Bimodal Observation Based USV Global Path Tracking and Local Collision Avoidance

This study proposes a robust end-to-end navigation framework for unmanned surface vehicles (USVs) that integrates global path tracking and local obstacle avoidance using bimodal input data. Global reference paths are generated using a Dijkstra-based algorithm, leveraging static map information, while dynamic obstacles are identified via simulated LiDAR data. The state space is designed to capture environmental and motion attributes, and a distributional reinforcement learning algorithm is employed for robust decision-making. A regression mechanism in the reward function encourages safe obstacle avoidance while maintaining path adherence. The algorithm is validated in simulated environments with increasing complexity and a real multi-island background, demonstrating effective collision avoidance and path tracking. qualitative and quantitative experiments show the algorithm's capability to adapt to dynamic conditions while maintaining real-time efficiency.