<p>Robot welding is the main form of automatic welding system. Although the application of machine vision, intelligent teaching, offline programming, and other technologies had greatly improved the manufacturing flexibility of the welding system, the existing robot welding was not fully adapted to the manufacture of products with small batch and customized needs, and the manufacturing flexibility needed to be further improved. To solve this problem, a flexible stacking welding manufacturing method known as direct stacking welding manufacturing (DSWM) has been proposed as a cost-effective method for constructing common metal structures. DSWM fabricates metal structures by stacking welding of modular pieces one-by-one bottom-to-top. A multi-robot system with vision guidance has been developed to conduct the DSWM process. This system sequentially combines the operations of cutting, grasping, assembling, and welding the modular pieces. 3D visions were employed to recognize modular pieces and intermediate states of the structure, thereby providing guidance for the grasping, assembling, and welding procedures. The proposed methodology and the developed system were validated through two manufacturing cases. The case studies demonstrated that the designed algorithms were capable of accurately recognizing and localizing the modular pieces. The manufacturing accuracy achieved was found to be compliant with relevant quality standards. In terms of production time, the DSWM process was markedly shorter than that of wire arc additive manufacturing (WAAM). In terms of production cost, it was estimated that the DSWM process incurred approximately one-fourth the cost of WAAM. For a detailed visual demonstration of the project, please refer to the accompanying video available at (<a href="https://youtu.be/gN-wKduHa3s?si=0xLW0Nd6sX8sm1r2">https://youtu.be/gN-wKduHa3s?si=0xLW0Nd6sX8sm1r2</a>).</p>