Researchers Introduce Perceptive Humanoid Parkour Framework

Researchers at Amazon Frontier AI & Robotics and UC Berkeley introduce perceptive humanoid parkour (PHP), a framework published on arXiv that combines motion-matching and reinforcement learning to generate human-like agile locomotion from parkour video data. Distilled into a unified controller and validated on a Unitree G1, PHP climbs obstacles up to 1.25m and performs perception-driven, long-horizon multi-obstacle traversal using onboard depth sensing and a discrete 2D velocity command.