Layered Multi-Agent Reactive-Behavior Learning in a Robotic Soccer

Abstract

Robotic soccer is an interesting test bench for the field of self-organizing and cooperating multi-agent systems. This paper deals with learning of two basic low-level behaviors that will enable the robotic player to participate further in higher-level collaborative and adversarial learning situations. First, a ball interception and obstacle avoidance behavior is learned. Then the acquired skills are incorporated into a next higher-level multi-agent learning scenario, namely the shooting ball behavior. The proposed control scheme for these behaviors consists of a trajectory generator with a layered structure, which supplies data to a trajectory-tracking controller.