Imitation learning for path tracking control in a multi-actuated vehicle

Autonomous driving has the potential to increase road safety and efficiency, but it requires robust control strategies to ensure accurate path tracking, particularly under demanding manoeuvres. This thesis focuses on the development of a path-tracking controller for a multi-actuated vehicle equipped with independent wheel motors and steer-by-wire on both axles. A high-fidelity vehicle model, including tire and actuator dynamics, was identified and validated. As a benchmark, a Proportional-Derivative controller with feedforward contributions was designed and experimentally tested. Nonlinear Model Predictive Control (NMPC) was employed in simulation to leverage its capability in handling nonlinear dynamics and constraints, thereby achieving optimal performance. Due to its high computational cost and sensitivity to sensor/actuator delays, NMPC is unsuitable for direct real-time implementation on the experimental platform. To overcome these limitations, imitation learning was employed: a neural network was trained on NMPC-generated datasets to replicate its control policies with negligible computational effort. Then, the Neural Network has been tested in the field, showing consistent improvements with respect to classical feedforward plus feedback controllers.