edit holosoma

* fix(ci): move hub & lerobot artefacts to /mnt to avoid No space left on device in the future

* chore(ci): remove dh -h steps

.github/workflows/fast_tests.yml

@@ -60,12 +60,19 @@ jobs:
     runs-on: ubuntu-latest
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           persist-credentials: false
           lfs: true
 
+      # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
+      # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
+      - name: Setup /mnt storage
+        run: sudo chown -R $USER:$USER /mnt
+
       # TODO(Steven): Evaluate the need of these dependencies
       - name: Install apt dependencies
         run: |

.github/workflows/full_tests.yml

@@ -58,12 +58,19 @@ jobs:
       github.event_name == 'workflow_dispatch'
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
+      # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
+      # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
+      - name: Setup /mnt storage
+        run: sudo chown -R $USER:$USER /mnt
+
       - name: Install apt dependencies
         run: |
           sudo apt-get update && sudo apt-get install -y build-essential \

.github/workflows/unbound_deps_tests.yml

@@ -45,12 +45,19 @@ jobs:
     runs-on: ubuntu-latest
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
+      # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
+      # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
+      - name: Setup /mnt storage
+        run: sudo chown -R $USER:$USER /mnt
+
       - name: Install apt dependencies
         run: |
           sudo apt-get update && sudo apt-get install -y build-essential \

docs/source/_toctree.yml

@@ -79,6 +79,8 @@
     title: Hope Jr
   - local: reachy2
     title: Reachy 2
+  - local: unitree_g1
+    title: Unitree G1
   title: "Robots"
 - sections:
   - local: phone_teleop

docs/source/unitree_g1.mdx

@@ -0,0 +1,203 @@
+# Unitree G1 Robot Setup and Control
+
+This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
+
+## About the Unitree G1
+
+We offer support for both 29 and 23 DOF G1. In this first PR we introduce:
+
+- **`unitree g1` robot class, handling low level communication with the humanoid**
+- **ZMQ socket bridge** for remote communication over WiFi, allowing one to deploy policies remotely instead of over ethernet or directly on the Orin
+- **GR00T locomotion policy** for bipedal walking and balance
+
+---
+
+## Part 1: Connect to Robot over Ethernet
+
+### Step 1: Configure Your Computer's Ethernet Interface
+
+Set a static IP on the same subnet as the robot:
+
+```bash
+# Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
+sudo ip addr flush dev enp131s0
+sudo ip addr add 192.168.123.200/24 dev enp131s0
+sudo ip link set enp131s0 up
+```
+
+**Note**: The robot's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` where x ≠ 164.
+
+### Step 2: SSH into the Robot
+
+```bash
+ssh unitree@192.168.123.164
+# Password: 123
+```
+
+You should now be connected to the robot's onboard computer.
+
+---
+
+## Part 2: Enable WiFi on the Robot
+
+Once connected via Ethernet, follow these steps to enable WiFi:
+
+### Step 1: Enable WiFi Hardware
+
+```bash
+# Unblock WiFi radio
+sudo rfkill unblock wifi
+sudo rfkill unblock all
+
+# Bring up WiFi interface
+sudo ip link set wlan0 up
+
+# Enable NetworkManager control
+sudo nmcli radio wifi on
+sudo nmcli device set wlan0 managed yes
+sudo systemctl restart NetworkManager
+```
+
+### Step 2: Enable Internet Forwarding
+
+**On your laptop:**
+
+```bash
+# Enable IP forwarding
+sudo sysctl -w net.ipv4.ip_forward=1
+
+# Set up NAT (replace wlp132s0f0 with your WiFi interface)
+sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
+sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
+sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
+```
+
+**On the robot:**
+
+```bash
+# Add laptop as default gateway
+sudo ip route del default 2>/dev/null || true
+sudo ip route add default via 192.168.123.200 dev eth0
+echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
+
+# Test connection
+ping -c 3 8.8.8.8
+```
+
+### Step 3: Connect to WiFi Network
+
+```bash
+# List available networks
+nmcli device wifi list
+
+# Connect to your WiFi (example)
+sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
+sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
+sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
+sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
+sudo nmcli connection up "YourNetwork"
+
+# Check WiFi IP address
+ip a show wlan0
+```
+
+### Step 4: SSH Over WiFi
+
+Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+# Password: 123
+```
+
+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address (e.g., `172.18.129.215`).
+
+---
+
+## Part 3: Robot Server Setup
+
+### Step 1: Install LeRobot on the Orin
+
+SSH into the robot and install LeRobot:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+
+### Step 2: Run the Robot Server
+
+On the robot:
+
+```bash
+python src/lerobot/robots/unitree_g1/run_g1_server.py
+```
+
+**Important**: Keep this terminal running. The server must be active for remote control.
+
+---
+
+## Part 4: Running GR00T Locomotion
+
+With the robot server running, you can now control the robot from your laptop.
+
+### Step 1: Install LeRobot on your machine
+
+```bash
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+### Step 2: Update Robot IP in Config
+
+Edit the config file to match your robot's WiFi IP:
+
+```python
+# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
+robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
+```
+
+**Note**: When running directly on the G1 (not remotely), set `robot_ip: str = "127.0.0.1"` instead.
+
+### Step 3: Run the Locomotion Policy
+
+```bash
+# Run GR00T locomotion controller
+python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
+```
+
+### Step 4: Control with Remote
+
+- **Left stick**: Forward/backward and left/right movement
+- **Right stick**: Rotation
+- **R1 button**: Raise waist height
+- **R2 button**: Lower waist height
+
+Press `Ctrl+C` to stop the policy.
+
+---
+
+## Additional Resources
+
+- [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
+- [GR00T Policy Repository](https://huggingface.co/nepyope/GR00T-WholeBodyControl_g1)
+- [LeRobot Documentation](https://github.com/huggingface/lerobot)
+- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
+
+---
+
+_Last updated: December 2025_

