Predicted Clearance and Realized Safety in Humanoid Hazard Avoidance

Embodied spatial reasoning is useful only when a geometrically favorable decision becomes favorable physical motion. We study this gap in simulated humanoid hazard avoidance. An onboard camera determines when a moving hazard could first be seen; from that time on, a privileged rule with perfect current geometry may choose between two fixed evasive routes, both executed by the same learned whole-body controller. This removes state-estimation error so that the experiment can ask whether spatial information remains meaningful after route selection, controller realization, and recovery. Across 480 paired scenarios, the informed intervention prevents 42 of 120 baseline contacts, reduces mean peak force by 54.9%, and reduces maximum-link force-time severity by 70.8%. However, this aggregate benefit is spatially inconsistent: avoidance is 63.3% for left-approach hazards and 6.7% for right-approach hazards, selected predicted clearance does not reliably track realized collision, and mirrored body-frame commands produce 0.15 versus 1.17m of stabilized lateral motion. Moreover, 69 of 78 residual contacts occur during recovery or stabilization; falls increase from 2 to 14 and goal successes decrease from 478 to 466. The results show that route ranking is not route safety. For embodied agents, spatial consistency must be evaluated through the physical trajectory the controller actually realizes, including the return to the task.