Proceedings of the Annual Meeting of the Cognitive Science Society

Humans demonstrate remarkable abilities to perceive physi-cal and social events based on very limited information (e.g.,movements of a few simple geometric shapes). However, thecomputational mechanisms underlying intuitive physics andsocial perception remain unclear. In an effort to identify thekey computational components, we propose a unified psycho-logical space that reveals the partition between the perceptionof physical events involving inanimate objects and the percep-tion of social events involving human interactions with otheragents. This unified space consists of two prominent dimen-sions: an intuitive sense of whether physical laws are obeyedor violated; and an impression of whether an agent possessesintentions, as inferred from movements. We adopt a physicsengine and a deep reinforcement learning model to synthe-size a rich set of motion patterns. In two experiments, humanjudgments were used to demonstrate that the constructed psy-chological space successfully partitions human perception ofphysical versus social events.