Multi-Fidelity Machine Learning Modeling for Wheeled Locomotion on Soft Soil

Abstract

Wheeled vehicles are the most convenient and widespread locomotion machines for the majority of research, industrial or private tasks. A perceptible share of wheeled vehicles is used on soft soil. Modelling wheel locomotion in these situations is challenging, because of the non-proportional relation between applied shear stress and the soil’s deformation. Currently, various conventional simulation approaches are used to describe wheel–soil interaction, ranging from detailed numerical methods with particle-level simulations to simpler empirical models, where a big part of physical formulas are set up a priori, empirically. The ultimate wheel locomotion modelling tool should have high-quality onboard predictions but within a reasonable time. The trade-off is unachievable with the current simulation tools. In this project, we argue that using Machine Learning (ML) we can build a tool with the quality of high-fidelity and speed of lower-fidelity simulations. To fit this requirement, we are combining data from several models with different fidelities, in order to build a multi-fidelity ML model. In the model, forces and torques acting on the wheel are predicted using input data like the wheel’s trajectory, surface and soil characteristics. The quality of this model will be validated by Terramechanics Robotics Locomotion Laboratory (TROLL) at Deutsche Zentrum für Luft- und Raumfahrt (DLR), a robotic single-wheel test bed designed to perform wheel–soil interaction experiments automatically. Early results show that, in simplified scenarios, our proposed method can be used to create efficient, multi-fidelity numerical models for locomotion prediction, including uncertainty estimation for the predictions.