Safe Robot Planning Through Reachability Analysis of Human Behaviors

This paper introduces a novel approach to human-robot interaction that enables robots to safely navigate shared spaces by actively gathering information about human internal states. We model human behavior as influenced by latent variables representing attentiveness and driving style, which cannot be directly observed by the robot. Our key contribution is a sampling-based reachability analysis method that integrates belief updates over human internal states with adaptive interval refinement, allowing robots to maintain probabilistic safety guarantees while efficiently planning their actions. Unlike passive observation approaches, our framework enables robots to purposefully execute probing actions that clarify human internal states, accelerating belief convergence. We implement our approach in an autonomous driving scenario at intersections, where the robot vehicle can actively test whether a human driver is attentive or distracted through subtle movements. Experimental results demonstrate that our method significantly outperforms passive estimation techniques in accuracy and convergence speed while maintaining safety constraints. The framework generalizes to various human-robot interaction domains where understanding human internal states is critical for effective collaboration and safety,