On the simulation and experimental analysis of imitation learning for preview-based torque vectoring control

Nonlinear model predictive control (NMPC) is a promising technology for chassis control applications, including torque vectoring control (TVC) of electric vehicles with multiple powertrains. Although NMPC can incorporate preview-based information, expected to augment performance of future active safety systems, its practical applications remain limited due to complexity and computational cost. In parallel, a few artificial intelligence (AI) TVC methods have been recently explored, but without considering imitation learning (IL). To cover the gap, this study proposes deep neural network (DNN)-based TVC, where the DNN is trained through IL of an NMPC algorithm including a 7-degree-of-freedom prediction model, and the preview of vehicle trajectory and tyre-road friction level. The application is an in-wheel motor-driven vehicle prototype operating in varying friction conditions. The simulation results highlight: (i) performance comparable to the NMPC, with turnaround time reductions exceeding 150 times; and (ii) TVC robustness to parameter uncertainties, evaluated through Monte Carlo analyses. Moreover, proof-of-concept experimental vehicle tests show that in absence of sideslip angle feedback, which significantly simplifies the estimation requirements, the proposed DNN reduces the yaw rate tracking error by >75%, compared with a real-time implementable benchmarking TVC system based on a yaw moment observer and a rule-based longitudinal tyre slip controller.