Research On Robot Grasping Based On Dilated Residual Convolutional Neural Network

At present,the research of robot grasping still faces challenges in multi-target grasping in scenarios such as small size,arbitrary direction,and dense distribution.Aiming at the problem that the robot grasping system has poor recognition and positioning effect in multi-target complex scenarios with multiple irregularities and large scale differences,a robot grasping system based on two-stage grasping detection network is designed.The first stage is the improved BCNet target segmentation network,and the second stage is the proposed robot grabbing detection network based on void residual convolutional neural network(DR-Conv Net).Based on the improved network structure,the performance and system capture effect of the two-stage network are experimentally verified on the designed physical platform.The main work content of this article includes:1)A robot handling system based on a two-stage grasping detection algorithm is designed.The visually guided robot grasping system is analyzed from three aspects:camera module,grasping detection module and grasping execution module,and the advantages and disadvantages of different series of robot grasping detection methods are compared,which shows the characteristics of the two-stage grasping detection method in this paper.2)In order to eliminate the interference of background information on grasping detection and improve the detection ability of small-sized objects,the BCNet instance segmentation network model is improved.The backbone network of the network is optimized,and after the Conv＿5 convolution of the Res Net50 residual network,hole convolution is used instead of the operation of averaging pooling.The improved network makes the output feature map have a better recognition effect on small-sized objects,and effectively reduces the interference of background information and irrelevant object information on the subsequent grasping detection network recognition.3)A robot grasping detection model based on void residual convolutional neural network is proposed.The input data is the image of the target segmentation network segmented only containing the target object,and the feature extraction is carried out through multi-layer convolution,and the Leaky-Relu activation function is added to enhance the ability of the network model to extract features and expression,and the combination of void convolution and residual connection is used to improve the detection accuracy of small target objects,which significantly improves the accuracy and detection speed of robot grasping detection.4)Finally,a physical grasping platform is built based on the Kinect depth camera and UR5 robotic arm,and a software control framework based on ROS operating system and Tensor Flow architecture is designed.The grasping performance of the robot handling system for single target and multi-target is tested,and the results show the practicability of the proposed method.