Title: Transfer learning-based and originally-designed CNNs for robotic pick and place operation

Authors: Fusaomi Nagata; Maki K. Habib; Keigo Watanabe

Addresses: Graduate School of Engineering, Sanyo-Onoda City University, 1-1-1 Daigaku-Dori, Sanyo-Onoda 756-0884, Japan ' Mechanical Engineering Department, School of Sciences and Engineering, American University in Cairo, AUC Avenue, P.O. Box 74, New Cairo 11835, Egypt ' Graduate School of Natural Science and Technology, Okayama University, 3-1-1 Tsushima-Naka, Kita-Ku, Okayama 700-8530, Japan

Abstract: The authors have developed a CNN and SVM design and training application for defect detection, and the effectiveness and the usefulness have been proved through several design, training and classification experiments. In this paper, the application further enables to facilitate the design of transfer learning-based CNNs. After introducing the application, a pick and place robot system based on DOBOT is proposed while implementing a visual feedback controller and a transfer learning-based CNN. The visual feedback controller is applied to avoiding the complicated calibration task between image and robot coordinate systems, also the transfer learning-based CNN allows to detect the orientation of target objects for dexterous picking operation. The effectiveness of the proposed system is demonstrated through pick and place tests using gripper type and suction cup type tools. Finally, an originally designed CNN with shallower layers is compared with the AlexNet's transfer learning-based CNN in terms of classification scores.

Keywords: convolutional neural network; CNN; transfer learning; pick and place; robot.

DOI: 10.1504/IJMA.2021.118430

International Journal of Mechatronics and Automation, 2021 Vol.8 No.3, pp.142 - 150

Received: 07 Dec 2020
Accepted: 20 Aug 2021
Published online: 25 Oct 2021 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article