Robot learns to walk on 1K+ moving cylinders via 10M RL steps

hbouammar · x · 2026-09-21

A reinforcement learning demo where a quadruped robot learns to walk on a shifting ground made of 1,000+ moving cylinders in MuJoCo, forcing the policy to adapt in real time.

Training took roughly 10 million steps with the terrain maze constantly changing; the robot survived several near-falls before converging to a stable policy. The video shows real policy rollouts, not animation.

Original post →

More from Embodied

Embodied channel →