0x08, 0x18, "dm4310"),  # J8 - Gripper (DM4310)
+        }
+    )
+
+    # MIT control parameters for position control (used in send_action)
+    # List of 8 values: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
+    position_kp: list[float] = field(
+        default_factory=lambda: [240.0, 240.0, 240.0, 240.0, 24.0, 31.0, 25.0, 25.0]
+    )
+    position_kd: list[float] = field(default_factory=lambda: [5.0, 5.0, 3.0, 5.0, 0.3, 0.3, 0.3, 0.3])
+
+    # Values for joint limits. Can be overridden via CLI (for custom values) or by setting config.side to either 'left' or 'right'.
+    # If config.side is left set to None and no CLI values are passed, the default joint limit values are small for safety.
+    joint_limits: dict[str, tuple[float, float]] = field(
+        default_factory=lambda: {
+            "joint_1": (-5.0, 5.0),
+            "joint_2": (-5.0, 5.0),
+            "joint_3": (-5.0, 5.0),
+            "joint_4": (0.0, 5.0),
+            "joint_5": (-5.0, 5.0),
+            "joint_6": (-5.0, 5.0),
+            "joint_7": (-5.0, 5.0),
+            "gripper": (-5.0, 0.0),
+        }
+    )
+
+
+@RobotConfig.register_subclass("openarm_follower")
+@dataclass
+class OpenArmFollowerConfig(RobotConfig, OpenArmFollowerConfigBase):
+    pass

src/lerobot/robots/openarm_follower/openarm_follower.py

@@ -0,0 +1,348 @@
+#!/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.
+
+import logging
+import time
+from functools import cached_property
+from typing import Any
+
+from lerobot.cameras.utils import make_cameras_from_configs
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode
+from lerobot.motors.damiao import DamiaoMotorsBus
+from lerobot.processor import RobotAction, RobotObservation
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..robot import Robot
+from ..utils import ensure_safe_goal_position
+from .config_openarm_follower import (
+    LEFT_DEFAULT_JOINTS_LIMITS,
+    RIGHT_DEFAULT_JOINTS_LIMITS,
+    OpenArmFollowerConfig,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class OpenArmFollower(Robot):
+    """
+    OpenArms Follower Robot which uses CAN bus communication to control 7 DOF arm with a gripper.
+    The arm uses Damiao motors in MIT control mode.
+    """
+
+    config_class = OpenArmFollowerConfig
+    name = "openarm_follower"
+
+    def __init__(self, config: OpenArmFollowerConfig):
+        super().__init__(config)
+        self.config = config
+
+        # Arm motors
+        motors: dict[str, Motor] = {}
+        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
+            motor = Motor(
+                send_id, motor_type_str, MotorNormMode.DEGREES
+            )  # Always use degrees for Damiao motors
+            motor.recv_id = recv_id
+            motor.motor_type_str = motor_type_str
+            motors[motor_name] = motor
+
+        self.bus = DamiaoMotorsBus(
+            port=self.config.port,
+            motors=motors,
+            calibration=self.calibration,
+            can_interface=self.config.can_interface,
+            use_can_fd=self.config.use_can_fd,
+            bitrate=self.config.can_bitrate,
+            data_bitrate=self.config.can_data_bitrate if self.config.use_can_fd else None,
+        )
+
+        if config.side is not None:
+            if config.side == "left":
+                config.joint_limits = LEFT_DEFAULT_JOINTS_LIMITS
+            elif config.side == "right":
+                config.joint_limits = RIGHT_DEFAULT_JOINTS_LIMITS
+            else:
+                raise ValueError(
+                    "config.side must be either 'left', 'right' (for default values) or 'None' (for CLI values)"
+                )
+        else:
+            logger.info(
+                "Set config.side to either 'left' or 'right' to use pre-configured values for joint limits."
+            )
+        logger.info(f"Values used for joint limits: {config.joint_limits}.")
+
+        # Initialize cameras
+        self.cameras = make_cameras_from_configs(config.cameras)
+
+    @property
+    def _motors_ft(self) -> dict[str, type]:
+        """Motor features for observation and action spaces."""
+        features: dict[str, type] = {}
+        for motor in self.bus.motors:
+            features[f"{motor}.pos"] = float
+            features[f"{motor}.vel"] = float  # Add this
+            features[f"{motor}.torque"] = float  # Add this
+        return features
+
+    @property
+    def _cameras_ft(self) -> dict[str, tuple]:
+        """Camera features for observation space."""
+        return {
+            cam: (self.config.cameras[cam].height, self.config.cameras[cam].width, 3) for cam in self.cameras
+        }
+
+    @cached_property
+    def observation_features(self) -> dict[str, type | tuple]:
+        """Combined observation features from motors and cameras."""
+        return {**self._motors_ft, **self._cameras_ft}
+
+    @cached_property
+    def action_features(self) -> dict[str, type]:
+        """Action features."""
+        return self._motors_ft
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if robot is connected."""
+        return self.bus.is_connected and all(cam.is_connected for cam in self.cameras.values())
+
+    def connect(self, calibrate: bool = True) -> None:
+        """
+        Connect to the robot and optionally calibrate.
+
+        We assume that at connection time, the arms are in a safe rest position,
+        and torque can be safely disabled to run calibration if needed.
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} already connected")
+
+        # Connect to CAN bus
+        logger.info(f"Connecting arm on {self.config.port}...")
+        self.bus.connect()
+
+        # Run calibration if needed
+        if not self.is_calibrated and calibrate:
+            logger.info(
+                "Mismatch between calibration values in the motor and the calibration file or no calibration file found"
+            )
+            self.calibrate()
+
+        for cam in self.cameras.values():
+            cam.connect()
+
+        self.configure()
+
+        if self.is_calibrated:
+            self.bus.set_zero_position()
+
+        self.bus.enable_torque()
+
+        logger.info(f"{self} connected.")
+
+    @property
+    def is_calibrated(self) -> bool:
+        """Check if robot is calibrated."""
+        return self.bus.is_calibrated
+
+    def calibrate(self) -> None:
+        """
+        Run calibration procedure for OpenArms robot.
+
+        The calibration procedure:
+        1. Disable torque
+        2. Ask user to position arms in hanging position with grippers closed
+        3. Set this as zero position
+        4. Record range of motion for each joint
+        5. Save calibration
+        """
+        if self.calibration:
+            # Calibration file exists, ask user whether to use it or run new calibration
+            user_input = input(
+                f"Press ENTER to use provided calibration file associated with the id {self.id}, or type 'c' and press ENTER to run calibration: "
+            )
+            if user_input.strip().lower() != "c":
+                logger.info(f"Writing calibration file associated with the id {self.id} to the motors")
+                self.bus.write_calibration(self.calibration)
+                return
+
+        logger.info(f"\nRunning calibration for {self}")
+        self.bus.disable_torque()
+
+        # Step 1: Set zero position
+        input(
+            "\nCalibration: Set Zero Position)\n"
+            "Position the arm in the following configuration:\n"
+            "  - Arm hanging straight down\n"
+            "  - Gripper closed\n"
+            "Press ENTER when ready..."
+        )
+
+        # Set current position as zero for all motors
+        self.bus.set_zero_position()
+        logger.info("Arm zero position set.")
+
+        logger.info("Setting range: -90° to +90° for safety by default for all joints")
+        for motor_name, motor in self.bus.motors.items():
+            self.calibration[motor_name] = MotorCalibration(
+                id=motor.id,
+                drive_mode=0,
+                homing_offset=0,
+                range_min=-90,
+                range_max=90,
+            )
+
+        self.bus.write_calibration(self.calibration)
+        self._save_calibration()
+        print(f"Calibration saved to {self.calibration_fpath}")
+
+    def configure(self) -> None:
+        """Configure motors with appropriate settings."""
+        # TODO(Steven, Pepijn): Slightly different from what it is happening in the leader
+        with self.bus.torque_disabled():
+            self.bus.configure_motors()
+
+    def setup_motors(self) -> None:
+        raise NotImplementedError(
+            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
+        )
+
+    def get_observation(self) -> RobotObservation:
+        """
+        Get current observation from robot including position, velocity, and torque.
+
+        Reads all motor states (pos/vel/torque) in one CAN refresh cycle
+        instead of 3 separate reads.
+        """
+        start = time.perf_counter()
+
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        obs_dict: dict[str, Any] = {}
+
+        states = self.bus.sync_read_all_states()
+
+        for motor in self.bus.motors:
+            state = states.get(motor, {})
+            obs_dict[f"{motor}.pos"] = state.get("position", 0.0)
+            obs_dict[f"{motor}.vel"] = state.get("velocity", 0.0)
+            obs_dict[f"{motor}.torque"] = state.get("torque", 0.0)
+
+        # Capture images from cameras
+        for cam_key, cam in self.cameras.items():
+            start = time.perf_counter()
+            obs_dict[cam_key] = cam.async_read()
+            dt_ms = (time.perf_counter() - start) * 1e3
+            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
+
+        dt_ms = (time.perf_counter() - start) * 1e3
+        logger.debug(f"{self} get_observation took: {dt_ms:.1f}ms")
+
+        return obs_dict
+
+    def send_action(
+        self,
+        action: RobotAction,
+        custom_kp: dict[str, float] | None = None,
+        custom_kd: dict[str, float] | None = None,
+    ) -> RobotAction:
+        """
+        Send action command to robot.
+
+        The action magnitude may be clipped based on safety limits.
+
+        Args:
+            action: Dictionary with motor positions (e.g., "joint_1.pos", "joint_2.pos")
+            custom_kp: Optional custom kp gains per motor (e.g., {"joint_1": 120.0, "joint_2": 150.0})
+            custom_kd: Optional custom kd gains per motor (e.g., {"joint_1": 1.5, "joint_2": 2.0})
+
+        Returns:
+            The action actually sent (potentially clipped)
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        goal_pos = {key.removesuffix(".pos"): val for key, val in action.items() if key.endswith(".pos")}
+
+        # Apply joint limit clipping to arm
+        for motor_name, position in goal_pos.items():
+            if motor_name in self.config.joint_limits:
+                min_limit, max_limit = self.config.joint_limits[motor_name]
+                clipped_position = max(min_limit, min(max_limit, position))
+                if clipped_position != position:
+                    logger.debug(f"Clipped {motor_name} from {position:.2f}° to {clipped_position:.2f}°")
+                goal_pos[motor_name] = clipped_position
+
+        # Cap goal position when too far away from present position.
+        # /!\ Slower fps expected due to reading from the follower.
+        if self.config.max_relative_target is not None:
+            present_pos = self.bus.sync_read("Present_Position")
+            goal_present_pos = {key: (g_pos, present_pos[key]) for key, g_pos in goal_pos.items()}
+            goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
+
+        # TODO(Steven, Pepijn): Refactor writing
+        # Motor name to index mapping for gains
+        motor_index = {
+            "joint_1": 0,
+            "joint_2": 1,
+            "joint_3": 2,
+            "joint_4": 3,
+            "joint_5": 4,
+            "joint_6": 5,
+            "joint_7": 6,
+            "gripper": 7,
+        }
+
+        # Use batch MIT control for arm (sends all commands, then collects responses)
+        commands = {}
+        for motor_name, position_degrees in goal_pos.items():
+            idx = motor_index.get(motor_name, 0)
+            # Use custom gains if provided, otherwise use config defaults
+            if custom_kp is not None and motor_name in custom_kp:
+                kp = custom_kp[motor_name]
+            else:
+                kp = (
+                    self.config.position_kp[idx]
+                    if isinstance(self.config.position_kp, list)
+                    else self.config.position_kp
+                )
+            if custom_kd is not None and motor_name in custom_kd:
+                kd = custom_kd[motor_name]
+            else:
+                kd = (
+                    self.config.position_kd[idx]
+                    if isinstance(self.config.position_kd, list)
+                    else self.config.position_kd
+                )
+            commands[motor_name] = (kp, kd, position_degrees, 0.0, 0.0)
+
+        self.bus._mit_control_batch(commands)
+
+        return {f"{motor}.pos": val for motor, val in goal_pos.items()}
+
+    def disconnect(self):
+        """Disconnect from robot."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        # Disconnect CAN bus
+        self.bus.disconnect(self.config.disable_torque_on_disconnect)
+
+        # Disconnect cameras
+        for cam in self.cameras.values():
+            cam.disconnect()
+
+        logger.info(f"{self} disconnected.")

src/lerobot/robots/unitree_g1/config_unitree_g1.py

@@ -65,3 +65,6 @@ class UnitreeG1Config(RobotConfig):
 
     # Cameras (ZMQ-based remote cameras)
     cameras: dict[str, CameraConfig] = field(default_factory=dict)
+
+    # Compensates for gravity on the unitree's arms using the arm ik solver
+    gravity_compensation: bool = False

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -18,7 +18,7 @@ from enum import IntEnum
 
