Research On Robot Intelligent Grasping Method Based On Meta-learning

With the rapid development of robot technology and artificial intelligence technology,service robots are widely used in assisted medical care,home services,smart factories and other fields,which greatly expands the application scope and application prospects of robots.Among them,the robot's grabbing operation is the main way for the service robot to interact with the complex external environment.Traditional deep learning-based object recognition and crawling methods rely on a large number of data samples and time training,and service robots usually face unknown objects,and it is difficult to obtain a large number of data samples of target objects.Aiming at the above problems,this paper conducts research on the robot crawling problem with a small number of samples,and builds a simulation and experiment platform for robot crawling to verify the effectiveness of the proposed algorithm.In this paper,on the basis of robot based on visual recognition grab,a fast recognition grab method based on small sample learning is proposed.Using deep vision sensor as the main perception tool of the robot,build a hand-eye calibration system for the robot and the vision sensor,use a small amount of sample information to train the object recognition model,on the basis of object recognition,explore the optimal grasping area of the unknown model object,and finally combine the grasping Perceive grabbing unknown objects.The main research results are as follows:(1)The establishment of the hand-eye calibration system of robot and visual sensor is the basis of robot grasping operation.The hand-eye calibration of robot and visual sensor is completed by means of the Kinect visual sensor mounted on the outside of the robot and the Aruco Marker tag,so as to facilitate the subsequent robot's recognition and grasping.(2)Using Model-Agnostic Meta-Learning classified recognition algorithm for few shot learning,in Omniglot and Mini Image Net respectively on two data sets for training.On this basis,the MAML algorithm was optimized by adding memory penalty terms,and the accuracy of different methods was compared.(3)Build a two-step cascade system of object recognition with few shot learning and object grasping pose detection.Use the few shot recognition algorithm based on MAML to complete object recognition,transfer the recognized image to the grasping pose detection network based on deep learning,and finally send the grasping pose of object to the robot through the upper computer.(4)Combine the few shot object identification and detection method proposed in this paper to design the robot's intelligent gripping system,so that the service robot can quickly identify and detect when facing unknown objects,and use the current change of the robot finger to design the gripping perception module,so that the robot Apply different forces when grasping objects of different sizes to achieve intelligent grasping of service robots.