Grasping Detection Of Dual-manipulators Based On Deep Learning

Grasping detection is the research focus of robot intelligence,the purpose is to allow robots to make judgments on the external environment like humans,and grasp a variety of objects.However,the steps of the traditional grasping pose detection method are cumbersome,and it is difficult to realize intelligent grasping of unknown objects.Moreover,at present,a single grasping method is often used,the grasping effect is not satisfactory in the face of objects with large shape differences.Therefore,the establishment of a robot grasping detection model that includes multiple grasping methods is of great significance for realizing effective grasping detection of unknown objects with large shape differences.Aiming at the above problems,a dual-manipulators grasping detection method based on deep learning is proposed.To study the grasping detection problem of the dual-manipulators with two kinds of grasping methods,and solve the problem of poor universality of the single grasping method due to the large shape difference of objects.A grasping quality detection model based on convolutional neural network is proposed.The convolutional neural network is trained with the dataset generated by the grasping analysis model,to obtain the dual-manipulators grasping scheme with the highest grasping efficiency.To realize the grasping detection of the robot for unknown objects,and improve the reliability and efficiency of the grasping of the manipulator.The main research work is as follows:(1)Synthesize the dataset based on the grasping analysis model.The method of domain randomization is used to construct the grasping analysis model of two grasping methods,and combine the model obtained from 3Dnet and Kit to construct a dataset consisting of depth images,grasping poses and reward tags.Use this dataset to train the grasping detection model,so that it has a certain degree of robustness to perception,gripe and physical properties.(2)Grasping detection model of dual-manipulators grasping based on deep learning.Take the Markov decision process as the theoretical framework,and the convolutional neural network is used to parameterize it.Taking the grasping quality as the metric,firstly,the grasping pose detection models of the two grasping methods are optimized to obtain the best grasping pose,and then the grasping quality is compared to determine the final grasping method and grasping pose of the dual-manipulators grasping detection model.(3)Grasping quality detection model based on convolutional neural network.Based on the cross-entropy convolution neural network and the full convolution neural network,two kinds of grasping quality detection models are constructed respectively.A group of samples are used to carry out the grasping detection test,the more efficient grasping detection network of two kinds of grasping methods are obtained,which constitute the dual-manipulators grasping detection model.The grasping experiment was carried out on the Dobot robot.The experimental results show that the dual-manipulators grasping detection method based on this study has achieved a success rate of 98.0% and a task completion rate of 90.0%,which verifies the effectiveness of the dual-manipulators grasping detection method based on deep learning.