A Comparative Study of Evolutionary Strategies for Aggregation Tasks in Robot Swarms: Macro- and Micro-Level Behavioral Analysis

Collective decision-making is widely observed in natural organisms, especially insects and animals. In this regard, aggregation represents one of the paramount behaviors, as it can be useful for protecting groups against predators or speeding up the foraging process. In the field of autonomous robotics, aggregation is often studied through various paradigms, with evolutionary algorithms being one of the most widely used tools for evolving this collective behavior. In this work, we compared three modern evolutionary strategies — Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES), Exponential Natural Evolution Strategies (xNES) and OpenAI Evolutionary Strategy (OpenAI-ES) — for their ability to evolve an aggregation behavior in a swarm of robots. Specifically, we systematically varied the number of agents in the group, the environmental conditions (i.e., the number of target nests) and the parameters tuning the fitness function. Our aim is to verify whether and how the selected methods are effective at addressing the problem. The results we obtained indicate how OpenAI-ES achieves better performance in all the considered scenarios. Furthermore, it displays qualitatively more diverse strategies than the other two methods.