The Robot’s Camera Could Be Clear and Still Mislead

LIBERO-VPro perturbs what robot foundation models see while they act, testing degraded evidence, stale cameras, inconsistent sources and task-relevant scene changes. The authors built 96 experimental settings and 3,296 task-condition cases, evaluated six model families in roughly 196,000 simulated episodes and added 200 real-world rollouts. Some models tolerated severe object occlusion yet failed when local interaction cues were disrupted. The results show that clean-image benchmark scores can hide weaknesses in closed-loop control, though performance varies by model and perturbation.