3D Part Scanning, Inspection, and Representation

Tianshuang Qiu

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-96
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-96.pdf

High-quality 3D part inspection, scanning, and representation are critical for applications across manufacturing, robotics, and virtual environments. This work presents two complementary approaches to advance these capabilities. First, we introduce a novel bimanual robotic system that generates high-quality 3D Gaussian Splat models of physical objects using a single stationary camera, achieving full 360° scanning through handover regrasping to reveal occluded surfaces. This system demonstrates the potential for effective defect detection across diverse objects. Second, we systematically evaluate deep learning pipelines for predicting additive manufacturing part quality using various 3D shape representations including voxels, depth images, distance fields, and point clouds. Our research investigates the impact of dataset size, input resolution, and hyperparameter choices on model performance. Together, these approaches advance the state of automated, cost-effective 3D scanning and quality assessment critical for modern manufacturing and robotics applications.

Advisor: Ken Goldberg

\"Edit"; ?>

BibTeX citation:

@mastersthesis{Qiu:EECS-2025-96,
    Author = {Qiu, Tianshuang},
    Title = {3D Part Scanning, Inspection, and Representation},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-96.html},
    Number = {UCB/EECS-2025-96},
    Abstract = {High-quality 3D part inspection, scanning, and representation are critical for applications across manufacturing, robotics, and virtual environments. This work presents two complementary approaches to advance these capabilities. First, we introduce a novel bimanual robotic system that generates high-quality 3D Gaussian Splat models of physical objects using a single stationary camera, achieving full 360° scanning through handover regrasping to reveal occluded surfaces. This system demonstrates the potential for effective defect detection across diverse objects. Second, we systematically evaluate deep learning pipelines for predicting additive manufacturing part quality using various 3D shape representations including voxels, depth images, distance fields, and point clouds. Our research investigates the impact of dataset size, input resolution, and hyperparameter choices on model performance. Together, these approaches advance the state of automated, cost-effective 3D scanning and quality assessment critical for modern manufacturing and robotics applications.}
}

EndNote citation:

%0 Thesis
%A Qiu, Tianshuang
%T 3D Part Scanning, Inspection, and Representation
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-96
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-96.html
%F Qiu:EECS-2025-96