A multi-objective deep reinforcement learning framework for energy efficiency of autonomous harbor crafts

This study proposes a novel preference-conditioned multi-objective deep reinforcement learning (MODRL) framework for adaptive navigation of unmanned surface vehicles in dynamic harbor environments. By integrating a multi-head Q-network and objective preference vectors, the framework enables flexible trade-offs among energy efficiency, time minimization, and navigational safety, conditioned on task-specific priorities. Simulation tasks include rule-constrained collision avoidance (COLREGs), wave-disturbed in-port delivery, and wind-affected return-to-dock scenarios. Compared to conventional DQN variants and additional baselines such as MO-A2C and QR-DQN, the MODRL agent achieves superior adaptability, with up to 23 % higher success rates, 0.15 improvements in safety score, and 20 %–35 % reductions in simulated energy usage based on a cubic-speed surrogate model. Policy behaviors under different preferences remain interpretable via trajectory visualizations and reward decomposition, revealing context-aware adjustments in heading and speed. The results confirm the framework’s robustness, transparency, and practicality for deploying intelligent harbor craft in real-world smart port operations.