APSIPA Transactions on Signal and Information Processing > Vol 12 > Issue 1

GP-Net: A Lightweight Generative Convolutional Neural Network with Grasp Priority

Yuxiang Yang, School of Electronics and Information, Hangzhou Dianzi University, China and School of Computer Science, The University of Sydney, Australia, Yuhu Xing, School of Electronics and Information, Hangzhou Dianzi University, China, Jing Zhang, School of Computer Science, The University of Sydney, Australia, Dacheng Tao, School of Computer Science, The University of Sydney, Australia, dacheng.tao@gmail.com
 
Suggested Citation
Yuxiang Yang, Yuhu Xing, Jing Zhang and Dacheng Tao (2023), "GP-Net: A Lightweight Generative Convolutional Neural Network with Grasp Priority", APSIPA Transactions on Signal and Information Processing: Vol. 12: No. 1, e26. http://dx.doi.org/10.1561/116.00000002

Publication Date: 25 Apr 2023
© 2023 Y. Yang, Y. Xing, J. Zhang and D. Tao
 
Subjects
 
Keywords
Robotic graspingGenerative convolutional networkCalibrated global context moduleGrasp priority
 

Share

Open Access

This is published under the terms of CC BY-NC.

Downloaded: 701 times

In this article:
Introduction 
Related Work 
The Proposed Method 
Experiment 
Conclusions 
References 

Abstract

Grasping densely stacked objects may cause collisions and result in failures, degenerating the functionality of robotic arms. In this paper, we propose a novel lightweight generative convolutional neural network with grasp priority called GP-Net to solve multiobject grasp tasks in densely stacked environments. Specifically, a calibrated global context (CGC) module is devised to model the global context while obtaining long-range dependencies to achieve salient feature representation. A grasp priority prediction (GPP) module is designed to assign high grasp priorities to top-level objects, resulting in better grasp performance. Moreover, a new loss function is proposed, which can guide the network to focus on high-priority objects effectively. Extensive experiments on several challenging benchmarks including REGRAD and VMRD demonstrate the superiority of our proposed GP-Net over representative state-of-the-art methods. We also tested our model in a real-world environment and obtained an average success rate of 83.3%, demonstrating that GP-Net has excellent generalization capabilities in real-world environments as well. The source code will be made publicly available.

DOI:10.1561/116.00000002