Degree

Doctor of Philosophy (PhD)

Department

Division of Electrical and Computer Engineering

Document Type

Dissertation

Abstract

Autonomous vehicles and electric mobility are converging to redefine modern transportation, yet their full potential is constrained by challenges in motion planning, energy-efficient motion control, and reliable operation in complex and unstructured environments. Existing trajectory planning approaches sacrifice optimality or rely on simplified approximations that over-constrain the feasible space. Motion control strategies lack adaptability to varying operating conditions. Data-driven discrete-time predictive models compromise accuracy for nonlinear dynamics and are rarely evaluated across safety-critical scenarios. To address these gaps, three contributions are presented.

First, a bi-stage optimal trajectory planning framework is developed for autonomous vehicles in unstructured environments. A feasible path is generated using an optimal RRT-based sampling planner, then used as a warm-start for a constrained optimal control problem solved through direct multiple shooting. Obstacles and the vehicle are represented as polytopes, and collision avoidance is reformulated using convex duality to handle the full vehicle shape. An obstacle region reduction strategy and first-stage planner reduce complexity and accelerate convergence. The framework generates smooth, collision-free, dynamically feasible trajectories, achieving up to 60% reduction in computation time over the baseline.

Second, a reinforcement learning-based adaptive motion control strategy is proposed for four in-wheel motor actuated electric vehicles. A TD3-QuadPID architecture is developed in which an actor-critic agent tunes multiple PID controllers for torque allocation and steering subject to tight motion constraints. Evaluated on a high-fidelity Blackbox vehicle model, the controller achieves more than 89% improvement in energy efficiency while generalizing to unseen environments.

Third, a data-driven nonlinear model predictive control framework is developed for autonomous vehicles when no physical model is available. A continuous-time neural statespace model is learned using the Neural Ordinary Differential Equation framework and embedded into a nonlinear model predictive control formulation. A time-scheduled horizon strategy improves feasibility and closed-loop performance. The model achieves prediction accuracy averaging 84%, tracking errors below 0.04 RMSE, and complete collision avoidance in safety-critical situations.

Overall, this dissertation contributes new methods for trajectory planning and motion control by leveraging control theory, optimization, reinforcement learning, and data-driven neural dynamic modeling, advancing autonomous driving systems that are safe, efficient, adaptive, and applicable in realistic scenarios.

Date

3-27-2026

Committee Chair

Meng, Xiangyu

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Available for download on Monday, March 26, 2029

Share

COinS