Object Recognition And Grasping Of Manipulator Based On Shape Priori

In the process of robot grasping,there are three key steps: target detection,pose estimation and manipulator path planning.Combined with the research status at home and abroad,this paper mainly proposes a 6D pose estimation method for manipulator recognition and grasping based on shape prior,which can estimate the 6D pose of object instances from RGB-D images and grasp objects in real time.The work content and achievements of the thesis mainly include:(1)A complete mechanical arm grasping software and hardware system platform is designed.Among them,the system hardware selection,system calibration,object instance segmentation,6D pose estimation algorithm,robot kinematics modeling and path planning are studied.In addition,an upper computer software is developed to control the system,so as to verify the proposed 6D pose estimation method and experiment in real scenes.(2)Hardware selection and system calibration.After selecting the hardware required by the system,the calibration algorithm of tool center point of the robot is introduced.Then,after the camera calibration,the registration principle of color camera and depth camera in Real Sense is carried out.The eye-in-hand calibration theory based on HALCON method was elaborated,and the relatively accurate rotation and translation transformation relationship between the manipulator tool coordinate system and the camera coordinate system was obtained.(3)To meet the real-time requirements of the system,a deep learning algorithm combining SOLOv2 instance segmentation and shape priority-based deformation network is proposed to perform class-level 6D pose estimation of the object,which can identify the class of the instance object and the pose in the camera coordinate system.Aiming at the problem of insufficient characterization ability of local feature extraction,the Point Net++ network was proposed as the main frame of point cloud feature extraction in multi-mode feature network,and the feature images sampled step by step covered both local and global features of point cloud,so as to complete point-to-point recovery of point cloud features.After verification,the improved 6D pose estimation algorithm improves the accuracy of the pose estimation of the target object.(4)In the experimental scene and environment set up in this paper,aiming at the problem that the manipulator may encounter obstacles in the process of removement,the kinematic simulation model of Yaskawa MH12 manipulator is established in MATLAB software,and the improved RRT algorithm is used to effectively avoid collisions;Aiming at the practical application problem that the robot arm may collide with the object in the process of grasping multi-object,an effective target grasping strategy is proposed,which can grasp the object accurately and quickly.The success rate of grasping is more than 90%.