Data-driven social-cognitive navigation for energy-efficient and low-carbon autonomous shipping

Abstract

The digital transformation of maritime transportation creates new opportunities to improve the safety, efficiency, and sustainability of autonomous shipping. However, existing navigation methods often emphasize collision avoidance and trajectory feasibility while paying less attention to unnecessary maneuvering, traffic disturbance, and associated operational demand. This study proposes a data-driven Social-Cognitive Navigation Framework (SCNF). Historical Automatic Identification System (AIS) trajectories are used to extract representative encounter patterns, while online AIS and onboard-sensor data are treated as partial and asynchronous observations. SCNF integrates a Recursive Hypergraph Transformer, recursive Level-k reasoning, and a legibility-aware game-theoretic planner. Performance is evaluated using safety, efficiency, interaction, and disturbance metrics across 500 Monte Carlo trials. SCNF outperforms representative reactive, optimization-based, and learning-based baselines. In the complex crossing scenario, it achieves a 0% collision rate, 15.2 m minimum distance of approach (MDA), 1.08 normalized path length, 4.1° Average Avoidance Magnitude by Others (AAMO), and 4.6/5 Trajectory Legibility Score (TLS). In the heterogeneous overtaking scenario, it maintains a 0% collision rate and 18.5 m MDA. Ablation results confirm complementary contributions from the three core modules. SCNF improves navigation-level safety, efficiency, and interaction coordination while reducing unnecessary maneuvering-related operational demand. Fuel consumption and carbon emissions were not directly measured; therefore, the sustainability benefits should be interpreted as navigation-level evidence rather than quantified real-ship emission reductions.