非结构化环境下的机械臂抓取位姿检测系统研究OA

In order to enhance the accuracy and robustness of robotic grasp pose detection in unstructured environments,a method of grasping pose detection integrating segment anything model(SAM)image segmentation with the GraspNet network is proposed.SAM model is used to segment target masks to filter out the interference from non-target areas,and combined with the GraspNet network to predict the grasping pose.The high-robustness scheme is screened based on force closure theory.The calibrated RealSense camera is used to obtain the RGB-D data and intrinsic parameters,so as to establish coordinate transformation relationship between the camera and robot base frames.The robotic arm is controlled by means of coordinate transformation and inverse kinematics for the grasping.The Sawyer 7-axic robotic arm and D435i camera are used to establish the verification platform for the experiments.The results show that the target detection rate and average grasp success rate of the improved algorithm can reach 91.7%and 81.6%,respectively,which are 7.5%and 8.3%higher than those of the original GraspNet algorithm and can significantly reduce the incorrect poses caused by interference from non-target objects.The proposed method can improve the grasping accuracy and generalization ability in unstructured environments by means of the collaborative optimization of GraspNet and zero-shot segmentation,providing a feasible technical solution for industrial automation and robot grasping tasks.