Suitable ASP U-Net training algorithms for grasping point detection of nontrivial objects

Abstract

Robotic manipulation with nontrivial or irregular objects, which provide various types of grasping points, is of both academic and industrial interest. Recently, a powerful data-driven ASP U-Net deep neural network has been proposed to detect feasible grasping points of manipulated objects using RGB data. The ASP U-Net showed the ability to detect feasible grasping points with exceptional accuracy and more than acceptable inference times. So far, the network has been trained using an Adam optimizer only. However, in order to optimally utilize the potential of ASP U-Net, it was necessary to perform a systematic investigation of suitable training algorithms. Therefore, the aim of this contribution was to extend the impact of ASP U-Net by recommending suitable training algorithms and their parameters based on the result of training experiments.