Abstract: Aiming at the problems of low efficiency and poor accuracy of teaching programming in complex welds with variable magnetic pole length and gap of large hydropower generators, a technology of welding seam identification and robot track generation without teaching was developed based on grating visual sensing. A grating sensor installed at the end of the robot was used to obtain the point cloud of the magnetic pole weld at different positions. A point cloud registration algorithm combining the robot tool pose transformation matrix and iterative closest point algorithm (ICP) was proposed to obtain the complete point cloud data of the large size magnetic pole weld. Based on random sampling consistency (RANSAC), a weld recognition algorithm was developed to realize the automatic generation of robot welding trajectories. The results show that the algorithm can identify a variety of complex magnetic pole welds with high recognition rate and strong anti-interference ability, and the average recognition error is with in ± 0.4 mm, which meets the welding requirements.