debug rtc

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

* add delete button

* fixes

* remove viz

* unused import

docs/source/openarms.mdx

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

docs/source/rac.mdx

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

examples/openarms/debug_can_communication.py

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

examples/openarms/evaluate.py

@@ -0,0 +1,360 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+OpenArms Policy Evaluation
+
+Evaluates a trained policy on the OpenArms robot by running inference and recording
+the evaluation episodes to a dataset. Supports optional leader arm for manual resets.
+
+Example usage:
+    python examples/openarms/evaluate.py
+"""
+
+import time
+from pathlib import Path
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.processor import make_default_processors
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+
+HF_MODEL_ID = "lerobot-data-collection/three-folds-pi0"  # TODO: Replace with your trained model
+HF_EVAL_DATASET_ID = "lerobot-data-collection/three-folds-pi0_eval7"  # TODO: Replace with your eval dataset name
+TASK_DESCRIPTION = "three-folds-dataset"  # TODO: Replace with your task, this should match!!
+
+NUM_EPISODES = 1
+FPS = 30
+EPISODE_TIME_SEC = 300
+RESET_TIME_SEC = 60
+
+# Robot CAN interfaces
+FOLLOWER_LEFT_PORT = "can0"
+FOLLOWER_RIGHT_PORT = "can1"
+
+# If enabled, you can manually reset the environment between evaluation episodes
+USE_LEADER_FOR_RESETS = True  # Set to False if you don't want to use leader
+LEADER_LEFT_PORT = "can2"
+LEADER_RIGHT_PORT = "can3"
+
+# Camera configuration
+CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video5", width=640, height=480, fps=FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video1", width=640, height=480, fps=FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video3", width=640, height=480, fps=FPS),
+}
+
+def main():
+    """Main evaluation function."""
+    print("OpenArms Policy Evaluation")
+    print(f"\nModel: {HF_MODEL_ID}")
+    print(f"Evaluation Dataset: {HF_EVAL_DATASET_ID}")
+    print(f"Task: {TASK_DESCRIPTION}")
+    print(f"Episodes: {NUM_EPISODES}")
+    print(f"Episode Duration: {EPISODE_TIME_SEC}s")
+    print(f"Reset Duration: {RESET_TIME_SEC}s")
+    print(f"Use Leader for Resets: {USE_LEADER_FOR_RESETS}")
+    
+    follower_config = OpenArmsFollowerConfig(
+        port_left=FOLLOWER_LEFT_PORT,
+        port_right=FOLLOWER_RIGHT_PORT,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=CAMERA_CONFIG,
+    )
+    
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+    
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    
+    leader = None
+    if USE_LEADER_FOR_RESETS:
+        leader_config = OpenArmsLeaderConfig(
+            port_left=LEADER_LEFT_PORT,
+            port_right=LEADER_RIGHT_PORT,
+            can_interface="socketcan",
+            id="openarms_leader",
+            manual_control=False,  # Enable torque control for gravity compensation
+        )
+        
+        leader = OpenArmsLeader(leader_config)
+        leader.connect(calibrate=False)
+        
+        if not leader.is_connected:
+            raise RuntimeError("Leader robot failed to connect!")
+        
+        # Enable gravity compensation
+        if leader.pin_robot is not None:
+            leader.bus_right.enable_torque()
+            leader.bus_left.enable_torque()
+            time.sleep(0.1)
+            print(f"Leader connected with gravity compensation ({LEADER_LEFT_PORT}, {LEADER_RIGHT_PORT})")
+        else:
+            print(f"Leader connected but gravity compensation unavailable (no URDF)")
+
+    # Build default processors for action and observation
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+    
+    # Build dataset features from robot features and processors
+    # For actions, only include positions (no velocity or torque)
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+    
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+    
+    # Check if dataset already exists
+    dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / HF_EVAL_DATASET_ID
+    if dataset_path.exists():
+        print(f"Evaluation dataset already exists at: {dataset_path}")
+        print("This will append new episodes to the existing dataset.")
+        choice = input("  Continue? (y/n): ").strip().lower()
+        if choice != 'y':
+            print("  Aborting evaluation.")
+            follower.disconnect()
+            if leader:
+                leader.disconnect()
+            return
+    
+    # Create dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_EVAL_DATASET_ID,
+        fps=FPS,
+        features=dataset_features,
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_processes=0,
+        image_writer_threads=12, 
+    )
+    
+    # Load policy config from pretrained model and create policy using factory
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+    policy = make_policy(policy_config, ds_meta=dataset.meta)
+    
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy.config,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        preprocessor_overrides={
+            "device_processor": {"device": str(policy.config.device)}
+        },
+    )
+
+    print(f"\nRunning evaluation...")
+    # Initialize keyboard listener and visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="openarms_evaluation")
+    episode_idx = 0
+    
+    try:
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Evaluating episode {episode_idx + 1} of {NUM_EPISODES}")
+            print(f"\nRunning inference for episode {episode_idx + 1}...")
+            
+            # Run inference with policy
+            record_loop(
+                robot=follower,
+                events=events,
+                fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+                policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
+            
+            # Handle re-recording
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
+            
+            # Save episode
+            if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                print(f"Saving episode {episode_idx + 1} ({dataset.episode_buffer['size']} frames)...")
+                dataset.save_episode()
+                episode_idx += 1
+            
+            # Reset environment between episodes (if not last episode)
+            if not events["stop_recording"] and episode_idx < NUM_EPISODES:
+                if USE_LEADER_FOR_RESETS and leader:
+                    log_say("Reset the environment using leader arms")
+                    print(f"\nManual reset period ({RESET_TIME_SEC}s)...")
+                    
+                    # Use leader for manual reset with gravity compensation
+                    import numpy as np
+                    
+                    dt = 1 / FPS
+                    reset_start_time = time.perf_counter()
+                    
+                    while time.perf_counter() - reset_start_time < RESET_TIME_SEC:
+                        if events["exit_early"] or events["stop_recording"]:
+                            break
+                        
+                        loop_start = time.perf_counter()
+                        
+                        # Get leader state
+                        leader_action = leader.get_action()
+                        
+                        # Extract positions and velocities
+                        leader_positions_deg = {}
+                        leader_velocities_deg_per_sec = {}
+                        
+                        for motor in leader.bus_right.motors:
+                            pos_key = f"right_{motor}.pos"
+                            vel_key = f"right_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+                        
+                        for motor in leader.bus_left.motors:
+                            pos_key = f"left_{motor}.pos"
+                            vel_key = f"left_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+                        
+                        # Calculate gravity and friction torques
+                        leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+                        leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+                        
+                        leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+                        leader_friction_torques_nm = leader._friction_from_velocity(
+                            leader_velocities_rad_per_sec,
+                            friction_scale=1.0
+                        )
+                        
+                        # Combine torques
+                        leader_total_torques_nm = {}
+                        for motor_name in leader_gravity_torques_nm:
+                            gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+                            friction = leader_friction_torques_nm.get(motor_name, 0.0)
+                            leader_total_torques_nm[motor_name] = gravity + friction
+                        
+                        # Apply compensation
+                        for motor in leader.bus_right.motors:
+                            full_name = f"right_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_right._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        for motor in leader.bus_left.motors:
+                            full_name = f"left_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_left._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        # Send leader positions to follower
+                        follower_action = {}
+                        for joint in leader_positions_deg.keys():
+                            pos_key = f"{joint}.pos"
+                            if pos_key in leader_action:
+                                follower_action[pos_key] = leader_action[pos_key]
+                        
+                        if follower_action:
+                            follower.send_action(follower_action)
+                        
+                        # Maintain loop rate
+                        loop_duration = time.perf_counter() - loop_start
+                        sleep_time = dt - loop_duration
+                        if sleep_time > 0:
+                            time.sleep(sleep_time)
+                    
+                    print("Reset complete")
+                else:
+                    log_say("Waiting for manual reset")
+                    print(f"Manually reset the environment and press ENTER to continue")
+                    input("Press ENTER when ready...")
+        
+        print(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+    
+    except KeyboardInterrupt:
+        print("\n\nEvaluation interrupted by user")
+    
+    finally:
+        if leader:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+
+        follower.disconnect()
+        
+        if listener is not None:
+            listener.stop()
+        
+        dataset.finalize()
+        print("\nUploading to Hugging Face Hub...")
+        dataset.push_to_hub(private=True)
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/evaluate_with_rtc.py

@@ -0,0 +1,703 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+OpenArms Policy Evaluation with Real-Time Chunking (RTC)
+
+Evaluates a trained policy on the OpenArms robot using RTC for smooth, continuous motion.
+RTC enables large flow-matching policies (Pi0, Pi0.5, SmolVLA) to produce reactive motion
+despite high inference latency by asynchronously generating action chunks.
+
+Features:
+- Thread-based asynchronous action generation and execution
+- RTC for smooth transitions between action chunks
+- Dataset recording for evaluation episodes
+
+Example usage:
+    python examples/openarms/evaluate_with_rtc.py
+
+    # With custom RTC parameters
+    python examples/openarms/evaluate_with_rtc.py \
+        --rtc.execution_horizon=12 \
+        --rtc.max_guidance_weight=10.0
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from pathlib import Path
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor import make_default_processors
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging, log_say
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+# ============================================================================
+# Default Configuration Constants
+# ============================================================================
+
+DEFAULT_HF_MODEL_ID = "lerobot-data-collection/three-folds-pi0"
+DEFAULT_HF_EVAL_DATASET_ID = "lerobot-data-collection/three-folds-pi0_eval_rtc"
+DEFAULT_TASK_DESCRIPTION = "three-folds-dataset"
+
+DEFAULT_NUM_EPISODES = 1
+DEFAULT_FPS = 30
+DEFAULT_EPISODE_TIME_SEC = 300
+DEFAULT_RESET_TIME_SEC = 60
+
+DEFAULT_FOLLOWER_LEFT_PORT = "can0"
+DEFAULT_FOLLOWER_RIGHT_PORT = "can1"
+
+DEFAULT_CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video5", width=640, height=480, fps=DEFAULT_FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video1", width=640, height=480, fps=DEFAULT_FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video3", width=640, height=480, fps=DEFAULT_FPS),
+}
+
+
+# ============================================================================
+# Thread-Safe Robot Wrapper
+# ============================================================================
+
+
+class RobotWrapper:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: OpenArmsFollower):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self.lock:
+            self.robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        with self.lock:
+            return self.robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        with self.lock:
+            return self.robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self.robot.name
+
+
+# ============================================================================
+# Configuration
+# ============================================================================
+
+
+@dataclass
+class OpenArmsRTCEvalConfig(HubMixin):
+    """Configuration for OpenArms evaluation with RTC."""
+
+    policy: PreTrainedConfig | None = None
+
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=10,
+            max_guidance_weight=10.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+        )
+    )
+
+    model_id: str = DEFAULT_HF_MODEL_ID
+    eval_dataset_id: str = DEFAULT_HF_EVAL_DATASET_ID
+    task: str = DEFAULT_TASK_DESCRIPTION
+
+    num_episodes: int = DEFAULT_NUM_EPISODES
+    fps: float = DEFAULT_FPS
+    episode_time_sec: float = DEFAULT_EPISODE_TIME_SEC
+    reset_time_sec: float = DEFAULT_RESET_TIME_SEC
+
+    follower_left_port: str = DEFAULT_FOLLOWER_LEFT_PORT
+    follower_right_port: str = DEFAULT_FOLLOWER_RIGHT_PORT
+
+    device: str = "cuda"
+
+    # Should be higher than inference_delay + execution_horizon
+    action_queue_size_to_get_new_actions: int = 30
+
+    record_dataset: bool = True
+    push_to_hub: bool = True
+
+    interpolation: bool = False
+
+    use_torch_compile: bool = False
+    torch_compile_backend: str = "inductor"
+    torch_compile_mode: str = "default"
+    torch_compile_disable_cudagraphs: bool = True
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+            self.model_id = policy_path
+        elif self.model_id:
+            self.policy = PreTrainedConfig.from_pretrained(self.model_id)
+            self.policy.pretrained_path = self.model_id
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+# ============================================================================
+# Action Generation Thread
+# ============================================================================
+
+
+def get_actions_thread(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+):
+    """Thread function to asynchronously generate action chunks from the policy."""
+    try:
+        logger.info("[GET_ACTIONS] Starting action generation thread")
+
+        latency_tracker = LatencyTracker()
+        time_per_chunk = 1.0 / cfg.fps
+
+        hw_features = hw_to_dataset_features(robot.observation_features, "observation")
+        policy_device = policy.config.device
+
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                time.sleep(0.01)
+                continue
+
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk) if inference_latency else 0
+
+                obs = robot.get_observation()
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    hw_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]
+                obs_with_policy_features["robot_type"] = robot.name
+
+                preprocessed_obs = preprocessor(obs_with_policy_features)
+
+                actions = policy.predict_action_chunk(
+                    preprocessed_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                original_actions = actions.squeeze(0).clone()
+                postprocessed_actions = postprocessor(actions).squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] action_queue_size_to_get_new_actions too small. "
+                        "Should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+
+                logger.debug(
+                    f"[GET_ACTIONS] Generated chunk, latency={new_latency:.3f}s, "
+                    f"delay={new_delay}, queue_size={action_queue.qsize()}"
+                )
+            else:
+                time.sleep(0.01)
+
+        logger.info("[GET_ACTIONS] Action generation thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Action Execution Thread
+# ============================================================================
+
+
+def actor_thread(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+    dataset: LeRobotDataset | None,
+    dataset_lock: Lock,
+    teleop_action_processor,
+    robot_observation_processor,
+):
+    """Thread function to execute actions on the robot."""
+    try:
+        action_keys = [k for k in robot.action_features.keys() if k.endswith(".pos")]
+
+        if cfg.interpolation:
+            interp_factor = 2
+            robot_interval = 1.0 / (cfg.fps * interp_factor)
+            logger.info(f"[ACTOR] Interpolation ON: policy={cfg.fps}Hz -> robot={cfg.fps * interp_factor}Hz (2x)")
+        else:
+            interp_factor = 1
+            robot_interval = 1.0 / cfg.fps
+            logger.info(f"[ACTOR] Interpolation OFF: policy={cfg.fps}Hz, robot={cfg.fps}Hz")
+
+        prev_action: Tensor | None = None
+        interpolated_actions: list[Tensor] = []
+        interp_idx = 0
+
+        robot_send_count = 0
+        policy_consume_count = 0
+        last_hz_print = time.perf_counter()
+        last_dataset_time = 0.0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                prev_action = None
+                interpolated_actions = []
+                interp_idx = 0
+                robot_send_count = 0
+                policy_consume_count = 0
+                last_hz_print = time.perf_counter()
+                time.sleep(0.01)
+                continue
+
+            start_time = time.perf_counter()
+
+            if interp_idx >= len(interpolated_actions):
+                new_action = action_queue.get()
+                if new_action is not None:
+                    current_action = new_action.cpu()
+                    policy_consume_count += 1
+
+                    if cfg.interpolation and prev_action is not None:
+                        mid = prev_action + 0.5 * (current_action - prev_action)
+                        interpolated_actions = [mid, current_action]
+                    else:
+                        interpolated_actions = [current_action]
+
+                    prev_action = current_action
+                    interp_idx = 0
+
+            if interp_idx < len(interpolated_actions):
+                action_to_send = interpolated_actions[interp_idx]
+                interp_idx += 1
+
+                action_dict = {}
+                for i, key in enumerate(action_keys):
+                    if i < len(action_to_send):
+                        action_dict[key] = action_to_send[i].item()
+
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+                robot_send_count += 1
+
+                if cfg.record_dataset and dataset is not None:
+                    now = time.perf_counter()
+                    if now - last_dataset_time >= (1.0 / cfg.fps):
+                        last_dataset_time = now
+                        with dataset_lock:
+                            obs = robot.get_observation()
+                            obs_processed = robot_observation_processor(obs)
+                            action_for_dataset = teleop_action_processor((action_dict, None))
+                            frame = {}
+                            for key, value in obs_processed.items():
+                                frame[f"observation.{key}"] = value
+                            for key, value in action_for_dataset.items():
+                                frame[f"action.{key}"] = value
+                            frame["task"] = cfg.task
+                            dataset.add_frame(frame)
+
+            now = time.perf_counter()
+            if now - last_hz_print >= 5.0:
+                elapsed = now - last_hz_print
+                actual_robot_hz = robot_send_count / elapsed if elapsed > 0 else 0
+                actual_policy_hz = policy_consume_count / elapsed if elapsed > 0 else 0
+                logger.info(f"[ACTOR] Actual Hz - Robot: {actual_robot_hz:.1f}, Policy: {actual_policy_hz:.1f}")
+                robot_send_count = 0
+                policy_consume_count = 0
+                last_hz_print = now
+
+            dt_s = time.perf_counter() - start_time
+            sleep_time = max(0, robot_interval - dt_s - 0.001)
+            if sleep_time > 0:
+                time.sleep(sleep_time)
+
+        logger.info("[ACTOR] Shutting down")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Main Evaluation Function
+# ============================================================================
+
+
+def _apply_torch_compile(policy, cfg: OpenArmsRTCEvalConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method."""
+    if policy.name in ["pi05", "pi0"]:
+        return policy
+
+    try:
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def main(cfg: OpenArmsRTCEvalConfig):
+    """Main evaluation function with RTC."""
+    init_logging()
+
+    print("=" * 60)
+    print("OpenArms Policy Evaluation with RTC")
+    print("=" * 60)
+    print(f"\nModel: {cfg.model_id}")
+    print(f"Evaluation Dataset: {cfg.eval_dataset_id}")
+    print(f"Task: {cfg.task}")
+    print(f"Episodes: {cfg.num_episodes}")
+    print(f"Episode Duration: {cfg.episode_time_sec}s")
+    print(f"RTC Enabled: {cfg.rtc.enabled}")
+    print(f"RTC Execution Horizon: {cfg.rtc.execution_horizon}")
+    print(f"RTC Max Guidance Weight: {cfg.rtc.max_guidance_weight}")
+    print(f"Policy Hz: {cfg.fps}")
+    print(f"Robot Hz: {cfg.fps * 2 if cfg.interpolation else cfg.fps}")
+    print(f"Interpolation: {cfg.interpolation}")
+    print(f"Device: {cfg.device}")
+    print("=" * 60)
+
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+    episode_active = Event()
+
+    # Initialize Robot
+    follower_config = OpenArmsFollowerConfig(
+        port_left=cfg.follower_left_port,
+        port_right=cfg.follower_right_port,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=DEFAULT_CAMERA_CONFIG,
+    )
+
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    robot = RobotWrapper(follower)
+    logger.info("Follower robot connected")
+
+    # Build Processors and Dataset Features
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+
+    # Create or Load Dataset
+    dataset = None
+    dataset_lock = Lock()
+
+    if cfg.record_dataset:
+        dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / cfg.eval_dataset_id
+        if dataset_path.exists():
+            logger.info(f"Evaluation dataset exists at: {dataset_path}")
+            logger.info("New episodes will be appended.")
+            choice = input("Continue? (y/n): ").strip().lower()
+            if choice != "y":
+                logger.info("Aborting evaluation.")
+                follower.disconnect()
+                return
+
+        dataset = LeRobotDataset.create(
+            repo_id=cfg.eval_dataset_id,
+            fps=int(cfg.fps),
+            features=dataset_features,
+            robot_type=follower.name,
+            use_videos=True,
+            image_writer_processes=0,
+            image_writer_threads=12,
+        )
+        logger.info(f"Dataset created: {cfg.eval_dataset_id}")
+
+    # Load Policy
+    logger.info(f"Loading policy from: {cfg.model_id}")
+
+    policy_class = get_policy_class(cfg.policy.type)
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type in ["pi05", "pi0"]:
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    policy.config.rtc_config = cfg.rtc
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    logger.info(f"Policy loaded: {policy.name}")
+
+    # Create Action Queue and Start Threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    get_actions_t = Thread(
+        target=get_actions_thread,
+        args=(
+            policy,
+            robot,
+            robot_observation_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+        ),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_t.start()
+    logger.info("Started action generation thread")
+
+    actor_t = Thread(
+        target=actor_thread,
+        args=(
+            robot,
+            robot_action_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+            dataset,
+            dataset_lock,
+            teleop_action_processor,
+            robot_observation_processor,
+        ),
+        daemon=True,
+        name="Actor",
+    )
+    actor_t.start()
+    logger.info("Started actor thread")
+
+    # Run Evaluation Episodes
+    episode_idx = 0
+
+    try:
+        while episode_idx < cfg.num_episodes and not shutdown_event.is_set():
+            log_say(f"Evaluating episode {episode_idx + 1} of {cfg.num_episodes}")
+            logger.info(f"\n{'='*40}")
+            logger.info(f"Episode {episode_idx + 1} / {cfg.num_episodes}")
+            logger.info(f"{'='*40}")
+
+            action_queue = ActionQueue(cfg.rtc)
+            episode_active.set()
+            episode_start_time = time.time()
+
+            while (time.time() - episode_start_time) < cfg.episode_time_sec:
+                if shutdown_event.is_set():
+                    break
+
+                elapsed = time.time() - episode_start_time
+                if int(elapsed) % 10 == 0 and int(elapsed) > 0:
+                    logger.info(
+                        f"[MAIN] Episode progress: {elapsed:.0f}/{cfg.episode_time_sec}s, "
+                        f"queue_size={action_queue.qsize()}"
+                    )
+
+                time.sleep(0.5)
+
+            episode_active.clear()
+            logger.info(f"Episode {episode_idx + 1} completed")
+
+            if cfg.record_dataset and dataset is not None:
+                with dataset_lock:
+                    if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                        logger.info(
+                            f"Saving episode {episode_idx + 1} "
+                            f"({dataset.episode_buffer['size']} frames)"
+                        )
+                        dataset.save_episode()
+
+            episode_idx += 1
+
+            # Manual reset between episodes
+            if not shutdown_event.is_set() and episode_idx < cfg.num_episodes:
+                log_say("Waiting for manual reset")
+                logger.info("Manually reset the environment and press ENTER to continue")
+                input("Press ENTER when ready...")
+
+        logger.info(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+
+    except KeyboardInterrupt:
+        logger.info("\n\nEvaluation interrupted by user")
+
+    finally:
+        shutdown_event.set()
+        episode_active.clear()
+
+        if get_actions_t.is_alive():
+            logger.info("Waiting for action generation thread to finish...")
+            get_actions_t.join(timeout=5.0)
+
+        if actor_t.is_alive():
+            logger.info("Waiting for actor thread to finish...")
+            actor_t.join(timeout=5.0)
+
+        follower.disconnect()
+        logger.info("Follower disconnected")
+
+        if cfg.record_dataset and dataset is not None:
+            dataset.finalize()
+            if cfg.push_to_hub:
+                logger.info("Uploading to Hugging Face Hub...")
+                dataset.push_to_hub(private=True)
+
+        logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    main()

