A Robotic Semantic Grasping Method For Pick-and-place Tasks

Abstract

In recent years, it has attracted significant interest for mobile robots to complete the grasping tasks. A fully autonomous robotic pick-and-place system requires dependable object recognition and localization in cluttered environments. Most of current robot grasping methods based on learning mostly train the models with exploiting human-labeled datasets, then grasping perpendicularly to the plane where the object is placed. In this paper, a robotic semantic grasping method is proposed to estimate six-degree-of-freedom (6DOF) grasping pose for the robotic manipulator, thereby a grip perpendicularly to the surface of the object can be achieved. Firstly, the graspable location is detected by an instance segmentation detection model and ellipse fitting method. Secondly, by processing the pixel mask and point clouds of the target object, the 6DOF grasping pose is estimated. Finally, the proposed method was implemented and tested in a real robot manipulator. Experimental results show that the proposed semantic grasping method enables to accomplish pick-and-place tasks in an unstructured scene.