Using Inverse Planning and Theory of Mind for Social Goal Inference

Previous research shows that people assign latent goals or intentions to simple animated agents based on the motion behavior of these agents. We propose that human observers can infer that an animated agent has a partial state of belief about its environment and that observers use this information – in combination with the agent's observable behavior – to infer its goals. We conducted an experiment that showed that observers used line-of-sight cues – an agent's orientation relative to various objects in the environment, and the presence or absence of visual obstructions – to determine the content of an agent's state of belief about the location of objects. Our results are consistent with the hypothesis that human observers use line-of-sight cues to assign belief states to agents and that these belief states can be used to interpret agent behavior. We found that observer models that incorporated inferences about agents’ beliefs outperformed an all-knowing observer model in describing human responses. Additionally, we found that human responses were most consistent with the behavior of a model that incorporates information about both orientation and line-of-sight obstructions.