Date of Award
12-2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Automotive Engineering
Committee Chair/Advisor
Venkat Krovi
Committee Member
Umesh Vaidya
Committee Member
Yunyi Jia
Committee Member
Rahul Rai
Abstract
This dissertation advances data-driven modeling and adaptive control techniques for Uncrewed Ground Vehicles (UGVs), with a focus on autonomy in mission-critical and safety sensitive environments. UGVs are deployed across a wide spectrum of domains, from structured manufacturing shop floors to unstructured off-road terrains, including planetary exploration, precision agriculture, and disaster response. These platforms, operating in dull, dirty, and dangerous conditions, demand autonomy that is both adaptable and robust. While traditional model-based control methods offer interpretability and robustness, they struggle with unmodeled dynamics, parameter variations, and integration of high-dimensional sensing. Conversely, modern machine learning approaches can directly exploit sensory data but often lack generalizability, predictive power, and safety guarantees. This work addresses the critical challenge of unifying these complementary paradigms through the Koopman operator framework, which enables linear representations of nonlinear dynamics within a data-driven, mathematically rigorous setting.
The dissertation presents a comprehensive methodological pathway for data-driven modeling, motion planning, and control of on-road and off-road UGV platforms across multiple scales. It begins with a novel workflow for data-driven discovery of vehicle dynamics from temporal snapshots of states and control inputs, framed through Koopman operator theory. A systematic pipeline for model discovery and control design is introduced, encompassing the construction of observables, feedback controller synthesis, and validation from simulation to physical platforms. This foundation is further extended by introducing spectral reduction techniques that extract dominant eigenfunctions, producing stable reduced-order models that retain predictive accuracy while remaining computationally efficient. Building on these advances, the Multi-Model Parameterized Koopman (MMPK) framework is developed to overcome explicit pose dependency, enabling dynamics representation across the entire configuration space through curvature-parameterized multi-model switching. This approach not only achieves pose-agnostic modeling but also enhances fault tolerance by supporting reachability aware trajectory planning. The framework is further extended to off-road domains by embedding terrain-induced load transfer effects and exogenous disturbances within adaptive Koopman formulations, allowing real-time model updates for improved robustness. These contributions culminate in Project Varuna, a standalone Koopman-based autonomy stack that unifies perception, planning, and control, and is demonstrated through extensive simulation studies and real-world UGV deployments.
Collectively, this research demonstrates how Koopman Operator Theory can serve as a unifying foundation for explainable, data-driven autonomy. By combining model discovery, spectral reduction, adaptive multi-model planning, and real-world validation, the dissertation contributes both theoretical advancements and practical tools that expand the operational boundaries of UGVs in complex, uncertain environments.
Recommended Citation
Joglekar, Ajinkya, "Data-Driven Discovery of Finite-Dimensional Koopman Operator for Modeling and Control of Uncrewed Ground Vehicles" (2025). All Dissertations. 4173.
https://open.clemson.edu/all_dissertations/4173
Author ORCID Identifier
0000-0002-7602-9796