Adaptive Behavior Learning for a Distributed Autonomous Swimming Robot in an Environment Including a Narrow Gate

Abstract

This paper describes a design method for an autonomous robot system consisting of plural agents that enable the robot to behave adaptively in the real world. One difficulty in designing a distributed autonomous system is how to embed dynamics for self-organizing environment-oriented action rules in the system. We, therefore, propose a robot system consisting of mechanically constrained agents that have been identically designed, which are controlled by a decision-making method that has an oscillator and learning methods based on the same object functions, such that each agent actively interacts with other agents and with the outer world. In order to verify the usefulness of this system, tests on behavioral acquisition in target approaching and obstacle avoidance were carried out by using a distributed autonomous swimming robot. Moreover, for learning efficiency in complex tasks, such as obstacle avoidance, we propose “Switching-Q learning,” in which previously obtained action rules are effectively used. It was found that the robot acquired simple obstacle avoidance behavior and dynamics based on interaction among agents and the outer world. It was, therefore, verified that the proposed robot design method is one solution for a system that can adapt to complex environments.