Increasing the Graspable Objects by Controlling the Errors in the Grasping Points of a Suction Pad Unit and Selecting an Optimal Hand

In this study, we developed a new method utilizing two types of hands to select a two-finger gripper or a suction pad unit depending on the object, aiming to automate the task of displaying products, i.e., stacking shelves. Although a grasping point must be estimated to grasp an object, large errors occur in estimating the grasping point for conventional suction pad units. Moreover, in prior works, hand selection was often performed by stipulating rules according to the shape and posture of a given object, but these methods were unable to handle increasing number of objects. Accordingly, in this study, to improve the performance of object-grasping systems, we dealt with two problems. The first is improving the robustness of deep neural network (DNN) models toward errors in the estimation of grasping points for a suction pad unit. The second is selecting a hand without utilizing fixed rules. For the first problem, we developed a new method that is imprting flexibility to the suction pad unit and processing the input to the DNN in order to focus on the object. For the second problem, we developed a DNN to appropriately select the optimal grasping hand even with limited training data. The results of experiments that the product conveyance success rate of the proposed method was 90%, which was 44pt better than that of the conventional method.