Contact-graph-guided digital twin construction and incremental update for discrete manufacturing cells

Workpiece states change frequently on discrete manufacturing lines, requiring digital twins to maintain continuous consistency between physical and virtual scenes. Automatically constructing and maintaining instance-level digital twins with low-cost RGB cameras is a key route for reducing deployment cost, yet existing one-shot reconstruction and object-wise synchronization mechanisms have difficulty handling local changes such as coupled multi-object motion and movement of underlying support objects. This paper proposes a digital twin construction and incremental update framework for fixed-view monocular RGB sequences, using a contact graph as a unified intermediate representation. In the initial stage, the framework combines open-vocabulary perception, CAD retrieval, pose estimation, depth scale calibration based on CAD geometric constraints, and contact-graph-based hierarchical layout optimization to build a physically consistent instance-level scene. In the update stage, the contact graph restricts processing to the affected subgraph, where motion-anchor pose estimation, rigid-motion propagation, and local relation maintenance avoid full-scene reconstruction. Experiments show that the proposed method achieves higher pose and geometric accuracy than Diorama in initial construction and substantially reduces computation time in incremental updates while maintaining geometric quality close to full reconstruction.