Scene Prediction and Manipulator Grasp Pose Estimation Based on YOLO-GraspNet

Aiming at the problems that the robotic arm grasps irregularly shaped objects in complex scenes, it cannot be accurately positioned, and the recognition speed is slow. This paper proposes a scene prediction-based location estimation algorithm YOLO-GraspNet. In the first stage of the method, the YOLOv5s model is used to identify and locate the grasping target, and then the depth information of the target area is marked. In the second stage, the GraspNet network is used to process only the data of the marked area, thereby improving the running speed of the model. This method uses GraspNet based on grasping pose prediction network, combined with YOLOv5s to estimate the pose of complex-shaped objects. It takes advantage of YOLOv5s' high detection accuracy, high grasping accuracy and other advantages. At the same time, it overcomes the problems of huge amount of input data, slow calculation speed and inability to accurately locate irregularly shaped objects in traditional methods. Finally, the prediction of grasping pose based on the proposed method and the grasping of the robotic arm are tested through experiments. The results show that the YOLO-GraspNet network can accurately and quickly predict the grasping pose of objects with complex shapes, and can be efficiently applied to robotic arm grasping tasks.