MUTP-LLM: Empowering Multi-UAV Task Planning with Large Language Models

Although large language models (LLMs) have succeeded in natural language understanding, there are still many challenges in converting natural language instructions into understandable and executable action plans for UAVs while generalizing across different missions. This paper introduces MUTP-LLM, a novel hybrid framework that addresses this gap by integrating the semantic flexibility of LLMs with the formal rigor of traditional planners through a structured, multi-stage architecture. Specifically, our framework first employs an large language model (LLM) to translate ambiguous human commands into a structured symbolic task sequence. A hierarchical planner then generates a high-level plan by allocating and sequencing these tasks. Subsequently, another LLM instance grounds the abstract plan into concrete navigational waypoints. Crucially, a two-stage validation mechanism verifies the plan’s logical coherence and physical safety before it is dispatched for execution. Simulation experiments demonstrate that MUTP-LLM achieves superior performance in task success, planning robustness, and safety compared to end-to-end LLM or purely traditional approaches.