Merge branch 'feat/mirror' into openarms_wallx_rebased_3

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

* add delete button

* fixes

* remove viz

* unused import

STREAMING_ENCODING_PR.md

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

docs/source/_toctree.yml

@@ -57,6 +57,8 @@
     title: Use Async Inference
   - local: rtc
     title: Real-Time Chunking (RTC)
+  - local: training_time_rtc
+    title: Training-Time RTC
   title: "Inference"
 - sections:
   - local: envhub

docs/source/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}
+}
+```
+

docs/source/training_time_rtc.mdx

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

examples/openarms/debug_can_communication.py

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

examples/openarms/evaluate.py

@@ -0,0 +1,360 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+OpenArms Policy Evaluation
+
+Evaluates a trained policy on the OpenArms robot by running inference and recording
+the evaluation episodes to a dataset. Supports optional leader arm for manual resets.
+
+Example usage:
+    python examples/openarms/evaluate.py
+"""
+
+import time
+from pathlib import Path
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.processor import make_default_processors
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.scripts.lerobot_record import record_loop
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.utils.control_utils import init_keyboard_listener
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import init_rerun
+
+
+HF_MODEL_ID = "lerobot-data-collection/level1_rac2_100k"  # TODO: Replace with your trained model
+HF_EVAL_DATASET_ID = "lerobot-data-collection/three-folds-pi0_eval_raccc3"  # TODO: Replace with your eval dataset name
+TASK_DESCRIPTION = "Fold the T-shirt properly" # TODO: Replace with your task, this should match!!
+
+NUM_EPISODES = 1
+FPS = 30
+EPISODE_TIME_SEC = 1000
+RESET_TIME_SEC = 60
+
+# Robot CAN interfaces
+FOLLOWER_LEFT_PORT = "can0"
+FOLLOWER_RIGHT_PORT = "can1"
+
+# If enabled, you can manually reset the environment between evaluation episodes
+USE_LEADER_FOR_RESETS = False  # Set to False if you don't want to use leader
+LEADER_LEFT_PORT = "can2"
+LEADER_RIGHT_PORT = "can3"
+
+# Camera configuration
+CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video0", width=1280, height=720, fps=FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video5", width=1280, height=720, fps=FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video2", width=640, height=480, fps=FPS),
+}
+
+def main():
+    """Main evaluation function."""
+    print("OpenArms Policy Evaluation")
+    print(f"\nModel: {HF_MODEL_ID}")
+    print(f"Evaluation Dataset: {HF_EVAL_DATASET_ID}")
+    print(f"Task: {TASK_DESCRIPTION}")
+    print(f"Episodes: {NUM_EPISODES}")
+    print(f"Episode Duration: {EPISODE_TIME_SEC}s")
+    print(f"Reset Duration: {RESET_TIME_SEC}s")
+    print(f"Use Leader for Resets: {USE_LEADER_FOR_RESETS}")
+    
+    follower_config = OpenArmsFollowerConfig(
+        port_left=FOLLOWER_LEFT_PORT,
+        port_right=FOLLOWER_RIGHT_PORT,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=CAMERA_CONFIG,
+    )
+    
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+    
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    
+    leader = None
+    if USE_LEADER_FOR_RESETS:
+        leader_config = OpenArmsLeaderConfig(
+            port_left=LEADER_LEFT_PORT,
+            port_right=LEADER_RIGHT_PORT,
+            can_interface="socketcan",
+            id="openarms_leader",
+            manual_control=False,  # Enable torque control for gravity compensation
+        )
+        
+        leader = OpenArmsLeader(leader_config)
+        leader.connect(calibrate=False)
+        
+        if not leader.is_connected:
+            raise RuntimeError("Leader robot failed to connect!")
+        
+        # Enable gravity compensation
+        if leader.pin_robot is not None:
+            leader.bus_right.enable_torque()
+            leader.bus_left.enable_torque()
+            time.sleep(0.1)
+            print(f"Leader connected with gravity compensation ({LEADER_LEFT_PORT}, {LEADER_RIGHT_PORT})")
+        else:
+            print(f"Leader connected but gravity compensation unavailable (no URDF)")
+
+    # Build default processors for action and observation
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+    
+    # Build dataset features from robot features and processors
+    # For actions, only include positions (no velocity or torque)
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+    
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+    
+    # Check if dataset already exists
+    dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / HF_EVAL_DATASET_ID
+    if dataset_path.exists():
+        print(f"Evaluation dataset already exists at: {dataset_path}")
+        print("This will append new episodes to the existing dataset.")
+        choice = input("  Continue? (y/n): ").strip().lower()
+        if choice != 'y':
+            print("  Aborting evaluation.")
+            follower.disconnect()
+            if leader:
+                leader.disconnect()
+            return
+    
+    # Create dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_EVAL_DATASET_ID,
+        fps=FPS,
+        features=dataset_features,
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_processes=0,
+        image_writer_threads=12, 
+    )
+    
+    # Load policy config from pretrained model and create policy using factory
+    policy_config = PreTrainedConfig.from_pretrained(HF_MODEL_ID)
+    policy_config.pretrained_path = HF_MODEL_ID
+    policy = make_policy(policy_config, ds_meta=dataset.meta)
+    
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy.config,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        preprocessor_overrides={
+            "device_processor": {"device": str(policy.config.device)}
+        },
+    )
+
+    print(f"\nRunning evaluation...")
+    # Initialize keyboard listener and visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="openarms_evaluation")
+    episode_idx = 0
+    
+    try:
+        while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+            log_say(f"Evaluating episode {episode_idx + 1} of {NUM_EPISODES}")
+            print(f"\nRunning inference for episode {episode_idx + 1}...")
+            
+            # Run inference with policy
+            record_loop(
+                robot=follower,
+                events=events,
+                fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+                policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                dataset=dataset,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
+            
+            # Handle re-recording
+            if events["rerecord_episode"]:
+                log_say("Re-recording episode")
+                events["rerecord_episode"] = False
+                events["exit_early"] = False
+                dataset.clear_episode_buffer()
+                continue
+            
+            # Save episode
+            if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                print(f"Saving episode {episode_idx + 1} ({dataset.episode_buffer['size']} frames)...")
+                dataset.save_episode()
+                episode_idx += 1
+            
+            # Reset environment between episodes (if not last episode)
+            if not events["stop_recording"] and episode_idx < NUM_EPISODES:
+                if USE_LEADER_FOR_RESETS and leader:
+                    log_say("Reset the environment using leader arms")
+                    print(f"\nManual reset period ({RESET_TIME_SEC}s)...")
+                    
+                    # Use leader for manual reset with gravity compensation
+                    import numpy as np
+                    
+                    dt = 1 / FPS
+                    reset_start_time = time.perf_counter()
+                    
+                    while time.perf_counter() - reset_start_time < RESET_TIME_SEC:
+                        if events["exit_early"] or events["stop_recording"]:
+                            break
+                        
+                        loop_start = time.perf_counter()
+                        
+                        # Get leader state
+                        leader_action = leader.get_action()
+                        
+                        # Extract positions and velocities
+                        leader_positions_deg = {}
+                        leader_velocities_deg_per_sec = {}
+                        
+                        for motor in leader.bus_right.motors:
+                            pos_key = f"right_{motor}.pos"
+                            vel_key = f"right_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"right_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"right_{motor}"] = leader_action[vel_key]
+                        
+                        for motor in leader.bus_left.motors:
+                            pos_key = f"left_{motor}.pos"
+                            vel_key = f"left_{motor}.vel"
+                            if pos_key in leader_action:
+                                leader_positions_deg[f"left_{motor}"] = leader_action[pos_key]
+                            if vel_key in leader_action:
+                                leader_velocities_deg_per_sec[f"left_{motor}"] = leader_action[vel_key]
+                        
+                        # Calculate gravity and friction torques
+                        leader_positions_rad = {k: np.deg2rad(v) for k, v in leader_positions_deg.items()}
+                        leader_gravity_torques_nm = leader._gravity_from_q(leader_positions_rad)
+                        
+                        leader_velocities_rad_per_sec = {k: np.deg2rad(v) for k, v in leader_velocities_deg_per_sec.items()}
+                        leader_friction_torques_nm = leader._friction_from_velocity(
+                            leader_velocities_rad_per_sec,
+                            friction_scale=1.0
+                        )
+                        
+                        # Combine torques
+                        leader_total_torques_nm = {}
+                        for motor_name in leader_gravity_torques_nm:
+                            gravity = leader_gravity_torques_nm.get(motor_name, 0.0)
+                            friction = leader_friction_torques_nm.get(motor_name, 0.0)
+                            leader_total_torques_nm[motor_name] = gravity + friction
+                        
+                        # Apply compensation
+                        for motor in leader.bus_right.motors:
+                            full_name = f"right_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_right._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        for motor in leader.bus_left.motors:
+                            full_name = f"left_{motor}"
+                            position = leader_positions_deg.get(full_name, 0.0)
+                            torque = leader_total_torques_nm.get(full_name, 0.0)
+                            kd = leader.get_damping_kd(motor)
+                            
+                            leader.bus_left._mit_control(
+                                motor=motor, kp=0.0, kd=kd,
+                                position_degrees=position,
+                                velocity_deg_per_sec=0.0,
+                                torque=torque,
+                            )
+                        
+                        # Send leader positions to follower
+                        follower_action = {}
+                        for joint in leader_positions_deg.keys():
+                            pos_key = f"{joint}.pos"
+                            if pos_key in leader_action:
+                                follower_action[pos_key] = leader_action[pos_key]
+                        
+                        if follower_action:
+                            follower.send_action(follower_action)
+                        
+                        # Maintain loop rate
+                        loop_duration = time.perf_counter() - loop_start
+                        sleep_time = dt - loop_duration
+                        if sleep_time > 0:
+                            time.sleep(sleep_time)
+                    
+                    print("Reset complete")
+                else:
+                    log_say("Waiting for manual reset")
+                    print(f"Manually reset the environment and press ENTER to continue")
+                    input("Press ENTER when ready...")
+        
+        print(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+    
+    except KeyboardInterrupt:
+        print("\n\nEvaluation interrupted by user")
+    
+    finally:
+        if leader:
+            leader.bus_right.disable_torque()
+            leader.bus_left.disable_torque()
+            time.sleep(0.1)
+            leader.disconnect()
+
+        follower.disconnect()
+        
+        if listener is not None:
+            listener.stop()
+        
+        dataset.finalize()
+        print("\nUploading to Hugging Face Hub...")
+        dataset.push_to_hub(private=True)
+
+
+if __name__ == "__main__":
+    main()
+

