roboterax/humanoid-gym

Builds on NVIDIA Isaac Gym to train humanoid locomotion policies using PPO with multi-frame observation stacking, then validates robustness through a Mujoco sim-to-sim pipeline before real-world deployment. Includes a denoising world model learning framework for enhanced state estimation and system identification during sim-to-real transfer. Verified on RobotEra's XBot-S and XBot-L humanoids with zero-shot real-world execution.

1,894 stars. No commits in the last 6 months.