A grasp detection method for industrial robots using a ConvNet

Abstract

In the near future, most of the industrial robots will serve as assistants involved in targeted complex manufacturing tasks which are difficult to be automated. To achieve this, it is crucial to enhance the ability of manipulators to pick and place objects from the assembly line. Reorienting and picking up pieces for assembly are difficult to be done by manipulators since, for different pieces, shapes and physical properties vary. In this work, we use Convolutional Neural Networks for recognizing a selected production piece on a cluster. Once the selected piece has been recognized, a grasping algorithm estimates the best
gripper configuration so that the robot is able to pick the piece up. We tested our algorithm on grasping experiments with an ABB robot and using a common webcam as image input. We found that our implementations perform well and the robot was able to pick up a variety of objects.