Objective A semantic-guided risk-sensitive autonomous navigation method based on a large language model (LLM) is proposed to accurately characterize collision-avoidance decision risk for multiple unmanned surface vehicles (USVs) under the coupled effects of vessel encounters, static obstacles, and local vortex-flow disturbances in complex confined waters.
Method An IQN-based decentralized navigation policy is developed to enable each USV to select actions independently according to its local observations, including the relative target direction, obstacle information, and motion states of neighboring vehicles. 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 quantile-sampling range 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 LLM-CVaR-IQN achieves task success rates of no lower than 0.92 in high-risk scenarios involving 3–6 USVs, outperforming Rule–CVaR-IQN, Geo-CVaR-IQN, IQN, DQN, RVO, and APF. In the six-USV scenario, its success rate reaches 0.92, exceeding those of Rule–CVaR-IQN, Geo-CVaR-IQN, and IQN by 4, 6, and 12 percentage points, respectively.The proposed method also achieves shorter navigation times, lower cumulative action costs, and reduced collision risk while maintaining robust navigation under non-ideal disturbances.
Conclusion LLM-derived semantic knowledge can provide high-level encounter information and risk priors that complement local numerical observations. By combining IQN-based return distribution modeling with adaptive CVaR risk evaluation, the proposed method improves task completion reliability, navigation efficiency, and collision avoidance safety for multiple USVs in complex and restricted waters.
DownLoad: