Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : September 11, 2022 - September 14, 2022
Bin picking is a parts handling process in factory production lines. There are various challenges in automating bin picking. In particular, the more complex the object shape is, the more difficult it becomes to recognize the position and posture of the object using a camera and to determine the appropriate position and posture of the end-effector. Using a Convolutional Neural Network (CNN), a deep learning technique for image recognition, the characteristics of the grasping objects, and grasping with a high success rate would expected. However, due to the limited number of simulation trial, the estimated position and posture is not always the optimal grasping posture. In this study, we modify the position and posture of the end-effector for handling robots using deep learning. Several candidates for position and posture of the end-effector are estimated using three types of CNNs. For each candidate for position and posture of the end-effector, the actual grasping position is determined using the balance between the estimated grasping possibility and the posture of the end-effector. Experimental results showed that integration of the CNN result and posture of the end-effector improved grasp success rate.
