wty-yy/go2_rl_lab

What it solves

This project provides a framework for training and deploying high-speed, agile locomotion policies for the Unitree Go2 quadruped robot. It bridges the gap between simulation and real-world performance by enabling robust reinforcement learning (RL) training in Isaac Lab and MuJoCo, followed by deployment on physical hardware.

How it works

It uses Reinforcement Learning (specifically PPO and Mixture-of-Experts/CTS algorithms) to train robot policies within NVIDIA's Isaac Lab simulation environment. The framework supports multiple physics backends, including Newton MuJoCo-Warp and Isaac Sim's PhysX. Once trained, policies can be exported as TorchScript or ONNX files for simulation-to-real (Sim2Real) deployment. The project also integrates with RoboGauge for asynchronous, automated evaluation of policy performance and safety during the training process.

Who it’s for

This is for robotics researchers and engineers working on legged locomotion, reinforcement learning, and Sim2Real transfer who need a structured environment to train and benchmark quadrupedal robot controllers.

Highlights

• Supports high-speed locomotion (up to 3.5 m/s forward and 4.4 m/s backward).
• Built on Isaac Lab 3.0 EA and Isaac Sim 6.1, with support for Newton MuJoCo-Warp physics.
• Includes a Mixture-of-Experts (MoE-CTS) training approach for improved robustness on complex terrains like stairs and slopes.
• Provides a unified CLI for training, playback, and Sim2Sim/Sim2Real deployment.
• Integrated with RoboGauge for automated, asynchronous policy performance evaluation.