Learning for cooperation in multirobot team competitions

Abstract

We propose a learning architecture for cooperation in multirobot team competitions. This is a fully distributed, behavior-based software architecture, which facilitates flexible and reliable coordination of a team of robots performing tasks that may be subverted by another team of robots. Through the use of genetic algorithms, the robot team learns from past task execution experiences and improves its cooperation between the robots. The team performance in a game competition can be effectively improved. The feasibility of this architecture is demonstrated through simulation and practical experiments on a team of robots performing 3-on-3 robot soccer game.