Highway Autopilot using Robust Model Predictive Control with a Transformer based Observer

Abstract

Abstract In this thesis, the area of robust trajectory planning through model predictive control in combination with trajectory predictions is explored. The combination is used to obtain an autopilot algorithm controlling a truck that is operating in a simulation of a highway scenario with surrounding traffic. A novel transformer-based observer has been developed to address the problem of generating trajectory predictions for the robust model predictive controller. The observer is also capable of expressing uncertainty through the Monte Carlo dropout methodology. Further, the model predictive controller extends its collision avoidance constraints based on these uncertainties to increase the robustness of the solution. The results of the developed observer showed that it was capable of giving more accurate multi-modal predictions compared to a constant velocity baseline, in terms of ADE and FDE. Indications for a more robust solution in terms of safety were observed for the robust model predictive controller combined with the developed trajectory predictor. However, the robustness of the new combination came at the cost of reduced drivability performance caused by overcompensation for future trajectory predictions of surrounding vehicles. It was concluded that the results signify the usefulness of the developed combination in safety-critical scenarios and that a foundation for future work has been provided.