Humanoid robots learn parkour maneuvers from human motion recordings

Researchers from Amazon Frontier AI & Robotics and UC Berkeley have developed a novel framework called Perceptive Humanoid Parkour (PHP) that enables humanoid robots to perform dynamic parkour maneuvers with human-like agility. By leveraging video recordings of human parkour movements, the team decomposed complex actions into atomic skills, which were then recombined using a motion matching approach based on nearest-neighbor search in feature space. This method allows robots to execute smooth, expressive, and adaptable sequences of movements such as running, jumping, climbing, vaulting, and rolling over obstacles in both urban and natural environments. The PHP framework integrates reinforcement learning to train controllers that use visual inputs from onboard depth sensors to plan and coordinate actions autonomously. This perception-driven decision-making enables the robot to select appropriate maneuvers in real time based on the geometry and height of obstacles. Validation experiments on a Unitree G1 humanoid robot demonstrated the framework’s effectiveness, with the robot successfully climbing obstacles up to 1.