humanoid-gym

Why we included this project

Teams training legged or humanoid robots will find this a solid starting point: it packages the whole reinforcement learning pipeline around Isaac Gym rather than leaving you to wire up the pieces yourself. The framework adds rewards tuned for humanoid robots, and zero-shot sim-to-real transfer has been verified on RobotEra's XBot-S and XBot-L, so the trained policies demonstrably survive contact with real hardware. Its Isaac Gym to MuJoCo sim-to-sim bridge lets you check a policy in a second physics engine before touching a real robot, which catches a lot of surprises early. The repo also walks through the Denoising World Model Learning approach from RSS 2024, adding state estimation and system identification techniques for more robust deployment. If you already work in Isaac Gym and your humanoid is close to the XBot family, the setup guides and ready-to-run training scripts get you from an empty workspace to a working locomotion policy fairly quickly.