Evolution of Hierarchical Controllers for Multirobot Systems

Abstract

Decentralized control for multirobot systems is difficult to design by hand because the behavioral rules for individual robots cannot, in general, be derived from a desired collective behavior. System designers have therefore resorted to evolutionary computation as a means to heuristically synthesize self-organized behaviors for robot collectives. Evolutionary computation is typically applied by putting the rules governing the individual robots under evolutionary control and by assigning fitness scores based on collective performance. Scaling evolutionary approaches to complex tasks has, however, proven challenging due to issues related to bootstrapping and premature convergence. In this paper, we show how hierarchical task decomposition and the combination of evolved and preprogrammed control can overcome these issues. We apply our approach to a complex multirobot task that requires a high degree of coordination and collective decision making, and we synthesize controllers capable of solving the task.