Research On 6D Pose Estimation And Grasping System For Robot

Nowadays,robot grasping technology has a wide range of applications in production and life,such as courier sorting,parts processing,garbage sorting,etc.With the development of artificial intelligence,deep learning positional estimation technology can be combined with robot grasping to make robot grasping technology more intelligent and improve robot grasping accuracy.The deep learning-based method reduces the algorithm cost compared with the traditional method and has high practical value.In this thesis we investigate the deep learning-based object 6D pose estimation technology,and use computer vision technology to make the robot have the scene perception ability,build a robot grasping experimental platform,and further verify the feasibility of the pose estimation technology in the real environment.The 6D pose contains 3D position and 3D pose,and a new solution is proposed for the problems of high cost,low accuracy of pose estimation,and poor robustness in extreme scenes of traditional robot grasping algorithms.The main research work is as follows:In terms of 6D pose estimation methods for objects,this thesis proposes an attention mechanism-based pose estimation method.The method fuses the attention mechanism with ResNet as the backbone network,and uses a neural network to regress object poses by establishing a dense 2D-3D correspondence between 2D image planes and 3D object models.The addition of the lightweight ECA attention module enhances the effective features and removes irrelevant information to help the network extract features efficiently.To evaluate the performance of the method,it is trained on three different publicly available datasets in the field of pose estimation,and compared and analyzed with similar recent pose estimation methods.The experiments show that the method can estimate the poses of objects in complex situations such as occlusion stacking,and performs well on three public datasets of pose-estimation.In terms of the robot grasping system,this thesis verifies the feasibility and practicality of the grasping system by creating a homemade grasping dataset,building a robot grasping experimental platform,and completing grasping experiments in different scenarios.Firstly,in response to the difficulty of manually annotating datasets,synthetic datasets are proposed to train the pose estimation network.By using the open-source tool Blender Proc to produce synthetic datasets,the efficiency of dataset production is improved.Then the camera calibration experiment and the hand-eye calibration experiment are used to obtain the camera parameters and the conversion relationship between the camera and the robot,respectively.Next,the grasping points were designed on the surface of the object to accommodate the grasping posture of the two-finger gripper to obtain the coordinates of the object grasping position and the grasping angle applicable to the robot.Finally,through a series of experiments in real scenes,the average object grasping success rate of the robot grasping system is 95%,which verifies the feasibility of the pose estimation method applied to robot grasping and proves that the grasping system can achieve the expected results.