Risk-Aware SAC Algorithm for Path Planning of Unmanned Surface Vehicles in the Presence of High-Speed Dynamic Obstacles

This paper proposes the Risk-Aware SAC (RA-SAC) algorithm for unmanned surface vehicle (USV) path planning in the presence of high-speed dynamic obstacle interference. This method integrates risk perception, obstacle motion state extrapolation and heading constraint reward into the SAC framework. While preserving the entropy-maximization mechanism of SAC, it enhances long-horizon obstacle avoidance awareness through reward shaping, and enables active predictive obstacle avoidance. Simulation results show that RA-SAC converges around 500 episodes with a final reward of approximately 4800, significantly outperforming traditional SAC and TD3. Ablation studies show that removing the heading constraint module reduces the task success rate by over 90 percentage points. Monte Carlo tests with 30 independent runs yielded a coefficient of variation of 0.135, demonstrating strong robustness. The proposed method effectively transforms passive responses into active predictive planning, thereby enhancing navigation safety and path efficiency in high-speed dynamic scenarios.