examples/openarms/evaluate_with_rtc.py

@@ -0,0 +1,703 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+OpenArms Policy Evaluation with Real-Time Chunking (RTC)
+
+Evaluates a trained policy on the OpenArms robot using RTC for smooth, continuous motion.
+RTC enables large flow-matching policies (Pi0, Pi0.5, SmolVLA) to produce reactive motion
+despite high inference latency by asynchronously generating action chunks.
+
+Features:
+- Thread-based asynchronous action generation and execution
+- RTC for smooth transitions between action chunks
+- Dataset recording for evaluation episodes
+
+Example usage:
+    python examples/openarms/evaluate_with_rtc.py
+
+    # With custom RTC parameters
+    python examples/openarms/evaluate_with_rtc.py \
+        --rtc.execution_horizon=12 \
+        --rtc.max_guidance_weight=10.0
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from pathlib import Path
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import build_dataset_frame, combine_feature_dicts, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor import make_default_processors
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging, log_say
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+# ============================================================================
+# Default Configuration Constants
+# ============================================================================
+
+DEFAULT_HF_MODEL_ID = "lerobot-data-collection/level1_rac3_100k"
+DEFAULT_HF_EVAL_DATASET_ID = "lerobot-data-collection/test"
+DEFAULT_TASK_DESCRIPTION = "Fold the T-shirt properly"
+
+DEFAULT_NUM_EPISODES = 1
+DEFAULT_FPS = 30
+DEFAULT_EPISODE_TIME_SEC = 1000
+DEFAULT_RESET_TIME_SEC = 60
+
+DEFAULT_FOLLOWER_LEFT_PORT = "can0"
+DEFAULT_FOLLOWER_RIGHT_PORT = "can1"
+
+DEFAULT_CAMERA_CONFIG = {
+    "left_wrist": OpenCVCameraConfig(index_or_path="/dev/video0", width=1280, height=720, fps=DEFAULT_FPS),
+    "right_wrist": OpenCVCameraConfig(index_or_path="/dev/video4", width=1280, height=720, fps=DEFAULT_FPS),
+    "base": OpenCVCameraConfig(index_or_path="/dev/video2", width=640, height=480, fps=DEFAULT_FPS),
+}
+
+
+# ============================================================================
+# Thread-Safe Robot Wrapper
+# ============================================================================
+
+
+class RobotWrapper:
+    """Thread-safe wrapper for robot operations."""
+
+    def __init__(self, robot: OpenArmsFollower):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: dict) -> None:
+        with self.lock:
+            self.robot.send_action(action)
+
+    @property
+    def observation_features(self) -> dict:
+        with self.lock:
+            return self.robot.observation_features
+
+    @property
+    def action_features(self) -> dict:
+        with self.lock:
+            return self.robot.action_features
+
+    @property
+    def name(self) -> str:
+        return self.robot.name
+
+
+# ============================================================================
+# Configuration
+# ============================================================================
+
+
+@dataclass
+class OpenArmsRTCEvalConfig(HubMixin):
+    """Configuration for OpenArms evaluation with RTC."""
+
+    policy: PreTrainedConfig | None = None
+
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=20,
+            max_guidance_weight=5.0,
+            prefix_attention_schedule=RTCAttentionSchedule.LINEAR,
+        )
+    )
+
+    model_id: str = DEFAULT_HF_MODEL_ID
+    eval_dataset_id: str = DEFAULT_HF_EVAL_DATASET_ID
+    task: str = DEFAULT_TASK_DESCRIPTION
+
+    num_episodes: int = DEFAULT_NUM_EPISODES
+    fps: float = DEFAULT_FPS
+    episode_time_sec: float = DEFAULT_EPISODE_TIME_SEC
+    reset_time_sec: float = DEFAULT_RESET_TIME_SEC
+
+    follower_left_port: str = DEFAULT_FOLLOWER_LEFT_PORT
+    follower_right_port: str = DEFAULT_FOLLOWER_RIGHT_PORT
+
+    device: str = "cuda"
+
+    # Should be higher than inference_delay + execution_horizon
+    action_queue_size_to_get_new_actions: int = 30
+
+    record_dataset: bool = True
+    push_to_hub: bool = True
+
+    interpolation: bool = True
+
+    use_torch_compile: bool = False
+    torch_compile_backend: str = "inductor"
+    torch_compile_mode: str = "default"
+    torch_compile_disable_cudagraphs: bool = True
+
+    def __post_init__(self):
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+            self.model_id = policy_path
+        elif self.model_id:
+            self.policy = PreTrainedConfig.from_pretrained(self.model_id)
+            self.policy.pretrained_path = self.model_id
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        return ["policy"]
+
+
+# ============================================================================
+# Action Generation Thread
+# ============================================================================
+
+
+def get_actions_thread(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+):
+    """Thread function to asynchronously generate action chunks from the policy."""
+    try:
+        logger.info("[GET_ACTIONS] Starting action generation thread")
+
+        latency_tracker = LatencyTracker()
+        time_per_chunk = 1.0 / cfg.fps
+
+        hw_features = hw_to_dataset_features(robot.observation_features, "observation")
+        policy_device = policy.config.device
+
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                time.sleep(0.01)
+                continue
+
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk) if inference_latency else 0
+
+                obs = robot.get_observation()
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    hw_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]
+                obs_with_policy_features["robot_type"] = robot.name
+
+                preprocessed_obs = preprocessor(obs_with_policy_features)
+
+                actions = policy.predict_action_chunk(
+                    preprocessed_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                original_actions = actions.squeeze(0).clone()
+                postprocessed_actions = postprocessor(actions).squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] action_queue_size_to_get_new_actions too small. "
+                        "Should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+
+                logger.debug(
+                    f"[GET_ACTIONS] Generated chunk, latency={new_latency:.3f}s, "
+                    f"delay={new_delay}, queue_size={action_queue.qsize()}"
+                )
+            else:
+                time.sleep(0.01)
+
+        logger.info("[GET_ACTIONS] Action generation thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Action Execution Thread
+# ============================================================================
+
+
+def actor_thread(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: OpenArmsRTCEvalConfig,
+    episode_active: Event,
+    dataset: LeRobotDataset | None,
+    dataset_lock: Lock,
+    teleop_action_processor,
+    robot_observation_processor,
+):
+    """Thread function to execute actions on the robot."""
+    try:
+        action_keys = [k for k in robot.action_features.keys() if k.endswith(".pos")]
+
+        if cfg.interpolation:
+            interp_factor = 2
+            robot_interval = 1.0 / (cfg.fps * interp_factor)
+            logger.info(f"[ACTOR] Interpolation ON: policy={cfg.fps}Hz -> robot={cfg.fps * interp_factor}Hz (2x)")
+        else:
+            interp_factor = 1
+            robot_interval = 1.0 / cfg.fps
+            logger.info(f"[ACTOR] Interpolation OFF: policy={cfg.fps}Hz, robot={cfg.fps}Hz")
+
+        prev_action: Tensor | None = None
+        interpolated_actions: list[Tensor] = []
+        interp_idx = 0
+
+        robot_send_count = 0
+        policy_consume_count = 0
+        last_hz_print = time.perf_counter()
+        last_dataset_time = 0.0
+
+        while not shutdown_event.is_set():
+            if not episode_active.is_set():
+                prev_action = None
+                interpolated_actions = []
+                interp_idx = 0
+                robot_send_count = 0
+                policy_consume_count = 0
+                last_hz_print = time.perf_counter()
+                time.sleep(0.01)
+                continue
+
+            start_time = time.perf_counter()
+
+            if interp_idx >= len(interpolated_actions):
+                new_action = action_queue.get()
+                if new_action is not None:
+                    current_action = new_action.cpu()
+                    policy_consume_count += 1
+
+                    if cfg.interpolation and prev_action is not None:
+                        mid = prev_action + 0.5 * (current_action - prev_action)
+                        interpolated_actions = [mid, current_action]
+                    else:
+                        interpolated_actions = [current_action]
+
+                    prev_action = current_action
+                    interp_idx = 0
+
+            if interp_idx < len(interpolated_actions):
+                action_to_send = interpolated_actions[interp_idx]
+                interp_idx += 1
+
+                action_dict = {}
+                for i, key in enumerate(action_keys):
+                    if i < len(action_to_send):
+                        action_dict[key] = action_to_send[i].item()
+
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+                robot_send_count += 1
+
+                if cfg.record_dataset and dataset is not None:
+                    now = time.perf_counter()
+                    if now - last_dataset_time >= (1.0 / cfg.fps):
+                        last_dataset_time = now
+                        with dataset_lock:
+                            obs = robot.get_observation()
+                            obs_processed = robot_observation_processor(obs)
+                            action_for_dataset = teleop_action_processor((action_dict, None))
+                            frame = {}
+                            for key, value in obs_processed.items():
+                                frame[f"observation.{key}"] = value
+                            for key, value in action_for_dataset.items():
+                                frame[f"action.{key}"] = value
+                            frame["task"] = cfg.task
+                            dataset.add_frame(frame)
+
+            now = time.perf_counter()
+            if now - last_hz_print >= 5.0:
+                elapsed = now - last_hz_print
+                actual_robot_hz = robot_send_count / elapsed if elapsed > 0 else 0
+                actual_policy_hz = policy_consume_count / elapsed if elapsed > 0 else 0
+                logger.info(f"[ACTOR] Actual Hz - Robot: {actual_robot_hz:.1f}, Policy: {actual_policy_hz:.1f}")
+                robot_send_count = 0
+                policy_consume_count = 0
+                last_hz_print = now
+
+            dt_s = time.perf_counter() - start_time
+            sleep_time = max(0, robot_interval - dt_s - 0.001)
+            if sleep_time > 0:
+                time.sleep(sleep_time)
+
+        logger.info("[ACTOR] Shutting down")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception: {e}")
+        logger.error(traceback.format_exc())
+        shutdown_event.set()
+        sys.exit(1)
+
+
+# ============================================================================
+# Main Evaluation Function
+# ============================================================================
+
+
+def _apply_torch_compile(policy, cfg: OpenArmsRTCEvalConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method."""
+    if policy.name in ["pi05", "pi0"]:
+        return policy
+
+    try:
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def main(cfg: OpenArmsRTCEvalConfig):
+    """Main evaluation function with RTC."""
+    init_logging()
+
+    print("=" * 60)
+    print("OpenArms Policy Evaluation with RTC")
+    print("=" * 60)
+    print(f"\nModel: {cfg.model_id}")
+    print(f"Evaluation Dataset: {cfg.eval_dataset_id}")
+    print(f"Task: {cfg.task}")
+    print(f"Episodes: {cfg.num_episodes}")
+    print(f"Episode Duration: {cfg.episode_time_sec}s")
+    print(f"RTC Enabled: {cfg.rtc.enabled}")
+    print(f"RTC Execution Horizon: {cfg.rtc.execution_horizon}")
+    print(f"RTC Max Guidance Weight: {cfg.rtc.max_guidance_weight}")
+    print(f"Policy Hz: {cfg.fps}")
+    print(f"Robot Hz: {cfg.fps * 2 if cfg.interpolation else cfg.fps}")
+    print(f"Interpolation: {cfg.interpolation}")
+    print(f"Device: {cfg.device}")
+    print("=" * 60)
+
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+    episode_active = Event()
+
+    # Initialize Robot
+    follower_config = OpenArmsFollowerConfig(
+        port_left=cfg.follower_left_port,
+        port_right=cfg.follower_right_port,
+        can_interface="socketcan",
+        id="openarms_follower",
+        disable_torque_on_disconnect=True,
+        max_relative_target=10.0,
+        cameras=DEFAULT_CAMERA_CONFIG,
+    )
+
+    follower = OpenArmsFollower(follower_config)
+    follower.connect(calibrate=False)
+
+    if not follower.is_connected:
+        raise RuntimeError("Follower robot failed to connect!")
+
+    robot = RobotWrapper(follower)
+    logger.info("Follower robot connected")
+
+    # Build Processors and Dataset Features
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    action_features_hw = {}
+    for key, value in follower.action_features.items():
+        if key.endswith(".pos"):
+            action_features_hw[key] = value
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=action_features_hw),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_observation_processor,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    )
+
+    # Create or Load Dataset
+    dataset = None
+    dataset_lock = Lock()
+
+    if cfg.record_dataset:
+        dataset_path = Path.home() / ".cache" / "huggingface" / "lerobot" / cfg.eval_dataset_id
+        if dataset_path.exists():
+            logger.info(f"Evaluation dataset exists at: {dataset_path}")
+            logger.info("New episodes will be appended.")
+            choice = input("Continue? (y/n): ").strip().lower()
+            if choice != "y":
+                logger.info("Aborting evaluation.")
+                follower.disconnect()
+                return
+
+        dataset = LeRobotDataset.create(
+            repo_id=cfg.eval_dataset_id,
+            fps=int(cfg.fps),
+            features=dataset_features,
+            robot_type=follower.name,
+            use_videos=True,
+            image_writer_processes=0,
+            image_writer_threads=12,
+        )
+        logger.info(f"Dataset created: {cfg.eval_dataset_id}")
+
+    # Load Policy
+    logger.info(f"Loading policy from: {cfg.model_id}")
+
+    policy_class = get_policy_class(cfg.policy.type)
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type in ["pi05", "pi0"]:
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    policy.config.rtc_config = cfg.rtc
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    logger.info(f"Policy loaded: {policy.name}")
+
+    # Create Action Queue and Start Threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    get_actions_t = Thread(
+        target=get_actions_thread,
+        args=(
+            policy,
+            robot,
+            robot_observation_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+        ),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_t.start()
+    logger.info("Started action generation thread")
+
+    actor_t = Thread(
+        target=actor_thread,
+        args=(
+            robot,
+            robot_action_processor,
+            action_queue,
+            shutdown_event,
+            cfg,
+            episode_active,
+            dataset,
+            dataset_lock,
+            teleop_action_processor,
+            robot_observation_processor,
+        ),
+        daemon=True,
+        name="Actor",
+    )
+    actor_t.start()
+    logger.info("Started actor thread")
+
+    # Run Evaluation Episodes
+    episode_idx = 0
+
+    try:
+        while episode_idx < cfg.num_episodes and not shutdown_event.is_set():
+            log_say(f"Evaluating episode {episode_idx + 1} of {cfg.num_episodes}")
+            logger.info(f"\n{'='*40}")
+            logger.info(f"Episode {episode_idx + 1} / {cfg.num_episodes}")
+            logger.info(f"{'='*40}")
+
+            action_queue = ActionQueue(cfg.rtc)
+            episode_active.set()
+            episode_start_time = time.time()
+
+            while (time.time() - episode_start_time) < cfg.episode_time_sec:
+                if shutdown_event.is_set():
+                    break
+
+                elapsed = time.time() - episode_start_time
+                if int(elapsed) % 10 == 0 and int(elapsed) > 0:
+                    logger.info(
+                        f"[MAIN] Episode progress: {elapsed:.0f}/{cfg.episode_time_sec}s, "
+                        f"queue_size={action_queue.qsize()}"
+                    )
+
+                time.sleep(0.5)
+
+            episode_active.clear()
+            logger.info(f"Episode {episode_idx + 1} completed")
+
+            if cfg.record_dataset and dataset is not None:
+                with dataset_lock:
+                    if dataset.episode_buffer is not None and dataset.episode_buffer.get("size", 0) > 0:
+                        logger.info(
+                            f"Saving episode {episode_idx + 1} "
+                            f"({dataset.episode_buffer['size']} frames)"
+                        )
+                        dataset.save_episode()
+
+            episode_idx += 1
+
+            # Manual reset between episodes
+            if not shutdown_event.is_set() and episode_idx < cfg.num_episodes:
+                log_say("Waiting for manual reset")
+                logger.info("Manually reset the environment and press ENTER to continue")
+                input("Press ENTER when ready...")
+
+        logger.info(f"Evaluation complete! {episode_idx} episodes recorded")
+        log_say("Evaluation complete", blocking=True)
+
+    except KeyboardInterrupt:
+        logger.info("\n\nEvaluation interrupted by user")
+
+    finally:
+        shutdown_event.set()
+        episode_active.clear()
+
+        if get_actions_t.is_alive():
+            logger.info("Waiting for action generation thread to finish...")
+            get_actions_t.join(timeout=5.0)
+
+        if actor_t.is_alive():
+            logger.info("Waiting for actor thread to finish...")
+            actor_t.join(timeout=5.0)
+
+        follower.disconnect()
+        logger.info("Follower disconnected")
+
+        if cfg.record_dataset and dataset is not None:
+            dataset.finalize()
+            if cfg.push_to_hub:
+                logger.info("Uploading to Hugging Face Hub...")
+                dataset.push_to_hub(private=True)
+
+        logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    main()

examples/openarms/friction_compensation.py

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

examples/openarms/gravity_compensation.py

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

examples/openarms/record_with_compensation.py

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

examples/openarms/replay.py

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

examples/openarms/setup_can.sh

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

examples/openarms/teleop.py

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

examples/openarms/teleop_openarms_mini.py

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

examples/openarms/teleop_with_compensation.py

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

examples/openarms/unify_task.py

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

examples/openarms_web_interface/App.css

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

examples/openarms_web_interface/App.jsx

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

examples/openarms_web_interface/README.md

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

examples/openarms_web_interface/index.html

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

examples/openarms_web_interface/launch.sh

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

examples/openarms_web_interface/main.jsx

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

examples/openarms_web_interface/package.json

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

examples/openarms_web_interface/vite.config.js

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

examples/rac/cmd.sh

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

examples/rac/rac_data_collection.py

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

examples/rac/rac_data_collection_openarms.py

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

examples/rac/rac_data_collection_openarms_rtc.py

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

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/configs/types.py

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

src/lerobot/datasets/compute_stats.py

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

src/lerobot/datasets/dataset_tools.py

@@ -37,7 +37,7 @@ import torch
 from tqdm import tqdm
 
 from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.datasets.compute_stats import aggregate_stats
+from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats, get_feature_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
     DATA_DIR,
@@ -1522,6 +1522,122 @@ def modify_tasks(
     return dataset
 
 
+def recompute_stats(
+    dataset: LeRobotDataset,
+    skip_image_video: bool = True,
+    delta_action: bool = False,
+    delta_exclude_joints: list[str] | None = None,
+) -> LeRobotDataset:
+    """Recompute stats.json from scratch by iterating all episodes.
+
+    Args:
+        dataset: The LeRobotDataset to recompute stats for.
+        skip_image_video: If True (default), only recompute stats for numeric features
+            (action, state, etc.) and keep existing image/video stats unchanged.
+        delta_action: If True, compute action stats as delta (action - state).
+            Useful when training with use_delta_actions=True so normalization matches.
+        delta_exclude_joints: Joint names to exclude from delta conversion when
+            delta_action=True. These dims keep absolute stats. Uses dataset's
+            action feature names to build the mask. Default: ["gripper"].
+
+    Returns:
+        The same dataset with updated stats.
+    """
+    features = dataset.meta.features
+    numeric_features = {
+        k: v for k, v in features.items()
+        if v["dtype"] not in ["image", "video", "string"]
+        and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+    }
+
+    if skip_image_video:
+        features_to_compute = numeric_features
+    else:
+        features_to_compute = {
+            k: v for k, v in features.items()
+            if v["dtype"] != "string"
+            and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+        }
+
+    # Build delta mask if delta_action is enabled
+    delta_mask = None
+    if delta_action and "action" in features and "observation.state" in features:
+        if delta_exclude_joints is None:
+            delta_exclude_joints = ["gripper"]
+        action_names = features["action"].get("names")
+        if action_names is not None:
+            exclude = set(delta_exclude_joints)
+            delta_mask = [n not in exclude for n in action_names]
+        else:
+            action_dim = features["action"]["shape"][0]
+            delta_mask = [True] * action_dim
+        # Only recompute action stats when delta is enabled — state stays unchanged
+        features_to_compute = {"action": features["action"]}
+        logging.info(f"Recomputing action stats as delta (exclude: {delta_exclude_joints})")
+    else:
+        logging.info(f"Recomputing stats for features: {list(features_to_compute.keys())}")
+
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    all_episode_stats = []
+    numeric_keys = [k for k, v in features_to_compute.items() if v["dtype"] not in ["image", "video"]]
+    # Also need state for delta computation even though we don't recompute state stats
+    needs_state = delta_mask is not None
+
+    for parquet_path in tqdm(parquet_files, desc="Computing stats from data files"):
+        df = pd.read_parquet(parquet_path)
+
+        for ep_idx in sorted(df["episode_index"].unique()):
+            ep_df = df[df["episode_index"] == ep_idx]
+            episode_data = {}
+            for key in numeric_keys:
+                if key in ep_df.columns:
+                    values = ep_df[key].values
+                    if hasattr(values[0], "__len__"):
+                        episode_data[key] = np.stack(values)
+                    else:
+                        episode_data[key] = np.array(values)
+
+            # Apply delta conversion to actions before computing stats
+            if delta_mask is not None and "action" in episode_data:
+                from lerobot.processor.delta_action_processor import to_delta_actions
+
+                # Load state for delta even if we're not computing state stats
+                if needs_state and "observation.state" in ep_df.columns:
+                    state_values = ep_df["observation.state"].values
+                    if hasattr(state_values[0], "__len__"):
+                        states = np.stack(state_values)
+                    else:
+                        states = np.array(state_values)
+                    actions_t = torch.from_numpy(episode_data["action"]).float()
+                    states_t = torch.from_numpy(states).float()
+                    episode_data["action"] = to_delta_actions(actions_t, states_t, delta_mask).numpy()
+
+            ep_stats = compute_episode_stats(episode_data, features_to_compute)
+            all_episode_stats.append(ep_stats)
+
+    if not all_episode_stats:
+        logging.warning("No episode stats computed")
+        return dataset
+
+    new_stats = aggregate_stats(all_episode_stats)
+
+    # Merge: keep existing stats for features we didn't recompute
+    if dataset.meta.stats:
+        for key, value in dataset.meta.stats.items():
+            if key not in new_stats:
+                new_stats[key] = value
+
+    write_stats(new_stats, dataset.root)
+    dataset.meta.stats = new_stats
+
+    logging.info(f"Stats recomputed for {len(all_episode_stats)} episodes")
+    return dataset
+
+
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path,

src/lerobot/datasets/lerobot_dataset.py

@@ -18,22 +18,30 @@ import contextlib
 import logging
 import shutil
 import tempfile
+import time
 from collections.abc import Callable
 from pathlib import Path
 
 import datasets
 import numpy as np
+import os
 import packaging.version
 import pandas as pd
 import PIL.Image
 import pyarrow as pa
 import pyarrow.parquet as pq
+from concurrent.futures import ProcessPoolExecutor
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
+from lerobot.datasets.compute_stats import (
+    RunningQuantileStats,
+    aggregate_stats,
+    auto_downsample_height_width,
+    compute_episode_stats,
+)
 from lerobot.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
@@ -68,6 +76,7 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
+    StreamingVideoEncoder,
     VideoFrame,
     concatenate_video_files,
     decode_video_frames,
@@ -79,7 +88,6 @@ from lerobot.datasets.video_utils import (
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -420,8 +428,10 @@ class LeRobotDatasetMetadata:
 
         write_info(self.info, self.root)
 
+        t0 = time.perf_counter()
         self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
         write_stats(self.stats, self.root)
+        logging.info(f"[meta.save_episode] aggregate+write_stats: {time.perf_counter() - t0:.2f}s")
 
     def update_video_info(self, video_key: str | None = None) -> None:
         """
@@ -544,13 +554,11 @@ class LeRobotDatasetMetadata:
         return obj
 
 
-def _encode_video_worker(
-    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
-) -> Path:
+def _encode_video_worker(video_key: str, episode_index: int, root: Path, fps: int) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
+    encode_video_frames(img_dir, temp_path, fps, overwrite=True)
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -569,7 +577,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         download_videos: bool = True,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
-        vcodec: str = "libsvtav1",
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -682,13 +689,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
-            vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
-                encoding is CPU-heavy.
         """
         super().__init__()
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -700,7 +702,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = vcodec
 
         # Unused attributes
         self.image_writer = None
@@ -708,6 +709,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
+        self._streaming_encoder = None
+        self._running_video_stats = {}
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -938,30 +941,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
         else:
             return get_hf_features_from_features(self.features)
 
-    def _get_query_indices(
-        self, abs_idx: int, ep_idx: int
-    ) -> tuple[dict[str, list[int]], dict[str, torch.Tensor]]:
-        """Compute query indices for delta timestamps.
-
-        Args:
-            abs_idx: The absolute index in the full dataset (not the relative index in filtered episodes).
-            ep_idx: The episode index.
-
-        Returns:
-            A tuple of (query_indices, padding) where:
-            - query_indices: Dict mapping keys to lists of absolute indices to query
-            - padding: Dict mapping "{key}_is_pad" to boolean tensors indicating padded positions
-        """
+    def _get_query_indices(self, idx: int, ep_idx: int) -> tuple[dict[str, list[int | bool]]]:
         ep = self.meta.episodes[ep_idx]
         ep_start = ep["dataset_from_index"]
         ep_end = ep["dataset_to_index"]
         query_indices = {
-            key: [max(ep_start, min(ep_end - 1, abs_idx + delta)) for delta in delta_idx]
+            key: [max(ep_start, min(ep_end - 1, idx + delta)) for delta in delta_idx]
             for key, delta_idx in self.delta_indices.items()
         }
         padding = {  # Pad values outside of current episode range
             f"{key}_is_pad": torch.BoolTensor(
-                [(abs_idx + delta < ep_start) | (abs_idx + delta >= ep_end) for delta in delta_idx]
+                [(idx + delta < ep_start) | (idx + delta >= ep_end) for delta in delta_idx]
             )
             for key, delta_idx in self.delta_indices.items()
         }
@@ -1053,12 +1043,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self._ensure_hf_dataset_loaded()
         item = self.hf_dataset[idx]
         ep_idx = item["episode_index"].item()
-        # Use the absolute index from the dataset for delta timestamp calculations
-        abs_idx = item["index"].item()
 
         query_indices = None
         if self.delta_indices is not None:
-            query_indices, padding = self._get_query_indices(abs_idx, ep_idx)
+            query_indices, padding = self._get_query_indices(idx, ep_idx)
             query_result = self._query_hf_dataset(query_indices)
             item = {**item, **padding}
             for key, val in query_result.items():
@@ -1102,6 +1090,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Close the parquet writers. This function needs to be called after data collection/conversion, else footer metadata won't be written to the parquet files.
         The dataset won't be valid and can't be loaded as ds = LeRobotDataset(repo_id=repo, root=HF_LEROBOT_HOME.joinpath(repo))
         """
+        if self._streaming_encoder:
+            self._streaming_encoder.close()
         self._close_writer()
         self.meta._close_writer()
 
@@ -1153,6 +1143,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Automatically add frame_index and timestamp to episode buffer
         frame_index = self.episode_buffer["size"]
+        if frame_index == 0 and self._streaming_encoder:
+            self._streaming_encoder.start_episode(self.meta.video_keys, self.root)
+            self._init_running_video_stats()
         timestamp = frame.pop("timestamp") if "timestamp" in frame else frame_index / self.fps
         self.episode_buffer["frame_index"].append(frame_index)
         self.episode_buffer["timestamp"].append(timestamp)
@@ -1166,14 +1159,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 )
 
             if self.features[key]["dtype"] in ["image", "video"]:
-                img_path = self._get_image_file_path(
-                    episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
-                )
-                if frame_index == 0:
-                    img_path.parent.mkdir(parents=True, exist_ok=True)
-                compress_level = 1 if self.features[key]["dtype"] == "video" else 6
-                self._save_image(frame[key], img_path, compress_level)
-                self.episode_buffer[key].append(str(img_path))
+                if self._streaming_encoder and self.features[key]["dtype"] == "video":
+                    self._feed_streaming_frame(key, frame[key])
+                    self.episode_buffer[key].append(None)
+                else:
+                    img_path = self._get_image_file_path(
+                        episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
+                    )
+                    if frame_index == 0:
+                        img_path.parent.mkdir(parents=True, exist_ok=True)
+                    compress_level = 1 if self.features[key]["dtype"] == "video" else 6
+                    self._save_image(frame[key], img_path, compress_level)
+                    self.episode_buffer[key].append(str(img_path))
             else:
                 self.episode_buffer[key].append(frame[key])
 
@@ -1224,53 +1221,50 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 continue
             episode_buffer[key] = np.stack(episode_buffer[key])
 
-        # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-        ep_stats = compute_episode_stats(episode_buffer, self.features)
 
+        t0 = time.perf_counter()
+        if self._streaming_encoder:
+            filtered = {k: v for k, v in episode_buffer.items() if k not in self.meta.video_keys}
+            ep_stats = compute_episode_stats(filtered, self.features)
+            for key in self.meta.video_keys:
+                stats = self._running_video_stats[key].get_statistics()
+                ep_stats[key] = {
+                    k: v if k == "count" else (v.reshape(-1, 1, 1) / 255.0)
+                    for k, v in stats.items()
+                }
+        else:
+            ep_stats = compute_episode_stats(episode_buffer, self.features)
+        t_stats = time.perf_counter() - t0
+
+        t0 = time.perf_counter()
         ep_metadata = self._save_episode_data(episode_buffer)
+        t_save_data = time.perf_counter() - t0
+
         has_video_keys = len(self.meta.video_keys) > 0
         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,
-                            self.vcodec,
-                        ): 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))
+        t0 = time.perf_counter()
+        if has_video_keys and self._streaming_encoder:
+            video_paths = self._streaming_encoder.finish_episode()
+            for video_key in self.meta.video_keys:
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, video_paths[video_key]))
+        elif has_video_keys and not use_batched_encoding:
+            video_paths = self._encode_multiple_temporary_episode_videos(self.meta.video_keys, episode_index)
+            for video_key, video_path in zip(self.meta.video_keys, video_paths):
+                ep_metadata.update(self._save_episode_video(video_key, episode_index, video_path))
+        t_video = time.perf_counter() - t0
 
         # `meta.save_episode` need to be executed after encoding the videos
+        t0 = time.perf_counter()
         self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
+        t_meta = time.perf_counter() - t0
+
+        logging.info(
+            f"[save_episode] ep={episode_index} frames={episode_length} | "
+            f"stats={t_stats:.2f}s data={t_save_data:.2f}s video={t_video:.2f}s meta={t_meta:.2f}s "
+            f"total={t_stats + t_save_data + t_video + t_meta:.2f}s"
+        )
 
         if has_video_keys and use_batched_encoding:
             # Check if we should trigger batch encoding
@@ -1438,6 +1432,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episode_index: int,
         temp_path: Path | None = None,
     ) -> dict:
+        t0 = time.perf_counter()
         # Encode episode frames into a temporary video
         if temp_path is None:
             ep_path = self._encode_temporary_episode_video(video_key, episode_index)
@@ -1511,9 +1506,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
             f"videos/{video_key}/from_timestamp": latest_duration_in_s,
             f"videos/{video_key}/to_timestamp": latest_duration_in_s + ep_duration_in_s,
         }
+        save_time = time.perf_counter() - t0
+        rate = ep_duration_in_s / save_time if save_time > 0 else float("inf")
+        logging.info(
+            f"[save_episode_video] {video_key} ep={episode_index} "
+            f"save={save_time:.2f}s video_dur={ep_duration_in_s:.1f}s "
+            f"size={ep_size_in_mb:.1f}MB rate={rate:.2f}x realtime"
+        )
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
+        if self._streaming_encoder:
+            self._streaming_encoder.stop_episode()
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1522,7 +1526,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             episode_index = self.episode_buffer["episode_index"]
             if isinstance(episode_index, np.ndarray):
                 episode_index = episode_index.item() if episode_index.size == 1 else episode_index[0]
-            for cam_key in self.meta.image_keys:
+            for cam_key in self.meta.camera_keys:
                 img_dir = self._get_image_file_dir(episode_index, cam_key)
                 if img_dir.is_dir():
                     shutil.rmtree(img_dir)
@@ -1555,13 +1559,66 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
+    def start_streaming_encoder(self):
+        """Enable streaming video encoding for recording."""
+        if len(self.meta.video_keys) > 0:
+            self._streaming_encoder = StreamingVideoEncoder(fps=self.fps)
+            self._running_video_stats = {}
+
+    def _init_running_video_stats(self):
+        self._running_video_stats = {key: RunningQuantileStats() for key in self.meta.video_keys}
+
+    def _feed_streaming_frame(self, key: str, image) -> None:
+        """Feed image to streaming encoder and accumulate running stats."""
+        if isinstance(image, np.ndarray):
+            if image.ndim == 3 and image.shape[0] in (1, 3, 4):
+                img_chw = image
+            else:
+                img_chw = image.transpose(2, 0, 1)
+        else:
+            img_chw = np.array(image).transpose(2, 0, 1)
+
+        self._streaming_encoder.feed_frame(key, image)
+        img_ds = auto_downsample_height_width(img_chw)
+        c, h, w = img_ds.shape
+        self._running_video_stats[key].update(
+            img_ds.transpose(1, 2, 0).reshape(-1, c).astype(np.float64)
+        )
+
     def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps)
+
+    def _encode_multiple_temporary_episode_videos(self, video_keys, episode_index):
+        temp_paths = []
+        img_dirs = []
+        for video_key in video_keys:
+            temp_paths.append(Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4")
+            img_dirs.append(self._get_image_file_dir(episode_index, video_key))
+        fps = [self.fps]*len(video_keys)
+
+        t0 = time.perf_counter()
+        with ProcessPoolExecutor(max_workers=len(video_keys)) as executor:
+            executor.map(encode_video_frames,img_dirs,temp_paths,fps)
+        encode_time = time.perf_counter() - t0
+
+        n_frames = len(list(img_dirs[0].glob("*"))) if img_dirs and img_dirs[0].exists() else 0
+        video_duration_s = n_frames / self.fps if n_frames > 0 else 0
+        rate = video_duration_s / encode_time if encode_time > 0 else float("inf")
+        logging.info(
+            f"[encode_videos] ep={episode_index} keys={len(video_keys)} "
+            f"encode={encode_time:.2f}s video_dur={video_duration_s:.1f}s "
+            f"rate={rate:.2f}x realtime"
+        )
+
+        for img_dir in img_dirs:
+            shutil.rmtree(img_dir)
+
+        return temp_paths
 
     @classmethod
     def create(
@@ -1577,11 +1634,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_threads: int = 0,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
-        vcodec: str = "libsvtav1",
+        streaming_encoding: bool = False,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        if vcodec not in VALID_VIDEO_CODECS:
-            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1598,7 +1653,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_writer = None
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
-        obj.vcodec = vcodec
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1616,6 +1670,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.writer = None
         obj.latest_episode = None
         obj._current_file_start_frame = None
+        obj._streaming_encoder = None
+        obj._running_video_stats = {}
+        if streaming_encoding and len(obj.meta.video_keys) > 0:
+            obj._streaming_encoder = StreamingVideoEncoder(fps=fps)
         # Initialize tracking for incremental recording
         obj._lazy_loading = False
         obj._recorded_frames = 0

src/lerobot/datasets/video_utils.py

@@ -16,16 +16,18 @@
 import glob
 import importlib
 import logging
+import queue
 import shutil
 import tempfile
 import warnings
 from dataclasses import dataclass, field
 from pathlib import Path
-from threading import Lock
+from threading import Lock, Thread
 from typing import Any, ClassVar
 
 import av
 import fsspec
+import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
@@ -310,7 +312,7 @@ 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 +357,9 @@ 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"
@@ -397,6 +402,141 @@ def encode_video_frames(
         raise OSError(f"Video encoding did not work. File not found: {video_path}.")
 
