Autonomous unmanned surface vehicle docking using large language model guide reinforcement learning

Autonomous docking of unmanned surface vehicles (USVs) represents the critical ”last mile” of intelligent navigation, presenting two main challenges: traditional control methods lack robustness in dynamic environments with disturbances such as wind and currents, while reinforcement learning (RL) methods suffer from low efficiency and often fail to transfer effectively from simulation to real-world applications. To tackle these issues, we propose LLM4SAC, a novel algorithm that integrates Large Language Models (LLMs) with the Soft Actor–Critic (SAC) framework to achieve USV autonomous docking tasks. LLM4SAC addresses these issues by leveraging the advanced contextual understanding and adaptive decision-making capabilities of LLMs. By providing high-level, context-specific guidance, LLMs enhance the RL agent's ability to interpret complex environmental data and adjust strategies in real time. This reduces the reliance on extensive simulated training datasets and increases the robustness and accuracy of the system under actual operating conditions. The dynamic request policy further refines the system's efficiency, querying LLMs only when necessary to minimize computational demands and interaction costs. Experiments in both simulation and real-world environments show that LLM4SAC significantly improves docking success rates, reduces computational costs, and enhances adaptability to dynamic conditions. Full implementation and resources are available on GitHub: https://github.com/RyanXu0428/LLM4SAC.