In order to function safely and autonomously, modern robotic systems need to understand other agents' mental states, including their beliefs and desires about the shared environment. This ability, known as Theory of Mind (TOM), is crucial for self-driving vehicles, as the exchange of beliefs through instantaneous maneuvers gives rise to nuanced social behavior in human driving and the lack of such exchange can lead to traffic conflicts and crashes. For social scientists and engineers, mental states serve as a compact representation of agent behavior, which can be used to understand human cognition, devise interventions on human cognitive limitations, and build autonomous agents to assimilate human behavior. However, the TOM ability in both humans and machines is not well-understood, with an important question being the unbounded possibility of agent beliefs leading to degenerate inference. In this dissertation, I study the possibility and advantages of TOM in the context of modeling human driving behavior. By proposing a set of algorithms to make inference about human drivers' mental states, I elicit the implicit assumptions in human's TOM ability. I show that human TOM likely involves a delicate balance between being realist about the environment and the unbounded imagination in a Bayesian fashion. These observations were engineered into the proposed algorithms, resulting in substantial improvements in interpreting abnormal human behavior, inspecting model failures, and robustifying control policies.
