GMATP-LLM: A General Multi-Agent Task Dynamic Planning Method using Large Language Models

In this work, we introduce a generalized task planning innovation framework named GMATP-LLM, which is designed for multi-agent systems. This method utilizes Chain-of-Thought (CoT) prompting to enable Large Language Models (LLMs) to perform task decomposition and assignment processes, transforming high-level task instructions into a set of sub-tasks. Based on the assignment strategy, it generates a PDDL goal plan, which is solved by an intelligent planner to generate a sequence of actions. The framework introduces a multi-agent three-dimensional Spatio-Temporal Motion Corridor (STMC) to constrain and optimize the parallel motion of the agents, improving system efficiency. This method combines the reasoning capabilities of LLMs and the fast solving advantage of the intelligent PDDL planners. It has been verified through simulation and real-world experiments across various task categories, achieving favorable results in multi-agent task planning.