Research On Robotic Arm Grasping Method Based On Deep Convolutional Network

With the development of science and technology,the application of robotic arms in production and life is becoming more and more extensive.However,in actual operation scenarios,the application of robotic arms faces severe challenges such as complex and changeable environments and unknown targets.Especially in an unstructured environment,object grasping pose detection is still a difficult task.Grasp pose detection for transparent objects is made more difficult due to inaccurate depth information that can occur during imaging with commercial depth cameras.This thesis conducts research on the pose detection of the robot arm grasp and the depth map repair for transparent objects.The main research contents are as follows:(1)A depthwise separable convolution based grasp pose detection network DSCGrasp Net is proposed.Design a lightweight backbone network based on depth-separable convolution to extract multi-level features of the image.After the features are fused and upsampled by the CSP(Cross Stage Partial)module,the grasping pose is predicted by the grasping generation head.In the network training phase,the soft coding quality map based on Gaussian distribution is used to highlight the importance of the center position of the grasping area,and a position-enhanced P-Huber loss function is designed to make the learning of the network more focused on the target grasping area.It achieved 98.8% image segmentation detection accuracy and 98.3% object segmentation detection accuracy on the Cornell dataset,and 95.3% detection accuracy on the Jacquard dataset.At the same time,the model parameters are 0.698 M.Inference a RGB-D image takes 14 ms.(2)A lightweight transparent object depth map repairing network E2 EDepth Net is designed.In order to increase the information flow path inside the network and improve the learning ability of the network,the encoder-decoder architecture is designed based on the dense skip connection structure.The feature extraction unit in the network is designed based on a single aggregation module that fuses e SE(effective Squeeze-Excite)attention.A multilayer feature fusion module is designed in the encoder,and dense up-sampling convolution is used in the decoder design to obtain More refined depth restoration results.The experimental results on the Trans CG transparent object depth repair dataset show that the RMSE of E2 EDepth Net is 0.014,the accuracy rate reaches 91.2% when the precision threshold is 1.05,and the time of inferring a image is 12.2ms.At the same time,in order to verify the contribution of the deep map inpainting network in the task of grasping and detecting transparent objects,a dataset of grasping transparent objects was constructed.The experimental results on this dataset show that E2 EDepth Net can effectively improve the accuracy of the grasp detection network.(3)The robotic arm grasping system was built,the hardware equipment was calibrated,and the software platform was built based on ROS.Finally,the grasping experiment was carried out based on the grasping detection network and depth map repair network in this thesis.100 grasping attempts were made in the non-transparent object grasping task and achieved an average grasping success rate of 84%,and in the transparent object grasping task 80 grasping attempts were performed and an 81.25% success rate was achieved,Verified the feasibility of the two networks in this thesis in the robot arm grasping task.