Perceptive Humanoid Motion from Simulation

Researchers have introduced HALO, a humanoid autonomy system that synthesizes vision-language-motion data across procedurally generated environments to overcome the data bottleneck in robotics. The system acquires a repertoire of physical skills without relying on human teleoperation or real-world training data. Without prior exposure to real homes during training, HALO achieved a 91% success rate when transferred zero-shot to unmapped physical environments. Open article →