Nonparametric dynamics modeling for underwater vehicles using local adaptive moment estimation Gaussian processes learning

This paper investigates a nonparametric modeling scheme for underwater vehicles to achieve continuous-time dynamics modeling, which is essential for various marine missions, control, and navigation of these vehicles. The proposed scheme addresses the challenges posed by the nonlinearity, strong coupling, and complex structure of underwater vehicles through local adaptive moment estimation Gaussian processes learning. This approach constructs mappings between hydrodynamics and motion states while providing uncertainty estimates of the dynamics model. A local weighted strategy is used to construct local models to localize Gaussian processes learning, and an adaptive moment estimation method is designed using gradients of innovation to tune hyperparameters of Gaussian processes automatically. Moreover, a subspace index is created and updated based on feature distance measures to improve the computational efficiency of Gaussian processes learning in each local model. The developed scheme can perform real-time simulation considering environmental disturbances and is applied to a 6 degree-of-freedom autonomous underwater vehicle. The results demonstrate that this scheme is an effective mathematical modeling tool for underwater vehicles dynamics.