Abstract: The research focuses on detecting robotic arm grasping poses using RGB-D cameras. A novel framework combining YOLO segmentation and GSNet grasp detection is designed to identify and localize target objects for robotic arm grasping. First, the YOLOv8n-seg segmentation model extracts and segments target objects in RGB images. The segmented features are then mapped to the depth map, enhancing the 3D positioning accuracy of the target. Next, the Iterative Closest Point (ICP) algorithm registers the RGB and depth data into a fused point cloud, which is fed into the GSNet model for grasp pose detection. To address GSNet's limitation in handling multi-scale objects—particularly the biased sampling distribution that leads to unreliable grasping of small objects—a multi-scale cylindrical grouping method (RANSNC) is introduced. This optimization adaptively aggregates features across scales, improving the success rate of grasp pose detection for objects of varying sizes.
颜维凤, 缑燕飞, 甘树坤, 吕雪飞. 基于RGB-D相机的机械臂抓取位姿检测[J]. 吉林化工学院学报, 2025, 42(5): 77-83.
YAN Weifeng, GOU Yanfei, GAN Shukun, LV Xuefei. Robotic arm grasping pose detection based on RGB-D camera. Journal of Jilin Institute of Chemical Technology, 2025, 42(5): 77-83.
