Learning car dynamics from repeated tasks

I reproduced the numerical results from "Cautious Model Predictive Control using Gaussian Process Regression " We start with an initial physics-based kinematic bicycle model with Pacejka tire model for friction. Due to modeling limitations/error, we then learn the predictive error of the model using Gaussian process regression. This regressor not only produces an expected value, but also uncertainty quantification in the form of a variance estimate. The model predictive control uses chance-constraints to act on this uncertainty quantification. The resulting controller written in FORCES PRO solves on the order of milliseconds but decreases in speed as the size of the training dataset increases, a known shortcoming of this regressor. These results motivated my work in experiment design for collecting the most useful data.