Learned corrections keep a small underwater robot's controller on track in six degrees of freedom

Small holonomous underwater robots can inspect structures from any angle, but accurate dynamics models are hard to build underwater, according to a paper on arXiv. David Dörner and colleagues present a nonlinear model predictive contouring controller in all six degrees of freedom for such a vehicle.

The controller's nominal model is corrected with residual dynamics learned through Gaussian process regression, embedded in the optimization through a zero-order scheme that keeps it running at 25 Hz in real time.

A simulation ablation study shows the learned corrections improve tracking of position and attitude under state and input constraints.