Research On Key Technology Of Target Grabbing Based On 3D Visual Robot

With the development of technology in recent years,more and more researchers are studying the field of robot grasping,making robot intelligent grasping a hot topic.At present,there have been significant breakthroughs and practical applications in the research of executing grab tasks on a single known target in structured environments.However,executing grab tasks on unknown objects in unconstrained and partially occluded complex environments is a challenge.Generally speaking,scholars mostly use depth cameras to obtain point cloud information in the scene,analyze the six-dimensional posture of the target in the scene,and finally send the pose information of the target in the scene to the robot after a series of coordinate transformations.The robot then performs the grab action to grab the target.In the whole process,obtaining the sixdimensional posture of the target is the most important link.Based on this point and combined with the analysis of the two-finger gripper,this dissertation mainly studies the grabbing posture of the target,not just the posture of the target.For target grabbing posture This dissertation designs an autonomous detection method for target grabbing posture based on deep learning technology,and completes the process of autonomous grabbing for unknown target in a chaotic and unconstrained environment.In order for the robot to better perform the grab,the target instance segmentation algorithm Mask R-CNN is also studied.The main research contents include:First of all,according to the characteristics of the robot’s grabbing environment and the two-finger grabbing posture,the grabbing posture of the target in the scene and the instance partition of the target in the scene are analyzed to provide theoretical support for the orderly grabbing of the robot.Then,the in-depth learning technology is used to segment the target instances in the scene in view of the background interference in the work scene and the situation of multi-target capture.By improving the feature extraction module in Mask R-CNN network,the features in two different situations of extracting target are extracted in two ways,then the two features are fused to get new features,and then the features are regressed.The improved network can detect and segment small targets better.Finally,an end-to-end method for capturing posture detection is designed by using the target mask obtained from the target instance split network and matching RGB-D information.Based on the Point Net++network and the structure characteristics of the two-finger gripper,a sevendegree-of-freedom capturing posture network is designed to obtain objects in three-dimensional scene directly.After training in the synthetic scene,the model can be directly applied to the real scene.The average success rate of grasping a single object is about 90%,and the average success rate of grasping a complex scene with multiple objects is about 75%.