Visual Perception System for Randomized Picking Task

Abstract

Randomized picking is a classical, high practical value, but complex mission for collaborative robots. In this paper we presents an efficient and robust perception system for this task, which is capable of detecting and classifying each instance in occlusion environment as well as outputting the 6D pose of target object for grasping. For system running efficiency, we design a grasping strategy which can automatically select appropriate target object among multiple ones, deal with the situation of object point cloud insufficient in amount of points and correct a particular wrong registration result that violate common sense. A gripper open degree estimation algorithm is also presented so as to prevent fingers colliding with neighbor objects of the target. Finally, to test the effectiveness and robustness of our proposed approaches, we show an experimental result of the whole robot system.