KungfuAthleteBot: learning high-dynamic humanoid motion from video with unified robust recovery
KungfuAthleteBot is a preprint pipeline that turns monocular video of elite martial artists into high-dynamic motions for a Unitree G1 humanoid by repairing trajectory physics, using physics-driven low-kinetic-energy initialization, and training one unified policy for tracking plus fall recovery. On real hardware the robot tracks a five-minute Tai Chi sequence and recovers from arbitrary falls in about 0.7 s, reported as the fastest for a unified policy. This matters because it shows video can replace expensive motion capture for extreme humanoid skills when reconstruction, missing actuation, and recovery are addressed together.