examples/openarms/friction_compensation.py

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

examples/openarms/gravity_compensation.py

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

examples/openarms/record_with_compensation.py

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

examples/openarms/replay.py

@@ -0,0 +1,166 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+OpenArms Dataset Replay Example
+
+Replays position actions from a recorded dataset on an OpenArms follower robot.
+Only position commands (ending with .pos) are replayed, not velocity or torque.
+
+Example usage:
+    python examples/openarms/replay.py
+"""
+
+import time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.constants import ACTION
+from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.utils import log_say
+
+# Configuration
+EPISODE_IDX = 0
+DATASET_REPO_ID = "lerobot-data-collection/replay-this-2025-11-02-17-58"  # TODO: Replace with your dataset
+DATASET_ROOT = None  # Use default cache location, or specify custom path
+
+# Robot configuration - adjust these to match your setup
+ROBOT_CONFIG = OpenArmsFollowerConfig(
+    port_left="can2",      # CAN interface for left arm
+    port_right="can3",     # CAN interface for right arm
+    can_interface="socketcan", 
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=10.0,  # Safety limit: max degrees to move per step
+)
+
+
+def main():
+    """Main replay function."""
+    print("=" * 70)
+    print("OpenArms Dataset Replay")
+    print("=" * 70)
+    print(f"\nDataset: {DATASET_REPO_ID}")
+    print(f"Episode: {EPISODE_IDX}")
+    print(f"Robot: {ROBOT_CONFIG.id}")
+    print(f"  Left arm: {ROBOT_CONFIG.port_left}")
+    print(f"  Right arm: {ROBOT_CONFIG.port_right}")
+    print("\n" + "=" * 70)
+    
+    # Initialize the robot
+    print("\n[1/3] Initializing robot...")
+    robot = OpenArmsFollower(ROBOT_CONFIG)
+    
+    # Load the dataset
+    print(f"\n[2/3] Loading dataset '{DATASET_REPO_ID}'...")
+    dataset = LeRobotDataset(
+        DATASET_REPO_ID,
+        root=DATASET_ROOT,
+        episodes=[EPISODE_IDX]
+    )
+    
+    # Filter dataset to only include frames from the specified episode
+    # (required for dataset V3.0 where episodes are chunked)
+    episode_frames = dataset.hf_dataset.filter(
+        lambda x: x["episode_index"] == EPISODE_IDX
+    )
+    
+    if len(episode_frames) == 0:
+        raise ValueError(
+            f"No frames found for episode {EPISODE_IDX} in dataset {DATASET_REPO_ID}"
+        )
+    
+    print(f"  Found {len(episode_frames)} frames in episode {EPISODE_IDX}")
+    
+    # Extract action features from dataset
+    action_features = dataset.features.get(ACTION, {})
+    action_names = action_features.get("names", [])
+    
+    # Filter to only position actions (ending with .pos)
+    position_action_names = [name for name in action_names if name.endswith(".pos")]
+    
+    if not position_action_names:
+        raise ValueError(
+            f"No position actions found in dataset. Action names: {action_names}"
+        )
+    
+    print(f"  Found {len(position_action_names)} position actions to replay")
+    print(f"  Actions: {', '.join(position_action_names[:5])}{'...' if len(position_action_names) > 5 else ''}")
+    
+    # Select only action columns from dataset
+    actions = episode_frames.select_columns(ACTION)
+    
+    # Connect to the robot
+    print(f"\n[3/3] Connecting to robot...")
+    robot.connect(calibrate=False)  # Skip calibration for replay
+    
+    if not robot.is_connected:
+        raise RuntimeError("Robot failed to connect!")
+    
+    print("\n" + "=" * 70)
+    print("Ready to replay!")
+    print("=" * 70)
+    print("\nThe robot will replay the recorded positions.")
+    print("Press Ctrl+C to stop at any time.\n")
+    
+    input("Press ENTER to start replaying...")
+    
+    # Replay loop
+    log_say(f"Replaying episode {EPISODE_IDX}", blocking=True)
+    
+    try:
+        for idx in range(len(episode_frames)):
+            loop_start = time.perf_counter()
+            
+            # Extract action array from dataset
+            action_array = actions[idx][ACTION]
+            
+            # Build action dictionary, but only include position actions
+            action = {}
+            for i, name in enumerate(action_names):
+                # Only include position actions (ending with .pos)
+                if name.endswith(".pos"):
+                    action[name] = float(action_array[i])
+            
+            # Send action to robot
+            robot.send_action(action)
+            
+            # Maintain replay rate (use dataset fps)
+            loop_duration = time.perf_counter() - loop_start
+            dt_s = 1.0 / dataset.fps - loop_duration
+            busy_wait(dt_s)
+            
+            # Progress indicator every 100 frames
+            if (idx + 1) % 100 == 0:
+                progress = (idx + 1) / len(episode_frames) * 100
+                print(f"Progress: {idx + 1}/{len(episode_frames)} frames ({progress:.1f}%)")
+        
+        print(f"\n✓ Successfully replayed {len(episode_frames)} frames")
+        log_say("Replay complete", blocking=True)
+        
+    except KeyboardInterrupt:
+        print("\n\nReplay interrupted by user")
+    finally:
+        # Disconnect robot
+        print("\nDisconnecting robot...")
+        robot.disconnect()
+        print("✓ Replay complete!")
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/setup_can.sh

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

examples/openarms/teleop.py

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

examples/openarms/teleop_openarms_mini.py

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

examples/openarms/teleop_with_compensation.py

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

examples/openarms/unify_task.py

@@ -0,0 +1,152 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Unify all tasks in a dataset to a single task (modifies in-place).
+
+This script:
+1. Loads a dataset
+2. Sets all task_index to 0 and task description to "fold"
+3. Updates tasks.parquet and task_index in data files (in-place, no copying)
+
+Usage:
+    python examples/openarms/unify_task.py --repo-id lerobot-data-collection/level1_rac1
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+from pathlib import Path
+
+import pandas as pd
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.utils import (
+    DATA_DIR,
+    write_info,
+    write_tasks,
+)
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+
+# Single unified task
+UNIFIED_TASK = "fold"
+
+
+def unify_dataset_tasks(
+    repo_id: str,
+    root: Path | None = None,
+    push_to_hub: bool = False,
+) -> None:
+    """Unify all tasks in a dataset to a single task (modifies in-place).
+
+    Args:
+        repo_id: Dataset repository ID.
+        root: Optional root path for dataset.
+        push_to_hub: Whether to push the result to HuggingFace Hub.
+    """
+    input_root = root if root else HF_LEROBOT_HOME / repo_id
+    input_repo_id = repo_id
+
+    logging.info(f"Loading metadata from {repo_id}")
+
+    # Load source metadata
+    src_meta = LeRobotDatasetMetadata(repo_id, root=input_root)
+
+    logging.info(f"Source dataset: {src_meta.total_episodes} episodes, {src_meta.total_frames} frames")
+    logging.info(f"Original tasks: {len(src_meta.tasks)}")
+
+    # Modify in-place (input_root == output_root supported)
+    data_dir = input_root / DATA_DIR
+
+    # Process data files - set all task_index to 0
+    logging.info("Processing data files (in-place)...")
+    for parquet_file in tqdm(sorted(data_dir.rglob("*.parquet")), desc="Processing data"):
+        df = pd.read_parquet(parquet_file)
+        df["task_index"] = 0  # All tasks unified to index 0
+        df.to_parquet(parquet_file)
+
+    # Process episodes metadata - set all tasks to unified task
+    logging.info("Processing episodes metadata (in-place)...")
+    episodes_dir = input_root / "meta" / "episodes"
+    if episodes_dir.exists():
+        for parquet_file in tqdm(sorted(episodes_dir.rglob("*.parquet")), desc="Processing episodes"):
+            df = pd.read_parquet(parquet_file)
+            df["tasks"] = [[UNIFIED_TASK]] * len(df)  # All episodes get the unified task
+            df.to_parquet(parquet_file)
+    else:
+        logging.warning(f"No episodes directory found at {episodes_dir}, skipping")
+
+    # Update tasks.parquet with single task
+    logging.info(f"Creating single task: {UNIFIED_TASK}")
+    new_tasks = pd.DataFrame({"task_index": [0]}, index=[UNIFIED_TASK])
+    write_tasks(new_tasks, input_root)
+
+    # Update info.json
+    new_info = src_meta.info.copy()
+    new_info["total_tasks"] = 1
+    write_info(new_info, input_root)
+
+    logging.info(f"Dataset modified in-place at {input_root}")
+    logging.info(f"Task: {UNIFIED_TASK}")
+
+    if push_to_hub:
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        logging.info(f"Pushing {input_repo_id} to hub")
+        dataset = LeRobotDataset(input_repo_id, root=input_root)
+        dataset.push_to_hub(private=True)
+        logging.info("Push complete!")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Unify all tasks in a dataset to a single task 'fold' (modifies in-place)."
+    )
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Dataset repository ID",
+    )
+    parser.add_argument(
+        "--root",
+        type=Path,
+        default=None,
+        help="Optional root path (defaults to HF_LEROBOT_HOME/repo_id)",
+    )
+    parser.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Push result to HuggingFace Hub",
+    )
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
+
+    unify_dataset_tasks(
+        repo_id=args.repo_id,
+        root=args.root,
+        push_to_hub=args.push_to_hub,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/openarms_web_interface/App.css

