Memory Efficient Deep Learning-Based Grasping Point Detection of Nontrivial Objects for Robotic Bin Picking

Abstract

Picking up non-trivial objects from a bin with a robotic arm is a common task of modern industrial processes. Here, an efficient data-driven method of grasping point detection, based on an attention squeeze parallel U-shaped neural network (ASP U-Net) for the bin picking task, is proposed. The method directly provides all necessary information about the feasible grasping points of objects, which are randomly or regularly arranged in a bin with side walls. Moreover, the method is able to evaluate and select the optimal grasping point among the feasible ones for two types of end effectors, i.e., a vacuum cup and a parallel gripper. The key element of the utilized ASP U-Net neural network is the transformation of a single RGB-Depth image of the bin containing nontrivial objects into a schematic grey-scale frame, where the positions and poses of the grasping points are coded into gradient geometric shapes. The experiments carried out in this study include a comprehensive set of scenes with randomly scattered, ordered, and semi-ordered objects arranged in impeccable or deformed bins. The results indicate outstanding accuracy with more than acceptable computational requirements. Additionally, the scaling possibilities of the method can offer extremely lightweight implementations, applicable, for example, to battery-powered edge-computing devices with low RAM capacity.