A Hierarchical Logic-Guided Reinforcement Learning Framework for Interpretable Autonomous Ship Collision Avoidance

Autonomous ship collision avoidance requires rule-consistent, interpretable, and effective decisions across different encounter situations. This paper proposes the Logic-Guided Interpretable Deep Reinforcement Learning framework (LG-IDRL) that connects Drools-based COLREGs reasoning with continuous proximal policy optimization (PPO). The rule engine identifies encounter responsibilities and urgency, supplies dynamically activated reward terms, and defines an admissible set for projecting the proposed two-dimensional rudder–acceleration action into the intersection of rule and physical limits when this set is nonempty. In five-seed held-out evaluations, LG-IDRL achieved a task success rate of 92.9±2.1%, a collision rate of 1.6±0.4%, an automatically and independently assessed behaviour-level rule-compliance rate of 98.7±0.4%, and a mean path length of 16.7±0.3 nm. An adapted recent GRU–PPO–COLREG baseline achieved 89.1±1.7%, 3.9±0.9%, 95.8±1.2%, and 17.4±0.3 nm, respectively, under the same evaluation protocol. The results characterize performance within six specified two-ship-encounter families; they do not establish guaranteed collision avoidance or generalization to unrestricted maritime traffic.