Physics-guided neural networks for motion estimation of an underwater vehicle with a complex propulsion system

This study investigates planar motion estimation for an underwater robot with bioinspired propulsion, where the tight coupling between hydrodynamics and actuation makes accurate dynamics modeling challenging. We implement and compare three frameworks: Physics Model Error Prediction (PMEP), a hybrid approach where a neural network learns the errors of a simplified physics model; Sim2Real fine-tuning, where a network is pretrained in a simulator and adapted with real-world data; and a purely data-driven Long-Short-Term Memory (LSTM) baseline. We also propose an inexpensive pipeline for generating underwater reference labels using ArUco markers and an onboard camera. To evaluate these models, we compare their performance on a motion prediction task using real-world experimental datasets of varying sizes. The results show that physics-guided learning can improve motion-estimation performance in data-limited regimes, but the benefit is not uniform across motion components or training-data ratios. PMEP is most effective when the analytical model provides a reasonable baseline for residual correction, especially for surge-related prediction, whereas Sim2Real fine-tuning provides more consistent trajectory-level performance when limited real-world data are available. These findings support the use of physics-guided learning as a forward motion-estimation component for underwater navigation, while also showing that open-loop dead reckoning should be integrated with feedback-based localization or sensor-fusion methods for long-duration deployment.