One robot policy walks, runs, and gets up—no mode switching

Researchers extended Adversarial Motion Priors to train a unified policy that handles multiple locomotion modes on the Unitree G1 humanoid. The key insight: route training data to different discriminators based on body tilt angle (a simple threshold at 37° determines recovery vs. locomotion) and commanded velocity (which selects between walk and run). With just three reference motion clips, the frozen policy runs at 50 Hz on hardware, executing seamless walk-to-run transitions and recovering from falls without any mode-switching logic at runtime.