Instrumentation, Sensing, and Off-Road Vehicle Performance Modeling; Integrating Classical Terramechanics and Machine Learning

TR Number

Date

2026-06-11

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Estimating tire forces on deformable terrains in real time is essential for mobility prediction and traction control. The complex off-road tire-terrain interaction leads to dynamically varying contact geometry, shear stress distribution, and resistive forces. Existing empirical formulations and high-fidelity simulations based on classical terramechanics or finite element models either lack robustness across terrain types or are computationally expensive for onboard deployment. These limitations motivate the central question of this dissertation: how do machine learning (ML) and deep learning (DL) methods fare against classical physics models in terramechanics, and under what conditions can they meaningfully extend their reach? The study begins with simple regression-based forward prediction models to predict traction under standardized test conditions where the input-output relationship is analytically unknown. Here, ML serves as a predictor as well as a diagnostic tool, ranking which input parameters most influence traction behavior. The work then moves to problems where the correlation between inputs and outputs is physically understood but the output is too high-dimensional for scalar models. Contact patch parameters are significant to predict traction, but estimating such spatially distributed quantities from one-dimensional vector input requires large datasets. The models evaluate methods to use empirical-physical relationships to backfill sparse experimental datasets and keep the models tractable. The next step addresses the inverse problem, where forward mechanics is well established, but the inverse has no solution. This implements a physics-informed network guided by a modified analytical formulation and a physics-based loss term. This estimated stress and contact geometry feeds into a semi-empirical drawbar pull framework that partitions traction forces into shear, normal resistance, bulldozing, and lug effects with uncertainty quantification throughout. This progression from sensing contact conditions, to estimating distributed forces, to predicting traction output maps directly onto the sensing and estimation stages an onboard system need. For control applications, the requirement changes as models must predict accurately and preserve the mathematical structure like Jacobian and Hessian properties that gradient-based optimization and model predictive control depend on. The work evaluates how data-driven models fare against physics-based models for such structural requirements, completing the sensing-estimation-control loop that real-world mobility applications demand. A consistent observation is that ML augmented by physics performs better across all fidelity levels, as physical constraints keep outputs consistent and preserve mathematical structure. Also, ML models interpolate within their training distribution but cannot extrapolate beyond it they learn patterns physics produces without understanding the mechanisms behind them. In terramechanics, where incorrect force estimates affect vehicle safety, physics-based constraints become essential. This work demonstrates that as problem complexity grows from forward prediction through inverse modeling to real-time control, ML methods remain effective only when implemented with increasing physical constraints. Together, these contributions advance the practical deployment of terramechanics knowledge in autonomous and human-operated off-road vehicles.

Description

Keywords

Terramechanics, Tire-Terrain Interaction, Machine Learning, Subsoil stress, Physics-Informed Neural Networks, Model Predictive Control

Citation