“Robust AI emerges where simulation and reality are brought together systematically”

Dr. Voit, why is synthetic data gaining so much importance right now?

Michael Voit: Today, many AI systems fail not because of algorithms, but because of the limits of available data. In safety-critical domains, rare or privacy-sensitive situations are difficult to capture in real-world data. Simulation and synthetic data help to close this gap by enabling relevant scenarios to be generated, varied and used for training and testing under controlled conditions. As a result, the focus is shifting from pure data collection towards strategic data development.

What does this mean from the perspective of the AI and Robotics business unit?

We develop AI systems for real, dynamic environments. These systems must operate reliably even under changing weather conditions, incomplete sensor data or unexpected behavior by other actors. In simulations, such parameters can be varied systematically to generate training and test data. At the same time, we are aware that a domain gap often exists between simulation and reality. For this reason, we combine virtual data with privacypreserving real-world data. This creates a feedback loop in which real observations improve simulation, while simulation in turn helps make AI systems more robust.

Where does Fraunhofer IOSB come in?

Across the entire chain from data acquisition to application. With OCTAS®, we are developing an open, modular simulation platform for autonomous mobility systems. It supports not only virtual validation, but also the generation of synthetic datasets. In AVEAS, we have demonstrated how critical traffic situations can be captured in a privacy compliant manner and transferred into data-driven simulation environments—including behavior, dynamics and sensor models. This transforms isolated datasets into a transparent, reproducible basis for developing and testing AI systems.

What added value does this offer for industry and users?

A good example is our AktiMeter, a system we developed for intelligent in-cabin behavior analysis. It enables real-time analysis of body poses, gestures and activities inside the vehicle for applications such as user research, ergonomics or adaptive assistance systems. Early development stages were based on driving simulator data; today, it is close to market readiness. While new approaches such as Vision Language Models are emerging that require less task-specific training data, reliable testing and validation remain essential − where simulation and synthetic data are key instruments.

Dr.-Ing. Michael Voit is spokesperson for the business unit AI and Robotics and head of the department Human-AI Interaction (HAI).