examples/unitree_g1/gr00t_locomotion.py

@@ -0,0 +1,347 @@
+#!/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.
+"""
+Example: GR00T Locomotion with Pre-loaded Policies
+
+This example demonstrates the NEW pattern for loading GR00T policies externally
+and passing them to the robot class.
+"""
+
+import argparse
+import logging
+import threading
+import time
+from collections import deque
+
+import numpy as np
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
+
+logger = logging.getLogger(__name__)
+
+GROOT_DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
+GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # hip pitch
+GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # knee
+GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # ankle pitch
+
+MISSING_JOINTS = []
+G1_MODEL = "g1_23"  # or "g1_29"
+if G1_MODEL == "g1_23":
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # waist yaw/pitch, wrist pitch/yaw
+
+LOCOMOTION_ACTION_SCALE = 0.25
+
+LOCOMOTION_CONTROL_DT = 0.02
+
+ANG_VEL_SCALE: float = 0.25
+DOF_POS_SCALE: float = 1.0
+DOF_VEL_SCALE: float = 0.05
+CMD_SCALE: list = [2.0, 2.0, 0.25]
+
+
+DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
+
+
+def load_groot_policies(
+    repo_id: str = DEFAULT_GROOT_REPO_ID,
+) -> tuple[ort.InferenceSession, ort.InferenceSession]:
+    """Load GR00T dual-policy system (Balance + Walk) from Hugging Face Hub.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID containing the ONNX policies.
+    """
+    logger.info(f"Loading GR00T dual-policy system from Hugging Face Hub ({repo_id})...")
+
+    # Download ONNX policies from Hugging Face Hub
+    balance_path = hf_hub_download(
+        repo_id=repo_id,
+        filename="GR00T-WholeBodyControl-Balance.onnx",
+    )
+    walk_path = hf_hub_download(
+        repo_id=repo_id,
+        filename="GR00T-WholeBodyControl-Walk.onnx",
+    )
+
+    # Load ONNX policies
+    policy_balance = ort.InferenceSession(balance_path)
+    policy_walk = ort.InferenceSession(walk_path)
+
+    logger.info("GR00T policies loaded successfully")
+
+    return policy_balance, policy_walk
+
+
+class GrootLocomotionController:
+    """
+    Handles GR00T-style locomotion control for the Unitree G1 robot.
+
+    This controller manages:
+    - Dual-policy system (Balance + Walk)
+    - 29-joint observation processing
+    - 15D action output (legs + waist)
+    - Policy inference and motor command generation
+    """
+
+    def __init__(self, policy_balance, policy_walk, robot, config):
+        self.policy_balance = policy_balance
+        self.policy_walk = policy_walk
+        self.robot = robot
+        self.config = config
+
+        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
+
+        # GR00T-specific state
+        self.groot_qj_all = np.zeros(29, dtype=np.float32)
+        self.groot_dqj_all = np.zeros(29, dtype=np.float32)
+        self.groot_action = np.zeros(15, dtype=np.float32)
+        self.groot_obs_single = np.zeros(86, dtype=np.float32)
+        self.groot_obs_history = deque(maxlen=6)
+        self.groot_obs_stacked = np.zeros(516, dtype=np.float32)
+        self.groot_height_cmd = 0.74  # Default base height
+        self.groot_orientation_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
+
+        # input to gr00t is 6 frames (6*86D=516)
+        for _ in range(6):
+            self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
+
+        # Thread management
+        self.locomotion_running = False
+        self.locomotion_thread = None
+
+        logger.info("GrootLocomotionController initialized")
+
+    def groot_locomotion_run(self):
+        # get current observation
+        robot_state = self.robot.get_observation()
+
+        if robot_state is None:
+            return
+
+        # get command from remote controller
+        if robot_state.wireless_remote is not None:
+            self.robot.remote_controller.set(robot_state.wireless_remote)
+            if self.robot.remote_controller.button[0]:  # R1 - raise waist
+                self.groot_height_cmd += 0.001
+                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+            if self.robot.remote_controller.button[4]:  # R2 - lower waist
+                self.groot_height_cmd -= 0.001
+                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+        else:
+            self.robot.remote_controller.lx = 0.0
+            self.robot.remote_controller.ly = 0.0
+            self.robot.remote_controller.rx = 0.0
+            self.robot.remote_controller.ry = 0.0
+
+        self.locomotion_cmd[0] = self.robot.remote_controller.ly  # forward/backward
+        self.locomotion_cmd[1] = self.robot.remote_controller.lx * -1  # left/right
+        self.locomotion_cmd[2] = self.robot.remote_controller.rx * -1  # rotation rate
+
+        for i in range(29):
+            self.groot_qj_all[i] = robot_state.motor_state[i].q
+            self.groot_dqj_all[i] = robot_state.motor_state[i].dq
+
+        # adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.groot_qj_all[idx] = 0.0
+            self.groot_dqj_all[idx] = 0.0
+
+        # Scale joint positions and velocities
+        qj_obs = self.groot_qj_all.copy()
+        dqj_obs = self.groot_dqj_all.copy()
+
+        # express imu data in gravity frame of reference
+        quat = robot_state.imu_state.quaternion
+        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
+        gravity_orientation = self.robot.get_gravity_orientation(quat)
+
+        # scale joint positions and velocities before policy inference
+        qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
+        dqj_obs = dqj_obs * DOF_VEL_SCALE
+        ang_vel_scaled = ang_vel * ANG_VEL_SCALE
+
+        # build single frame observation
+        self.groot_obs_single[:3] = self.locomotion_cmd * np.array(CMD_SCALE)
+        self.groot_obs_single[3] = self.groot_height_cmd
+        self.groot_obs_single[4:7] = self.groot_orientation_cmd
+        self.groot_obs_single[7:10] = ang_vel_scaled
+        self.groot_obs_single[10:13] = gravity_orientation
+        self.groot_obs_single[13:42] = qj_obs
+        self.groot_obs_single[42:71] = dqj_obs
+        self.groot_obs_single[71:86] = self.groot_action  # 15D previous actions
+
+        # Add to history and stack observations (6 frames × 86D = 516D)
+        self.groot_obs_history.append(self.groot_obs_single.copy())
+
+        # Stack all 6 frames into 516D vector
+        for i, obs_frame in enumerate(self.groot_obs_history):
+            start_idx = i * 86
+            end_idx = start_idx + 86
+            self.groot_obs_stacked[start_idx:end_idx] = obs_frame
+
+        # Run policy inference (ONNX) with 516D stacked observation
+
+        cmd_magnitude = np.linalg.norm(self.locomotion_cmd)
+
+        selected_policy = (
+            self.policy_balance if cmd_magnitude < 0.05 else self.policy_walk
+        )  # balance/standing policy for small commands, walking policy for movement commands
+
+        # run policy inference
+        ort_inputs = {selected_policy.get_inputs()[0].name: np.expand_dims(self.groot_obs_stacked, axis=0)}
+        ort_outs = selected_policy.run(None, ort_inputs)
+        self.groot_action = ort_outs[0].squeeze()
+
+        # transform action back to target joint positions
+        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * LOCOMOTION_ACTION_SCALE
+
+        # command motors
+        for i in range(15):
+            motor_idx = i
+            self.robot.msg.motor_cmd[motor_idx].q = target_dof_pos_15[i]
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
+            self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # adapt action for g1_23dof
+        for joint_idx in MISSING_JOINTS:
+            self.robot.msg.motor_cmd[joint_idx].q = 0.0
+            self.robot.msg.motor_cmd[joint_idx].qd = 0
+            self.robot.msg.motor_cmd[joint_idx].kp = self.robot.kp[joint_idx]
+            self.robot.msg.motor_cmd[joint_idx].kd = self.robot.kd[joint_idx]
+            self.robot.msg.motor_cmd[joint_idx].tau = 0
+
+        # send action to robot
+        self.robot.send_action(self.robot.msg)
+
+    def _locomotion_thread_loop(self):
+        """Background thread that runs the locomotion policy at specified rate."""
+        logger.info("Locomotion thread started")
+        while self.locomotion_running:
+            start_time = time.time()
+            try:
+                self.groot_locomotion_run()
+            except Exception as e:
+                logger.error(f"Error in locomotion loop: {e}")
+
+            # Sleep to maintain control rate
+            elapsed = time.time() - start_time
+            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+        logger.info("Locomotion thread stopped")
+
+    def start_locomotion_thread(self):
+        if self.locomotion_running:
+            logger.warning("Locomotion thread already running")
+            return
+
+        logger.info("Starting locomotion control thread...")
+        self.locomotion_running = True
+        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
+        self.locomotion_thread.start()
+
+        logger.info("Locomotion control thread started!")
+
+    def stop_locomotion_thread(self):
+        if not self.locomotion_running:
+            return
+
+        logger.info("Stopping locomotion control thread...")
+        self.locomotion_running = False
+        if self.locomotion_thread:
+            self.locomotion_thread.join(timeout=2.0)
+        logger.info("Locomotion control thread stopped")
+
+    def reset_robot(self):
+        """Move robot legs to default standing position over 2 seconds (arms are not moved)."""
+        total_time = 3.0
+        num_step = int(total_time / self.robot.control_dt)
+
+        # Only control legs, not arms (first 12 joints)
+        default_pos = GROOT_DEFAULT_ANGLES  # First 12 values are leg angles
+        dof_size = len(default_pos)
+
+        # Get current lowstate
+        robot_state = self.robot.get_observation()
+
+        # Record the current leg positions
+        init_dof_pos = np.zeros(dof_size, dtype=np.float32)
+        for i in range(dof_size):
+            init_dof_pos[i] = robot_state.motor_state[i].q
+
+        # Move legs to default pos
+        for i in range(num_step):
+            alpha = i / num_step
+            for motor_idx in range(dof_size):
+                target_pos = default_pos[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].q = (
+                    init_dof_pos[motor_idx] * (1 - alpha) + target_pos * alpha
+                )
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+        logger.info("Reached default position (legs only)")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="GR00T Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_GROOT_REPO_ID,
+        help=f"Hugging Face Hub repo ID for GR00T policies (default: {DEFAULT_GROOT_REPO_ID})",
+    )
+    args = parser.parse_args()
+
+    # load policies
+    policy_balance, policy_walk = load_groot_policies(repo_id=args.repo_id)
+
+    # initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    # initialize gr00t locomotion controller
+    groot_controller = GrootLocomotionController(
+        policy_balance=policy_balance,
+        policy_walk=policy_walk,
+        robot=robot,
+        config=config,
+    )
+
+    # reset legs and start locomotion thread
+    try:
+        groot_controller.reset_robot()
+        groot_controller.start_locomotion_thread()
+
+        # log status
+        logger.info("Robot initialized with GR00T locomotion policies")
+        logger.info("Locomotion controller running in background thread")
+        logger.info("Press Ctrl+C to stop")
+
+        # keep robot alive
+        while True:
+            time.sleep(1.0)
+    except KeyboardInterrupt:
+        print("\nStopping locomotion...")
+        groot_controller.stop_locomotion_thread()
+        print("Done!")

