Physics informed Gaussian process models for real-time simulation of tire terrain interactions for off- road conditions

Abstract

We propose a Gaussian process machine learning model (GPM) for real-time simulation of tire-terrain interactions, especially under off-road conditions. Compared to purely empirical models or classical Neural Networks, the GPM requires much less input data for training, has greater ability to explain, and is able to quantify uncertainty in predictions. The model can seamlessly incorporate any combination of physics-based numerical simulations, empirical or semi-empirical models, and experimental data and produce real-time predictions of the interaction parameters (such as normal and shear forces, tire sinkage, etc.) along with uncertainty estimates on its predictions. The key idea is to use empirical models such as the steady-state Becker-Wong model as the baseline and “learn” the difference due to the dynamic response of the tire from detailed physics-based models or experimental data or any combination. We show that the result is able to make highly accurate predictions of the tire response in real-time. Such simplified models can be useful for training autonomous off-road vehicles under various conditions. They are also useful for virtual testing of different vehicle designs on different terrain.