Silent Failures in Physical AI: A Literature Review of Runtime Action Authorization for Autonomous Systems

The paper reviews runtime guardrails for Physical AI, focusing on how autonomous systems can still output unsafe actions even when they appear correct.

It argues that current safety methods do not yet provide a complete boundary between black-box models and real-world physical action.

The framework centers on runtime authorization, uncertainty estimation, and verification to catch harmful actions before execution.

The research highlights the risk of silent failures in robotics foundation models and vision-language-action systems.

Its core message: safer deployment of Physical AI needs stronger runtime assurance, not just better model accuracy.