The traditional robot grinding of large heterogeneous workpieces is processed according to the unified teaching path of a single workpiece model. Due to the dimensional deviation between the actual workpiece and the theoretical model, there are problems such as grinding empty stroke or excessive removal. In this paper, the method of laser scanning the workpiece with moving line by robot is proposed to obtain the point cloud daa from different perspectives. The combination of SAC-IA rough registration based on FPFH and ICP fine registration is used for point cloud stitching. The KD-tree corresponding point distance algorithm is used to delete the redundant points in the overlapping area, and then the 3D model reconstruction is completed by the greedy projection algorithm based on Delaunay triangle growth principle, which can be provided to the robot for adaptive grinding. The blade with length of 600 mm and width of 200 mm is used for the experiment. Compared with ICP and NDT + ICP, the registration accuracy of this method is improved by 71.90% and 31.53% respectively. The size deviation between the reconstructed model and the actual workpiece is less than 0.15 mm.