What it does

Rex-gym is a collection of OpenAI Gym environments built around a 12-joint quadruped robot called Rex, based on the open-source SpotMicro design. It simulates gaits and poses in PyBullet, trains policies with PPO, and aims to transfer those policies directly to a real 3D-printed robot without re-tuning. A companion CLI handles batch training, policy replay, and rendered single-agent sessions.

The interesting bit

The project uses a hybrid policy that blends hand-designed trajectories with learned feedback: a(t, o) = a(t) + π(o). You can dial this from fully specified (inverse kinematics Bezier curves) to fully learned (open loop, from scratch) by adjusting bounds on the feedback component. The README frames this as a controlled experiment in how much prior knowledge helps.

Key highlights

• Supports both inverse-kinematics Bezier controllers and open-loop learning for gaits like walking, galloping, turning, and standing up
• Includes terrain randomization: flat plane, random heightfields, hills, mounts, and mazes
• Simulates the Poppy Ergo Jr robotic arm mounted on Rex’s back via --mark flag
• Ships as a PyPI package (pip install rex_gym) with a CLI for training and policy playback
• Targets Python 3.7 and uses 20 parallel agents by default for PPO training

Caveats

• The task checklist shows many unfinished items: backward walking, stairs, door opening, object grabbing, fall recovery, and mapping are all unchecked
• The README is vague on actual sim-to-real transfer results; it states the aim but not whether it has been achieved
• Requires Python 3.7 specifically, which is well past end-of-life

Verdict

Worth a look if you’re building legged RL from scratch on a budget and want a pre-built PyBullet environment with gait ablations. Skip it if you need mature sim-to-real validation or modern Python support.