Target Driven Object Grasping in Highly Cluttered Scenarios through Domain Randomization and Active Segmentation

Mak, Ivar (2022) Target Driven Object Grasping in Highly Cluttered Scenarios through Domain Randomization and Active Segmentation. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text MASTER_THESIS-IVARMAK.pdf Download (6MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (130kB)

Abstract

Robots nowadays are demanded to perform tasks that divert from controlled envi- ronments. With regards to robotic grasping systems, dynamic situations pose prob- lems in multiple ways. Grasping a target object from a pile might be impeded by other objects that obscure the robot’s view, or restrict the grasping motion. In this work, a deep learning approach is presented that is capable of flexible object ma- nipulation in highly cluttered environments. Using a Neural Network (NN) for ob- ject detection and -segmentation, and a second NN for grasp synthesis, a system is built that handles robust object grasping in human-centric domains. These consist of household objects located on a table, which are manipulated by a robotic arm with a two-fingered grasping hand in order to single out a target object. To evaluate the performance of the proposed approach, four sets of experiments are performed, with the objects being in isolated, packed, piled, and cluttered scenarios. An experi- ment is considered a success when the target object is placed in its corresponding tray, located next to the table. Experimental results show accuracy scores of 94% for isolated, 89% for packed, 85% for piled, and 80% for cluttered environments.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Mohades Kasaei, S.H. and Carloni, R.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	30 Aug 2022 13:53
Last Modified:	30 Aug 2022 13:53
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28346

Actions (login required)

View Item