Abstract: Objectives To address the issues of low path planning efficiency and poor robustness caused by the limited sensing range in maritime rescue missions., a local observation-based path planning method for unmanned surface vehicle is proposed. MethodsThe Soft Actor-Critic algorithm is employed, with a reward function based on local observation designed and combined with a feature-enhanced training method. By extracting key environmental features and training in a randomized feature environment, the path planning sampling efficiency and robustness under limited perception conditions are improved. An adaptive waypoint planning algorithm based on local perception domains is developed, coordinating local and global planning for efficient path planning. Results Simulation results show the proposed method achieves a success rate of over 98% in feature environments and a completion rate of over 93% in rescue missions, demonstrating strong robustness and adaptability. Conclusions The proposed local observation path planning method for unmanned surface vehicle addresses challenges in maritime rescue missions and provides technical support for applying reinforcement learning in real-world scenarios.