A grasping posture estimation method based on 3D detection network

Abstract

In robotic grasping tasks, different kinds of objects with different poses and positions are often distributed in the scene, which makes the design a feasible visual solution for robotic grasping very challenging. In this paper, we propose a robotic grasping method based on a 3D detection network, which reduces the influence of different camera positions in image recognition. First, a convolutional neural network is designed to recognize objects in RGB images and calculate the 3D bounding boxes and center points of objects. Second, to further improve the stability of the whole grasping system, we propose a strategy for calculating the best object grasping posture. Finally, robotic control is implemented to grasp objects in real scenes. Experimental results show that the 3D detection accuracy reached 88%, and the successful grasp rate reached 94% in real scenes. In summary, the grasping system designed in this paper can generate appropriate grasping postures effectively and achieve high successful grasp rates, meeting the higher requirements of grasping tasks.