A Robot Grasping System With Single-Stage Anchor-Free Deep Grasp Detector
Hu Cheng, Yingying Wang, Max Q.‐H. Meng
Abstract
Robot grasp estimation in an unstructured environment for complex objects is a challenging and not solved topic. In this article, we propose a novel grasp pose detection deep model that can generate grasp poses for unknown objects using a single RGB or depth image. By regarding the grasp poses as rotated bounding boxes in the image plane, we design an efficient single-stage anchor-free fully convolutional style grasp generation model, which outputs a six-channel image representing the keypoints and the corresponding geometries of the grasp rectangles. Our model generates an enormous number of grasps in the pixel level without any time-consuming intermediate grasp candidate generation procedures. Moreover, a U-shaped model structure extracts and combines low-level high-resolution features and high-level abstract features to improve the grasp detection accuracy. We validate our model on two public grasp datasets and achieve state-of-the-art grasp detection accuracy and efficiency. Practical grasp experiments further prove the effectiveness of our grasp system in unstructured environments.