PROTEA robot hazard memory can over-report risks while tracking delayed effects

A robot action can be locally acceptable and still create a later unsafe state. PROTEA proposes carrying forward environmental state and asking a safety checker about the consequences of the entire sequence. The same memory can accumulate mistaken hazard assumptions and cause excessive refusal of otherwise safe actions.

The PROTEA paper reports tracking object and task state across robotic actions to detect deferred hazards, while some evaluated LLM judges exhibit materially more false-positive alerts in its benchmark. The source was published in January 2026 and resurfaced in October conference reporting. Demonstrations with toy objects and selected judges cannot be generalized to injury rates on deployed robots.

Use deterministic state constraints for known physical hazards and evaluate judge precision as well as recall across longer trajectories and changed objects. Safety memory needs verifiable state updates. More retained context by itself does not guarantee better decisions. Primary evidence.

Additional evidence and trust boundary

A robot plan can be locally benign at each step yet unsafe when object state accumulates across a longer task. Keeping memory of objects and their permitted uses helps a judge notice delayed risks, while a language-model judge can also invent hazards. For irreversible motions, a deterministic interlock and verified environmental state must remain separate from a probabilistic text-based safety verdict. Supporting record.