Signal Humanoid Robot Achieves 95% Real-Time Dodge Rate Using Only Onboard Camera Safety

Summary

A preprint posted July 30, 2026 introduces PAC-MAN, a perception-aware control-barrier-function reinforcement learning (CBF-RL) framework for whole-body safety in humanoid robots, tested via a dodgeball task. The deployed policy perceives an incoming ball only as segmentation-masked depth from a head-mounted camera, while training-time guidance uses a control-barrier function representing clearance to every body link, regularized by an adversarial motion prior. Evaluated on a benchmark with seeded throws across single-throw and deployment-loop regimes, the perception-only policy performs within a few points of a privileged state oracle, showing that a fixed onboard camera alone can be adequate for evasion. The researchers found that barrier structure choice depends on perceptual observability — a joint-level barrier performs best with accurate ball states but degrades under fixed-camera-only observation, recovering with a tracking gimbal or privileged filter. A lightweight link-level barrier policy was deployed zero-shot on a real Unitree G1 humanoid, tolerating imperfect perception and succeeding on 95% of throws using semantic segmentation to distinguish different balls.