examples/unitree_g1/holosoma_locomotion.py

@@ -0,0 +1,479 @@
+#!/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.
+"""
+Example: Holosoma Whole-Body Locomotion (23-DOF and 29-DOF)
+
+This example demonstrates loading Holosoma whole-body locomotion policies
+and running them on the Unitree G1 robot.
+
+Supports both:
+- 23-DOF native policies (82D observations, 23D actions)
+- 29-DOF policies (100D observations, 29D actions)
+"""
+
+import argparse
+import logging
+import threading
+import time
+
+import numpy as np
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+# =============================================================================
+# 29-DOF Configuration
+# =============================================================================
+# fmt: off
+HOLOSOMA_29DOF_DEFAULT_ANGLES = np.array([
+    -0.312, 0.0, 0.0, 0.669, -0.363, 0.0,  # left leg
+    -0.312, 0.0, 0.0, 0.669, -0.363, 0.0,  # right leg
+    0.0, 0.0, 0.0,                          # waist (yaw, roll, pitch)
+    0.2, 0.2, 0.0, 0.6, 0.0, 0.0, 0.0,     # left arm
+    0.2, -0.2, 0.0, 0.6, 0.0, 0.0, 0.0,    # right arm
+], dtype=np.float32)
+
+HOLOSOMA_29DOF_KP = np.array([
+    40.179238471, 99.098427777, 40.179238471, 99.098427777, 28.501246196, 28.501246196,  # left leg
+    40.179238471, 99.098427777, 40.179238471, 99.098427777, 28.501246196, 28.501246196,  # right leg
+    40.179238471, 28.501246196, 28.501246196,  # waist
+    14.250623098, 14.250623098, 14.250623098, 14.250623098, 14.250623098, 16.778327481, 16.778327481,  # left arm
+    14.250623098, 14.250623098, 14.250623098, 14.250623098, 14.250623098, 16.778327481, 16.778327481,  # right arm
+], dtype=np.float32)
+
+HOLOSOMA_29DOF_KD = np.array([
+    2.557889765, 6.308801854, 2.557889765, 6.308801854, 1.814445687, 1.814445687,  # left leg
+    2.557889765, 6.308801854, 2.557889765, 6.308801854, 1.814445687, 1.814445687,  # right leg
+    2.557889765, 1.814445687, 1.814445687,  # waist
+    0.907222843, 0.907222843, 0.907222843, 0.907222843, 0.907222843, 1.068141502, 1.068141502,  # left arm
+    0.907222843, 0.907222843, 0.907222843, 0.907222843, 0.907222843, 1.068141502, 1.068141502,  # right arm
+], dtype=np.float32)
+
+# =============================================================================
+# 23-DOF Configuration (native G1-23: no waist_roll/pitch, no wrist_pitch/yaw)
+# Derived from 29-DOF Holosoma values
+# =============================================================================
+# Joint order: 6 left leg, 6 right leg, 1 waist_yaw, 5 left arm, 5 right arm
+HOLOSOMA_23DOF_DEFAULT_ANGLES = np.array([
+    -0.312, 0.0, 0.0, 0.669, -0.363, 0.0,  # left leg (from 29-DOF)
+    -0.312, 0.0, 0.0, 0.669, -0.363, 0.0,  # right leg (from 29-DOF)
+    0.0,                                    # waist_yaw only (from 29-DOF)
+    0.2, 0.2, 0.0, 0.6, 0.0,               # left arm first 5 joints (from 29-DOF)
+    0.2, -0.2, 0.0, 0.6, 0.0,              # right arm first 5 joints (from 29-DOF)
+], dtype=np.float32)
+
+HOLOSOMA_23DOF_KP = np.array([
+    40.179238471, 99.098427777, 40.179238471, 99.098427777, 28.501246196, 28.501246196,  # left leg
+    40.179238471, 99.098427777, 40.179238471, 99.098427777, 28.501246196, 28.501246196,  # right leg
+    40.179238471,                                                                         # waist_yaw
+    14.250623098, 14.250623098, 14.250623098, 14.250623098, 14.250623098,                 # left arm
+    14.250623098, 14.250623098, 14.250623098, 14.250623098, 14.250623098,                 # right arm
+], dtype=np.float32)
+
+HOLOSOMA_23DOF_KD = np.array([
+    2.557889765, 6.308801854, 2.557889765, 6.308801854, 1.814445687, 1.814445687,  # left leg
+    2.557889765, 6.308801854, 2.557889765, 6.308801854, 1.814445687, 1.814445687,  # right leg
+    2.557889765,                                                                    # waist_yaw
+    0.907222843, 0.907222843, 0.907222843, 0.907222843, 0.907222843,               # left arm
+    0.907222843, 0.907222843, 0.907222843, 0.907222843, 0.907222843,               # right arm
+], dtype=np.float32)
+
+# Maps 23-DOF policy index → 29-DOF motor index
+# 23-DOF: legs(0-11), waist_yaw(12), L_arm(13-17), R_arm(18-22)
+# 29-DOF: legs(0-11), waist(12-14), L_arm(15-21), R_arm(22-28)
+DOF_23_TO_MOTOR_MAP = [
+    0, 1, 2, 3, 4, 5,       # left leg → motor 0-5
+    6, 7, 8, 9, 10, 11,     # right leg → motor 6-11
+    12,                      # waist_yaw → motor 12
+    15, 16, 17, 18, 19,     # left arm (skip wrist_pitch/yaw) → motor 15-19
+    22, 23, 24, 25, 26,     # right arm (skip wrist_pitch/yaw) → motor 22-26
+]
+# fmt: on
+
+# Control parameters
+LOCOMOTION_CONTROL_DT = 0.02  # 50Hz
+LOCOMOTION_ACTION_SCALE = 0.25
+ANG_VEL_SCALE = 0.25
+DOF_POS_SCALE = 1.0
+DOF_VEL_SCALE = 0.05
+GAIT_PERIOD = 1.0
+
+DEFAULT_HOLOSOMA_REPO_ID = "nepyope/holosoma_locomotion"
+
+
+def load_holosoma_policy(
+    repo_id: str = DEFAULT_HOLOSOMA_REPO_ID,
+    policy_name: str = "fastsac",
+    local_path: str | None = None,
+) -> tuple[ort.InferenceSession, int]:
+    """Load Holosoma policy and detect observation dimension.
+
+    Returns:
+        (policy, obs_dim) tuple where obs_dim is 82 (23-DOF) or 100 (29-DOF)
+    """
+    if local_path is not None:
+        logger.info(f"Loading policy from local path: {local_path}")
+        policy_path = local_path
+    else:
+        logger.info(f"Loading policy from Hugging Face Hub: {repo_id}")
+        policy_path = hf_hub_download(repo_id=repo_id, filename=f"{policy_name}_g1_29dof.onnx")
+
+    policy = ort.InferenceSession(policy_path)
+
+    # Detect observation dimension from model input shape
+    input_shape = policy.get_inputs()[0].shape
+    obs_dim = input_shape[1] if len(input_shape) > 1 else input_shape[0]
+
+    logger.info(f"Policy loaded successfully")
+    logger.info(f"  Input: {policy.get_inputs()[0].name}, shape: {input_shape} → obs_dim={obs_dim}")
+    logger.info(f"  Output: {policy.get_outputs()[0].name}, shape: {policy.get_outputs()[0].shape}")
+
+    return policy, obs_dim
+
+
+class HolosomaLocomotionController:
+    """
+    Handles Holosoma whole-body locomotion for Unitree G1.
+    Supports both 23-DOF (82D obs) and 29-DOF (100D obs) policies.
+    """
+
+    def __init__(self, policy, robot, config, obs_dim: int = 100):
+        self.policy = policy
+        self.robot = robot
+        self.config = config
+        self.obs_dim = obs_dim
+
+        # Detect policy type from observation dimension
+        self.is_23dof = (obs_dim == 82)
+        self.num_dof = 23 if self.is_23dof else 29
+
+        # Velocity commands
+        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
+
+        # State variables sized for policy type
+        self.qj = np.zeros(self.num_dof, dtype=np.float32)
+        self.dqj = np.zeros(self.num_dof, dtype=np.float32)
+        self.locomotion_action = np.zeros(self.num_dof, dtype=np.float32)
+        self.locomotion_obs = np.zeros(obs_dim, dtype=np.float32)
+        self.last_unscaled_action = np.zeros(self.num_dof, dtype=np.float32)
+
+        # Select config based on DOF
+        if self.is_23dof:
+            self.default_angles = HOLOSOMA_23DOF_DEFAULT_ANGLES
+            self.kp = HOLOSOMA_23DOF_KP
+            self.kd = HOLOSOMA_23DOF_KD
+            self.motor_map = DOF_23_TO_MOTOR_MAP
+        else:
+            self.default_angles = HOLOSOMA_29DOF_DEFAULT_ANGLES
+            self.kp = HOLOSOMA_29DOF_KP
+            self.kd = HOLOSOMA_29DOF_KD
+            self.motor_map = list(range(29))  # Identity map for 29-DOF
+
+        # Phase state for gait
+        self.phase = np.zeros((1, 2), dtype=np.float32)
+        self.phase[0, 0] = 0.0
+        self.phase[0, 1] = np.pi
+        self.phase_dt = 2 * np.pi / (50.0 * GAIT_PERIOD)
+        self.is_standing = False
+
+        self.counter = 0
+        self.locomotion_running = False
+        self.locomotion_thread = None
+
+        logger.info(f"HolosomaLocomotionController initialized")
+        logger.info(f"  Mode: {'23-DOF (82D obs)' if self.is_23dof else '29-DOF (100D obs)'}")
+        logger.info(f"  Action dim: {self.num_dof}")
+
+    def holosoma_locomotion_run(self):
+        """Main locomotion loop - handles both 23-DOF and 29-DOF."""
+        self.counter += 1
+
+        if self.counter == 1:
+            print("\n" + "=" * 60)
+            print(f"🚀 RUNNING HOLOSOMA {self.num_dof}-DOF LOCOMOTION POLICY")
+            print(f"   {self.obs_dim}D observations → {self.num_dof}D actions")
+            print("=" * 60 + "\n")
+
+        robot_state = self.robot.get_observation()
+        if robot_state is None:
+            return
+
+        # Remote controller
+        if robot_state.wireless_remote is not None:
+            self.robot.remote_controller.set(robot_state.wireless_remote)
+        else:
+            self.robot.remote_controller.lx = 0.0
+            self.robot.remote_controller.ly = 0.0
+            self.robot.remote_controller.rx = 0.0
+            self.robot.remote_controller.ry = 0.0
+
+        # Deadzone
+        ly = self.robot.remote_controller.ly if abs(self.robot.remote_controller.ly) > 0.1 else 0.0
+        lx = self.robot.remote_controller.lx if abs(self.robot.remote_controller.lx) > 0.1 else 0.0
+        rx = self.robot.remote_controller.rx if abs(self.robot.remote_controller.rx) > 0.1 else 0.0
+
+        self.locomotion_cmd[0] = ly
+        self.locomotion_cmd[1] = -lx
+        self.locomotion_cmd[2] = -rx
+
+        # Read joint states using motor map
+        for i in range(self.num_dof):
+            motor_idx = self.motor_map[i]
+            self.qj[i] = robot_state.motor_state[motor_idx].q
+            self.dqj[i] = robot_state.motor_state[motor_idx].dq
