Evaluating grasp quality metrics of cloth like deformable objects in simulation

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Insinööritieteiden korkeakoulu | Master's thesis

Date

2024-06-10

Department

Major/Subject

Mechanical Engineering

Mcode

Degree programme

Master's Programme in Mechanical Engineering (MEC)

Language

en

Pages

59

Series

Abstract

Robotic cloth manipulation poses unique challenges due to the complexity and flexibility of textiles. Cloth manipulation involves a delicate interplay of grasping, folding, draping, and arranging fabrics, which requires sophisticated robotic systems capable of precise control and adaptability. Although evaluation of grasp quality metrics on deformable solids has recently started to interest researchers, the evaluation of standard grasp quality metrics on clothes remains underexplored. In this thesis, we evaluate well-known grasp quality metrics in simulation across different cloth configurations and orientations. We then compare the data collected with the qualitative assessment of grasp by studying individual images and correlate the grasp quality metric with human intuition. The experimental evaluation setup uses the MuJoCo physics simulator with the Robosuite manipulation-specific framework to instantiate reinforcement learning-ready environments. The primary research questions are how well the standard grasp quality metrics perform when grasping cloth-like deformableobjects compared to human i ntuition of what constitutes a good grasp. Upon comparing images of grasping and the corresponding quality metric, it is evident that using the standard metrics for cloth manipulation is not suitable as they do not provide enough discrimination across different grasps in a smooth manner. Although there are some benefits of using the standard metrics, such as being easier to interpret, they cannot be relied upon to give full confidence on the grasp quality, and more research is needed on grasp quality metrics for deformable manipulation.

Description

Supervisor

Kyrki, Ville

Thesis advisor

Blanco Mulero, David
Nguyen Le, Tran

Keywords

physics simulation, grasping, mujoco, robosuite, manipulation, reinforcement learning

Other note

Citation