Proceedings of the XMO Industrial Seminar 2026: Excellence in Manufacturing and Operations
Keywords
Immersive digital twin; Multimodal sensing; Point cloud registration; Plane segmentation
Tracks
DIGITAL MANUFACTURING
DOI
10.5703/1288284318686
Abstract
This paper presents an automated workflow for constructing an immersive digital twin by combining multimodal field sensing, ground-based CAD-to-point-cloud registration, and physics engine-based scene generation. To reduce human intervention during digital twin construction, the proposed pipeline estimates machine poses by registering CAD-derived point clouds to a colorized point cloud map. Instead of relying on unconstrained six-degree-of-freedom point cloud registration algorithms, such as Iterative Closest Point (ICP) or TEASER++, the proposed method detects the ground plane, reduces the search space to planar translation and yaw, and applies coarse-to-fine slicing followed by 3-DoF planar ICP refinement. The method was evaluated using two manufacturing machines in the Manufacturing and Materials Research Laboratories (MMRL) at Purdue University. Comparative trials using conventional ICP and TEASER++ failed to converge to correct alignments under the same conditions. In contrast, the proposed algorithm achieved an average rotation error of 1.582° and an average translation error norm of 0.2325 m across both test cases. These results demonstrate that the proposed ground-based registration method provides a practical and automated pipeline for constructing immersive digital twin scenes in large indoor manufacturing environments.
An Automated Construction Framework for Immersive Digital Twins with Ground-based CAD Pose Estimation
This paper presents an automated workflow for constructing an immersive digital twin by combining multimodal field sensing, ground-based CAD-to-point-cloud registration, and physics engine-based scene generation. To reduce human intervention during digital twin construction, the proposed pipeline estimates machine poses by registering CAD-derived point clouds to a colorized point cloud map. Instead of relying on unconstrained six-degree-of-freedom point cloud registration algorithms, such as Iterative Closest Point (ICP) or TEASER++, the proposed method detects the ground plane, reduces the search space to planar translation and yaw, and applies coarse-to-fine slicing followed by 3-DoF planar ICP refinement. The method was evaluated using two manufacturing machines in the Manufacturing and Materials Research Laboratories (MMRL) at Purdue University. Comparative trials using conventional ICP and TEASER++ failed to converge to correct alignments under the same conditions. In contrast, the proposed algorithm achieved an average rotation error of 1.582° and an average translation error norm of 0.2325 m across both test cases. These results demonstrate that the proposed ground-based registration method provides a practical and automated pipeline for constructing immersive digital twin scenes in large indoor manufacturing environments.