Research On Robot Detection And Grasping Strategy Based On Deep Learning

Intelligent grasping is an important function of intelligent robots.In actual life scenarios,the types and grasping points are frequently changed.Participating in the setting of grasping parameters,which will lead to seriously lack of grasping intelligence.In this paper,aiming at the two intelligent grasping tasks of regular objects and irregular objects,the grasping strategies are studied for the respective tasks.And under different tasks the grasping parameters are predicted based on the deep learning method,finally finish smartly grasp regular objects and irregular objects.The main research contents of this paper are as follows:1.Studying the grasping strategy of regular objects,by the second-class angle to grasp the center point of the bounding box of the target as the grasping strategy of regular objects.Based on the analysis of the current popular detection’s networks,the YOLOv5 is used to generate grasping parameters,and the model’s accuracy is 96% after training.2.In usual life,not only have regular objects but also exist a large number of irregular objects,using the best grasping point of the target and the rotation angle of the object and the level to grasp to grasp as the grasping strategy of irregular objects,And building a special grasping detection’s network to generate grasping parameters.The grasping detection’s network is constructed with the Res Net-50 backbone network of Inception ideas.And for the problem that the proportion of positive and negative samples in the data set is quite different during network’s training,Focal Loss is used as the loss function of the network’s model during update weight’s stage.3.Discussing the effect among the data sets of RGB,D and RGB-D channels in training the grasping detection’s network.The data sets of three different channels are trained under the same conditions and the results are analyzed,finally the average accuracy of the grasping detection’s network trained by the RGB-D channel data set is 92.6% after 300 iterations of training.4.Clearing the mapping relationship from target’s coordinate system to the robot’s base coordinate system,so obtain the spatial coordinates of the target’s grasping parameters in the robot’s base coordinate system.After real scene’s verification,the successful rate of intelligent grasping of the grasping strategy for regular objects is 90.7%,and the successful rate of intelligent grasping of the grasping strategy for irregular objects is 84%.It lays the foundation for finishing a fully intelligent robot grasping system.