| Robot grasping operation is one of the most basic ways for robots to interact with the outside world.Traditional robot grasping methods have low success rate and poor generality in the face of today’s more common unstructured scenarios.Deep learning and vision technologies enable robots with perception and decision-making capabilities to accomplish grasping tasks in various complex environments.In this thesis,we focus on the core task of grasping detection,which is the core task of intelligent robots to complete grasping operations,and construct a grasping detection model based on convolutional neural network for different grasping scenarios,and build a robot grasping experiment platform to verify the model.The main research work of this paper includes:(1)To address the problem that the accuracy of current mainstream 2D grasping detection methods is easily affected by the grasping background noise,a planar 2D grasping detection model based on hybrid network architecture is carried out.Firstly,the target detection frame of the object to be grasped is obtained based on the improved YOLO V5 algorithm,then the lightweight convolutional neural network Mobile Net V3 is used as the backbone feature extraction network to design the grasp detection network to generate the grasp detection frame of the object to be grasped,and finally an IOU region evaluation method is designed to filter the optimal grasp detection frame with the target detection frame of the object,to be grasped as the background.thereby calculating the final grasping pose and avoiding the influence of grasping background noise on the grasping detection results;(2)To meet the requirements of performing grasping on objects from different angles,a 6-Dof grasping detection model based on Point Net and point cloud attention mechanism is constructed.Firstly,the point cloud attention mechanism is designed and embedded into the Point Net point cloud classification network,and the grasping pose quality evaluation network is designed,then the grasping dataset is generated based on the YCB dataset to train and test the model;finally,the best model structure is obtained by comparison experiments;(3)The proposed model is validated by building a robot grasping experiment platform,and the trained model is deployed on the robot platform for grasping experiments respectively,and the rationality of the proposed model is verified by setting different grasping scenarios and evaluation indexes.The experimental comparison shows that the planar 2D grasping detection model based on hybrid network architecture has a high success rate for single-target grasping,and for multi-target grasping,the 6-Dof grasping detection model based on Point Net and point cloud attention mechanism has a better performance in terms of both success rate and completion rate. |
