Objective Unmanned surface vehicle (USV) swarms are increasingly deployed for complex maritime missions such as continuous environmental monitoring, island patrol, and strategic surveillance. However, in open and uncertain environments, existing methods still face severe limitations in online autonomous task allocation, dynamic replanning, and real-time adaptation to unexpected events. This study aims to enhance the decision-making autonomy, robustness, and task continuity of USV swarms during long-duration area coverage operations. To this end, we propose a real-time decision-making and cooperative control framework based on large language models (LLMs).

Method The proposed method introduces a hierarchical decomposition framework that integrates semantic interpretation with structured task planning. By embedding LLMs, natural language mission instructions are dynamically translated into executable objectives, enabling USV swarms to autonomously generate efficient and conflict-free navigation paths. A hybrid architecture is adopted, where a centralized planner performs global task allocation, while distributed planners on individual USVs compute local collision-free trajectories in parallel. To handle uncertainty, a dual closed-loop feedback mechanism is designed: (i) at the task level, dynamic prompt generation supports reallocation when resources change or failures occur; (ii) at the action level, incremental trajectory correction through geometric collision detection ensures safe adaptation without full replanning. This multilayered structure provides scalability, responsiveness, and resilience for long-term multi-agent operations.

Result The proposed framework was validated in a simulated maritime island patrol scenario, where multiple USVs collaboratively encircled randomly generated islands and returned to base. Comparative experiments against classical methods (A*, genetic algorithms, and ant colony optimization) demonstrate that our approach achieves comparable efficiency in terms of path length and travel time, while uniquely maintaining strict adherence to natural language mission goals. Cross-model evaluation with phi-4, Qwen-3, and DeepSeek R1 further shows that lightweight reasoning-oriented models strike the best balance between planning accuracy and computational efficiency. Experimental results indicate that the framework achieves an average task success rate of over 80% in complex, long-term, multi-task settings, while maintaining a success rate above 70% for real-time responses and handling of abnormal events (e.g., random USV failures). Ablation studies confirm that hierarchical task decomposition and multi-level feedback contribute significantly to overall performance improvement.

Conclusion This study demonstrates that LLMs can effectively facilitate decision-making for USV swarms in dynamic maritime environments. By combining semantic instruction understanding, hybrid centralized–distributed planning, and dual closed-loop feedback, the proposed framework ensures task continuity, adaptability, and safety. These findings highlight a promising pathway for deploying USV swarms in reliable, long-endurance missions and lay the foundation for extending this approach to large-scale, real-world multi-agent systems.