 # ruff: noqa: N801, N815
 
-NUM_MOTORS = 35
+NUM_MOTORS = 29
 
 
 class G1_29_JointArmIndex(IntEnum):

src/lerobot/robots/unitree_g1/robot_kinematic_processor.py

@@ -0,0 +1,313 @@
+#!/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.
+
+import logging
+import os
+import sys
+
+import numpy as np
+
+logger = logging.getLogger(__name__)
+parent2_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+sys.path.append(parent2_dir)
+
+
+class WeightedMovingFilter:
+    def __init__(self, weights, data_size=14):
+        self._window_size = len(weights)
+        self._weights = np.array(weights)
+        self._data_size = data_size
+        self._filtered_data = np.zeros(self._data_size)
+        self._data_queue = []
+
+    def _apply_filter(self):
+        if len(self._data_queue) < self._window_size:
+            return self._data_queue[-1]
+
+        data_array = np.array(self._data_queue)
+        temp_filtered_data = np.zeros(self._data_size)
+        for i in range(self._data_size):
+            temp_filtered_data[i] = np.convolve(data_array[:, i], self._weights, mode="valid")[-1]
+
+        return temp_filtered_data
+
+    def add_data(self, new_data):
+        assert len(new_data) == self._data_size
+
+        if len(self._data_queue) > 0 and np.array_equal(
+            new_data, self._data_queue[-1]
+        ):  # skip duplicate data
+            return
+
+        if len(self._data_queue) >= self._window_size:
+            self._data_queue.pop(0)
+
+        self._data_queue.append(new_data)
+        self._filtered_data = self._apply_filter()
+
+    @property
+    def filtered_data(self):
+        return self._filtered_data
+
+
+class G1_29_ArmIK:  # noqa: N801
+    def __init__(self, unit_test=False):
+        import casadi
+        import pinocchio as pin
+        from huggingface_hub import snapshot_download
+        from pinocchio import casadi as cpin
+
+        self._pin = pin
+        np.set_printoptions(precision=5, suppress=True, linewidth=200)
+
+        self.unit_test = unit_test
+
+        self.repo_path = snapshot_download("lerobot/unitree-g1-mujoco")
+        urdf_path = os.path.join(self.repo_path, "assets", "g1_body29_hand14.urdf")
+        mesh_dir = os.path.join(self.repo_path, "assets")
+
+        self.robot = self._pin.RobotWrapper.BuildFromURDF(urdf_path, mesh_dir)
+
+        self.mixed_jointsToLockIDs = [
+            "left_hip_pitch_joint",
+            "left_hip_roll_joint",
+            "left_hip_yaw_joint",
+            "left_knee_joint",
+            "left_ankle_pitch_joint",
+            "left_ankle_roll_joint",
+            "right_hip_pitch_joint",
+            "right_hip_roll_joint",
+            "right_hip_yaw_joint",
+            "right_knee_joint",
+            "right_ankle_pitch_joint",
+            "right_ankle_roll_joint",
+            "waist_yaw_joint",
+            "waist_roll_joint",
+            "waist_pitch_joint",
+            "left_hand_thumb_0_joint",
+            "left_hand_thumb_1_joint",
+            "left_hand_thumb_2_joint",
+            "left_hand_middle_0_joint",
+            "left_hand_middle_1_joint",
+            "left_hand_index_0_joint",
+            "left_hand_index_1_joint",
+            "right_hand_thumb_0_joint",
+            "right_hand_thumb_1_joint",
+            "right_hand_thumb_2_joint",
+            "right_hand_index_0_joint",
+            "right_hand_index_1_joint",
+            "right_hand_middle_0_joint",
+            "right_hand_middle_1_joint",
+        ]
+
+        self.reduced_robot = self.robot.buildReducedRobot(
+            list_of_joints_to_lock=self.mixed_jointsToLockIDs,
+            reference_configuration=np.array([0.0] * self.robot.model.nq),
+        )
+
+        # Arm joint names in G1 motor order (G1_29_JointArmIndex)
+        self._arm_joint_names_g1 = [
+            "left_shoulder_pitch_joint",
+            "left_shoulder_roll_joint",
+            "left_shoulder_yaw_joint",
+            "left_elbow_joint",
+            "left_wrist_roll_joint",
+            "left_wrist_pitch_joint",
+            "left_wrist_yaw_joint",
+            "right_shoulder_pitch_joint",
+            "right_shoulder_roll_joint",
+            "right_shoulder_yaw_joint",
+            "right_elbow_joint",
+            "right_wrist_roll_joint",
+            "right_wrist_pitch_joint",
+            "right_wrist_yaw_joint",
+        ]
+        # Pinocchio uses its own joint order in q; build index mapping.
+        self._arm_joint_names_pin = sorted(
+            self._arm_joint_names_g1,
+            key=lambda name: self.reduced_robot.model.idx_qs[self.reduced_robot.model.getJointId(name)],
+        )
+        logger.info(f"Pinocchio arm joint order: {self._arm_joint_names_pin}")
+        self._arm_reorder_g1_to_pin = [
+            self._arm_joint_names_g1.index(name) for name in self._arm_joint_names_pin
+        ]
+        # Inverse mapping to return tau in G1 motor order.
+        self._arm_reorder_pin_to_g1 = np.argsort(self._arm_reorder_g1_to_pin)
+
+        self.reduced_robot.model.addFrame(
+            self._pin.Frame(
+                "L_ee",
+                self.reduced_robot.model.getJointId("left_wrist_yaw_joint"),
+                self._pin.SE3(np.eye(3), np.array([0.05, 0, 0]).T),
+                self._pin.FrameType.OP_FRAME,
+            )
+        )
+
+        self.reduced_robot.model.addFrame(
+            self._pin.Frame(
+                "R_ee",
+                self.reduced_robot.model.getJointId("right_wrist_yaw_joint"),
+                self._pin.SE3(np.eye(3), np.array([0.05, 0, 0]).T),
+                self._pin.FrameType.OP_FRAME,
+            )
+        )
+
+        # Creating Casadi models and data for symbolic computing
+        self.cmodel = cpin.Model(self.reduced_robot.model)
+        self.cdata = self.cmodel.createData()
+
+        # Creating symbolic variables
+        self.cq = casadi.SX.sym("q", self.reduced_robot.model.nq, 1)
+        self.cTf_l = casadi.SX.sym("tf_l", 4, 4)
+        self.cTf_r = casadi.SX.sym("tf_r", 4, 4)
+        cpin.framesForwardKinematics(self.cmodel, self.cdata, self.cq)
+
+        # Get the hand joint ID and define the error function
+        self.L_hand_id = self.reduced_robot.model.getFrameId("L_ee")
+        self.R_hand_id = self.reduced_robot.model.getFrameId("R_ee")
+
+        self.translational_error = casadi.Function(
+            "translational_error",
+            [self.cq, self.cTf_l, self.cTf_r],
+            [
+                casadi.vertcat(
+                    self.cdata.oMf[self.L_hand_id].translation - self.cTf_l[:3, 3],
+                    self.cdata.oMf[self.R_hand_id].translation - self.cTf_r[:3, 3],
+                )
+            ],
+        )
+        self.rotational_error = casadi.Function(
+            "rotational_error",
+            [self.cq, self.cTf_l, self.cTf_r],
+            [
+                casadi.vertcat(
+                    cpin.log3(self.cdata.oMf[self.L_hand_id].rotation @ self.cTf_l[:3, :3].T),
+                    cpin.log3(self.cdata.oMf[self.R_hand_id].rotation @ self.cTf_r[:3, :3].T),
+                )
+            ],
+        )
+
+        # Defining the optimization problem
+        self.opti = casadi.Opti()
+        self.var_q = self.opti.variable(self.reduced_robot.model.nq)
+        self.var_q_last = self.opti.parameter(self.reduced_robot.model.nq)  # for smooth
+        self.param_tf_l = self.opti.parameter(4, 4)
+        self.param_tf_r = self.opti.parameter(4, 4)
+        self.translational_cost = casadi.sumsqr(
+            self.translational_error(self.var_q, self.param_tf_l, self.param_tf_r)
+        )
+        self.rotation_cost = casadi.sumsqr(
+            self.rotational_error(self.var_q, self.param_tf_l, self.param_tf_r)
+        )
+        self.regularization_cost = casadi.sumsqr(self.var_q)
+        self.smooth_cost = casadi.sumsqr(self.var_q - self.var_q_last)
+
+        # Setting optimization constraints and goals
+        self.opti.subject_to(
+            self.opti.bounded(
+                self.reduced_robot.model.lowerPositionLimit,
+                self.var_q,
+                self.reduced_robot.model.upperPositionLimit,
+            )
+        )
+        self.opti.minimize(
+            50 * self.translational_cost
+            + self.rotation_cost
+            + 0.02 * self.regularization_cost
+            + 0.1 * self.smooth_cost
+        )
+
+        opts = {
+            "ipopt": {"print_level": 0, "max_iter": 50, "tol": 1e-6},
+            "print_time": False,  # print or not
+            "calc_lam_p": False,  # https://github.com/casadi/casadi/wiki/FAQ:-Why-am-I-getting-%22NaN-detected%22in-my-optimization%3F
+        }
+        self.opti.solver("ipopt", opts)
+
+        self.init_data = np.zeros(self.reduced_robot.model.nq)
+        self.smooth_filter = WeightedMovingFilter(np.array([0.4, 0.3, 0.2, 0.1]), 14)
+
+    def solve_ik(self, left_wrist, right_wrist, current_lr_arm_motor_q=None, current_lr_arm_motor_dq=None):
+        if current_lr_arm_motor_q is not None:
+            self.init_data = current_lr_arm_motor_q
+        self.opti.set_initial(self.var_q, self.init_data)
+
+        self.opti.set_value(self.param_tf_l, left_wrist)
+        self.opti.set_value(self.param_tf_r, right_wrist)
+        self.opti.set_value(self.var_q_last, self.init_data)  # for smooth
+
+        try:
+            self.opti.solve()
+
+            sol_q = self.opti.value(self.var_q)
+            self.smooth_filter.add_data(sol_q)
+            sol_q = self.smooth_filter.filtered_data
+
+            if current_lr_arm_motor_dq is not None:
+                v = current_lr_arm_motor_dq * 0.0
+            else:
+                v = (sol_q - self.init_data) * 0.0
+
+            self.init_data = sol_q
+
+            sol_tauff = self._pin.rnea(
+                self.reduced_robot.model,
+                self.reduced_robot.data,
+                sol_q,
+                v,
+                np.zeros(self.reduced_robot.model.nv),
+            )
+
+            return sol_q, sol_tauff
+
+        except Exception as e:
+            logger.error(f"ERROR in convergence, plotting debug info.{e}")
+
+            sol_q = self.opti.debug.value(self.var_q)
+            self.smooth_filter.add_data(sol_q)
+            sol_q = self.smooth_filter.filtered_data
+
+            if current_lr_arm_motor_dq is not None:
+                v = current_lr_arm_motor_dq * 0.0
+            else:
+                v = (sol_q - self.init_data) * 0.0
+
+            self.init_data = sol_q
+
+            logger.error(
+                f"sol_q:{sol_q} \nmotorstate: \n{current_lr_arm_motor_q} \nleft_pose: \n{left_wrist} \nright_pose: \n{right_wrist}"
+            )
+
+            return current_lr_arm_motor_q, np.zeros(self.reduced_robot.model.nv)
+
+    def solve_tau(self, current_lr_arm_motor_q=None, current_lr_arm_motor_dq=None):
+        try:
+            q_g1 = np.array(current_lr_arm_motor_q, dtype=float)
+            if q_g1.shape[0] != len(self._arm_joint_names_g1):
+                raise ValueError(f"Expected {len(self._arm_joint_names_g1)} arm joints, got {q_g1.shape[0]}")
+            q_pin = q_g1[self._arm_reorder_g1_to_pin]
+            sol_tauff = self._pin.rnea(
+                self.reduced_robot.model,
+                self.reduced_robot.data,
+                q_pin,
+                np.zeros(self.reduced_robot.model.nv),
+                np.zeros(self.reduced_robot.model.nv),
+            )
+            return sol_tauff[self._arm_reorder_pin_to_g1]
+
+        except Exception as e:
+            logger.error(f"ERROR in convergence, plotting debug info.{e}")
+            return np.zeros(self.reduced_robot.model.nv)

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -27,7 +27,8 @@ import numpy as np
 from lerobot.cameras.utils import make_cameras_from_configs
 from lerobot.envs.factory import make_env
 from lerobot.processor import RobotAction, RobotObservation
-from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointArmIndex, G1_29_JointIndex
+from lerobot.robots.unitree_g1.robot_kinematic_processor import G1_29_ArmIK
 
 from ..robot import Robot
 from .config_unitree_g1 import UnitreeG1Config
@@ -127,6 +128,8 @@ class UnitreeG1(Robot):
         self.subscribe_thread = None
         self.remote_controller = self.RemoteController()
 
+        self.arm_ik = G1_29_ArmIK()
+
     def _subscribe_motor_state(self):  # polls robot state @ 250Hz
         while not self._shutdown_event.is_set():
             start_time = time.time()
@@ -361,6 +364,20 @@ class UnitreeG1(Robot):
                 self.msg.motor_cmd[motor.value].kd = self.kd[motor.value]
                 self.msg.motor_cmd[motor.value].tau = 0
 
+        if self.config.gravity_compensation:
+            # Build action_np from motor commands (arm joints are indices 15-28, local indices 0-13)
+            action_np = np.zeros(14)
+            arm_start_idx = G1_29_JointArmIndex.kLeftShoulderPitch.value  # 15
+            for joint in G1_29_JointArmIndex:
+                local_idx = joint.value - arm_start_idx
+                action_np[local_idx] = self.msg.motor_cmd[joint.value].q
+            tau = self.arm_ik.solve_tau(action_np)
+
+            # Apply tau back to motor commands
+            for joint in G1_29_JointArmIndex:
+                local_idx = joint.value - arm_start_idx
+                self.msg.motor_cmd[joint.value].tau = tau[local_idx]
+
         self.msg.crc = self.crc.Crc(self.msg)
         self.lowcmd_publisher.Write(self.msg)
         return action

src/lerobot/robots/utils.py

@@ -60,6 +60,14 @@ def make_robot_from_config(config: RobotConfig) -> Robot:
         from .reachy2 import Reachy2Robot
 
         return Reachy2Robot(config)
+    elif config.type == "openarm_follower":
+        from .openarm_follower import OpenArmFollower
+
+        return OpenArmFollower(config)
+    elif config.type == "bi_openarm_follower":
+        from .bi_openarm_follower import BiOpenArmFollower
+
+        return BiOpenArmFollower(config)
     elif config.type == "mock_robot":
         from tests.mocks.mock_robot import MockRobot
 

src/lerobot/scripts/lerobot_calibrate.py

@@ -36,23 +36,28 @@ from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraCon
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
+    bi_openarm_follower,
     bi_so_follower,
     hope_jr,
     koch_follower,
     lekiwi,
     make_robot_from_config,
     omx_follower,
+    openarm_follower,
     so_follower,
 )
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
+    bi_openarm_leader,
     bi_so_leader,
     homunculus,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
+    openarm_leader,
     so_leader,
+    unitree_g1,
 )
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.utils import init_logging
@@ -81,8 +86,11 @@ def calibrate(cfg: CalibrateConfig):
         device = make_teleoperator_from_config(cfg.device)
 
     device.connect(calibrate=False)
-    device.calibrate()
-    device.disconnect()
+
+    try:
+        device.calibrate()
+    finally:
+        device.disconnect()
 
 
 def main():

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -18,7 +18,7 @@
 Edit LeRobot datasets using various transformation tools.
 