 
+_DONE = object()
+
+
+class _CameraEncoder:
+    """Encodes frames for one camera in a daemon thread."""
+
+    def __init__(self, video_path, fps, vcodec, pix_fmt, g, crf):
+        self.video_path = Path(video_path)
+        self.fps = fps
+        self.vcodec = vcodec
+        self.pix_fmt = pix_fmt
+        self.g = g
+        self.crf = crf
+        self.queue = queue.Queue()
+        self._thread = None
+        self._cancelled = False
+
+    def start(self):
+        self.video_path.parent.mkdir(parents=True, exist_ok=True)
+        self._thread = Thread(target=self._run, daemon=True)
+        self._thread.start()
+
+    def finish(self) -> Path:
+        self.queue.put(_DONE)
+        self._thread.join(timeout=120)
+        return self.video_path
+
+    def cancel(self):
+        self._cancelled = True
+        while not self.queue.empty():
+            try:
+                self.queue.get_nowait()
+            except queue.Empty:
+                break
+        self.queue.put(_DONE)
+        if self._thread:
+            self._thread.join(timeout=5)
+        if self.video_path.parent.exists():
+            shutil.rmtree(self.video_path.parent, ignore_errors=True)
+
+    def _run(self):
+        options = {}
+        if self.g is not None:
+            options["g"] = str(self.g)
+        if self.crf is not None:
+            options["crf"] = str(self.crf)
+        if self.vcodec == "libsvtav1":
+            options["preset"] = "12"
+
+        output = None
+        output_stream = None
+        try:
+            while True:
+                data = self.queue.get()
+                if data is _DONE or self._cancelled:
+                    break
+
+                if isinstance(data, np.ndarray):
+                    if data.ndim == 3 and data.shape[0] in (1, 3, 4):
+                        data = data.transpose(1, 2, 0)
+                    pil = Image.fromarray(data.astype(np.uint8)).convert("RGB")
+                else:
+                    pil = data.convert("RGB")
+
+                if output is None:
+                    w, h = pil.size
+                    output = av.open(str(self.video_path), "w")
+                    output_stream = output.add_stream(self.vcodec, self.fps, options=options)
+                    output_stream.pix_fmt = self.pix_fmt
+                    output_stream.width = w
+                    output_stream.height = h
+
+                pkt = output_stream.encode(av.VideoFrame.from_image(pil))
+                if pkt:
+                    output.mux(pkt)
+
+            if output_stream and not self._cancelled:
+                pkt = output_stream.encode()
+                if pkt:
+                    output.mux(pkt)
+        except Exception as e:
+            logging.error(f"[StreamingEncoder] {e}")
+        finally:
+            if output:
+                output.close()
+
+
+class StreamingVideoEncoder:
+    """Encodes video on-the-fly using one background thread per camera.
+
+    PyAV releases the GIL during encoding, so Python threads give true
+    parallelism for the CPU-intensive codec work. The queue is unbounded
+    so feed_frame never blocks the caller (teleop thread always has priority).
+    """
+
+    def __init__(self, fps, vcodec="libsvtav1", pix_fmt="yuv420p", g=2, crf=30):
+        self.fps = fps
+        self._vcodec = vcodec
+        self._pix_fmt = pix_fmt
+        self._g = g
+        self._crf = crf
+        self._encoders: dict[str, _CameraEncoder] = {}
+
+    def start_episode(self, video_keys, temp_dir):
+        self.stop_episode()
+        for key in video_keys:
+            path = Path(tempfile.mkdtemp(dir=temp_dir)) / f"{key}_stream.mp4"
+            enc = _CameraEncoder(path, self.fps, self._vcodec, self._pix_fmt, self._g, self._crf)
+            enc.start()
+            self._encoders[key] = enc
+
+    def feed_frame(self, video_key, image):
+        """Non-blocking: put frame on unbounded queue (never blocks caller)."""
+        enc = self._encoders.get(video_key)
+        if enc:
+            enc.queue.put(image)
+
+    def finish_episode(self) -> dict[str, Path]:
+        """Flush all encoders, wait for completion, return {key: video_path}."""
+        paths = {}
+        for key, enc in self._encoders.items():
+            paths[key] = enc.finish()
+        self._encoders.clear()
+        return paths
+
+    def stop_episode(self):
+        """Cancel current episode encoding (for re-record)."""
+        for enc in self._encoders.values():
+            enc.cancel()
+        self._encoders.clear()
+
+    def close(self):
+        self.stop_episode()
+
+
 def concatenate_video_files(
     input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
 ):

src/lerobot/motors/__init__.py

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

src/lerobot/motors/damiao/tables.py

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

src/lerobot/motors/dynamixel/dynamixel.py

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

src/lerobot/motors/feetech/feetech.py

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

src/lerobot/policies/factory.py

@@ -470,6 +470,13 @@ def make_policy(
     cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
     if not cfg.input_features:
         cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
+
+    # Store action feature names for delta_exclude_joints support
+    if ds_meta is not None and hasattr(cfg, "action_feature_names"):
+        action_names = ds_meta.features.get(ACTION, {}).get("names")
+        if action_names is not None:
+            cfg.action_feature_names = list(action_names)
+
     kwargs["config"] = cfg
 
     # Pass dataset_stats to the policy if available (needed for some policies like SARM)

src/lerobot/policies/pi0/configuration_pi0.py

@@ -50,6 +50,13 @@ class PI0Config(PreTrainedConfig):
     min_period: float = 4e-3
     max_period: float = 4.0
 
+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+    # Joint names to exclude from delta (kept absolute). Empty list = all dims delta.
+    delta_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
+    # Populated at runtime from dataset metadata by make_policy.
+    action_feature_names: list[str] | None = None
+
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 

src/lerobot/policies/pi0/modeling_pi0.py

@@ -44,6 +44,12 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+from lerobot.policies.rtc.training_time import (
+    apply_rtc_training_time,
+    apply_training_time_rtc_inference,
+    masked_mean,
+    sample_rtc_delay,
+)
 from lerobot.utils.constants import (
     ACTION,
     OBS_LANGUAGE_ATTENTION_MASK,
@@ -79,8 +85,8 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
     if dimension % 2 != 0:
         raise ValueError(f"dimension ({dimension}) must be divisible by 2")
 
-    if time.ndim != 1:
-        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
+    if time.ndim not in (1, 2):
+        raise ValueError("The time tensor is expected to be of shape `(batch_size,)` or `(batch_size, T)`.")
 
     dtype = get_safe_dtype(torch.float64, device.type)
     fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
@@ -88,8 +94,14 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
     # Compute the outer product
     scaling_factor = 1.0 / period * 2 * math.pi
-    sin_input = scaling_factor[None, :] * time[:, None]
-    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    if time.ndim == 1:
+        sin_input = scaling_factor[None, :] * time[:, None]
+        return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+
+    time_flat = time.reshape(-1)
+    sin_input = scaling_factor[None, :] * time_flat[:, None]
+    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    return pos_emb.reshape(*time.shape, dimension)
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
@@ -605,6 +617,9 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
     def _rtc_enabled(self):
         return self.config.rtc_config is not None and self.config.rtc_config.enabled
 
+    def _training_time_rtc_inference_enabled(self):
+        return self.config.rtc_training_config is not None and self.config.rtc_training_config.enabled
+
     def _apply_checkpoint(self, func, *args, **kwargs):
         """Helper method to apply gradient checkpointing if enabled."""
         if self.gradient_checkpointing_enabled and self.training:
@@ -714,7 +729,10 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         action_emb = self._apply_checkpoint(action_proj_func, noisy_actions)
 
-        time_emb = time_emb[:, None, :].expand_as(action_emb)
+        if time_emb.dim() == 2:
+            time_emb = time_emb[:, None, :].expand_as(action_emb)
+        elif time_emb.shape[:2] != action_emb.shape[:2]:
+            raise ValueError(f"Expected time_emb shape {action_emb.shape[:2]}, got {time_emb.shape[:2]}")
         action_time_emb = torch.cat([action_emb, time_emb], dim=2)
 
         def mlp_func(action_time_emb):
@@ -750,7 +768,12 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         if time is None:
             time = self.sample_time(actions.shape[0], actions.device)
 
-        time_expanded = time[:, None, None]
+        if time.ndim == 1:
+            time_expanded = time[:, None, None]
+        elif time.ndim == 2:
+            time_expanded = time[:, :, None]
+        else:
+            raise ValueError(f"Expected time shape (B,) or (B, T), got {time.shape}")
         x_t = time_expanded * noise + (1 - time_expanded) * actions
         u_t = noise - actions
 
@@ -846,24 +869,37 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         dt = -1.0 / num_steps
 
+        inference_delay = kwargs.get("inference_delay")
+        prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+        execution_horizon = kwargs.get("execution_horizon")
+        use_training_time_rtc = self._training_time_rtc_inference_enabled()
+
         x_t = noise
         for step in range(num_steps):
             time = 1.0 + step * dt
-            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
-                return self.denoise_step(
+            if use_training_time_rtc:
+                x_t_cond, time_tensor = apply_training_time_rtc_inference(
+                    x_t, time, inference_delay, prev_chunk_left_over, self.config.chunk_size
+                )
+                v_t = self.denoise_step(
                     state=state,
                     prefix_pad_masks=prefix_pad_masks,
                     past_key_values=past_key_values,
-                    x_t=input_x_t,
-                    timestep=current_timestep,
+                    x_t=x_t_cond,
+                    timestep=time_tensor,
                 )
+            elif self._rtc_enabled():
+                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-            if self._rtc_enabled():
-                inference_delay = kwargs.get("inference_delay")
-                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
-                execution_horizon = kwargs.get("execution_horizon")
+                def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                    return self.denoise_step(
+                        state=state,
+                        prefix_pad_masks=prefix_pad_masks,
+                        past_key_values=past_key_values,
+                        x_t=input_x_t,
+                        timestep=current_timestep,
+                    )
 
                 v_t = self.rtc_processor.denoise_step(
                     x_t=x_t,
@@ -874,7 +910,14 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
                     execution_horizon=execution_horizon,
                 )
             else:
-                v_t = denoise_step_partial_call(x_t)
+                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+                v_t = self.denoise_step(
+                    state=state,
+                    prefix_pad_masks=prefix_pad_masks,
+                    past_key_values=past_key_values,
+                    x_t=x_t,
+                    timestep=time_tensor,
+                )
 
             x_t = x_t + dt * v_t
 
@@ -1277,7 +1320,19 @@ class PI0Policy(PreTrainedPolicy):
         actions = self.prepare_action(batch)
 
         # Compute loss
-        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions)
+        postfix_mask = None
+        rtc_cfg = self.config.rtc_training_config
+        if rtc_cfg is not None and rtc_cfg.enabled and self.training:
+            batch_size = actions.shape[0]
+            time = self.model.sample_time(batch_size, actions.device)
+            noise = self.model.sample_noise(actions.shape, actions.device)
+            delay = sample_rtc_delay(rtc_cfg, batch_size, actions.device)
+            time, postfix_mask = apply_rtc_training_time(time, delay, actions.shape[1])
+            losses = self.model.forward(
+                images, img_masks, lang_tokens, lang_masks, state, actions, noise=noise, time=time
+            )
+        else:
+            losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions)
 
         # Truncate losses to actual action dimensions
         original_action_dim = self.config.output_features[ACTION].shape[0]
@@ -1289,12 +1344,12 @@ class PI0Policy(PreTrainedPolicy):
 
         if reduction == "none":
             # Return per-sample losses (B,) by averaging over time and action dims
-            per_sample_loss = losses.mean(dim=(1, 2))
+            per_sample_loss = masked_mean(losses, postfix_mask, reduce_dims=(1, 2))
             loss_dict["loss"] = per_sample_loss.mean().item()
             return per_sample_loss, loss_dict
         else:
             # Default: return scalar mean loss
-            loss = losses.mean()
+            loss = masked_mean(losses, postfix_mask, reduce_dims=(0, 1, 2))
             loss_dict["loss"] = loss.item()
             return loss, loss_dict
 

src/lerobot/policies/pi0/processor_pi0.py

@@ -21,8 +21,10 @@ import torch
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
     AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
+    DeltaActionsProcessorStep,
     DeviceProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
@@ -126,7 +128,13 @@ def make_pi0_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
-    # Add remaining processors
+    delta_step = DeltaActionsProcessorStep(
+        enabled=config.use_delta_actions,
+        exclude_joints=getattr(config, "delta_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    # OpenPI order: raw → delta → normalize → model → unnormalize → absolute
     input_steps: list[ProcessorStep] = [
         RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
         AddBatchDimensionProcessorStep(),
@@ -138,6 +146,7 @@ def make_pi0_pre_post_processors(
             padding="max_length",
         ),
         DeviceProcessorStep(device=config.device),
+        delta_step,
         NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
@@ -149,6 +158,7 @@ def make_pi0_pre_post_processors(
         UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
+        AbsoluteActionsProcessorStep(enabled=config.use_delta_actions, delta_step=delta_step),
         DeviceProcessorStep(device="cpu"),
     ]
 

src/lerobot/policies/pi05/configuration_pi05.py

@@ -20,7 +20,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig, RTCTrainingConfig
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
 
 DEFAULT_IMAGE_SIZE = 224
@@ -50,8 +50,16 @@ class PI05Config(PreTrainedConfig):
     min_period: float = 4e-3
     max_period: float = 4.0
 
+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+    # Joint names to exclude from delta (kept absolute). Empty list = all dims delta.
+    delta_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
+    # Populated at runtime from dataset metadata by make_policy.
+    action_feature_names: list[str] | None = None
+
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
+    rtc_training_config: RTCTrainingConfig | None = None
 
     image_resolution: tuple[int, int] = (
         DEFAULT_IMAGE_SIZE,

src/lerobot/policies/pi05/modeling_pi05.py

@@ -44,6 +44,12 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+from lerobot.policies.rtc.training_time import (
+    apply_rtc_training_time,
+    apply_training_time_rtc_inference,
+    masked_mean,
+    sample_rtc_delay,
+)
 from lerobot.utils.constants import (
     ACTION,
     OBS_LANGUAGE_ATTENTION_MASK,
@@ -78,8 +84,8 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
     if dimension % 2 != 0:
         raise ValueError(f"dimension ({dimension}) must be divisible by 2")
 
