Suction Grasping Detection for Items Sorting in Warehouse Logistics using Deep Convolutional Neural Networks

Abstract

Items sorting in warehouse logistics is a labor-intensive and time-consuming work. Combined with computer vision and real-time motion planning technologies, industrial robots have been ideal substitutes for human beings in that cases. But picking and placing a large quantity of object categories including known and novel objects in heavily cluttered environments is really a challenging task. This paper proposes a pipeline to address suction grasping detection for isolated objects. Firstly, a two-dimensional suction configuring is proposed. Secondly, we establish a dataset including depth images, color images and suction labels for logistics warehouse scenario. Thirdly, a lightweight network named Generative Grasp Convolutional Neural Network (GG-CNN) intended for planar antipodal grasp is adapted for predicting spatial suction affordance in pixel. Finally, we get a accuracy of 91.45% on test data sets. Primary contributions of our work are: (1) a practical annotation method and dataset collecting from retail industry, (2) an innovative application of GG-CNN.