An imitation learning framework for neural network-based path tracking control of multi-actuated electric vehicles

This paper presents a path tracking (PT) control framework for electric vehicles equipped with multiple chassis actuators. The methodology combines nonlinear model predictive controllers (NMPCs) acting as experts, with the derivation and application of imitation learning (IL) neural policies. Three actuation configurations are considered: (i) front-wheel-steering (2WS); (ii) four-wheel steering (4WS); and (iii) front-wheel steering and torque-vectoring (TV). The NMPC formulations use an accurate vehicle dynamics model to achieve precise trajectory tracking under aggressive driving conditions. Their high computational load, however, limits deployment in real-time embedded systems. A dedicated IL toolchain trains neural networks (NNs) to reproduce NMPC behaviour with drastically lower turnaround time, also in the presence of a reduced set of input variables with significant latencies. The resulting controllers are evaluated through simulations and experiments on a prototype vehicle platform. In simulation, the NNs match NMPC tracking performance. In vehicle experiments in limit handling conditions, the maximum lateral error is reduced by an average of 58% across the three configurations (2WS, 4WS and TV) relative to the corresponding benchmarking controllers (BCs). Robustness is assessed through: (a) a Monte Carlo analysis; (b) manoeuvres with swiftly variable tyre-road friction conditions; and (c) trajectories substantially different from those in the training set, all confirming the generalisation capability of the NN-based approach.