<p>High-fidelity digital scene generation is critical for robotic offline programming (OLP) and agile manufacturing. However, manufacturing and installation errors often cause spatial discrepancies between physical and virtual environments, degrading simulation fidelity and necessitating time-consuming manual calibration. To address this, this paper proposes a multi-instance point cloud registration method for multi-robot manufacturing scenes, aiming to enhance geometric consistency between the virtual and real worlds. First, an instance-focused transformer module is proposed to handle manufacturing scenes in which similar local geometric structures and indistinct boundaries are exhibited by robots of different models. This module effectively identifies instance boundaries and accurately models spatial correlations between local regions. Second, an instance hypothesis generation module is proposed to address the complexity of industrial object geometries. Coarse correspondences and a neighbor mask matrix are comprehensively integrated, thereby promoting a balanced distribution of correspondences under multi-instance conditions. Finally, an efficient instance filtering and pose estimation optimization algorithm is proposed, through which accurate estimation of multi-instance pose transformations is achieved while computational efficiency is maintained. Operationally, automating calibration reduces commissioning time and labor, supporting agile manufacturing. Experiments on Scan2CAD and Welding-Station datasets confirm that our method outperforms state-of-the-art techniques. Specifically, MR and MP improved by 12.15% and 17.79% on Scan2CAD, and by 16.95% and 24.15% on Welding-Station, respectively.</p>