30 October 2019 in IEEE Access
This paper proposes GraspCNN, an approach to grasp detection where a feasible robotic grasp is detected as an oriented diameter circle in RGB image, using a single convolutional neural network. By detecting robotic grasps as oriented diameter circles, grasp representation is thereby simplified. In addition to our novel grasp representation, a grasp pose localization algorithm is proposed to project an oriented diameter circle back to a 6D grasp pose in point cloud. GraspCNN predicts feasible grasping circles and grasp probabilities directly from RGB image. Experiments show that GraspCNN achieves a 96.5% accuracy on the Cornell Grasping Dataset, outperforming existing one-stage detectors for grasp detection. GraspCNN is fast and stable, which can process RGB image at 50 fps and meet the requirements of real-time applications. To detect objects and locate feasible grasps simultaneously, GraspCNN is executed in parallel with YOLO, which achieves outstanding performance on both object detection and grasp detection.