Automated Discovery of Efficient Behavior Strategies for Distributed Shape Formation of Swarm Robots by Genetic Programming

The distributed shape formation (DSF) of swarm robots is to form a specific shape, where each robot autonomously selects and moves to one target position in the desired shape. In this paper, we design some basic behaviors to construct behavior strategies, in which the swap selector and the obstacle avoidance planner are designed to guide robots in mitigating the negative effects of deadlock and collision during the DSF task. We propose a framework based on multi-agent simulation and genetic programming (GP) to automatically discover efficient behavior strategies followed by each robot to achieve DSF efficiently. The framework is distinguished by a centralized optimization of behavior strategies followed by each robot, along with strategy-driven decentralized decision-making and execution processes. In centralized optimization, to find behavior strategies with better generalization regarding shape types, a shape generator is designed to generate diverse training instances for comprehensively evaluating each behavior strategy in multi-agent simulation. Then, GP is used to evolve behavior strategies. In decentralized decision-making and execution, each robot can be guided by the same behavior strategy to form the shape. In terms of DSF completion time, the behavior strategies discovered by GP outperform the state-of-the-art distributed algorithm significantly across all test instances. These strategies can be well applied to different instances beyond the training instances. Through the statistical analysis, we also identify some crucial behaviors for realizing DSF.