Model predictive path integral control: Theoretical foundations and applications to autonomous driving
Abstract
This thesis presents a new approach for stochastic model predictive (optimal) control: model predictive path integral control, which is based on massive parallel sampling of control trajectories. We first show the theoretical foundations of model predictive path integral control, which are based on a combination of path integral control theory and an information theoretic interpretation of stochastic optimal control. We then apply the method to high speed autonomous driving on a 1/5 scale vehicle and analyze the performance and robustness of the method. Extensive experimental results are used to identify and solve key problems relating to robustness of the approach, which leads to a robust stochastic model predictive control algorithm capable of consistently pushing the limits of performance on the 1/5 scale vehicle.