-    if time.ndim != 1:
-        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
+    if time.ndim not in (1, 2):
+        raise ValueError("The time tensor is expected to be of shape `(batch_size,)` or `(batch_size, T)`.")
 
     dtype = get_safe_dtype(torch.float64, device.type)
     fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
@@ -87,8 +93,14 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
     # Compute the outer product
     scaling_factor = 1.0 / period * 2 * math.pi
-    sin_input = scaling_factor[None, :] * time[:, None]
-    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    if time.ndim == 1:
+        sin_input = scaling_factor[None, :] * time[:, None]
+        return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+
+    time_flat = time.reshape(-1)
+    sin_input = scaling_factor[None, :] * time_flat[:, None]
+    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    return pos_emb.reshape(*time.shape, dimension)
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
@@ -602,6 +614,9 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
     def _rtc_enabled(self):
         return self.config.rtc_config is not None and self.config.rtc_config.enabled
 
+    def _training_time_rtc_inference_enabled(self):
+        return self.config.rtc_training_config is not None and self.config.rtc_training_config.enabled
+
     def _apply_checkpoint(self, func, *args, **kwargs):
         """Helper method to apply gradient checkpointing if enabled."""
         if self.gradient_checkpointing_enabled and self.training:
@@ -729,7 +744,12 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         if time is None:
             time = self.sample_time(actions.shape[0], actions.device)
 
-        time_expanded = time[:, None, None]
+        if time.ndim == 1:
+            time_expanded = time[:, None, None]
+        elif time.ndim == 2:
+            time_expanded = time[:, :, None]
+        else:
+            raise ValueError(f"Expected time shape (B,) or (B, T), got {time.shape}")
         x_t = time_expanded * noise + (1 - time_expanded) * actions
         u_t = noise - actions
 
@@ -820,23 +840,35 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         dt = -1.0 / num_steps
 
+        inference_delay = kwargs.get("inference_delay")
+        prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+        execution_horizon = kwargs.get("execution_horizon")
+        use_training_time_rtc = self._training_time_rtc_inference_enabled()
+
         x_t = noise
         for step in range(num_steps):
             time = 1.0 + step * dt
-            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
-                return self.denoise_step(
+            if use_training_time_rtc:
+                x_t_cond, time_tensor = apply_training_time_rtc_inference(
+                    x_t, time, inference_delay, prev_chunk_left_over, self.config.chunk_size
+                )
+                v_t = self.denoise_step(
                     prefix_pad_masks=prefix_pad_masks,
                     past_key_values=past_key_values,
-                    x_t=input_x_t,
-                    timestep=current_timestep,
+                    x_t=x_t_cond,
+                    timestep=time_tensor,
                 )
+            elif self._rtc_enabled():
+                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
 
-            if self._rtc_enabled():
-                inference_delay = kwargs.get("inference_delay")
-                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
-                execution_horizon = kwargs.get("execution_horizon")
+                def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                    return self.denoise_step(
+                        prefix_pad_masks=prefix_pad_masks,
+                        past_key_values=past_key_values,
+                        x_t=input_x_t,
+                        timestep=current_timestep,
+                    )
 
                 v_t = self.rtc_processor.denoise_step(
                     x_t=x_t,
@@ -847,7 +879,13 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
                     execution_horizon=execution_horizon,
                 )
             else:
-                v_t = denoise_step_partial_call(x_t)
+                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+                v_t = self.denoise_step(
+                    prefix_pad_masks=prefix_pad_masks,
+                    past_key_values=past_key_values,
+                    x_t=x_t,
+                    timestep=time_tensor,
+                )
 
             x_t = x_t + dt * v_t
 
@@ -1250,7 +1288,17 @@ class PI05Policy(PreTrainedPolicy):
         actions = self.prepare_action(batch)
 
         # Compute loss (no separate state needed for PI05)
-        losses = self.model.forward(images, img_masks, tokens, masks, actions)
+        postfix_mask = None
+        rtc_cfg = self.config.rtc_training_config
+        if rtc_cfg is not None and rtc_cfg.enabled and self.training:
+            batch_size = actions.shape[0]
+            time = self.model.sample_time(batch_size, actions.device)
+            noise = self.model.sample_noise(actions.shape, actions.device)
+            delay = sample_rtc_delay(rtc_cfg, batch_size, actions.device)
+            time, postfix_mask = apply_rtc_training_time(time, delay, actions.shape[1])
+            losses = self.model.forward(images, img_masks, tokens, masks, actions, noise=noise, time=time)
+        else:
+            losses = self.model.forward(images, img_masks, tokens, masks, actions)
 
         # Truncate losses to actual action dimensions
         original_action_dim = self.config.output_features[ACTION].shape[0]
@@ -1262,12 +1310,12 @@ class PI05Policy(PreTrainedPolicy):
 
         if reduction == "none":
             # Return per-sample losses (B,) by averaging over time and action dims
-            per_sample_loss = losses.mean(dim=(1, 2))
+            per_sample_loss = masked_mean(losses, postfix_mask, reduce_dims=(1, 2))
             loss_dict["loss"] = per_sample_loss.mean().item()
             return per_sample_loss, loss_dict
         else:
             # Default: return scalar mean loss
-            loss = losses.mean()
+            loss = masked_mean(losses, postfix_mask, reduce_dims=(0, 1, 2))
             loss_dict["loss"] = loss.item()
             return loss, loss_dict
 

src/lerobot/policies/pi05/processor_pi05.py

@@ -25,7 +25,9 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pi05.modeling_pi05 import pad_vector
 from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
     AddBatchDimensionProcessorStep,
+    DeltaActionsProcessorStep,
     DeviceProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
@@ -129,10 +131,19 @@ def make_pi05_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
-    # Add remaining processors
+    delta_step = DeltaActionsProcessorStep(
+        enabled=config.use_delta_actions,
+        exclude_joints=getattr(config, "delta_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    # OpenPI order: raw → delta → normalize → model → unnormalize → absolute
+    # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
+    # because the tokenizer step expects normalized state in [-1, 1] range for discretization
     input_steps: list[ProcessorStep] = [
         RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
         AddBatchDimensionProcessorStep(),
+        delta_step,
         # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
         # because the tokenizer step expects normalized state in [-1, 1] range for discretization
         NormalizerProcessorStep(
@@ -154,6 +165,7 @@ def make_pi05_pre_post_processors(
         UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
+        AbsoluteActionsProcessorStep(enabled=config.use_delta_actions, delta_step=delta_step),
         DeviceProcessorStep(device="cpu"),
     ]
 

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -41,6 +41,9 @@ class PI0FastConfig(PreTrainedConfig):
     max_action_dim: int = 32
     max_action_tokens: int = 256
 
+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -48,12 +48,14 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+from lerobot.processor.delta_action_processor import to_absolute_actions
 from lerobot.utils.constants import (
     ACTION,
     ACTION_TOKEN_MASK,
     ACTION_TOKENS,
     OBS_LANGUAGE_ATTENTION_MASK,
     OBS_LANGUAGE_TOKENS,
+    OBS_STATE,
     OPENPI_ATTENTION_MASK_VALUE,
 )
 
@@ -1315,6 +1317,12 @@ class PI0FastPolicy(PreTrainedPolicy):
             action_tokens, action_horizon=action_horizon, action_dim=action_dim
         )
 
+        if self.config.use_delta_actions and OBS_STATE in batch:
+            state = pad_vector(batch[OBS_STATE], self.config.max_state_dim)
+            continuous_actions = to_absolute_actions(
+                continuous_actions, state, [True] * continuous_actions.shape[-1]
+        )
+
         return continuous_actions
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -27,6 +27,7 @@ from lerobot.policies.pi0_fast.modeling_pi0_fast import pad_vector
 from lerobot.processor import (
     ActionTokenizerProcessorStep,
     AddBatchDimensionProcessorStep,
+    DeltaActionsProcessorStep,
     DeviceProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
@@ -147,6 +148,7 @@ def make_pi0_fast_pre_post_processors(
             padding_side="right",
             padding="max_length",
         ),
+        DeltaActionsProcessorStep(enabled=config.use_delta_actions),
         ActionTokenizerProcessorStep(
             action_tokenizer_name=config.action_tokenizer_name,
             max_action_tokens=config.max_action_tokens,

src/lerobot/policies/rtc/configuration_rtc.py

@@ -23,7 +23,7 @@ Based on:
 
 from dataclasses import dataclass
 
-from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.configs.types import RTCAttentionSchedule, RTCTrainingDelayDistribution
 
 
 @dataclass
@@ -53,3 +53,22 @@ class RTCConfig:
             raise ValueError(f"max_guidance_weight must be positive, got {self.max_guidance_weight}")
         if self.debug_maxlen <= 0:
             raise ValueError(f"debug_maxlen must be positive, got {self.debug_maxlen}")
+
+
+@dataclass
+class RTCTrainingConfig:
+    """Configuration for training-time RTC action prefix conditioning."""
+
+    enabled: bool = False
+    min_delay: int = 0
+    max_delay: int = 0
+    delay_distribution: RTCTrainingDelayDistribution = RTCTrainingDelayDistribution.UNIFORM
+    exp_decay: float = 1.0
+
+    def __post_init__(self):
+        if self.min_delay < 0:
+            raise ValueError(f"min_delay must be >= 0, got {self.min_delay}")
+        if self.max_delay < self.min_delay:
+            raise ValueError(f"max_delay ({self.max_delay}) must be >= min_delay ({self.min_delay})")
+        if self.exp_decay <= 0:
+            raise ValueError(f"exp_decay must be positive, got {self.exp_decay}")

src/lerobot/policies/rtc/training_time.py

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

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -20,7 +20,7 @@ from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import (
     CosineDecayWithWarmupSchedulerConfig,
 )
-from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig, RTCTrainingConfig
 from lerobot.utils.constants import OBS_IMAGES
 
 
@@ -103,8 +103,9 @@ class SmolVLAConfig(PreTrainedConfig):
     min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding
     max_period: float = 4.0
 
-    # Real-Time Chunking (RTC) configuration
+    # Real-Time Chunking (RTC) configurations
     rtc_config: RTCConfig | None = None
+    rtc_training_config: RTCTrainingConfig | None = None
 
     def __post_init__(self):
         super().__post_init__()

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/processor/__init__.py

@@ -28,7 +28,14 @@ from .core import (
     RobotObservation,
     TransitionKey,
 )
-from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorToDeltaActionDictStep
+from .delta_action_processor import (
+    AbsoluteActionsProcessorStep,
+    DeltaActionsProcessorStep,
+    MapDeltaActionToRobotActionStep,
+    MapTensorToDeltaActionDictStep,
+    to_absolute_actions,
+    to_delta_actions,
+)
 from .device_processor import DeviceProcessorStep
 from .factory import (
     make_default_processors,
@@ -97,6 +104,8 @@ __all__ = [
     "make_default_teleop_action_processor",
     "make_default_robot_action_processor",
     "make_default_robot_observation_processor",
+    "AbsoluteActionsProcessorStep",
+    "DeltaActionsProcessorStep",
     "MapDeltaActionToRobotActionStep",
     "MapTensorToDeltaActionDictStep",
     "NormalizerProcessorStep",
@@ -126,6 +135,8 @@ __all__ = [
     "transition_to_batch",
     "TransitionKey",
     "TruncatedProcessorStep",
+    "to_absolute_actions",
+    "to_delta_actions",
     "UnnormalizerProcessorStep",
     "VanillaObservationProcessorStep",
 ]

src/lerobot/processor/delta_action_processor.py

@@ -14,12 +14,54 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from dataclasses import dataclass
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+import torch
+from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.utils.constants import OBS_STATE
 
-from .core import PolicyAction, RobotAction
-from .pipeline import ActionProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
+from .core import EnvTransition, PolicyAction, RobotAction, TransitionKey
+from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
+
+
+def to_delta_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) -> Tensor:
+    """Convert absolute actions to delta: delta = action - state (for masked dims).
+
+    Args:
+        actions: (B, T, action_dim) or (B, action_dim).
+        state: (B, state_dim). Broadcast across time dimension.
+        mask: Which dims to convert. Can be shorter than action_dim.
+    """
+    mask_t = torch.tensor(mask, dtype=actions.dtype, device=actions.device)
+    dims = mask_t.shape[0]
+    state_offset = state[..., :dims] * mask_t
+    if actions.ndim == 3:
+        state_offset = state_offset.unsqueeze(-2)
+    actions = actions.clone()
+    actions[..., :dims] -= state_offset
+    return actions
+
+
+def to_absolute_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) -> Tensor:
+    """Convert delta actions back to absolute: absolute = delta + state (for masked dims).
+
+    Args:
+        actions: (B, T, action_dim) or (B, action_dim).
+        state: (B, state_dim). Broadcast across time dimension.
+        mask: Which dims to convert. Can be shorter than action_dim.
+    """
+    mask_t = torch.tensor(mask, dtype=actions.dtype, device=actions.device)
+    dims = mask_t.shape[0]
+    state_offset = state[..., :dims] * mask_t
+    if actions.ndim == 3:
+        state_offset = state_offset.unsqueeze(-2)
+    actions = actions.clone()
+    actions[..., :dims] += state_offset
+    return actions
 
 
 @ProcessorStepRegistry.register("map_tensor_to_delta_action_dict")
@@ -141,3 +183,126 @@ class MapDeltaActionToRobotActionStep(RobotActionProcessorStep):
             )
 
         return features
+
+
+@ProcessorStepRegistry.register("delta_actions_processor")
+@dataclass
+class DeltaActionsProcessorStep(ProcessorStep):
+    """Converts absolute actions to delta actions (action -= state) for masked dimensions.
+
+    Mirrors OpenPI's DeltaActions transform. Applied during preprocessing so the model
+    trains on relative offsets instead of absolute positions.
+    Caches the last seen state so a paired AbsoluteActionsProcessorStep can reverse
+    the conversion during postprocessing.
+
+    Attributes:
+        enabled: Whether to apply the delta conversion.
+        exclude_joints: Joint names to keep absolute (not converted to delta).
+        action_names: Action dimension names from dataset metadata, used to build
+            the mask from exclude_joints. If None, all dims are converted.
+    """
+
+    enabled: bool = False
+    exclude_joints: list[str] = field(default_factory=list)
+    action_names: list[str] | None = None
+    _last_state: torch.Tensor | None = field(default=None, init=False, repr=False)
+
+    def _build_mask(self, action_dim: int) -> list[bool]:
+        if not self.exclude_joints or self.action_names is None:
+            return [True] * action_dim
+
+        exclude_tokens = [str(name).lower() for name in self.exclude_joints if name]
+        if not exclude_tokens:
+            return [True] * action_dim
+
+        mask = []
+        for name in self.action_names[:action_dim]:
+            action_name = str(name).lower()
+            is_excluded = any(token == action_name or token in action_name for token in exclude_tokens)
+            mask.append(not is_excluded)
+
+        if len(mask) < action_dim:
+            mask.extend([True] * (action_dim - len(mask)))
+
+        return mask
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = transition.get(TransitionKey.OBSERVATION, {})
+        state = observation.get(OBS_STATE) if observation else None
+
+        # Always cache state for the paired AbsoluteActionsProcessorStep
+        if state is not None:
+            self._last_state = state
+
+        if not self.enabled:
+            return transition
+
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None or state is None:
+            return new_transition
+
+        mask = self._build_mask(action.shape[-1])
+        new_transition[TransitionKey.ACTION] = to_delta_actions(action, state, mask)
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {"enabled": self.enabled, "exclude_joints": self.exclude_joints}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("absolute_actions_processor")
+@dataclass
+class AbsoluteActionsProcessorStep(ProcessorStep):
+    """Converts delta actions back to absolute actions (action += state) for all dimensions.
+
+    Mirrors OpenPI's AbsoluteActions transform. Applied during postprocessing so
+    predicted deltas are converted back to absolute positions for execution.
+    Reads the cached state from its paired DeltaActionsProcessorStep.
+
+    Attributes:
+        enabled: Whether to apply the absolute conversion.
+        delta_step: Reference to the paired DeltaActionsProcessorStep that caches state.
+    """
+
+    enabled: bool = False
+    delta_step: DeltaActionsProcessorStep | None = field(default=None, repr=False)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        if not self.enabled:
+            return transition
+
+        if self.delta_step is None:
+            raise RuntimeError(
+                "AbsoluteActionsProcessorStep requires a paired DeltaActionsProcessorStep "
+                "but delta_step is None. Ensure delta_step is set when constructing the postprocessor."
+            )
+
+        if self.delta_step._last_state is None:
+            raise RuntimeError(
+                "AbsoluteActionsProcessorStep requires state from DeltaActionsProcessorStep "
+                "but no state has been cached. Ensure the preprocessor runs before the postprocessor."
+            )
+
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None:
+            return new_transition
+
+        mask = self.delta_step._build_mask(action.shape[-1])
+        new_transition[TransitionKey.ACTION] = to_absolute_actions(
+            action, self.delta_step._last_state, mask
+        )
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {"enabled": self.enabled}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/processor/normalize_processor.py

