Digital factory 2026
How the car plant becomes a learning system
Digital twins and physical AI are currently the biggest game changers in automotive production. Examples from BMW, Volkswagen or Schaeffler make clear why some technological promises still have to stand the test of reality.
TL;DR - Digital factory 2026
The digital factory is becoming increasingly operational in the automotive industry: Digital twins, AI-supported planning and live data from production are intended not only to visualise processes, but also to prepare decisions and make systems controllable. BMW relies on AI in factory planning, Schaeffler uses digital images and real-time data to master complex e-mobility programmes, Volkswagen and Audi are driving forward physical AI and software-defined robotics. Humanoid robots are regarded as an exciting field of the future, but still remain far from robust series maturity. The decisive factor will be to turn demos into scalable, economical and safe solutions for the shop floor.
Automotive production is facing a new surge in digitalisation. At the 21st Digitale Fabrik congress in Herzogenaurach, experts including from BMW, Volkswagen and Schaeffler showed how digital twins, AI-supported planning and Physical AI could change the factory. But it also became clear: much work remains between impressive demos and robust series deployment.
At Schaeffler, this year’s host of the event, this shift became particularly tangible. Silvio Zamzow, Senior Vice President Operations & SCM E-Mobility at the supplier, described digitalisation not as an abstract strategic project, but as an operational necessity.
The E-Mobility division, which was restructured two years ago through the acquisition of Vitesco, had moved beyond the start-up phase and now had a significant series business. As a result, complexity was increasing massively: an electric axle could consist of around 1,200 components that would have to be brought together from numerous sites, plants and regions.
Zamzow cited an example with 1,280 components from 17 Schaeffler locations, seven plants and three regions. “All the influences that you can have are in there,” he said. Quality, material, regional risks and supply chains had to be managed in such a way that, in the end, there was secure supply to the customer.
From digital master to controllable twin
For Schaeffler, the answer to this begins with a digital master. What is meant is as complete a digital image as possible of the site, building, plant planning and product development before investments in “steel and iron” are made. The second step is the digital shadow, that is, the live view of machine and process data. Only from the connection of both worlds does the digital twin arise for Zamzow: a system with which products and production processes can be controlled independently of location.
The claim here is clearly more operational than in earlier digitalisation programmes. Zamzow compared the digital shadow with a navigation system: if a blockage occurred, the system would have to suggest a new route. Transferred to production, this means: disruptions, bottlenecks and deviations should no longer become visible only in subsequent reports, but should be controllable as far as possible in real time.
In the E-Mobility division, Schaeffler is already well advanced in this regard. Zamzow explained that 100 per cent of the E-Mobility areas are digitally mapped, with a defined annual re-scan. Across the group, around 70 per cent of the production areas are accessible via Navis. In addition, there are applications such as the Manufacturing Intelligence Assistant, or MIA for short. Via this system, those responsible can access station and plant data, analyse output, OEE, downtime or error patterns and derive escalations. “I want to discuss with live data in meetings and not with data that someone entered,” said Zamzow.
Especially in the brownfield, however, this is not a sure-fire success. In the discussion, it became clear that Schaeffler in recent years first had to create the technical basis. More than 10,000 machines had been brought to a uniform standard via the company’s own machine connectivity programme. Only this creates that data layer on which dashboards, digital twins and future AI applications can be built at all.
BMW relies on AI in production planning
At BMW, too, the digital twin is the basis for the next step. Tobias Delago, Researcher Virtual Factory at the BMW Group, made it clear that the use of AI in production planning is subject to special conditions. The complexity is high, but the data basis is limited compared with classical machine-learning applications. BMW does indeed have large production structures worldwide, but not thousands of identical systems as a training basis. “We have 16 production sites,” said Delago. For a machine-learning model, that is not a “cool number”.
This is why, according to Delago's account, BMW is pursuing a neuro-symbolic approach. In this, probabilistic AI methods are combined with deterministic expert knowledge. That sounds abstract, but it can be explained using a simple planning example: an AI must not only generate suggestions for plant layouts, but also observe rules that are self-evident to humans. A production facility should, for example, not be placed on a hall pillar. It is precisely this knowledge that must be incorporated symbolically in order to make AI systems robust in planning.
Delago classified the path within a multi-stage digitalisation logic. First, BMW scanned plants worldwide and connected them to planning systems. This was followed by holistic planning environments, simulation and optimisation in the Industrial Metaverse. Now, AI-supported and increasingly agentic planning is moving into focus. The aim is not to generatively design any arbitrary plants, but to arrange available operating resources sensibly and assess planning variants more quickly. The benchmark is high here: an AI proposal must be at least as good as copying and adapting a proven plant from an earlier project.
