Physical AI Eval Atlas: A Map of 94 Safety Evaluations for Robots and Embodied AI

At Convergent Robotics, we have released the Physical AI Eval Atlas ↗, a searchable catalogue of published safety evaluations for robots and embodied AI systems.

The Atlas maps how well the existing evaluations cover the risks of advanced physical AI. It currently lists 94 evaluations tagged by application domain, system architecture, threat model and harm class. Evals are scored on task complexity, dexterity demand, cognition demand and relevance to catastrophic risk. The catalogue can be filtered and sorted along all of these dimensions.

The main gap is that no current safety evaluation combines dexterous manipulation with causal or strategic reasoning. Safe behaviour in the physical world needs both strong cognitive abilities and dexterous manipulation. The majority of current benchmarks test either one of these dimensions or the other. The literature also focuses primarily on everyday physical harm in simulation: 70 of the 94 evaluations target local physical harm, 64 run in closed-loop simulation, and only 10 score 2 or higher (out of 3) for relevance to catastrophic risks. Only a handful of evaluations test whether a robot policy accepts correction or shutdown. Learned policies on real humanoids have not yet been evaluated from a safety perspective.

The full analysis is in the accompanying blog post, “Mapping Physical AI Safety Evaluations” ↗ (also reposted here). The Atlas will be updated as new evaluations appear on arXiv. If we have missed or mis-scored an evaluation, please get in touch!