@@ -331,11 +331,9 @@ class _NormalizationMixin:
                 )
 
             mean, std = stats["mean"], stats["std"]
-            # Avoid division by zero by adding a small epsilon.
-            denom = std + self.eps
             if inverse:
-                return tensor * std + mean
-            return (tensor - mean) / denom
+                return tensor * (std + 1e-6) + mean
+            return (tensor - mean) / (std + 1e-6)
 
         if norm_mode == NormalizationMode.MIN_MAX:
             min_val = stats.get("min", None)
@@ -367,11 +365,7 @@ class _NormalizationMixin:
                     "QUANTILES normalization mode requires q01 and q99 stats, please update the dataset with the correct stats using the `augment_dataset_quantile_stats.py` script"
                 )
 
-            denom = q99 - q01
-            # Avoid division by zero by adding epsilon when quantiles are identical
-            denom = torch.where(
-                denom == 0, torch.tensor(self.eps, device=tensor.device, dtype=tensor.dtype), denom
-            )
+            denom = q99 - q01 + 1e-6
             if inverse:
                 return (tensor + 1.0) * denom / 2.0 + q01
             return 2.0 * (tensor - q01) / denom - 1.0

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 +1,120 @@
+#!/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])
+
+    #position_kp: list[float] = field(default_factory=lambda: [200.0, 200.0, 240.0, 200.0, 24.0, 31.0, 25.0, 25.0])
+    
+    # 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_mirror_dataset.py

@@ -0,0 +1,366 @@
+# 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.
+
+"""
+Mirror a bimanual robot dataset by swapping left/right arms and inverting joint values.
+
+This script creates a mirrored version of a dataset where:
+1. Left and right arm observations/actions are swapped
+2. Joint values are inverted according to a mirroring mask
+3. Video frames are horizontally flipped
+
+Example usage:
+```shell
+python -m lerobot.scripts.lerobot_mirror_dataset \
+    --repo_id=pepijn/openarm_bimanual \
+    --output_repo_id=pepijn/openarm_bimanual_mirrored
+```
+"""
+
+import argparse
+import logging
+import os
+import subprocess
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import (
+    DATA_DIR,
+    DEFAULT_DATA_PATH,
+    write_info,
+    write_stats,
+    write_tasks,
+)
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+logger = logging.getLogger(__name__)
+
+OPENARM_MIRRORING_MASK = {
+    "joint_1": -1,  # Pan - invert
+    "joint_2": -1,  # Lift - invert
+    "joint_3": -1,  # Roll - invert
+    "joint_4": 1,   # Elbow - no invert
+    "joint_5": -1,  # W-Roll - invert
+    "joint_6": -1,  # W-Pitch - invert
+    "joint_7": -1,  # W-Yaw - invert
+    "gripper": 1,   # Gripper - no invert
+}
+
+
+def get_mirroring_mask(robot_type: str) -> dict[str, int]:
+    """Get the mirroring mask for a given robot type."""
+    if robot_type in ["bi_openarm_follower", "openarm_follower", "bi_openarms_follower", "openarms_follower"]:
+        return OPENARM_MIRRORING_MASK
+    raise ValueError(f"Unknown robot type: {robot_type}. Add a mirroring mask for this robot.")
+
+
+def swap_left_right_name(name: str) -> str:
+    """Swap 'left' and 'right' in a feature name."""
+    # Use placeholder to avoid double-swap
+    result = name.replace("left_", "LEFT_PLACEHOLDER_")
+    result = result.replace("right_", "left_")
+    result = result.replace("LEFT_PLACEHOLDER_", "right_")
+    return result
+
+
+def mirror_feature_names(names: list[str]) -> tuple[list[str], dict[int, int]]:
+    """Mirror feature names by swapping left/right and return the new names and index mapping."""
+    mirrored_names = [swap_left_right_name(n) for n in names]
+    old_to_new_idx = {}
+    for old_idx, old_name in enumerate(names):
+        new_name = swap_left_right_name(old_name)
+        new_idx = mirrored_names.index(new_name)
+        old_to_new_idx[old_idx] = new_idx
+    return mirrored_names, old_to_new_idx
+
+
+def apply_mirroring_mask(
+    value: float,
+    feature_name: str,
+    mirroring_mask: dict[str, int],
+) -> float:
+    """Apply mirroring mask to a joint value."""
+    name_without_prefix = feature_name.split("_", 1)[1] if "_" in feature_name else feature_name
+    joint_name = name_without_prefix.split(".")[0]
+    if joint_name in mirroring_mask:
+        return value * mirroring_mask[joint_name]
+    return value
+
+
+def mirror_array(
+    array: np.ndarray,
+    names: list[str],
+    mirroring_mask: dict[str, int],
+) -> np.ndarray:
+    """Mirror an array of values (action or state) by swapping left/right and applying mask."""
+    mirrored_names, idx_mapping = mirror_feature_names(names)
+    result = np.zeros_like(array)
+    for old_idx, new_idx in idx_mapping.items():
+        old_name = names[old_idx]
+        new_name = mirrored_names[new_idx]
+        value = array[old_idx]
+        mirrored_value = apply_mirroring_mask(value, new_name, mirroring_mask)
+        result[new_idx] = mirrored_value
+    return result
+
+
+def flip_video_frames(
+    input_path: Path,
+    output_path: Path,
+    fps: float,
+    vcodec: str = "libsvtav1",
+):
+    """Flip video frames horizontally using FFmpeg with same settings as encode_video_frames."""
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    cmd = [
+        "ffmpeg", "-y", "-i", str(input_path),
+        "-vf", "hflip",
+        "-c:v", vcodec,
+        "-g", "2",
+        "-crf", "30",
+        "-r", str(int(fps)),
+        "-pix_fmt", "yuv420p",
+        "-loglevel", "error",
+    ]
+    if vcodec == "libsvtav1":
+        cmd.extend(["-preset", "12"])
+    cmd.append(str(output_path))
+    result = subprocess.run(cmd, capture_output=True, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"FFmpeg failed: {result.stderr}")
+
+
+def mirror_dataset(
+    repo_id: str,
+    output_repo_id: str,
+    root: str | Path | None = None,
+    output_root: str | Path | None = None,
+    mirroring_mask: dict[str, int] | None = None,
+    vcodec: str = "libsvtav1",
+    num_workers: int | None = None,
+) -> LeRobotDataset:
+    """Mirror a bimanual robot dataset."""
+    logger.info(f"Loading dataset: {repo_id}")
+    dataset = LeRobotDataset(repo_id, root=root)
+
+    if mirroring_mask is None:
+        robot_type = dataset.meta.robot_type or "bi_openarms_follower"
+        mirroring_mask = get_mirroring_mask(robot_type)
+        logger.info(f"Using mirroring mask for robot type: {robot_type}")
+
+    output_root = Path(output_root) if output_root else HF_LEROBOT_HOME / output_repo_id
+
+    mirrored_features = {}
+    for key, feat in dataset.meta.features.items():
+        new_key = swap_left_right_name(key)
+        new_feat = feat.copy()
+        if "names" in new_feat and new_feat["names"]:
+            new_feat["names"] = [swap_left_right_name(n) for n in new_feat["names"]]
+        mirrored_features[new_key] = new_feat
+
+    logger.info("Creating mirrored dataset metadata...")
+    new_meta = LeRobotDatasetMetadata.create(
+        repo_id=output_repo_id,
+        fps=dataset.meta.fps,
+        features=mirrored_features,
+        robot_type=dataset.meta.robot_type,
+        root=output_root,
+        use_videos=len(dataset.meta.video_keys) > 0,
+    )
+
+    if dataset.meta.tasks is not None:
+        write_tasks(dataset.meta.tasks, new_meta.root)
+        new_meta.tasks = dataset.meta.tasks.copy()
+
+    _mirror_data(dataset, new_meta, mirroring_mask)
+    _mirror_videos(dataset, new_meta, vcodec, num_workers)
+    _copy_episodes_metadata(dataset, new_meta)
+
+    logger.info(f"Mirrored dataset saved to: {output_root}")
+    return LeRobotDataset(output_repo_id, root=output_root)
+
+
+def _mirror_data(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    mirroring_mask: dict[str, int],
+) -> None:
+    """Mirror parquet data files."""
+    data_dir = src_dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    action_names = src_dataset.meta.features.get("action", {}).get("names", [])
+    state_names = src_dataset.meta.features.get("observation.state", {}).get("names", [])
+
+    for src_path in tqdm(parquet_files, desc="Mirroring data files"):
+        df = pd.read_parquet(src_path).reset_index(drop=True)
+        relative_path = src_path.relative_to(src_dataset.root)
+        chunk_dir = relative_path.parts[1]
+        file_name = relative_path.parts[2]
+        chunk_idx = int(chunk_dir.split("-")[1])
+        file_idx = int(file_name.split("-")[1].split(".")[0])
+
+        if "action" in df.columns and action_names:
+            actions = np.stack(df["action"].values)
+            mirrored_actions = np.array([
+                mirror_array(row, action_names, mirroring_mask) for row in actions
+            ])
+            df["action"] = list(mirrored_actions)
+
+        if "observation.state" in df.columns and state_names:
+            states = np.stack(df["observation.state"].values)
+            mirrored_states = np.array([
+                mirror_array(row, state_names, mirroring_mask) for row in states
+            ])
+            df["observation.state"] = list(mirrored_states)
+
+        dst_path = dst_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        dst_path.parent.mkdir(parents=True, exist_ok=True)
+        df.to_parquet(dst_path, index=False)
+
+
+def _mirror_videos(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    vcodec: str,
+    num_workers: int | None = None,
+) -> None:
+    """Mirror video files by flipping horizontally and swapping left/right names."""
+    if not src_dataset.meta.video_keys:
+        return
+
+    video_tasks = []
+    for old_video_key in src_dataset.meta.video_keys:
+        new_video_key = swap_left_right_name(old_video_key)
+        for ep_idx in range(src_dataset.meta.total_episodes):
+            try:
+                src_path = src_dataset.root / src_dataset.meta.get_video_file_path(ep_idx, old_video_key)
+                dst_relative = src_dataset.meta.get_video_file_path(ep_idx, old_video_key)
+                dst_relative_str = str(dst_relative).replace(old_video_key, new_video_key)
+                dst_path = dst_meta.root / dst_relative_str
+                if src_path.exists():
+                    video_tasks.append((src_path, dst_path))
+            except KeyError:
+                continue
+
+    def process_video(task, pbar):
+        src_path, dst_path = task
+        pbar.set_postfix_str(src_path.name)
+        flip_video_frames(src_path, dst_path, src_dataset.meta.fps, vcodec)
+        return src_path
+
+    if num_workers is None:
+        num_workers = os.cpu_count() or 16
+    num_workers = min(len(video_tasks), num_workers)
+    logger.info(f"Processing {len(video_tasks)} videos with {num_workers} workers")
+    with tqdm(total=len(video_tasks), desc="Mirroring videos") as pbar:
+        with ThreadPoolExecutor(max_workers=num_workers) as executor:
+            futures = {executor.submit(process_video, t, pbar): t for t in video_tasks}
+            for future in as_completed(futures):
+                task = futures[future]
+                future.result()
+                pbar.set_postfix_str(f"done: {task[0].name}")
+                pbar.update(1)
+
+
+def _copy_episodes_metadata(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+) -> None:
+    """Copy episodes metadata with swapped video keys."""
+    episodes_dir = src_dataset.root / "meta/episodes"
+    dst_episodes_dir = dst_meta.root / "meta/episodes"
+
+    if episodes_dir.exists():
+        dst_episodes_dir.mkdir(parents=True, exist_ok=True)
+        for src_parquet in episodes_dir.glob("*/*.parquet"):
+            df = pd.read_parquet(src_parquet)
+            columns_to_rename = {}
+            for col in df.columns:
+                if col.startswith("videos/"):
+                    parts = col.split("/")
+                    if len(parts) >= 2:
+                        video_key = parts[1]
+                        new_video_key = swap_left_right_name(video_key)
+                        new_col = col.replace(f"videos/{video_key}/", f"videos/{new_video_key}/")
+                        columns_to_rename[col] = new_col
+            if columns_to_rename:
+                df = df.rename(columns=columns_to_rename)
+            dst_parquet = dst_episodes_dir / src_parquet.relative_to(episodes_dir)
+            dst_parquet.parent.mkdir(parents=True, exist_ok=True)
+            df.to_parquet(dst_parquet, index=False)
+
+    dst_meta.info.update({
+        "total_episodes": src_dataset.meta.total_episodes,
+        "total_frames": src_dataset.meta.total_frames,
+        "total_tasks": src_dataset.meta.total_tasks,
+        "total_videos": src_dataset.meta.total_videos,
+        "total_chunks": src_dataset.meta.total_chunks,
+    })
+    write_info(dst_meta.info, dst_meta.root)
+
+    if src_dataset.meta.stats is not None:
+        mirrored_stats = _mirror_stats(src_dataset.meta.stats)
+        write_stats(mirrored_stats, dst_meta.root)
+
+
+def _mirror_stats(stats: dict) -> dict:
+    """Mirror stats by swapping left/right in feature names."""
+    mirrored = {}
+    for key, value in stats.items():
+        new_key = swap_left_right_name(key)
+        if isinstance(value, dict):
+            mirrored[new_key] = _mirror_stats(value)
+        else:
+            mirrored[new_key] = value
+    return mirrored
+
+
+def main():
+    logging.basicConfig(level=logging.INFO)
+    parser = argparse.ArgumentParser(description="Mirror a bimanual robot dataset")
+    parser.add_argument("--repo_id", type=str, required=True, help="Source dataset repo_id")
+    parser.add_argument("--output_repo_id", type=str, required=True, help="Output dataset repo_id")
+    parser.add_argument("--root", type=str, default=None, help="Source dataset root directory")
+    parser.add_argument("--output_root", type=str, default=None, help="Output dataset root directory")
+    parser.add_argument("--vcodec", type=str, default="libsvtav1", help="Video codec (libsvtav1, h264, hevc)")
+    parser.add_argument("--num_workers", type=int, default=None, help="Number of parallel workers for video processing")
+    parser.add_argument("--push-to-hub", action="store_true", help="Push mirrored dataset to HuggingFace Hub")
+    args = parser.parse_args()
+
+    dataset = mirror_dataset(
+        repo_id=args.repo_id,
+        output_repo_id=args.output_repo_id,
+        root=args.root,
+        output_root=args.output_root,
+        vcodec=args.vcodec,
+        num_workers=args.num_workers,
+    )
+
+    if getattr(args, "push_to_hub", False):
+        logger.info(f"Pushing dataset to HuggingFace Hub: {args.output_repo_id}")
+        dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()
+

src/lerobot/scripts/lerobot_record.py

@@ -182,6 +182,10 @@ class DatasetRecordConfig:
     # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1'.
     # Use 'h264' for faster encoding on systems where AV1 encoding is CPU-heavy.
     vcodec: str = "libsvtav1"
+    # Encode video on-the-fly during recording using background threads (one per camera).
+    # Eliminates PNG writing and post-episode encoding, reducing save_episode from ~79s to <1s.
+    # Set to False on weaker PCs to fall back to the old post-episode encoding pipeline.
+    streaming_encoding: bool = True
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
 
