Recognition and robot grasping of disordered workpieces with 3D laser line profile sensor

Abstract

Disorderly grasping of mixed or stacked workpieces are widely used in the manufacturing industry, especially in customized assembly lines. The mutual occlusion between workpieces causes geometrical information loss, thus making it difficult to obtain precise poses when grasping the workpieces. To provide a better robotic disorderly grasping solution, we firstly use a 3D laser line profile sensor to acquire the point clouds of the workpieces. Then the point clouds are inputted into a relation-shape convolutional neural network (RS-CNN) for instance-level recognition. According to the recognition results, coarse point cloud registration as well as fine point cloud registration are employed to estimate the pose of each workpiece. The disordered workpieces will be picked up by a 6-axis industrial robot in priority order. Experimental results show that the proposed recognition and grasping system can achieve a higher grasping success rate on a set of four types of common workpieces with adversarial geometry.