Research On Robot Grasp Pose Detection System Based On Deep Learning

With the proposal of the "Made in China 2025" strategy,robots are used more and more widely in various industries,especially in the field of industrial production.The application of robots to achieve automated production can not only effectively solve the problem of labor shortage,but also improve production efficiency and product quality.Robot grasping is an important function of industrial robots,the traditional way is to make the robot repeat the specified grasping actions according to the established program by teaching programming,which can only grasp the single type object with fixed position.In order to adapt to the rapid and ever-changing characteristics of modern industry and to meet the increasing complexity requirements,robots must not only have the ability to perform repetitive tasks for a long time,but also have intelligent characteristics.Aiming at the visual detection scene of the industrial robot using the two-finger gripper for grasping of scattered multi-objects,the application of convolutional neural network in multi-object grasped planning scene is studied in detail.This paper proposes a method for the detection of the target's grasp pose and verifies the effectiveness of this method through the actual detection experiments.This paper mainly completed the following works:1.Firstly,carry out the overall scheme design of the grasp pose detection system and determine the algorithm processing flow.The advantages and disadvantages of the two different hardware platform design are analyzed.According to the specific selection of the vision sensor and the robot,select Eye-to-Hand platform as the hardware platform in this paper.2.For the problem that the color image acquired by the visual sensor does not match the depth image and the edge information is missing after the image registration preprocess,the process of image registration and repair is carried out.The object detection network model based on SSD framework is built and trained in the way of transfer learning.Propose a graph representation method of grasp pose,represent the grasp pose in the form of a vector which provide the feature for the calculation of subsequent algorithm.Then the candidate grasp poses are sampled base on the contour features of the object,the sampling result satisfying the conditions of distance,direction and force closure.3.Improve the structure of the classical residual block and propose two new residual block structures as the convolution module of the grasp pose classification network(GPCN).The optimized network structure is determined through comparative analysis,compare with the network only with classical residual block,the GPCN network's accuracy of verification set increased by 1.42% to 98.80%.In order to solve the problem that the output of grasp pose may deviate from the center of weight,a grasp pose optimization algorithm is proposed,corresponding optimization calculations are carried out for different kinds of items.The experimental results show the effectiveness of this optimization algorithm.4.Finally,in order to verify the performance of the grasp pose detection system,build a robot experiment platform,complete the camera calibration,hand-eye calibration and determine the mapping relationship between the grasp pose and the robot grasp parameters,design the pose detection experiment and robot grasp experiment,then the various objects of different shapes are used as test samples for pose detection and grasping experiments.The accuracy of effective grasp pose detection is 91%,and the success rate of grasping is 84%,this result shows that the proposed pose detection system has certain feasibility and effectiveness.