Learning-based Model Predictive Controller for Lateral Control of Autonomous Vehicles

Lateral control of an Autonomous Vehicle (AV) presents significant challenges, including trajectory tracking while respecting actuator and dynamic limits. The presence of nonlinear dynamics, coupled with challenging-to-model and uncertain parameters (such as friction forces and mass variations), further complicate the task. Model Predictive Control (MPC) emerges as a popular choice in autonomous vehicle applications due to its ability to handle multiple variables and constraints. However, the effectiveness of the controller relies heavily on the accuracy of the vehicle model. To address this, a precise dynamic vehicle model becomes crucial. A data-driven approach, leveraging vehicle operation data, offers a promising solution. By learning vehicle dynamics, this solution strikes a balance between accurate state predictions and computational cost, enhancing the effectiveness of MPC. This work proposes a framework (NNMPC) for designing an MPC with a Neural Network (NN)-based learned dynamic model of the vehicle, utilizing data from simulations. By training a neural network model with historical states and controls, NNMPC predicts vehicle dynamics even in changing and complex operating conditions. In particular, this thesis work focuses on reference tracking without prior knowledge of the road-tire friction coefficient and of the number of passengers, hence of the vehicle mass. The performance of the NNMPC is then compared with other classical control strategies.