Optimizing vehicle handling through Koopman-based model predictive torque vectoring: An experimental investigation

Abstract

This paper presents a Koopman operator model predictive control (KMPC) torque vectoring where the vehicle model is represented by a finite-dimensional approximation of the Koopman operator obtained by using the extended dynamic mode decomposition. The Koopman operator acts like a linear predictor for a nonlinear dynamical system by lifting the nonlinear dynamics into a higher dimensional space where its evolution becomes linear. Different scenarios are simulated using the nonlinear vehicle model to generate the required data set and to obtain the Koopman model used for the KMPC. The KMPC was implemented on the dSPACE MicroLabBox platform, followed by its evaluation in two different experiments. These experiments are conducted using a scaled four-wheel-drive electric vehicle driving on a treadmill that serves as a surrogate for the roadway. The performance of KMPC is compared to that of linear time-varying model predictive controller (LTV-MPC), and nonlinear model predictive controller (NMPC). The results show not only the real-time applicability of KMPC but also a comparable performance and lower computational complexity compared to NMPC. Additionally, an interesting effect of discretization and communication delay on the performance of both LTV-MPC, and NMPC is observed, whereas KMPC demonstrates robustness in this scenario.