Multi-object Grasping Detection in Cluttered Scenes Based on Deep Learning

In this paper, we mainly improved the grasp detection network based on the grasp pose detection (GPD) algorithm. Three Network in Network (NIN) structure blocks are used as feature extraction modules, and a fully connected layer is used for classification. And trained the network on our own dataset which these 3D models are all from GraspNetAPI, and the accuracy rate on the test set is 2% higher than the original network. And the amount of its parameters has dropped by 38%. In addition, we combine the Mask R-CNN semantic segmentation network to perform target detection and pixel-level instance segmentation on target objects, which can grasp objects of interest. Finally, we carried out grasping experiments on real robots, and the total grasping success rate was over 70%.