Abstract: To solve the problem that it is difficult to distinguish multiple adjacent objects and the accuracy of high-dimensional pose learning is poor, a pose estimation method combining on semantic instance reconstruction is proposed. The semantic instance reconstruction branch is added to complete implicit 3D reconstruction of the foreground, and the center coordinate of each foreground point belongs to the instance is predicted by the voting method to distinguish adjacent objects. A pose dimensionality reduction learning method is proposed. Two orthogonal unit vectors are used to decompose the three- dimensional rotation matrix to improve the accuracy of pose learning. A semantic instance reconstruction grasping network (SIRGN) is proposed, and the training is completed on VGN simulation grasping dataset. The experimental results show that the grasping success rate of SIRGN in Packed and Pile environment is 89.5% and 78.1% respectively, and it has good applicability in real environment.

Key words: grasp pose estimation, implicit 3D reconstruction, voting, dimensionality reduction, rotation matrix