 This script allows you to delete episodes, split datasets, merge datasets,
-remove features, and convert image datasets to video format.
+remove features, modify tasks, and convert image datasets to video format.
 When new_repo_id is specified, creates a new dataset.
 
 Usage Examples:
@@ -66,6 +66,25 @@ Remove camera feature:
         --operation.type remove_feature \
         --operation.feature_names "['observation.images.top']"
 
+Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --operation.type modify_tasks \
+        --operation.new_task "Pick up the cube and place it"
+
+Modify tasks - set different tasks for specific episodes (WARNING: modifies in-place):
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --operation.type modify_tasks \
+        --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'
+
+Modify tasks - set default task with overrides for specific episodes (WARNING: modifies in-place):
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --operation.type modify_tasks \
+        --operation.new_task "Default task" \
+        --operation.episode_tasks '{"5": "Special task for episode 5"}'
+
 Convert image dataset to video format and save locally:
     python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht_image \
@@ -100,6 +119,7 @@ from lerobot.datasets.dataset_tools import (
     convert_image_to_video_dataset,
     delete_episodes,
     merge_datasets,
+    modify_tasks,
     remove_feature,
     split_dataset,
 )
@@ -132,6 +152,13 @@ class RemoveFeatureConfig:
     feature_names: list[str] | None = None
 
 
+@dataclass
+class ModifyTasksConfig:
+    type: str = "modify_tasks"
+    new_task: str | None = None
+    episode_tasks: dict[str, str] | None = None
+
+
 @dataclass
 class ConvertImageToVideoConfig:
     type: str = "convert_image_to_video"
@@ -151,7 +178,12 @@ class ConvertImageToVideoConfig:
 class EditDatasetConfig:
     repo_id: str
     operation: (
-        DeleteEpisodesConfig | SplitConfig | MergeConfig | RemoveFeatureConfig | ConvertImageToVideoConfig
+        DeleteEpisodesConfig
+        | SplitConfig
+        | MergeConfig
+        | RemoveFeatureConfig
+        | ModifyTasksConfig
+        | ConvertImageToVideoConfig
     )
     root: str | None = None
     new_repo_id: str | None = None
@@ -296,6 +328,48 @@ def handle_remove_feature(cfg: EditDatasetConfig) -> None:
         LeRobotDataset(output_repo_id, root=output_dir).push_to_hub()
 
 
+def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
+    if not isinstance(cfg.operation, ModifyTasksConfig):
+        raise ValueError("Operation config must be ModifyTasksConfig")
+
+    new_task = cfg.operation.new_task
+    episode_tasks_raw = cfg.operation.episode_tasks
+
+    if new_task is None and episode_tasks_raw is None:
+        raise ValueError("Must specify at least one of new_task or episode_tasks for modify_tasks operation")
+
+    # Warn about in-place modification behavior
+    if cfg.new_repo_id is not None:
+        logging.warning("modify_tasks modifies datasets in-place. The --new_repo_id parameter is ignored.")
+
+    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
+    logging.warning(f"Modifying dataset in-place at {dataset.root}. Original data will be overwritten.")
+
+    # Convert episode_tasks keys from string to int if needed (CLI passes strings)
+    episode_tasks: dict[int, str] | None = None
+    if episode_tasks_raw is not None:
+        episode_tasks = {int(k): v for k, v in episode_tasks_raw.items()}
+
+    logging.info(f"Modifying tasks in {cfg.repo_id}")
+    if new_task:
+        logging.info(f"  Default task: '{new_task}'")
+    if episode_tasks:
+        logging.info(f"  Episode-specific tasks: {episode_tasks}")
+
+    modified_dataset = modify_tasks(
+        dataset,
+        new_task=new_task,
+        episode_tasks=episode_tasks,
+    )
+
+    logging.info(f"Dataset modified at {dataset.root}")
+    logging.info(f"Tasks: {list(modified_dataset.meta.tasks.index)}")
+
+    if cfg.push_to_hub:
+        logging.info(f"Pushing to hub as {cfg.repo_id}")
+        modified_dataset.push_to_hub()
+
+
 def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
     # Note: Parser may create any config type with the right fields, so we access fields directly
     # instead of checking isinstance()
@@ -371,12 +445,14 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
         handle_merge(cfg)
     elif operation_type == "remove_feature":
         handle_remove_feature(cfg)
+    elif operation_type == "modify_tasks":
+        handle_modify_tasks(cfg)
     elif operation_type == "convert_image_to_video":
         handle_convert_image_to_video(cfg)
     else:
         raise ValueError(
             f"Unknown operation type: {operation_type}\n"
-            f"Available operations: delete_episodes, split, merge, remove_feature, convert_to_video"
+            f"Available operations: delete_episodes, split, merge, remove_feature, modify_tasks, convert_image_to_video"
         )
 
 

src/lerobot/scripts/lerobot_find_joint_limits.py

@@ -44,19 +44,23 @@ import numpy as np
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.robots import (  # noqa: F401
     RobotConfig,
+    bi_openarm_follower,
     bi_so_follower,
     koch_follower,
     make_robot_from_config,
     omx_follower,
+    openarm_follower,
     so_follower,
 )
 from lerobot.teleoperators import (  # noqa: F401
     TeleoperatorConfig,
+    bi_openarm_leader,
     bi_so_leader,
     gamepad,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
+    openarm_leader,
     so_leader,
 )
 from lerobot.utils.robot_utils import precise_sleep

src/lerobot/scripts/lerobot_record.py

@@ -98,26 +98,31 @@ from lerobot.processor.rename_processor import rename_stats
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
+    bi_openarm_follower,
     bi_so_follower,
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
     make_robot_from_config,
     omx_follower,
+    openarm_follower,
     reachy2,
     so_follower,
-    unitree_g1,
+    unitree_g1 as unitree_g1_robot,
 )
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
+    bi_openarm_leader,
     bi_so_leader,
     homunculus,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
+    openarm_leader,
     reachy2_teleoperator,
     so_leader,
+    unitree_g1,
 )
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop
 from lerobot.utils.constants import ACTION, OBS_STR
@@ -179,6 +184,9 @@ class DatasetRecordConfig:
     vcodec: str = "libsvtav1"
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
+    # Expert noise injection scale. Noise is added to robot actions but not recorded in dataset.
+    # This forces recovery behavior for more robust learned policies. 0.0 means no noise. #https://arxiv.org/pdf/1703.09327, https://arxiv.org/abs/2507.09061
+    noise_scale: float = 0.0
 
     def __post_init__(self):
         if self.single_task is None:
@@ -278,6 +286,7 @@ def record_loop(
     single_task: str | None = None,
     display_data: bool = False,
     display_compressed_images: bool = False,
+    noise_scale: float = 0.0,
 ):
     if dataset is not None and dataset.fps != fps:
         raise ValueError(f"The dataset fps should be equal to requested fps ({dataset.fps} != {fps}).")
@@ -375,18 +384,27 @@ def record_loop(
             action_values = act_processed_teleop
             robot_action_to_send = robot_action_processor((act_processed_teleop, obs))
 
+        # Write clean action to dataset (before noise injection)
+        if dataset is not None:
+            action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
+            frame = {**observation_frame, **action_frame, "task": single_task}
+            dataset.add_frame(frame)
+
+        # Expert noise injection: add noise to motor commands but not to recorded labels
+        if noise_scale > 0:
+            import torch
+
+            for key in robot_action_to_send:
+                if isinstance(robot_action_to_send[key], torch.Tensor):
+                    noise = torch.randn_like(robot_action_to_send[key]) * noise_scale
+                    robot_action_to_send[key] = robot_action_to_send[key] + noise
+
         # Send action to robot
         # Action can eventually be clipped using `max_relative_target`,
         # so action actually sent is saved in the dataset. action = postprocessor.process(action)
         # TODO(steven, pepijn, adil): we should use a pipeline step to clip the action, so the sent action is the action that we input to the robot.
         _sent_action = robot.send_action(robot_action_to_send)
 
-        # Write to dataset
-        if dataset is not None:
-            action_frame = build_dataset_frame(dataset.features, action_values, prefix=ACTION)
-            frame = {**observation_frame, **action_frame, "task": single_task}
-            dataset.add_frame(frame)
-
         if display_data:
             log_rerun_data(
                 observation=obs_processed, action=action_values, compress_images=display_compressed_images
@@ -505,6 +523,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                     single_task=cfg.dataset.single_task,
                     display_data=cfg.display_data,
                     display_compressed_images=display_compressed_images,
+                    noise_scale=cfg.dataset.noise_scale,
                 )
 
                 # Execute a few seconds without recording to give time to manually reset the environment

src/lerobot/scripts/lerobot_replay.py

@@ -53,12 +53,14 @@ from lerobot.processor import (
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
+    bi_openarm_follower,
     bi_so_follower,
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
     make_robot_from_config,
     omx_follower,
+    openarm_follower,
     reachy2,
     so_follower,
     unitree_g1,
@@ -108,25 +110,26 @@ def replay(cfg: ReplayConfig):
 
     robot.connect()
 
-    log_say("Replaying episode", cfg.play_sounds, blocking=True)
-    for idx in range(len(episode_frames)):
-        start_episode_t = time.perf_counter()
+    try:
+        log_say("Replaying episode", cfg.play_sounds, blocking=True)
+        for idx in range(len(episode_frames)):
+            start_episode_t = time.perf_counter()
 
-        action_array = actions[idx][ACTION]
-        action = {}
-        for i, name in enumerate(dataset.features[ACTION]["names"]):
-            action[name] = action_array[i]
+            action_array = actions[idx][ACTION]
+            action = {}
+            for i, name in enumerate(dataset.features[ACTION]["names"]):
+                action[name] = action_array[i]
 
-        robot_obs = robot.get_observation()
+            robot_obs = robot.get_observation()
 
-        processed_action = robot_action_processor((action, robot_obs))
+            processed_action = robot_action_processor((action, robot_obs))
 
-        _ = robot.send_action(processed_action)
+            _ = robot.send_action(processed_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()
 
 
 def main():