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

examples/openarms_web_interface/App.jsx

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

examples/openarms_web_interface/README.md

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

examples/openarms_web_interface/index.html

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

examples/openarms_web_interface/launch.sh

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

examples/openarms_web_interface/main.jsx

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

examples/openarms_web_interface/package.json

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

examples/openarms_web_interface/vite.config.js

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

examples/rac/rac_data_collection.py

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

examples/rac/rac_data_collection_openarms.py

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

examples/rac/rac_data_collection_openarms_rtc.py

@@ -0,0 +1,877 @@
+#!/usr/bin/env python
+"""
+RaC (Recovery and Correction) Data Collection for OpenArms Robot with RTC.
+
+This combines RaC data collection with Real-Time Chunking (RTC) for smooth policy execution.
+RTC enables large flow-matching policies (Pi0, Pi0.5, SmolVLA) to produce reactive motion
+despite high inference latency by asynchronously generating action chunks.
+
+The workflow:
+1. Policy runs autonomously with RTC (teleop is idle/free)
+2. Press SPACE to pause - teleop moves to match robot position
+3. Press 'c' to take control - teleop is free, human provides RECOVERY + CORRECTION
+4. Press → to end episode (save and continue to next)
+5. Reset, then do next rollout
+
+Usage:
+    python examples/rac/rac_data_collection_openarms_rtc.py \
+        --robot.port_right=can0 \
+        --robot.port_left=can1 \
+        --teleop.port_right=/dev/ttyUSB0 \
+        --teleop.port_left=/dev/ttyUSB1 \
+        --policy.path=outputs/train/my_policy/checkpoints/last/pretrained_model \
+        --dataset.repo_id=my_user/rac_openarms_dataset \
+        --dataset.single_task="Pick up the cube"
+"""
+
+import logging
+import math
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from pprint import pformat
+from threading import Event, Lock, Thread
+from typing import Any
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.image_writer import safe_stop_image_writer
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.datasets.video_utils import VideoEncodingManager
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.policies.utils import make_robot_action
+from lerobot.processor import (
+    IdentityProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.processor.rename_processor import rename_stats
+from lerobot.robots import Robot, RobotConfig, make_robot_from_config
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig  # noqa: F401
+from lerobot.teleoperators import Teleoperator, TeleoperatorConfig, make_teleoperator_from_config
+from lerobot.teleoperators.openarms_mini.config_openarms_mini import OpenArmsMiniConfig  # noqa: F401
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.control_utils import is_headless, predict_action
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import get_safe_torch_device, init_logging, log_say
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+# ============================================================================
+# Configuration
+# ============================================================================
+
+@dataclass
+class RaCRTCDatasetConfig:
+    repo_id: str = "lerobot/rac_openarms_rtc"
+    single_task: str = "default task"
+    root: str | Path | None = None
+    fps: int = 30
+    episode_time_s: float = 500
+    reset_time_s: float = 30
+    num_episodes: int = 50
+    video: bool = True
+    push_to_hub: bool = True
+    private: bool = False
+    tags: list[str] | None = None
+    num_image_writer_processes: int = 0
+    num_image_writer_threads_per_camera: int = 4
+    video_encoding_batch_size: int = 1
+    rename_map: dict[str, str] = field(default_factory=dict)
+
+
+@dataclass
+class RaCRTCConfig:
+    robot: RobotConfig = field(default_factory=lambda: OpenArmsFollowerConfig(
+        port_left="can0",
+        port_right="can1",
+    ))
+    teleop: TeleoperatorConfig = field(default_factory=lambda: OpenArmsMiniConfig(
+        port_left="/dev/ttyUSB1",
+        port_right="/dev/ttyUSB0",
+    ))
+    dataset: RaCRTCDatasetConfig = field(default_factory=RaCRTCDatasetConfig)
+    policy: PreTrainedConfig | None = None
+    
+    rtc: RTCConfig = field(default_factory=lambda: RTCConfig(
+        enabled=True, 
+        execution_horizon=20,
+        max_guidance_weight=5.0,
+        prefix_attention_schedule=RTCAttentionSchedule.LINEAR,
+    ))
+    
+    interpolation: bool = True
+    display_data: bool = True
+    play_sounds: bool = True
+    resume: bool = False
+    device: str = "cuda"
+    action_queue_size_to_get_new_actions: int = 30
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        if self.policy is None:
+            raise ValueError("policy.path is required")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+# ============================================================================
+# Thread-Safe Robot Wrapper (from evaluate_with_rtc.py)
+# ============================================================================
+
+class RobotWrapper:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: Robot):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self.lock:
+            self.robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        return self.robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        return self.robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self.robot.name
+    
+    @property
+    def robot_type(self) -> str:
+        return self.robot.robot_type
+
+
+# ============================================================================
+# Keyboard/Pedal Listeners
+# ============================================================================
+
+def init_rac_keyboard_listener():
+    """Initialize keyboard listener with RaC-specific controls."""
+    events = {
+        "exit_early": False,
+        "rerecord_episode": False,
+        "stop_recording": False,
+        "policy_paused": False,
+        "correction_active": False,
+        "in_reset": False,
+        "start_next_episode": False,
+    }
+
+    if is_headless():
+        logging.warning("Headless environment - keyboard controls unavailable")
+        return None, events
+
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if events["in_reset"]:
+                if key == keyboard.Key.space or key == keyboard.Key.right:
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    print("\n[RaC] Starting next episode...")
+                    events["start_next_episode"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["start_next_episode"] = True
+            else:
+                if key == keyboard.Key.space:
+                    if not events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ⏸ PAUSED - Policy stopped, teleop moving to robot position")
+                        print("      Press 'c' or START to take control")
+                        events["policy_paused"] = True
+                elif hasattr(key, 'char') and key.char == 'c':
+                    if events["policy_paused"] and not events["correction_active"]:
+                        print("\n[RaC] ▶ START pressed - taking control")
+                        events["start_next_episode"] = True
+                elif key == keyboard.Key.right:
+                    print("[RaC] → End episode")
+                    events["exit_early"] = True
+                elif key == keyboard.Key.left:
+                    print("[RaC] ← Re-record episode")
+                    events["rerecord_episode"] = True
+                    events["exit_early"] = True
+                elif key == keyboard.Key.esc:
+                    print("[RaC] ESC - Stop recording, pushing to hub...")
+                    events["stop_recording"] = True
+                    events["exit_early"] = True
+        except Exception as e:
+            print(f"Key error: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+    
+    start_pedal_listener(events)
+    
+    return listener, events
+
+
+def start_pedal_listener(events: dict):
+    """Start foot pedal listener thread if evdev is available."""
+    import threading
+    
+    try:
+        from evdev import InputDevice, ecodes  # noqa: F401
+    except ImportError:
+        logging.info("[Pedal] evdev not installed - pedal support disabled")
+        return
+    
+    PEDAL_DEVICE = "/dev/input/by-id/usb-PCsensor_FootSwitch-event-kbd"
+    KEY_LEFT = "KEY_A"
+    KEY_RIGHT = "KEY_C"
+    
+    def pedal_reader():
+        try:
+            dev = InputDevice(PEDAL_DEVICE)
+            print(f"[Pedal] Connected: {dev.name}")
+            
+            for ev in dev.read_loop():
+                if ev.type != ecodes.EV_KEY:
+                    continue
+                
+                from evdev import categorize  # noqa: F401
+                key = categorize(ev)
+                code = key.keycode
+                if isinstance(code, (list, tuple)):
+                    code = code[0]
+                
+                if key.keystate != 1:
+                    continue
+                
+                if events["in_reset"]:
+                    if code in [KEY_LEFT, KEY_RIGHT]:
+                        events["start_next_episode"] = True
+                else:
+                    if code == KEY_RIGHT:
+                        if events["correction_active"]:
+                            events["exit_early"] = True
+                        elif not events["policy_paused"]:
+                            events["policy_paused"] = True
+                    elif code == KEY_LEFT:
+                        if events["policy_paused"] and not events["correction_active"]:
+                            events["start_next_episode"] = True
+                        
+        except FileNotFoundError:
+            logging.info(f"[Pedal] Device not found: {PEDAL_DEVICE}")
+        except PermissionError:
+            logging.warning(f"[Pedal] Permission denied for {PEDAL_DEVICE}")
+        except Exception as e:
+            logging.debug(f"[Pedal] Error: {e}")
+    
+    thread = threading.Thread(target=pedal_reader, daemon=True)
+    thread.start()
+
+
+def make_identity_processors():
+    """Create identity processors for RaC recording."""
+    teleop_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    robot_proc = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[IdentityProcessorStep()],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+    obs_proc = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[IdentityProcessorStep()],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+    return teleop_proc, robot_proc, obs_proc
+
+
+# ============================================================================
+# RTC Inference Thread (from evaluate_with_rtc.py)
+# ============================================================================
+
+def rtc_inference_thread(
+    policy,
+    obs_holder: dict,
+    hw_features: dict,
+    preprocessor,
+    postprocessor,
+    queue_holder: dict,
+    shutdown_event: Event,
+    policy_active: Event,
+    cfg: RaCRTCConfig,
+):
+    """Background thread that generates action chunks using RTC."""
+    try:
+        logger.info("[RTC] ========== INFERENCE THREAD STARTED ==========")
+        logger.info(f"[RTC] policy={policy.name}, hw_features has {len(hw_features)} keys")
+        
+        latency_tracker = LatencyTracker()
+        time_per_chunk = 1.0 / cfg.dataset.fps
+        policy_device = policy.config.device
+        
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+        
+        inference_count = 0
+        wait_logged = False
+        
+        while not shutdown_event.is_set():
+            if not policy_active.is_set():
+                if not wait_logged:
+                    logger.info("[RTC] Waiting for policy_active...")
+                    wait_logged = True
+                time.sleep(0.01)
+                continue
+            
+            wait_logged = False
+            
+            action_queue = queue_holder["queue"]
+            if action_queue is None:
+                logger.warning("[RTC] queue_holder['queue'] is None!")
+                time.sleep(0.01)
+                continue
+            
+            obs_filtered = obs_holder.get("obs")
+            if obs_filtered is None:
+                logger.warning("[RTC] obs_holder['obs'] is None!")
+                time.sleep(0.01)
+                continue
+            
+            qsize = action_queue.qsize()
+            if qsize <= get_actions_threshold:
+                try:
+                    if inference_count == 0:
+                        logger.info(f"[RTC] Starting first inference, obs keys={len(obs_filtered)}, qsize={qsize}")
+                    
+                    current_time = time.perf_counter()
+                    action_index_before_inference = action_queue.get_action_index()
+                    prev_actions = action_queue.get_left_over()
+                    
+                    inference_latency = latency_tracker.max()
+                    inference_delay = math.ceil(inference_latency / time_per_chunk) if inference_latency else 0
+                    
+                    obs_with_policy_features = build_dataset_frame(hw_features, obs_filtered, prefix="observation")
+                    
+                    for name in obs_with_policy_features:
+                        obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                        if "image" in name:
+                            obs_with_policy_features[name] = obs_with_policy_features[name].float() / 255
+                            obs_with_policy_features[name] = obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0).to(policy_device)
+                    
+                    obs_with_policy_features["task"] = [cfg.dataset.single_task]
+                    obs_with_policy_features["robot_type"] = obs_holder.get("robot_type", "openarms_follower")
+                    
+                    preprocessed_obs = preprocessor(obs_with_policy_features)
+                    
+                    actions = policy.predict_action_chunk(
+                        preprocessed_obs,
+                        inference_delay=inference_delay,
+                        prev_chunk_left_over=prev_actions,
+                    )
+                    
+                    original_actions = actions.squeeze(0).clone()
+                    postprocessed_actions = postprocessor(actions).squeeze(0)
+                    
+                    new_latency = time.perf_counter() - current_time
+                    new_delay = math.ceil(new_latency / time_per_chunk)
+                    latency_tracker.add(new_latency)
+                    
+                    action_queue.merge(original_actions, postprocessed_actions, new_delay, action_index_before_inference)
+                    
+                    inference_count += 1
+                    logger.info(f"[RTC] Inference #{inference_count}, latency={new_latency:.2f}s, queue={action_queue.qsize()}")
+                except Exception as e:
+                    logger.error(f"[RTC] Inference error: {e}")
+                    import traceback
+                    traceback.print_exc()
+                    time.sleep(1.0)
+            else:
+                time.sleep(0.01)
+        
+        logger.info("[RTC] Inference thread shutting down")
+    except Exception as e:
+        logger.error(f"[RTC] THREAD CRASHED: {e}")
+        import traceback
+        traceback.print_exc()
+
+
+# ============================================================================
+# Main Rollout Loop
+# ============================================================================
+
+@safe_stop_image_writer
+def rac_rtc_rollout_loop(
+    robot: RobotWrapper,
+    teleop: Teleoperator,
+    policy: PreTrainedPolicy,
+    preprocessor,
+    postprocessor,
+    dataset: LeRobotDataset,
+    events: dict,
+    cfg: RaCRTCConfig,
+    queue_holder: dict,
+    obs_holder: dict,  # Main loop writes obs here for RTC thread to read
+    policy_active: Event,
+    hw_features: dict,
+) -> dict:
+    """RaC rollout loop with RTC for smooth policy execution."""
+    fps = cfg.dataset.fps
+    single_task = cfg.dataset.single_task
+    control_time_s = cfg.dataset.episode_time_s
+    device = get_safe_torch_device(cfg.device)
+    
+    # Reset policy state
+    policy.reset()
+    preprocessor.reset()
+    postprocessor.reset()
+    
+    frame_buffer = []
+    stats = {
+        "total_frames": 0,
+        "autonomous_frames": 0,
+        "paused_frames": 0,
+        "correction_frames": 0,
+    }
+
+    teleop.disable_torque()
+    was_paused = False
+    waiting_for_takeover = False
+    
+    # Action keys for converting tensor to dict
+    action_keys = [k for k in robot.action_features.keys() if k.endswith(".pos")]
+    
+    # Interpolation state
+    prev_action: Tensor | None = None
+    interpolated_actions: list[Tensor] = []
+    interp_idx = 0
+    
+    if cfg.interpolation:
+        control_interval = 1.0 / (fps * 2)  # 2x rate
+    else:
+        control_interval = 1.0 / fps
+    
+    robot_action = {}
+    timestamp = 0
+    start_t = time.perf_counter()
+
+    while timestamp < control_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["policy_paused"] = False
+            events["correction_active"] = False
+            break
+
+        # State transition: entering paused state
+        if events["policy_paused"] and not was_paused:
+            policy_active.clear()  # Stop RTC inference
+            obs = robot.get_observation()
+            obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+            robot_pos = {k: v for k, v in obs_filtered.items() if k.endswith(".pos")}
+            print("[RaC] Moving teleop to robot position...")
+            teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+            print("[RaC] Teleop aligned. Press 'c' to take control.")
+            events["start_next_episode"] = False
+            waiting_for_takeover = True
+            was_paused = True
+            # Reset interpolation
+            prev_action = None
+            interpolated_actions = []
+            interp_idx = 0
+
+        # Wait for takeover
+        if waiting_for_takeover and events["start_next_episode"]:
+            print("[RaC] Taking control...")
+            teleop.disable_torque()
+            events["start_next_episode"] = False
+            events["correction_active"] = True
+            waiting_for_takeover = False
+
+        # Get observation (ONLY the main loop reads from robot!)
+        obs = robot.get_observation()
+        obs_filtered = {k: v for k, v in obs.items() if k in robot.observation_features}
+        obs_frame = build_dataset_frame(dataset.features, obs_filtered, prefix=OBS_STR)
+        
+        # Share observation with RTC thread (thread reads, main loop writes)
+        obs_holder["obs"] = obs_filtered
+
+        if events["correction_active"]:
+            # Human controlling
+            robot_action = teleop.get_action()
+            for key in robot_action:
+                if "gripper" in key:
+                    robot_action[key] = -0.65 * robot_action[key]
+            robot.send_action(robot_action)
+            stats["correction_frames"] += 1
+            
+            action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+            frame = {**obs_frame, **action_frame, "task": single_task}
+            frame_buffer.append(frame)
+            stats["total_frames"] += 1
+            
+        elif waiting_for_takeover:
+            stats["paused_frames"] += 1
+            
+        elif events["policy_paused"]:
+            robot_pos = {k: v for k, v in obs_filtered.items() if k.endswith(".pos")}
+            teleop.send_feedback(robot_pos)
+            stats["paused_frames"] += 1
+            
+        else:
+            # Policy execution with RTC
+            if not policy_active.is_set():
+                policy_active.set()
+                logger.info("[ROLLOUT] Policy activated, waiting for first actions...")
+            
+            action_queue = queue_holder["queue"]
+            
+            # Get action from queue (with interpolation)
+            if interp_idx >= len(interpolated_actions):
+                new_action = action_queue.get() if action_queue else None
+                
+                # Log queue status periodically
+                if stats["autonomous_frames"] == 0 and new_action is None:
+                    qsize = action_queue.qsize() if action_queue else -1
+                    if timestamp < 0.5 or int(timestamp * 10) % 10 == 0:
+                        logger.info(f"[ROLLOUT] Waiting for actions... queue_size={qsize}, obs_set={obs_holder.get('obs') is not None}")
+                
+                if new_action is not None:
+                    current_action = new_action.cpu()
+                    
+                    if cfg.interpolation and prev_action is not None:
+                        mid = prev_action + 0.5 * (current_action - prev_action)
+                        interpolated_actions = [mid, current_action]
+                    else:
+                        interpolated_actions = [current_action]
+                    
+                    prev_action = current_action
+                    interp_idx = 0
+                    
+                    if stats["autonomous_frames"] == 0:
+                        logger.info(f"[ROLLOUT] Got first action! Starting robot motion.")
+            
+            if interp_idx < len(interpolated_actions):
+                action_to_send = interpolated_actions[interp_idx]
+                interp_idx += 1
+                
+                robot_action = {}
+                for i, key in enumerate(action_keys):
+                    if i < len(action_to_send):
+                        robot_action[key] = action_to_send[i].item()
+                
+                robot.send_action(robot_action)
+                stats["autonomous_frames"] += 1
+                
+                # Record at original fps
+                action_frame = build_dataset_frame(dataset.features, robot_action, prefix=ACTION)
+                frame = {**obs_frame, **action_frame, "task": single_task}
+                frame_buffer.append(frame)
+                stats["total_frames"] += 1
+
+        if cfg.display_data:
+            log_rerun_data(observation=obs_filtered, action=robot_action)
+
+        dt = time.perf_counter() - loop_start
+        sleep_time = control_interval - dt
+        if sleep_time > 0:
+            precise_sleep(sleep_time)
+        timestamp = time.perf_counter() - start_t
+
+    policy_active.clear()
+    teleop.disable_torque()
+
+    for frame in frame_buffer:
+        dataset.add_frame(frame)
+
+    return stats
+
+
+def reset_loop(robot: RobotWrapper, teleop: Teleoperator, events: dict, fps: int):
+    """Reset period where human repositions environment."""
+    print("\n" + "=" * 65)
+    print("  [RaC] RESET")
+    print("=" * 65)
+    
+    events["in_reset"] = True
+    events["start_next_episode"] = False
+    
+    obs = robot.get_observation()
+    robot_pos = {k: v for k, v in obs.items() if k.endswith(".pos") and k in robot.observation_features}
+    teleop.smooth_move_to(robot_pos, duration_s=2.0, fps=50)
+    
+    print("  Press any key/pedal to enable teleoperation")
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        precise_sleep(0.05)
+    
+    if events["stop_recording"]:
+        return
+    
+    events["start_next_episode"] = False
+    teleop.disable_torque()
+    print("  Teleop enabled - press any key/pedal to start episode")
+
+    while not events["start_next_episode"] and not events["stop_recording"]:
+        loop_start = time.perf_counter()
+        action = teleop.get_action()
+        for key in action:
+            if "gripper" in key:
+                action[key] = -0.65 * action[key]
+        robot.send_action(action)
+        dt = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt)
+    
+    events["in_reset"] = False
+    events["start_next_episode"] = False
+    events["exit_early"] = False
+    events["policy_paused"] = False
+    events["correction_active"] = False
+
+
+# ============================================================================
+# Main Entry Point
+# ============================================================================
+
+@parser.wrap()
+def rac_rtc_collect(cfg: RaCRTCConfig) -> LeRobotDataset:
+    """Main RaC data collection function with RTC."""
+    init_logging()
+    logging.info(pformat(cfg.__dict__))
+
+    if cfg.display_data:
+        init_rerun(session_name="rac_rtc_collection_openarms")
+
+    robot_raw = make_robot_from_config(cfg.robot)
+    teleop = make_teleoperator_from_config(cfg.teleop)
+    
+    teleop_proc, robot_proc, obs_proc = make_identity_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot_raw.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot_raw.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    dataset = None
+    listener = None
+    shutdown_event = Event()
+    policy_active = Event()
+    rtc_thread = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+            if hasattr(robot_raw, "cameras") and robot_raw.cameras:
+                dataset.start_image_writer(
+                    num_processes=cfg.dataset.num_image_writer_processes,
+                    num_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot_raw.cameras),
+                )
+        else:
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot_raw.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera
+                * len(robot_raw.cameras if hasattr(robot_raw, "cameras") else []),
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+            )
+
+        # Load policy
+        logger.info(f"Loading policy from: {cfg.policy.pretrained_path}")
+        policy_class = get_policy_class(cfg.policy.type)
+        policy = policy_class.from_pretrained(cfg.policy.pretrained_path)
+        policy.config.rtc_config = cfg.rtc
+        policy.init_rtc_processor()
+        policy = policy.to(cfg.device)
+        policy.eval()
+        logger.info(f"Policy loaded: {policy.name}")
+
+        # Setup preprocessor/postprocessor
+        hw_features = hw_to_dataset_features(robot_raw.observation_features, "observation")
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=rename_stats(dataset.meta.stats, cfg.dataset.rename_map),
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+                "rename_observations_processor": {"rename_map": cfg.dataset.rename_map},
+            },
+        )
+
+        # Connect robot and wrap for thread safety
+        robot_raw.connect()
+        robot = RobotWrapper(robot_raw)
+        
+        teleop.connect()
+        listener, events = init_rac_keyboard_listener()
+
+        # Shared state holders (main loop writes, RTC thread reads)
+        queue_holder = {"queue": ActionQueue(cfg.rtc)}
+        obs_holder = {"obs": None, "robot_type": robot.robot_type}  # Main loop updates obs
+
+        # Start RTC inference thread
+        # NOTE: Thread does NOT access robot directly - reads from obs_holder
+        rtc_thread = Thread(
+            target=rtc_inference_thread,
+            args=(
+                policy,
+                obs_holder,  # Thread reads obs from here (set by main loop)
+                hw_features,
+                preprocessor,
+                postprocessor,
+                queue_holder,
+                shutdown_event,
+                policy_active,
+                cfg,
+            ),
+            daemon=True,
+            name="RTCInference",
+        )
+        rtc_thread.start()
+        logger.info("Started RTC inference thread")
+
+        print("\n" + "=" * 65)
+        print("  RaC Data Collection with RTC")
+        print("=" * 65)
+        print(f"  Policy: {cfg.policy.pretrained_path}")
+        print(f"  Task: {cfg.dataset.single_task}")
+        print(f"  FPS: {cfg.dataset.fps}")
+        print(f"  Interpolation: {cfg.interpolation}")
+        print()
+        print("  Controls:")
+        print("    SPACE  - Pause policy")
+        print("    c      - Take control")
+        print("    →      - End episode")
+        print("    ESC    - Stop and push to hub")
+        print("=" * 65 + "\n")
+
+        with VideoEncodingManager(dataset):
+            recorded = 0
+            while recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                log_say(f"RaC episode {dataset.num_episodes}", cfg.play_sounds)
+                
+                # Fresh action queue per episode (update holder so thread sees it)
+                queue_holder["queue"] = ActionQueue(cfg.rtc)
+                
+                logger.info(f"Episode {recorded + 1} / {cfg.dataset.num_episodes}")
+
+                stats = rac_rtc_rollout_loop(
+                    robot=robot,
+                    teleop=teleop,
+                    policy=policy,
+                    preprocessor=preprocessor,
+                    postprocessor=postprocessor,
+                    dataset=dataset,
+                    events=events,
+                    cfg=cfg,
+                    queue_holder=queue_holder,
+                    obs_holder=obs_holder,
+                    policy_active=policy_active,
+                    hw_features=hw_features,
+                )
+
+                logging.info(f"Episode stats: {stats}")
+
+                if events["rerecord_episode"]:
+                    log_say("Re-recording", cfg.play_sounds)
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                dataset.save_episode()
+                recorded += 1
+
+                if recorded < cfg.dataset.num_episodes and not events["stop_recording"]:
+                    reset_loop(robot, teleop, events, cfg.dataset.fps)
+
+    finally:
+        log_say("Stop recording", cfg.play_sounds, blocking=True)
+        
+        shutdown_event.set()
+        policy_active.clear()
+        
+        if rtc_thread and rtc_thread.is_alive():
+            rtc_thread.join(timeout=2.0)
+
+        if dataset:
+            dataset.finalize()
+
+        if robot_raw.is_connected:
+            robot_raw.disconnect()
+        if teleop.is_connected:
+            teleop.disconnect()
+
+        if not is_headless() and listener:
+            listener.stop()
+
+        if cfg.dataset.push_to_hub:
+            dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+def main():
+    from lerobot.utils.import_utils import register_third_party_plugins
+    register_third_party_plugins()
+    rac_rtc_collect()
+
+
+if __name__ == "__main__":
+    main()

loop_datasets.py

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

src/lerobot/datasets/lerobot_dataset.py

@@ -23,11 +23,13 @@ from pathlib import Path
 
 import datasets
 import numpy as np
+import os
 import packaging.version
 import pandas as pd
 import PIL.Image
 import pyarrow as pa
 import pyarrow.parquet as pq
+from concurrent.futures import ProcessPoolExecutor
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
@@ -1199,40 +1201,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         use_batched_encoding = self.batch_encoding_size > 1
 
         if has_video_keys and not use_batched_encoding:
-            num_cameras = len(self.meta.video_keys)
-            if parallel_encoding and num_cameras > 1:
-                # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
-                # num_cameras * num_threads = (total_cpu -1)
-                with concurrent.futures.ProcessPoolExecutor(max_workers=num_cameras) as executor:
-                    future_to_key = {
-                        executor.submit(
-                            _encode_video_worker,
-                            video_key,
-                            episode_index,
-                            self.root,
-                            self.fps,
-                        ): video_key
-                        for video_key in self.meta.video_keys
-                    }
-
-                    results = {}
-                    for future in concurrent.futures.as_completed(future_to_key):
-                        video_key = future_to_key[future]
-                        try:
-                            temp_path = future.result()
-                            results[video_key] = temp_path
-                        except Exception as exc:
-                            logging.error(f"Video encoding failed for {video_key}: {exc}")
-                            raise exc
-
-                for video_key in self.meta.video_keys:
-                    temp_path = results[video_key]
-                    ep_metadata.update(
-                        self._save_episode_video(video_key, episode_index, temp_path=temp_path)
-                    )
-            else:
-                for video_key in self.meta.video_keys:
-                    ep_metadata.update(self._save_episode_video(video_key, episode_index))
+            video_paths = self._encode_multiple_temporary_episode_videos(self.meta.video_keys, episode_index)
+            for video_key, video_path in zip(self.meta.video_keys, video_paths):
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, video_path))
 
         # `meta.save_episode` need to be executed after encoding the videos
         self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
@@ -1528,6 +1499,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         return _encode_video_worker(video_key, episode_index, self.root, self.fps)
 
