Deep physics for deep seas: Estimating underwater vehicle dynamics with PINNs

The RUVIFIST (Reconfigurable Underwater Vehicle for Inspection, Free-floating Intervention, and Survey Tasks) platform was developed to combine the long-range endurance of AUVs with the precision and manoeuvrability of ROVs. Its reconfigurable design enables both a streamlined “survey” mode and a compact “hovering” mode for inspection and intervention. The main contribution of this paper is the development of a Physics-Informed Neural Network (PINN) to estimate key hydrodynamic parameters, including added mass and linear damping coefficients, directly from field data, improving the accuracy of the vehicle’s dynamic model and enabling a feedforward controller alongside PID control. Experiments focused on estimating surge, sway, and yaw parameters. Performance was assessed by comparing the feedforward control effort with that of a baseline PID controller. In the “survey” configuration, the method achieved high accuracy, with force discrepancies below 5 N during constant-speed surge and below 2 N during heading rotations. In the “hovering” configuration, the estimations remained reliable, with the feedforward controller alone achieving velocity errors of about 0.03 m/s during constant-speed surge and sway motions.