SwarMind: Harnessing Large Language Models for Flock Dynamics

Abstract

The deployment of autonomous agent swarms has witnessed a rapid increase in a variety of fields, from logistics to surveillance. This prevalence stems from the complex objectives they can accomplish through simple interactions, their adaptive behavior, and their inherent robustness to potential disturbances. Despite these advantages, achieving control over such systems remains a formidable challenge, often necessitating the use of approximations or domain-specific heuristics. As demonstrated by recent works, large language models (LLMs) display a robust capacity to excel across a diverse array of tasks. In this work, we extend the exploration of LLM's capabilities into the niche domain of flock driving. Specifically, our study presents a comparative analysis of LLMs and reinforcement learning (RL) in a fair setting, scrutinizing their performances under various prompting strategies. Furthermore, it investigates the potential of eliciting more sophisticated behaviors from LLMs through textual instructions, offering a deeper understanding of their limitations and strengths in swarm control and management. The results illuminate several potential shortcomings while con-currently uncovering exciting prospects and extensions. This research therefore advances our understanding of the applicability of LLMs to the intricate field of swarm control, opening doors to their potential use in domains hitherto unexplored.