Automated Tuning of Nonlinear Kalman Filters for Optimal Trajectory Tracking Performance of AUVs

The performance of navigation algorithms significantly determines the trajectory tracking accuracy of the guidance, navigation, and control (GNC) system of an autonomous underwater vehicle (AUV). In closed-loop operation, the interaction among path planning, control, and navigation plays a crucial role in the tracking accuracy of the overall GNC system. A Doppler velocity log (DVL) is often used for AUVs to measure velocity over the ground, positively affecting the closed-loop tracking error. However, a DVL may not be installed in miniaturized AUVs due to limited space and energy. In this paper, a navigation filter for an underactuated miniature AUV (nanoAUV) is considered that is mainly based on acoustic localization using a novel highly-miniaturized ultra-short baseline (USBL) system and a depth pressure sensor. The nanoAUV is being developed for subglacial lake exploration. We compare two unscented Kalman filters (UKF) with different prediction models - the classical strapdown inertial navigation systems (SINS) model and a hydrodynamic motion model (HMM). To enable a fair comparison, filter parameters are auto-tuned with Bayesian optimization (BO) for open and closed-loop performance, which is novel in AUV navigation. The results indicate that BO performs similarly to particle swarm optimization (PSO) regarding sample efficiency for the proposed problem. To quantify the GNC tracking performance, we use extensive Monte Carlo simulations. Results suggest that with BO-tuned navigation filter parameters, the median tracking error is reduced by up to 50% compared to default parametrization.