Breaking Down and Scaling Up: Game-Theoretic Algorithms for Open-world HRI

For robots to be reliably deployed in our homes, city streets, and airways, they must have a strong understanding of how they are influenced by-and can influence-people and the environment. In theory, foundational work in dynamic games offers us a rigorous mathematical formalism to model these feedback loops. However, in practice, bringing game-theoretic algorithms to truly open-world human-robot interactions has been limited by unrealistic assumptions about people and handcrafted representations of the world. My research systematically breaks down these rigid assumptions to enable robots to 1) infer game-theoretic information structures from deployment-time interaction data, 2) adapt their safety strategies using data-driven models of human behavior, and 3) generalize safety strategies to prevent human-intuitive but hard-to-specify failures.