Consortium from industry develops robust AI systems together with Karlsruhe University of Applied Sciences (IRAS) – Training with virtual worlds instead of expensive real data.

Karlsruhe, 19 February 2026 – Imagine a screw sorting system suddenly no longer recognizing screws because their color shade has changed, even though this is not relevant to their function. Or a quality inspection in production failing because strong reflections occur depending on the position of the sun. These are exactly the kinds of problems the new research project “DeepFrame” aims to solve, carried out by a consortium of science and industry.

Strong consortium for applied AI research

In addition to the Karlsruhe University of Applied Sciences / IRAS (Institute for Robotics and Autonomous Systems), the companies VisionTools Bildanalyse Systeme GmbH, Elma Electronic GmbH and DE software & control GmbH are involved as partners. BMW AG supports the project as an associated partner and provides industrial use cases.
The project is funded by the Federal Ministry for Research, Technology and Space (BMFTR). Together, the consortium is working to make AI systems more reliable and robust for industrial applications.
Artificial intelligence often works excellently today, but only under ideal conditions. As soon as lighting conditions change, sensors become dirty, or unforeseen situations arise, many systems reach their limits. It’s like a student who has memorized everything but can’t cope with a slightly changed question. This is where DeepFrame comes in: The research team is developing AI systems that not only work, but also remain reliable under difficult conditions.

Training in virtual worlds

A special highlight of the project: Instead of laboriously recording and describing thousands of real images, the consortium trains AI systems with virtually generated data from simulations. This is comparable to flight simulators for pilots. We can run through any number of scenarios, from perfect to extreme conditions. No real component needs to be photographed.
These synthetic training data are not only more cost-effective, but also make it possible to specifically simulate critical situations:
What happens with backlighting? How does the system react to shadows or reflections? Can missing parts be reliably detected? “We know how demanding our customers’ challenges are,” explains Josef Djulic, Managing Director of VisionTools. “With the innovations from DeepFrame we offer solutions that are not only more reliable, but also easier to integrate. Our products such as VisionCockpit enable easy training of complex image processing systems, while the VoE-AIBox provides AI-based evaluation algorithms directly on the line. In this way, we support our customers in increasing their productivity, optimizing processes, and securing decisive competitive advantages.”

A second research focus is the combination of different sensor types – similar to how humans perceive their environment through the interplay of eyes, ears, and sense of touch. DeepFrame is developing AI models that intelligently link data from multiple sensors, such as color cameras or thermal imaging cameras. The advantage: If one sensor fails or delivers inaccurate data, the others can compensate.

“Our expertise in embedded computing solutions enables us to transfer AI developments directly onto powerful hardware,” explains Aksel Saltuklar, CTO of Elma Electronic. “We are proud to provide the technological foundation for these novel applications with our systems.”
“We integrate the AI functionalities into our smart worker assistance system,” explains Friedrich Steininger, CEO of DE software & control. “With the workstAItion 5.0 DE offers the ideal framework to use technological innovations in better image recognition and multimodal sensors in a process-oriented way. This enables us to make our worker assistance system even more resilient and at the same time easier to configure.”

Possible applications extend far beyond industry: In agriculture, robust AI systems could monitor plants under a wide range of weather conditions. And in rescue operations, they could keep search and rescue robots operational even in poor visibility or smoke-filled rooms.

Tailored for German SMEs

While American tech giants such as Google or Meta rely on huge amounts of data, the DeepFrame consortium develops solutions for small and medium-sized enterprises that cannot invest millions of euros in training data. The methods are resource-efficient and practical – exactly what German industry needs.

The DeepFrame Consortium:

VisionTools Bildanalyse Systeme GmbH
One of the leading system houses for industrial image processing in Germany.

Elma Electronic GmbH
Worldwide leading provider of embedded computing solutions.

DE software & control GmbH
Specialist for software and highly flexible automation solutions, especially smart worker assistance systems.

BMW AG (associated partner)
Global automobile and motorcycle manufacturer of innovative premium vehicles.

Karlsruhe University of Applied Sciences – Institute for Robotics and Autonomous Systems (IRAS)
Institute for applied AI research with a focus on robotics and intelligent systems.