Humanoid-Gym

RobotEra's open-source reinforcement-learning framework for humanoid walking, built around zero-shot sim-to-real transfer.

Humanoid-Gym is a reinforcement-learning framework open-sourced in 2024 by the RobotEra (星动纪元) team; the paper's authors are Xinyang Gu, Yen-Jen Wang, and Jianyu Chen (Tsinghua IIIS), and it appeared at the ICRA 2024 workshop on agile robotics. It trains humanoid walking policies (with PPO) at massive scale on NVIDIA Isaac Gym, with code borrowed from legged_gym and rsl_rl. Its distinguishing feature is an Isaac Gym-to-MuJoCo sim2sim pipeline: retesting a trained policy in a different physics engine can expose behavior that only worked because it exploited quirks of the single original simulator, improving the odds of success on the real robot. The authors validated zero-shot transfer to the real world on two humanoid robots, the 1.2-meter XBot-S and the 1.65-meter XBot-L. It's also commonly used as a starting point for getting other humanoid robots walking with reinforcement learning.

ExampleA walking policy for XBot-L is first trained with PPO in Isaac Gym, then dropped into MuJoCo to run sim2sim checks for continued stability, and deployed to the real robot once it passes.

Also called

Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer

Related

legged_gym · Isaac Gym · Sim-to-Sim Transfer · Sim-to-Real Transfer · RobotEra · RL-based Locomotion Control

Sources

Humanoid-Gym (arXiv 2404.05695)
roboterax/humanoid-gym (GitHub)

As of

2024-04