Abstract:Aiming at the problems of low programming efficiency, low intelligence and weak human-computer interaction performance of industrial robots, a vision-based industrial robot assembly demonstration and teaching system was proposed. The system included a target detection and center point positioning module, an assembly action classification and recognition module, and a robot action execution module. In the target detection and center point positioning module, a method of target object center point positioning and robot grasping was proposed. The instance segmentation algorithm was used to identify the object category, and the 3D pose information of the object was calculated by mask averaging and coordinate transformation. In the assembly action classification and recognition module, an action classification and recognition model based on a deep learning network was established. The input of the model was the assembly action video frame, and the output was the action classification label. Finally, the robot assembly action was planned according to the object category, object 3D pose and action classification label and other information of the robot action execution module, and the robot was controlled to perform the assembly task.Taking the shaft-hole assembly as an example,the effectiveness of the above method is verified, the imitation programming of robot assembly is realized based on visual demonstration.It has certain reference value for the research of robot demonstration and teaching.