Self-Aware Intelligent Medical Rescue Unmanned Team via Large Language Model and Multi-Agent Reinforcement Learning

Abstract: The evolution of medical rescue operations from traditional to future scenarios demands innovative approaches to intelligent automation. This paper proposes a self-award medical rescue team, leveraging cutting-edge technologies such as Large Language Models (LLMs), Multi-Agent Reinforcement Learning (MARL), and unmanned equipment to revolutionize frontline medical support. Our proposal of Unmanned and Digitalized Medical Rescue Team (UDMRT) emphasizes the decomposition of complex rescue tasks into manageable subtasks, facilitating improved learning and coordination among agents. This paper first implemented the rational assignment of roles for unmanned intelligent agents under complex tasks based on CMA-LLM (Cooperative Multi-Agent Large Language Models), forming combat teams with different functions. As for these teams, the paper proposed a novel hierarchical learning method designed for composite multi-agent tasks which addresses the challenge of coordination in complex domains by leveraging subtasks assignment. This method reduces observation spaces and encourages the reuse of subtask-specific policies, leading to more efficient learning and enhanced generalization capabilities. This architecture's modularity allows for better generalization to new environment configurations. Our work can adapt to new scenarios for practical applications in real-world multi-agent systems, where tasks are frequently comprised of discrete instances of localized interactions.