Evolving cooperative robotic behaviour using distributed genetic programming

Cooperative robotic systems provide design and implementation challenges that are not easy to solve. This paper describes a parallel implementation for evolving cooperative robotic behaviour using an island model based genetic program on a cluster computer system. The application domain is robot soccer in which two robots must cooperate to avoid collisions with each other and score goals. The system has access to robot and ball positions and outputs velocity set points for the robot wheel motors. The evolved controllers are evaluated on a kinematic model that has been optimized to improve the time complexity of the genetic programming algorithm. The inter-process communication on the cluster is implemented using the message passing interface (MPI).