Evolving Adaptive Foraging Robot Swarms with NEAT in Environments with Obstacles

Abstract

We apply NeuroEvolution of Augmented Topologies (NEAT) to evolve adaptive and efficient swarm foraging behaviors in unknown environments with randomly placed obstacles. By rewarding effective actions and penalizing inefficient ones using the proposed strategy P-NeatFA, the training generates efficient foraging and obstacle avoidance strategies, reducing redundancy and outperforming traditional stochastic foraging algorithms. Optimization is guided by cumulative reward-based fitness, evaluated through simulations involving three types of distributed resources. Foraging performance is assessed in terms of resource retrieval rates. We compare the performance of our proposed P-NeatFA with that of CPFA and NeatFA. Experimental results show that P-NeatFA significantly outperforms the two strategies. The experimental results with three different numbers of robots in swarms show that the proposed strategy has better collision avoidance strategies and scalability when the swarm size increases. In future work, our goal is to integrate Federated Learning (FL) to develop a secure, scalable, and distributed swarm framework for real-world foraging applications.