nicrusso7/rex-gym

Implements hybrid RL policies combining user-specified open-loop trajectories (Bezier inverse kinematics gaits) with learned feedback components using PPO, enabling sim-to-real transfer without manual tuning. Provides PyBullet simulation with multiple terrain types (heightfields, hills, mazes) and a CLI for batch training, policy playback, and rendered debugging. Supports both base quadruped and extended arm configurations, with modular environment parameters for exploring locomotion tasks from static poses to terrain navigation.

1,084 stars and 50 monthly downloads. No commits in the last 6 months. Available on PyPI.