In the field of powder packaging,manual grasping valve pocket is often used to complete the loading of powder,which is low in automation and harmful to human health.In order to solve this problem,a dynamic grasping system of valve pocket based on machine vision was designed.The camera calibration and image feature information extraction were completed by machine vision,and the pose data of the valve pocket in the robot coordinate system were obtained.In order to improve the positioning accuracy of the valve pocket in the moving state,Kalman filter was used to correct the position of the feature points.After that the data fitting was used to describe the change of the valve pockets pose,then the pose of the valve pocket was predicted at the moment of grasping.Finally,under the control of PLC,the industrial robot realized the dynamic grasping of the moving valve pocket on the conveyor belt.The system runs stably and the grasping error is less than 1 mm,which provides a feasible scheme for automatic powder packaging.