Industrial

Reality Capture in the Industrial Metaverse Real-World Lab

Creating a digital twin in industry requires the transformation of physical assets, manufacturing cells, and complex systems into precise and operational three-dimensional representations directly from the shop floor. Choosing the appropriate 3D capture technology is by no means straightforward: it depends directly on the requirements of the specific use case, the required level of metric accuracy, the available budget, and the implementation time. 

While simple shop floor layout planning may require only volumetric or two-dimensional information, metrological inspections and physical simulations in NVIDIA Omniverse or the training of models in Physical AI require high-precision geometries at the millimeter level or advanced photorealistic accuracy. 

Benchmark Test System: SPIE Ventilation System

To objectively evaluate every data collection technology on the shop floor, the ventilation system was used as the reference assetventilation system SPIE ventilation system on the shop floor of the ARENA2036. 

Industrial Challenges of the Benchmark Asset:

• Mirror-like reflections and sheet metal: The stainless steel and aluminum channels have a high gloss finish. This leads to errors in feature matching (feature matching) in photogrammetry, as well as signal deviations caused by multipath reflections during laser scanning. 

• Flat, featureless surfaces: Large, smooth sections make it difficult to identify landmarks (keypoints) by passive computer vision algorithms. 

• Changing ambient reflections: The dynamic environment of the shop floor creates moving reflections that interfere with photometric reconstruction. 

Analysis by Technology

Terrestrial Laser Scanning 

Using the laser scanner FARO Focus was used to SPIEfacility was captured from six strategically distributed scanning positions at various heights and perspectives on the shop floor. 

• Methodology: Registration of a high-density point cloud with manual cleaning and alignment using reference spheres. 

• Data collection time: approximately 30 minutes on-site on the shop floor. 

• Result: a point cloud with millimeter-level precision, suitable for conversion to BIM or CAD reconstruction. 

• Business Benefits & Added Value: Drastically reduces downtime during reengineering projects (retrofits). Scanning the entire installation in 30 minutes eliminates inaccurate manual measurements and ensures that new CAD components fit without collisions right from the start. 

Photogrammetry

A high-resolution photogrammetry pipeline was implemented, in which the asset was photographed from multiple angles on the shop floor. 

Technical Consideration: Use a smaller aperture (f/8 – f/11) to ensure a deep field of depth. This prevents blurred areas at the corners and edges, which could impair the alignment of key points. 

Limitation: Smooth metal surfaces required light polarization techniques to reduce reflections and glare. 

Business Benefits & Added Value: Good cost-benefit ratio for documentation and remote visual inspection without the need for expensive, dedicated laser hardware. 

3D Gaussian Splatting (3DGS)

Gaussian splatting represents the scene not with conventional polygons, but with millions of spatially optimized three-dimensional Gaussian functions (splats). 

Historical Data Capture and Demonstrator: Video-based capture using a mobile device and photorealistic processing. Provides a highly immersive visual experience for inspections and virtual tours on the shop floor. The integration of advanced image calibration pipelines has resulted in greater geometric precision and reproducibility. 

Business Benefits & Added Value: Enables immediate, real-time virtual tours on standard hardware (web browsers, VR headsets) without requiring supercomputers for rendering. This facilitates collaborative decision-making among remote teams. 

Foundation Models for Generative 3D AI

Models were evaluated that are capable of inferring 3D geometry and texture from sparse visual inputs or single images infer, comparing technological developments between 2025 and 2026. 

• Benchmark Evolution:
  • Initial evaluation (September 2025): Reconstruction using first-generation tools (e.g., Polycam), which generated simplified meshes with structural constraints: primarily useful as placeholders. 
  • Current Benchmark (2026 - Hi3D 3.0): Inference using state-of-the-art models that generate complete geometries with texture maps in 8K resolution. This produces static assets that can be integrated directly into the digital twin. 

• Business Benefits & Added Value: Exponentially accelerates the design and simulation phase by allowing virtual scenarios to be populated with 3D assets in minutes using simple photographs. This eliminates weeks of manual CAD modeling of secondary objects. 

Evaluation Matrix and Decision-Making Framework

The qualitative evaluation matrix developed in the ARENA2036 real-world lab is presented below. This table summarizes the data acquisition methods based on their geometric accuracy, operating/hardware costs, reproducibility, and ideal application on the shop floor, and takes into account the evolution of the state of the art between September 2025 and August 2026.

*Development Note (2025–2026): 3D Gaussian Splatting has evolved through advanced pipelines that even incorporate calibration images, resulting in higher geometric accuracy (3/5) and reproducibility (3/5). 

**Development Note (2026): AI-based 3D foundation models (e.g., Hi3D 3.0) have significantly improved texture quality (up to 8K) and visual consistency compared to the tools from 2025 (e.g., Polycam). This facilitates the direct integration of assets into the digital twin, although certain scale discrepancies and geometric artifacts may still occur. 

Summary of Technology Selection

The following simplified decision-making rule is recommended for selecting the appropriate technology within the ARENA2036 ecosystem: 

1. For requirements related to PLC integration, metrology, or reverse engineering: Use laser scanning (TLS) or direct CAD modeling.
2. For applications requiring fast, immersive visual inspection without physics, use 3D Gaussian Splatting (3DGS).
3. To populate simulation scenarios and integrate assets into digital twins: Use generative 3D AI models (after verifying the scale).
4. Hybrid/Combinatorial Strategy: For complex shop floor projects, the combination of laser scanning (for rigid geometry and metrology) and 3DGS or generative 3D AI (for rapid aesthetic contextualization and secondary assets) offers the optimal balance of precision, development time, and ROI.