BMW is reinventing the digital twin by embedding AI at its core, enabling intelligent assistants, global collaboration, and process optimisation across its industrial estate. The company’s work redefines the future of smart manufacturing and highlights how AI and simulation can converge in real time at an enterprise scale.
BMW is reinventing the digital twin by embedding AI at its core, enabling intelligent assistants, global collaboration, and process optimisation across its industrial estate. The company’s work redefines the future of smart manufacturing and highlights how AI and simulation can converge in real time at an enterprise scale.
Originally developed for video game design, Unreal Engine is a high-fidelity 3D creation tool that BMW first used in 2018 to explore immersive factory planning scenarios, an early step in what would become a radically ambitious digital transformation. By 2020, BMW had formalised its virtual factory initiative and began working with NVIDIA Omniverse. Just a year later, a strategic partnership was in place, and by 2022, the company had fully mapped its Debrecen plant in Hungary into Omniverse.
This was not just a visualisation milestone. It marked a shift in mindset from isolated digital replicas to dynamic, data-rich digital twins embedded in core business processes. “To scale usage and make the platform more accessible, we transitioned from on-premise hardware to Microsoft Azure cloud infrastructure in 2023,” Felix Theurer, Digital Solutions Architect Virtual Factory at BMW, explains. “This shift also allowed us to expand our development capacity through long-term DevOps partnerships. In 2024, we are continuing development using NVIDIA’s cloud infrastructure, which provides further flexibility and performance improvements for our virtual factory.”
The move to the cloud was pivotal. It enabled the company to support hundreds of simultaneous users while introducing a degree of operational resilience and compute elasticity impossible with on-premise hardware. Most importantly, it laid the foundations for bringing AI into the heart of factory planning.
A network of intelligence
BMW’s current scale is significant. With over 600 factory planners, each making around three modifications per week across 33 global sites, the virtual factory must support an enormous volume of activity. Each action, whether rerouting AGVs, checking for collisions, or simulating human workflows, must be tracked, visualised, and validated in real time.
“By increasing the number of users and the amount of data in the virtual factory, we gain network effects,” Theurer adds. “We began with the Debrecen plant, then progressively integrated additional plants into the virtual environment, each with specific use cases tailored to local requirements.” These include AGV mapping, semi-automated clash detection, and ergonomic analysis via human digital simulations.
But these are not siloed use cases. What is emerging is a globally connected network of intelligence. “With Omniverse and OpenUSD, we enable fully virtual factory planning and optimisation,” Theurer continues. “Multi-user collaboration capabilities mean planners can connect worldwide in real time, significantly accelerating optimisation cycles. Each BMW plant becomes a globally connected intelligence network node accessible via a standard web browser.”
Accessibility is more than just a convenience; it is a strategic asset. It supports organisational agility, enhances cross-border collaboration, and reduces time to implementation. The OpenUSD data format is central to this flexibility. Designed for extensibility and compatibility, USD allows BMW to customise workflows, extend functionality, and interoperate across tools without vendor lock-in.
Turning complexity into clarity
The technology stack beneath this ecosystem is deceptively sophisticated. At its core is a robust data integration pipeline, capable of ingesting and harmonising a wide range of 3D file formats, Revit, JT, DGN and others, into a unified USD structure. Access control, system connectors, file verification, and quality assurance are all automated within this pipeline.
“The virtual factory at BMW is built on two fundamental value propositions,” Harshit Raikar, Lead for Virtual Apps and Collaboration (Omniverse), BMW Group explains. “The first is a holistic file type, based on Universal Scene Description (USD), supported by a robust data integration pipeline. The second is customisability.” That customisability is not theoretical. BMW has developed a suite of Omniverse extensions, including point cloud visualisation, live sessions, and advanced filters, to adapt the platform to its operations.
This technical architecture supports digital twins that are both vast and accurate. More than one million square metres of production space is now modelled and validated, enabling end-to-end virtual replicating BMW’s manufacturing environment. These are not passive models. They are interactive, intelligent environments that reflect the state of production with a high degree of fidelity and can be used for real-time decision-making.
Intelligence for human-machine collaboration
At the heart of BMW’s next leap is the Intelligent Assistant. Designed to navigate the complexity of large-scale 3D environments, the assistant applies the semantic power of large language models to factory planning. “Factory planners need to find relevant information quickly within enormous 3D spaces, often labelled inconsistently or in multiple languages,” Tobias Delago, Researcher at BMW, says. “Using semantic capabilities of LLMs, we can identify user intent, even when it is expressed imprecisely, and direct them to the correct location or data.”
The assistant is not just a search function. It interprets queries, understands context, and adapts responses based on the environment loaded into the session. “Our architecture starts with the user inputting a prompt,” Delago continues. “This prompt is processed by a hybrid model combining the LLM with filtering mechanisms for validation and contextual narrowing. If only a factory segment is loaded, the assistant will limit its search accordingly. This conserves resources and improves result accuracy.”
Critically, the assistant maintains a principle of human-in-the-loop. It suggests actions, but it does not execute them autonomously. This balance between automation and control is essential in complex industrial environments, where multiple valid interpretations can exist, and human judgment is often required.
Bridging the knowledge gap
Beyond navigation, the Intelligent Assistant also acts as a bridge between domain expertise and technical complexity. Users can describe what they want to achieve rather than having to know which tool to use or how to use it. “Everything in our implementation is an Omniverse extension, built using NVIDIA’s Kit SDK framework,” Raikar continues. “But this level of customisation means tools often require expert knowledge to use effectively. This is where AI again plays a role.”
The assistant leverages retrieval-augmented generation (RAG) to tap into internal knowledge bases, combining structured operational data with unstructured documentation. This blend of contextual awareness and domain-specific intelligence is what allows the assistant to operate at an enterprise scale. “Years of groundwork in structuring our data now enable us to deliver highly accurate, context-aware AI assistance,” Raikar adds.
This ability to understand organisational context, not just user queries, differentiates BMW’s approach. AI is not simply a layer on top of existing systems. It is integrated into how the company designs, plans, and operates its factories.
The future of intelligent planning
BMW’s virtual factory is no longer just a vision. It is a functioning, AI-powered platform that supports real work, solves real problems, and delivers measurable benefits. It demonstrates what happens when AI is used not to replace human planners but to enhance their capabilities, shorten iteration loops, and reduce friction across global operations.
This is not a laboratory experiment. It is a new operational layer that spans facilities, departments, and time zones. It signals a broader shift in how industrial AI is being deployed: not as a standalone product but as a deeply embedded infrastructure that learns, adapts, and evolves with the business.
The company is now laying the foundation for a future where users communicate outcomes, not commands. This subtle but profound shift from tool-driven workflows to intent-driven collaboration represents the next phase in AI’s industrial evolution. It is a vision of smarter factories and factories that can think, interpret, and respond alongside their human counterparts.
BMW’s work proves that AI can be practical, intelligible, and useful at scale. The lesson for other organisations is clear: planning becomes transformation when digital twins become intelligent.
