Communication-Free Adaptive Swarm Robotic System: LLM-Based Decision Making and MARL-Based Multi-Policy Control

Swarm robotic systems consist of a large number of distributed autonomous robots that coordinate their actions to accomplish diverse tasks beyond the capabilities of a single robot. These systems have recently been considered for deployment in disaster scenarios, where communication is often unstable, making it necessary to achieve adaptive cooperative behavior without relying on explicit communication between robots. In the context of multi-robot systems—including swarm robotic systems—some studies have explored approaches utilizing large language models (LLMs) or other learning-based methods, but few have proposed systems that enable communication-free coordination. In this paper, we propose a system incorporating a novel method that combines high-level decision-making via LLM-based policy selection—guided by questionnaire-style prompts—with low-level control using multiple MARL-trained policies. We consider a complex task scenario in which robots search for a target object and transport it to a designated destination. To evaluate the method, we define implicit consensus as a condition in which a robot selects the same policy as its nearby robots without any explicit communication. The effectiveness of the proposed method is demonstrated through simulated task execution, with particular emphasis on implicit consensus as a key evaluation metric.