src/lerobot/scripts/lerobot_setup_can.py

@@ -0,0 +1,360 @@
+# 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.
+
+"""
+Setup and debug CAN interfaces for Damiao motors (e.g., OpenArms).
+
+Examples:
+
+Setup CAN interfaces with CAN FD:
+```shell
+lerobot-setup-can --mode=setup --interfaces=can0,can1,can2,can3
+```
+
+Test motors on a single interface:
+```shell
+lerobot-setup-can --mode=test --interfaces=can0
+```
+
+Test motors on all interfaces:
+```shell
+lerobot-setup-can --mode=test --interfaces=can0,can1,can2,can3
+```
+
+Speed test:
+```shell
+lerobot-setup-can --mode=speed --interfaces=can0
+```
+"""
+
+import subprocess
+import sys
+import time
+from dataclasses import dataclass, field
+
+import draccus
+
+from lerobot.utils.import_utils import is_package_available
+
+MOTOR_NAMES = {
+    0x01: "joint_1",
+    0x02: "joint_2",
+    0x03: "joint_3",
+    0x04: "joint_4",
+    0x05: "joint_5",
+    0x06: "joint_6",
+    0x07: "joint_7",
+    0x08: "gripper",
+}
+
+
+@dataclass
+class CANSetupConfig:
+    mode: str = "test"
+    interfaces: str = "can0"  # Comma-separated, e.g. "can0,can1,can2,can3"
+    bitrate: int = 1000000
+    data_bitrate: int = 5000000
+    use_fd: bool = True
+    motor_ids: list[int] = field(default_factory=lambda: list(range(0x01, 0x09)))
+    timeout: float = 1.0
+    speed_iterations: int = 100
+
+    def get_interfaces(self) -> list[str]:
+        return [i.strip() for i in self.interfaces.split(",") if i.strip()]
+
+
+def check_interface_status(interface: str) -> tuple[bool, str, bool]:
+    """Check if CAN interface is UP and configured."""
+    try:
+        result = subprocess.run(["ip", "link", "show", interface], capture_output=True, text=True)  # nosec B607
+        if result.returncode != 0:
+            return False, "Interface not found", False
+
+        output = result.stdout
+        is_up = "UP" in output
+        is_fd = "fd on" in output.lower() or "canfd" in output.lower()
+        status = "UP" if is_up else "DOWN"
+        if is_fd:
+            status += " (CAN FD)"
+
+        return is_up, status, is_fd
+    except FileNotFoundError:
+        return False, "ip command not found", False
+
+
+def setup_interface(interface: str, bitrate: int, data_bitrate: int, use_fd: bool) -> bool:
+    """Configure a CAN interface."""
+    try:
+        subprocess.run(["sudo", "ip", "link", "set", interface, "down"], check=False, capture_output=True)  # nosec B607
+
+        cmd = ["sudo", "ip", "link", "set", interface, "type", "can", "bitrate", str(bitrate)]
+        if use_fd:
+            cmd.extend(["dbitrate", str(data_bitrate), "fd", "on"])
+
+        result = subprocess.run(cmd, capture_output=True, text=True)  # nosec B607
+        if result.returncode != 0:
+            print(f"  ✗ Failed to configure: {result.stderr}")
+            return False
+
+        result = subprocess.run(  # nosec B607
+            ["sudo", "ip", "link", "set", interface, "up"], capture_output=True, text=True
+        )
+        if result.returncode != 0:
+            print(f"  ✗ Failed to bring up: {result.stderr}")
+            return False
+
+        return True
+    except Exception as e:
+        print(f"  ✗ Error: {e}")
+        return False
+
+
+def test_motor(bus, motor_id: int, timeout: float, use_fd: bool):
+    """Test a single motor and return responses."""
+    import can
+
+    enable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+        is_extended_id=False,
+        is_fd=use_fd,
+    )
+
+    try:
+        bus.send(enable_msg)
+    except Exception as e:
+        return None, f"Send error: {e}"
+
+    responses = []
+    start_time = time.time()
+
+    while time.time() - start_time < timeout:
+        msg = bus.recv(timeout=0.1)
+        if msg:
+            responses.append((msg.arbitration_id, msg.data.hex(), getattr(msg, "is_fd", False)))
+
+    disable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFD],
+        is_extended_id=False,
+        is_fd=use_fd,
+    )
+    try:
+        bus.send(disable_msg)
+    except Exception:
+        print(f"Error sending message to motor 0x{motor_id:02X}")
+
+    return responses, None
+
+
+def test_interface(cfg: CANSetupConfig, interface: str):
+    """Test all motors on a CAN interface."""
+    import can
+
+    is_up, status, _ = check_interface_status(interface)
+    print(f"\n{interface}: {status}")
+
+    if not is_up:
+        print(f"  ⚠ Interface is not UP. Run: lerobot-setup-can --mode=setup --interfaces {interface}")
+        return {}
+
+    try:
+        kwargs = {"channel": interface, "interface": "socketcan", "bitrate": cfg.bitrate}
+        if cfg.use_fd:
+            kwargs.update({"data_bitrate": cfg.data_bitrate, "fd": True})
+        bus = can.interface.Bus(**kwargs)
+    except Exception as e:
+        print(f"  ✗ Connection failed: {e}")
+        return {}
+
+    results = {}
+    try:
+        while bus.recv(timeout=0.01):
+            pass
+
+        for motor_id in cfg.motor_ids:
+            motor_name = MOTOR_NAMES.get(motor_id, f"motor_0x{motor_id:02X}")
+            responses, error = test_motor(bus, motor_id, cfg.timeout, cfg.use_fd)
+
+            if error:
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ {error}")
+                results[motor_id] = {"found": False, "error": error}
+            elif responses:
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✓ FOUND")
+                for resp_id, data, is_fd in responses:
+                    fd_flag = " [FD]" if is_fd else ""
+                    print(f"    → Response 0x{resp_id:02X}{fd_flag}: {data}")
+                results[motor_id] = {"found": True, "responses": responses}
+            else:
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ No response")
+                results[motor_id] = {"found": False}
+
+            time.sleep(0.05)
+    finally:
+        bus.shutdown()
+
+    found = sum(1 for r in results.values() if r.get("found"))
+    print(f"\n  Summary: {found}/{len(cfg.motor_ids)} motors found")
+    return results
+
+
+def speed_test(cfg: CANSetupConfig, interface: str):
+    """Test communication speed with motors."""
+    import can
+
+    is_up, status, _ = check_interface_status(interface)
+    if not is_up:
+        print(f"{interface}: {status} - skipping")
+        return
+
+    print(f"\n{interface}: Running speed test ({cfg.speed_iterations} iterations)...")
+
+    try:
+        kwargs = {"channel": interface, "interface": "socketcan", "bitrate": cfg.bitrate}
+        if cfg.use_fd:
+            kwargs.update({"data_bitrate": cfg.data_bitrate, "fd": True})
+        bus = can.interface.Bus(**kwargs)
+    except Exception as e:
+        print(f"  ✗ Connection failed: {e}")
+        return
+
+    responding_motor = None
+    for motor_id in cfg.motor_ids:
+        responses, _ = test_motor(bus, motor_id, 0.5, cfg.use_fd)
+        if responses:
+            responding_motor = motor_id
+            break
+
+    if not responding_motor:
+        print("  ✗ No responding motors found")
+        bus.shutdown()
+        return
+
+    print(f"  Testing with motor 0x{responding_motor:02X}...")
+    latencies = []
+
+    for _ in range(cfg.speed_iterations):
+        start = time.perf_counter()
+        msg = can.Message(
+            arbitration_id=responding_motor,
+            data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+            is_extended_id=False,
+            is_fd=cfg.use_fd,
+        )
+        bus.send(msg)
+        resp = bus.recv(timeout=0.1)
+        if resp:
+            latencies.append((time.perf_counter() - start) * 1000)
+
+    bus.shutdown()
+
+    if latencies:
+        avg_latency = sum(latencies) / len(latencies)
+        hz = 1000.0 / avg_latency if avg_latency > 0 else 0
+        print(f"  ✓ Success rate: {len(latencies)}/{cfg.speed_iterations}")
+        print(f"  ✓ Avg latency: {avg_latency:.2f} ms")
+        print(f"  ✓ Max frequency: {hz:.1f} Hz")
+    else:
+        print("  ✗ No successful responses")
+
+
+def run_setup(cfg: CANSetupConfig):
+    """Setup CAN interfaces."""
+    print("=" * 50)
+    print("CAN Interface Setup")
+    print("=" * 50)
+    print(f"Mode: {'CAN FD' if cfg.use_fd else 'CAN 2.0'}")
+    print(f"Bitrate: {cfg.bitrate / 1_000_000:.1f} Mbps")
+    if cfg.use_fd:
+        print(f"Data bitrate: {cfg.data_bitrate / 1_000_000:.1f} Mbps")
+    print()
+
+    interfaces = cfg.get_interfaces()
+    for interface in interfaces:
+        print(f"Configuring {interface}...")
+        if setup_interface(interface, cfg.bitrate, cfg.data_bitrate, cfg.use_fd):
+            is_up, status, _ = check_interface_status(interface)
+            print(f"  ✓ {interface}: {status}")
+        else:
+            print(f"  ✗ {interface}: Failed")
+
+    print("\nSetup complete!")
+    print("\nNext: Test motors with:")
+    print(f"  lerobot-setup-can --mode=test --interfaces {','.join(interfaces)}")
+
+
+def run_test(cfg: CANSetupConfig):
+    """Test motors on CAN interfaces."""
+    print("=" * 50)
+    print("CAN Motor Test")
+    print("=" * 50)
+    print(f"Testing motors 0x{min(cfg.motor_ids):02X}-0x{max(cfg.motor_ids):02X}")
+    print(f"Mode: {'CAN FD' if cfg.use_fd else 'CAN 2.0'}")
+    print()
+
+    interfaces = cfg.get_interfaces()
+    all_results = {}
+    for interface in interfaces:
+        all_results[interface] = test_interface(cfg, interface)
+
+    total_found = sum(sum(1 for r in res.values() if r.get("found")) for res in all_results.values())
+
+    print("\n" + "=" * 50)
+    print("Summary")
+    print("=" * 50)
+    print(f"Total motors found: {total_found}")
+
+    if total_found == 0:
+        print("\n⚠ No motors found! Check:")
+        print("  1. Motors are powered (24V)")
+        print("  2. CAN wiring (CANH, CANL, GND)")
+        print("  3. Motor timeout parameter > 0 (use Damiao tools)")
+        print("  4. 120Ω termination at both cable ends")
+        print(f"  5. Interface configured: lerobot-setup-can --mode=setup --interfaces {interfaces[0]}")
+
+
+def run_speed(cfg: CANSetupConfig):
+    """Run speed tests on CAN interfaces."""
+    print("=" * 50)
+    print("CAN Speed Test")
+    print("=" * 50)
+
+    for interface in cfg.get_interfaces():
+        speed_test(cfg, interface)
+
+
+@draccus.wrap()
+def setup_can(cfg: CANSetupConfig):
+    if not is_package_available("can"):
+        print("Error: python-can not installed. Install with: pip install python-can")
+        sys.exit(1)
+
+    if cfg.mode == "setup":
+        run_setup(cfg)
+    elif cfg.mode == "test":
+        run_test(cfg)
+    elif cfg.mode == "speed":
+        run_speed(cfg)
+    else:
+        print(f"Unknown mode: {cfg.mode}")
+        print("Available modes: setup, test, speed")
+        sys.exit(1)
+
+
+def main():
+    setup_can()
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/scripts/lerobot_teleoperate.py

@@ -70,18 +70,22 @@ from lerobot.processor import (
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
+    bi_openarm_follower,
     bi_so_follower,
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
     make_robot_from_config,
     omx_follower,
+    openarm_follower,
     reachy2,
     so_follower,
+    unitree_g1 as unitree_g1_robot,
 )
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
+    bi_openarm_leader,
     bi_so_leader,
     gamepad,
     homunculus,
@@ -89,8 +93,10 @@ from lerobot.teleoperators import (  # noqa: F401
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
+    openarm_leader,
     reachy2_teleoperator,
     so_leader,
+    unitree_g1,
 )
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.robot_utils import precise_sleep

src/lerobot/teleoperators/bi_openarm_leader/__init__.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .bi_openarm_leader import BiOpenArmLeader
+from .config_bi_openarm_leader import BiOpenArmLeaderConfig
+
+__all__ = ["BiOpenArmLeader", "BiOpenArmLeaderConfig"]

src/lerobot/teleoperators/bi_openarm_leader/bi_openarm_leader.py

@@ -0,0 +1,131 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from functools import cached_property
+
+from lerobot.processor import RobotAction
+from lerobot.teleoperators.openarm_leader import OpenArmLeaderConfig
+
+from ..openarm_leader import OpenArmLeader
+from ..teleoperator import Teleoperator
+from .config_bi_openarm_leader import BiOpenArmLeaderConfig
+
+logger = logging.getLogger(__name__)
+
+
+class BiOpenArmLeader(Teleoperator):
+    """
+    Bimanual OpenArm Leader Arms
+    """
+
+    config_class = BiOpenArmLeaderConfig
+    name = "bi_openarm_leader"
+
+    def __init__(self, config: BiOpenArmLeaderConfig):
+        super().__init__(config)
+        self.config = config
+
+        left_arm_config = OpenArmLeaderConfig(
+            id=f"{config.id}_left" if config.id else None,
+            calibration_dir=config.calibration_dir,
+            port=config.left_arm_config.port,
+            can_interface=config.left_arm_config.can_interface,
+            use_can_fd=config.left_arm_config.use_can_fd,
+            can_bitrate=config.left_arm_config.can_bitrate,
+            can_data_bitrate=config.left_arm_config.can_data_bitrate,
+            motor_config=config.left_arm_config.motor_config,
+            manual_control=config.left_arm_config.manual_control,
+            position_kd=config.left_arm_config.position_kd,
+            position_kp=config.left_arm_config.position_kp,
+        )
+
+        right_arm_config = OpenArmLeaderConfig(
+            id=f"{config.id}_right" if config.id else None,
+            calibration_dir=config.calibration_dir,
+            port=config.right_arm_config.port,
+            can_interface=config.right_arm_config.can_interface,
+            use_can_fd=config.right_arm_config.use_can_fd,
+            can_bitrate=config.right_arm_config.can_bitrate,
+            can_data_bitrate=config.right_arm_config.can_data_bitrate,
+            motor_config=config.right_arm_config.motor_config,
+            manual_control=config.right_arm_config.manual_control,
+            position_kd=config.right_arm_config.position_kd,
+            position_kp=config.right_arm_config.position_kp,
+        )
+
+        self.left_arm = OpenArmLeader(left_arm_config)
+        self.right_arm = OpenArmLeader(right_arm_config)
+
+    @cached_property
+    def action_features(self) -> dict[str, type]:
+        left_arm_features = self.left_arm.action_features
+        right_arm_features = self.right_arm.action_features
+
+        return {
+            **{f"left_{k}": v for k, v in left_arm_features.items()},
+            **{f"right_{k}": v for k, v in right_arm_features.items()},
+        }
+
+    @cached_property
+    def feedback_features(self) -> dict[str, type]:
+        return {}
+
+    @property
+    def is_connected(self) -> bool:
+        return self.left_arm.is_connected and self.right_arm.is_connected
+
+    def connect(self, calibrate: bool = True) -> None:
+        self.left_arm.connect(calibrate)
+        self.right_arm.connect(calibrate)
+
+    @property
+    def is_calibrated(self) -> bool:
+        return self.left_arm.is_calibrated and self.right_arm.is_calibrated
+
+    def calibrate(self) -> None:
+        self.left_arm.calibrate()
+        self.right_arm.calibrate()
+
+    def configure(self) -> None:
+        self.left_arm.configure()
+        self.right_arm.configure()
+
+    def setup_motors(self) -> None:
+        raise NotImplementedError(
+            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
+        )
+
+    def get_action(self) -> RobotAction:
+        action_dict = {}
+
+        # Add "left_" prefix
+        left_action = self.left_arm.get_action()
+        action_dict.update({f"left_{key}": value for key, value in left_action.items()})
+
+        # Add "right_" prefix
+        right_action = self.right_arm.get_action()
+        action_dict.update({f"right_{key}": value for key, value in right_action.items()})
+
+        return action_dict
+
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        # TODO: Implement force feedback
+        raise NotImplementedError
+
+    def disconnect(self) -> None:
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()

src/lerobot/teleoperators/bi_openarm_leader/config_bi_openarm_leader.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from lerobot.teleoperators.openarm_leader import OpenArmLeaderConfigBase
+
+from ..config import TeleoperatorConfig
+
+
+@TeleoperatorConfig.register_subclass("bi_openarm_leader")
+@dataclass
+class BiOpenArmLeaderConfig(TeleoperatorConfig):
+    """Configuration class for Bi OpenArm Follower robots."""
+
+    left_arm_config: OpenArmLeaderConfigBase
+    right_arm_config: OpenArmLeaderConfigBase

src/lerobot/teleoperators/omx_leader/config_omx_leader.py

@@ -27,4 +27,4 @@ class OmxLeaderConfig(TeleoperatorConfig):
 
