Robust sliding-mode control of an underwater ROV via neural differential identification of model uncertainties

This work proposes an adaptive controller for a low-cost remotely operated vehicle (ROV), the BlueROV2, using a hybrid approach based on a first-principles model and a neural differential equation. This architecture allows for the simulation of longer time horizons, leveraging the adaptive nature of neural differential equations. As a result, the controller demonstrates improved performance compared to controllers based solely on first-principles models, offering greater flexibility than traditional adaptive control alternatives. The mixed model incorporates the estimation of hydrodynamic interactions between the vehicle, the water, and the ocean currents, modeled as irrotational flow with constant velocity. A case study is presented in which a BlueROV2 follows a reference trajectory under different underwater currents and model uncertainties. The study uses a multivariable super-twisting sliding-mode controller, which reduces tracking error by an average of 8.5 % when no model uncertainty is present and up to 57 % in high-uncertainty scenarios compared with a purely first-principles-based controller.The proposed hybrid DNN approach also achieves 86 % improvement in linear accuracy and 71 % in angular accuracy compared to the best-performing radial basis function (RBF) network. Furthermore, it outperforms classical nonlinear controllers in high-noise conditions, delivering the lowest linear tracking error with competitive control effort. © 2025 Published by Elsevier Ltd.

Robust sliding-mode control of an underwater ROV via neural differential identification of model uncertainties Academic Article in Scopus