A Cluster-Aware Collision-Avoidance Framework for Unmanned Surface Vehicles

With growing global attention to marine resource exploitation and maritime security, unmanned surface vehicles (USVs) have come to play an increasingly important role in tasks such as ocean exploration, maritime patrol, and water-quality monitoring. In congested waters where multiple vessels interact simultaneously, the safe avoidance of clustered ships has emerged as a critical prerequisite for autonomous navigation, since the corresponding performance directly affects both navigational safety and mission efficiency. However, most existing studies are tailored to isolated targets and fall short of identifying latent fleets with coordinated kinematic patterns; meanwhile, conventional risk-domain models typically rely on symmetric geometries that lack directional adaptability, and the classical Dynamic Window Approach (DWA) suffers from weight sensitivity and a tendency to converge to local optima, rendering it inadequate for integrated decision-making in cluster scenarios. To address these issues, this paper proposes a cluster-aware USV collision-avoidance framework that integrates optimized spectral clustering, an asymmetric biased risk domain, and an improved DWA. At the perception layer, the original signed-mean neighbour representation is augmented with magnitude and dispersion descriptors to form Robust10, followed by data-dependent RBF scaling, normalized spectral clustering, and observation-based hazardous-cluster selection. At the risk-modeling layer, principal axis decomposition, four-direction asymmetric radii, and a direction-aligned centre offset define an explicit piecewise cluster domain. At the decision layer, the original three-score DWA and four-stage bypass state machine are retained, while latency-compensated vessel/domain prediction and hard 25 m separation checks are added. Spectral clustering, PCA, DWA, and finite-state control are established techniques; the contribution lies in the task-specific representation, risk-domain construction, safety modifications, and end-to-end coupling. In randomized experiments, Robust10 raises F1 from 50.73% to 73.99% over 360 perception scenes, while the original 150-realization structural planning ablation raises safe-mission success from 9.33% for PredictiveDWA to 54.0% for the complete default-weight framework. To address weight subjectivity separately from that structural ablation, a safety-first simulation-data calibration selected [0.75,0.15,0.10]; when frozen and evaluated on 60 previously unseen planning encounters, the calibrated vector raises safe-mission success from 41.67% for the engineering-default vector to 75.00%, with no observed 25 m separation violation. A matched non-DWA finite-control-set nonlinear model-predictive-control (FCS-NMPC) baseline achieves 81.67% safe-mission success on the same held-out encounters; the paired difference from the calibrated framework is not statistically significant (p=0.4545). In 60 end-to-end trials, Robust10 raises safe-mission success from 28.33% to 45.00% relative to Mean5. These results support competitive simulation-level performance under the tested conditions without implying global parameter optimality, universal superiority over MPC, or deployment-level reliability.