Objectives To address the poor adaptability and insufficient robustness of unmanned surface vehicle (USV) swarm path planning methods in complex environments, this paper proposes an large language model (LLM)-driven adaptive path planning with tool-function chains (APPT).

Methods The proposed method employs a planning encoder to assist the LLM in parsing environmental obstacle features. Combined with prompt engineering, it constructs an agent for USV swarm path planning, enabling dynamic assembly and optimization of classical path planning algorithms such as A, RRT, APF, and DWA. A similarity calculation strategy is used to achieve intelligent matching of tool chains, allowing flexible adaptation to diverse task requirements and complex obstacle environments while supporting user-guided adaptive iterative optimization.

Results Experimental results demonstrate that the APPT method achieves an average accuracy of 89.7% in effective tool selection across multiple scenarios, along with the capability for iterative optimization based on demands, reducing the total path length by 14.55%.

Conclusions The APPT method leverages the reasoning and analytical advantages of LLMs, significantly enhancing the intelligent decision-making ability of USV swarms. It provides a solution for path planning in complex environments that combines theoretical innovation and engineering practicality.