     # Sets the arm in torque mode with the gripper motor set to this value. This makes it possible to squeeze
     # the gripper and have it spring back to an open position on its own.
-    gripper_open_pos: float = 37.0
+    gripper_open_pos: float = 60.0

src/lerobot/teleoperators/omx_leader/omx_leader.py

@@ -103,7 +103,7 @@ class OmxLeader(Teleoperator):
             self.calibration[motor] = MotorCalibration(
                 id=m.id,
                 drive_mode=drive_modes[motor],
-                homing_offset=0,
+                homing_offset=0 if motor != "gripper" else 100,
                 range_min=0,
                 range_max=4095,
             )
@@ -123,12 +123,20 @@ class OmxLeader(Teleoperator):
                 # point
                 self.bus.write("Operating_Mode", motor, OperatingMode.EXTENDED_POSITION.value)
 
+            if motor == "gripper":
+                self.bus.write("Drive_Mode", motor, DriveMode.INVERTED.value)
+            else:
+                self.bus.write("Drive_Mode", motor, DriveMode.NON_INVERTED.value)
+
         # Use 'position control current based' for gripper to be limited by the limit of the current.
         # For the follower gripper, it means it can grasp an object without forcing too much even tho,
         # its goal position is a complete grasp (both gripper fingers are ordered to join and reach a touch).
         # For the leader gripper, it means we can use it as a physical trigger, since we can force with our finger
         # to make it move, and it will move back to its original target position when we release the force.
         self.bus.write("Operating_Mode", "gripper", OperatingMode.CURRENT_POSITION.value)
+        self.bus.write("Current_Limit", "gripper", 100)
+        self.bus.write("Goal_Current", "gripper", 100)
+        self.bus.write("Homing_Offset", "gripper", 100)
         # Set gripper's goal pos in current position mode so that we can use it as a trigger.
         self.bus.enable_torque("gripper")
         if self.is_calibrated:

src/lerobot/teleoperators/openarm_leader/__init__.py

@@ -0,0 +1,20 @@
+#!/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 .config_openarm_leader import OpenArmLeaderConfig, OpenArmLeaderConfigBase
+from .openarm_leader import OpenArmLeader
+
+__all__ = ["OpenArmLeader", "OpenArmLeaderConfig", "OpenArmLeaderConfigBase"]

src/lerobot/teleoperators/openarm_leader/config_openarm_leader.py

@@ -0,0 +1,75 @@
+#!/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 dataclasses import dataclass, field
+
+from ..config import TeleoperatorConfig
+
+
+@dataclass
+class OpenArmLeaderConfigBase:
+    """Base configuration for the OpenArms leader/teleoperator with Damiao motors."""
+
+    # CAN interfaces - one per arm
+    # Arm CAN interface (e.g., "can3")
+    # Linux: "can0", "can1", etc.
+    port: str
+
+    # CAN interface type: "socketcan" (Linux), "slcan" (serial), or "auto" (auto-detect)
+    can_interface: str = "socketcan"
+
+    # CAN FD settings (OpenArms uses CAN FD by default)
+    use_can_fd: bool = True
+    can_bitrate: int = 1000000  # Nominal bitrate (1 Mbps)
+    can_data_bitrate: int = 5000000  # Data bitrate for CAN FD (5 Mbps)
+
+    # Motor configuration for OpenArms (7 DOF per arm)
+    # Maps motor names to (send_can_id, recv_can_id, motor_type)
+    # Based on: https://docs.openarm.dev/software/setup/configure-test
+    # OpenArms uses 4 types of motors:
+    # - DM8009 (DM-J8009P-2EC) for shoulders (high torque)
+    # - DM4340P and DM4340 for shoulder rotation and elbow
+    # - DM4310 (DM-J4310-2EC V1.1) for wrist and gripper
+    motor_config: dict[str, tuple[int, int, str]] = field(
+        default_factory=lambda: {
+            "joint_1": (0x01, 0x11, "dm8009"),  # J1 - Shoulder pan (DM8009)
+            "joint_2": (0x02, 0x12, "dm8009"),  # J2 - Shoulder lift (DM8009)
+            "joint_3": (0x03, 0x13, "dm4340"),  # J3 - Shoulder rotation (DM4340)
+            "joint_4": (0x04, 0x14, "dm4340"),  # J4 - Elbow flex (DM4340)
+            "joint_5": (0x05, 0x15, "dm4310"),  # J5 - Wrist roll (DM4310)
+            "joint_6": (0x06, 0x16, "dm4310"),  # J6 - Wrist pitch (DM4310)
+            "joint_7": (0x07, 0x17, "dm4310"),  # J7 - Wrist rotation (DM4310)
+            "gripper": (0x08, 0x18, "dm4310"),  # J8 - Gripper (DM4310)
+        }
+    )
+
+    # Torque mode settings for manual control
+    # When enabled, motors have torque disabled for manual movement
+    manual_control: bool = True
+
+    # TODO(Steven, Pepijn): Not used ... ?
+    # MIT control parameters (used when manual_control=False for torque control)
+    # List of 8 values: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
+    position_kp: list[float] = field(
+        default_factory=lambda: [240.0, 240.0, 240.0, 240.0, 24.0, 31.0, 25.0, 16.0]
+    )
+    position_kd: list[float] = field(default_factory=lambda: [3.0, 3.0, 3.0, 3.0, 0.2, 0.2, 0.2, 0.2])
+
+
+@TeleoperatorConfig.register_subclass("openarm_leader")
+@dataclass
+class OpenArmLeaderConfig(TeleoperatorConfig, OpenArmLeaderConfigBase):
+    pass

src/lerobot/teleoperators/openarm_leader/openarm_leader.py

@@ -0,0 +1,225 @@
+#!/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.
+
+import logging
+import time
+from typing import Any
+
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode
+from lerobot.motors.damiao import DamiaoMotorsBus
+from lerobot.processor import RobotAction
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..teleoperator import Teleoperator
+from .config_openarm_leader import OpenArmLeaderConfig
+
+logger = logging.getLogger(__name__)
+
+
+class OpenArmLeader(Teleoperator):
+    """
+    OpenArm Leader/Teleoperator Arm with Damiao motors.
+
+    This teleoperator uses CAN bus communication to read positions from
+    Damiao motors that are manually moved (torque disabled).
+    """
+
+    config_class = OpenArmLeaderConfig
+    name = "openarm_leader"
+
+    def __init__(self, config: OpenArmLeaderConfig):
+        super().__init__(config)
+        self.config = config
+
+        # Arm motors
+        motors: dict[str, Motor] = {}
+        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
+            motor = Motor(
+                send_id, motor_type_str, MotorNormMode.DEGREES
+            )  # Always use degrees for Damiao motors
+            motor.recv_id = recv_id
+            motor.motor_type_str = motor_type_str
+            motors[motor_name] = motor
+
+        self.bus = DamiaoMotorsBus(
+            port=self.config.port,
+            motors=motors,
+            calibration=self.calibration,
+            can_interface=self.config.can_interface,
+            use_can_fd=self.config.use_can_fd,
+            bitrate=self.config.can_bitrate,
+            data_bitrate=self.config.can_data_bitrate if self.config.use_can_fd else None,
+        )
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        """Features produced by this teleoperator."""
+        features: dict[str, type] = {}
+        for motor in self.bus.motors:
+            features[f"{motor}.pos"] = float
+            features[f"{motor}.vel"] = float
+            features[f"{motor}.torque"] = float
+        return features
+
+    @property
+    def feedback_features(self) -> dict[str, type]:
+        """Feedback features (not implemented for OpenArms)."""
+        return {}
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if teleoperator is connected."""
+        return self.bus.is_connected
+
+    def connect(self, calibrate: bool = True) -> None:
+        """
+        Connect to the teleoperator.
+
+        For manual control, we disable torque after connecting so the
+        arm can be moved by hand.
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} already connected")
+
+        # Connect to CAN bus
+        logger.info(f"Connecting arm on {self.config.port}...")
+        self.bus.connect()
+
+        # Run calibration if needed
+        if not self.is_calibrated and calibrate:
+            logger.info(
+                "Mismatch between calibration values in the motor and the calibration file or no calibration file found"
+            )
+            self.calibrate()
+
+        self.configure()
+
+        if self.is_calibrated:
+            self.bus.set_zero_position()
+
+        logger.info(f"{self} connected.")
+
+    @property
+    def is_calibrated(self) -> bool:
+        """Check if teleoperator is calibrated."""
+        return self.bus.is_calibrated
+
+    def calibrate(self) -> None:
+        """
+        Run calibration procedure for OpenArms leader.
+
+        The calibration procedure:
+        1. Disable torque (if not already disabled)
+        2. Ask user to position arm in zero position (hanging with gripper closed)
+        3. Set this as zero position
+        4. Record range of motion for each joint
+        5. Save calibration
+        """
+        if self.calibration:
+            # Calibration file exists, ask user whether to use it or run new calibration
+            user_input = input(
+                f"Press ENTER to use provided calibration file associated with the id {self.id}, or type 'c' and press ENTER to run calibration: "
+            )
+            if user_input.strip().lower() != "c":
+                logger.info(f"Writing calibration file associated with the id {self.id} to the motors")
+                self.bus.write_calibration(self.calibration)
+                return
+
+        logger.info(f"\nRunning calibration for {self}")
+        self.bus.disable_torque()
+
+        # Step 1: Set zero position
+        input(
+            "\nCalibration: Set Zero Position)\n"
+            "Position the arm in the following configuration:\n"
+            "  - Arm hanging straight down\n"
+            "  - Gripper closed\n"
+            "Press ENTER when ready..."
+        )
+
+        # Set current position as zero for all motors
+        self.bus.set_zero_position()
+        logger.info("Arm zero position set.")
+
+        logger.info("Setting range: -90° to +90° by default for all joints")
+        # TODO(Steven, Pepijn): Check if MotorCalibration is actually needed here given that we only use Degrees
+        for motor_name, motor in self.bus.motors.items():
+            self.calibration[motor_name] = MotorCalibration(
+                id=motor.id,
+                drive_mode=0,
+                homing_offset=0,
+                range_min=-90,
+                range_max=90,
+            )
+
+        self.bus.write_calibration(self.calibration)
+        self._save_calibration()
+        print(f"Calibration saved to {self.calibration_fpath}")
+
+    def configure(self) -> None:
+        """
+        Configure motors for manual teleoperation.
+
+        For manual control, we disable torque so the arm can be moved by hand.
+        """
+
+        return self.bus.disable_torque() if self.config.manual_control else self.bus.configure_motors()
+
+    def setup_motors(self) -> None:
+        raise NotImplementedError(
+            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
+        )
+
+    def get_action(self) -> RobotAction:
+        """
+        Get current action from the leader arm.
+
+        This is the main method for teleoperators - it reads the current state
+        of the leader arm and returns it as an action that can be sent to a follower.
+
+        Reads all motor states (pos/vel/torque) in one CAN refresh cycle.
+        """
+        start = time.perf_counter()
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        action_dict: dict[str, Any] = {}
+
+        # Use sync_read_all_states to get pos/vel/torque in one go
+        states = self.bus.sync_read_all_states()
+        for motor in self.bus.motors:
+            state = states.get(motor, {})
+            action_dict[f"{motor}.pos"] = state.get("position")
+            action_dict[f"{motor}.vel"] = state.get("velocity")
+            action_dict[f"{motor}.torque"] = state.get("torque")
+
+        dt_ms = (time.perf_counter() - start) * 1e3
+        logger.debug(f"{self} read state: {dt_ms:.1f}ms")
+
+        return action_dict
+
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        raise NotImplementedError("Feedback is not yet implemented for OpenArm leader.")
+
+    def disconnect(self) -> None:
+        """Disconnect from teleoperator."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        # Disconnect CAN bus
+        # For manual control, ensure torque is disabled before disconnecting
+        self.bus.disconnect(disable_torque=self.config.manual_control)
+        logger.info(f"{self} disconnected.")

