Grasping System for Industrial Application Using Point Cloud-Based Clustering

Abstract

In recent years, numerous studies have been conducted on the robot grasping using deep learning, which requires a lot of data and training time. This study proposes a grasping algorithm that does not require data collection and training. In addition, the hardware of the proposed system is simply configured for a quick application in industrial fields. This algorithm is performed through clustering and grasping analysis based on point clouds. First, the point cloud obtained from the 3D camera is clustered, and the cluster most similar to the 3D CAD model is selected. Next, using the selected cluster, the object pose and the grasping pose are estimated. Finally, the target object is grasped through the estimated grasping pose, and the grasped object is loaded with a predetermined pose in consideration of the object pose. In order to evaluate the performance of the proposed algorithm, the grasping and loading of the target object with a product used on the actual industrial site and the loading jig of the object were tested. The algorithm showed the success rate of 95% in grasping, transporting and loading experiments.