SAS-Transformer: A Sensor-Aware Switchable Framework for Robust Underwater Navigation via Implicit Spatiotemporal Fusion

Reliable navigation for Autonomous Underwater Vehicles (AUVs) faces significant challenges due to heterogeneous multi-sensor data, intermittent signal failures, and physical noise. To address these challenges, we propose SAS-Transformer, a novel switchable Transformer architecture that fuses physics-based principles with deep learning to achieve robust implicit spatiotemporal fusion for navigation. Its core innovations include: (1) Physics-informed embedding layers that encode kinematic constraints into deep features; (2) A switchable attention mechanism that dynamically senses sensor states and adaptively adjusts information flow, enhancing robustness to sensor failures; (3) Cross-modal gating modules for effective fusion of spatiotemporal features from heterogeneous sensors; and (4) Uncertainty-aware dual output heads that jointly optimize state estimation and its uncertainty. Extensive experiments on real-world AUV datasets demonstrate that SAS-Transformer significantly outperforms existing methods, reducing position estimation Root Mean Square Error (RMSE) by 38.7% compared to Factor Graph Optimization (FGO) and by 62.1% compared to Long Short-Term Memory (LSTM) baselines. Crucially, under harsh conditions like LBL outages lasting 60 s, its error elevation remains below 15%, exhibiting remarkable robustneess. Furthermore, the model achieves exceptional real-time performance with inference latency below 20 ms, making it highly suitable for online AUV navigation.