Research On Deep Learning-based Grasp Detection Method For An Industrial Manipulator

As one of the most widely used industrial robots,the industrial manipulator can be applied in various skillful manufacturing operations.At present,the industrial manipulator has low-level intelligence.A traditional vision system of the industrial manipulator relies on the color image of the scene,which lack of the ability of disturbance rejection.The depth image of the scene collected by the RGB-D sensor is not affected by illumination changes,which could be used to improve the robustness of the application of the industrial manipulator.It is worth exploring the application of deep learning in the industrial manipulator,because of the excellent performance of depth learning methods in the computer vision.Taking the grasp detection problem of the industrial manipulator as the breakthrough point,this thesis designs a novel deep learning-based grasp detection method for the industrial manipulator with the RGB-D sensor,improving the intelligence level of the industrial manipulator.Firstly,the industrial manipulator with object classification capability will provide flexibility in the grasp task.This thesis presents a fast object classification method,which is suitable for the industrial manipulator.This method reduces the computing consuming of the standard convolution operation,and significantly improves the computing efficiency of the deep convolutional neural network.This method has obtained 94.2% accuracy rate in object classification task,which can process 137.42 pictures per second,which could be used in real time classification of the industrial manipulator.Secondly,to solve the image segmentation in the grasp detection problem,this paper studies the fully convolutional neural network-based image segmentation method and the depth image based segmentation method respectively.The full convolutional neural network-based image segmentation method achieve high accuracy rate but with complex training process.Applying the RGB-D image to the image segmentation,this thesis proposes the depth image based image segmentation method,which is easy to use with tolerable performance in the tests.This thesis proposes a novel method for grasp detection based on deep learning and proposal regions.First,according to the result of the image segmentation,a set of proposal grasp regions were selected by the grasp detection method.And then using the deep residual network to obtain the optimal grasp region,transforming the optimal grasp region to the target pose of the industrial manipulator.This grasp detection method has achieved the 86.49% accuracy rate.Finally,applying the three-dimensional simulation experiment in ROS robot research platform,this thesis designs a simple seven degree of freedom industrial manipulator with two finger gripper,and builds the grasp system to test the fast object classification method and the deep learning-based grasp detection method.The experiment results demonstrate that the effectiveness of the grasp detection method in this paper.