UC Berkeley

As human-robot systems make their ways into our every day life, safety has become a core concern of the learning algorithms used by such systems. Examples include semi-autonomous vehicles such as automobiles and aircrafts. The robustness of controllers in such systems relies on the accuracy of models of human behavior. In this paper, we propose a systematic methodology for analyzing the robustness of learning-based control of human-cyber-physical systems. We focus on the setting where human models are learned from data, with humans modeled as approximately rational agents optimizing their reward functions. In this setting, we provide a novel optimization-driven approach to find small deviations in learned human behavior that lead to violation of desired (safety) objectives. Our approach is experimentally validated via simulation for the application of autonomous driving.