High-precision 6D Pose Estimation For Manipulator Grasping

6D Pose estimation is the key to the interaction between the manipulator and the environment.In the real world,illumination change,lack of texture,data noise,and target occlusion will decrease robustness of the pose estimation algorithm,and the accuracy is difficult to be guaranteed,which leads to fail target capture for the manipulator.In addition,for high-efficiency application scenarios,such as assembly line of workpiece,bulk cargo sorting,etc.,the 6D pose estimation algorithm needs to meet the real-time requirement.Therefore,high performance and high efficiency 6D pose estimation algorithm has become a research focus in the field of robot.Traditional 6D pose estimation algorithms extract hand-craft features,perform template matching,key point matching or voting strategy based on hand-craft features,and indirectly solve the target pose.However,the hand-craft features are not robust to the complex changing environment and data noise.At the same time,methods based on template matching,key point matching and voting hypothesis are not efficient,because the search space is large in these methods.In recent years,data-driven pose estimation methods have been developed rapidly based on the powerful representation capability of deep learning.The algorithms based on deep learning has surpassed the traditional algorithms in performance with sufficient data,especially in the face of complex environment and data noise.Therefore,deep learning-based pose estimation algorithms are investigated in this thesis,and a high-performance and efficient full-convolution feature extraction network,named XYZNet,is designed for pose estimation.The XYZNet extracts dense point-wise features from RGB-D data.The features integrate RGB images and point cloud information to support downstream tasks to obtain more robust and accurate pose estimation results.In addition,thanks to the full convolution architecture,XYZnet is superior to other feature extraction networks in terms of computational efficiency,showing obvious advantages by taking account the trade-off between efficiency and precision.Further more,in industrial scene,it is necessary to estimate the 6D pose of symmetry objects.An object in different poses may have the same appearance,which means that one object may have multiple ground-truth poses,this one-to-many relationship leads to symmetry ambiguity in the training phase.The network will converge to the incorrect state with symmetry ambiguity.In order to solve the ambiguity problem,we design a symmetry-invariant pose distance metric,named average(maximum)grouped primitives distance(A(M)GPD).The proposed A(M)GPD loss can make the regression network converge to the correct state,i.e.,all minima in the A(M)GPD loss surface are mapped to the correct poses.