Date of Award

5-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering

Committee Chair/Advisor

Luyang Zhao

Committee Member

Yongqiang Wang

Committee Member

Adam W. Hoover

Abstract

Compared to rigid robot systems, soft robots and deformable objects provide clear benefits, such as adapting to the environment and moving safely around humans. However, these features make modeling and design difficult. In this thesis, we study the SoftSnap system, a modular cable-driven soft robot whose shape is defined by discrete threading patterns and continuous segment lengths. The original approach relies on an iterative optimization process that is time-consuming and may produce unstable results, making large-scale data generation and inverse design challenging.

To address these issues, this thesis proposes a learning-based framework to accelerate forward modeling and inverse design while maintaining physical accuracy. The framework includes data generation, data cleaning, forward modeling, inverse prediction, and optimization. The shape is represented by an 11-dimensional angle vector, and the structure is defined by 12 wire-routing positions and a wire length. One-hot encoding is used to handle mixed discrete and continuous inputs.

A multi-step data cleaning process removes invalid and unstable samples, reducing iterations from about 100 to 45 and doubling data generation speed. For forward modeling, multiple methods are tested, and MLP achieves the best performance, reducing computation time from minutes to near real-time.

For inverse design, results are verified using forward validation, and an uncertainty-based refinement method further improves performance. Experiments on the SoftSnap system show that the framework produces accurate and physically consistent results.

Download

Available for download on Monday, May 31, 2027

Share

COinS