Physics-Aware Vehicle Dynamics Modeling for Enhanced High-Speed Autonomous Navigation

Date of Award

8-15-2026

Degree Name

M.S. in Mechanical Engineering

Department

Department of Mechanical and Aerospace Engineering

Advisor/Chair

Krishna Bhavithavya Kidambi

Abstract

Autonomous racing provides a challenging platform for evaluating high-speed navigation algorithms as these vehicles operate near the limits of tire-road interaction, actuator authority, and dynamic stability. Under these operating conditions, accurate vehicle dynamics modeling is essential for reliable prediction, trajectory optimization, and closed-loop control. Classical physics-based models ofer interpretability and computational efciency, but they often rely on fxed parameters and simplifying assumptions that limit their accuracy during aggressive maneuvers. In contrast, purely data-driven models can capture complex nonlinear behavior, but they may lack physical consistency, interpretability, and robustness when used recursively inside model-based controllers. This thesis investigates physics-aware learning methodology for vehicle dynamics modeling and closed-loop controller evaluation in autonomous racing. A physics-aware state-space memory framework, DynSSM, is developed to combine structured vehicle dynamics with learned temporal representations. DynSSM integrates state-space sequence modeling and recurrent encoding to capture both long-term operating context and short-term actuator-induced transients. The learned temporal features are used to adapt physically meaningful vehicle parameters within bounded ranges, while a gated residual correction compensates for remaining local model mismatch without replacing the analytical dynamics pathway. The model is evaluated using simulated small-scale racing data and real-world full-scale autonomous racecar data through open-loop prediction, component-wise ablation studies, and closed-loop nonlinear model predictive control simulations. Results show that physics-aware learned dynamics can improve prediction accuracy, computational efciency, controller feasibility, and lap-completion performance. Overall, this thesis demonstrates that combining temporal memory, bounded physics-guided adaptation, and gated residual refnement provides a practical pathway toward accurate, interpretable, and control-ready vehicle dynamics learning for high-speed autonomous racing.

Keywords

Robotics

Rights Statement

Copyright 2026, author

Link to Full Text

Share

COinS