COLREGs-compliant unmanned sailboat navigation via safe reinforcement learning

Autonomous collision avoidance for unmanned sailboats is challenging because stochastic wind-driven maneuvering constraints must be reconciled with the strict requirements of the International Regulations for Preventing Collisions at Sea (COLREGs). Existing deep reinforcement learning (DRL) approaches often struggle in this setting due to non-stationary reward characteristics across mixed encounter scenarios, limited transparency in collision-risk inference, and insufficient guarantees of rule compliance. This paper proposes a unified decision-making framework for COLREGs-compliant collision avoidance of unmanned sailboats, in which the wind-dependent decision logic prescribed by COLREGs Rule 12 is first embedded into the DRL environment. The framework integrates three key components: a hierarchical risk-aware perception module that incorporates collision-risk indicators into the state representation, a scenario-balanced SAC-PopArt learning strategy to stabilize training across heterogeneous encounters, and a dual-layer safety mechanism combining reward shaping with model-predictive shielding. Simulation results show that the proposed approach maintains COLREGs-compliant collision avoidance in canonical scenarios and exhibits zero-shot generalization in complex, unseen multi-vessel environments.