FAITH preprint reports humanoid safety gains while preserving task performance

Researchers report that their FAITH safety filter achieved a 99.95% safety rate while retaining 97% of unfiltered task return on a 29-degree-of-freedom humanoid in Walking-Avoid. The preprint also describes demonstrations of the same policies on a real-world Unitree G1 humanoid. FAITH learns to filter actions without requiring an analytic dynamics model. When no action meets its learned safety condition, it approaches the action with the lowest predicted peak harm. In Push-Avoid, the system achieved the highest measured safety rate by sacrificing balance and falling away from the protected region.