Whom Will an Intrinsically Motivated Robot Learner Choose to Imitate from?

Abstract

This paper studies an interactive learning system that couples internally guided learning and social interaction in the case it can interact with several teachers. Socially Guided Intrinsic Motivation with Interactive learning at the Meta level (SGIMIM) is an algorithm for robot learning of motor skills in highdimensional, continuous and non-preset environments, with two levels of active learning: SGIM-IM actively decides at a metalevel when and to whom to ask for help; and an active choice of goals in autonomous exploration. We illustrate through an air hockey game that SGIM-IM efficiently chooses the best strategy.