motion_tracking_controller

ROS 2 Humble package for deploying reinforcement-learning (RL) policies to the Unitree G1 humanoid robot in MuJoCo simulation and on real hardware.
Policies trained in IsaacLab (and optionally stored in Weights & Biases) can be executed seamlessly using a unified ROS interface.

Features

• Load policies from WandB or local ONNX
• Run policies in LeggedGym/unitree_rl_gym MuJoCo simulation
• Deploy policies on the real Unitree G1 through SSH + g1_loco_client
• Real-time action publishing (joint torques / commands)
• Consistent sim-to-real pipeline

Repository Structure

motion_tracking_controller/
  package.xml
  setup.py
  CMakeLists.txt
  launch/
    mujoco.launch.py
    real.launch.py
  motion_tracking_controller/
    policy_node.py

Usage (Simulation)

ros2 launch motion_tracking_controller mujoco.launch.py wandb_path:=ENTITY/PROJECT/RUN

or

ros2 launch motion_tracking_controller mujoco.launch.py policy_path:=path/to/policy.onnx

Usage (Real Robot)

ros2 launch motion_tracking_controller real.launch.py robot_ip:=ROBOT_IP wandb_path:=ENTITY/PROJECT/RUN

or

ros2 launch motion_tracking_controller real.launch.py robot_ip:=ROBOT_IP policy_path:=path/to/policy.onnx

Requirements

• ROS 2 Humble
• Python 3.10+
• MuJoCo + LeggedGym (unitree_rl_gym) for simulation
• ONNX Runtime and WandB (if using WandB policies)

License

MIT License