BMW Group: planning car plants in an Omniverse-based Virtual Factory

Bringing a new or updated model into an existing plant used to mean changing real production lines and testing them physically. BMW says collision checks for a new model took almost four weeks of real testing, sometimes with vehicle bodies pushed through the line by hand over several weekends.

Planning teams also worked in separate tools (BMW names Bentley MicroStation, ipolog, Siemens Process Simulate, Dassault Systemes CATIA and Autodesk Revit), and NVIDIA's case study says linking teams that work on different parts of the same factory required a central platform. With more than 40 new or updated vehicles planned for its network by 2027, BMW wanted to validate each launch before touching the physical line.¹²³

2023: Debrecen, planned virtually according to BMW

In March 2023 BMW described its future plant in Debrecen, Hungary, as its first facility planned and validated fully in a virtual environment, using NVIDIA Omniverse Enterprise, more than two years before series production; construction had only just begun at the time. Since November 2020 BMW had 3D-scanned more than 7 million square meters of indoor and 15 million square meters of outdoor production space, and it connected its existing planning tools to the shared Omniverse scene.

FactoryExplorer on Omniverse Kit and OpenUSD

According to NVIDIA's case study, BMW built its own planning application, FactoryExplorer, with the Omniverse Kit SDK on OpenUSD (an open scene standard that originated at Pixar and is now developed through the Alliance for OpenUSD). Planners from different disciplines edit separate OpenUSD layers of the same factory at once. BMW added its own extensions for collision detection, filtering, loading factory states from different dates, navigation and measuring whether equipment fits a space.

Synthetic data for inspection

The same NVIDIA page says BMW used Omniverse Replicator and Isaac Sim to generate synthetic training images for an AI inspection system. It adds that BMW used NVIDIA DGX systems (Hopper architecture) to train the synthetic-data-generation models behind its SORDI dataset, and uses DGX to evaluate and test trained models.

2025: scaled across the network

In June 2025 BMW said digital twins covered more than 30 production sites. The Virtual Factory combines building, equipment, logistics and vehicle data and supports automated collision checks, simulation of manual work steps, and automated mapping of surroundings from 3D scans for transport systems. BMW described generative and agentic AI assistants as an extension under way, not a finished capability.¹²³⁴

What others can learn: start with a use that has a clear before and after (BMW's collision checks), reuse scan and CAD data you already have, and let teams work on separate layers of one shared scene instead of exchanging files. Limits: BMW spent years scanning its plants and built its own applications, with BMW and NVIDIA engineering and planning experts working closely together; that effort is not shown in the figures. The 30 percent cost figure is a projection, and the three-day collision check is BMW's own number without a published method or independent check. Manufacturers with one site or simple line changes may get enough from the simulation features of the CAD and process tools they already license, without a custom Omniverse application.