From Motor Learning to Social Learning: A Study of Development on a Humanoid Robot

From Motor Learning to Social Learning: A Study of Development on a Humanoid Robot In this thesis, we describe how a humanoid robot designed to match the kinematics of a one-year old infant can learn to reach to visual targets, point toward visual targets, and share attention with a human. These three skills span the domains of motor learning and social learning. Instead of developing each of these skills independently, the robot naturally progresses from reaching to pointing and then from pointing to joint attention by taking advantage of social conventions and the assistance of a human. We first present a biologically plausible model for learning to reach to visual targets. This model is then extended to allow the robot to point toward distant objects that are outside of its reach without requiring additional learning. We demonstrate that the development of joint attention can be greatly facilitated by using pointing gestures to actively direct the attention of the robot’s caregiver, resulting in learning times that are two orders of magnitude less than comparable published models and eliminating the need for hand-labeling training data. A dedicated system to evaluate this idea is presented and its performance is demonstrated. From Motor Learning to Social Learning: A Study of Development on a Humanoid Robot A Dissertation Presented to the Faculty of the Graduate School of Yale University in Candidacy for the Degree of Doctor of Philosophy