+
+        # IMU
+        quat = robot_state.imu_state.quaternion
+        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
+        gravity_orientation = self.robot.get_gravity_orientation(quat)
+
+        # Scale observations
+        qj_obs = (self.qj - self.default_angles) * DOF_POS_SCALE
+        dqj_obs = self.dqj * DOF_VEL_SCALE
+        ang_vel_scaled = ang_vel * ANG_VEL_SCALE
+
+        # Phase update
+        cmd_norm = np.linalg.norm(self.locomotion_cmd[:2])
+        ang_cmd_norm = np.abs(self.locomotion_cmd[2])
+
+        if cmd_norm < 0.01 and ang_cmd_norm < 0.01:
+            self.phase[0, :] = np.pi * np.ones(2)
+            self.is_standing = True
+        elif self.is_standing:
+            self.phase = np.array([[0.0, np.pi]], dtype=np.float32)
+            self.is_standing = False
+        else:
+            phase_tp1 = self.phase + self.phase_dt
+            self.phase = np.fmod(phase_tp1 + np.pi, 2 * np.pi) - np.pi
+
+        sin_phase = np.sin(self.phase[0, :])
+        cos_phase = np.cos(self.phase[0, :])
+
+        # Build observation (format depends on DOF)
+        if self.is_23dof:
+            # 82D: [23 actions, 3 ang_vel, 1 cmd_yaw, 2 cmd_lin, 2 cos, 23 pos, 23 vel, 3 grav, 2 sin]
+            self.locomotion_obs[0:23] = self.last_unscaled_action
+            self.locomotion_obs[23:26] = ang_vel_scaled
+            self.locomotion_obs[26] = self.locomotion_cmd[2]
+            self.locomotion_obs[27:29] = self.locomotion_cmd[:2]
+            self.locomotion_obs[29:31] = cos_phase
+            self.locomotion_obs[31:54] = qj_obs
+            self.locomotion_obs[54:77] = dqj_obs
+            self.locomotion_obs[77:80] = gravity_orientation
+            self.locomotion_obs[80:82] = sin_phase
+        else:
+            # 100D: [29 actions, 3 ang_vel, 1 cmd_yaw, 2 cmd_lin, 2 cos, 29 pos, 29 vel, 3 grav, 2 sin]
+            self.locomotion_obs[0:29] = self.last_unscaled_action
+            self.locomotion_obs[29:32] = ang_vel_scaled
+            self.locomotion_obs[32] = self.locomotion_cmd[2]
+            self.locomotion_obs[33:35] = self.locomotion_cmd[:2]
+            self.locomotion_obs[35:37] = cos_phase
+            self.locomotion_obs[37:66] = qj_obs
+            self.locomotion_obs[66:95] = dqj_obs
+            self.locomotion_obs[95:98] = gravity_orientation
+            self.locomotion_obs[98:100] = sin_phase
+
+        # Policy inference
+        obs_input = self.locomotion_obs.reshape(1, -1).astype(np.float32)
+        ort_inputs = {self.policy.get_inputs()[0].name: obs_input}
+        ort_outs = self.policy.run(None, ort_inputs)
+
+        raw_action = ort_outs[0].squeeze()
+        clipped_action = np.clip(raw_action, -100.0, 100.0)
+
+        self.last_unscaled_action = clipped_action.copy()
+        self.locomotion_action = clipped_action * LOCOMOTION_ACTION_SCALE
+
+        # Debug
+        if self.counter <= 3:
+            print(f"\n[Holosoma Debug #{self.counter}]")
+            print(f"  Phase: ({self.phase[0, 0]:.3f}, {self.phase[0, 1]:.3f})")
+            print(f"  Cmd: ({self.locomotion_cmd[0]:.2f}, {self.locomotion_cmd[1]:.2f}, {self.locomotion_cmd[2]:.2f})")
+            print(f"  Action range: [{raw_action.min():.3f}, {raw_action.max():.3f}]")
+
+        # Compute target positions
+        target_dof_pos = self.default_angles + self.locomotion_action
+
+        # Send commands to motors via motor map
+        for i in range(self.num_dof):
+            motor_idx = self.motor_map[i]
+            self.robot.msg.motor_cmd[motor_idx].q = target_dof_pos[i]
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = self.kp[i]
+            self.robot.msg.motor_cmd[motor_idx].kd = self.kd[i]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # For 23-DOF: zero out missing joints (waist_roll/pitch, wrist_pitch/yaw)
+        if self.is_23dof:
+            missing_motors = [13, 14, 20, 21, 27, 28]  # waist_roll, waist_pitch, wrist_pitch/yaw
+            for motor_idx in missing_motors:
+                self.robot.msg.motor_cmd[motor_idx].q = 0.0
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = 40.0
+                self.robot.msg.motor_cmd[motor_idx].kd = 2.0
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        self.robot.send_action(self.robot.msg)
+
+    def _locomotion_thread_loop(self):
+        logger.info("Locomotion thread started")
+        while self.locomotion_running:
+            start_time = time.time()
+            try:
+                self.holosoma_locomotion_run()
+            except Exception as e:
+                logger.error(f"Error in locomotion loop: {e}")
+                import traceback
+                traceback.print_exc()
+
+            elapsed = time.time() - start_time
+            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+        logger.info("Locomotion thread stopped")
+
+    def start_locomotion_thread(self):
+        if self.locomotion_running:
+            logger.warning("Locomotion thread already running")
+            return
+        logger.info("Starting locomotion control thread...")
+        self.locomotion_running = True
+        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
+        self.locomotion_thread.start()
+        logger.info("Locomotion control thread started!")
+
+    def stop_locomotion_thread(self):
+        if not self.locomotion_running:
+            return
+        logger.info("Stopping locomotion control thread...")
+        self.locomotion_running = False
+        if self.locomotion_thread:
+            self.locomotion_thread.join(timeout=2.0)
+        logger.info("Locomotion control thread stopped")
+
+    def reset_robot(self):
+        """Move joints to default position."""
+        logger.info(f"Moving {self.num_dof} joints to default position...")
+
+        total_time = 3.0
+        num_step = int(total_time / self.robot.control_dt)
+
+        robot_state = self.robot.get_observation()
+
+        # Record current positions
+        init_dof_pos = np.zeros(self.num_dof, dtype=np.float32)
+        for i in range(self.num_dof):
+            motor_idx = self.motor_map[i]
+            init_dof_pos[i] = robot_state.motor_state[motor_idx].q
+
+        # Interpolate to target
+        for step in range(num_step):
+            alpha = step / num_step
+            for i in range(self.num_dof):
+                motor_idx = self.motor_map[i]
+                target = self.default_angles[i]
+                self.robot.msg.motor_cmd[motor_idx].q = init_dof_pos[i] * (1 - alpha) + target * alpha
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = self.kp[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = self.kd[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            # Zero missing joints for 23-DOF
+            if self.is_23dof:
+                for motor_idx in [13, 14, 20, 21, 27, 28]:
+                    self.robot.msg.motor_cmd[motor_idx].q = 0.0
+                    self.robot.msg.motor_cmd[motor_idx].qd = 0
+                    self.robot.msg.motor_cmd[motor_idx].kp = 40.0
+                    self.robot.msg.motor_cmd[motor_idx].kd = 2.0
+                    self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+
+        logger.info(f"Reached default position ({self.num_dof} joints)")
+
+        # Hold for 2 seconds
+        logger.info("Holding default position for 2 seconds...")
+        hold_steps = int(2.0 / self.robot.control_dt)
+        for _ in range(hold_steps):
+            for i in range(self.num_dof):
+                motor_idx = self.motor_map[i]
+                self.robot.msg.motor_cmd[motor_idx].q = self.default_angles[i]
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = self.kp[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = self.kd[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            if self.is_23dof:
+                for motor_idx in [13, 14, 20, 21, 27, 28]:
+                    self.robot.msg.motor_cmd[motor_idx].q = 0.0
+                    self.robot.msg.motor_cmd[motor_idx].qd = 0
+                    self.robot.msg.motor_cmd[motor_idx].kp = 40.0
+                    self.robot.msg.motor_cmd[motor_idx].kd = 2.0
+                    self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+
+        logger.info("Ready to start locomotion!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Holosoma Locomotion Controller for Unitree G1")
+    parser.add_argument("--repo-id", type=str, default=DEFAULT_HOLOSOMA_REPO_ID)
+    parser.add_argument("--policy", type=str, default="fastsac", choices=["fastsac", "ppo"])
+    parser.add_argument("--local-path", type=str, default=None, help="Path to local ONNX file")
+    args = parser.parse_args()
+
+    # Load policy and detect dimensions
+    policy, obs_dim = load_holosoma_policy(
+        repo_id=args.repo_id,
+        policy_name=args.policy,
+        local_path=args.local_path,
+    )
+
+    # Initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    # Initialize controller with detected obs_dim
+    controller = HolosomaLocomotionController(
+        policy=policy,
+        robot=robot,
+        config=config,
+        obs_dim=obs_dim,
+    )
+
+    try:
+        controller.reset_robot()
+        controller.start_locomotion_thread()
+
+        logger.info(f"Robot initialized with Holosoma {'23-DOF' if obs_dim == 82 else '29-DOF'} policy")
+        logger.info("Use remote controller: LY=fwd/back, LX=left/right, RX=rotate")
+        logger.info("Press Ctrl+C to stop")
+
+        while True:
+            time.sleep(1.0)
+    except KeyboardInterrupt:
+        print("\nStopping locomotion...")
+        controller.stop_locomotion_thread()
+        print("Done!")

