Paper proposes automatic generation of robot constitutions and the ASIMOV semantic safety benchmark

Sermanet et al.'s paper proposes the ASIMOV benchmark for evaluating the semantic safety of foundation models acting as robot controllers, and a framework for automatically generating a constitution of behavior. The benchmark combines text and image generation, real visual data, and hospital injury reports, extending robot safety from geometric obstacle avoidance to semantic judgment—i.e., the model must judge whether an action is harmful in a specific context. The paper claims that a constitution iteratively refined through an automatic revision mechanism achieves an 84.3% alignment rate on ASIMOV, higher than baselines and human-written constitutions. The framework transfers Constitutional AI from language model alignment to robot control. Failure-First notes that the remaining 15.7% of failures may be concentrated in the most dangerous ambiguous edge cases.