When Does Test-Time Physical Diagnosis Pay? A Frozen Policy Buys Evidence It Never Reads

This paper addresses test-time physical diagnosis for robots operating under unfamiliar physical conditions. The authors identify six ordered empirical conditions that must hold for diagnosis to improve robot behavior: meaningful reference, identifiability of physical conditions, evidence utilization, decision value, selection value over fixed alternatives, and safe realization. The work validates this theoretical chain through experiments in both controlled simulation environments and public real-world settings, providing empirical guidance for when robot physical diagnosis is worthwhile.