A Bayesian Robot That Distinguishes “Self” from “Other”

A Bayesian kinesthetic-visual matching model allows a humanoid robot to perform mirror self-recognition without social understanding. The robot learns the relationship between its own motor activity and perceived motion by observing the movements of its arm for four minutes. Over this time, it builds a simple dynamic Bayesian model that relates these two events, and uses this self model to build a model of “animate others” that is a copy of its self model but with the motor state hidden. Presented with a mirror, the robot then judges its mirror image to match its “self” model, while people are judged to be “animate others.”