| Most deep learning models for vessel maneuvering are trained under calm water conditions but fail to generalize when exposed to open-sea disturbances such as wind. A common solution is to retrain or fine-tune models after gathering new data. However, modifying a validated model can reduce confidence in its previously verified behavior. This work proposes a trustworthy adaptation methodology in which the predictions of a calm water model are used as stable anchors, while a secondary model learns residual corrections between the anchor predictions and ground-truth values under new environmental conditions. By preserving the anchor model unchanged, the proposed method maintains the reliability of the validated model while enabling adaptation to previously unseen environments. We evaluate this approach using various architectures across simulated maneuvers under varying wind speeds. Experimental results demonstrate the effectiveness of the proposed method in predicting vessel dynamics and trajectory compared to both unadapted anchor models and retrained baselines. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.