Towards Black-Box Dynamics Modelling within Learning-based Nonlinear Model Predictive Control for Virtual Motorcycles

Abstract

In recent years, both academic and industrial effort has been devoted to developing Virtual Riding (VR) strategies, in line with the attempts towards virtual prototyping in the automotive business. The advantages of virtual prototyping include reduced costs, time-to-market, and increased safety. Nonlinear Model Predictive Control (NMPC) schemes have been utilized to achieve highperformance real-time control of the motorcycle system, while emulating a realistic riding behavior, as described, e.g., in Massaro (2011); Bruschetta et al. (2021). To apply NMPC techniques, a detailed model of the plant is necessary, and different dynamics models have been proposed in literature. However, the accuracy of physics-based motorcycle models that grant real-time feasibility of the NMPC remained questionable when compared to a realistic system behavior. Besides, they require specific measurements and identification procedures to be applied. In this sense, Learning-based NMPC (LbNMPC) framework, described in Hewing et al. (2020), could be a promising new direction. LbMPC is a broad research field that generally refers to the implementation of learning techniques into NMPC control schemes. Specifically, the learning dynamics sub-field covers the data-based adaptation of the NMPC prediction model. In this context, Gaussian Process Regression (GPR) resulted very effective in applicative scenarios, as shown in Kabzan et al. (2019);Carron et al. (2019);Picotti et al. (2023).