Research On Robot Grasping Simulation Training Technology Based On Deep Learning

Intelligent robot grasping plays an important role in robot intelligence.Due to the diversity of shapes and scales of objects and the influence of scenes,it is quite difficult to solve the task with accurate mathematical formulas.Most of the previous researches were based on computer vision,machine learning,and other related technologies.Although there are some achievements,the degree of intelligence is still in low stage.Since 2012,deep learning has gradually emerged because its good feature extraction performance and has been applied in various fields,such as medical images,automatic driving,data analysis and so on.In recent years,foreign scholars have begun to apply this technology to robot grasping,and achieved some results.In order to get good results,deep learning requires a large amount of data to train.Although some scholars have pointed out that pre training can reduce the size of dataset for training,the amount of data is still large for robot,and the intensive actions may cause a robot fault,so in recent years some scholars have begun to try to perform a large number of experiments in the virtual environment to collect data.The subject studied the calibration principle of the camera and completed the handeye calibration of robot.In order to complete the real experiment,the Kinect camera is used to provide the color depth picture.Before executing experiment,the color camera and the infrared camera and their relative position were calibrated.We studied two sampling based motion planning algorithms,PRM and RRT,to guide subsequent robot motion planning.We used nine point calibration method to calibrate the relative position between the manipulator and Kinect,and then combined with the depth map,the pixel points in the color map can be converted to the robot based coordinate system to guide the robot to perform grasping.Built a virtual environment for collecting robot grasping data based on VREP.The virtual environment includes depth camera,a robotic arm,and an object.This topic uses the method of random sampling in the object pixel set to generate the grasping center point.For a single grasping center point,a grasping direction generation algorithm based on depth map is adopted.This algorithm selects those edge points whose normal line pass through the grasp center point.These candidate grasping direction was reduced to several by fused with other points belong to a certain cluster.Finally,collect the grasping data of the object by executing a large number of grasp through an external simulation control client.Several kinds of object detection networks were studied,On this basis,a network of grasp detection was written.In this project,we have studied several object detection networks including Faster-RCNN and YOLO,which can be used to identify objects in images.Based on the SSD300,the feature map of middle layer is used to predict the feasible grasping pose.In the training process,we use the idea of SSD to get the coordinates of four points of groudtruth grasping box by direct regression.The final grasping pose is generated by evaluating the Io U and non maximum suppression.The modified network shows a good performance on multi-task.In the end,we provide an experiment in reality,The above network predicts the target grasp pose in the image provided by Kinect,and then converts it to the robot coordinate space.Then,grasping is executed by the AUBO-i5 robot and the ROBOTIQ-140 gripper.The experiment verified the feasibility of virtual data training neural network for real grasping...