DVL-DeepONet: A Physics-Guided Operator Learning for Resilient Underwater Navigation

Researchers have developed DVL-DeepONet, a physics-guided deep learning framework that improves underwater vehicle navigation by accurately estimating velocity from noisy or incomplete sensor data. The system outperforms traditional approaches by 40% in real-world testing, enabling autonomous underwater vehicles to operate reliably even with degraded sensor inputs or without expensive inertial measurement units.

DVL-DeepONet addresses a critical challenge in autonomous underwater vehicle (AUV) operations: maintaining accurate navigation when sensor data degrades. Traditional AUV systems depend on Doppler velocity logs (DVLs) and inertial measurement units working in tandem, but real-world underwater environments frequently introduce noise, incomplete beam measurements, and reflections that compromise data quality. This research bridges a significant gap by developing an operator learning framework that maintains navigation accuracy under these adverse conditions.

The innovation combines physics-guided constraints with deep neural networks, forcing the model to respect the underlying physics of DVL measurements while learning from temporal sensor patterns. This hybrid approach differs from purely data-driven methods that ignore physical principles. The framework demonstrates versatility across three operational scenarios: noise resilience with coupled sensors, DVL-only estimation for low-cost platforms, and active beam measurement recovery. Testing on approximately 10,000 meters of real AUV mission data validates the approach's practical effectiveness.

The 40% performance improvement over baseline and learning-based methods has meaningful implications for underwater robotics deployment. Cost reduction becomes possible when expensive inertial sensors become optional, expanding accessibility for research institutions and marine industries. Enhanced resilience in degraded sensing conditions increases operational reliability in challenging marine environments where equipment failures or environmental disturbances occur frequently.

This work represents incremental but meaningful progress in autonomous underwater systems. Future development likely includes integration with other sensor modalities, real-time deployment on resource-constrained underwater platforms, and adaptation to different AUV designs. The physics-informed learning approach may transfer to other robotic navigation domains facing similar sensor degradation challenges.

• →DVL-DeepONet achieves 40% better velocity estimation than baseline methods in degraded underwater sensing conditions
• →Physics-guided neural operators enable reliable AUV navigation even without expensive inertial measurement units
• →The framework successfully handles three scenarios: noise resilience, DVL-only learning, and beam measurement recovery
• →Real-world validation on 10,000 meters of AUV mission data demonstrates practical applicability in marine environments
• →Cost reduction potential could expand autonomous underwater vehicle deployment across research and commercial marine sectors