Semantic-Guided Risk-Sensitive Autonomous Navigation Method for Multiple Unmanned Surface Vehicles Using Large Language Models

Objectives A semantic-guided risk-sensitive autonomous navigation method based on a large language model is proposed to address the problem of difficulty in accurately characterizing collision avoidance decision-making risks in complex and restricted water areas where multiple unmanned vehicles are subject to multiple encounters, static obstacles, and local eddy current disturbances during autonomous navigation. Methods An IQN-based decentralized navigation policy is developed to enable each unmanned surface vehicle to select actions independently according to its local observations, including the relative target direction, obstacle information, and motion states of neighboring vehicles. A large language model (LLM) is employed as an offline semantic teacher to interpret local encounter situations and generate structured semantic labels, including encounter type, avoidance intention, interaction target, risk level, and confidence. These semantic labels are then distilled into a lightweight semantic guidance model for online decision-making. The semantic guidance vector is further incorporated into the policy input, and the CVaR parameter is adaptively adjusted according to the predicted semantic risk level and confidence, so that the policy places greater emphasis on lower-tail return estimates in high-risk states. Results Simulation results show that, in high-risk test scenarios involving 3~6 unmanned surface vehicles, LLM-CVaR-IQN consistently maintains a task success rate above 0.92 and outperforms Geo-CVaR-IQN, IQN, DQN, RVO, and APF. In the complex six-USV scenario, the proposed method still achieves a success rate of 0.92, which is 6% and 12% higher than those of Geo-CVaR-IQN and IQN, respectively. Meanwhile, the distributions of navigation time and action energy consumption are more concentrated, demonstrating better adaptability to complex encounters and stronger risk robustness. Conclusions The results indicate that the semantic guidance of the large language model can enhance the recognition ability of the strategy for local risks. Combined with IQN reward distribution modeling and CVaR risk-sensitive action selection, it can improve the task completion stability, navigation efficiency, and collision avoidance safety of multiple unmanned vehicles in complex and restricted waters.