Research on Dynamic Obstacle Avoidance Planning for Unmanned Surface Vessels Based on the Introduction of Attention Mechanism

To address the problem of local obstacle avoidance and path planning for Unmanned Surface Vehicles in complex dynamic marine environments, this paper proposes an end-to-end control method based on deep reinforcement learning. The method employs the Proximal Policy Optimization algorithm and designs a feature extraction network that integrates one-dimensional depthwise separable convolution with a multi-head attention mechanism. Given the characteristics of sparse LiDAR data, the Squeeze-and-Excitation (SE) module is first utilized to enhance the representational capacity of channel features. Subsequently, a multi-head attention module based on the Transformer architecture is constructed, which uses the vehicle's own motion states as query vectors to dynamically weight environmental observation features, thereby achieving effective fusion of proprioception and environmental perception. Finally, complex scenarios comprising static reefs and dynamic interfering obstacles are constructed within the PyBullet physics simulation environment. Experimental results demonstrate that the proposed method effectively handles dynamic obstacles and achieves smooth autonomous obstacle avoidance control . Compared with standard PPO, the proposed method improves the obstacle-avoidance planning success rate by 4%.