Single-Stage Grasping Algorithm for the Manipulator Based on Rotational Gaussian Encoding

As we all know, there is an unsolved problem for robots to accurately and quickly grasp unknown objects in an unstructured environment. In order to describe the pose information of objects more accurately, a rotational Gaussian encoding method is proposed in this paper. Different from the traditional 2-D Gaussian encoding to represent ground-truth, this method introduces rotation information based on 2-D Gaussian. Grasping is made more accurate and robust by treating the grasp pose as a rotating bounding box in the image plane. In addition, a balanced loss function is introduced in this paper. This loss function not only solves the problem of positive and negative sample imbalance in the traditional cross-entropy loss but also reduces the penalty of points around the Gaussian center location. In this paper, a pixel-level single-stage grasping algorithm is designed in an end-to-end manner without the intermediate process of using anchor points, which saves time and improves the accuracy of grasping simultaneously. Then, the proposed model in this paper is evaluated on two standard datasets, Cornell Dataset and Jacquard Dataset, achieving 96.8% and 94.7% accuracy, respectively. Finally, we use a 6 DoF UR5e robotic arm for real-world grasping experiments. The success rate of single-object grasping is 94.7% and the time to generate the grasp frame is 8 ms. Experiments demonstrat that the algorithm is equally effective in real-world environments and has a faster detection time.