NAViLoss: An Underwater Navigation-Aware Dual-Residual Objective for Physics-Consistent Learning

This research proposes NAViLoss, a robust and uncertainty-aware loss function for learning-based AUV velocity estimation. It jointly penalizes residuals in navigation-state and DVL measurement domains, and is integrated with a DeepONet architecture. Evaluated on 10,000m of semi-synthetic data, the model shows a 44% improvement in velocity-estimation accuracy compared to conventional and learning-based baselines.