+    def _encode_multiple_temporary_episode_videos(self, video_keys, episode_index):
+        temp_paths = []
+        img_dirs = []
+        for video_key in video_keys:
+            temp_paths.append(Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4")
+            img_dirs.append(self._get_image_file_dir(episode_index, video_key))
+        fps = [self.fps]*len(video_keys)
+
+        with ProcessPoolExecutor(max_workers=len(video_keys)) as executor:
+            executor.map(encode_video_frames,img_dirs,temp_paths,fps)
+
+        for img_dir in img_dirs:
+            shutil.rmtree(img_dir)
+
+        return temp_paths
+
     @classmethod
     def create(
         cls,

src/lerobot/datasets/video_utils.py

@@ -310,7 +310,7 @@ def encode_video_frames(
     crf: int | None = 30,
     fast_decode: int = 0,
     log_level: int | None = av.logging.ERROR,
-    overwrite: bool = False,
+    overwrite: bool = True,
     preset: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
@@ -355,6 +355,9 @@ def encode_video_frames(
     if crf is not None:
         video_options["crf"] = str(crf)
 
+    #TEMPORARY FIX
+    video_options["preset"] = "12"
+
     if fast_decode:
         key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"

src/lerobot/motors/__init__.py

@@ -14,4 +14,11 @@
 # 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,
+    MotorsBus,  # Backward compatibility (alias for SerialMotorsBus)
+    MotorsBusBase,
+    SerialMotorsBus,
+)

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 +1,905 @@
+# 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.
+
+# TODO(pepijn): add license of: https://github.com/cmjang/DM_Control_Python?tab=MIT-1-ov-file#readme
+
+import logging
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from functools import cached_property
+from typing import Dict, List, Optional, Tuple, Union
+
+import can
+import numpy as np
+
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode, MotorsBusBase
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.utils import enter_pressed, move_cursor_up
+
+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,
+    MODEL_RESOLUTION,
+    MOTOR_LIMIT_PARAMS,
+    NORMALIZED_DATA,
+    MotorType,
+)
+
+logger = logging.getLogger(__name__)
+
+NameOrID = Union[str, int]
+Value = Union[int, 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
+    
+    # Motor configuration
+    model_resolution_table = deepcopy(MODEL_RESOLUTION)
+    normalized_data = deepcopy(NORMALIZED_DATA)
+
+    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 = None
+        self._is_connected = False
+        
+        # Map motor names to CAN IDs
+        self._motor_can_ids = {}
+        self._recv_id_to_motor = {}
+        
+        # Store motor types and recv IDs
+        self._motor_types = {}
+        for name, motor in self.motors.items():
+            if hasattr(motor, "motor_type"):
+                self._motor_types[name] = motor.motor_type
+            else:
+                # Default to DM4310 if not specified
+                self._motor_types[name] = MotorType.DM4310
+            
+            # Map recv_id to motor name for filtering responses
+            if hasattr(motor, "recv_id"):
+                self._recv_id_to_motor[motor.recv_id] = name
+
+    @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/"):
+                    # Serial device (macOS/Windows)
+                    self.can_interface = "slcan"
+                    logger.info(f"Auto-detected slcan interface for port {self.port}")
+                else:
+                    # Network interface (Linux)
+                    self.can_interface = "socketcan"
+                    logger.info(f"Auto-detected socketcan interface for port {self.port}")
+            
+            # Connect to CAN bus
+            if self.can_interface == "socketcan":
+                # Linux SocketCAN with CAN FD support
+                if self.use_can_fd and self.data_bitrate is not None:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface="socketcan",
+                        bitrate=self.bitrate,
+                        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:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface="socketcan",
+                        bitrate=self.bitrate
+                    )
+                    logger.info(f"Connected to {self.port} with CAN 2.0 (bitrate={self.bitrate})")
+            elif self.can_interface == "slcan":
+                # Serial Line CAN (macOS, Windows, or USB adapters)
+                # Note: SLCAN typically doesn't support CAN FD
+                self.canbus = can.interface.Bus(
+                    channel=self.port,
+                    interface="slcan",
+                    bitrate=self.bitrate
+                )
+                logger.info(f"Connected to {self.port} with SLCAN (bitrate={self.bitrate})")
+            else:
+                # Generic interface (vector, pcan, etc.)
+                if self.use_can_fd and self.data_bitrate is not None:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface=self.can_interface,
+                        bitrate=self.bitrate,
+                        data_bitrate=self.data_bitrate,
+                        fd=True
+                    )
+                else:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface=self.can_interface,
+                        bitrate=self.bitrate
+                    )
+            
+            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}")
+
+    def _handshake(self) -> None:
+        """Verify all motors are present by refreshing their status."""
+        for motor_name in self.motors:
+            self._refresh_motor(motor_name)
+            time.sleep(0.01)  # Small delay between motors
+
+    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._enable_motor(motor)
+            time.sleep(0.01)
+
+    def _enable_motor(self, motor: NameOrID) -> None:
+        """Enable a single motor."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [CAN_CMD_ENABLE]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _disable_motor(self, motor: NameOrID) -> None:
+        """Disable a single motor."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [CAN_CMD_DISABLE]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._enable_motor(motor)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(0.01)
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._disable_motor(motor)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(0.01)
+
+    @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.
+        
+        Examples:
+            >>> with bus.torque_disabled():
+            ...     # Safe operations here with torque disabled
+            ...     pass
+        """
+        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."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            data = [0xFF] * 7 + [CAN_CMD_SET_ZERO]
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+            self._recv_motor_response(expected_recv_id=recv_id)
+            time.sleep(0.01)
+
+    def _refresh_motor(self, motor: NameOrID) -> Optional[can.Message]:
+        """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)
+        self.canbus.send(msg)
+        return self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _recv_motor_response(self, expected_recv_id: Optional[int] = None, timeout: float = 0.001) -> Optional[can.Message]:
+        """
+        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=0.0001)  # 100us timeout for fast polling
+                if msg:
+                    messages_seen.append(f"0x{msg.arbitration_id:02X}")
+                    # If no filter specified, return any message
+                    if expected_recv_id is None:
+                        return msg
+                    # Otherwise, only return if it matches the expected recv_id
+                    if msg.arbitration_id == expected_recv_id:
+                        return msg
+                    else:
+                        logger.debug(f"Ignoring message from CAN ID 0x{msg.arbitration_id:02X}, expected 0x{expected_recv_id:02X}")
+            
+            # Only log warnings if we're in debug mode to reduce overhead
+            if logger.isEnabledFor(logging.DEBUG):
+                if messages_seen:
+                    logger.debug(f"Received {len(messages_seen)} message(s) from IDs {set(messages_seen)}, but expected 0x{expected_recv_id:02X}")
+                else:
+                    logger.debug(f"No CAN messages received (expected from 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:
+                msg = self.canbus.recv(timeout=0.0002)  # 200us poll timeout (increased from 100us for better reliability)
+                if msg and msg.arbitration_id in expected_set:
+                    responses[msg.arbitration_id] = msg
+                    if len(responses) == len(expected_recv_ids):
+                        break  # Got all responses, exit early
+        except Exception as e:
+            logger.debug(f"Error receiving responses: {e}")
+        
+        return responses
+
+    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.
+
+        Args:
+            motor: Motor name or ID
+            kp: Position gain
+            kd: Velocity gain
+            position_degrees: Target position (degrees)
+            velocity_deg_per_sec: Target velocity (degrees/s)
+            torque: Target torque (N·m)
+        """
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        motor_type = self._motor_types.get(motor_name, MotorType.DM4310)
+        
+        # Convert degrees to radians for motor control
+        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, 0, 500, 12)
+        kd_uint = self._float_to_uint(kd, 0, 5, 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
+        
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        recv_id = self._get_motor_recv_id(motor)
+        self._recv_motor_response(expected_recv_id=recv_id)
+    
+    def _mit_control_batch(
+        self,
+        commands: Dict[NameOrID, Tuple[float, float, float, float, float]],
+    ) -> None:
+        """
+        Send MIT control commands to multiple motors in batch (optimized).
+        Sends all commands first, then collects responses. Much faster than sequential.
+        
+        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
+        
+        expected_recv_ids = []
+        
+        # Step 1: Send all MIT control commands (no waiting)
+        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.get(motor_name, MotorType.DM4310)
+            
+            # 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, 0, 500, 12)
+            kd_uint = self._float_to_uint(kd, 0, 5, 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
+            
+            # Send command
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+            
+            # Track expected response
+            recv_id = self._get_motor_recv_id(motor)
+            expected_recv_ids.append(recv_id)
+        
+        # Step 2: Collect all responses at once
+        self._recv_all_responses(expected_recv_ids, timeout=0.002)
+
+    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: bytes, motor_type: MotorType) -> Tuple[float, float, float, int, int]:
+        """
+        Decode motor state from CAN data.
+
+        Returns:
+            Tuple of (position_degrees, velocity_deg_per_sec, 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 (radians)
+        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)
+        
+        # Convert to degrees
+        position_degrees = np.degrees(position_rad)
+        velocity_deg_per_sec = np.degrees(velocity_rad_per_sec)
+        
+        return position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor
+
+    def read(
+        self,
+        data_name: str,
+        motor: str,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> 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"
+            )
+        
+        motor_type = self._motor_types.get(motor, MotorType.DM4310)
+        position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+        
+        # Return requested data (already in degrees for position/velocity)
+        if data_name == "Present_Position":
+            value = position_degrees
+        elif data_name == "Present_Velocity":
+            value = velocity_deg_per_sec
+        elif data_name == "Present_Torque":
+            value = torque
+        elif data_name == "Temperature_MOS":
+            value = t_mos
+        elif data_name == "Temperature_Rotor":
+            value = t_rotor
+        else:
+            raise ValueError(f"Unknown data_name: {data_name}")
+        
+        # For Damiao, positions are always in degrees, no normalization needed
+        # We keep the normalize parameter for compatibility but don't use it
+        return value
+
+    def write(
+        self,
+        data_name: str,
+        motor: str,
+        value: Value,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> None:
+        """Write a value to a single motor. Positions are always in degrees."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Value is expected to be in degrees for positions
+        if data_name == "Goal_Position":
+            # Use MIT control with position in degrees
+            self._mit_control(motor, 10.0, 0.5, value, 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,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> Dict[str, Value]:
+        """
+        Read the same value from multiple motors simultaneously.
+        Uses batched operations: sends all refresh commands, then collects all responses.
+        This is MUCH faster than sequential reads (OpenArms pattern).
+        """
+        motors = self._get_motors_list(motors)
+        result = {}
+
+        # Step 1: Send refresh commands to ALL motors first (no waiting)
+        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)
+            self.canbus.send(msg)
+        
+        # Step 2: Collect all responses at once (batch receive)
+        expected_recv_ids = [self._get_motor_recv_id(motor) for motor in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=0.01)  # 10ms total timeout
+        
+        # Step 3: Parse responses
+        for motor in motors:
+            try:
+                recv_id = self._get_motor_recv_id(motor)
+                msg = responses.get(recv_id)
+                
+                if msg is None:
+                    logger.warning(f"No response from motor '{motor}' (recv ID: 0x{recv_id:02X})")
+                    result[motor] = 0.0
+                    continue
+                
+                motor_type = self._motor_types.get(motor, MotorType.DM4310)
+                position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+                
+                # Return requested data
+                if data_name == "Present_Position":
+                    value = position_degrees
+                elif data_name == "Present_Velocity":
+                    value = velocity_deg_per_sec
+                elif data_name == "Present_Torque":
+                    value = torque
+                elif data_name == "Temperature_MOS":
+                    value = t_mos
+                elif data_name == "Temperature_Rotor":
+                    value = t_rotor
+                else:
+                    raise ValueError(f"Unknown data_name: {data_name}")
+                
+                result[motor] = value
+                
+            except Exception as e:
+                logger.warning(f"Failed to read {data_name} from {motor}: {e}")
+                result[motor] = 0.0
+        
+        return result
+    
+    def sync_read_all_states(
+        self,
+        motors: str | list[str] | None = None,
+        *,
+        num_retry: int = 0,
+    ) -> Dict[str, Dict[str, Value]]:
+        """
+        Read ALL motor states (position, velocity, torque) from multiple motors in ONE refresh cycle.
+        This is 3x faster than calling sync_read() three times separately.
+        
+        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}, ...}
+        """
+        motors = self._get_motors_list(motors)
+        result = {}
+
+        # Step 1: Send refresh commands to ALL motors first (with small delays to reduce bus congestion)
+        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)
+            self.canbus.send(msg)
+            time.sleep(0.0001)  # 100us delay between commands to reduce bus congestion
+        
+        # Step 2: Collect all responses at once (batch receive)
+        expected_recv_ids = [self._get_motor_recv_id(motor) for motor in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=0.015)  # 15ms timeout (increased for reliability)
+        
+        # Step 3: Parse responses and extract ALL state values
+        for motor in motors:
+            try:
+                recv_id = self._get_motor_recv_id(motor)
+                msg = responses.get(recv_id)
+                
+                if msg is None:
+                    logger.warning(f"No response from motor '{motor}' (recv ID: 0x{recv_id:02X})")
+                    result[motor] = {"position": 0.0, "velocity": 0.0, "torque": 0.0}
+                    continue
+                
+                motor_type = self._motor_types.get(motor, MotorType.DM4310)
+                position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+                
+                # Return all state values in one dict
+                result[motor] = {
+                    "position": position_degrees,
+                    "velocity": velocity_deg_per_sec,
+                    "torque": torque,
+                    "temp_mos": t_mos,
+                    "temp_rotor": t_rotor,
+                }
+                
+            except Exception as e:
+                logger.warning(f"Failed to read state from {motor}: {e}")
+                result[motor] = {"position": 0.0, "velocity": 0.0, "torque": 0.0}
+        
+        return result
+
+    def sync_write(
+        self,
+        data_name: str,
+        values: Dict[str, Value],
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> None:
+        """
+        Write different values to multiple motors simultaneously. Positions are always in degrees.
+        Uses batched operations: sends all commands first, then collects responses (OpenArms pattern).
+        """
+        if data_name == "Goal_Position":
+            # Step 1: Send all MIT control commands first (no waiting)
+            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.get(motor_name, MotorType.DM4310)
+                
+                # Convert degrees to radians
+                position_rad = np.radians(value_degrees)
+                
+                # Default gains for position control
+                kp, kd = 10.0, 0.5
+                
+                # Get motor limits and encode parameters
+                pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+                kp_uint = self._float_to_uint(kp, 0, 500, 12)
+                kd_uint = self._float_to_uint(kd, 0, 5, 12)
+                q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+                dq_uint = self._float_to_uint(0, -vmax, vmax, 12)
+                tau_uint = self._float_to_uint(0, -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
+                
+                msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+                self.canbus.send(msg)
+                time.sleep(0.0001)  # 100us delay between commands to reduce bus congestion
+            
+            # Step 2: Collect all responses at once
+            expected_recv_ids = [self._get_motor_recv_id(motor) for motor in values.keys()]
+            self._recv_all_responses(expected_recv_ids, timeout=0.015)  # 15ms timeout (increased for reliability)
+        else:
+            # Fall back to individual writes for other data types
+            for motor, value in values.items():
+                self.write(data_name, motor, value, normalize=normalize, num_retry=num_retry)
+
+    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.
+        """
+        if motors is None:
+            motors = list(self.motors.keys())
+        elif isinstance(motors, (str, int)):
+            motors = [motors]
+
+        # Disable torque for manual movement
+        self.disable_torque(motors)
+        time.sleep(0.1)
+
+        # Get initial positions (already in degrees)
+        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        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", motors, normalize=False)
+            
+            for motor in 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 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 to overwrite the previous output
+                move_cursor_up(len(motors) + 4)
+
+            time.sleep(0.05)
+
+        # Re-enable torque
+        self.enable_torque(motors)
+        
+        # Validate ranges
+        for motor in motors:
+            if motor in mins and motor in maxes:
+                if abs(maxes[motor] - mins[motor]) < 5.0:  # At least 5 degrees of range
+                    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) -> Optional[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 hasattr(motor_obj, "recv_id"):
+            return motor_obj.recv_id
+        return 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
+from typing import Dict, List, Tuple
+
+# 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
+
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
+# OpenArms specific configurations
+# Based on: https://docs.openarm.dev/software/setup/configure-test
+# OpenArms has 7 DOF per arm (14 total for dual arm)
+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

@@ -24,7 +24,7 @@ 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 ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     AVAILABLE_BAUDRATES,
     MODEL_BAUDRATE_TABLE,
@@ -100,7 +100,7 @@ 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:

src/lerobot/motors/feetech/feetech.py

@@ -19,7 +19,7 @@ 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 ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     FIRMWARE_MAJOR_VERSION,
     FIRMWARE_MINOR_VERSION,
@@ -96,7 +96,7 @@ 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.
@@ -165,7 +165,7 @@ class FeetechMotorsBus(MotorsBus):
 
     def _handshake(self) -> None:
         self._assert_motors_exist()
-        self._assert_same_firmware()
+        #self._assert_same_firmware()
 
     def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
         if self.protocol_version == 0:

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 +43,92 @@ 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, *, normalize: bool = True, num_retry: int = 0) -> Value:
+        """Read a value from a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def write(
+        self, data_name: str, motor: str, value: Value, *, normalize: bool = True, num_retry: int = 0
+    ) -> None:
+        """Write a value to a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def sync_read(
+        self, data_name: str, motors: str | list[str] | None = None, *, normalize: bool = True
+    ) -> dict[str, Value]:
+        """Read a value from multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def sync_write(
+        self,
+        data_name: str,
+        values: Value | dict[str, Value],
+        motors: str | list[str] | None = None,
+        *,
+        normalize: bool = True,
+    ) -> 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: int | 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:
@@ -203,15 +291,15 @@ 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:
@@ -1212,3 +1300,7 @@ 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 = SerialMotorsBus

src/lerobot/policies/wall_x/__init__.py

@@ -15,5 +15,7 @@
 # limitations under the License.
 
 from .configuration_wall_x import WallXConfig
+from .modeling_wall_x import WallXPolicy
+from .processor_wall_x import make_wall_x_pre_post_processors
 
