PHP brings parkour to humanoids

This is the kind of robotics paper that feels like a real capability jump, not just a better demo reel. PHP matters because it turns parkour from isolated scripted motions into perception-driven skill composition on a real humanoid.

• –The core trick is combining motion matching from human parkour data with RL-trained controllers, then distilling them into one depth-based multi-skill policy
• –The robot can decide when to step over, vault, climb, or roll off obstacles based on onboard sensing instead of replaying a fixed sequence
• –Real-world tests on Unitree G1 include roughly 3 m/s vaults, 1.25 m wall climbs, and minute-long traversals with obstacle perturbations
• –For robotics engineers, the bigger takeaway is long-horizon chaining and smooth transitions between skills, which is a harder problem than a single athletic maneuver
• –It is still a preprint on structured obstacle courses, so this is progress toward robust physical AI rather than proof that humanoids are ready for messy real-world jobs