Developing automatic driving solutions and driver support systems requires accurate vehicle specific models to describe and predict the associated motion dynamics of the vehicle. Despite of the mature understanding of ideal vehicle dynamics, which are inherently nonlinear, modern cars are equipped with a wide array of digital and mechatronic components that are difficult to model. Furthermore, due to manufacturing, each car has its personal motion characteristics which change over time. Hence, it is important to develop data-driven modelling methods that are capable to capture from data all relevant aspects of vehicle dynamics in a model that is directly utilisable for control. In this paper, we show how Linear Parameter-Varying (LPV) modelling and system identification can be applied to reliably capture personalised model of the steering system of an autonomous car based on measured data. Compared to other nonlinear identification techniques, the obtained LPV model is directly utilisable for powerful controller synthesis methods of the LPV framework.

Automated driving applications require accurate vehicle specific models to precisely predict and control the motion dynamics. However, modern vehicles have a wide array of digital and mechatronic components that are difficult to model, manufactures do not disclose all details required for modelling and even existing models of subcomponents require coefficient estimation to match the specific characteristics of each vehicle and their change over time. Hence, it is attractive to use data-driven modelling to capture the relevant vehicle dynamics and synthesise model-based control solutions. In this paper, we address identification of the steering system of an autonomous car based on measured data. We show that the underlying dynamics are highly nonlinear and challenging to be captured, necessitating the use of data-driven methods that fuse the approximation capabilities of learning and the efficiency of dynamic system identification. We demonstrate that such a neural network based subspace-encoder method can successfully capture the underlying dynamics while other methods fall short to provide reliable results.

The information about a system's dynamics represented by measurement data sets are often confined to regions of restricted operations where the system is not sufficiently excited for model identification purposes. Experiments performed in closed-loop with safety constraints allow only for reduced order modeling. In the paper, a set of low order models are identified from real experimental data of the lateral dynamics of an electric passenger car. Low order models are advantageous for on-line computation in model-based control, though uncertainty due to neglected dynamics may deteriorate control performance and constraint satisfaction. The effect of uncertainty is analyzed by controller cross-validation where a controller designed based on one model is evaluated on other models playing the role of the true system. This method allows us to qualify not only model-controller pairs, but to determine the properties of input data and model uncertainty, which lead to more useful data sets, more robust and better performing controllers than the others.