 __all__ = ["WallXConfig", "WallXPolicy", "make_wall_x_pre_post_processors"]

src/lerobot/robots/openarms/__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_openarms_follower import OpenArmsFollowerConfig
+from .openarms_follower import OpenArmsFollower
+
+__all__ = ["OpenArmsFollower", "OpenArmsFollowerConfig"]

src/lerobot/robots/openarms/config_openarms_follower.py

@@ -0,0 +1,118 @@
+#!/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 typing import Dict, Optional
+
+from lerobot.cameras import CameraConfig
+from lerobot.motors.damiao.tables import MotorType
+
+from ..config import RobotConfig
+
+
+@RobotConfig.register_subclass("openarms_follower")
+@dataclass
+class OpenArmsFollowerConfig(RobotConfig):
+    """Configuration for the OpenArms follower robot with Damiao motors."""
+    
+    # CAN interfaces - one per arm
+    # Right arm CAN interface (e.g., "can0")
+    # Left arm CAN interface (e.g., "can1")
+    # Linux: "can0", "can1", etc.
+    # macOS: "/dev/cu.usbmodem*" (serial device)
+    port_right: str = "can0"    # CAN interface for right arm
+    port_left: str = "can1"     # CAN interface for left arm
+    
+    # CAN interface type: "socketcan" (Linux), "slcan" (macOS/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: Optional[float | Dict[str, float]] = 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])
+    
+    # Damping gains for stability when applying torque compensation (gravity/friction)
+    # Used when kp=0 and only torque is applied
+    damping_kd: list[float] = field(default_factory=lambda: [0.5, 0.5, 0.5, 0.5, 0.1, 0.1, 0.1, 0.1])
+    
+    # Friction model parameters: τ_fric(ω) = Fo + Fv·ω + Fc·tanh(k·ω)
+    # From OpenArms config/follower.yaml
+    friction_fc: list[float] = field(default_factory=lambda: [0.306, 0.306, 0.40, 0.166, 0.050, 0.093, 0.172, 0.0512])  # Coulomb friction [Nm]
+    friction_k: list[float] = field(default_factory=lambda: [28.417, 28.417, 29.065, 130.038, 151.771, 242.287, 7.888, 4.000])  # tanh steepness
+    friction_fv: list[float] = field(default_factory=lambda: [0.063, 0.0630, 0.604, 0.813, 0.029, 0.072, 0.084, 0.084])  # Viscous friction [Nm·s/rad]
+    friction_fo: list[float] = field(default_factory=lambda: [0.088, 0.088, 0.008, -0.058, 0.005, 0.009, -0.059, -0.050])  # Offset torque [Nm]
+    
+    # Calibration parameters
+    calibration_mode: str = "manual"  # "manual" or "auto"
+    zero_position_on_connect: bool = False  # Set zero position on connect
+    
+    # Joint limits for position clipping (degrees)
+    # Format: [min, max] for each joint
+    # These limits clip commands in send_action to prevent mechanical damage
+    joint_limits_right: Dict[str, tuple[float, float]] = field(default_factory=lambda: {
+        "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),
+    })
+    
+    joint_limits_left: Dict[str, tuple[float, float]] = field(default_factory=lambda: {
+        "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),
+    })

src/lerobot/robots/openarms/openarms_follower.py

@@ -0,0 +1,698 @@
+#!/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, Dict, Optional
+
+import numpy as np
+import pinocchio as pin
+
+from lerobot.cameras.utils import make_cameras_from_configs
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode
+from lerobot.motors.damiao import DamiaoMotorsBus
+from lerobot.motors.damiao.tables import MotorType
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..robot import Robot
+from ..utils import ensure_safe_goal_position
+from .config_openarms_follower import OpenArmsFollowerConfig
+
+logger = logging.getLogger(__name__)
+
+
+class OpenArmsFollower(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 = OpenArmsFollowerConfig
+    name = "openarms_follower"
+
+    def __init__(self, config: OpenArmsFollowerConfig):
+        super().__init__(config)
+        self.config = config
+        
+        norm_mode_body = MotorNormMode.DEGREES  # Always use degrees for Damiao motors
+        
+        # Right arm motors (on port_right)
+        # Each arm uses the same CAN IDs since they're on separate buses
+        motors_right = {}
+        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
+            motor = Motor(send_id, motor_type_str, norm_mode_body)
+            motor.recv_id = recv_id
+            motor.motor_type = getattr(MotorType, motor_type_str.upper().replace("-", "_"))
+            motors_right[motor_name] = motor
+        
+        # Left arm motors (on port_left, same IDs as right since separate bus)
+        motors_left = {}
+        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
+            motor = Motor(send_id, motor_type_str, norm_mode_body)
+            motor.recv_id = recv_id
+            motor.motor_type = getattr(MotorType, motor_type_str.upper().replace("-", "_"))
+            motors_left[motor_name] = motor
+        
+        # Initialize separate Damiao motors buses (one per arm) with CAN FD support
+        self.bus_right = DamiaoMotorsBus(
+            port=self.config.port_right,
+            motors=motors_right,
+            calibration={k.replace("right_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("right_")},
+            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,
+        )
+        
+        self.bus_left = DamiaoMotorsBus(
+            port=self.config.port_left,
+            motors=motors_left,
+            calibration={k.replace("left_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("left_")},
+            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,
+        )
+        
+        # Initialize cameras
+        self.cameras = make_cameras_from_configs(config.cameras)
+        # Cache for last valid camera frames (to avoid blocking on slow USB reads)
+        self.camera_frame_cache = {key: None for key in self.cameras.keys()}
+        
+        # Initialize Pinocchio robot model for dynamics (optional)
+        self.pin_robot = None
+        try:
+            # Load URDF - try external path first (with meshes), then repository
+            import os
+            from os.path import expanduser, dirname
+            
+            # Try external URDF with meshes first
+            external_urdf_path = expanduser("~/Documents/openarm_description/openarm_bimanual_pybullet.urdf")
+            if os.path.exists(external_urdf_path):
+                urdf_path = external_urdf_path
+                urdf_dir = dirname(urdf_path)
+            
+                self.pin_robot = pin.RobotWrapper.BuildFromURDF(urdf_path, urdf_dir)
+                self.pin_robot.data = self.pin_robot.model.createData()
+                logger.info(f"Loaded OpenArms URDF for dynamics computation from {urdf_path}")
+        except Exception as e:
+            logger.warning(f"Could not load URDF for dynamics: {e}. Gravity compensation will not be available.")
+
+    @property
+    def _motors_ft(self) -> Dict[str, type]:
+        """Motor features for observation and action spaces."""
+        features = {}
+        # Right arm motors - only positions stored in dataset
+        for motor in self.bus_right.motors:
+            features[f"right_{motor}.pos"] = float
+        # Left arm motors - only positions stored in dataset
+        for motor in self.bus_left.motors:
+            features[f"left_{motor}.pos"] = float
+        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 (motor positions only)."""
+        return self._motors_ft
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if robot is connected."""
+        return (self.bus_right.is_connected and 
+                self.bus_left.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 both CAN buses
+        logger.info(f"Connecting right arm on {self.config.port_right}...")
+        self.bus_right.connect()
+        logger.info(f"Connecting left arm on {self.config.port_left}...")
+        self.bus_left.connect()
+        
+        # Run calibration if needed
+        if calibrate:
+            logger.info(
+                "No calibration found or overwriting calibration. Running calibration..."
+            )
+            self.calibrate()
+        
+        # Connect cameras
+        for cam in self.cameras.values():
+            cam.connect()
+        
+        # Configure motors
+        self.configure()
+        
+        # Optionally set zero position
+        if self.config.zero_position_on_connect:
+            logger.info("Setting current position as zero...")
+            self.bus_right.set_zero_position()
+            self.bus_left.set_zero_position()
+        
+        logger.info(f"{self} connected.")
+
+    @property
+    def is_calibrated(self) -> bool:
+        """Check if robot is calibrated."""
+        return self.bus_right.is_calibrated and self.bus_left.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:
+            # Ask user whether to use existing calibration
+            user_input = input(
+                f"Press ENTER to use existing calibration for {self.id}, "
+                f"or type 'c' and press ENTER to run new calibration: "
+            )
+            if user_input.strip().lower() != "c":
+                logger.info(f"Using existing calibration for {self.id}")
+                # Split calibration for each bus
+                cal_right = {k.replace("right_", ""): v for k, v in self.calibration.items() if k.startswith("right_")}
+                cal_left = {k.replace("left_", ""): v for k, v in self.calibration.items() if k.startswith("left_")}
+                self.bus_right.write_calibration(cal_right)
+                self.bus_left.write_calibration(cal_left)
+                return
+        
+        logger.info(f"\nRunning calibration for {self}")
+        
+        # Calibrate each arm separately
+        self._calibrate_arm("right", self.bus_right)
+        self._calibrate_arm("left", self.bus_left)
+        
+        print(f"\nCalibration complete and saved to {self.calibration_fpath}")
+    
+    def _calibrate_arm(self, arm_name: str, bus: DamiaoMotorsBus) -> None:
+        """Calibrate a single arm."""
+        logger.info(f"\n=== Calibrating {arm_name.upper()} arm ===")
+        
+        # Disable torque for manual positioning
+        bus.disable_torque()
+        time.sleep(0.1)
+        
+        # Step 1: Set zero position
+        input(
+            f"\nCalibration: Zero Position ({arm_name.upper()} arm)\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
+        bus.set_zero_position()
+        logger.info(f"{arm_name.capitalize()} arm zero position set.")
+        
+        # Automatically set range to -90° to +90° for all joints
+        print(
+            f"\nAutomatically setting range: -90° to +90° for all joints"
+        )
+        
+        # Create calibration data with fixed ranges
+        if self.calibration is None:
+            self.calibration = {}
+            
+        for motor_name, motor in bus.motors.items():
+            # Prefix motor name with arm name for storage
+            prefixed_name = f"{arm_name}_{motor_name}"
+            
+            # Use -90 to +90 for all joints and gripper (integers required)
+            self.calibration[prefixed_name] = MotorCalibration(
+                id=motor.id,
+                drive_mode=0,  # Normal direction
+                homing_offset=0,  # Already set via set_zero_position
+                range_min=-90,  # -90 degrees (integer)
+                range_max=90,   # +90 degrees (integer)
+            )
+            logger.info(f"  {prefixed_name}: range set to [-90°, +90°]")
+        
+        # Write calibration to this arm's motors
+        cal_for_bus = {k.replace(f"{arm_name}_", ""): v for k, v in self.calibration.items() if k.startswith(f"{arm_name}_")}
+        bus.write_calibration(cal_for_bus)
+        
+        # Re-enable torque
+        bus.enable_torque()
+        
+        # Save calibration after each arm
+        self._save_calibration()
+
+    def configure(self) -> None:
+        """Configure motors with appropriate settings."""
+        # Configure right arm
+        with self.bus_right.torque_disabled():
+            self.bus_right.configure_motors()
+        
+        # Configure left arm
+        with self.bus_left.torque_disabled():
+            self.bus_left.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) -> Dict[str, Any]:
+        """
+        Get current observation from robot including position, velocity, and torque.
+        
+        OPTIMIZED: Reads all motor states (pos/vel/torque) in one CAN refresh cycle
+        instead of 3 separate reads.
+        
+        Note: Velocity and torque are read but not stored in dataset (only used for
+        internal calculations). Only positions and camera images are stored.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        obs_dict = {}
+        
+        # Detailed profiling for bottleneck analysis
+        timings = {}
+        
+        # OPTIMIZED: Use sync_read_all_states to get pos/vel/torque in one go
+        t0 = time.perf_counter()
+        right_states = self.bus_right.sync_read_all_states()
+        timings["right_motors"] = (time.perf_counter() - t0) * 1000
+        
+        for motor in self.bus_right.motors:
+            state = right_states.get(motor, {})
+            obs_dict[f"right_{motor}.pos"] = state.get("position", 0.0)
+            obs_dict[f"right_{motor}.vel"] = state.get("velocity", 0.0)
+            obs_dict[f"right_{motor}.torque"] = state.get("torque", 0.0)
+        
+        # OPTIMIZED: Use sync_read_all_states to get pos/vel/torque in one go
+        t0 = time.perf_counter()
+        left_states = self.bus_left.sync_read_all_states()
+        timings["left_motors"] = (time.perf_counter() - t0) * 1000
+        
+        for motor in self.bus_left.motors:
+            state = left_states.get(motor, {})
+            obs_dict[f"left_{motor}.pos"] = state.get("position", 0.0)
+            obs_dict[f"left_{motor}.vel"] = state.get("velocity", 0.0)
+            obs_dict[f"left_{motor}.torque"] = state.get("torque", 0.0)
+        
+        # Capture images from cameras (with individual timing)
+        # Use async_read with very short timeout to avoid blocking on slow USB cameras
+        for cam_key, cam in self.cameras.items():
+            t0 = time.perf_counter()
+            try:
+                # Use 5ms timeout - if frame isn't ready, reuse last frame
+                frame = cam.async_read(timeout_ms=5)
+                self.camera_frame_cache[cam_key] = frame  # Update cache
+                obs_dict[cam_key] = frame
+            except TimeoutError:
+                # If no new frame available, reuse last valid frame from cache
+                # This prevents blocking the entire control loop on slow USB reads
+                if self.camera_frame_cache[cam_key] is not None:
+                    obs_dict[cam_key] = self.camera_frame_cache[cam_key]
+                    logger.debug(f"Camera {cam_key} timeout, reusing cached frame")
+
+            # Store timing with padded name to align output (e.g. "left_wrist    ")
+            timings[f"{cam_key:14s}"] = (time.perf_counter() - t0) * 1000
+        
+        # Log detailed timings (for debugging slow observations)
+        if logger.isEnabledFor(logging.DEBUG):
+            total_time = sum(timings.values())
+            breakdown = " | ".join([f"{k}: {v:.1f}ms" for k, v in timings.items()])
+            logger.debug(f"{self} get_observation: {total_time:.1f}ms total | {breakdown}")
+        
+        # Store timings in obs_dict for external profiling
+        obs_dict["_timing_breakdown"] = timings
+        
+        return obs_dict
+
+    def send_action(
+        self, 
+        action: Dict[str, Any], 
+        custom_kp: Optional[Dict[str, float]] = None,
+        custom_kd: Optional[Dict[str, float]] = None
+    ) -> Dict[str, Any]:
+        """
+        Send action command to robot.
+        
+        The action magnitude may be clipped based on safety limits.
+        
+        Args:
+            action: Dictionary with motor positions (e.g., "right_joint_1.pos", "left_joint_2.pos")
+            custom_kp: Optional custom kp gains per motor (e.g., {"right_joint_1": 120.0, "left_joint_2": 150.0})
+            custom_kd: Optional custom kd gains per motor (e.g., {"right_joint_1": 1.5, "left_joint_2": 2.0})
+            
+        Returns:
+            The action actually sent (potentially clipped)
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Extract motor positions from action and split by arm
+        goal_pos_right = {}
+        goal_pos_left = {}
+        
+        for key, val in action.items():
+            if key.endswith(".pos"):
+                motor_name = key.removesuffix(".pos")
+                if motor_name.startswith("right_"):
+                    # Remove "right_" prefix for bus access
+                    goal_pos_right[motor_name.removeprefix("right_")] = val
+                elif motor_name.startswith("left_"):
+                    # Remove "left_" prefix for bus access
+                    goal_pos_left[motor_name.removeprefix("left_")] = val
+        
+        # Apply joint limit clipping to right arm
+        for motor_name, position in goal_pos_right.items():
+            if motor_name in self.config.joint_limits_right:
+                min_limit, max_limit = self.config.joint_limits_right[motor_name]
+                clipped_position = max(min_limit, min(max_limit, position))
+                if clipped_position != position:
+                    logger.debug(f"Clipped right_{motor_name} from {position:.2f}° to {clipped_position:.2f}°")
+                goal_pos_right[motor_name] = clipped_position
+        
+        # Apply joint limit clipping to left arm
+        for motor_name, position in goal_pos_left.items():
+            if motor_name in self.config.joint_limits_left:
+                min_limit, max_limit = self.config.joint_limits_left[motor_name]
+                clipped_position = max(min_limit, min(max_limit, position))
+                if clipped_position != position:
+                    logger.debug(f"Clipped left_{motor_name} from {position:.2f}° to {clipped_position:.2f}°")
+                goal_pos_left[motor_name] = clipped_position
+        
+        # Apply safety limits if configured
+        if self.config.max_relative_target is not None:
+            # Get current positions
+            present_pos_right = self.bus_right.sync_read("Present_Position")
+            present_pos_left = self.bus_left.sync_read("Present_Position")
+            
+            # Apply safety limits to right arm
+            if goal_pos_right:
+                goal_present_pos_right = {
+                    key: (g_pos, present_pos_right.get(key, 0.0)) 
+                    for key, g_pos in goal_pos_right.items()
+                }
+                goal_pos_right = ensure_safe_goal_position(
+                    goal_present_pos_right, 
+                    self.config.max_relative_target
+                )
+            
+            # Apply safety limits to left arm
+            if goal_pos_left:
+                goal_present_pos_left = {
+                    key: (g_pos, present_pos_left.get(key, 0.0)) 
+                    for key, g_pos in goal_pos_left.items()
+                }
+                goal_pos_left = ensure_safe_goal_position(
+                    goal_present_pos_left, 
+                    self.config.max_relative_target
+                )
+        
+        # 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 right arm (sends all commands, then collects responses)
+        if goal_pos_right:
+            commands_right = {}
+            for motor_name, position_degrees in goal_pos_right.items():
+                idx = motor_index.get(motor_name, 0)
+                
+                # Use custom gains if provided, otherwise use config defaults
+                full_motor_name = f"right_{motor_name}"
+                if custom_kp is not None and full_motor_name in custom_kp:
+                    kp = custom_kp[full_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 full_motor_name in custom_kd:
+                    kd = custom_kd[full_motor_name]
+                else:
+                    kd = self.config.position_kd[idx] if isinstance(self.config.position_kd, list) else self.config.position_kd
+                
+                commands_right[motor_name] = (kp, kd, position_degrees, 0.0, 0.0)
+            self.bus_right._mit_control_batch(commands_right)
+        
+        # Use batch MIT control for left arm (sends all commands, then collects responses)
+        if goal_pos_left:
+            commands_left = {}
+            for motor_name, position_degrees in goal_pos_left.items():
+                idx = motor_index.get(motor_name, 0)
+                
+                # Use custom gains if provided, otherwise use config defaults
+                full_motor_name = f"left_{motor_name}"
+                if custom_kp is not None and full_motor_name in custom_kp:
+                    kp = custom_kp[full_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 full_motor_name in custom_kd:
+                    kd = custom_kd[full_motor_name]
+                else:
+                    kd = self.config.position_kd[idx] if isinstance(self.config.position_kd, list) else self.config.position_kd
+                
+                commands_left[motor_name] = (kp, kd, position_degrees, 0.0, 0.0)
+            self.bus_left._mit_control_batch(commands_left)
+        
+        # Return the actions that were actually sent
+        result = {}
+        for motor, val in goal_pos_right.items():
+            result[f"right_{motor}.pos"] = val
+        for motor, val in goal_pos_left.items():
+            result[f"left_{motor}.pos"] = val
+        return result
+
+    def disconnect(self):
+        """Disconnect from robot."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Disconnect from CAN buses
+        self.bus_right.disconnect(self.config.disable_torque_on_disconnect)
+        self.bus_left.disconnect(self.config.disable_torque_on_disconnect)
+        
+        # Disconnect cameras
+        for cam in self.cameras.values():
+            cam.disconnect()
+        
+        logger.info(f"{self} disconnected.")
+    
+    def _deg_to_rad(self, deg: Dict[str, float | int]) -> Dict[str, float]:
+        """Convert degrees to radians for all motors."""
+        return {m: np.deg2rad(float(v)) for m, v in deg.items()}
+    
+    def _gravity_from_q(self, q_rad: Dict[str, float]) -> Dict[str, float]:
+        """
+        Compute g(q) [N·m] for all joints in the robot.
+        The order of joints in the URDF matches the concatenated motor lists (right then left).
+        
+        Args:
+            q_rad: Dictionary mapping motor names (with arm prefix) to positions in radians
+            
+        Returns:
+            Dictionary mapping motor names to gravity torques in N·m
+            
+        Raises:
+            RuntimeError: If URDF model is not loaded
+        """
+        if self.pin_robot is None:
+            raise RuntimeError(
+                "Cannot compute gravity: URDF model not loaded. "
+                "Ensure urdf/openarms.urdf exists and is valid."
+            )
+        
+        # Build position vector in the order of motors (left arm, then right arm)
+        # This order must match the URDF joint order
+        # URDF has: left_joint1-7, left_finger_joint1-2, right_joint1-7, right_finger_joint1-2
+        q = np.zeros(self.pin_robot.model.nq)
+        idx = 0
+        
+        # Left arm motors (first in URDF) - joints 1-7
+        for motor_name in self.bus_left.motors:
+            if motor_name == "gripper":
+                continue  # Skip gripper, will be handled separately
+            full_name = f"left_{motor_name}"
+            q[idx] = q_rad.get(full_name, 0.0)
+            idx += 1
+        
+        # Skip left finger joints (leave as zeros)
+        idx += 2
+        
+        # Right arm motors (second in URDF) - joints 1-7
+        for motor_name in self.bus_right.motors:
+            if motor_name == "gripper":
+                continue  # Skip gripper, will be handled separately
+            full_name = f"right_{motor_name}"
+            q[idx] = q_rad.get(full_name, 0.0)
+            idx += 1
+        
+        # Skip right finger joints (leave as zeros)
+        idx += 2
+        
+        # Compute generalized gravity vector
+        g = pin.computeGeneralizedGravity(self.pin_robot.model, self.pin_robot.data, q)
+        
+        # Map back to motor names (only arm joints, not fingers)
+        result = {}
+        idx = 0
+        
+        # Left arm torques (joints 1-7)
+        for motor_name in self.bus_left.motors:
+            if motor_name == "gripper":
+                result["left_gripper"] = 0.0  # No gravity compensation for gripper
+                continue
+            result[f"left_{motor_name}"] = float(g[idx])
+            idx += 1
+        
+        # Skip left finger joint torques in output
+        idx += 2
+        
+        # Right arm torques (joints 1-7)
+        for motor_name in self.bus_right.motors:
+            if motor_name == "gripper":
+                result["right_gripper"] = 0.0  # No gravity compensation for gripper
+                continue
+            result[f"right_{motor_name}"] = float(g[idx])
+            idx += 1
+        
+        # Skip right finger joint torques in output
+        idx += 2
+        
+        return result
+    
+    def _friction_from_velocity(
+        self, 
+        velocity_rad_per_sec: Dict[str, float],
+        friction_scale: float = 1.0,
+        amp_tmp: float = 1.0,
+        coef_tmp: float = 0.1
+    ) -> Dict[str, float]:
+        """
+        Compute friction torques for all joints in the robot using tanh friction model.
+        
+        Args:
+            velocity_rad_per_sec: Dictionary mapping motor names (with arm prefix) to velocities in rad/s
+            friction_scale: Scale factor for friction compensation (default 1.0, use 0.3 for stability)
+            amp_tmp: Amplitude factor for tanh term (default 1.0)
+            coef_tmp: Coefficient for tanh steepness (default 0.1)
+            
+        Returns:
+            Dictionary mapping motor names to friction torques in N·m
+        """
+        # Motor name to index mapping
+        motor_name_to_index = {
+            "joint_1": 0,
+            "joint_2": 1,
+            "joint_3": 2,
+            "joint_4": 3,
+            "joint_5": 4,
+            "joint_6": 5,
+            "joint_7": 6,
+            "gripper": 7,
+        }
+        
+        result = {}
+        
+        # Process all motors (left and right)
+        for motor_full_name, velocity in velocity_rad_per_sec.items():
+            # Extract motor name without arm prefix
+            if motor_full_name.startswith("right_"):
+                motor_name = motor_full_name.removeprefix("right_")
+            elif motor_full_name.startswith("left_"):
+                motor_name = motor_full_name.removeprefix("left_")
+            else:
+                result[motor_full_name] = 0.0
+                continue
+            
+            # Get motor index for friction parameters
+            motor_index = motor_name_to_index.get(motor_name, 0)
+            
+            # Get friction parameters from config
+            Fc = self.config.friction_fc[motor_index]
+            k = self.config.friction_k[motor_index]
+            Fv = self.config.friction_fv[motor_index]
+            Fo = self.config.friction_fo[motor_index]
+            
+            # Friction model: τ_fric = amp * Fc * tanh(coef * k * ω) + Fv * ω + Fo
+            friction_torque = (
+                amp_tmp * Fc * np.tanh(coef_tmp * k * velocity) +
+                Fv * velocity +
+                Fo
+            )
+            
+            # Apply scale factor
+            friction_torque *= friction_scale
+            
+            result[motor_full_name] = float(friction_torque)
+        
+        return result
+    
+    def get_damping_kd(self, motor_name: str) -> float:
+        """
+        Get damping gain (Kd) for a specific motor.
+        
+        Args:
+            motor_name: Motor name without arm prefix (e.g., "joint_1", "gripper")
+            
+        Returns:
+            Damping gain value
+        """
+        motor_name_to_index = {
+            "joint_1": 0,
+            "joint_2": 1,
+            "joint_3": 2,
+            "joint_4": 3,
+            "joint_5": 4,
+            "joint_6": 5,
+            "joint_7": 6,
+            "gripper": 7,
+        }
+        
+        motor_index = motor_name_to_index.get(motor_name, 0)
+        return self.config.damping_kd[motor_index]
+    

