Exploiting 'Like Me' Hypotheses in Learning Robots

University Of Vermont & State Agricultural College, Burlington VT

Investigators

Abstract

Exploiting 'Like Me' Hypotheses for Learning Robots Josh Bongard* & Andrew Meltzoff** *Department of Computer Science, University of Vermont **Institute for Learning & Brain Sciences, University of Washington All current robotic and computer technologies must be programmed by hand: such machines must not only be told what to do by a human, but how exactly to do it. Here it is proposed to construct a robotic device that can learn on its own by observing a human or robot teacher, inferring what that teacher is attempting to do, and then devising different or better actions to achieve the teacher's intended result. This surpasses simple copying, in which a robot reproduces the exact actions of a teacher. The proposed robot will first create internal simulations of itself and its teacher. It will then use these simulations to determine how the teacher is moving, and what physical consequences those movements will have. Finally, the robot can use the simulation of itself to find a new way of moving that will achieve the same goals. For instance, a human teacher may attempt and fail to lift a heavy object. The robot would infer the teacher's intent and use its grippers to lift the object in another way. Robotics has completely transformed heavy industry because such machines can perform the same actions repeatedly in a structured environment, such as a factory floor. Similarly, computers have revolutionized most sectors of human society by automating those aspects of them that can be performed in a repetitive manner. However, machines would be of equal or greater use in outdoor and unstructured environments--construction sites, homes, farms, and the surface of other planets--which constantly change, and therefore challenge the machine to continually alter how it achieves a given task. By observing others, a robot could learn both what to do (inferring the teacher's goals) and how to do it (creating new behaviors that achieve the intended result) on its own. Success could lead to intelligent devices capable to working alongside humans in the everyday world. The work would also contribute to our understanding of how the human brain develops a sense of 'self' and 'other' and how that supports social interaction and 'common ground' in communication. This work would also impact our understanding of a fundamental mechanism in human cognition, observational learning, by illuminating how students progress from simple rote copying to discovering novel solutions that improve upon solutions demonstrated by a teacher.