@@ -446,6 +450,8 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 batch_encoding_size=cfg.dataset.video_encoding_batch_size,
                 vcodec=cfg.dataset.vcodec,
             )
+            if cfg.dataset.streaming_encoding:
+                dataset.start_streaming_encoder()
 
             if hasattr(robot, "cameras") and len(robot.cameras) > 0:
                 dataset.start_image_writer(
@@ -467,6 +473,7 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
                 batch_encoding_size=cfg.dataset.video_encoding_batch_size,
                 vcodec=cfg.dataset.vcodec,
+                streaming_encoding=cfg.dataset.streaming_encoding,
             )
 
         # Load pretrained policy

src/lerobot/scripts/lerobot_setup_motors.py

@@ -30,20 +30,21 @@ import draccus
 
 from lerobot.robots import (  # noqa: F401
     RobotConfig,
-    bi_so_follower,
     koch_follower,
     lekiwi,
     make_robot_from_config,
     omx_follower,
-    so_follower,
+    so100_follower,
+    so101_follower,
 )
 from lerobot.teleoperators import (  # noqa: F401
     TeleoperatorConfig,
-    bi_so_leader,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
-    so_leader,
+    so100_leader,
+    so101_leader,
+    openarms_mini
 )
 
 COMPATIBLE_DEVICES = [
@@ -56,6 +57,7 @@ COMPATIBLE_DEVICES = [
     "so101_follower",
     "so101_leader",
     "lekiwi",
+    "openarms_mini",
 ]
 
 

src/lerobot/scripts/lerobot_train.py

@@ -175,8 +175,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         from accelerate.utils import DistributedDataParallelKwargs
 
         ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
-        # Accelerate auto-detects the device based on the available hardware and ignores the policy.device setting.
-        # Force the device to be CPU when policy.device is set to CPU.
         force_cpu = cfg.policy.device == "cpu"
         accelerator = Accelerator(
             step_scheduler_with_optimizer=False,
@@ -211,16 +209,98 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     torch.backends.cuda.matmul.allow_tf32 = True
 
     # Dataset loading synchronization: main process downloads first to avoid race conditions
+    delta_action_stats = None
     if is_main_process:
         logging.info("Creating dataset")
         dataset = make_dataset(cfg)
 
+        # Compute delta action stats BEFORE distributed sync to avoid NCCL timeout
+        if getattr(cfg.policy, "use_delta_actions", False):
+            import numpy as np
+
+            from lerobot.datasets.compute_stats import get_feature_stats
+            from lerobot.processor.delta_action_processor import DeltaActionsProcessorStep, to_delta_actions
+
+            chunk_size = cfg.policy.chunk_size
+            hf = dataset.hf_dataset
+            total_frames = len(hf)
+            sample_upper_bound = total_frames - chunk_size
+            if sample_upper_bound <= 0:
+                raise ValueError(
+                    f"Cannot compute delta action stats: total_frames={total_frames}, chunk_size={chunk_size}"
+                )
+
+            max_samples = min(100_000, sample_upper_bound)
+            indices = np.random.choice(sample_upper_bound, max_samples, replace=False)
+
+            action_names = dataset.meta.features.get("action", {}).get("names")
+            delta_mask_step = DeltaActionsProcessorStep(
+                enabled=True,
+                exclude_joints=getattr(cfg.policy, "delta_exclude_joints", []),
+                action_names=action_names,
+            )
+            delta_mask = delta_mask_step._build_mask(dataset.meta.features["action"]["shape"][0])
+            logging.info(
+                f"use_delta_actions is enabled — computing delta action stats "
+                f"from {max_samples} chunk samples (chunk_size={chunk_size})"
+            )
+
+            all_delta_actions = []
+            episode_indices = np.array(hf["episode_index"])
+            for idx in indices:
+                idx = int(idx)
+                ep_idx = episode_indices[idx]
+                end_idx = min(idx + chunk_size, total_frames)
+                if end_idx > idx and episode_indices[end_idx - 1] != ep_idx:
+                    continue
+
+                chunk_data = hf[idx:end_idx]
+                actions = torch.tensor(np.stack([np.asarray(a) for a in chunk_data["action"]])).float()
+                state = torch.tensor(np.asarray(chunk_data["observation.state"][0])).float()
+
+                delta = to_delta_actions(actions.unsqueeze(0), state.unsqueeze(0), delta_mask).squeeze(0)
+                all_delta_actions.append(delta.numpy())
+
+            if not all_delta_actions:
+                raise RuntimeError("Failed to compute delta action stats: no valid chunks found.")
+
+            all_delta = np.concatenate(all_delta_actions, axis=0)
+            delta_stats = get_feature_stats(all_delta, axis=0, keepdims=all_delta.ndim == 1)
+            delta_action_stats = delta_stats
+            dataset.meta.stats["action"] = delta_action_stats
+
+            norm_type = "UNKNOWN"
+            if hasattr(cfg.policy, "normalization_mapping"):
+                from lerobot.configs.types import NormalizationMode
+                action_norm = cfg.policy.normalization_mapping.get("ACTION", None)
+                norm_type = action_norm.value if action_norm else "UNKNOWN"
+
+            excluded_dims = len(delta_mask) - sum(delta_mask)
+            logging.info(
+                f"Delta action stats ({len(all_delta_actions)} chunks, {len(all_delta)} values, norm={norm_type}): "
+                f"delta_dims={sum(delta_mask)}/{len(delta_mask)} (excluded={excluded_dims}), "
+                f"mean={np.abs(delta_stats['mean']).mean():.4f}, std={delta_stats['std'].mean():.4f}, "
+                f"q01={delta_stats['q01'].mean():.4f}, q99={delta_stats['q99'].mean():.4f}"
+            )
+            if norm_type == "QUANTILES":
+                q_range = (delta_stats['q99'] - delta_stats['q01']).mean()
+                logging.info(f"  Quantile range (q99-q01): {q_range:.4f}")
+
     accelerator.wait_for_everyone()
 
     # Now all other processes can safely load the dataset
     if not is_main_process:
         dataset = make_dataset(cfg)
 
+    # Ensure all ranks use the exact same delta action stats.
+    if getattr(cfg.policy, "use_delta_actions", False):
+        if accelerator.num_processes > 1 and torch.distributed.is_initialized():
+            stats_list = [delta_action_stats]
+            torch.distributed.broadcast_object_list(stats_list, src=0)
+            delta_action_stats = stats_list[0]
+        if delta_action_stats is not None:
+            dataset.meta.stats["action"] = delta_action_stats
+
     # Create environment used for evaluating checkpoints during training on simulation data.
     # On real-world data, no need to create an environment as evaluations are done outside train.py,
     # using the eval.py instead, with gym_dora environment and dora-rs.
@@ -246,10 +326,22 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     # Wait for all processes to finish policy creation before continuing
     accelerator.wait_for_everyone()
 
+    processor_pretrained_path = cfg.policy.pretrained_path
+    if (
+        getattr(cfg.policy, "use_delta_actions", False)
+        and processor_pretrained_path is not None
+        and not cfg.resume
+    ):
+        logging.warning(
+            "use_delta_actions=true with pretrained processors can skip delta transforms if "
+            "the checkpoint processors do not define them. Building processors from current policy config."
+        )
+        processor_pretrained_path = None
+
     # Create processors - only provide dataset_stats if not resuming from saved processors
     processor_kwargs = {}
     postprocessor_kwargs = {}
-    if (cfg.policy.pretrained_path and not cfg.resume) or not cfg.policy.pretrained_path:
+    if (processor_pretrained_path and not cfg.resume) or not processor_pretrained_path:
         # Only provide dataset_stats when not resuming from saved processor state
         processor_kwargs["dataset_stats"] = dataset.meta.stats
 
@@ -257,7 +349,7 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     if cfg.policy.type == "sarm":
         processor_kwargs["dataset_meta"] = dataset.meta
 
-    if cfg.policy.pretrained_path is not None:
+    if processor_pretrained_path is not None:
         processor_kwargs["preprocessor_overrides"] = {
             "device_processor": {"device": device.type},
             "normalizer_processor": {
@@ -279,7 +371,7 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
 
     preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=cfg.policy,
-        pretrained_path=cfg.policy.pretrained_path,
+        pretrained_path=processor_pretrained_path,
         **processor_kwargs,
         **postprocessor_kwargs,
     )
@@ -397,7 +489,36 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     for _ in range(step, cfg.steps):
         start_time = time.perf_counter()
         batch = next(dl_iter)
+
+        # Debug logging for first few steps and periodically
+        if is_main_process and (step < 3 or (cfg.log_freq > 0 and step % (cfg.log_freq * 10) == 0)):
+            action = batch.get("action")
+            state = batch.get("observation.state")
+            if action is not None and state is not None:
+                logging.info(
+                    f"[DEBUG step={step}] PRE-PROCESSOR — "
+                    f"action: shape={tuple(action.shape)}, mean={action.mean():.4f}, std={action.std():.4f}, "
+                    f"min={action.min():.4f}, max={action.max():.4f} | "
+                    f"state: shape={tuple(state.shape)}, mean={state.mean():.4f}"
+                )
+
         batch = preprocessor(batch)
+
+        if is_main_process and (step < 3 or (cfg.log_freq > 0 and step % (cfg.log_freq * 10) == 0)):
+            action = batch.get("action")
+            state = batch.get("observation.state")
+            if action is not None:
+                logging.info(
+                    f"[DEBUG step={step}] POST-PROCESSOR — "
+                    f"action: shape={tuple(action.shape)}, mean={action.mean():.4f}, std={action.std():.4f}, "
+                    f"min={action.min():.4f}, max={action.max():.4f}"
+                )
+                if state is not None:
+                    logging.info(
+                        f"[DEBUG step={step}] POST-PROCESSOR — "
+                        f"state: shape={tuple(state.shape)}, mean={state.mean():.4f}, std={state.std():.4f}"
+                    )
+
         train_tracker.dataloading_s = time.perf_counter() - start_time
 
         train_tracker, output_dict = update_policy(

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.")
+

tests/conftest.py

@@ -31,6 +31,39 @@ 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

@@ -1,66 +1,338 @@
-"""Minimal test script for Damiao motor with ID 3."""
+#!/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.utils.import_utils import _can_available
-
-if not _can_available:
-    pytest.skip("python-can not available", allow_module_level=True)
-
-from lerobot.motors import Motor
+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.skip(reason="Requires physical Damiao motor and CAN interface")
-def test_damiao_motor():
-    motors = {
-        "joint_3": Motor(
-            id=0x03,
-            model="damiao",
-            norm_mode="degrees",
-            motor_type_str="dm4310",
-            recv_id=0x13,
-        ),
-    }
-
-    bus = DamiaoMotorsBus(port="can0", motors=motors)
-
+@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:
-        print("Connecting...")
-        bus.connect()
-        print("✓ Connected")
-
-        print("Enabling torque...")
-        bus.enable_torque()
-        print("✓ Torque enabled")
-
-        print("Reading all states...")
-        states = bus.sync_read_all_states()
-        print(f"✓ States: {states}")
-
-        print("Reading position...")
-        positions = bus.sync_read("Present_Position")
-        print(f"✓ Position: {positions}")
-
-        print("Testing MIT control batch...")
-        current_pos = states["joint_3"]["position"]
-        commands = {"joint_3": (10.0, 0.5, current_pos, 0.0, 0.0)}
-        bus._mit_control_batch(commands)
-        print("✓ MIT control batch sent")
-
-        print("Disabling torque...")
-        bus.disable_torque()
-        print("✓ Torque disabled")
-
-        print("Setting zero position...")
-        bus.set_zero_position()
-        print("✓ Zero position set")
-
+        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:
-        print("Disconnecting...")
-        bus.disconnect(disable_torque=True)
-        print("✓ Disconnected")
+        # 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__":
-    test_damiao_motor()
+    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/rtc/test_training_time_rtc.py

