3P2-Q02 Discovering novel actions in environment where little changes of action cause large changes of sensory information(Neurorobotics & Cognitive Robotics)

Abstract

Humans determine the action that they expect they can acquire unexperienced results. We try to give robots such ability. The ability is expected to make robots search things efficiently and discover an action to acquire unexperienced sensory information. We proposed a curiosity-based framework which decides a robot movement in a situation that has a locally linear relation between an action and the resulting sensory change. In this paper, we propose another curiositybased framework which generates a manipulation movement. In manipulation, there is a non-linear relation between an action and the resulting sensory change. In the experiment using a humanoid, it acquired sensory information which was unexperienced and hard to acquire by actions using a random search method.