Grinding is a traditional yet critical process in manufacturing, and the integration of robotic systems can significantly enhance grinding performance while ensuring worker safety. However, current robotic grinding systems primarily rely on manual teaching or offline programming for trajectory generation. Moreover, these methods are often cumbersome and lack scalability. In this paper, a refined trajectory planning approach based on machine vision perception is introduced, aiming to address the issue of limited autonomy in traditional trajectory planning methods. Specifically, for full-surface grinding, a detection algorithm is developed to identify the polishing area on raised, welded and rough surface, enabling precise positioning of the target areas. Two toolpath patterns, a “bow” shape and a “Z” shape, are designed to cover the surface efficiently. For point-by-point polishing, an improved Laplacian matrix is introduced into the Laplacian-Based Contraction pipeline to improve the detection accuracy of key polishing points. The shortest path in the graph is then transferred to 3D space, where a point cloud eight-neighborhood graph is constructed. A minimum spanning tree algorithm is used to simplify the graph, facilitating trajectory sorting and endpoint positioning. The proposed trajectory planning algorithm offers high adaptability to process requirements, including adjustments to grinding head size, depth, angle, path density, and fitting accuracy. Finally, the effectiveness and accuracy of trajectory planning are verified by applying the proposed algorithm to the actual grinding robot.