A Novel Two-Stage Robot Grasping Detection Method

Selecting predefined targets for grasping based on diverse requirements in actual environments is a critical issue in the field of robotic grasping. To address this problem, this paper propose a novel two-step cascading method. First, we propose a lightweight target detection backbone GhostNetV2S, which replaces the standard convolution with Ghost convolution to enable real-time detection of the region of interest in the RGB image. Then, the RGB image and the depth image are cropped with the region of interest as the center to reduce the interference between the object and the background. Finally, The cropped images are input into generative residual convolutional neural network to determine the optimal grasping position of the target. We conduct qualitative and quantitative comparative experiments on the Cornell and VMRD datasets. Simultaneously, we conduct a robot grasping experiment in a real-world scenario, and the success rate of robot grasping reaches 97%, experimentally verifying the reliability of the proposed method.