src/lerobot/scratch.txt

@@ -0,0 +1,41 @@
+# Eun visualizer locally
+
+# login to hf an set your access token
+hf auth login
+# if not installed, install with: pip install huggingface_hub
+git clone https://github.com/huggingface/lerobot-dataset-visualizer.git
+cd lerobot-dataset-visualizer
+python -m lerobot_dataset_viz --repo-id lerobot-data-collection/repo-id-nez --episode-index 0
+git checkout feat/private_repo_viz
+npm install
+npm run dev
+# open http://localhost:3000 in your browser
+
+
+# ======================================================
+
+
+# default merge command; copy your list of datasets ids in repo_ids
+
+python -m lerobot.scripts.lerobot_edit_dataset \
+--repo_id lerobot-data-collection/repo-id-nez \
+--operation.type merge --push_to_hub true \
+--operation.repo_ids "[]"
+
+
+# merge test datasets into one
+
+python -m lerobot.scripts.lerobot_edit_dataset \
+--repo_id lerobot-data-collection/test-2025-11-03-merged \
+--operation.type merge --push_to_hub true \
+--operation.repo_ids "['lerobot-data-collection/test-2025-11-03-13-18', 'lerobot-data-collection/test-2025-11-03-13-19', 'lerobot-data-collection/test-2025-11-03-13-20', 'lerobot-data-collection/test-2025-11-03-13-21', 'lerobot-data-collection/test-2025-11-03-13-23', 'lerobot-data-collection/test-2025-11-03-13-24', 'lerobot-data-collection/test-2025-11-03-13-25', 'lerobot-data-collection/test-2025-11-03-13-26', 'lerobot-data-collection/test-2025-11-03-13-27', 'lerobot-data-collection/test-2025-11-03-13-29', 'lerobot-data-collection/test-2025-11-03-13-30', 'lerobot-data-collection/test-2025-11-03-13-31', 'lerobot-data-collection/test-2025-11-03-13-34', 'lerobot-data-collection/test-2025-11-03-13-41', 'lerobot-data-collection/test-2025-11-03-13-42', 'lerobot-data-collection/test-2025-11-03-13-43', 'lerobot-data-collection/test-2025-11-03-13-44', 'lerobot-data-collection/test-2025-11-03-13-45', 'lerobot-data-collection/test-2025-11-03-13-46', 'lerobot-data-collection/test-2025-11-03-13-47', 'lerobot-data-collection/test-2025-11-03-13-48', 'lerobot-data-collection/test-2025-11-03-13-49']"
+
+# RUN loop_dataset.py to get your repo_ids
+
+# ========================================================= Two folds datasets
+
+#merge
+python -m lerobot.scripts.lerobot_edit_dataset \
+--repo_id lerobot-data-collection/two-folds-dataset-full-11-04 \
+--operation.type merge --push_to_hub true \
+--operation.repo_ids "['lerobot-data-collection/two-folds-dataset-2025-11-04-15-06', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-08', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-10', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-11', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-12', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-14', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-16', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-18', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-20', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-22', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-24', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-25', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-27', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-28', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-29', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-33', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-34', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-35', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-36', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-52', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-53', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-54', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-55', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-56', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-57', 'lerobot-data-collection/two-folds-dataset-2025-11-04-15-59', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-00', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-01', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-02', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-03', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-04', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-05', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-06', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-07', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-08', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-09', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-26', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-28', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-29', 'lerobot-data-collection/two-folds-dataset-2025-11-04-16-30']"

src/lerobot/scripts/lerobot_setup_motors.py

@@ -44,6 +44,7 @@ from lerobot.teleoperators import (  # noqa: F401
     omx_leader,
     so100_leader,
     so101_leader,
+    openarms_mini
 )
 
 COMPATIBLE_DEVICES = [
@@ -56,6 +57,7 @@ COMPATIBLE_DEVICES = [
     "so101_follower",
     "so101_leader",
     "lekiwi",
+    "openarms_mini",
 ]
 
 

src/lerobot/teleoperators/openarms/__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_openarms_leader import OpenArmsLeaderConfig
+from .openarms_leader import OpenArmsLeader
+
+__all__ = ["OpenArmsLeader", "OpenArmsLeaderConfig"]

src/lerobot/teleoperators/openarms/config_openarms_leader.py

@@ -0,0 +1,80 @@
+#!/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 typing import Dict
+
+from ..config import TeleoperatorConfig
+
+
+@TeleoperatorConfig.register_subclass("openarms_leader")
+@dataclass
+class OpenArmsLeaderConfig(TeleoperatorConfig):
+    """Configuration for the OpenArms leader/teleoperator with Damiao motors."""
+    
+    # CAN interfaces - one per arm
+    # Right arm CAN interface (e.g., "can2")
+    # Left arm CAN interface (e.g., "can3")
+    # Linux: "can0", "can1", etc.
+    # macOS: "/dev/cu.usbmodem*" (serial device)
+    port_right: str = "can2"    # CAN interface for right arm
+    port_left: str = "can3"     # CAN interface for left arm
+    
+    # CAN interface type: "socketcan" (Linux), "slcan" (macOS/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
+    
+    # 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])
+    
+    # Damping gains for stability when applying torque compensation (gravity/friction)
+    # Used when kp=0 and only torque is applied
+    damping_kd: list[float] = field(default_factory=lambda: [0.5, 0.5, 0.5, 0.5, 0.1, 0.1, 0.1, 0.1])
+    
+    # Friction model parameters: τ_fric(ω) = Fo + Fv·ω + Fc·tanh(k·ω)
+    # From OpenArms config/leader.yaml (note: Fc[5] is slightly different: 0.083 vs 0.093)
+    friction_fc: list[float] = field(default_factory=lambda: [0.306, 0.306, 0.40, 0.166, 0.050, 0.083, 0.172, 0.0512])  # Coulomb friction [Nm]
+    friction_k: list[float] = field(default_factory=lambda: [28.417, 28.417, 29.065, 130.038, 151.771, 242.287, 7.888, 4.000])  # tanh steepness
+    friction_fv: list[float] = field(default_factory=lambda: [0.063, 0.0630, 0.604, 0.813, 0.029, 0.072, 0.084, 0.084])  # Viscous friction [Nm·s/rad]
+    friction_fo: list[float] = field(default_factory=lambda: [0.088, 0.088, 0.008, -0.058, 0.005, 0.009, -0.059, -0.050])  # Offset torque [Nm]

src/lerobot/teleoperators/openarms/openarms_leader.py

@@ -0,0 +1,503 @@
+#!/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, Dict
+
+import numpy as np
+import pinocchio as pin
+
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode
+from lerobot.motors.damiao import DamiaoMotorsBus
+from lerobot.motors.damiao.tables import MotorType
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..teleoperator import Teleoperator
+from .config_openarms_leader import OpenArmsLeaderConfig
+
+logger = logging.getLogger(__name__)
+
+
+class OpenArmsLeader(Teleoperator):
+    """
+    OpenArms 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 = OpenArmsLeaderConfig
+    name = "openarms_leader"
+
+    def __init__(self, config: OpenArmsLeaderConfig):
+        super().__init__(config)
+        self.config = config
+        
+        norm_mode_body = MotorNormMode.DEGREES  # Always use degrees for Damiao motors
+        
+        # Right arm motors (on port_right)
+        # Each arm uses the same CAN IDs since they're on separate buses
+        motors_right = {}
+        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
+            motor = Motor(send_id, motor_type_str, norm_mode_body)
+            motor.recv_id = recv_id
+            motor.motor_type = getattr(MotorType, motor_type_str.upper().replace("-", "_"))
+            motors_right[motor_name] = motor
+        
+        # Left arm motors (on port_left, same IDs as right since separate bus)
+        motors_left = {}
+        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
+            motor = Motor(send_id, motor_type_str, norm_mode_body)
+            motor.recv_id = recv_id
+            motor.motor_type = getattr(MotorType, motor_type_str.upper().replace("-", "_"))
+            motors_left[motor_name] = motor
+        
+        # Initialize separate Damiao motors buses (one per arm) with CAN FD support
+        self.bus_right = DamiaoMotorsBus(
+            port=self.config.port_right,
+            motors=motors_right,
+            calibration={k.replace("right_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("right_")},
+            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,
+        )
+        
+        self.bus_left = DamiaoMotorsBus(
+            port=self.config.port_left,
+            motors=motors_left,
+            calibration={k.replace("left_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("left_")},
+            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,
+        )
+
+        # Initialize Pinocchio robot model for dynamics (optional)
+        self.pin_robot = None
+        try:
+            # Load URDF - try external path first (with meshes), then repository
+            import os
+            from os.path import expanduser, dirname
+            
+            # Try external URDF with meshes first
+            external_urdf_path = expanduser("~/Documents/openarm_description/openarm_bimanual_pybullet.urdf")
+            if os.path.exists(external_urdf_path):
+                urdf_path = external_urdf_path
+                urdf_dir = dirname(urdf_path)
+            
+                self.pin_robot = pin.RobotWrapper.BuildFromURDF(urdf_path, urdf_dir)
+                self.pin_robot.data = self.pin_robot.model.createData()
+                logger.info(f"Loaded OpenArms URDF for dynamics computation from {urdf_path}")
+        except Exception as e:
+            logger.warning(f"Could not load URDF for dynamics: {e}. Gravity compensation will not be available.")
+
+    @property
+    def action_features(self) -> Dict[str, type]:
+        """Features produced by this teleoperator."""
+        features = {}
+        # Right arm motors - only positions stored in dataset
+        for motor in self.bus_right.motors:
+            features[f"right_{motor}.pos"] = float
+        # Left arm motors - only positions stored in dataset
+        for motor in self.bus_left.motors:
+            features[f"left_{motor}.pos"] = 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_right.is_connected and self.bus_left.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 buses
+        logger.info(f"Connecting right arm on {self.config.port_right}...")
+        self.bus_right.connect()
+        logger.info(f"Connecting left arm on {self.config.port_left}...")
+        self.bus_left.connect()
+        
+        # Run calibration if needed
+        if calibrate:
+            logger.info(
+                "No calibration found or overwriting calibration. Running calibration..."
+            )
+            self.calibrate()
+        
+        # Configure for manual control
+        self.configure()
+        
+        logger.info(f"{self} connected.")
+
+    @property
+    def is_calibrated(self) -> bool:
+        """Check if teleoperator is calibrated."""
+        return self.bus_right.is_calibrated and self.bus_left.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:
+            # Ask user whether to use existing calibration
+            user_input = input(
+                f"Press ENTER to use existing calibration for {self.id}, "
+                f"or type 'c' and press ENTER to run new calibration: "
+            )
+            if user_input.strip().lower() != "c":
+                logger.info(f"Using existing calibration for {self.id}")
+                # Split calibration for each bus
+                cal_right = {k.replace("right_", ""): v for k, v in self.calibration.items() if k.startswith("right_")}
+                cal_left = {k.replace("left_", ""): v for k, v in self.calibration.items() if k.startswith("left_")}
+                self.bus_right.write_calibration(cal_right)
+                self.bus_left.write_calibration(cal_left)
+                return
+        
+        logger.info(f"\nRunning calibration for {self}")
+        
+        # Calibrate each arm separately
+        self._calibrate_arm("right", self.bus_right)
+        self._calibrate_arm("left", self.bus_left)
+        
+        print(f"\nCalibration complete and saved to {self.calibration_fpath}")
+    
+    def _calibrate_arm(self, arm_name: str, bus: DamiaoMotorsBus) -> None:
+        """Calibrate a single arm."""
+        logger.info(f"\n=== Calibrating {arm_name.upper()} arm ===")
+        
+        # Ensure torque is disabled for manual positioning
+        bus.disable_torque()
+        time.sleep(0.1)
+        
+        # Step 1: Set zero position
+        input(
+            f"\nCalibration: Zero Position ({arm_name.upper()} arm)\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
+        bus.set_zero_position()
+        logger.info(f"{arm_name.capitalize()} arm zero position set.")
+        
+        # Automatically set range to -90° to +90° for all joints
+        print(
+            f"\nAutomatically setting range: -90° to +90° for all joints"
+        )
+        
+        # Create calibration data with fixed ranges
+        if self.calibration is None:
+            self.calibration = {}
+            
+        for motor_name, motor in bus.motors.items():
+            # Prefix motor name with arm name for storage
+            prefixed_name = f"{arm_name}_{motor_name}"
+            
+            # Use -90 to +90 for all joints and gripper (integers required)
+            self.calibration[prefixed_name] = MotorCalibration(
+                id=motor.id,
+                drive_mode=0,  # Normal direction
+                homing_offset=0,  # Already set via set_zero_position
+                range_min=-90,  # -90 degrees (integer)
+                range_max=90,   # +90 degrees (integer)
+            )
+            logger.info(f"  {prefixed_name}: range set to [-90°, +90°]")
+        
+        # Write calibration to this arm's motors
+        cal_for_bus = {k.replace(f"{arm_name}_", ""): v for k, v in self.calibration.items() if k.startswith(f"{arm_name}_")}
+        bus.write_calibration(cal_for_bus)
+        
+        # Save calibration after each arm
+        self._save_calibration()
+
+    def configure(self) -> None:
+        """
+        Configure motors for manual teleoperation.
+        
+        For manual control, we disable torque so the arm can be moved by hand.
+        """
+        if self.config.manual_control:
+            # Disable torque for manual control
+            logger.info("Disabling torque for manual control...")
+            self.bus_right.disable_torque()
+            self.bus_left.disable_torque()
+        else:
+            # Configure motors normally
+            self.bus_right.configure_motors()
+            self.bus_left.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) -> Dict[str, Any]:
+        """
+        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.
+        Note: Velocity and torque are read but not stored in dataset (only used for
+        gravity/friction compensation during recording).
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        action_dict = {}
+        start = time.perf_counter()
+        
+        # OPTIMIZED: Use sync_read_all_states to get pos/vel/torque in one go
+        right_states = self.bus_right.sync_read_all_states()
+        for motor in self.bus_right.motors:
+            state = right_states.get(motor, {})
+            action_dict[f"right_{motor}.pos"] = state.get("position", 0.0)
+            action_dict[f"right_{motor}.vel"] = state.get("velocity", 0.0)
+            action_dict[f"right_{motor}.torque"] = state.get("torque", 0.0)
+        
+        # OPTIMIZED: Use sync_read_all_states to get pos/vel/torque in one go
+        left_states = self.bus_left.sync_read_all_states()
+        for motor in self.bus_left.motors:
+            state = left_states.get(motor, {})
+            action_dict[f"left_{motor}.pos"] = state.get("position", 0.0)
+            action_dict[f"left_{motor}.vel"] = state.get("velocity", 0.0)
+            action_dict[f"left_{motor}.torque"] = state.get("torque", 0.0)
+        
+        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 OpenArms leader.")
+
+    def disconnect(self) -> None:
+        """Disconnect from teleoperator."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # For manual control, ensure torque is disabled before disconnecting
+        if self.config.manual_control:
+            try:
+                self.bus_right.disable_torque()
+                self.bus_left.disable_torque()
+            except Exception as e:
+                logger.warning(f"Failed to disable torque during disconnect: {e}")
+        
+        # Disconnect from CAN buses
+        self.bus_right.disconnect(disable_torque=False)  # Already disabled above if needed
+        self.bus_left.disconnect(disable_torque=False)
+        
+        logger.info(f"{self} disconnected.")
+
+    def _deg_to_rad(self, deg: Dict[str, float | int]) -> Dict[str, float]:
+        """Convert degrees to radians for all motors."""
+        return {m: np.deg2rad(float(v)) for m, v in deg.items()}
+    
+    def _gravity_from_q(self, q_rad: Dict[str, float]) -> Dict[str, float]:
+        """
+        Compute g(q) [N·m] for all joints in the robot.
+        The order of joints in the URDF matches the concatenated motor lists (right then left).
+        
+        Args:
+            q_rad: Dictionary mapping motor names (with arm prefix) to positions in radians
+            
+        Returns:
+            Dictionary mapping motor names to gravity torques in N·m
+            
+        Raises:
+            RuntimeError: If URDF model is not loaded
+        """
+        if self.pin_robot is None:
+            raise RuntimeError(
+                "Cannot compute gravity: URDF model not loaded. "
+                "Ensure urdf/openarms.urdf exists and is valid."
+            )
+        
+        # Build position vector in the order of motors (left arm, then right arm)
+        # This order must match the URDF joint order
+        # URDF has: left_joint1-7, left_finger_joint1-2, right_joint1-7, right_finger_joint1-2
+        q = np.zeros(self.pin_robot.model.nq)
+        idx = 0
+        
+        # Left arm motors (first in URDF) - joints 1-7
+        for motor_name in self.bus_left.motors:
+            if motor_name == "gripper":
+                continue  # Skip gripper, will be handled separately
+            full_name = f"left_{motor_name}"
+            q[idx] = q_rad.get(full_name, 0.0)
+            idx += 1
+        
+        # Skip left finger joints (leave as zeros)
+        idx += 2
+        
+        # Right arm motors (second in URDF) - joints 1-7
+        for motor_name in self.bus_right.motors:
+            if motor_name == "gripper":
+                continue  # Skip gripper, will be handled separately
+            full_name = f"right_{motor_name}"
+            q[idx] = q_rad.get(full_name, 0.0)
+            idx += 1
+        
+        # Skip right finger joints (leave as zeros)
+        idx += 2
+        
+        # Compute generalized gravity vector
+        g = pin.computeGeneralizedGravity(self.pin_robot.model, self.pin_robot.data, q)
+        
+        # Map back to motor names (only arm joints, not fingers)
+        result = {}
+        idx = 0
+        
+        # Left arm torques (joints 1-7)
+        for motor_name in self.bus_left.motors:
+            if motor_name == "gripper":
+                result["left_gripper"] = 0.0  # No gravity compensation for gripper
+                continue
+            result[f"left_{motor_name}"] = float(g[idx])
+            idx += 1
+        
+        # Skip left finger joint torques in output
+        idx += 2
+        
+        # Right arm torques (joints 1-7)
+        for motor_name in self.bus_right.motors:
+            if motor_name == "gripper":
+                result["right_gripper"] = 0.0  # No gravity compensation for gripper
+                continue
+            result[f"right_{motor_name}"] = float(g[idx])
+            idx += 1
+        
+        # Skip right finger joint torques in output
+        idx += 2
+        
+        return result
+    
+    def _friction_from_velocity(
+        self, 
+        velocity_rad_per_sec: Dict[str, float],
+        friction_scale: float = 1.0,
+        amp_tmp: float = 1.0,
+        coef_tmp: float = 0.1
+    ) -> Dict[str, float]:
+        """
+        Compute friction torques for all joints in the robot using tanh friction model.
+        
+        Args:
+            velocity_rad_per_sec: Dictionary mapping motor names (with arm prefix) to velocities in rad/s
+            friction_scale: Scale factor for friction compensation (default 1.0, use 0.3 for stability)
+            amp_tmp: Amplitude factor for tanh term (default 1.0)
+            coef_tmp: Coefficient for tanh steepness (default 0.1)
+            
+        Returns:
+            Dictionary mapping motor names to friction torques in N·m
+        """
+        # Motor name to index mapping
+        motor_name_to_index = {
+            "joint_1": 0,
+            "joint_2": 1,
+            "joint_3": 2,
+            "joint_4": 3,
+            "joint_5": 4,
+            "joint_6": 5,
+            "joint_7": 6,
+            "gripper": 7,
+        }
+        
+        result = {}
+        
+        # Process all motors (left and right)
+        for motor_full_name, velocity in velocity_rad_per_sec.items():
+            # Extract motor name without arm prefix
+            if motor_full_name.startswith("right_"):
+                motor_name = motor_full_name.removeprefix("right_")
+            elif motor_full_name.startswith("left_"):
+                motor_name = motor_full_name.removeprefix("left_")
+            else:
+                result[motor_full_name] = 0.0
+                continue
+            
+            # Get motor index for friction parameters
+            motor_index = motor_name_to_index.get(motor_name, 0)
+            
+            # Get friction parameters from config
+            Fc = self.config.friction_fc[motor_index]
+            k = self.config.friction_k[motor_index]
+            Fv = self.config.friction_fv[motor_index]
+            Fo = self.config.friction_fo[motor_index]
+            
+            # Friction model: τ_fric = amp * Fc * tanh(coef * k * ω) + Fv * ω + Fo
+            friction_torque = (
+                amp_tmp * Fc * np.tanh(coef_tmp * k * velocity) +
+                Fv * velocity +
+                Fo
+            )
+            
+            # Apply scale factor
+            friction_torque *= friction_scale
+            
+            result[motor_full_name] = float(friction_torque)
+        
+        return result
+    
+    def get_damping_kd(self, motor_name: str) -> float:
+        """
+        Get damping gain (Kd) for a specific motor.
+        
+        Args:
+            motor_name: Motor name without arm prefix (e.g., "joint_1", "gripper")
+            
+        Returns:
+            Damping gain value
+        """
+        motor_name_to_index = {
+            "joint_1": 0,
+            "joint_2": 1,
+            "joint_3": 2,
+            "joint_4": 3,
+            "joint_5": 4,
+            "joint_6": 5,
+            "joint_7": 6,
+            "gripper": 7,
+        }
+        
+        motor_index = motor_name_to_index.get(motor_name, 0)
+        return self.config.damping_kd[motor_index]
+
+

