Robot Grasp Detection Based On Convolutional Neural Network

With the development of society and the growing problem of aging,service robots are increasingly used in many fields such as home services,warehousing and logistics,and property security.The most important way for robots to interact with the environment is grasping,so it is necessary to research the grasping problem of service robots.Unlike industrial robots,the environment faced by service robots is usually unstructured and complex.Traditional data-driven grasping methods rely heavily on the models of objects,while service robots often face with unknown objects,making it difficult to obtain accurate the objects' models.Therefore,in view of the above problems,this work has carried out research on the grasping problem of service robots in unstructured environments.It regards the grasping task as a detection problem and proposes a single object grasp detection network and a multi object grasp detection network,which achieve more reliable grasp detection effects.This work also builds a robot self-grasping simulation and experimental platform to further verify the performance of the proposed algorithm.First of all,this work is researched the basic theory of convolutional neural network and explores the basic principles and construction methods that constitute the various parts of the convolutional neural network.Then the different end-to-end object detection algorithms are researched.The performance of different algorithms is analyzed from the aspects of accuracy and real-time,which provides a basis for the construction and improvement of the subsequent grasp detection network model.Subsequently,this work is researched a single object grasp detection network based on convolutional neural network.Firstly,the grasping parameter model of the robot is determined and the darknet19 network for object detection is improved.The grasp detection network model that directly uses the regression method to output the grasping parameters is proposed,which can avoid the time-consuming sliding window algorithm and greatly accelerate the network forward prediction speed.Aiming at the problems of poor extraction ability,output the average value of the labels,model oscillation and only for single object,this work is researched a the multi objects grasp detection method.Firstly,the convolution layer of ResNet50 network is used as the feature extractor to enhance the feature extraction capability and the additional convolution layer is used to replace the fully connected layer as the network output,reducing the model parameters.Then it use the YOLO idea to divide the original image into multiple units,each unit corresponding to a set of grasp probability and grasp parameters,to avoid the network model to the average value of the labels.For the problem of unstable model training process,all output parameters are normalized.Finally,this work combines two methods of object detection and grasp detection to achieve multi objects grasp detection.Finally,a simulation platform and a experimental platform for robot self grasping are built,which can verify the feasibility of the proposed algorithms and evaluate the performance of the two algorithms.