Robot Grasping Detection Using Deep Learning

Since 1960s,the development of computer science,computer software and hard-ware,technology of control and communication has vastly changed the production mode in manufacture and the life of human beings.As an intelligent apparatus,robot arms imitate the operation mode of human beings and play an important role at home and manufacture.Robot grasping is a common task of robot arms,accurate and robust robot grasping is the guarantee of robot intelligence.The accuracy of robot grasping is based on the precision of the mechanics design and the superioty of control algorithm.The robustness of robot grasping is related to the analysis method of the sensor data.The pictures of scene include the most com-prehensive information,but it is the hardest for analysis.This article presents methods for robot grasping detection by analyzing the picture of objects,helping to finish robot grasping tasks.The first work in this article presents an end-to-end robot grasping detection method using convolutional neural networks and traditional vision algorithms.Firstly,give a robot grasping point detection using convolutional neural networks.Secondly,present a main direction detection method to detect robot grasping orientation and distance be-tween robot arm's parallel fingers.Finally,gives the complete robot grasping detection.The second work in this article presents an object detection method using convo-lutional neurak networks.The main aim of object detection is detecting the object in pictures,giving a rectangle to locate the position of obj ects and classifying these objects at the same time.According to common sense of life,we divide the grasping dataset into 20 categories.We give their labels fitting object detection refer to Pascal VOC dataset.Then,we train a convolutional neural network using pictures in the grasping dataset and their object detection label to detect objects in the pictures.Object detection helps robot grasping detection algorithm to handle the pictures including more than one object.Also,Object detection helps to shrink the scope of vision algorithm analysis and give a rough semantic label for every object in the picture.The third work in this article presents a robot grasping detection based on object detection.Compared to the first robot grasping detection method based on end-to-end,regional analysis can give more accurate grasping detection.But traditional grasping detection methods based on regional analysis gives a bad instantaneity because of its large scope search.We use object detection shrinking the search scope to cut down the detection time.Also,we use spatial pyramid pooling sharing the convolution layer between object detection network and feature extraction network to reduce computation time of convolution layers.Besides,we give a candidate rectangle extraction method to reduce the number of candidate rectangles,which also reduce the computation time.Finally,our grasping detection method can give higher detection precision,satisfy the instantaneity requirement in the meantime.We do some research in the field of robot grasping detection,give robot grasp-ing detection methods for common objects,make some contribution to robot grasping detection.