Machine Learning-Augmented Multi-Fidelity Tire-Soil Interaction Model for Autonomous Off-Road Mobility Prediction

Project begins in 2024.

A reliable simulation tool capable of predicting off-road mobility on complex granular deformable terrain is critical to the development of autonomous navigation algorithms for ground vehicles. In particular, consideration of vehicle-terrain interaction characteristics in path planning algorithms is crucial to successful autonomous navigation missions in highly stochastic terrain conditions, and the navigation algorithms must be rigorously assessed prior to field test and operation. The high-fidelity physics-based complex terramechanics (CT) models in NG-NRMM using the discrete element (DE) method have proven effective in predicting complex tire-soil interaction behavior accurately. In past years, extensive efforts have been made to improve the computational intensity of the CT model by leveraging high-performance computing and enhanced hierarchical multiscale modeling approaches. However, virtual autonomous navigation testing requires numerous simulation runs, involving long-distance driving over various topography in stochastic soil conditions, resulting in a prohibitively high computational cost.

To address modeling and computational challenges in the NG-NRMM terramechanics models for simulation-based assessment of autonomous mobility systems, this project proposes a new multi-fidelity mobility model that fuses ST and CT models through a machine learning (ML) approach to account for complex deformable terrain dynamics at a low computational cost. The novelty of this work lies in bridging two different fidelity models in terramechanics to eliminate limitations of the current NG-NRMM ST and CT models for virtual testing of autonomous mobility systems. Whereas improvements of the ST and CT models by themselves have been previously explored, the idea of leveraging the strengths of both approaches and pursuing multi-fidelity solutions has never been explored in the past. This project will, therefore, establish a new research foundation for off-road mobility simulations based on a multi-fidelity modeling approach. Furthermore, to address challenges in quantifying variability of soil model parameters for off-road mobility simulations, a Bayesian calibration procedure will be introduced to the proposed multi-fidelity mobility model.

The following fundamental research questions underpin the proposed work:

1. Can the computationally cheaper ST model be fused with the high-fidelity CT model to enhance its soil modeling capability while ensuring a low computational cost?
2. How do the quasi-static soil modeling assumptions in the ST model impact the accuracy of off-road mobility prediction?
3. How should the coupled transient soil material behavior resulting from the dynamic interaction with deformable tires be characterized?
4. How can the confidence level of calibrated soil parameters be quantified for off-road mobility simulations
5. How good is good enough in off-road mobility prediction to ensure a reliable simulation-based assessment of autonomous mobility systems?

#3.25