Model Predictive Control for Autonomous Driving via Machine Learning Approximation

Autonomous Driving is, nowadays, a topic widely studied by the most renowned automotive companies and universities. This interest is motivated by several advantages, which include fuel consumption optimization, pollution reduction, accidents decrease and comfort enhancement. In this work, the self-driving topic is applied to a vehicle moving in a highway environment. In particular, this thesis focuses on: 1. the design of a single layer predictive control architecture, based on Artificial Potential Fields techniques, that provides the optimal control inputs in terms of steering angle and acceleration; 2. the use of Machine Learning (ML) techniques to obtain a Neural Network (NN) approximation of the control law designed at the previous point. As to the first point, a Nonlinear Model Predictive Controller (NMPC) is designed to compute the needed control action based on the driving scenario data provided by the on-board camera and sensors. However, the NMPC online implementation requires a significant computational effort that the vehicle control unit can not handle. To overcome this problem, a NN approximation of the NMPC control function is computed, by exploiting a data-based ML approach. Such a NN is then employed to implement the NMPC controller with a lower computation burden. Extensive simulation tests are performed on a nonlinear 3-DOF vehicle model to show the effectiveness of the proposed approach in several highway driving scenarios.