czy213hd/go2_arx_mjlab

This repository adds support for a specific robot configuration to an existing simulation and training framework called mjlab. The robot being trained is a Unitree Go2, a four-legged walking robot, paired with an ARX L5 robotic arm mounted on top. Together the system has 18 joints the software can control: 12 in the legs and 6 in the arm. Reinforcement learning is a training technique where a simulated robot tries actions, receives feedback on how well it did, and gradually learns a useful behavior through millions of repeated attempts. This project defines two training tasks: one where the robot learns to walk and move its arm on flat ground, and one on rough terrain. The training runs inside a physics simulator rather than on a physical robot, and it requires an NVIDIA graphics card because the software runs thousands of simulated environments in parallel to speed up learning. Once training finishes, the project includes a sim-to-sim step: you run the trained policy inside a standard MuJoCo physics viewer to sanity-check behavior before putting it on real hardware. A keyboard-controlled version lets you steer the walking speed and move the arm's target position interactively from the terminal while watching the simulation. The keys map to things like forward velocity, side velocity, turning speed, and the arm's target coordinates along three axes. The simulation model, the task definitions, and the training scripts are all built as an extension on top of the mjlab framework, which handles the underlying physics and training loop. The README lists the relevant source files and explains how to resume from a checkpoint if a training run is interrupted. Real-robot deployment is described as still in progress. The project is released under the Apache 2.0 license, with separate license files for the Unitree Go2 and ARX L5 robot model assets that should be checked before any commercial use.