A Method for Object Recognition and Robot Grasping Detection in Multi-object Scenes

Abstract

AbstractDue to robot grasping has always been an open challenge and a difficult problem, it has attracted many researches. With the wide application of deep learning methods in robot grasping, the grasping performance of robots has been greatly improved. Traditionally, many robot grasp detection approaches focus on how to find the best grasp. However, it is very important for robots to have an object recognition function during the grasping process to meet more industrial requirements, such as industrial assembly tasks and sorting tasks. In addition, the problem of missed detection has always existed in the current multi-object grasping detection. To solve the above problems, this paper proposes a two-stage robot grasping method to recognize objects and detect the most likely grasp for every object in multi-object scenes. Our approach achieved a detection accuracy of 65.7% on the VMRD dataset and outperformed the benchmark algorithm by 11.2%. The simulation experimental results showed that our approach achieved a recognition success rate of 96.7% and a grasp success rate of 85% for robotic grasp detection in multi-object scenes.KeywordsRobot graspingObject detectionGrasping detectionDeep learning