Abstract
Perspective taking is the ability to take the point of view of another agent.This skill is not unique to humans as it is also displayed by other animalslike chimpanzees. It is an essential ability for efficient social interactions,including cooperation, competition, and communication. In this work, we presentour progress toward building artificial agents with such abilities. To this endwe implemented a perspective taking task that was inspired by experiments donewith chimpanzees. We show that agents controlled by artificial neural networkscan learn via reinforcement learning to pass simple tests that requireperspective taking capabilities. In particular, this ability is more readilylearned when the agent has allocentric information about the objects in theenvironment. Building artificial agents with perspective taking ability willhelp to reverse engineer how computations underlying theory of mind might beaccomplished in our brains.