Czy213hd / Go2_ARX_mjlab

What is Go2_ARX_mjlab?

This project adds a Unitree Go2 quadruped robot with an ARX L5 arm to the mjlab reinforcement learning framework. It provides ready-made RL tasks for controlling both the robot's locomotion and arm manipulation simultaneously. The system trains policies to track base velocity commands while positioning the arm's end-effector to target poses. You get two terrain variants out of the box: flat ground and rough terrain for curriculum learning. Training uses RSL-RL with GPU acceleration via MuJoCo-Warp, and you can log to TensorBoard or Weights & Biases. Deployment scripts let you test trained policies in native MuJoCo before touching real hardware.

Why is it gaining traction?

The combined quadruped-plus-arm setup fills a gap for whole-body manipulation research. Most RL robot projects focus on either locomotion or arm control in isolation. This handles both with a single policy, which is useful for tasks like picking up objects while walking. The sim-to-sim deployment pipeline is practical for bridging simulation to reality. The keyboard-controlled viewer lets you manually command the robot and visualize target vs. actual end-effector poses in real time.

Who should use this?

Robotics researchers working on quadruped manipulation or whole-body control should evaluate this. If you're studying how to make a legged robot carry or place objects, this provides a solid starting point. Hardware teams with Go2 and ARX L5 arms can use the deployment scripts for sim-to-hardware validation. It's not suitable for production use yet given the early stage.

Verdict

This is a niche but useful addition to the legged robotics ecosystem. The 0.899% credibility score reflects a very new project with limited community adoption (14 stars), so expect to do some digging and potentially contribute fixes. If you need Go2 arm control in MuJoCo, this is worth a closer look. For general-purpose quadruped RL, the parent mjlab framework may be a better starting point.