Goal-Oriented Control of Self-Organizing Behavior in Autonomous Robots

Abstract

We study adaptive control algorithms within a dynamical systems approach for autonomous robots that cause the self-organization of coordinated behaviors without specific goals or particular information about the physical body. A self-exploration is achieved that causes different body- and environment-related behaviors to emerge and to change during the learning process. We propose several methods to guide the self-organization towards specific behaviors, which is particularly useful in high-dimensional systems. An unsupervised extraction of behavioral primitives is achieved with a set of competing neural networks, where each network develops to a controller for one behavior. Finally we combine classical reinforcement learning with the behavioral primitives to obtain goal-oriented behaviors. The algorithms are verified with realistically simulated robots using our own simulator LpzRobots, which is briefly described.