@@ -0,0 +1,50 @@
+#!/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.
+
+"""Tests for training-time RTC helpers."""
+
+import torch
+
+from lerobot.configs.types import RTCTrainingDelayDistribution
+from lerobot.policies.rtc.configuration_rtc import RTCTrainingConfig
+from lerobot.policies.rtc.training_time import apply_rtc_training_time, sample_rtc_delay
+
+
+def test_rtc_training_config_defaults():
+    config = RTCTrainingConfig()
+    assert config.enabled is False
+    assert config.min_delay == 0
+    assert config.max_delay == 0
+    assert config.delay_distribution == RTCTrainingDelayDistribution.UNIFORM
+    assert config.exp_decay == 1.0
+
+
+def test_sample_rtc_delay_uniform_range():
+    cfg = RTCTrainingConfig(enabled=True, min_delay=1, max_delay=4)
+    delays = sample_rtc_delay(cfg, batch_size=100, device=torch.device("cpu"))
+    assert delays.min().item() >= 1
+    assert delays.max().item() <= 4
+
+
+def test_apply_rtc_training_time_prefix_mask():
+    time = torch.tensor([0.5])
+    delays = torch.tensor([2])
+    time_tokens, postfix_mask = apply_rtc_training_time(time, delays, seq_len=4)
+    assert time_tokens.shape == (1, 4)
+    assert postfix_mask.shape == (1, 4)
+    # Delay=2 means the first two steps are prefix (time forced to 0.0) and only the last two are postfix.
+    assert torch.allclose(time_tokens[0], torch.tensor([0.0, 0.0, 0.5, 0.5]))
+    assert torch.equal(postfix_mask[0], torch.tensor([False, False, True, True]))

tests/policies/test_delta_actions.py

@@ -0,0 +1,344 @@
+"""Tests for delta action transforms — full pipeline validation.
+
+Tests the complete flow matching OpenPI:
+  raw actions → DeltaActions → Normalize(delta_stats) → model → Unnormalize → AbsoluteActions
+
+Uses real dataset: lerobot-data-collection/dagger_final_1_21
+"""
+
+import numpy as np
+import pytest
+import torch
+
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.datasets.compute_stats import get_feature_stats
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.processor import TransitionKey, batch_to_transition
+from lerobot.processor.delta_action_processor import (
+    AbsoluteActionsProcessorStep,
+    DeltaActionsProcessorStep,
+    to_absolute_actions,
+    to_delta_actions,
+)
+from lerobot.processor.normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep
+from lerobot.utils.constants import ACTION, OBS_STATE
+
+CHUNK_SIZE = 10
+REPO_ID = "lerobot-data-collection/dagger_final_1_21"
+
+
+@pytest.fixture(scope="module")
+def dataset():
+    return LeRobotDataset(REPO_ID, episodes=[0])
+
+
+@pytest.fixture(scope="module")
+def action_dim(dataset):
+    return dataset.meta.features["action"]["shape"][0]
+
+
+def _build_action_chunks(dataset, chunk_size, max_chunks=50):
+    """Build action chunks from hf_dataset, like the training script does."""
+    hf = dataset.hf_dataset
+    total = len(hf)
+    all_ep = torch.tensor([int(hf[i]["episode_index"]) for i in range(total)])
+    chunks, states = [], []
+    for i in range(total - chunk_size + 1):
+        if all_ep[i] != all_ep[i + chunk_size - 1]:
+            continue
+        chunk_actions = torch.stack([hf[i + k]["action"] for k in range(chunk_size)]).float()
+        state = hf[i]["observation.state"].float()
+        chunks.append(chunk_actions)
+        states.append(state)
+        if len(chunks) >= max_chunks:
+            break
+    assert len(chunks) > 0, f"No valid chunks found. total={total}, ep_indices={all_ep.tolist()}"
+    return torch.stack(chunks), torch.stack(states)
+
+
+def _compute_delta_chunk_stats(action_chunks, states, mask):
+    all_deltas = []
+    for actions, state in zip(action_chunks, states):
+        delta = to_delta_actions(actions.unsqueeze(0), state.unsqueeze(0), mask).squeeze(0)
+        all_deltas.append(delta.numpy())
+    all_delta = np.concatenate(all_deltas, axis=0)
+    return get_feature_stats(all_delta, axis=0, keepdims=all_delta.ndim == 1)
+
+
+# --- Basic roundtrip tests ---
+
+def test_roundtrip_3d(action_dim):
+    actions = torch.randn(4, CHUNK_SIZE, action_dim)
+    state = torch.randn(4, action_dim)
+    mask = [True] * action_dim
+    recovered = to_absolute_actions(to_delta_actions(actions, state, mask), state, mask)
+    torch.testing.assert_close(recovered, actions)
+
+
+def test_roundtrip_2d(action_dim):
+    actions = torch.randn(4, action_dim)
+    state = torch.randn(4, action_dim)
+    mask = [True] * action_dim
+    recovered = to_absolute_actions(to_delta_actions(actions, state, mask), state, mask)
+    torch.testing.assert_close(recovered, actions)
+
+
+def test_no_mutation(action_dim):
+    actions = torch.randn(2, CHUNK_SIZE, action_dim)
+    original = actions.clone()
+    state = torch.randn(2, action_dim)
+    to_delta_actions(actions, state, [True] * action_dim)
+    torch.testing.assert_close(actions, original)
+
+
+def test_exclude_joints_supports_partial_name_matching():
+    names = [
+        "right_joint_1.pos",
+        "right_gripper.pos",
+        "left_joint_1.pos",
+        "left_gripper.pos",
+    ]
+    step = DeltaActionsProcessorStep(enabled=True, exclude_joints=["gripper"], action_names=names)
+    assert step._build_mask(len(names)) == [True, False, True, False]
+
+
+# --- Chunk-level delta stats test ---
+
+def test_chunk_stats_have_larger_std_than_frame_stats(dataset, action_dim):
+    """Chunk-level delta stats should have larger std than per-frame delta stats."""
+    action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)
+    mask = [True] * action_dim
+
+    chunk_stats = _compute_delta_chunk_stats(action_chunks, states, mask)
+
+    # Per-frame stats
+    hf = dataset.hf_dataset
+    n = min(500, len(hf))
+    frame_actions = torch.stack([hf[i]["action"] for i in range(n)]).float()
+    frame_states = torch.stack([hf[i]["observation.state"] for i in range(n)]).float()
+    frame_deltas = to_delta_actions(frame_actions, frame_states, mask).numpy()
+    frame_stats = get_feature_stats(frame_deltas, axis=0, keepdims=frame_deltas.ndim == 1)
+
+    assert chunk_stats["std"].mean() >= frame_stats["std"].mean(), (
+        f"Chunk std ({chunk_stats['std'].mean():.4f}) should be >= "
+        f"frame std ({frame_stats['std'].mean():.4f})"
+    )
+
+
+# --- Full pipeline roundtrip: delta → normalize → unnormalize → absolute ---
+
+def test_full_pipeline_roundtrip(dataset, action_dim):
+    """Test the complete OpenPI pipeline: delta → normalize → unnormalize → absolute."""
+    action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)
+    mask = [True] * action_dim
+
+    delta_stats = _compute_delta_chunk_stats(action_chunks, states, mask)
+    stats = {ACTION: {k: v for k, v in delta_stats.items()}}
+
+    features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,))}
+    norm_map = {FeatureType.ACTION: NormalizationMode.MEAN_STD}
+
+    delta_step = DeltaActionsProcessorStep(enabled=True)
+    normalizer = NormalizerProcessorStep(features=features, norm_map=norm_map, stats=stats)
+    unnormalizer = UnnormalizerProcessorStep(features=features, norm_map=norm_map, stats=stats)
+    absolute_step = AbsoluteActionsProcessorStep(enabled=True, delta_step=delta_step)
+
+    original_actions = action_chunks[0].unsqueeze(0)
+    state = states[0].unsqueeze(0)
+
+    batch = {ACTION: original_actions, OBS_STATE: state}
+    transition = batch_to_transition(batch)
+
+    # Forward: delta → normalize
+    t1 = delta_step(transition)
+    t2 = normalizer(t1)
+
+    normalized_action = t2[TransitionKey.ACTION]
+    assert normalized_action.abs().mean() < 10, (
+        f"Normalized actions should be in reasonable range, got mean abs {normalized_action.abs().mean():.2f}"
+    )
+
+    # Reverse: unnormalize → absolute
+    t3 = unnormalizer(t2)
+    t4 = absolute_step(t3)
+
+    recovered_actions = t4[TransitionKey.ACTION]
+    torch.testing.assert_close(recovered_actions, original_actions, atol=1e-4, rtol=1e-4)
+
+
+def test_normalized_delta_values_are_reasonable(dataset, action_dim):
+    """With correct chunk stats, normalized delta actions should be in a reasonable range."""
+    action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)
+    mask = [True] * action_dim
+
+    delta_stats = _compute_delta_chunk_stats(action_chunks, states, mask)
+    mean = torch.tensor(delta_stats["mean"]).float()
+    std = torch.tensor(delta_stats["std"]).float()
+
+    all_normalized = []
+    for actions, state in zip(action_chunks, states):
+        delta = to_delta_actions(actions.unsqueeze(0), state.unsqueeze(0), mask).squeeze(0)
+        normalized = (delta - mean) / (std + 1e-6)
+        all_normalized.append(normalized)
+
+    all_normalized = torch.cat(all_normalized, dim=0)
+
+    pct_in_range = (all_normalized.abs() < 5).float().mean()
+    assert pct_in_range > 0.9, (
+        f"Only {pct_in_range*100:.1f}% of normalized values in [-5, 5], expected >90%"
+    )
+
+    assert all_normalized.mean().abs() < 1.0, (
+        f"Mean of normalized deltas is {all_normalized.mean():.2f}, expected near 0"
+    )
+
+
+def test_processor_step_roundtrip(dataset, action_dim):
+    """DeltaActionsProcessorStep applies delta; to_absolute_actions recovers original."""
+    hf = dataset.hf_dataset
+    batch = {
+        ACTION: torch.stack([hf[i]["action"] for i in range(4)]),
+        OBS_STATE: torch.stack([hf[i]["observation.state"] for i in range(4)]),
+    }
+    original_actions = batch[ACTION].clone()
+    transition = batch_to_transition(batch)
+
+    step = DeltaActionsProcessorStep(enabled=True)
+    delta_transition = step(transition)
+    assert not torch.allclose(delta_transition[TransitionKey.ACTION], original_actions)
+
+    state = transition[TransitionKey.OBSERVATION][OBS_STATE]
+    mask = [True] * action_dim
+    recovered = to_absolute_actions(delta_transition[TransitionKey.ACTION], state, mask)
+    torch.testing.assert_close(recovered, original_actions)
+
+
+def test_processor_step_disabled_is_noop(dataset, action_dim):
+    """enabled=False should be a no-op."""
+    hf = dataset.hf_dataset
+    batch = {
+        ACTION: torch.stack([hf[i]["action"] for i in range(2)]),
+        OBS_STATE: torch.stack([hf[i]["observation.state"] for i in range(2)]),
+    }
+    original = batch[ACTION].clone()
+    transition = batch_to_transition(batch)
+    result = DeltaActionsProcessorStep(enabled=False)(transition)
+    torch.testing.assert_close(result[TransitionKey.ACTION], original)
+
+
+# --- Training batch shape validation ---
+
+def test_delta_with_action_chunks(dataset, action_dim):
+    """Verify delta works correctly with (B, chunk_size, action_dim) shaped actions."""
+    action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)
+
+    # Simulate a training batch: actions=(B, chunk_size, action_dim), state=(B, state_dim)
+    batch_actions = action_chunks[:4]  # (4, chunk_size, action_dim)
+    batch_states = states[:4]  # (4, state_dim)
+
+    mask = [True] * action_dim
+    delta = to_delta_actions(batch_actions, batch_states, mask)
+
+    # First action in each chunk should be close to zero (action[t] - state[t] ≈ small)
+    first_deltas = delta[:, 0, :]  # (B, action_dim)
+    assert first_deltas.abs().mean() < delta.abs().mean(), (
+        f"First action in chunk should have smaller delta than average. "
+        f"First: {first_deltas.abs().mean():.4f}, Average: {delta.abs().mean():.4f}"
+    )
+
+    # Later actions should have larger deltas
+    last_deltas = delta[:, -1, :]  # (B, action_dim)
+    assert last_deltas.abs().mean() >= first_deltas.abs().mean(), (
+        f"Last action in chunk should have >= delta than first. "
+        f"Last: {last_deltas.abs().mean():.4f}, First: {first_deltas.abs().mean():.4f}"
+    )
+
+    # Roundtrip
+    recovered = to_absolute_actions(delta, batch_states, mask)
+    torch.testing.assert_close(recovered, batch_actions)
+
+
+def test_delta_stats_match_actual_data_distribution(dataset, action_dim):
+    """Verify computed stats match the actual delta distribution."""
+    action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)
+    mask = [True] * action_dim
+
+    # Compute stats like the training script does
+    delta_stats = _compute_delta_chunk_stats(action_chunks, states, mask)
+
+    # Also compute directly
+    all_deltas = []
+    for actions, state in zip(action_chunks, states):
+        delta = to_delta_actions(actions.unsqueeze(0), state.unsqueeze(0), mask).squeeze(0)
+        all_deltas.append(delta)
+    all_deltas_tensor = torch.cat(all_deltas, dim=0)
+
+    # Compare mean
+    actual_mean = all_deltas_tensor.mean(dim=0).numpy()
+    np.testing.assert_allclose(delta_stats["mean"], actual_mean, atol=0.01)
+
+    # Compare std
+    actual_std = all_deltas_tensor.std(dim=0).numpy()
+    np.testing.assert_allclose(delta_stats["std"], actual_std, atol=0.1)
+
+    # Verify q01 < mean < q99
+    assert (delta_stats["q01"] < delta_stats["mean"]).all(), "q01 should be < mean"
+    assert (delta_stats["mean"] < delta_stats["q99"]).all(), "mean should be < q99"
+
+
+def test_quantile_normalization_roundtrip(dataset, action_dim):
+    """Full roundtrip with QUANTILES normalization (what OpenPI uses for pi05)."""
+    action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)
+    mask = [True] * action_dim
+
+    delta_stats = _compute_delta_chunk_stats(action_chunks, states, mask)
+    stats = {ACTION: {k: v for k, v in delta_stats.items()}}
+
+    features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,))}
+    norm_map = {FeatureType.ACTION: NormalizationMode.QUANTILES}
+
+    delta_step = DeltaActionsProcessorStep(enabled=True)
+    normalizer = NormalizerProcessorStep(features=features, norm_map=norm_map, stats=stats)
+    unnormalizer = UnnormalizerProcessorStep(features=features, norm_map=norm_map, stats=stats)
+    absolute_step = AbsoluteActionsProcessorStep(enabled=True, delta_step=delta_step)
+
+    original_actions = action_chunks[0].unsqueeze(0)
+    state = states[0].unsqueeze(0)
+
+    batch = {ACTION: original_actions, OBS_STATE: state}
+    transition = batch_to_transition(batch)
+
+    # Forward: delta → quantile normalize
+    t1 = delta_step(transition)
+    t2 = normalizer(t1)
+
+    normalized = t2[TransitionKey.ACTION]
+    # Most values should be in [-1, 1] with quantile normalization
+    pct_in_range = (normalized.abs() < 2).float().mean()
+    assert pct_in_range > 0.5, (
+        f"Only {pct_in_range*100:.1f}% in [-2, 2] after quantile norm, expected >50%"
+    )
+
+    # Reverse: unnormalize → absolute
+    t3 = unnormalizer(t2)
+    t4 = absolute_step(t3)
+
+    recovered = t4[TransitionKey.ACTION]
+    torch.testing.assert_close(recovered, original_actions, atol=1e-3, rtol=1e-3)
+
+
+def test_state_not_modified_by_delta(dataset, action_dim):
+    """State should never be modified by the delta processor."""
+    hf = dataset.hf_dataset
+    batch = {
+        ACTION: torch.stack([hf[i]["action"] for i in range(4)]),
+        OBS_STATE: torch.stack([hf[i]["observation.state"] for i in range(4)]),
+    }
+    original_state = batch[OBS_STATE].clone()
+    transition = batch_to_transition(batch)
+
+    step = DeltaActionsProcessorStep(enabled=True)
+    result = step(transition)
+
+    result_state = result[TransitionKey.OBSERVATION][OBS_STATE]
+    torch.testing.assert_close(result_state, original_state)

tests/policies/wall_x/test_wallx.py

@@ -19,24 +19,28 @@
 import os
 
 import pytest
-import torch
 
-# Skip if required dependencies are not available
-pytest.importorskip("peft")
-pytest.importorskip("transformers")
-pytest.importorskip("torchdiffeq")
+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():