Audit Before You Commit
Locating belief failures in active identification for one‑shot manipulation
Locating belief failures in active identification for one‑shot manipulation
ICRA 2027
Five taps, then one commit. The robot is never told where the pocket is. Each tap reads plate or hole through a touch switch, the belief updates, and the arm commits once. Played at 10×.
A robot that probes a few times before one irreversible action must decide when the evidence is enough to commit. We argue that this decision rests on two conditions that existing methods do not separate: the belief must still cover the truth in the coordinate that decides the action, and the failure model that scores actions must track realized failure. We audit both, separately and offline with ground truth, on a deployed probe‑then‑commit pipeline. On simulated insertion, more taps sharpen the belief while the truth leaves its support on 16.9% of episodes and the failure score turns optimistic by 0.31. Conformal calibration restores coverage but not the decision. The audit's signatures instead point at the observation model, where a hand scan finds a 2.1 mm error in the tap boundary; correcting that one number cuts failure from 0.354 to 0.112 on untouched instances and transfers unrefitted to a second engine. On a physical arm inserting a tool into a rigid pocket by touch, the gain and both audit conditions reproduce, and replaying the recorded taps under an injected model `error shows the audit's signature on real data.
Left, the blind arm never observes and commits at the prior mean. Right, five taps at fixed positions, each read as top or pocket, sharpen a particle belief over the hole's position before the one commit.
Two more positions, same crop, same clock. Between them the block was slid 32 mm by hand.
Twelve positions, slid by hand to places the robot was never told, 118 episodes at one to five taps. The truth is found afterwards by the same touch switch that took the taps.