src/lerobot/teleoperators/unitree_g1/__init__.py

@@ -0,0 +1,21 @@
+#!/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 .config_unitree_g1 import ExoskeletonArmPortConfig, UnitreeG1TeleoperatorConfig
+from .exo_calib import ExoskeletonCalibration, ExoskeletonJointCalibration
+from .exo_ik import ExoskeletonIKHelper
+from .exo_serial import ExoskeletonArm
+from .unitree_g1 import UnitreeG1Teleoperator

src/lerobot/teleoperators/unitree_g1/config_unitree_g1.py

@@ -0,0 +1,37 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from ..config import TeleoperatorConfig
+
+
+@dataclass
+class ExoskeletonArmPortConfig:
+    """Serial port configuration for individual exoskeleton arm."""
+
+    port: str = ""
+    baud_rate: int = 115200
+
+
+@TeleoperatorConfig.register_subclass("unitree_g1")
+@dataclass
+class UnitreeG1TeleoperatorConfig(TeleoperatorConfig):
+    left_arm_config: ExoskeletonArmPortConfig = field(default_factory=ExoskeletonArmPortConfig)
+    right_arm_config: ExoskeletonArmPortConfig = field(default_factory=ExoskeletonArmPortConfig)
+
+    # Frozen joints (comma-separated joint names that won't be moved by IK)
+    frozen_joints: str = ""

src/lerobot/teleoperators/unitree_g1/exo_calib.py

@@ -0,0 +1,446 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This module handles calibration of hall effect sensors used in the exoskeleton.
+Each joint has a pair of ADC channels outputting sin and cos values that trace an ellipse
+as the joint rotates due to imprecision in magnet/sensor placement. We fit this ellipse to a unit circle,
+and calculate arctan2 of the unit circle to get the joint angle.
+We then store the ellipse parameters and the zero offset for each joint to be used at runtime.
+"""
+
+import json
+import logging
+import time
+from collections import deque
+from dataclasses import dataclass, field
+from pathlib import Path
+
+import numpy as np
+import serial
+
+logger = logging.getLogger(__name__)
+
+
+# exoskeleton joint names -> ADC channel pairs. TODO: add wrist pitch and wrist yaw
+JOINTS = {
+    "shoulder_pitch": (0, 1),
+    "shoulder_yaw": (2, 3),
+    "shoulder_roll": (4, 5),
+    "elbow_flex": (6, 7),
+    "wrist_roll": (14, 15),
+}
+
+
+@dataclass
+class ExoskeletonJointCalibration:
+    name: str  # joint name
+    center_fit: list[float]  # center of the ellipse
+    T: list[list[float]]  # 2x2 transformation matrix
+    zero_offset: float = 0.0  # angle at neutral pose
+
+
+@dataclass
+class ExoskeletonCalibration:
+    """Full calibration data for an exoskeleton arm."""
+
+    version: int = 2
+    side: str = ""
+    adc_max: int = 2**12 - 1
+    joints: list[ExoskeletonJointCalibration] = field(default_factory=list)
+
+    def to_dict(self) -> dict:
+        return {
+            "version": self.version,
+            "side": self.side,
+            "adc_max": self.adc_max,
+            "joints": [
+                {
+                    "name": j.name,
+                    "center_fit": j.center_fit,
+                    "T": j.T,
+                    "zero_offset": j.zero_offset,
+                }
+                for j in self.joints
+            ],
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "ExoskeletonCalibration":
+        joints = [
+            ExoskeletonJointCalibration(
+                name=j["name"],
+                center_fit=j["center_fit"],
+                T=j["T"],
+                zero_offset=j.get("zero_offset", 0.0),
+            )
+            for j in data.get("joints", [])
+        ]
+        return cls(
+            version=data.get("version", 2),
+            side=data.get("side", ""),
+            adc_max=data.get("adc_max", 2**12 - 1),
+            joints=joints,
+        )
+
+
+@dataclass(frozen=True)
+class CalibParams:
+    fit_every: float = 0.15
+    min_fit_points: int = 60
+    fit_window: int = 900
+    max_fit_points: int = 300
+    trim_low: float = 0.05
+    trim_high: float = 0.95
+    median_window: int = 5
+    history: int = 3500
+    draw_hz: float = 120.0
+    sample_count: int = 50
+
+
+def normalize_angle(angle: float) -> float:
+    while angle > np.pi:
+        angle -= 2 * np.pi
+    while angle < -np.pi:
+        angle += 2 * np.pi
+    return angle
+
+
+def joint_z_and_angle(raw16: list[int], j: ExoskeletonJointCalibration) -> tuple[np.ndarray, float]:
+    """
+    Applies calibration to each joint: raw → centered → ellipse-to-circle → angle.
+    """
+    pair = JOINTS[j.name]
+    s, c = raw16[pair[0]], raw16[pair[1]]  # get sin and cos
+    p = np.array([float(c) - (2**12 - 1) / 2, float(s) - (2**12 - 1) / 2])  # center the raw values
+    z = np.asarray(j.T) @ (
+        p - np.asarray(j.center_fit)
+    )  # center the ellipse and invert the transformation matrix to get unit circle coords
+    ang = float(np.arctan2(z[1], z[0])) - j.zero_offset  # calculate the anvgle and apply the zero offset
+    return z, normalize_angle(-ang)  # ensure range is [-pi, pi]
+
+
+def exo_raw_to_angles(raw16: list[int], calib: ExoskeletonCalibration) -> dict[str, float]:
+    """Convert raw sensor readings to joint angles using calibration."""
+    return {j.name: joint_z_and_angle(raw16, j)[1] for j in calib.joints}
+
+
+def run_exo_calibration(
+    ser: serial.Serial,
+    side: str,
+    save_path: Path,
+    params: CalibParams | None = None,
+) -> ExoskeletonCalibration:
+    """
+    Run interactive calibration for an exoskeleton arm.
+    """
+    try:
+        import cv2
+        import matplotlib.pyplot as plt
+    except ImportError as e:
+        raise ImportError(
+            "Calibration requires matplotlib and opencv-python. "
+            "Install with: pip install matplotlib opencv-python"
+        ) from e
+
+    from .exo_serial import read_raw_from_serial
+
+    params = params or CalibParams()
+    joint_list = list(JOINTS.items())  # Convert dict to list for indexing
+    logger.info(f"Starting calibration for {side} exoskeleton arm")
+
+    def running_median(win: deque) -> float:
+        return float(np.median(np.fromiter(win, dtype=float)))
+
+    def read_joint_point(raw16: list[int], pair: tuple[int, int]):
+        s, c = raw16[pair[0]], raw16[pair[1]]
+        return float(c) - (2**12 - 1) / 2, float(s) - (2**12 - 1) / 2, float(s), float(c)
+
+    def select_fit_subset(xs, ys):
+        """Select and filter points for ellipse fitting. Trims outliers by radius and downsamples."""
+        n = min(params.fit_window, len(xs))
+        if n <= 0:
+            return None, None
+        x = np.asarray(list(xs)[-n:], dtype=float)  # most recent n samples
+        y = np.asarray(list(ys)[-n:], dtype=float)
+        r = np.sqrt(x * x + y * y)  # radius from origin
+        if len(r) >= 20:
+            lo, hi = np.quantile(r, params.trim_low), np.quantile(r, params.trim_high)  # outlier bounds
+            keep = (r >= lo) & (r <= hi)
+            x, y = x[keep], y[keep]  # remove outliers
+        if len(x) > params.max_fit_points:
+            idx = np.linspace(0, len(x) - 1, params.max_fit_points).astype(int)  # downsample evenly
+            x, y = x[idx], y[idx]
+        return x, y
+
+    def fit_ellipse_opencv(x, y):
+        """Fit ellipse to (x,y) points using OpenCV. Returns center, axes, rotation matrix, and outline."""
+        x, y = np.asarray(x, dtype=float), np.asarray(y, dtype=float)
+        if len(x) < 5:
+            return None
+        pts = np.stack([x, y], axis=1).astype(np.float32).reshape(-1, 1, 2)
+        try:
+            (xc, yc), (w, h), angle_deg = cv2.fitEllipse(pts)  # returns center, axes, rotation in degrees
+        except cv2.error:
+            return None
+        a, b = float(w) * 0.5, float(h) * 0.5  # get ellipse major and minor semi-axes
+        phi = np.deg2rad(float(angle_deg))  # to rad
+        if b > a:  # ensure major axis is a
+            a, b = b, a
+            phi += np.pi / 2.0
+        if not np.isfinite(a) or not np.isfinite(b) or a <= 1e-6 or b <= 1e-6:
+            return None
+        cp, sp = float(np.cos(phi)), float(np.sin(phi))  #
+        rot = np.array([[cp, -sp], [sp, cp]], dtype=float)  # 2x2 rotation matrix
+        center = np.array([float(xc), float(yc)], dtype=float)  # offset vector
+        tt = np.linspace(0, 2 * np.pi, 360)
+        outline = (rot @ np.stack([a * np.cos(tt), b * np.sin(tt)])).T + center  # for viz
+        return {"center": center, "a": a, "b": b, "R": rot, "ex": outline[:, 0], "ey": outline[:, 1]}
+
+    # Setup matplotlib
+    plt.ion()
+    fig, (ax0, ax1) = plt.subplots(1, 2, figsize=(12, 6))
+    ax0.set_xlabel("cos - center")
+    ax0.set_ylabel("sin - center")
+    ax0.grid(True, alpha=0.25)
+    ax0.set_aspect("equal", adjustable="box")
+    ax1.set_title("Unit circle + angle")
+    ax1.set_xlabel("x")
+    ax1.set_ylabel("y")
+    ax1.grid(True, alpha=0.25)
+    ax1.set_aspect("equal", adjustable="box")
+    tt = np.linspace(0, 2 * np.pi, 360)
+    ax1.plot(np.cos(tt), np.sin(tt), "k-", linewidth=1)
+    ax0.set_xlim(-2200, 2200)
+    ax0.set_ylim(-2200, 2200)
+    ax1.set_xlim(-1.4, 1.4)
+    ax1.set_ylim(-1.4, 1.4)
+
+    sc0 = ax0.scatter([], [], s=6, animated=True)
+    (ell_line,) = ax0.plot([], [], "r-", linewidth=2, animated=True)
+    sc1 = ax1.scatter([], [], s=6, animated=True)
+    (radius_line,) = ax1.plot([], [], "g-", linewidth=2, animated=True)
+    angle_text = ax1.text(
+        0.02, 0.98, "", transform=ax1.transAxes, va="top", ha="left", fontsize=12, animated=True
+    )
+
+    fig.canvas.draw()
+    bg0 = fig.canvas.copy_from_bbox(ax0.bbox)
+    bg1 = fig.canvas.copy_from_bbox(ax1.bbox)
+
+    # State
+    joints_out = []
+    joint_idx = 0
+    phase = "ellipse"
+    advance_requested = False
+    zero_samples = []
+
+    def on_key(event):
+        nonlocal advance_requested
+        if event.key in ("n", "N", "enter", " "):
+            advance_requested = True
+
+    fig.canvas.mpl_connect("key_press_event", on_key)
+
+    def reset_state():
+        return {
+            "xs": deque(maxlen=params.history),
+            "ys": deque(maxlen=params.history),
+            "xu": deque(maxlen=params.history),
+            "yu": deque(maxlen=params.history),
+            "win_s": deque(maxlen=params.median_window),
+            "win_c": deque(maxlen=params.median_window),
+            "ellipse_cache": None,
+            "T": None,
+            "center_fit": None,
+            "have_transform": False,
+            "latest_z": None,
+            "last_fit": 0.0,
+        }
+
+    state = reset_state()
+    last_draw = 0.0
+    name, pair = joint_list[joint_idx]
+    fig.canvas.manager.set_window_title(f"[{joint_idx + 1}/{len(joint_list)}] {name} - ELLIPSE")
+    ax0.set_title(f"{name} raw (filtered)")
+    logger.info(f"[{joint_idx + 1}/{len(joint_list)}] Calibrating {name}")
+    logger.info("Step 1: Move joint around to map ellipse, then press 'n'")
+
+    try:
+        while plt.fignum_exists(fig.number):
+            name, pair = joint_list[joint_idx]
+
+            # Handles calibration GUI state: ellipse → zero_pose → next joint -> ellipse -> ...
+            if phase == "ellipse" and advance_requested and state["have_transform"]:
+                joints_out.append(
+                    {
+                        "name": name,
+                        "center_fit": state["center_fit"].tolist(),
+                        "T": state["T"].tolist(),
+                    }
+                )
+                logger.info(f"  -> Ellipse saved for {name}")
+                phase, zero_samples, advance_requested = "zero_pose", [], False
+                fig.canvas.manager.set_window_title(f"[{joint_idx + 1}/{len(joint_list)}] {name} - ZERO POSE")
+                ax0.set_title(f"{name} - hold zero pose")
+                fig.canvas.draw()
+                bg0, bg1 = fig.canvas.copy_from_bbox(ax0.bbox), fig.canvas.copy_from_bbox(ax1.bbox)
+                logger.info(f"Step 2: Hold {name} in zero position, then press 'n'")
+
+            elif phase == "ellipse" and advance_requested and not state["have_transform"]:
+                logger.info("  (Need valid fit first - keep moving the joint)")
+                advance_requested = False
+
+            elif phase == "zero_pose" and advance_requested:
+                if len(zero_samples) >= params.sample_count:
+                    zero_offset = float(np.mean(zero_samples[-params.sample_count :]))
+                    joints_out[-1]["zero_offset"] = zero_offset
+                    logger.info(f"  -> {name} zero: {zero_offset:+.3f} rad ({np.degrees(zero_offset):+.1f}°)")
+                    joint_idx += 1
+                    advance_requested = False
+
+                    if joint_idx >= len(joint_list):
+                        # All joints done
+                        calib = ExoskeletonCalibration(
+                            version=2,
+                            side=side,
+                            adc_max=2**12 - 1,
+                            joints=[
+                                ExoskeletonJointCalibration(
+                                    name=j["name"],
+                                    center_fit=j["center_fit"],
+                                    T=j["T"],
+                                    zero_offset=j.get("zero_offset", 0.0),
+                                )
+                                for j in joints_out
+                            ],
+                        )
+                        save_path.parent.mkdir(parents=True, exist_ok=True)
+                        with open(save_path, "w") as f:
+                            json.dump(calib.to_dict(), f, indent=2)
+                        logger.info(f"Saved calibration to {save_path}")
+                        logger.info("Calibration complete!")
+                        plt.close(fig)
+                        return calib
+
+                    # Next joint
+                    phase, state = "ellipse", reset_state()
+                    name, pair = joint_list[joint_idx]
+                    fig.canvas.manager.set_window_title(
+                        f"[{joint_idx + 1}/{len(joint_list)}] {name} - ELLIPSE"
+                    )
+                    ax0.set_title(f"{name} raw (filtered)")
+                    fig.canvas.draw()
+                    bg0, bg1 = fig.canvas.copy_from_bbox(ax0.bbox), fig.canvas.copy_from_bbox(ax1.bbox)
+                    logger.info(f"[{joint_idx + 1}/{len(joint_list)}] Calibrating {name}")
+                    logger.info("Step 1: Move joint around to map ellipse, then press 'n'")
+                else:
+                    logger.info(
+                        f"  (Collecting samples: {len(zero_samples)}/{params.sample_count} - hold still)"
+                    )
+                    advance_requested = False
+
+            # Read sensor
+            raw16 = read_raw_from_serial(ser)
+            if raw16 is not None:
+                x_raw, y_raw, s_raw, c_raw = read_joint_point(raw16, pair)
+
+                if phase == "ellipse":
+                    if state["have_transform"]:
+                        z = state["T"] @ (np.array([x_raw, y_raw]) - state["center_fit"])
+                        state["xu"].append(float(z[0]))
+                        state["yu"].append(float(z[1]))
+                        state["latest_z"] = (float(z[0]), float(z[1]))
+                    state["win_s"].append(s_raw)
+                    state["win_c"].append(c_raw)
+                    if len(state["win_s"]) >= max(3, params.median_window):
+                        state["ys"].append(running_median(state["win_s"]) - (2**12 - 1) / 2)
+                        state["xs"].append(running_median(state["win_c"]) - (2**12 - 1) / 2)
+                else:
+                    jdata = joints_out[-1]
+                    z = np.array(jdata["T"]) @ (np.array([x_raw, y_raw]) - np.array(jdata["center_fit"]))
+                    zero_samples.append(float(np.arctan2(z[1], z[0])))
+                    state["latest_z"] = (float(z[0]), float(z[1]))
+
+            # Ellipse fitting
+            t = time.time()
+            if (
+                phase == "ellipse"
+                and (t - state["last_fit"]) >= params.fit_every
+                and len(state["xs"]) >= params.min_fit_points
+            ):
+                xfit, yfit = select_fit_subset(state["xs"], state["ys"])
+                if xfit is not None and len(xfit) >= params.min_fit_points:
+                    fit = fit_ellipse_opencv(xfit, yfit)
+                    if fit is not None:
+                        state["center_fit"] = fit["center"]
+                        state["T"] = np.diag([1.0 / fit["a"], 1.0 / fit["b"]]) @ fit["R"].T
+                        state["ellipse_cache"] = (fit["ex"], fit["ey"])
+                        state["have_transform"] = True
+                state["last_fit"] = t
+
+            # Drawing
+            if (t - last_draw) >= 1.0 / params.draw_hz:
+                fig.canvas.restore_region(bg0)
+                fig.canvas.restore_region(bg1)
+
+                if phase == "ellipse":
+                    sc0.set_offsets(np.c_[state["xs"], state["ys"]] if state["xs"] else np.empty((0, 2)))
+                    ax0.draw_artist(sc0)
+                    ell_line.set_data(*state["ellipse_cache"] if state["ellipse_cache"] else ([], []))
+                    ax0.draw_artist(ell_line)
+                    sc1.set_offsets(np.c_[state["xu"], state["yu"]] if state["xu"] else np.empty((0, 2)))
+                    ax1.draw_artist(sc1)
+                    if state["latest_z"]:
+                        zx, zy = state["latest_z"]
+                        radius_line.set_data([0.0, zx], [0.0, zy])
+                        ang = float(np.arctan2(zy, zx))
+                        angle_text.set_text(
+                            f"angle: {ang:+.3f} rad  ({np.degrees(ang):+.1f}°)\nmove {name}, press 'n' to advance"
+                        )
+                    else:
+                        radius_line.set_data([], [])
+                        angle_text.set_text("(waiting for fit)")
+                else:
+                    sc0.set_offsets(np.empty((0, 2)))
+                    ax0.draw_artist(sc0)
+                    ell_line.set_data([], [])
+                    ax0.draw_artist(ell_line)
+                    if state["latest_z"]:
+                        zx, zy = state["latest_z"]
+                        sc1.set_offsets([[zx, zy]])
+                        radius_line.set_data([0.0, zx], [0.0, zy])
+                        ang = float(np.arctan2(zy, zx))
+                        angle_text.set_text(
+                            f"Zero pose for {name}\nangle: {ang:+.3f} rad\nsamples: {len(zero_samples)}/{params.sample_count}\nhold still, press 'n'"
+                        )
+                    else:
+                        sc1.set_offsets(np.empty((0, 2)))
+                        radius_line.set_data([], [])
+                        angle_text.set_text("(waiting for data)")
+                    ax1.draw_artist(sc1)
+
+                ax1.draw_artist(radius_line)
+                ax1.draw_artist(angle_text)
+                fig.canvas.blit(ax0.bbox)
+                fig.canvas.blit(ax1.bbox)
+                fig.canvas.flush_events()
+                last_draw = t
+
+            plt.pause(0.001)
+
+    finally:
+        plt.close(fig)

