Enhancing Aggregation in Locomotor Multi-Agent Systems: a Theoretical Framework

The synthesis of collective behaviors in Multi-Agent Systems is typically approached using various methods, with Evolutionary Algorithms being among the most prevalent. In these systems, agents engage in local interactions with their peers and collectively adopt strategies that manifest at a group level, resembling social behaviors seen in animal societies. We extended the AntBullet problem, which is part of the PyBullet simulation tool, to a collective scenario involving a group of five homogeneous robots to aggregate during locomotion. To evolve this behavior, we employed the OpenAI-ES algorithm alongside a multi-objective fitness function. Our findings indicate that while the robots developed successful locomotion behaviors, they did not exhibit aggregation. This discrepancy is attributed to design choices that unintentionally emphasized locomotion over aggregation capabilities. We discuss the dynamic interplay induced by the fitness function to validate our results and outline future directions. Ultimately, our goal is a first attempt to establish a framework for analyzing collective behaviors using advanced algorithms within modern simulation environments.