examples/unitree_g1/unitree_rl_locomotion.py

@@ -0,0 +1,447 @@
+#!/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.
+"""
+Example: Unitree RL 12-DOF Legs-Only Locomotion (TorchScript)
+
+This example demonstrates loading a 12-DOF legs-only locomotion policy
+(TorchScript .pt format) and running it on the Unitree G1 robot.
+
+Key characteristics:
+- Single TorchScript policy (.pt)
+- 47D observations, 12D actions (legs only)
+- Phase-based gait timing
+- Arms and waist held at fixed positions
+"""
+
+import argparse
+import logging
+import threading
+import time
+
+import numpy as np
+import torch
+from huggingface_hub import hf_hub_download
+from scipy.spatial.transform import Rotation as R
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+# 12-DOF leg joint configuration
+# Joint order: [L_hip_pitch, L_hip_roll, L_hip_yaw, L_knee, L_ankle_pitch, L_ankle_roll,
+#               R_hip_pitch, R_hip_roll, R_hip_yaw, R_knee, R_ankle_pitch, R_ankle_roll]
+LEG_JOINT_INDICES = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]
+
+# Default leg angles for standing
+DEFAULT_LEG_ANGLES = np.array([
+    -0.1, 0.0, 0.0, 0.3, -0.2, 0.0,   # left leg
+    -0.1, 0.0, 0.0, 0.3, -0.2, 0.0,   # right leg
+], dtype=np.float32)
+
+# KP/KD for leg joints
+LEG_KPS = np.array([150, 150, 150, 300, 40, 40, 150, 150, 150, 300, 40, 40], dtype=np.float32)
+LEG_KDS = np.array([6, 6, 6, 4, 2, 2, 6, 6, 6, 4, 2, 2], dtype=np.float32)
+
+# Waist configuration (held at zero)
+WAIST_JOINT_INDICES = [12, 13, 14]  # yaw, roll, pitch
+WAIST_KPS = np.array([250, 250, 250], dtype=np.float32)
+WAIST_KDS = np.array([5, 5, 5], dtype=np.float32)
+
+# Arm configuration (indices 15-28, held at initial position)
+ARM_JOINT_INDICES = list(range(15, 29))
+ARM_KPS = np.array([80, 80, 80, 80, 40, 40, 40,   # left arm (shoulder + wrist)
+                   80, 80, 80, 80, 40, 40, 40], dtype=np.float32)  # right arm
+ARM_KDS = np.array([3, 3, 3, 3, 1.5, 1.5, 1.5,
+                   3, 3, 3, 3, 1.5, 1.5, 1.5], dtype=np.float32)
+
+# Control parameters
+LOCOMOTION_CONTROL_DT = 0.02  # 50Hz control rate
+LOCOMOTION_ACTION_SCALE = 0.25
+ANG_VEL_SCALE = 0.25
+DOF_POS_SCALE = 1.0
+DOF_VEL_SCALE = 0.05
+CMD_SCALE = np.array([2.0, 2.0, 0.25], dtype=np.float32)
+MAX_CMD = np.array([0.8, 0.5, 1.57], dtype=np.float32)  # max vx, vy, yaw_rate
+
+# Gait parameters
+GAIT_PERIOD = 0.8  # seconds
+
+DEFAULT_REPO_ID = "nepyope/unitree_rl_locomotion"
+
+
+def load_torchscript_policy(
+    repo_id: str = DEFAULT_REPO_ID,
+    filename: str = "motion.pt",
+) -> torch.jit.ScriptModule:
+    """Load TorchScript locomotion policy from Hugging Face Hub.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID containing the policy.
+        filename: Policy filename (default: motion.pt).
+    """
+    logger.info(f"Loading TorchScript policy from Hugging Face Hub ({repo_id}/{filename})...")
+
+    policy_path = hf_hub_download(
+        repo_id=repo_id,
+        filename=filename,
+    )
+
+    policy = torch.jit.load(policy_path)
+    policy.eval()
+
+    logger.info("TorchScript policy loaded successfully")
+
+    return policy
+
+
+class UnitreeRLLocomotionController:
+    """
+    Handles 12-DOF legs-only locomotion control for the Unitree G1 robot.
+
+    This controller manages:
+    - Single TorchScript policy
+    - 47D observations (single frame)
+    - 12D action output (legs only)
+    - Arms and waist held at fixed positions
+    - Phase-based gait timing
+    """
+
+    def __init__(self, policy, robot, config):
+        self.policy = policy
+        self.robot = robot
+        self.config = config
+
+        # Velocity commands (vx, vy, yaw_rate)
+        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
+
+        # State variables (12 DOF legs)
+        self.qj = np.zeros(12, dtype=np.float32)
+        self.dqj = np.zeros(12, dtype=np.float32)
+        self.locomotion_action = np.zeros(12, dtype=np.float32)
+        self.locomotion_obs = np.zeros(47, dtype=np.float32)
+
+        # Initial arm positions (captured on reset)
+        self.initial_arm_positions = np.zeros(14, dtype=np.float32)
+
+        # Counter for phase calculation
+        self.counter = 0
+
+        # Thread management
+        self.locomotion_running = False
+        self.locomotion_thread = None
+
+        logger.info("UnitreeRLLocomotionController initialized")
+        logger.info("  Observation dim: 47, Action dim: 12 (legs only)")
+
+    def locomotion_run(self):
+        """12-DOF legs-only locomotion policy loop."""
+        self.counter += 1
+
+        if self.counter == 1:
+            print("\n" + "=" * 60)
+            print("🚀 RUNNING UNITREE RL 12-DOF LOCOMOTION POLICY")
+            print("   47D observations → 12D actions (legs only)")
+            print("   Arms and waist held at fixed positions")
+            print("=" * 60 + "\n")
+
+        # Get current observation
+        robot_state = self.robot.get_observation()
+        if robot_state is None:
+            return
+
+        # Get command from remote controller
+        if robot_state.wireless_remote is not None:
+            self.robot.remote_controller.set(robot_state.wireless_remote)
+        else:
+            self.robot.remote_controller.lx = 0.0
+            self.robot.remote_controller.ly = 0.0
+            self.robot.remote_controller.rx = 0.0
+            self.robot.remote_controller.ry = 0.0
+
+        self.locomotion_cmd[0] = self.robot.remote_controller.ly       # forward/backward
+        self.locomotion_cmd[1] = self.robot.remote_controller.lx * -1  # left/right (inverted)
+        self.locomotion_cmd[2] = self.robot.remote_controller.rx * -1  # yaw (inverted)
+
+        # Get leg joint positions and velocities (12 DOF)
+        for i, motor_idx in enumerate(LEG_JOINT_INDICES):
+            self.qj[i] = robot_state.motor_state[motor_idx].q
+            self.dqj[i] = robot_state.motor_state[motor_idx].dq
+
+        # Get IMU data
+        quat = robot_state.imu_state.quaternion
+        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
+
+        # Scale observations
+        gravity_orientation = self.robot.get_gravity_orientation(quat)
+        qj_obs = (self.qj - DEFAULT_LEG_ANGLES) * DOF_POS_SCALE
+        dqj_obs = self.dqj * DOF_VEL_SCALE
+        ang_vel_scaled = ang_vel * ANG_VEL_SCALE
+
+        # Calculate phase
+        count = self.counter * LOCOMOTION_CONTROL_DT
+        phase = (count % GAIT_PERIOD) / GAIT_PERIOD
+        sin_phase = np.sin(2 * np.pi * phase)
+        cos_phase = np.cos(2 * np.pi * phase)
+
+        # Build 47D observation vector
+        # [0:3]   - angular velocity (scaled)
+        # [3:6]   - gravity orientation
+        # [6:9]   - velocity command (scaled)
+        # [9:21]  - joint positions (12D, relative to default)
+        # [21:33] - joint velocities (12D, scaled)
+        # [33:45] - previous actions (12D)
+        # [45]    - sin_phase
+        # [46]    - cos_phase
+        self.locomotion_obs[0:3] = ang_vel_scaled
+        self.locomotion_obs[3:6] = gravity_orientation
+        self.locomotion_obs[6:9] = self.locomotion_cmd * CMD_SCALE * MAX_CMD
+        self.locomotion_obs[9:21] = qj_obs
+        self.locomotion_obs[21:33] = dqj_obs
+        self.locomotion_obs[33:45] = self.locomotion_action
+        self.locomotion_obs[45] = sin_phase
+        self.locomotion_obs[46] = cos_phase
+
+        # Run policy inference (TorchScript)
+        obs_tensor = torch.from_numpy(self.locomotion_obs).unsqueeze(0).float()
+        with torch.no_grad():
+            action_tensor = self.policy(obs_tensor)
+        self.locomotion_action = action_tensor.squeeze().numpy()
+
+        # Transform action to target joint positions
+        target_leg_pos = DEFAULT_LEG_ANGLES + self.locomotion_action * LOCOMOTION_ACTION_SCALE
+
+        # Debug logging (first 3 iterations)
+        if self.counter <= 3:
+            print(f"\n[Unitree RL Debug #{self.counter}]")
+            print(f"  Phase: {phase:.3f} (sin={sin_phase:.3f}, cos={cos_phase:.3f})")
+            print(f"  Cmd (vx, vy, yaw): ({self.locomotion_cmd[0]:.2f}, {self.locomotion_cmd[1]:.2f}, {self.locomotion_cmd[2]:.2f})")
+            print(f"  Action range: [{self.locomotion_action.min():.3f}, {self.locomotion_action.max():.3f}]")
+
+        # Send commands to LEG motors (0-11)
+        for i, motor_idx in enumerate(LEG_JOINT_INDICES):
+            self.robot.msg.motor_cmd[motor_idx].q = target_leg_pos[i]
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = LEG_KPS[i]
+            self.robot.msg.motor_cmd[motor_idx].kd = LEG_KDS[i]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # Hold WAIST motors at zero (12, 13, 14)
+        for i, motor_idx in enumerate(WAIST_JOINT_INDICES):
+            self.robot.msg.motor_cmd[motor_idx].q = 0.0
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = WAIST_KPS[i]
+            self.robot.msg.motor_cmd[motor_idx].kd = WAIST_KDS[i]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # Hold ARM motors at initial position (15-28)
+        for i, motor_idx in enumerate(ARM_JOINT_INDICES):
+            self.robot.msg.motor_cmd[motor_idx].q = self.initial_arm_positions[i]
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = ARM_KPS[i]
+            self.robot.msg.motor_cmd[motor_idx].kd = ARM_KDS[i]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # Send command
+        self.robot.send_action(self.robot.msg)
+
+    def _locomotion_thread_loop(self):
+        """Background thread that runs the locomotion policy at specified rate."""
+        logger.info("Locomotion thread started")
+        while self.locomotion_running:
+            start_time = time.time()
+            try:
+                self.locomotion_run()
+            except Exception as e:
+                logger.error(f"Error in locomotion loop: {e}")
+                import traceback
+                traceback.print_exc()
+
+            # Sleep to maintain control rate
+            elapsed = time.time() - start_time
+            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+        logger.info("Locomotion thread stopped")
+
+    def start_locomotion_thread(self):
+        if self.locomotion_running:
+            logger.warning("Locomotion thread already running")
+            return
+
+        logger.info("Starting locomotion control thread...")
+        self.locomotion_running = True
+        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
+        self.locomotion_thread.start()
+
+        logger.info("Locomotion control thread started!")
+
+    def stop_locomotion_thread(self):
+        if not self.locomotion_running:
+            return
+
+        logger.info("Stopping locomotion control thread...")
+        self.locomotion_running = False
+        if self.locomotion_thread:
+            self.locomotion_thread.join(timeout=2.0)
+        logger.info("Locomotion control thread stopped")
+
+    def reset_robot(self):
+        """Move legs to default standing position over 2 seconds (arms are captured and held)."""
+        logger.info("Moving legs to default position...")
+
+        total_time = 2.0
+        num_step = int(total_time / self.robot.control_dt)
+
+        # Get current state
+        robot_state = self.robot.get_observation()
+
+        # Capture initial arm positions (to hold during locomotion)
+        for i, motor_idx in enumerate(ARM_JOINT_INDICES):
+            self.initial_arm_positions[i] = robot_state.motor_state[motor_idx].q
+        logger.info(f"Captured initial arm positions: {self.initial_arm_positions[:4]}...")
+
+        # Record current leg positions
+        init_leg_pos = np.zeros(12, dtype=np.float32)
+        for i, motor_idx in enumerate(LEG_JOINT_INDICES):
+            init_leg_pos[i] = robot_state.motor_state[motor_idx].q
+
+        # Interpolate legs to default position
+        for step in range(num_step):
+            alpha = step / num_step
+
+            # Interpolate leg positions
+            for i, motor_idx in enumerate(LEG_JOINT_INDICES):
+                target_pos = DEFAULT_LEG_ANGLES[i]
+                self.robot.msg.motor_cmd[motor_idx].q = (
+                    init_leg_pos[i] * (1 - alpha) + target_pos * alpha
+                )
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = LEG_KPS[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = LEG_KDS[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            # Hold waist at zero
+            for i, motor_idx in enumerate(WAIST_JOINT_INDICES):
+                self.robot.msg.motor_cmd[motor_idx].q = 0.0
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = WAIST_KPS[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = WAIST_KDS[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            # Hold arms at initial position
+            for i, motor_idx in enumerate(ARM_JOINT_INDICES):
+                self.robot.msg.motor_cmd[motor_idx].q = self.initial_arm_positions[i]
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = ARM_KPS[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = ARM_KDS[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+
+        logger.info("Reached default leg position")
+
+        # Hold position for 2 seconds
+        logger.info("Holding default position for 2 seconds...")
+        hold_time = 2.0
+        num_hold_steps = int(hold_time / self.robot.control_dt)
+
+        for _ in range(num_hold_steps):
+            # Hold legs at default
+            for i, motor_idx in enumerate(LEG_JOINT_INDICES):
+                self.robot.msg.motor_cmd[motor_idx].q = DEFAULT_LEG_ANGLES[i]
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = LEG_KPS[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = LEG_KDS[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            # Hold waist at zero
+            for i, motor_idx in enumerate(WAIST_JOINT_INDICES):
+                self.robot.msg.motor_cmd[motor_idx].q = 0.0
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = WAIST_KPS[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = WAIST_KDS[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            # Hold arms at initial position
+            for i, motor_idx in enumerate(ARM_JOINT_INDICES):
+                self.robot.msg.motor_cmd[motor_idx].q = self.initial_arm_positions[i]
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = ARM_KPS[i]
+                self.robot.msg.motor_cmd[motor_idx].kd = ARM_KDS[i]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+
+        logger.info("Ready to start locomotion!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Unitree RL 12-DOF Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_REPO_ID,
+        help=f"Hugging Face Hub repo ID for policy (default: {DEFAULT_REPO_ID})",
+    )
+    parser.add_argument(
+        "--filename",
+        type=str,
+        default="motion.pt",
+        help="Policy filename (default: motion.pt)",
+    )
+    args = parser.parse_args()
+
+    # Load policy
+    policy = load_torchscript_policy(repo_id=args.repo_id, filename=args.filename)
+
+    # Initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    # Initialize locomotion controller
+    locomotion_controller = UnitreeRLLocomotionController(
+        policy=policy,
+        robot=robot,
+        config=config,
+    )
+
+    # Reset robot and start locomotion thread
+    try:
+        locomotion_controller.reset_robot()
+        locomotion_controller.start_locomotion_thread()
+
+        # Log status
+        logger.info("Robot initialized with Unitree RL locomotion policy")
+        logger.info("Locomotion controller running in background thread")
+        logger.info("Use remote controller to command velocity:")
+        logger.info("  Left stick Y: forward/backward")
+        logger.info("  Left stick X: left/right")
+        logger.info("  Right stick X: rotate")
+        logger.info("Press Ctrl+C to stop")
+
+        # Keep robot alive
+        while True:
+            time.sleep(1.0)
+    except KeyboardInterrupt:
+        print("\nStopping locomotion...")
+        locomotion_controller.stop_locomotion_thread()
+        print("Done!")
+