src/lerobot/teleoperators/unitree_g1/exo_ik.py

@@ -0,0 +1,353 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+IK helper for exoskeleton-to-G1 teleoperation. We map Exoskeleton joint angles to end-effector pose in world frame,
+visualizing the result in meshcat after calibration.
+"""
+
+import logging
+import os
+from dataclasses import dataclass
+
+import numpy as np
+
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointArmIndex
+from lerobot.robots.unitree_g1.robot_kinematic_processor import G1_29_ArmIK
+
+from .exo_calib import JOINTS
+
+logger = logging.getLogger(__name__)
+
+
+def _frame_id(model, name: str) -> int | None:
+    try:
+        fid = model.getFrameId(name)
+        return fid if 0 <= fid < model.nframes else None
+    except Exception:
+        return None
+
+
+@dataclass
+class ArmCfg:
+    side: str  # "left" | "right"
+    urdf: str  # exo_left.urdf / exo_right.urdf
+    root: str  # "exo_left" / "exo_right"
+    g1_ee: str  # "l_ee" / "r_ee"
+    offset: np.ndarray  # world offset for viz + target
+    marker_prefix: str  # "left" / "right"
+
+
+class Markers:
+    """Creates meshcat visualization primitives, showing end-effector frames of exoskeleton and G1"""
+
+    def __init__(self, viewer):
+        self.v = viewer
+
+    def sphere(self, path: str, r: float, rgba: tuple[float, float, float, float]):
+        import meshcat.geometry as mg
+
+        c = (int(rgba[0] * 255) << 16) | (int(rgba[1] * 255) << 8) | int(rgba[2] * 255)
+        self.v[path].set_object(
+            mg.Sphere(r),
+            mg.MeshPhongMaterial(color=c, opacity=rgba[3], transparent=rgba[3] < 1.0),
+        )
+
+    def axes(self, path: str, axis_len: float = 0.1, axis_w: int = 6):
+        import meshcat.geometry as mg
+
+        pts = np.array(
+            [[0, 0, 0], [axis_len, 0, 0], [0, 0, 0], [0, axis_len, 0], [0, 0, 0], [0, 0, axis_len]],
+            dtype=np.float32,
+        ).T
+        cols = np.array(
+            [[1, 0, 0], [1, 0, 0], [0, 1, 0], [0, 1, 0], [0, 0, 1], [0, 0, 1]],
+            dtype=np.float32,
+        ).T
+        self.v[path].set_object(
+            mg.LineSegments(
+                mg.PointsGeometry(position=pts, color=cols),
+                mg.LineBasicMaterial(linewidth=axis_w, vertexColors=True),
+            )
+        )
+
+    def tf(self, path: str, mat: np.ndarray):
+        self.v[path].set_transform(mat)
+
+
+class ExoskeletonIKHelper:
+    """
+    - Loads G1 robot and exoskeleton URDF models via Pinocchio
+    - Computes forward kinematics on exoskeleton to get end-effector poses
+    - Solves inverse kinematics on G1 to match those poses
+    - Provides meshcat visualization showing both robots and targets
+
+    Args:
+        frozen_joints: List of G1 joint names to exclude from IK (kept at neutral).
+    """
+
+    def __init__(self, frozen_joints: list[str] | None = None):
+        try:
+            import pinocchio as pin
+        except ImportError as e:
+            raise ImportError("ik mode needs pinocchio: pip install pin") from e
+
+        self.pin = pin
+        self.frozen_joints = frozen_joints or []
+
+        self.g1_ik = G1_29_ArmIK()
+        self.robot_g1 = self.g1_ik.reduced_robot
+        self.robot_g1.data = self.robot_g1.model.createData()
+        self.q_g1 = pin.neutral(self.robot_g1.model)
+
+        assets_dir = os.path.join(self.g1_ik.repo_path, "assets")
+
+        self.frozen_idx = self._frozen_joint_indices()
+
+        self.arms = [
+            ArmCfg(
+                side="left",
+                urdf=os.path.join(assets_dir, "exo_left.urdf"),
+                root="exo_left",
+                g1_ee="L_ee",
+                offset=np.array([0.6, 0.3, 0.0]),
+                marker_prefix="left",
+            ),
+            ArmCfg(
+                side="right",
+                urdf=os.path.join(assets_dir, "exo_right.urdf"),
+                root="exo_right",
+                g1_ee="R_ee",
+                offset=np.array([0.6, -0.3, 0.0]),
+                marker_prefix="right",
+            ),
+        ]
+
+        self.exo = {}  # side -> pin.RobotWrapper
+        self.q_exo = {}  # side -> q
+        self.ee_id_exo = {}  # side -> frame id
+        self.qmap = {}  # side -> {joint_name: q_idx}
+        self.ee_id_g1 = {}  # side -> frame id
+
+        self._load_exo_models(assets_dir)
+        for a in self.arms:
+            self.ee_id_g1[a.side] = _frame_id(self.robot_g1.model, a.g1_ee)
+
+        self.viewer = None
+        self.markers: Markers | None = None
+        self.viz_g1 = None
+        self.viz_exo = {}  # side -> viz
+
+    def _frozen_joint_indices(self) -> dict[str, int]:
+        out = {}
+        m = self.robot_g1.model
+        for name in self.frozen_joints:
+            if name in m.names:
+                jid = m.getJointId(name)
+                out[name] = m.idx_qs[jid]
+                logger.info(f"freezing joint: {name} (q_idx={out[name]})")
+        return out
+
+    def _find_exo_ee(self, model, ee_name: str = "ee") -> int:
+        ee = _frame_id(model, ee_name)
+        if ee is not None:
+            return ee
+        for fid in reversed(range(model.nframes)):
+            if model.frames[fid].type == self.pin.FrameType.BODY:
+                return fid
+        return 0
+
+    def _build_joint_map(self, robot) -> dict[str, int]:
+        m = robot.model
+        return {n: m.idx_qs[m.getJointId(n)] for n in JOINTS if n in m.names}
+
+    def _load_exo_models(self, assets_dir: str):
+        pin = self.pin
+        for a in self.arms:
+            if not os.path.exists(a.urdf):
+                logger.warning(f"{a.side} exo urdf not found: {a.urdf}")
+                continue
+            r = pin.RobotWrapper.BuildFromURDF(a.urdf, assets_dir)
+            self.exo[a.side] = r
+            self.q_exo[a.side] = pin.neutral(r.model)
+            self.ee_id_exo[a.side] = self._find_exo_ee(r.model)
+            self.qmap[a.side] = self._build_joint_map(r)
+            logger.info(f"loaded {a.side} exo urdf: {a.urdf}")
+
+    def init_visualization(self):
+        """
+        Creates a browser-based visualization of exoskeleton and G1 robot,
+        highlighting end-effector frames and target positions.
+        """
+        try:
+            from pinocchio.visualize import MeshcatVisualizer
+        except ImportError as e:
+            logger.warning(f"meshcat viz unavailable: {e}")
+            return
+
+        # g1
+        self.viz_g1 = MeshcatVisualizer(
+            self.robot_g1.model, self.robot_g1.collision_model, self.robot_g1.visual_model
+        )
+        self.viz_g1.initViewer(open=True)
+        self.viz_g1.loadViewerModel("g1")
+        self.viz_g1.display(self.q_g1)
+
+        self.viewer = self.viz_g1.viewer
+        self.markers = Markers(self.viewer)
+
+        # exos
+        for a in self.arms:
+            if a.side not in self.exo:
+                continue
+            r = self.exo[a.side]
+            v = MeshcatVisualizer(r.model, r.collision_model, r.visual_model)
+            v.initViewer(open=False)
+            v.viewer = self.viewer
+            v.loadViewerModel(a.root)
+            offset_tf = np.eye(4)
+            offset_tf[:3, 3] = a.offset
+            self.viewer[a.root].set_transform(offset_tf)
+            v.display(self.q_exo[a.side])
+            self.viz_exo[a.side] = v
+
+        # markers
+        for a in self.arms:
+            p = a.marker_prefix
+            self.markers.sphere(f"markers/{p}_exo_ee", 0.012, (0.2, 1.0, 0.2, 0.9))
+            self.markers.sphere(f"markers/{p}_g1_ee", 0.015, (1.0, 0.2, 0.2, 0.9))
+            self.markers.sphere(f"markers/{p}_ik_target", 0.015, (0.1, 0.3, 1.0, 0.9))
+            self.markers.axes(f"markers/{p}_exo_axes", 0.06)
+            self.markers.axes(f"markers/{p}_g1_axes", 0.08)
+
+        logger.info(f"meshcat viz initialized: {self.viewer.url()}")
+        print(f"\nmeshcat url: {self.viewer.url()}\n")
+
+    def _fk_target_world(self, side: str, angles: dict[str, float]) -> np.ndarray | None:
+        """returns wrist frame target to be used for G1 IK in 4x4 homogeneous transform. Takes offset into account."""
+        if side not in self.exo or not angles:
+            return None
+
+        pin = self.pin
+        q = self.q_exo[side]
+        qmap = self.qmap[side]
+
+        for name, ang in angles.items():
+            idx = qmap.get(name)
+            if idx is not None:
+                q[idx] = float(ang)
+
+        r = self.exo[side]
+        pin.forwardKinematics(r.model, r.data, q)
+        pin.updateFramePlacements(r.model, r.data)
+
+        ee = r.data.oMf[self.ee_id_exo[side]]
+        target = np.eye(4)
+        target[:3, :3] = ee.rotation
+        # offset gets applied in world space
+        cfg = next(a for a in self.arms if a.side == side)
+        target[:3, 3] = cfg.offset + ee.translation
+        return target
+
+    def update_visualization(self):
+        if self.viewer is None or self.markers is None:
+            return
+
+        pin = self.pin
+
+        # g1
+        if self.viz_g1 is not None:
+            self.viz_g1.display(self.q_g1)
+            pin.forwardKinematics(self.robot_g1.model, self.robot_g1.data, self.q_g1)
+            pin.updateFramePlacements(self.robot_g1.model, self.robot_g1.data)
+
+            for a in self.arms:
+                fid = self.ee_id_g1.get(a.side)
+                if fid is None:
+                    continue
+                ee_tf = self.robot_g1.data.oMf[fid].homogeneous
+                p = a.marker_prefix
+                self.markers.tf(f"markers/{p}_g1_ee", ee_tf)
+                self.markers.tf(f"markers/{p}_g1_axes", ee_tf)
+
+        # exos
+        for a in self.arms:
+            side = a.side
+            v = self.viz_exo.get(side)
+            if v is None:
+                continue
+
+            v.display(self.q_exo[side])
+            r = self.exo[side]
+            pin.forwardKinematics(r.model, r.data, self.q_exo[side])
+            pin.updateFramePlacements(r.model, r.data)
+
+            ee = r.data.oMf[self.ee_id_exo[side]]
+            world_tf = (pin.SE3(np.eye(3), a.offset) * ee).homogeneous
+            p = a.marker_prefix
+            self.markers.tf(f"markers/{p}_exo_ee", world_tf)
+            self.markers.tf(f"markers/{p}_exo_axes", world_tf)
+
+            target_tf = np.eye(4)
+            target_tf[:3, :3] = ee.rotation
+            target_tf[:3, 3] = a.offset + ee.translation
+            self.markers.tf(f"markers/{p}_ik_target", target_tf)
+
+    def compute_g1_joints_from_exo(
+        self,
+        left_angles: dict[str, float],
+        right_angles: dict[str, float],
+    ) -> dict[str, float]:
+        """
+        Performs FK on exoskeleton to get end-effector poses in world frame,
+        after which it solves IK on G1 to return joint angles matching those poses in G1 motor order.
+        """
+        pin = self.pin
+
+        targets = {
+            "left": self._fk_target_world("left", left_angles),
+            "right": self._fk_target_world("right", right_angles),
+        }
+
+        # fallback to current g1 ee pose if missing target
+        pin.forwardKinematics(self.robot_g1.model, self.robot_g1.data, self.q_g1)
+        pin.updateFramePlacements(self.robot_g1.model, self.robot_g1.data)
+
+        for a in self.arms:
+            if targets[a.side] is not None:
+                continue
+            fid = self.ee_id_g1.get(a.side)
+            if fid is not None:
+                targets[a.side] = self.robot_g1.data.oMf[fid].homogeneous
+
+        if targets["left"] is None or targets["right"] is None:
+            logger.warning("missing ik targets, returning current pose")
+            return {}
+
+        frozen_vals = {n: self.q_g1[i] for n, i in self.frozen_idx.items()}
+
+        self.q_g1, _ = self.g1_ik.solve_ik(
+            targets["left"], targets["right"], current_lr_arm_motor_q=self.q_g1
+        )
+
+        for n, i in self.frozen_idx.items():
+            self.q_g1[i] = frozen_vals[n]
+
+        return {
+            f"{j.name}.q": float(self.q_g1[i])
+            for i, j in enumerate(G1_29_JointArmIndex)
+            if i < len(self.q_g1)
+        }

