This research introduces the ASIMOV Benchmark to evaluate and improve semantic safety of foundation models in robotics applications.

• Creates "robot constitutions" to define safe operational boundaries
• Develops testable safety scenarios in human-robot interactions
• Establishes a comprehensive benchmark for assessing semantic understanding and safety in robotic systems
• Addresses critical security vulnerabilities like hallucinations when LLMs control physical robots

This research is crucial for security as it provides a systematic approach to evaluate and prevent dangerous behaviors when AI systems control physical robots that interact with humans and the environment.

Generating Robot Constitutions & Benchmarks for Semantic Safety