Abstraction of state-action space by utilizing properties of the body and the environment - Application to an autonomous snake-like robot controlled by reinforcement learning -

Abstract

In this paper we consider autonomous control of a real snake-like robot using reinforcement learning. We focus on curse of dimensionality and lack of generality, and point out that the causes of these problems are not in learning algorithm but in neglect of properties of the real world. To solve these problems we propose new framework in which the body of the robot abstracts general meanings by interacting with environment. We apply the proposed framework to control of a real snake-like robot and confirm that the two problems are solved.