Simulation Environments for Learning Manipulation of Cloth-like 3D Bags
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2024-08-19
Department
Major/Subject
Control, Robotics and Autonomous Systems
Mcode
ELEC3025
Degree programme
AEE - Master’s Programme in Automation and Electrical Engineering (TS2013)
Language
en
Pages
55
Series
Abstract
There has been increased use of robots in households compared to the past decade. However, robots still suffer from being unable to manipulate common household deformable objects such as bags with the same dexterity as rigid objects. This thesis aims to bridge the gap by developing a simulation framework that can be used to train reinforcement learning algorithms for bag manipulation. It creates four bag manipulation environments with varying difficulties and a common goal of lifting the deformable bag to enclose a rigid cylinder. These environments help to understand the interactions between a rigid manipulator and a deformable object. Two environments use the full state of the bag in the form of vertex positions, and the other two environments use visual inputs as observations, making learning difficult. Rewards are engineered to split the manipulation into subtasks. The effects of using final sparse reward and subtask rewards are compared, and it is observed that the "sparse rewards" approach does not converge to learn the task while the latter approach learns to perform the task successfully. The effects of using full vertex position and RGB observations are also compared. When vertex positions are used as observations, the agent accumulates a higher reward than when using RGB observations. The stability and effective frame rate of the simulations are also measured for a random agent across the four environments, it worked without potential crashes and unretrievable scenarios.Description
Supervisor
Kyrki, VilleThesis advisor
Shintemirov, AlmasKeywords
deformable objects, manipulation, reinforcement learning, simulations