Autonomous racing has emerged as a challenging platform for evaluating state-of-the-art control algorithms used in self-driving vehicles. In this context, Model Predictive Control (MPC) is commonly utilized to meet the complex control requirements inherent to the task. However, accurate real-time modeling of nonlinear vehicle dynamics poses a significant challenge and frequently results in model discrepancies. These modeling errors can not only adversely affect control performance, but also compromise the safety of the control system.

Recognizing this limitation, recent works have demonstrated how data-driven dynamics models such as neural networks and Gaussian processes can be utilized to reduce modeling error. Concurrently, robust control techniques such as tube-MPC and stochastic MPC have been shown to mitigate the impact of modeling errors effectively, by taking the imperfect nature of the dynamics model in to consideration during the design phase.

This thesis enriches existing research by analyzing the robustness properties of an uncertainty-aware MPC. The adopted approach combines a first-principles based dynamics model with a sparse Gaussian process (GP), which is trained online to correct errors present in the nominal dynamics model. The model uncertainty is taken in to account by tightening the MPC constraints based on the variance predictions of the GP, which are propagated through the prediction horizon.

While previous research has investigated this combined approach, it has primarily emphasized lap time improvements, rather than offering a detailed analysis of its robustness properties. This thesis aims, through simulation experiments, to investigate whether this integrated approach manifests increased robustness, notably in slippery road conditions and when the tire properties change due to degradation. Unlike lap times, which hold limited relevance for commercial applications, robustness serves as a more pertinent metric for improving commercial autonomous vehicles. Consequently, the findings may offer valuable insights into the broader applicability of these methods beyond autonomous racing.