A novel vision-based multi-task robotic grasp detection method for multi-object scenes

Yanan Song, Liang Gao, Xinyu Li, Weiming Shen, Kunkun Peng

Abstract: Grasping a specified object from multi-object scenes is an essential ability for intelligent robots. This ability depends on the affiliation between the grasp position and the object category. Most existing multi-object grasp detection methods considering the affiliation rely on object detection results, thus limiting the improvement of robotic grasp detection accuracy. This paper proposes a decoupled single-stage multi-task robotic grasp detection method based on the Faster R-CNN framework for multi-object scenes. The designed network independently detects the category of an object and its possible grasp positions by using one loss function. A new grasp matching strategy is designed to determine the relationship between object categories and predicted grasp positions. The VMRD grasp dataset is used to test the performance of the proposed method. Compared with other grasp detection methods, the proposed method achieves higher object detection accuracy and grasp detection accuracy.