Dual-mode deep reinforcement learning for safety-oriented MASS collision avoidance

Area of Science:

• Maritime Safety
• Artificial Intelligence
• Autonomous Systems

Background:

• Collision avoidance for Maritime Autonomous Surface Ships (MASS) is complex due to the need for predictable yet adaptive behavior in uncertain traffic.
• Existing autonomous systems struggle to balance rule-guided operations with dynamic risk assessment.

Purpose of the Study:

• To introduce a dual-mode, safety-oriented deep reinforcement learning (DRL) framework for MASS collision avoidance.
• To integrate model-based predictability with data-driven adaptability for enhanced maritime navigation safety.

Main Methods:

• A dual-mode DRL framework combining routine-navigation (Proximal Policy Optimisation - PPO) and heightened-safety modes.
• Quantification of encounter scenarios using International Regulations for Preventing Collisions at Sea (COLREGs).
• A multi-objective reward function integrating rule-guided reasoning and dynamic risk awareness, with a tree-based safety filter using Reachable Envelope Pruning (REP) and Terminal Safety Criterion (TSC).

Main Results:

• The safety layer demonstrated no collisions in two-ship encounters.
• Collision rates were reduced by 80-90% in congested scenarios compared to the baseline PPO.
• Maintained an average minimum passing distance exceeding 0.5 nautical miles even in six-ship traffic.
• Achieved over 90% reduction in node expansions, improving computational efficiency.

Conclusions:

• The proposed DRL framework effectively enhances MASS collision avoidance by balancing predictability and adaptability.
• The heightened-safety mode significantly improves safety margins and computational efficiency.
• The framework shows decision-level computational feasibility for real-world maritime applications.