Zero-shot ship berthing trajectory prediction: A Target-Driven Residual Bi-LSTM with ego-centric spatial awareness based on real berthing data and port geometry

Accurate prediction of vessel trajectories during berthing is important for maritime safety. Models trained on absolute coordinates tend to memorize their training ports’ geometry. We propose a Target-Driven Residual Bi-LSTM that decouples control logic from absolute geography. Ego-centric Signed Distance Fields (SDF), relative goal vectors, and scale-invariant kinematics drive non-linear residual corrections over a deterministic kinematic baseline, with a Dynamic L1 sparsity penalty, yaw-rate filtering, and a velocity-scaled deadband regularizing open-water transit. In zero-shot spatial Leave-One-Out Cross-Validation (LOOCV) across six ports over a 100-step horizon, the model outperformed kinematic baselines by up to 12.2% in Average Displacement Error (ADE) within constrained ports, with +6.4% longitudinal braking accuracy (Along-Track Error). It exceeded Transformer and Quaternion Ship Domain (QSD) LSTM baselines by factors of 2.8 and 3.2 in ADE, and its advantage over kinematic extrapolation grew monotonically with horizon length, reaching +19.2% at 300 steps. On outbound maneuvers the kinematic baseline remained superior, delimiting the framework to arrivals. A seven-configuration sensitivity analysis (42 LOOCV models, 4.7 × parameter range) found macro ADE varying within 0.52 m; a 100-step trajectory is forecast in 2.1 ms on a laptop-class GPU (45 MB memory), supporting Vessel Traffic Service and Maritime Autonomous Surface Ship (MASS) applications.