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.2"
+version = "0.4.3"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -107,6 +107,10 @@ dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
+unitree_g1 = [
+    "pyzmq>=26.2.1,<28.0.0",
+    "onnxruntime>=1.16.0"
+]
 reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [

src/lerobot/policies/pi05/modeling_pi05.py

@@ -538,6 +538,8 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         if config.compile_model:
             torch.set_float32_matmul_precision("high")
             self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
+            # Also compile the main forward pass used during training
+            self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
         msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
 

src/lerobot/robots/unitree_g1/__init__.py

@@ -0,0 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .config_unitree_g1 import UnitreeG1Config
+from .unitree_g1 import UnitreeG1

src/lerobot/robots/unitree_g1/config_unitree_g1.py

@@ -0,0 +1,55 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from ..config import RobotConfig
+
+_GAINS: dict[str, dict[str, list[float]]] = {
+    "left_leg": {
+        "kp": [150, 150, 150, 300, 40, 40],
+        "kd": [2, 2, 2, 4, 2, 2],
+    },  # pitch, roll, yaw, knee, ankle_pitch, ankle_roll
+    "right_leg": {"kp": [150, 150, 150, 300, 40, 40], "kd": [2, 2, 2, 4, 2, 2]},
+    "waist": {"kp": [250, 250, 250], "kd": [5, 5, 5]},  # yaw, roll, pitch
+    "left_arm": {"kp": [80, 80, 80, 80], "kd": [3, 3, 3, 3]},  # shoulder_pitch/roll/yaw, elbow
+    "left_wrist": {"kp": [40, 40, 40], "kd": [1.5, 1.5, 1.5]},  # roll, pitch, yaw
+    "right_arm": {"kp": [80, 80, 80, 80], "kd": [3, 3, 3, 3]},
+    "right_wrist": {"kp": [40, 40, 40], "kd": [1.5, 1.5, 1.5]},
+    "other": {"kp": [80, 80, 80, 80, 80, 80], "kd": [3, 3, 3, 3, 3, 3]},
+}
+
+
+def _build_gains() -> tuple[list[float], list[float]]:
+    """Build kp and kd lists from body-part groupings."""
+    kp = [v for g in _GAINS.values() for v in g["kp"]]
+    kd = [v for g in _GAINS.values() for v in g["kd"]]
+    return kp, kd
+
+
+_DEFAULT_KP, _DEFAULT_KD = _build_gains()
+
+
+@RobotConfig.register_subclass("unitree_g1")
+@dataclass
+class UnitreeG1Config(RobotConfig):
+    kp: list[float] = field(default_factory=lambda: _DEFAULT_KP.copy())
+    kd: list[float] = field(default_factory=lambda: _DEFAULT_KD.copy())
+
+    control_dt: float = 1.0 / 250.0  # 250Hz
+
+    # socket config for ZMQ bridge
+    robot_ip: str = "172.18.129.215"

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -0,0 +1,89 @@
+#!/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 enum import IntEnum
+
+# ruff: noqa: N801, N815
+
+NUM_MOTORS = 35
+
+
+class G1_29_JointArmIndex(IntEnum):
+    # Left arm
+    kLeftShoulderPitch = 15
+    kLeftShoulderRoll = 16
+    kLeftShoulderYaw = 17
+    kLeftElbow = 18
+    kLeftWristRoll = 19
+    kLeftWristPitch = 20
+    kLeftWristyaw = 21
+
+    # Right arm
+    kRightShoulderPitch = 22
+    kRightShoulderRoll = 23
+    kRightShoulderYaw = 24
+    kRightElbow = 25
+    kRightWristRoll = 26
+    kRightWristPitch = 27
+    kRightWristYaw = 28
+
+
+class G1_29_JointIndex(IntEnum):
+    # Left leg
+    kLeftHipPitch = 0
+    kLeftHipRoll = 1
+    kLeftHipYaw = 2
+    kLeftKnee = 3
+    kLeftAnklePitch = 4
+    kLeftAnkleRoll = 5
+
+    # Right leg
+    kRightHipPitch = 6
+    kRightHipRoll = 7
+    kRightHipYaw = 8
+    kRightKnee = 9
+    kRightAnklePitch = 10
+    kRightAnkleRoll = 11
+
+    kWaistYaw = 12
+    kWaistRoll = 13
+    kWaistPitch = 14
+
+    # Left arm
+    kLeftShoulderPitch = 15
+    kLeftShoulderRoll = 16
+    kLeftShoulderYaw = 17
+    kLeftElbow = 18
+    kLeftWristRoll = 19
+    kLeftWristPitch = 20
+    kLeftWristyaw = 21
+
+    # Right arm
+    kRightShoulderPitch = 22
+    kRightShoulderRoll = 23
+    kRightShoulderYaw = 24
+    kRightElbow = 25
+    kRightWristRoll = 26
+    kRightWristPitch = 27
+    kRightWristYaw = 28
+
+    # not used
+    kNotUsedJoint0 = 29
+    kNotUsedJoint1 = 30
+    kNotUsedJoint2 = 31
+    kNotUsedJoint3 = 32
+    kNotUsedJoint4 = 33
+    kNotUsedJoint5 = 34

