Learning From the Wizard: Programming Social Interaction through Teleoperated Demonstrations (Extended Abstract)

W. Bradley Knox (Massachusetts Institute of Technology), Samuel Spaulding (Massachusetts Institute of Technology), Cynthia Breazeal (Massachusetts Institute of Technology)

Abstract

This paper considers the question of whether robots can be effectively programmed for autonomous social interaction through learning from demonstrations recorded via Wizardof-Oz teleoperation. We present a novel LfW system for educational play between young children and a robot and results from a randomized experiment comparing a teleoperated robot and a robot with autonomous behavior derived by LfW. Across numerous metrics, the teleoperated robot and the autonomous robot programmed by LfW elicit similar behavior from their human interaction partners. Additionally, when children were asked whether the robot was humancontrolled or autonomous, approximately half in each condition thought it was human-controlled.