Scrap Steel Grabbing Robot System Based on Grasp Quality Convolutional Neural Network

With the rapid development of the steel industry, the output of scrap steel has increased greatly which raises the necessity of promoting the recycling of scrap steel resources. In the traditional scrap steel detection and sampling scheme, workers are needed to operate the grasping machine, and then experts are required to manually sample scrap steel, which is more manpower and time consuming. In scrap steel inspection involving mechanical arm, due to similar physical and chemical properties of scrap steel samples, similar surface texture, unstructured stacking mode, specialized samples to be tested and other factors, the traditional grasping solution has the characteristics of long time and low generalization ability in the occasion of scrap steel inspection. In this paper, based on deep learning algorithm, the object capture detection in scrap quality inspection scene image is studied. Firstly, the surface point cloud information and image information of the target scrap heap are obtained by using linear structured light sensor, and preprocessing operations such as filtering and denoising are carried out. Secondly, aiming at the complex structure of scrap mixed stack, OTSU algorithm is used to segment the foreground and background to determine the actual boundary of the target in the scene and obtain the location information of the object. Then, according to the obtained point cloud information, several grasp candidates are generated and ranked by the trained grasp quality convolutional neural network (GQ-CNN) to obtain the optimal grasp under the current evaluation network. Finally, the manipulator is driven to grasp the target object. The experimental results show that the method in this paper can achieve a relatively robust grasp of objects in scrap steel stack.