src/lerobot/teleoperators/openarms_mini/__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_openarms_mini import OpenArmsMiniConfig
+from .openarms_mini import OpenArmsMini
+
+__all__ = ["OpenArmsMini", "OpenArmsMiniConfig"]
+

src/lerobot/teleoperators/openarms_mini/config_openarms_mini.py

@@ -0,0 +1,33 @@
+#!/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
+
+from ..teleoperator import TeleoperatorConfig
+
+
+@TeleoperatorConfig.register_subclass("openarms_mini")
+@dataclass
+class OpenArmsMiniConfig(TeleoperatorConfig):
+    """Configuration for OpenArms Mini teleoperator with Feetech motors (dual arms)."""
+    
+    # Serial ports for left and right arms
+    port_right: str = "/dev/ttyUSB0"  # Serial port for right arm
+    port_left: str = "/dev/ttyUSB1"   # Serial port for left arm
+    
+    # Whether to use degrees mode (True) or normalized mode (False)
+    use_degrees: bool = True
+

src/lerobot/teleoperators/openarms_mini/openarms_mini.py

@@ -0,0 +1,485 @@
+#!/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 lerobot.motors import Motor, MotorCalibration, MotorNormMode
+from lerobot.motors.feetech import (
+    FeetechMotorsBus,
+    OperatingMode,
+)
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..teleoperator import Teleoperator
+from .config_openarms_mini import OpenArmsMiniConfig
+
+logger = logging.getLogger(__name__)
+
+
+class OpenArmsMini(Teleoperator):
+    """
+    OpenArms Mini Teleoperator with dual Feetech-based arms (8 motors per arm).
+    Each arm has 7 joints plus a gripper, using the same DOF as OpenArms.
+    """
+
+    config_class = OpenArmsMiniConfig
+    name = "openarms_mini"
+
+    def __init__(self, config: OpenArmsMiniConfig):
+        super().__init__(config)
+        self.config = config
+        
+        # Use degrees mode for all motors
+        norm_mode_body = MotorNormMode.DEGREES
+        
+        # Right arm motors (8 motors: joint_1 to joint_7 + gripper)
+        motors_right = {
+            "joint_1": Motor(1, "sts3215", norm_mode_body),
+            "joint_2": Motor(2, "sts3215", norm_mode_body),
+            "joint_3": Motor(3, "sts3215", norm_mode_body),
+            "joint_4": Motor(4, "sts3215", norm_mode_body),
+            "joint_5": Motor(5, "sts3215", norm_mode_body),
+            "joint_6": Motor(6, "sts3215", norm_mode_body),
+            "joint_7": Motor(7, "sts3215", norm_mode_body),
+            "gripper": Motor(8, "sts3215", MotorNormMode.RANGE_0_100),
+        }
+        
+        # Left arm motors (8 motors: joint_1 to joint_7 + gripper)
+        # Note: Left arm uses IDs 11-18 to avoid conflicts if on same bus
+        motors_left = {
+            "joint_1": Motor(1, "sts3215", norm_mode_body),
+            "joint_2": Motor(2, "sts3215", norm_mode_body),
+            "joint_3": Motor(3, "sts3215", norm_mode_body),
+            "joint_4": Motor(4, "sts3215", norm_mode_body),
+            "joint_5": Motor(5, "sts3215", norm_mode_body),
+            "joint_6": Motor(6, "sts3215", norm_mode_body),
+            "joint_7": Motor(7, "sts3215", norm_mode_body),
+            "gripper": Motor(8, "sts3215", MotorNormMode.RANGE_0_100),
+        }
+        
+        # Initialize Feetech motor buses for each arm
+        self.bus_right = FeetechMotorsBus(
+            port=self.config.port_right,
+            motors=motors_right,
+            calibration={k.replace("right_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("right_")},
+        )
+        
+        self.bus_left = FeetechMotorsBus(
+            port=self.config.port_left,
+            motors=motors_left,
+            calibration={k.replace("left_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("left_")},
+        )
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        """Action features include positions for all motors (16 total: 8 per arm)."""
+        features = {}
+        # Right arm motors
+        for motor in self.bus_right.motors:
+            features[f"right_{motor}.pos"] = float
+        # Left arm motors
+        for motor in self.bus_left.motors:
+            features[f"left_{motor}.pos"] = float
+        return features
+
+    @property
+    def feedback_features(self) -> dict[str, type]:
+        """No feedback features for now."""
+        return {}
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if both arms are connected."""
+        return self.bus_right.is_connected and self.bus_left.is_connected
+
+    def connect(self, calibrate: bool = True) -> None:
+        """Connect to both arms and optionally calibrate."""
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} already connected")
+
+        # Connect to both buses
+        logger.info(f"Connecting right arm on {self.config.port_right}...")
+        self.bus_right.connect()
+        logger.info(f"Connecting left arm on {self.config.port_left}...")
+        self.bus_left.connect()
+        
+        # Calibrate if requested (always prompt user)
+        if calibrate:
+            self.calibrate()
+
+        # Configure motors
+        self.configure()
+        logger.info(f"{self} connected.")
+
+    @property
+    def is_calibrated(self) -> bool:
+        """Check if both arms are calibrated."""
+        return self.bus_right.is_calibrated and self.bus_left.is_calibrated
+
+    def calibrate(self) -> None:
+        """
+        Run calibration procedure for OpenArms Mini.
+        
+        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. Set fixed range of -90° to +90° for all joints (0-100 for gripper)
+        5. Save calibration
+        """
+        if self.calibration:
+            # Ask user whether to use existing calibration
+            user_input = input(
+                f"Press ENTER to use existing calibration for {self.id}, "
+                f"or type 'c' and press ENTER to run new calibration: "
+            )
+            if user_input.strip().lower() != "c":
+                logger.info(f"Using existing calibration for {self.id}")
+                # Split calibration for each bus
+                cal_right = {k.replace("right_", ""): v for k, v in self.calibration.items() if k.startswith("right_")}
+                cal_left = {k.replace("left_", ""): v for k, v in self.calibration.items() if k.startswith("left_")}
+                self.bus_right.write_calibration(cal_right)
+                self.bus_left.write_calibration(cal_left)
+                return
+
+        logger.info(f"\nRunning calibration for {self}")
+        
+        # Calibrate each arm separately
+        self._calibrate_arm("right", self.bus_right)
+        self._calibrate_arm("left", self.bus_left)
+        
+        self._save_calibration()
+        print(f"\nCalibration complete and saved to {self.calibration_fpath}")
+
+    def _calibrate_arm(self, arm_name: str, bus: FeetechMotorsBus) -> None:
+        """Calibrate a single arm with Feetech motors."""
+        logger.info(f"\n=== Calibrating {arm_name.upper()} arm ===")
+        
+        # Disable torque for manual positioning
+        bus.disable_torque()
+        
+        # Set Phase to 12 for all motors
+        logger.info(f"Setting Phase to 12 for all motors in {arm_name.upper()} arm...")
+        for motor in bus.motors:
+            bus.write("Phase", motor, 12)
+        logger.info(f"Phase set to 12 for all motors in {arm_name.upper()} arm")
+        
+        # Set all motors to position mode
+        for motor in bus.motors:
+            bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
+        
+        # Step 1: Set zero position
+        input(
+            f"\nCalibration: Zero Position ({arm_name.upper()} arm)\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 (half-turn homing)
+        homing_offsets = bus.set_half_turn_homings()
+        logger.info(f"{arm_name.capitalize()} arm zero position set.")
+        
+        # Step 2: Set ranges for joints and gripper
+        print(f"\nSetting motor ranges for {arm_name.upper()} arm\n")
+        
+        # Create calibration data with full motor ranges
+        if self.calibration is None:
+            self.calibration = {}
+        
+        # Get motor resolution
+        motor_resolution = bus.model_resolution_table[list(bus.motors.values())[0].model]
+        max_res = motor_resolution - 1
+        
+        for motor_name, motor in bus.motors.items():
+            # Prefix motor name with arm name for storage
+            prefixed_name = f"{arm_name}_{motor_name}"
+            
+            if motor_name == "gripper":
+                # Interactive calibration for gripper
+                input(
+                    f"\nGripper Calibration ({arm_name.upper()} arm)\n"
+                    f"Step 1: CLOSE the gripper fully\n"
+                    f"Press ENTER when gripper is closed..."
+                )
+                closed_pos = bus.read("Present_Position", motor_name, normalize=False)
+                logger.info(f"  Gripper closed position recorded: {closed_pos}")
+                
+                input(
+                    f"\nStep 2: OPEN the gripper fully\n"
+                    f"Press ENTER when gripper is fully open..."
+                )
+                open_pos = bus.read("Present_Position", motor_name, normalize=False)
+                logger.info(f"  Gripper open position recorded: {open_pos}")
+                
+                # For RANGE_0_100: range_min maps to 0 (closed), range_max maps to 100 (open)
+                # If gripper motor direction is reversed (closed > open), we need to swap and use drive_mode=1
+                if closed_pos < open_pos:
+                    # Normal direction: 0=closed, 100=open
+                    range_min = int(closed_pos)
+                    range_max = int(open_pos)
+                    drive_mode = 0
+                else:
+                    # Reversed direction: swap so min < max, and set drive_mode=1 to reverse
+                    range_min = int(open_pos)
+                    range_max = int(closed_pos)
+                    drive_mode = 1
+                
+                logger.info(f"  {prefixed_name}: range set to [{range_min}, {range_max}] (0=closed, 100=open, drive_mode={drive_mode})")
+            else:
+                # Use full motor range for joint motors (will use degrees normalization)
+                range_min = 0
+                range_max = max_res
+                drive_mode = 0  # Normal direction for non-gripper motors
+                logger.info(f"  {prefixed_name}: range set to [0, {max_res}] (full motor range)")
+            
+            self.calibration[prefixed_name] = MotorCalibration(
+                id=motor.id,
+                drive_mode=drive_mode,
+                homing_offset=homing_offsets[motor_name],
+                range_min=range_min,
+                range_max=range_max,
+            )
+        
+        # Write calibration to this arm's motors
+        cal_for_bus = {k.replace(f"{arm_name}_", ""): v for k, v in self.calibration.items() if k.startswith(f"{arm_name}_")}
+        bus.write_calibration(cal_for_bus)
+
+    def configure(self) -> None:
+        """Configure motors with appropriate settings."""
+        # Configure right arm
+        self.bus_right.disable_torque()
+        self.bus_right.configure_motors()
+        for motor in self.bus_right.motors:
+            self.bus_right.write("Operating_Mode", motor, OperatingMode.POSITION.value)
+        
+        # Configure left arm
+        self.bus_left.disable_torque()
+        self.bus_left.configure_motors()
+        for motor in self.bus_left.motors:
+            self.bus_left.write("Operating_Mode", motor, OperatingMode.POSITION.value)
+
+    def setup_motors(self) -> None:
+        """Setup motor IDs for both arms."""
+        print("\nSetting up RIGHT arm motors...")
+        for motor in reversed(self.bus_right.motors):
+            input(f"Connect the controller board to the RIGHT '{motor}' motor only and press enter.")
+            self.bus_right.setup_motor(motor)
+            print(f"RIGHT '{motor}' motor id set to {self.bus_right.motors[motor].id}")
+        
+        print("\nSetting up LEFT arm motors...")
+        for motor in reversed(self.bus_left.motors):
+            input(f"Connect the controller board to the LEFT '{motor}' motor only and press enter.")
+            self.bus_left.setup_motor(motor)
+            print(f"LEFT '{motor}' motor id set to {self.bus_left.motors[motor].id}")
+
+    def get_action(self) -> dict[str, float]:
+        """Get current action from both arms (read positions from all motors)."""
+        start = time.perf_counter()
+        
+        # Motors to flip (invert direction) - different for each arm
+        right_motors_to_flip = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_7"]
+        left_motors_to_flip = ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]
+        
+        # Joint mapping: leader joint 6 -> follower joint 7, leader joint 7 -> follower joint 6
+        joint_remap = {"joint_6": "joint_7", "joint_7": "joint_6"}
+        
+        # Read positions from both arms
+        right_positions = self.bus_right.sync_read("Present_Position")
+        left_positions = self.bus_left.sync_read("Present_Position")
+        
+        # Combine into single action dict with arm prefixes, flip and remap joints
+        action = {}
+        for motor, val in right_positions.items():
+            target_motor = joint_remap.get(motor, motor)
+            if motor in right_motors_to_flip:
+                action[f"right_{target_motor}.pos"] = -val
+            else:
+                action[f"right_{target_motor}.pos"] = val
+        for motor, val in left_positions.items():
+            target_motor = joint_remap.get(motor, motor)
+            if motor in left_motors_to_flip:
+                action[f"left_{target_motor}.pos"] = -val
+            else:
+                action[f"left_{target_motor}.pos"] = val
+        
+        dt_ms = (time.perf_counter() - start) * 1e3
+        logger.debug(f"{self} read action: {dt_ms:.1f}ms")
+        return action
+
+    def send_feedback(self, feedback: dict[str, float]) -> None:
+        """
+        Send position feedback to teleoperator motors.
+        
+        Args:
+            feedback: Dictionary with motor positions (e.g., "right_joint_1.pos", "left_joint_2.pos")
+        """
+        # Motors to flip (invert direction) -> matches get_action()
+        right_motors_to_flip = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_7"]
+        left_motors_to_flip = ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]
+        
+        # Reverse joint mapping: follower joint 7 -> leader joint 6, follower joint 6 -> leader joint 7
+        joint_remap = {"joint_7": "joint_6", "joint_6": "joint_7"}
+        
+        # Split feedback by arm and flip/remap motors as needed
+        right_positions = {}
+        left_positions = {}
+        
+        for key, val in feedback.items():
+            if key.endswith(".pos"):
+                motor_name = key.removesuffix(".pos")
+                if motor_name.startswith("right_"):
+                    base_name = motor_name.removeprefix("right_")
+                    # Convert gripper from robot degrees (-65 to 0) to teleop range (0 to 100)
+                    if "gripper" in base_name:
+                        val = val / -0.65
+                    target_motor = joint_remap.get(base_name, base_name)
+                    # Check flip on target_motor (leader joint) to match get_action() logic
+                    if target_motor in right_motors_to_flip:
+                        right_positions[target_motor] = -val
+                    else:
+                        right_positions[target_motor] = val
+                elif motor_name.startswith("left_"):
+                    base_name = motor_name.removeprefix("left_")
+                    # Convert gripper from robot degrees (-65 to 0) to teleop range (0 to 100)
+                    if "gripper" in base_name:
+                        val = val / -0.65
+                    target_motor = joint_remap.get(base_name, base_name)
+                    # Check flip on target_motor (leader joint) to match get_action() logic
+                    if target_motor in left_motors_to_flip:
+                        left_positions[target_motor] = -val
+                    else:
+                        left_positions[target_motor] = val
+        
+        # Write positions to both arms
+        if right_positions:
+            self.bus_right.sync_write("Goal_Position", right_positions)
+        if left_positions:
+            self.bus_left.sync_write("Goal_Position", left_positions)
+    
+    def enable_torque(self) -> None:
+        """Enable torque on both arms (for position tracking)."""
+        self.bus_right.enable_torque()
+        self.bus_left.enable_torque()
+    
+    def disable_torque(self) -> None:
+        """Disable torque on both arms (for free movement)."""
+        self.bus_right.disable_torque()
+        self.bus_left.disable_torque()
+    
+    def smooth_move_to(self, target: dict[str, float], duration_s: float = 1.0, fps: int = 50) -> None:
+        """
+        Smoothly move teleoperator to target position over specified duration.
+        Uses linear interpolation to avoid sudden jerky movements.
+        
+        Args:
+            target: Target positions dict (e.g., {"right_joint_1.pos": 30.0, ...})
+            duration_s: Duration of movement in seconds (default 1.0)
+            fps: Control frequency (default 50 Hz)
+        """
+        # Read current positions
+        current_right = self.bus_right.sync_read("Present_Position")
+        current_left = self.bus_left.sync_read("Present_Position")
+        
+        # Build current position dict with full names
+        current = {}
+        for motor, val in current_right.items():
+            current[f"right_{motor}"] = val
+        for motor, val in current_left.items():
+            current[f"left_{motor}"] = val
+        
+        # Reverse joint mapping for target (follower joint 7 -> leader joint 6)
+        joint_remap = {"joint_7": "joint_6", "joint_6": "joint_7"}
+        right_motors_to_flip = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_7"]
+        left_motors_to_flip = ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]
+        
+        # Parse target and apply remapping/flipping
+        target_parsed = {}
+        for key, val in target.items():
+            if key.endswith(".pos"):
+                motor_name = key.removesuffix(".pos")
+                if motor_name.startswith("right_"):
+                    base_name = motor_name.removeprefix("right_")
+                    # Convert gripper from robot degrees (-65 to 0) to teleop range (0 to 100)
+                    if "gripper" in base_name:
+                        val = val / -0.65
+                    target_motor = joint_remap.get(base_name, base_name)
+                    # Check flip on target_motor (leader joint) to match get_action() logic
+                    if target_motor in right_motors_to_flip:
+                        target_parsed[f"right_{target_motor}"] = -val
+                    else:
+                        target_parsed[f"right_{target_motor}"] = val
+                elif motor_name.startswith("left_"):
+                    base_name = motor_name.removeprefix("left_")
+                    # Convert gripper from robot degrees (-65 to 0) to teleop range (0 to 100)
+                    if "gripper" in base_name:
+                        val = val / -0.65
+                    target_motor = joint_remap.get(base_name, base_name)
+                    # Check flip on target_motor (leader joint) to match get_action() logic
+                    if target_motor in left_motors_to_flip:
+                        target_parsed[f"left_{target_motor}"] = -val
+                    else:
+                        target_parsed[f"left_{target_motor}"] = val
+        
+        # Enable torque for movement
+        self.enable_torque()
+        
+        # Interpolate over duration
+        num_steps = int(duration_s * fps)
+        step_duration = 1.0 / fps
+        
+        for step in range(num_steps + 1):
+            t = step / num_steps  # 0.0 to 1.0
+            
+            # Smooth easing (ease-in-out)
+            t_smooth = t * t * (3 - 2 * t)
+            
+            # Interpolate positions
+            right_positions = {}
+            left_positions = {}
+            
+            for motor_full, target_val in target_parsed.items():
+                current_val = current.get(motor_full, target_val)
+                interp_val = current_val + t_smooth * (target_val - current_val)
+                
+                if motor_full.startswith("right_"):
+                    motor = motor_full.removeprefix("right_")
+                    right_positions[motor] = interp_val
+                elif motor_full.startswith("left_"):
+                    motor = motor_full.removeprefix("left_")
+                    left_positions[motor] = interp_val
+            
+            # Send interpolated positions
+            if right_positions:
+                self.bus_right.sync_write("Goal_Position", right_positions)
+            if left_positions:
+                self.bus_left.sync_write("Goal_Position", left_positions)
+            
+            if step < num_steps:
+                time.sleep(step_duration)
+
+    def disconnect(self) -> None:
+        """Disconnect from both arms."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        # Disconnect both buses
+        self.bus_right.disconnect()
+        self.bus_left.disconnect()
+        logger.info(f"{self} disconnected.")
+

