Research on multi-agent collaboration in path planning based on large language model

Abstract

To address the challenges of traditional path planning's lack of adaptability in dynamic environments and the limited decision-making capabilities of single agents, this paper proposes a multi-agent collaborative path planning framework, MAGIC–PPF. This framework, comprised of collaborative, navigation, state, and decision-making agents, forms a closed-loop process of "static computation—dynamic correction—multi-objective decision-making," achieving efficient and robust path planning. Experiments were conducted using a real-world geographic path using the LLM-as-a-Judge method to evaluate different model combinations based on accuracy, efficiency, robustness, and feasibility. Results demonstrate that MAGIC–PPF outperforms single-model approaches in overall performance, particularly in terms of adaptability and user experience under dynamic conditions. This provides a viable solution for applications in intelligent transportation, emergency dispatch, and unmanned systems.