Referentenportraits

A Technical Introduction to Neural Networks

In this fast-paced keynote, Prof. Dr. Guido M. Schuster introduces the fundamentals of neural networks, covering neuron structure, activation functions, and supervised learning with gradient descent. He demonstrates how single-layer networks can learn simple patterns but fail on non-linear problems like XOR.

To address this, he introduces multi-layer networks and the back-propagation algorithm, enabling effective training of deeper architectures. Using basic MATLAB code and mini-batch training, he illustrates key concepts such as error propagation, weight updates, and convergence behavior.

The session ends with a successful demonstration of solving XOR using a two-layer network, showing how neural networks can learn complex patterns when properly structured and trained.

Advancing AI-Guided Robotic, the Role of Synthetic Data

A major challenge in robotics lies in the significant implementation effort required to adapt a system for a specific task—an effort that demands substantial engineering expertise. AI-guided robotics is poised to change this by enabling robots to understand tasks and autonomously determine how to perform them, without explicit instructions. However, training AI systems typically requires large volumes of data, which can be costly, time-consuming, and difficult to acquire.

Synthetic data is transforming the development of artificial intelligence in robotics and industrial automation. By providing abundant, annotated, and customizable training datasets, it addresses the limitations of real-world data, which is often scarce or hard to capture.

One compelling example is the development of a robot designed to automatically unscrew and dismantle used car batteries for recycling. In this case, synthetic data played a key role in training the robot’s AI system, enabling more efficient and flexible performance. This example highlights the potential of synthetic data to accelerate development cycles, improve system robustness, and support automation in complex or safety-critical environments.

An automated 3D organ model framework as basis for robust AI data sets replicating human diversity

Within the European Chips JU project UNLOOC we aim at «unlocking the data content of organ-on-chips». Together with partners in the automation, bio and data analysis space, we aim at developing a technology platform for high-throughput cryopreservation of panels of 3D organ models, ready to use in screening. Advanced multi-parametric and multi-modal read out methods will allow to resolve, how pharmaceutical compounds react different human populations.

Mastering Real-World Challenges to Deploy AI

Exploring hybrid AI approaches for automation—combining data-driven models with physical knowledge. Applications include anomaly detection, predictive maintenance, and quality optimization. We demonstrate how AI enables adaptive control and transparent decision-making in dynamic production environments.

AI in Industrial Automation: Why the Models Are the Easy Part

The excitement around Artificial Intelligence in industrial automation is tangible—but we're often focusing on the wrong layer of the problem. While transformer architectures and generative models advance rapidly, the real challenge in our industry lies elsewhere: in unlocking the data behind proprietary barriers and connecting AI meaningfully to the physical world.

In this talk, I’ll argue that we must think about AI in industrial automation from the end goal backwards. The true IP of manufacturing lies not in algorithms, but in the experience of operators, the nuance of machine configurations, and decades of process tuning—most of which is trapped in PLCs and vendor-specific control systems.

To enable AI that truly enhances operations, we must first dismantle these silos, normalize and structure data at scale, and create plant-specific models that reflect the unique DNA of each facility. This isn’t about building a general AI overlord for all factories. It’s about building the systems that enable each factory to train its own expert.

Software Defined Automation enables this shift—decoupling control logic from hardware, liberating data, and making automation systems programmable and observable like modern software. These systems, not the AI models themselves, will be the true drivers of competitive advantage in the next industrial era.

Key Takeaways:

• Why «getting the data» is still the hardest part of AI in automation
• The risk of betting on general AI before solving infrastructure and integration
• How operator-centric, facility-specific AI can outperform one-size-fits-all models
• A vision for an open, software-defined control layer as the foundation for AI-powered manufacturing

Let’s move beyond the hype—and get to work on what really matters.

Lightweight Neural Networks in the field: Smart Cameras learn to monitor product quality

Instead of cost-intensive search for visual features by hand, convolutional neural networks (CNNs) learn such features automatically. New Smart Cameras like the Baumer AX series equipped with onboard GPUs allow for executing CNNs in the field. Neural Architecture Search (NAS) can help to find lightweight architectures to meet the requirements in typical production scenarios. We present an example project.

B&R integrate AI for developer in the daily usage

B&R has integrated AI into its current development toolbox, Automation Studio Code. We demonstrate how this makes daily work much easier and how quickly we can now write the glue code. We also demonstrate the AI camera and present a use case of a solution we have already implemented for a machine builder.