Archived

This content is available here for research, reference, and/or recordkeeping.

Author ORCID Identifier

https://orcid.org/0009-0006-0906-0431

Date Available

5-18-2026

Year of Publication

2026

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

Engineering

Department/School/Program

Electrical and Computer Engineering

Faculty

Biyun Xie

Faculty

Daniel Lau

Abstract

Safe and efficient motion planning is a core requirement for robots operating in complex real-world environments, yet traditional model-based approaches struggle to simultaneously achieve computational efficiency, solution optimality, and the ability to handle complex constraints and potential hardware failures. This dissertation investigates how neural networks, as general purpose function approximators, can be leveraged in both traditional and novel ways to overcome these limitations and advance beyond what purely model-based methods can offer. Three distinct contributions are presented. First, a new motion planner is developed that uses ReLU neural networks to decompose the configuration space into linear cost regions, enabling existing convex optimization frameworks to handle arbitrary non-convex cost functions. Second, the piecewise-linear structure of ReLU neural networks is exploited to construct a complete geometric representation of constraint manifolds offline, decoupling manifold computation from online planning and enabling fast trajectory generation. Third, a learning-based algorithm is introduced that uses Fourier series and neural networks to efficiently approximate self-motion manifolds, supporting real-time global fault-tolerant motion planning for kinematically redundant robots. The effectiveness of each contribution is demonstrated through a series of simulation and hardware experiments, validating the proposed methods across a range of robotic systems and planning scenarios.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2026.322

Archival?

Archival

Funding Information

This research was supported by the Graduate Assistance in Areas of National Need fellowship in 2024-2025, the National Science Foundation under Grant #2205292 in 2022-2025, and the National Aeronautics and Space Administration (NASA) and the NASA Kentucky Established Program to Stimulate Competitive Research award under NASA award number 80NSSC22M0034 in 2020-2021.

Included in

Robotics Commons

Share

COinS