Lars Fritzsche, Managing Director of IMK Industrial Intelligence, showed together with Delago how AI can concretely accelerate planning processes. In IMK's EMA software, manual activities, ergonomics, productivity and material flows can be simulated. Until now, it was above all the creation of such simulations that was time-consuming. This is precisely where AI comes in: From text information, work instructions and 3D data, an initial simulation is to be created automatically.
Fritzsche spoke of a new quality, because AI could also assign objects in the 3D scene when designations do not exactly match the work instruction, for example in the case of abbreviations or different languages. In one example, an experienced user had needed around nine to ten minutes, the AI about one minute. That corresponded to a potential of 80 to 90 per cent in the creation of an initial simulation. Nevertheless, the human remained decisive. Fritzsche emphasised that, in his view, “in the end, even when everyone simply uses AI, it is still the human who makes the difference”. The human defines premises, initiates optimisation loops and validates results.
Physical AI becomes the next automation step
Beyond simulation and planning, Volkswagen showed how AI is already growing into real automation tasks. Timon Thomaser, Head of Analytical & Physical AI Production at Volkswagen, and Henning Löser, Head of Innovation Management at Audi, placed the term “AI Robotics” at the centre. Physical AI meant that AI acts in the physical world. When this happens in robots, Volkswagen speaks of AI Robotics.
Important for Thomaser was the distinction from the hype around humanoid robots. Humanoids were only a sub-area for Volkswagen. The form factor was of secondary importance; what was decisive were the capabilities integrated into a robot system. The aim was to automate processes that had previously not been automatable, to increase quality and to reduce engineering effort. Thomaser spoke of the expectation of being able to reduce the engineering effort for classical robotics by up to 70 per cent. Especially under the cost pressure on German plants, this potential was strategically relevant.
The Volkswagen Group has organised the topic in a three-shell model. A lead team from Volkswagen and Audi coordinates, a core team involves the twelve automotive brands, and a community is intended to carry knowledge more broadly into the Group. The roadmap has three stages: first implement use cases and learn what works; then roll out successful applications to further sites; finally significantly increase the degree of automation. Thomaser spoke of the goal of automating 50 per cent additional processes in the longer term.
As an example, he mentioned the handling of cables in battery assembly at the VWN plant in Hanover. There, one robot takes a cable from a box, a second takes it over and inserts it into the battery in real production. The basis is camera data, CAD models and synthetic training data. The second use case concerned unpacking ball bearings from cartons and separating packaging materials. Here, Volkswagen worked together with the US start-up Rhoda AI. After around ten to 15 hours of additional training, the system, with two stationary robots, had stably mastered the task. Thomaser said he was impressed by the possibilities and announced that he wanted to bring the topic into series production in Wolfsburg at the end of the year.
Audi wants to detach robotics from the automation box
Löser then broadened the view to the architecture of automation. Traditionally, in industry, a box is bought for one function: camera, computer, controller, robot hardware. But this logic is becoming less and less suited to a world in which software and AI hardware are developing in short cycles. A robot that is bought today must still be able in three years' time to cope with new AI models and more powerful hardware.
Löser therefore argued for a new paradigm. Functions should not be permanently tied to proprietary hardware, but should be able to run as software, virtualised or containerised, in the local data centre. Together with partners, Audi has already shown that a programmable logic controller with Safety Integrity Level 3 can be operated in such an architecture. From this, Löser derives the next step: If a virtual PLC is possible, why should robot path planning not also be separated from the robot and operated centrally on scalable hardware?
His vision extends to the real-time coupling of the digital twin and the real robot cell. A central control system could then know where each robot is, how process times fluctuate and how programmes must be adjusted “on the fly”. Classic interlocks, in which one robot waits until the other has reached its position, could thereby be reduced. Löser sees in this an opportunity to operate existing and future installations more flexibly and cost-effectively.
Here too, he put the focus on humanoid systems into perspective. On the shopfloor there were hardly any activities in which employees would have to climb stairs. “So why necessarily legs?”, Löser asked the audience. And why limit oneself to human reach when four metres are more sensible for a side wall frame? Löser put it pointedly: “Humanoid robotics provide important impulses in perception, reaction to environmental influences and software. For production, however, these capabilities must be transferred to designs that actually fit the process.”
