A decision architecture for unmanned surface vehicle swarms using large language models and prompt engineering

In unmanned surface vehicle (USV) swarm missions, challenges such as insufficient generalization, slow response in scenarios with small samples, limited resources, and complex dynamics often hinder target recognition, situational awareness, and autonomous decision-making control. This paper aims to develop an end-to-end multi-agent architecture that combines the reasoning capabilities of large language models with real-time control. The goal is to improve the collaborative combat and autonomous control capabilities of USV swarms in complex environments.

This study proposes a three-tier collaborative 'perception–understanding–decision' architecture for the maritime commander agent (MCA), integrating the Qwen2.5-72B large language model throughout the decision-making process of USV swarms. By combining prompt engineering with proportional-integral-derivative (PID) control, the approach achieves a deep integration of high-level strategic planning and low-level precise control. The system comprises: a target perception agent, which employs YOLOv8l to achieve high-precision detection and localization of multiple types of maritime targets; a situational understanding agent, which utilizes structured prompt templates to translate perception results into high-level natural language situational descriptions; and a decision-making agent, which combines large language model inference with external computational tools to generate control commands and optimizes response speed via PID regulation. This architecture requires no additional model fine-tuning and exhibits low latency and excellent adaptability.

Experimental results demonstrate that the system achieves high accuracy in object detection on the public ABOships dataset and under degraded conditions such as heavy fog and heavy rain. The semantic conversion accuracy in situational understanding reaches 93.5%. Simulation experiment results show that the success rate of the 4v1 encirclement task improves from 20% (with traditional rule-based methods) to 80%, while the success rate of the 10v10 adversarial task increases from 25% to 75%. These results validate the system's robustness and cross-domain generalization capabilities in complex maritime environments.

The proposed MCA architecture maintains the high-level cognitive reasoning capabilities of large language models while enhancing real-time response and execution accuracy through PID control, significantly improving the collaborative decision-making abilities of unmanned vehicle swarms in dynamic tasks. This research offers a practical technical framework for intelligent maritime swarm systems.