Design Of Robot Dexterous Hand Recognition And Grasping Based On Deep Learning

Dexterous hand is an important research field of robotics.Based on dexterous platform,it can be used to develop new control algorithm and manufacture dexterous operating system.The classification of grasping action of dexterous hand is an important link in the grasp planning process of dexterous robot hand Dexterous handhas more freedom and different grasping modes.Based on the deep learning algorithm,this thesis extract s the features of the captured object from the image,finds out the mapping relationship between the object and the grasping action,and carries out the grasping feedback through the tactile sensor,which is of great significance for the robot to realize fine operation and human-computer interaction.This paper is ba sed on chl-ds-01-a industrial robot PCB special-shaped plug-in workstation platform,which covers a number of engineering fields,including the physical design of mechanical system,the design of printed circuit board,the control theory of control system,and the realization of control software.In order to verify the accurate judgment of object grasping type in complex environment,combined with,image recognition and deep learning technology,and drawing on the functional characteristics of human grasping objects,a dexterous hand grasping classifier which can predict the grasping types of different objects is designed and implem ented,and its performance is verified in the simulation environm ent.The main work of this paper is as follows,(1)Path data collection of dexterous hand platform grasping objects.In this paper,based on chl-ds-01-a industrial robot platform,a large number of grabs are carried out for different objects in different positions to collect the grasping path parameters.It mainly includes the data of the process of moving from the starting position of the dexterous hand to the position of the object and returning to the starting position after grasping the object.The amount of path data collection is very important for the network model to automatically generate the path,and it can make the dexterous hand learn the way and habit of human grasping objects.(2)MobileNetV3 network model is improved and trained.Aiming at the problem that dexterous hand is difficult to locate accurately,especially when grasping objects in complex environment,the depth convolution algorithm is used to separate the grasped object from the complex environment.In order to improve the speed of using deep network model in mobile terminal,this paper improves the mobile netv3 network model combined with the characteristics of food industry Through the experimental comparison,the image of the object recognized by the improved network model is more accurate.(3)The bi-directional recurrent neural network is used to simplify the complex mechanical motion control,and the BRNN model is trained by self collecting data.Aiming at the problem of dexterous mobile phone's motion planning,a bi-directional recurrent neural network model is designed.The model can grasp any position and make the dexterous hand have certain generalization ability.