src/lerobot/teleoperators/unitree_g1/exo_serial.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+
+import serial
+
+from .exo_calib import ExoskeletonCalibration, exo_raw_to_angles, run_exo_calibration
+
+logger = logging.getLogger(__name__)
+
+
+def parse_raw16(line: bytes) -> list[int] | None:
+    try:
+        parts = line.decode("utf-8", errors="ignore").split()
+        if len(parts) < 16:
+            return None
+        return [int(x) for x in parts[:16]]
+    except Exception:
+        return None
+
+
+def read_raw_from_serial(ser) -> list[int] | None:
+    """Read latest sample from serial; if buffer is backed up, keep only the newest."""
+    last = None
+    while ser.in_waiting > 0:
+        b = ser.readline()
+        if not b:
+            break
+        raw16 = parse_raw16(b)
+        if raw16 is not None:
+            last = raw16
+    if last is None:
+        b = ser.readline()
+        if b:
+            last = parse_raw16(b)
+    return last
+
+
+@dataclass
+class ExoskeletonArm:
+    port: str
+    calibration_fpath: Path
+    side: str
+    baud_rate: int = 115200
+
+    _ser: serial.Serial | None = None
+    calibration: ExoskeletonCalibration | None = None
+
+    def __post_init__(self):
+        if self.calibration_fpath.is_file():
+            self._load_calibration()
+
+    @property
+    def is_connected(self) -> bool:
+        return self._ser is not None and getattr(self._ser, "is_open", False)
+
+    @property
+    def is_calibrated(self) -> bool:
+        return self.calibration is not None
+
+    def connect(self, calibrate: bool = True) -> None:
+        if self.is_connected:
+            return
+        try:
+            self._ser = serial.Serial(self.port, self.baud_rate, timeout=0.02)
+            self._ser.reset_input_buffer()
+            logger.info(f"connected: {self.port}")
+        except serial.SerialException as e:
+            raise ConnectionError(f"failed to connect to {self.port}: {e}") from e
+
+        if calibrate and not self.is_calibrated:
+            self.calibrate()
+
+    def disconnect(self) -> None:
+        if self._ser:
+            try:
+                self._ser.close()
+            finally:
+                self._ser = None
+
+    def _load_calibration(self) -> None:
+        try:
+            data = json.loads(self.calibration_fpath.read_text())
+            self.calibration = ExoskeletonCalibration.from_dict(data)
+            logger.info(f"loaded calibration: {self.calibration_fpath}")
+        except Exception as e:
+            logger.warning(f"failed to load calibration: {e}")
+
+    def read_raw(self) -> list[int] | None:
+        if not self._ser:
+            return None
+        return read_raw_from_serial(self._ser)
+
+    def get_angles(self) -> dict[str, float]:
+        if not self.calibration:
+            raise RuntimeError("exoskeleton not calibrated")
+        raw = self.read_raw()
+        return {} if raw is None else exo_raw_to_angles(raw, self.calibration)
+
+    def calibrate(self) -> None:
+        ser = self._ser
+        self.calibration = run_exo_calibration(ser, self.side, self.calibration_fpath)

src/lerobot/teleoperators/unitree_g1/unitree_g1.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import time
+from functools import cached_property
+
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
+from lerobot.utils.constants import HF_LEROBOT_CALIBRATION, TELEOPERATORS
+
+from ..teleoperator import Teleoperator
+from .config_unitree_g1 import UnitreeG1TeleoperatorConfig
+from .exo_ik import ExoskeletonIKHelper
+from .exo_serial import ExoskeletonArm
+
+logger = logging.getLogger(__name__)
+
+
+class UnitreeG1Teleoperator(Teleoperator):
+    """
+    Bimanual exoskeleton arms teleoperator for Unitree G1 arms.
+
+    Uses inverse kinematics: exoskeleton FK computes end-effector pose,
+    G1 IK solves for joint angles.
+    """
+
+    config_class = UnitreeG1TeleoperatorConfig
+    name = "unitree_g1"
+
+    def __init__(self, config: UnitreeG1TeleoperatorConfig):
+        super().__init__(config)
+        self.config = config
+
+        # Setup calibration directory
+        self.calibration_dir = (
+            config.calibration_dir
+            if config.calibration_dir
+            else HF_LEROBOT_CALIBRATION / TELEOPERATORS / self.name
+        )
+        self.calibration_dir.mkdir(parents=True, exist_ok=True)
+
+        left_id = f"{config.id}_left" if config.id else "left"
+        right_id = f"{config.id}_right" if config.id else "right"
+
+        # Create exoskeleton arm instances
+        self.left_arm = ExoskeletonArm(
+            port=config.left_arm_config.port,
+            baud_rate=config.left_arm_config.baud_rate,
+            calibration_fpath=self.calibration_dir / f"{left_id}.json",
+            side="left",
+        )
+        self.right_arm = ExoskeletonArm(
+            port=config.right_arm_config.port,
+            baud_rate=config.right_arm_config.baud_rate,
+            calibration_fpath=self.calibration_dir / f"{right_id}.json",
+            side="right",
+        )
+
+        self.ik_helper: ExoskeletonIKHelper | None = None
+
+    @cached_property
+    def action_features(self) -> dict[str, type]:
+        return {f"{name}.q": float for name in self._g1_joint_names}
+
+    @cached_property
+    def feedback_features(self) -> dict[str, type]:
+        return {}
+
+    @property
+    def is_connected(self) -> bool:
+        return self.left_arm.is_connected and self.right_arm.is_connected
+
+    @property
+    def is_calibrated(self) -> bool:
+        return self.left_arm.is_calibrated and self.right_arm.is_calibrated
+
+    def connect(self, calibrate: bool = True) -> None:
+        self.left_arm.connect(calibrate)
+        self.right_arm.connect(calibrate)
+
+        frozen_joints = [j.strip() for j in self.config.frozen_joints.split(",") if j.strip()]
+        self.ik_helper = ExoskeletonIKHelper(frozen_joints=frozen_joints)
+        logger.info("IK helper initialized")
+
+    def calibrate(self) -> None:
+        if not self.left_arm.is_calibrated:
+            logger.info("Starting calibration for left arm...")
+            self.left_arm.calibrate()
+        else:
+            logger.info("Left arm already calibrated. Skipping.")
+
+        if not self.right_arm.is_calibrated:
+            logger.info("Starting calibration for right arm...")
+            self.right_arm.calibrate()
+        else:
+            logger.info("Right arm already calibrated. Skipping.")
+
+        logger.info("Starting visualization to verify calibration...")
+        self.run_visualization_loop()
+
+    def configure(self) -> None:
+        pass
+
+    def get_action(self) -> dict[str, float]:
+        left_angles = self.left_arm.get_angles()
+        right_angles = self.right_arm.get_angles()
+        return self.ik_helper.compute_g1_joints_from_exo(left_angles, right_angles)
+
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        raise NotImplementedError("Exoskeleton arms do not support feedback")
+
+    def disconnect(self) -> None:
+        self.left_arm.disconnect()
+        self.right_arm.disconnect()
+
+    def run_visualization_loop(self):
+        """Run interactive Meshcat visualization loop to verify tracking."""
+        if self.ik_helper is None:
+            frozen_joints = [j.strip() for j in self.config.frozen_joints.split(",") if j.strip()]
+            self.ik_helper = ExoskeletonIKHelper(frozen_joints=frozen_joints)
+
+        self.ik_helper.init_visualization()
+
+        print("\n" + "=" * 60)
+        print("Visualization running! Move the exoskeletons to test tracking.")
+        print("Press Ctrl+C to exit.")
+        print("=" * 60 + "\n")
+
+        try:
+            while True:
+                left_angles = self.left_arm.get_angles()
+                right_angles = self.right_arm.get_angles()
+
+                self.ik_helper.compute_g1_joints_from_exo(left_angles, right_angles)
+                self.ik_helper.update_visualization()
+
+                time.sleep(0.01)
+
+        except KeyboardInterrupt:
+            print("\n\nVisualization stopped.")
+
+    @cached_property
+    def _g1_joint_names(self) -> list[str]:
+        return [joint.name for joint in G1_29_JointIndex]