Deep learning of grasping detection for a robot used in sorting construction and demolition waste

The traditional construction and demolition waste (CDW) treatment process adopts the method of crushing and screening after mixing and combines the method with manual sorting for resource recycling. However, there is a problem of low recycling purity and low efficiency of manual sorting after mixed screening. This paper proposes a robot for sorting CDW, which is used to finely sort a large number of objects before mixing and crushing. The use of the robot improves the level of resource utilization of CDW. However, under actual working conditions, the adhesion and stacking of CDW on the conveyor belt and the irregularity of the shapes of CDW lead to errors in grasping-information. Thus, a deep learning method for grasping detection is proposed. The method generates some grasping rectangles through a searching algorithm, and inputs the rectangles to the neural network. Then, the network outputs the optimal grasping pose. The experiment demonstrated that the original accuracy of robotic grasping was only 70%. After deep learning for grasping detection, the accuracy was over 90%, which thoroughly meets the requirements of efficiency and accuracy for sorting CDW under actual working conditions.