Humanoid robots: great potential, major gaps
Sebastian Reitelshöfer from the Chair of Manufacturing Automation and Production Systems at FAU Erlangen-Nuremberg elaborated precisely on this classification. He drew an ambivalent picture of humanoid robotics. On the one hand, such systems could become a new industrial field of the future for Europe and Germany. The automotive and supplier industry in particular has the know-how to build complex mechatronic systems in high quality. On the other hand, industrial use is still far from being suitable for everyday use.
Reitelshöfer warned against confusing demo videos with productive maturity. At the beginning of 2025, there were practically no market-available humanoid systems. In the meantime, Unitree and Agibot now offer the first purchasable platforms, but these are not yet comparable with robust industrial robots. With the Unitree G1, for example, CE marking and industrial maturity were lacking; after just a few minutes, even simple loads could already reach limits. Such systems are suitable for practising and preparing use cases, but not yet for broad productive use.
At the same time, Reitelshöfer pointed to the dynamics of Chinese manufacturers. Companies such as Agibot had only been on the market for a short time, but were already communicating high unit volumes. China was investing massively in humanoid robotics. For Europe, therefore, the question was not only one of use in factories, but also of industrial sovereignty. If robot systems were in future to become a kind of workforce in care, public space or production, a strong dependence on non-European providers could become problematic.
AI was a central enabler in this, but not the whole solution. Visual-language-action models worked impressively in demos, but under laboratory conditions had not yet achieved the reliability required in production. Reitelshöfer spoke of success rates of around 80 to 85 per cent for individual actions. For industrial processes, in which scrap and downtime are expensive, that was not sufficient. In addition, there were standardisation, safety, social acceptance and mechatronics. He sees planned standards for stabilised robot systems as particularly critical if they transfer requirements from collaborative robotics almost unchanged. Proof that a freely acting AI system complies with defined pain tolerance limits upon physical contact could considerably complicate deployment in Europe.
Social interaction is also underestimated. Reitelshöfer described an example from a plant in which mobile robots are no longer allowed to speak due to a works council decision. The reason: The systems had repeatedly disturbed employees without scene understanding. For humanoid robots, expectations are even higher because their form suggests human behaviour. If these systems are to be accepted in production environments, they must act not only safely, but also in a socially appropriate manner.
Simulation as a bridge between hype and series capability
A common denominator of the presentations was the role of simulation. Whether at BMW, Volkswagen or Schaeffler: Digital models are intended to not only visualise planning, but also prepare decisions, evaluate variants and make real systems controllable. Fritzsche showed how humanoid robots can be simulated in EMA scenarios and compared with manual activities or cobots. This is not only about cycle time, but also about walking routes, energy consumption, ergonomics and cost-benefit assessments. Even for robots, workplace design remains relevant, Fritzsche said in essence: A humanoid robot takes longer, consumes more energy or wears out more quickly if it constantly has to pick up heavy parts from the floor.
This makes simulation a translating instance between technological euphoria and industrial reality. It helps to identify use cases, estimate economic viability and make physical limits visible before expensive hardware is integrated into lines. Precisely because many humanoid systems are not yet available or not sufficiently robust, the virtual evaluation can help to prepare the right applications.
Research warns that humanoid robotics will only become an industrial field of the future if safety, standardisation, mechatronics and acceptance grow along with it
The digital factory is becoming more operational
What linked the contributions in Herzogenaurach was a sober view of the benefits. The digital factory is no longer understood primarily as a visualisation environment, but as an operational control instrument. At Schaeffler, the aim is to make global e-mobility programmes manageable with live data. At BMW, AI and simulation are intended to accelerate planning. Volkswagen is looking for AI robotics use cases that close real automation gaps. Audi is rethinking the architecture of automation, away from fixed hardware boxes towards software-defined functions. And research warns that humanoid robotics will only become an industrial field of the future if safety, standardisation, mechatronics and acceptance grow along with it.
This shows that Digitale Fabrik 2026 presents an industry in transition. Automotive production has created many of the foundations of the digital factory: scanned plants, networked machines, data platforms, virtual models. The next step is to use these foundations intelligently. AI is intended not only to analyse, but to plan, optimise, control and act physically. Yet the path to the autonomous factory will not be a leap, but a sequence of hard integration work. The decisive question is therefore not whether AI and Physical AI will change production. But how quickly it succeeds in turning impressive demonstrators into robust, scalable and economical production solutions.
