Robotic perception and manipulation of garments

Duan, Li (2023) Robotic perception and manipulation of garments. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2023DuanPhD.pdf] PDF
Download (18MB)


This thesis introduces an effective robotic garment flattening pipeline and robotic perception paradigms for predicting garments’ geometric (shape) and physics properties.

Robotic garment manipulation is a popular and challenging task in robotic research. Due to the high dimensionality of garments, object states of garments are infinite. Also, garments deform irregularly during manipulations, which makes predicting their deformations difficult. However, robotic garment manipulation is an essential topic in robotic research. Robotic laundry and household sorting play a vital role in an ageing society, and automated manufacturing requires robots to be able to grasp different mechanical components, some of which are deformable objects. Also, robot-aided garment dressing is essential for the community with disabilities. Therefore, designing and implementing effective robotic garment manipulation pipelines are necessary but challenging.

This thesis mainly focuses on designing an effective robotic garment flattening pipeline. Therefore, this thesis is divided into two main parts: robotic perception and robotic manipulation. Below is a summary of the research in this PhD thesis:

• Robotic perception provides prior knowledge on garment attributes (geometrical (shape)
and physics properties) that facilitates robotic garment flattening. Continuous perception
paradigms are introduced for predicting shapes and visually perceived garments weights.

• A reality-simulation knowledge transferring paradigm for predicting the physics properties of real garments and fabrics has been proposed in this thesis.

• The second part of this thesis is robotic manipulation. This thesis suggests learning the known configurations of garments with prior knowledge of garments’ geometric (shape) properties and selecting pre-designed manipulation strategies to flatten garments. The robotic manipulation part takes advantage of the geometric (shape) properties learned from the robotic perception part to recognise the known configurations of garments, demonstrating
the importance of robotic perception in robotic manipulation.

The experiment results of this thesis revealed that: 1). A robot gains confidence in prediction (shapes and visually perceived weights of unseen garments) from continuously perceiving video frames of unseen garments being grasped, where high accuracies on predictions (93% for shapes and 98.5 % for visually perceived weights) are obtained; 2). Predicting the physics properties of real garments and fabrics can be realised by learning physics similarities between simulated fabrics. The approach in this thesis outperforms SOTA (34 % improvement on real fabrics and 68.1 % improvement for real garments); 3). Compared with state-of-the-art robotic garment flattening, this thesis enables the flattening of garments of various shapes (five shapes) and fast and effective manipulations. Therefore, this thesis advanced SOTA of robotic perception and manipulation (flattening) of garments.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Colleges/Schools: College of Science and Engineering > School of Computing Science
Supervisor's Name: Aragon Camarasa, Dr. Gerardo
Date of Award: 2023
Depositing User: Theses Team
Unique ID: glathesis:2023-83456
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 27 Feb 2023 16:23
Last Modified: 28 Feb 2023 09:01
Thesis DOI: 10.5525/gla.thesis.83456
Related URLs:

Actions (login required)

View Item View Item


Downloads per month over past year