Unmanned Underwater Vehicle Autonomy and Control near Submarines Using Actively Sampled Surrogates

Abstract

Many tools have been developed to simulate unmanned underwater vehicle (UUV) motion and autonomous behaviors to evaluate UUV capabilities. However, there is no simulator that performs real-time modeling of the complex hydrodynamic interaction forces that a UUV experiences when operating near a moving submarine. These hydrodynamic interactions must be determined in real time to simulate the launch and recovery of UUVs from submarines. Potential flow models may be fast enough to solve the hydrodynamic interactions in real time, but by oversimplifying the physics and neglecting viscosity, they introduce inaccuracies into the simulations. Computational fluid dynamics (CFD) is capable of accurately modeling these hydrodynamic interactions, but simulations take hours or days to solve. To overcome this obstacle, a machine learning method known as Gaussian process (GP) regression is used to create a surrogate reduced-order-model that predicts the hydrodynamic interactions in real time. The GP regression model is trained by actively sampling CFD simulations in order to accurately model complex hydrodynamic interactions. This new approach allows the GP regression model to be incorporated into a UUV motion simulator and evaluate how the UUV is affected by the hydrodynamic interactions. Operating envelopes are developed that outline regions where the UUV safely overcomes the hydrodynamic interactions and where the UUV is overpowered and collides with the submarine. By incorporating this surrogate model into the autonomy architecture, new autonomous behaviors are created that compensate for the hydrodynamic interactions by adjusting the desired UUV heading and speed which allows it to better stay on course.