Evolving real-time behavioral modules for a robot with GP

In this paper we demonstrate an eecient method which divides a control task into smaller sub{tasks. We use a Genetic Programming system that rst learns the sub-tasks and then evolves a higher{level action selection strategy for deciding which of the evolved lower{level algorithms should be in control. The Swiss miniature robot Khepera is employed as the experimental platform. Results are presented which show that the robot is indeed able to evolve both the control algorithms for the diierent lower{level tasks and the strategy for the selection of tasks. The evolved solutions also show robust performance even if the robot is lifted and placed in a completely diierent environment or if obstacles are moved around.