Research On Dexterous Hand Grasp Classifier Based On Improved ShuffleNet Network In Complex Environment

With the extensive application of robot technology in complex realistic scenes,people hope to complete some dangerous,repetitive and difficult tasks with the help of robot dexterous hands,which puts forward the requirements of flexibility,intelligence and ease of control.The grasp taxonomy is one of the key problems in grasp planning of dexterous hands.However,because the dexterous hand has more degrees of freedom and the diversity of grasping modes exists,it is difficult to predict the suitable grasp type for different objects.In this paper,we use deep learning technology to extract the characteristics of the grasped object from the image,and find out the mapping relationship from the image to the dexterous grasp.In order to accurately judge the grasping types of objects in complex environments,a dexterous hand grasping classifier is designed and implemented in this paper.By referring to the characteristics of human grasping objects,this dexterous grasp classifier is designed with image segmentation and deep learning techniques to predict the grasp types of different objects.The performance of the classifier is verified in the simulation environment.The main contents of this paper are as follows:(1)To solve the problem that the neural network is difficult to accurately extract the object features when the dexterous hands grasp the objects in the complex environment,Grab Cut segmentation algorithm combined with superpixel SLIC is used to extract the objects from the complex environment.In order to improve the segmentation effect of grasped object image,the minkowski distance is used in SLIC algorithm instead of Euclidean distance to describe the color similarity measure between two points.At the same time,color weight and spatial weight are introduced into the distance measurement formula of SLIC to improve the effect of superpixel.After experimental comparison,the image of the grasped object is more accurate with the improved segmentation algorithm.(2)A classification model of dexterous grasp based on improved Shuffle Net network is designed to predict the optimal grasp type of dexterous hand.The Shuffle Net was built and trained under Paddle Paddle,an open source deep learning framework in China.The network was improved with the SE-Res Net module to optimize the characteristics of network extraction,effectively improve the accuracy of the model.The focus loss function is used to replace the traditional cross entropy loss function to optimize the model.Experimental analysis shows that the improved Shuffle Net network model with focus loss has the best classification performance,and the model size is only 11.3M,which is easy to deploy in small and medium embedded systems.(3)At present,real dexterous hand is limited by cost and experimental environment,so a simulation platform for dexterous hand grasping is built based on ROS system.Public data set is used to verify the designed dexterous grasp classifier,and the results show that the object picture after background removal can effectively reduce the influence of complex background on the recognition accuracy.The results of the simulation test on the simulation platform show that the grasping classifier can be used to get the dexterous grasping type which conforms to the habit of manual grasping.The simulation results verify the usability of the dexterous hand grasping classifier.