Tipping the scales: guidance and intrinsically motivated behavior

Abstract

We propose a novel approach to learning in autonomous robots that relies on the dynamical maintenance of an actively sensitized sensorimotor loop.Very weak learning cues are sufficient to orient a robot towards the desired behavior which is then selected from the intrinsic exploratory movements rather than imposed by a control command.The learning paradigm is a form of guided self-organization and is complementary to both active and intrinsically motivated learning.We present a systematic analysis of the learning algorithm in a robot control task and demonstrate its remarkable scalability with respect to the degrees of freedom of the system.