src/lerobot/robots/unitree_g1/run_g1_server.py

@@ -0,0 +1,212 @@
+#!/usr/bin/env python3
+
+# 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.
+
+"""
+DDS-to-ZMQ bridge server for Unitree G1 robot.
+
+This server runs on the robot and forwards:
+- Robot state (LowState) from DDS to ZMQ (for remote clients)
+- Robot commands (LowCmd) from ZMQ to DDS (from remote clients)
+
+Uses JSON for secure serialization instead of pickle.
+"""
+
+import base64
+import contextlib
+import json
+import threading
+import time
+from typing import Any
+
+import zmq
+from unitree_sdk2py.comm.motion_switcher.motion_switcher_client import MotionSwitcherClient
+from unitree_sdk2py.core.channel import ChannelFactoryInitialize, ChannelPublisher, ChannelSubscriber
+from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
+from unitree_sdk2py.idl.unitree_hg.msg.dds_ import LowCmd_ as hg_LowCmd, LowState_ as hg_LowState
+from unitree_sdk2py.utils.crc import CRC
+
+# DDS topic names follow Unitree SDK naming conventions
+# ruff: noqa: N816
+kTopicLowCommand_Debug = "rt/lowcmd"  # action to robot
+kTopicLowState = "rt/lowstate"  # observation from robot
+
+LOWCMD_PORT = 6000
+LOWSTATE_PORT = 6001
+NUM_MOTORS = 35
+
+
+def lowstate_to_dict(msg: hg_LowState) -> dict[str, Any]:
+    """Convert LowState SDK message to a JSON-serializable dictionary."""
+    motor_states = []
+    for i in range(NUM_MOTORS):
+        temp = msg.motor_state[i].temperature
+        avg_temp = float(sum(temp) / len(temp)) if isinstance(temp, list) else float(temp)
+        motor_states.append(
+            {
+                "q": float(msg.motor_state[i].q),
+                "dq": float(msg.motor_state[i].dq),
+                "tau_est": float(msg.motor_state[i].tau_est),
+                "temperature": avg_temp,
+            }
+        )
+
+    return {
+        "motor_state": motor_states,
+        "imu_state": {
+            "quaternion": [float(x) for x in msg.imu_state.quaternion],
+            "gyroscope": [float(x) for x in msg.imu_state.gyroscope],
+            "accelerometer": [float(x) for x in msg.imu_state.accelerometer],
+            "rpy": [float(x) for x in msg.imu_state.rpy],
+            "temperature": float(msg.imu_state.temperature),
+        },
+        # Encode bytes as base64 for JSON compatibility
+        "wireless_remote": base64.b64encode(bytes(msg.wireless_remote)).decode("ascii"),
+        "mode_machine": int(msg.mode_machine),
+    }
+
+
+def dict_to_lowcmd(data: dict[str, Any]) -> hg_LowCmd:
+    """Convert dictionary back to LowCmd SDK message."""
+    cmd = unitree_hg_msg_dds__LowCmd_()
+    cmd.mode_pr = data.get("mode_pr", 0)
+    cmd.mode_machine = data.get("mode_machine", 0)
+
+    for i, motor_data in enumerate(data.get("motor_cmd", [])):
+        cmd.motor_cmd[i].mode = motor_data.get("mode", 0)
+        cmd.motor_cmd[i].q = motor_data.get("q", 0.0)
+        cmd.motor_cmd[i].dq = motor_data.get("dq", 0.0)
+        cmd.motor_cmd[i].kp = motor_data.get("kp", 0.0)
+        cmd.motor_cmd[i].kd = motor_data.get("kd", 0.0)
+        cmd.motor_cmd[i].tau = motor_data.get("tau", 0.0)
+
+    return cmd
+
+
+def state_forward_loop(
+    lowstate_sub: ChannelSubscriber,
+    lowstate_sock: zmq.Socket,
+    state_period: float,
+    shutdown_event: threading.Event,
+) -> None:
+    """Read observation from DDS and forward to ZMQ clients."""
+    last_state_time = 0.0
+
+    while not shutdown_event.is_set():
+        # read from DDS
+        msg = lowstate_sub.Read()
+        if msg is None:
+            continue
+
+        now = time.time()
+        # optional downsampling (if robot dds rate > state_period)
+        if now - last_state_time >= state_period:
+            # Convert to dict and serialize with JSON
+            state_dict = lowstate_to_dict(msg)
+            payload = json.dumps({"topic": kTopicLowState, "data": state_dict}).encode("utf-8")
+            # if no subscribers / tx buffer full, just drop
+            with contextlib.suppress(zmq.Again):
+                lowstate_sock.send(payload, zmq.NOBLOCK)
+            last_state_time = now
+
+
+def cmd_forward_loop(
+    lowcmd_sock: zmq.Socket,
+    lowcmd_pub_debug: ChannelPublisher,
+    crc: CRC,
+) -> None:
+    """Receive commands from ZMQ and forward to DDS."""
+    while True:
+        try:
+            payload = lowcmd_sock.recv()
+        except zmq.ContextTerminated:
+            break
+        msg_dict = json.loads(payload.decode("utf-8"))
+
+        topic = msg_dict.get("topic", "")
+        cmd_data = msg_dict.get("data", {})
+
+        # Reconstruct LowCmd object from dict
+        cmd = dict_to_lowcmd(cmd_data)
+
+        # recompute crc
+        cmd.crc = crc.Crc(cmd)
+
+        if topic == kTopicLowCommand_Debug:
+            lowcmd_pub_debug.Write(cmd)
+
+
+def main() -> None:
+    """Main entry point for the robot server bridge."""
+    # initialize DDS
+    ChannelFactoryInitialize(0)
+
+    # stop all active publishers on the robot
+    msc = MotionSwitcherClient()
+    msc.SetTimeout(5.0)
+    msc.Init()
+
+    status, result = msc.CheckMode()
+    while result is not None and "name" in result and result["name"]:
+        msc.ReleaseMode()
+        status, result = msc.CheckMode()
+        time.sleep(1.0)
+
+    crc = CRC()
+
+    # initialize DDS publisher
+    lowcmd_pub_debug = ChannelPublisher(kTopicLowCommand_Debug, hg_LowCmd)
+    lowcmd_pub_debug.Init()
+
+    # initialize DDS subscriber
+    lowstate_sub = ChannelSubscriber(kTopicLowState, hg_LowState)
+    lowstate_sub.Init()
+
+    # initialize ZMQ
+    ctx = zmq.Context.instance()
+
+    # receive commands from remote client
+    lowcmd_sock = ctx.socket(zmq.PULL)
+    lowcmd_sock.bind(f"tcp://0.0.0.0:{LOWCMD_PORT}")
+
+    # publish state to remote clients
+    lowstate_sock = ctx.socket(zmq.PUB)
+    lowstate_sock.bind(f"tcp://0.0.0.0:{LOWSTATE_PORT}")
+
+    state_period = 0.002  # ~500 hz
+    shutdown_event = threading.Event()
+
+    # start observation forwarding in background thread
+    t_state = threading.Thread(
+        target=state_forward_loop,
+        args=(lowstate_sub, lowstate_sock, state_period, shutdown_event),
+    )
+    t_state.start()
+
+    print("bridge running (lowstate -> zmq, lowcmd -> dds)")
+
+    # run command forwarding in main thread
+    try:
+        cmd_forward_loop(lowcmd_sock, lowcmd_pub_debug, crc)
+    except KeyboardInterrupt:
+        print("shutting down bridge...")
+    finally:
+        shutdown_event.set()
+        ctx.term()  # terminates blocking zmq.recv() calls
+        t_state.join(timeout=2.0)
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -0,0 +1,268 @@
+#!/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 struct
+import threading
+import time
+from dataclasses import dataclass, field
+from functools import cached_property
+from typing import Any
+
+import numpy as np
+from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
+from unitree_sdk2py.idl.unitree_hg.msg.dds_ import (
+    LowCmd_ as hg_LowCmd,
+    LowState_ as hg_LowState,
+)
+from unitree_sdk2py.utils.crc import CRC
+
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
+from lerobot.robots.unitree_g1.unitree_sdk2_socket import (
+    ChannelFactoryInitialize,
+    ChannelPublisher,
+    ChannelSubscriber,
+)
+
+from ..robot import Robot
+from .config_unitree_g1 import UnitreeG1Config
+
+logger = logging.getLogger(__name__)
+
+# DDS topic names follow Unitree SDK naming conventions
+# ruff: noqa: N816
+kTopicLowCommand_Debug = "rt/lowcmd"
+kTopicLowState = "rt/lowstate"
+
+G1_29_Num_Motors = 35
+G1_23_Num_Motors = 35
+H1_2_Num_Motors = 35
+H1_Num_Motors = 20
+
+
+@dataclass
+class MotorState:
+    q: float | None = None  # position
+    dq: float | None = None  # velocity
+    tau_est: float | None = None  # estimated torque
+    temperature: float | None = None  # motor temperature
+
+
+@dataclass
+class IMUState:
+    quaternion: np.ndarray | None = None  # [w, x, y, z]
+    gyroscope: np.ndarray | None = None  # [x, y, z] angular velocity (rad/s)
+    accelerometer: np.ndarray | None = None  # [x, y, z] linear acceleration (m/s²)
+    rpy: np.ndarray | None = None  # [roll, pitch, yaw] (rad)
+    temperature: float | None = None  # IMU temperature
+
+
+# g1 observation class
+@dataclass
+class G1_29_LowState:  # noqa: N801
+    motor_state: list[MotorState] = field(
+        default_factory=lambda: [MotorState() for _ in range(G1_29_Num_Motors)]
+    )
+    imu_state: IMUState = field(default_factory=IMUState)
+    wireless_remote: Any = None  # Raw wireless remote data
+    mode_machine: int = 0  # Robot mode
+
+
+class DataBuffer:
+    def __init__(self):
+        self.data = None
+        self.lock = threading.Lock()
+
+    def get_data(self):
+        with self.lock:
+            return self.data
+
+    def set_data(self, data):
+        with self.lock:
+            self.data = data
+
+
+class UnitreeG1(Robot):
+    config_class = UnitreeG1Config
+    name = "unitree_g1"
+
+    # unitree remote controller
+    class RemoteController:
+        def __init__(self):
+            self.lx = 0
+            self.ly = 0
+            self.rx = 0
+            self.ry = 0
+            self.button = [0] * 16
+
+        def set(self, data):
+            # wireless_remote
+            keys = struct.unpack("H", data[2:4])[0]
+            for i in range(16):
+                self.button[i] = (keys & (1 << i)) >> i
+            self.lx = struct.unpack("f", data[4:8])[0]
+            self.rx = struct.unpack("f", data[8:12])[0]
+            self.ry = struct.unpack("f", data[12:16])[0]
+            self.ly = struct.unpack("f", data[20:24])[0]
+
+    def __init__(self, config: UnitreeG1Config):
+        super().__init__(config)
+
+        logger.info("Initialize UnitreeG1...")
+
+        self.config = config
+
+        self.control_dt = config.control_dt
+
+        # connect robot
+        self.connect()
+
+        # initialize direct motor control interface
+        self.lowcmd_publisher = ChannelPublisher(kTopicLowCommand_Debug, hg_LowCmd)
+        self.lowcmd_publisher.Init()
+        self.lowstate_subscriber = ChannelSubscriber(kTopicLowState, hg_LowState)
+        self.lowstate_subscriber.Init()
+        self.lowstate_buffer = DataBuffer()
+
+        # initialize subscribe thread to read robot state
+        self._shutdown_event = threading.Event()
+        self.subscribe_thread = threading.Thread(target=self._subscribe_motor_state)
+        self.subscribe_thread.start()
+
+        while not self.is_connected:
+            time.sleep(0.1)
+
+        # initialize hg's lowcmd msg
+        self.crc = CRC()
+        self.msg = unitree_hg_msg_dds__LowCmd_()
+        self.msg.mode_pr = 0
+
+        # Wait for first state message to arrive
+        lowstate = None
+        while lowstate is None:
+            lowstate = self.lowstate_buffer.get_data()
+            if lowstate is None:
+                time.sleep(0.01)
+            logger.warning("[UnitreeG1] Waiting for robot state...")
+        logger.warning("[UnitreeG1] Connected to robot.")
+        self.msg.mode_machine = lowstate.mode_machine
+
+        # initialize all motors with unified kp/kd from config
+        self.kp = np.array(config.kp, dtype=np.float32)
+        self.kd = np.array(config.kd, dtype=np.float32)
+
+        for id in G1_29_JointIndex:
+            self.msg.motor_cmd[id].mode = 1
+            self.msg.motor_cmd[id].kp = self.kp[id.value]
+            self.msg.motor_cmd[id].kd = self.kd[id.value]
+            self.msg.motor_cmd[id].q = lowstate.motor_state[id.value].q
+
+        # Initialize remote controller
+        self.remote_controller = self.RemoteController()
+
+    def _subscribe_motor_state(self):  # polls robot state @ 250Hz
+        while not self._shutdown_event.is_set():
+            start_time = time.time()
+            msg = self.lowstate_subscriber.Read()
+            if msg is not None:
+                lowstate = G1_29_LowState()
+
+                # Capture motor states
+                for id in range(G1_29_Num_Motors):
+                    lowstate.motor_state[id].q = msg.motor_state[id].q
+                    lowstate.motor_state[id].dq = msg.motor_state[id].dq
+                    lowstate.motor_state[id].tau_est = msg.motor_state[id].tau_est
+                    lowstate.motor_state[id].temperature = msg.motor_state[id].temperature
+
+                # Capture IMU state
+                lowstate.imu_state.quaternion = list(msg.imu_state.quaternion)
+                lowstate.imu_state.gyroscope = list(msg.imu_state.gyroscope)
+                lowstate.imu_state.accelerometer = list(msg.imu_state.accelerometer)
+                lowstate.imu_state.rpy = list(msg.imu_state.rpy)
+                lowstate.imu_state.temperature = msg.imu_state.temperature
+
+                # Capture wireless remote data
+                lowstate.wireless_remote = msg.wireless_remote
+
+                # Capture mode_machine
+                lowstate.mode_machine = msg.mode_machine
+
+                self.lowstate_buffer.set_data(lowstate)
+
+            current_time = time.time()
+            all_t_elapsed = current_time - start_time
+            sleep_time = max(0, (self.control_dt - all_t_elapsed))  # maintain constant control dt
+            time.sleep(sleep_time)
+
+    @cached_property
+    def action_features(self) -> dict[str, type]:
+        return {f"{G1_29_JointIndex(motor).name}.pos": float for motor in G1_29_JointIndex}
+
+    def calibrate(self) -> None:  # robot is already calibrated
+        pass
+
+    def configure(self) -> None:
+        pass
+
+    def connect(self, calibrate: bool = True) -> None:  # connect to DDS
+        ChannelFactoryInitialize(0)
+
+    def disconnect(self):
+        self._shutdown_event.set()
+        self.subscribe_thread.join(timeout=2.0)
+
+    def get_observation(self) -> dict[str, Any]:
+        return self.lowstate_buffer.get_data()
+
+    @property
+    def is_calibrated(self) -> bool:
+        return True
+
+    @property
+    def is_connected(self) -> bool:
+        return self.lowstate_buffer.get_data() is not None
+
+    @property
+    def _motors_ft(self) -> dict[str, type]:
+        return {f"{G1_29_JointIndex(motor).name}.pos": float for motor in G1_29_JointIndex}
+
+    @property
+    def _cameras_ft(self) -> dict[str, tuple]:
+        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]:
+        return {**self._motors_ft, **self._cameras_ft}
+
+    def send_action(self, action: dict[str, Any]) -> dict[str, Any]:
+        self.msg.crc = self.crc.Crc(action)
+        self.lowcmd_publisher.Write(action)
+        return action
+
+    def get_gravity_orientation(self, quaternion):  # get gravity orientation from quaternion
+        """Get gravity orientation from quaternion."""
+        qw = quaternion[0]
+        qx = quaternion[1]
+        qy = quaternion[2]
+        qz = quaternion[3]
+
+        gravity_orientation = np.zeros(3)
+        gravity_orientation[0] = 2 * (-qz * qx + qw * qy)
+        gravity_orientation[1] = -2 * (qz * qy + qw * qx)
+        gravity_orientation[2] = 1 - 2 * (qw * qw + qz * qz)
+        return gravity_orientation

