Multi-Task Learning for Ship Trajectory Prediction and Motion Planning via Node Relationship Modeling

This study proposes an integrated framework for multi-ship trajectory prediction and motion planning based on joint learning, which significantly enhances the overall performance in multi-ship interaction scenarios by combining the global information from trajectory prediction with the local optimization capabilities of motion planning. In the trajectory prediction task, the proposed MTSGCN model achieves an average performance improvement of 26.3% compared to the Rain model, while the joint multi-task learning strategy yields a 26.9% performance gain over single-task learning, demonstrating the effectiveness of the multi-ship interaction adjacency matrix extraction module. In the motion planning task, the MTSGCN model outperforms the TFT model with an average performance improvement of 13.2%, whereas the MTSGCN-T model without joint learning experiences an 11.2% performance degradation. Furthermore, the study reveals that ship speed and heading decisions are influenced by multiple factors, with the inertial effect of historical parameters being the most significant. Additionally, the feature distributions of different datasets have a substantial impact on model performance.