Agile perceptive multi-skill locomotion for quadrupedal robots in the wild

arXiv cs.AI · 18 hours ago

Researchers developed APT-RL, a reinforcement learning framework that enables quadrupedal robots to perform multiple locomotion skills and navigate complex terrains using only onboard sensors and computation. The system generates 2D motion datasets through trajectory optimization to train reusable skills that transfer to real robots, achieving peak speeds of 6 meters per second through dynamic maneuvers. A single onboard policy can now robustly traverse diverse obstacles including stairs, hurdles, gaps, and fallen branches in both indoor and outdoor environments.