src/lerobot/teleoperators/utils.py

@@ -81,6 +81,10 @@ def make_teleoperator_from_config(config: TeleoperatorConfig) -> Teleoperator:
         from .reachy2_teleoperator import Reachy2Teleoperator
 
         return Reachy2Teleoperator(config)
+    elif config.type == "openarms_mini":
+        from .openarms_mini import OpenArmsMini
+
+        return OpenArmsMini(config)
     else:
         try:
             return cast(Teleoperator, make_device_from_device_class(config))

tests/conftest.py

@@ -32,6 +32,39 @@ pytest_plugins = [
 ]
 
 
+def pytest_addoption(parser):
+    """Add custom command line option for hardware tests."""
+    parser.addoption(
+        "--run-hardware",
+        action="store_true",
+        default=False,
+        help="Run hardware tests that require actual motors connected",
+    )
+    parser.addoption(
+        "--can-port",
+        action="store",
+        default=None,
+        help="CAN interface port (e.g., 'can0' for Linux, '/dev/cu.usbmodem*' for macOS)",
+    )
+
+
+def pytest_configure(config):
+    """Register custom marker for hardware tests."""
+    config.addinivalue_line("markers", "hardware: mark test as requiring hardware")
+
+
+def pytest_collection_modifyitems(config, items):
+    """Skip hardware tests unless --run-hardware flag is provided."""
+    if config.getoption("--run-hardware"):
+        # --run-hardware given in cli: do not skip hardware tests
+        return
+    
+    skip_hardware = pytest.mark.skip(reason="need --run-hardware option to run")
+    for item in items:
+        if "hardware" in item.keywords:
+            item.add_marker(skip_hardware)
+
+
 def pytest_collection_finish():
     print(f"\nTesting with {DEVICE=}")
 

tests/motors/test_damiao.py

@@ -0,0 +1,338 @@
+#!/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.
+
+"""
+Test script for Damiao motor communication and control.
+
+This script tests basic functionality of a single Damiao motor via CAN bus:
+1. Connects to CAN interface
+2. Discovers and enables the motor
+3. Reads current position
+4. Sets zero position
+5. Writes target positions
+6. Disables torque
+
+Requirements:
+- Motor must be connected and powered (24V)
+- CAN interface must be configured (e.g., can0)
+- Motor ID must be set to 0x01 (send) and 0x11 (receive)
+
+Setup CAN interface:
+    sudo ip link set can0 type can bitrate 1000000
+    sudo ip link set can0 up
+    
+Verify connection:
+    candump can0  # In another terminal
+    cansend can0 001#FFFFFFFFFFFFFFFC  # Should enable motor and LED turns green
+"""
+
+import time
+
+import pytest
+
+from lerobot.motors import Motor, MotorNormMode
+from lerobot.motors.damiao import DamiaoMotorsBus
+from lerobot.motors.damiao.tables import MotorType
+
+@pytest.fixture
+def can_port(request):
+    """Get CAN port from command line or raise error if not provided."""
+    port = request.config.getoption("--can-port")
+    if port is None:
+        pytest.skip("CAN port not specified. Use --can-port to specify the CAN interface.")
+    return port
+
+
+@pytest.mark.hardware
+def test_single_motor_basic_operations(can_port):
+    """
+    Test basic operations with a single Damiao motor.
+    
+    This test requires actual hardware and is skipped by default.
+    To run with hardware, use: pytest tests/motors/test_damiao.py --run-hardware --can-port PORT
+    """
+    
+    # Configuration
+    MOTOR_ID = 0x01  # Sender CAN ID
+    MOTOR_RECV_ID = 0x11  # Receiver/Master ID
+    MOTOR_TYPE = "dm4310"
+    MOTOR_NAME = "test_motor"
+    
+    print(f"\n{'='*60}")
+    print("Damiao Motor Test - Single Motor Basic Operations")
+    print(f"{'='*60}\n")
+    
+    # Step 1: Create motor configuration
+    print(f"Step 1: Creating motor configuration...")
+    print(f"  - Motor ID: 0x{MOTOR_ID:02X} (send) / 0x{MOTOR_RECV_ID:02X} (recv)")
+    print(f"  - Motor Type: {MOTOR_TYPE}")
+    print(f"  - CAN Port: {can_port}")
+    
+    motor = Motor(MOTOR_ID, MOTOR_TYPE, MotorNormMode.DEGREES)
+    motor.recv_id = MOTOR_RECV_ID
+    motor.motor_type = MotorType.DM4310
+    
+    motors = {MOTOR_NAME: motor}
+    
+    # Step 2: Connect to CAN bus
+    print(f"\nStep 2: Connecting to CAN bus...")
+    bus = DamiaoMotorsBus(port=can_port, motors=motors)
+    
+    try:
+        bus.connect(handshake=True)
+        print(f"  ✓ Connected to {can_port}")
+    except Exception as e:
+        print(f"  ✗ Failed to connect: {e}")
+        print("\nTroubleshooting:")
+        print(f"  1. Check CAN interface is up: ip link show {can_port}")
+        print(f"  2. Setup if needed: sudo ip link set {can_port} type can bitrate 1000000")
+        print(f"  3. Bring up: sudo ip link set {can_port} up")
+        print(f"  4. Test with: cansend {can_port} 001#FFFFFFFFFFFFFFFC")
+        return
+    
+    try:
+        # Step 3: Enable motor (torque on)
+        print(f"\nStep 3: Enabling motor...")
+        bus.enable_torque(MOTOR_NAME)
+        time.sleep(0.1)
+        print(f"  ✓ Motor enabled (LED should be green)")
+        
+        # Step 4: Read current position
+        print(f"\nStep 4: Reading current position...")
+        current_pos = bus.read("Present_Position", MOTOR_NAME, normalize=False)
+        current_vel = bus.read("Present_Velocity", MOTOR_NAME, normalize=False)
+        current_torque = bus.read("Present_Torque", MOTOR_NAME, normalize=False)
+        
+        print(f"  Current State:")
+        print(f"    Position: {current_pos:8.2f}°")
+        print(f"    Velocity: {current_vel:8.2f}°/s")
+        print(f"    Torque:   {current_torque:8.3f} N·m")
+        
+        # Step 5: Set zero position
+        print(f"\nStep 5: Setting current position as zero...")
+        bus.set_zero_position([MOTOR_NAME])
+        time.sleep(0.2)
+        
+        new_pos = bus.read("Present_Position", MOTOR_NAME, normalize=False)
+        print(f"  Position after zero: {new_pos:8.2f}°")
+        print(f"  ✓ Zero position set")
+        
+        # Step 6: Test position commands
+        print(f"\nStep 6: Testing position control...")
+        
+        test_positions = [0.0, 45.0, -45.0, 0.0]
+        
+        for target_pos in test_positions:
+            print(f"\n  Moving to {target_pos:6.1f}°...")
+            bus.write("Goal_Position", MOTOR_NAME, target_pos, normalize=False)
+            time.sleep(1.0)  # Allow motor to move
+            
+            actual_pos = bus.read("Present_Position", MOTOR_NAME, normalize=False)
+            error = abs(actual_pos - target_pos)
+            
+            print(f"    Target:   {target_pos:8.2f}°")
+            print(f"    Actual:   {actual_pos:8.2f}°")
+            print(f"    Error:    {error:8.2f}°")
+            
+            if error > 10.0:
+                print(f"    ⚠ Large position error!")
+            else:
+                print(f"    ✓ Position reached")
+        
+        # Step 7: Test MIT control with custom gains
+        print(f"\nStep 7: Testing MIT control with custom gains...")
+        print(f"  Using lower gains for gentler movement...")
+        
+        # Lower gains for smoother motion
+        bus._mit_control(
+            MOTOR_NAME,
+            kp=5.0,   # Lower position gain
+            kd=0.3,   # Lower damping
+            position_degrees=30.0,
+            velocity_deg_per_sec=0.0,
+            torque=0.0
+        )
+        time.sleep(1.5)
+        
+        final_pos = bus.read("Present_Position", MOTOR_NAME, normalize=False)
+        print(f"  Final position: {final_pos:8.2f}°")
+        print(f"  ✓ MIT control test complete")
+        
+        # Step 8: Return to zero
+        print(f"\nStep 8: Returning to zero position...")
+        bus.write("Goal_Position", MOTOR_NAME, 0.0, normalize=False)
+        time.sleep(1.0)
+        
+        final_pos = bus.read("Present_Position", MOTOR_NAME, normalize=False)
+        print(f"  Final position: {final_pos:8.2f}°")
+        
+    finally:
+        # Step 9: Disable motor
+        print(f"\nStep 9: Disabling motor...")
+        if bus.is_connected:
+            bus.disable_torque(MOTOR_NAME)
+            time.sleep(0.1)
+            print(f"  ✓ Motor disabled (torque off)")
+        
+        # Step 10: Disconnect
+        print(f"\nStep 10: Disconnecting...")
+        if bus.is_connected:
+            bus.disconnect(disable_torque=False)  # Already disabled
+            print(f"  ✓ Disconnected from {can_port}")
+    
+    print(f"\n{'='*60}")
+    print("Test completed successfully!")
+    print(f"{'='*60}\n")
+
+
+@pytest.mark.hardware
+def test_motor_discovery_and_setup(can_port):
+    """
+    Test motor discovery and ID configuration.
+    
+    Note: This test requires the Damiao Debugging Tools for actual ID changes.
+    This test only demonstrates the bus scan functionality.
+    """
+    
+    print(f"\n{'='*60}")
+    print("Damiao Motor Discovery Test")
+    print(f"{'='*60}\n")
+    
+    print("Note: Motor ID configuration must be done via Damiao Debugging Tools")
+    print("See: https://docs.openarm.dev/software/setup/motor-id")
+    print()
+    
+    # Test if CAN interface is accessible
+    print(f"Testing CAN interface: {can_port}")
+    
+    # Create a minimal motor bus for testing
+    test_motor = Motor(0x01, "dm4310", MotorNormMode.DEGREES)
+    test_motor.recv_id = 0x11
+    test_motor.motor_type = MotorType.DM4310
+    
+    bus = DamiaoMotorsBus(port=can_port, motors={"test": test_motor})
+    
+    try:
+        bus.connect(handshake=False)
+        print(f"✓ CAN interface {can_port} is accessible")
+        
+        # Try to communicate with motor at 0x01
+        print(f"\nLooking for motor at ID 0x01...")
+        try:
+            bus._refresh_motor("test")
+            msg = bus._recv_motor_response(timeout=0.5)
+            if msg:
+                print(f"✓ Motor found at ID 0x01, response ID: 0x{msg.arbitration_id:02X}")
+            else:
+                print(f"✗ No response from motor")
+                print("\nTroubleshooting:")
+                print("  1. Verify motor is powered (24V)")
+                print("  2. Check CAN wiring (CANH, CANL)")
+                print("  3. Verify motor ID is set to 0x01")
+                print("  4. Enable with: cansend can0 001#FFFFFFFFFFFFFFFC")
+        except Exception as e:
+                print(f"✗ Error communicating with motor: {e}")
+        
+    except Exception as e:
+        print(f"✗ Failed to access CAN interface: {e}")
+        print("\nSetup CAN interface:")
+        print(f"  sudo ip link set {can_port} type can bitrate 1000000")
+        print(f"  sudo ip link set {can_port} up")
+    
+    finally:
+        if bus.is_connected:
+            bus.disconnect(disable_torque=True)
+    
+    print(f"\n{'='*60}\n")
+
+
+@pytest.mark.hardware
+def test_multi_motor_sync_operations(can_port):
+    """
+    Test synchronized read/write with multiple motors.
+    
+    This demonstrates how to control multiple motors simultaneously.
+    """
+    
+    print(f"\n{'='*60}")
+    print("Damiao Multi-Motor Sync Test")
+    print(f"{'='*60}\n")
+    
+    # Setup motors (adjust IDs as needed)
+    motors = {
+        "joint_1": Motor(0x01, "dm4310", MotorNormMode.DEGREES),
+        "joint_2": Motor(0x02, "dm4310", MotorNormMode.DEGREES),
+    }
+    
+    motors["joint_1"].recv_id = 0x11
+    motors["joint_1"].motor_type = MotorType.DM4310
+    motors["joint_2"].recv_id = 0x12
+    motors["joint_2"].motor_type = MotorType.DM4310
+    
+    bus = DamiaoMotorsBus(port=can_port, motors=motors)
+    
+    try:
+        bus.connect()
+        bus.enable_torque()
+        
+        print("Reading all motor positions...")
+        positions = bus.sync_read("Present_Position")
+        for motor, pos in positions.items():
+            print(f"  {motor}: {pos:.2f}°")
+        
+        print("\nMoving all motors to 45°...")
+        target_positions = {motor: 45.0 for motor in motors}
+        bus.sync_write("Goal_Position", target_positions)
+        time.sleep(2.0)
+        
+        positions = bus.sync_read("Present_Position")
+        print("Final positions:")
+        for motor, pos in positions.items():
+            print(f"  {motor}: {pos:.2f}°")
+        
+    except Exception as e:
+        print(f"✗ Test failed: {e}")
+        print("\nThis is expected on macOS without proper CAN hardware.")
+        print("macOS does not support SocketCAN natively (Linux-only feature).")
+        print("For macOS, you need a USB-CAN adapter with SLCAN support.")
+    finally:
+        if bus.is_connected:
+            bus.disable_torque()
+            bus.disconnect()
+    
+    print(f"\n{'='*60}\n")
+
+
+if __name__ == "__main__":
+    print("Damiao Motor Test Suite")
+    print("=" * 60)
+    print("\nThese tests require actual hardware to run.")
+    print("Please ensure:")
+    print("  1. Motor is connected and powered (24V)")
+    print("  2. CAN interface is configured")
+    print("  3. Motor ID is set to 0x01/0x11")
+    print("\nTo run tests with hardware:")
+    print("\n  Linux (SocketCAN):")
+    print("    sudo ip link set can0 type can bitrate 1000000")
+    print("    sudo ip link set can0 up")
+    print("    pytest tests/motors/test_damiao.py --run-hardware --can-port can0")
+    print("\n  macOS (USB-CAN adapter with SLCAN):")
+    print("    pytest tests/motors/test_damiao.py --run-hardware --can-port /dev/cu.usbmodem00000000050C1")
+    print("\nTo run without hardware (tests will be skipped):")
+    print("  pytest tests/motors/test_damiao.py")
+    print("\nNote: The --run-hardware and --can-port flags are configured in tests/conftest.py")
+    print("=" * 60)
+

tests/policies/wall_x/test_wallx.py

@@ -19,23 +19,28 @@
 import os
 
 import pytest
-import torch
 
-# Skip if openpi or transformers is not available
-pytest.importorskip("peft")
-pytest.importorskip("transformers==4.49.0")
+try:
+    import peft  # noqa: F401
 
-# Skip this entire module in CI
+    WALLX_AVAILABLE = True
+except ImportError:
+    WALLX_AVAILABLE = False
+
+# Skip this entire module if wallx deps not available or in CI
 pytestmark = pytest.mark.skipif(
-    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="This test requires local Wall-X installation and is not meant for CI",
+    not WALLX_AVAILABLE or os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
+    reason="This test requires Wall-X dependencies (peft) and is not meant for CI",
 )
 
-from lerobot.policies.factory import make_policy_config  # noqa: E402
-from lerobot.policies.wall_x import WallXConfig  # noqa: E402
-from lerobot.policies.wall_x.modeling_wall_x import WallXPolicy  # noqa: E402
-from lerobot.policies.wall_x.processor_wall_x import make_wall_x_pre_post_processors  # noqa: E402
-from lerobot.utils.random_utils import set_seed  # noqa: E402
+if WALLX_AVAILABLE:
+    import torch
+
+    from lerobot.policies.factory import make_policy_config
+    from lerobot.policies.wall_x import WallXConfig
+    from lerobot.policies.wall_x.modeling_wall_x import WallXPolicy
+    from lerobot.policies.wall_x.processor_wall_x import make_wall_x_pre_post_processors
+    from lerobot.utils.random_utils import set_seed
 
 
 def test_policy_instantiation():