Predictive and Inertia-Aware Motion Planning for USV Navigation in Cluttered Harbors

Autonomous navigation for Unmanned Surface Vehicles (USVs) in cluttered harbors presents a dual challenge: the rigorous constraints of hydrodynamic inertia and the complexity of perceiving diverse surface and underwater hazards amidst wave clutter. While recent neuro-symbolic planners excel on ground robots, they often fail in maritime settings due to kinematic mismatches and sensitivity to environmental noise. This paper presents a Predictive and Inertia-Aware Neuro-Symbolic framework designed to bridge this gap. First, to address the heterogeneity of maritime perception, we replace raw sensor inputs with a high-level representation based on Potential Collision Areas (PCAs). By aggregating object detections from multiple modalities (e.g., LiDAR, Sonar) and predicting their future trajectories, we construct dynamic 2D bounding boxes in the navigation plane that encapsulate future collision risks. This object-level abstraction unifies diverse sensor data and effectively filters out transient wave clutter that confuses standard planners. Second, we propose a physics-informed domain adaptation strategy where the neural policy is trained via constrained imitation learning. By subjecting the expert demonstrator to strict acceleration limits, the network implicitly internalizes hydrodynamic inertia, learning to initiate avoidance maneuvers well in advance. Validation in high-fidelity simulations demonstrates that our method successfully compensates for drift, handles multi-modal obstacle data, and proactively avoids dynamic threats, achieving safe navigation where standard kinematic baselines fail. The source code is available at https://gitlab.cylab.be/t.nguyen/usv-motion-planning.