Reviewing robot plans before they run

Language and vision models can now propose multi-step plans for robots. On screen, the plans look sensible. In a lab, a sensible-looking plan can still break something.

Success in simulation is not safety

A simulator can report that a task was completed without modelling that glass cracks under a grip tuned for plastic, or that a short path crosses a heated surface. High simulated success tells you the plan finishes. It does not tell you the plan is safe.

A practical review checklist

Experienced technicians tend to ask the same questions: What is being touched, and is the force right for that material? What does the path pass over or near? What happens if a step fails halfway? Where is the clearance, and who verifies it?

Why this needs people

These are physical judgments, learned by working with real equipment. Reviewers who have that experience can catch hazards before a robot ever moves, and their notes become the safety criteria the next model is trained against.