Seek-and-Take Games of Heterogeneous Agent Teams with Large Language Model

The problem of confrontation games between multi-agent teams has attracted considerable interest, and the question of how to ensure effective coordination of heterogeneous agents in dynamic and adversarial environments remains a significant challenge. In this paper, we present a decision-making methodology for a confrontation game between two teams, which is based on large language models (LLMs). The game is designed as a competitive Seek-and-Take task comprising three distinct robot types: reconnaissance, grabbing, and blocking robots. The LLMs enhance the robot team's decision-making efficiency, delivering high-level commands that significantly improve coordination and adversarial performance. The game is initially formalized as a partially observable Markov decision process. Subsequently, a systematic construction of the LLM agents is provided, enabling real-time decisions based on the state information of a heterogeneous robot team (HRT). Detailed prompt design and the “Zero-shot-CoT” procedure are provided, and a cue word iteration method with reflections is proposed as a means of enhancing the decision-making efficiency of the LLMs and generating optimal collaboration strategies. Finally, the efficacy of the methodology is validated through experimentation with multiple LLMs in a multi-robot simulation environment. The results demonstrate that LLMs improve the efficiency and adaptability of decision-making in dynamic environments by serving as a core decision module, generating optimal strategies through real-time analysis.