Safety-Enhanced COLREGs-Compliant Path Planning for USVs with a CBF-Based Safety Shield

This study proposes a safety-enhanced path planning system that integrates a Control Barrier Function (CBF)-based Safety Shield with Deep Reinforcement Learning (DRL). This framework addresses the critical limitations of conventional DRL-based Unmanned Surface Vehicle (USV) navigation models, which can output hazardous control commands in edge cases and violate the International Regulations for Preventing Collisions at Sea (COLREGs). The proposed system continuously operates during navigation via an Encounter Classifier that identifies multi-vessel situations (such as Head-on, Crossing, and Overtaking) in real time. The nominal control inputs generated by the DRL policy are verified and safely filtered through a Control Barrier Function-Quadratic Programming (CBF-QP) optimization layer immediately prior to execution, incorporating ship safety radii and asymmetric COLREGs constraints. Furthermore, we introduce a ‘Shielded Training’ mechanism that penalizes the agent based on the magnitude of the shield’s interventions during the training loop. This effectively diminishes the policy’s over-reliance on the safety filter and guides the network toward discovering robust, inherently safe trajectories. Extensive simulations conducted under diverse single- and multi-vessel encounter scenarios quantitatively demonstrate that the proposed method substantially reduces collision and COLREGs violation rates compared to baseline DRL-only or reward-shaping methods, while maintaining excellent computational scalability and real-time responsiveness. Consequently, by unifying the adaptive environmental exploration of reinforcement learning with model-based runtime safety constraints derived from control theory, this study provides a practical runtime assurance framework for future marine deployment.