Abstract: Welding a lifting-eye plate onto underwater structures such as sunken ships to serve as anchoring points for underwater salvage can significantly improve salvage efficiency and reduce operational time. Under unmanned collaboration conditions, the welding robot needs to autonomously acquire the three-dimensional information of the lifting-eye plate and determine the welding position. Therefore, a three-dimensional reconstruction and welding guidance method for the lifting-eye plate based on instant neural graphics primitives (Instant-NGP) was proposed. Multi-view images of the lifting-eye plate were captured in an underwater hyperbaric dry chamber environment by a depth camera mounted on the welding robot, and the Instant-NGP algorithm was utilized to perform three-dimensional reconstruction within a large field of view. The three-dimensional point cloud of the lifting-eye plate was extracted, filtered, registered, coordinate-transformed, and mapped to the robot base coordinate system. The LO-RANSAC algorithm was used to perform planar segmentation on the point cloud on both sides of the weld groove; the weld seam centerline was obtained by calculating the intersection line of the fitted planes on both sides, and the trajectory starting point was determined by combining the point cloud boundary range of the groove region, thereby generating the initial welding guidance trajectory. Welding guidance and accuracy verification experiments were conducted through an experimental platform. The experiments indicate that the method can effectively accomplish the initial welding guidance; the average errors of the trajectory starting point in the x, y, and z directions are 1.17 mm, 1.09 mm, and 1.36 mm, respectively, meeting the requirements for the initial welding guidance of the underwater lifting-eye plate.