RoboRisk benchmark evaluates safety in agentic robotics
November 9, 2026
RoboRisk is a safety benchmark for agentic robotics featuring 92 simulated and 13 physical-robot tasks. It evaluates five critical categories: task-level refusal, plan repair, execution-time hazards, injection, and sequential memory.
HOW THIS AFFECTS YOU
●
researcherYou can test if your models fail during execution or when faced with malicious injections.
●
policyThis gives you a way to measure the actual safety risks of deploying foundation models on robots.