src/lerobot/robots/unitree_g1/unitree_sdk2_socket.py

@@ -0,0 +1,168 @@
+#!/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 base64
+import json
+from typing import Any
+
+import zmq
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+
+_ctx: zmq.Context | None = None
+_lowcmd_sock: zmq.Socket | None = None
+_lowstate_sock: zmq.Socket | None = None
+
+LOWCMD_PORT = 6000
+LOWSTATE_PORT = 6001
+
+# DDS topic names follow Unitree SDK naming conventions
+# ruff: noqa: N816
+kTopicLowCommand_Debug = "rt/lowcmd"
+
+
+class LowStateMsg:
+    """
+    Wrapper class that mimics the Unitree SDK LowState_ message structure.
+
+    Reconstructs the message from deserialized JSON data to maintain
+    compatibility with existing code that expects SDK message objects.
+    """
+
+    class MotorState:
+        """Motor state data for a single joint."""
+
+        def __init__(self, data: dict[str, Any]) -> None:
+            self.q: float = data.get("q", 0.0)
+            self.dq: float = data.get("dq", 0.0)
+            self.tau_est: float = data.get("tau_est", 0.0)
+            self.temperature: float = data.get("temperature", 0.0)
+
+    class IMUState:
+        """IMU sensor data."""
+
+        def __init__(self, data: dict[str, Any]) -> None:
+            self.quaternion: list[float] = data.get("quaternion", [1.0, 0.0, 0.0, 0.0])
+            self.gyroscope: list[float] = data.get("gyroscope", [0.0, 0.0, 0.0])
+            self.accelerometer: list[float] = data.get("accelerometer", [0.0, 0.0, 0.0])
+            self.rpy: list[float] = data.get("rpy", [0.0, 0.0, 0.0])
+            self.temperature: float = data.get("temperature", 0.0)
+
+    def __init__(self, data: dict[str, Any]) -> None:
+        """Initialize from deserialized JSON data."""
+        self.motor_state = [self.MotorState(m) for m in data.get("motor_state", [])]
+        self.imu_state = self.IMUState(data.get("imu_state", {}))
+        # Decode base64-encoded wireless_remote bytes
+        wireless_b64 = data.get("wireless_remote", "")
+        self.wireless_remote: bytes = base64.b64decode(wireless_b64) if wireless_b64 else b""
+        self.mode_machine: int = data.get("mode_machine", 0)
+
+
+def lowcmd_to_dict(topic: str, msg: Any) -> dict[str, Any]:
+    """Convert LowCmd message to a JSON-serializable dictionary."""
+    motor_cmds = []
+    # Iterate over all motor commands in the message
+    for i in range(len(msg.motor_cmd)):
+        motor_cmds.append(
+            {
+                "mode": int(msg.motor_cmd[i].mode),
+                "q": float(msg.motor_cmd[i].q),
+                "dq": float(msg.motor_cmd[i].dq),
+                "kp": float(msg.motor_cmd[i].kp),
+                "kd": float(msg.motor_cmd[i].kd),
+                "tau": float(msg.motor_cmd[i].tau),
+            }
+        )
+
+    return {
+        "topic": topic,
+        "data": {
+            "mode_pr": int(msg.mode_pr),
+            "mode_machine": int(msg.mode_machine),
+            "motor_cmd": motor_cmds,
+        },
+    }
+
+
+def ChannelFactoryInitialize(*args: Any, **kwargs: Any) -> None:  # noqa: N802
+    """
+    Initialize ZMQ sockets for robot communication.
+
+    This function mimics the Unitree SDK's ChannelFactoryInitialize but uses
+    ZMQ sockets to connect to the robot server bridge instead of DDS.
+    """
+    global _ctx, _lowcmd_sock, _lowstate_sock
+
+    # read socket config
+    config = UnitreeG1Config()
+    robot_ip = config.robot_ip
+
+    ctx = zmq.Context.instance()
+    _ctx = ctx
+
+    # lowcmd: send robot commands
+    lowcmd_sock = ctx.socket(zmq.PUSH)
+    lowcmd_sock.setsockopt(zmq.CONFLATE, 1)  # keep only last message
+    lowcmd_sock.connect(f"tcp://{robot_ip}:{LOWCMD_PORT}")
+    _lowcmd_sock = lowcmd_sock
+
+    # lowstate: receive robot observations
+    lowstate_sock = ctx.socket(zmq.SUB)
+    lowstate_sock.setsockopt(zmq.CONFLATE, 1)  # keep only last message
+    lowstate_sock.connect(f"tcp://{robot_ip}:{LOWSTATE_PORT}")
+    lowstate_sock.setsockopt_string(zmq.SUBSCRIBE, "")
+    _lowstate_sock = lowstate_sock
+
+
+class ChannelPublisher:
+    """ZMQ-based publisher that sends commands to the robot server."""
+
+    def __init__(self, topic: str, msg_type: type) -> None:
+        self.topic = topic
+        self.msg_type = msg_type
+
+    def Init(self) -> None:  # noqa: N802
+        """Initialize the publisher (no-op for ZMQ)."""
+        pass
+
+    def Write(self, msg: Any) -> None:  # noqa: N802
+        """Serialize and send a command message to the robot."""
+        if _lowcmd_sock is None:
+            raise RuntimeError("ChannelFactoryInitialize must be called first")
+
+        payload = json.dumps(lowcmd_to_dict(self.topic, msg)).encode("utf-8")
+        _lowcmd_sock.send(payload)
+
+
+class ChannelSubscriber:
+    """ZMQ-based subscriber that receives state from the robot server."""
+
+    def __init__(self, topic: str, msg_type: type) -> None:
+        self.topic = topic
+        self.msg_type = msg_type
+
+    def Init(self) -> None:  # noqa: N802
+        """Initialize the subscriber (no-op for ZMQ)."""
+        pass
+
+    def Read(self) -> LowStateMsg:  # noqa: N802
+        """Receive and deserialize a state message from the robot."""
+        if _lowstate_sock is None:
+            raise RuntimeError("ChannelFactoryInitialize must be called first")
+
+        payload = _lowstate_sock.recv()
+        msg_dict = json.loads(payload.decode("utf-8"))
+        return LowStateMsg(msg_dict.get("data", {}))

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -65,7 +65,6 @@ import argparse
 import gc
 import logging
 import time
-from collections.abc import Iterator
 from pathlib import Path
 
 import numpy as np
@@ -78,19 +77,6 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD
 
 
-class EpisodeSampler(torch.utils.data.Sampler):
-    def __init__(self, dataset: LeRobotDataset, episode_index: int):
-        from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-        to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
-        self.frame_ids = range(from_idx, to_idx)
-
-    def __iter__(self) -> Iterator:
-        return iter(self.frame_ids)
-
-    def __len__(self) -> int:
-        return len(self.frame_ids)
-
-
 def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
     assert chw_float32_torch.dtype == torch.float32
     assert chw_float32_torch.ndim == 3
@@ -119,12 +105,10 @@ def visualize_dataset(
     repo_id = dataset.repo_id
 
     logging.info("Loading dataloader")
-    episode_sampler = EpisodeSampler(dataset, episode_index)
     dataloader = torch.utils.data.DataLoader(
         dataset,
         num_workers=num_workers,
         batch_size=batch_size,
-        sampler=episode_sampler,
     )
 
     logging.info("Starting Rerun")