Hanno Gottschalk (TU Berlin) on Automated Driving Using AI

The image shows a prediction of the quality of image recognition (semantic segmentation): Bottom left: Image annotated by a human, the so-called ground trouth. Bottom right: The prediction of the image recognition network. Top left: Comparison of the prediction quality with the ground trouth. Well-predicted segments are in green, and poorly predicted segments are in red. Top right: Quality prediction by the MetaSeg algorithm [*].

Hanno Gottschalk has been a professor for “Mathematical Modeling of Industrial Life Cycles” at the Institute of Mathematics at TU Berlin since April 2023, when he also became a member of MATH+. His appointment marked the first of three new TU professorships for the Werner-von-Siemens Centre for Industry and Science (WvSC), where he supervises the technology field “Digitalization.” A significant focus of his research is the application of data-driven modeling using neural networks to automated driving, particularly specializing in its safety.


© Sebastian Jarych

Hanno Gottschalk prefers the term “automated driving” over “autonomous driving,” noting that, unlike the Greek concept of autonomy, which implies freedom and sovereignty, automated driving involves algorithms acting automatically without these notions. His research centers on the errors of AI-based perception systems, which transform sensor data into meaningful representations through a process known as semantic segmentation. His team focuses on identifying and addressing errors in perception systems, which are critical for ensuring the safety of automated driving.


Gottschalk heads two MATH+ projects related to automated driving. The first project (AA5-8), in collaboration with Gabriele Steidl (TU Berlin), involves generative AI and transport theory leading to new AI architectures for generating learning scenarios. The aim is to use this for generative counterfactuals, i.e., hypothetical scenarios exploring alternative realities relevant to street scenes, thus connecting the topic to automated driving. The second project (PaP-5) applies computer vision experiences with large models and big clusters to more classical applied mathematics in a field called scientific machine learning, using solutions of partial differential equations for training data.


The benefits of automated driving are evident, such as enhancing mobility, particularly for the elderly, and reducing transportation costs and the number of cars needed. However, additional problems can arise, for example, cutting transport costs can increase micro-transports, leading to more crowded streets. Gottschalk reflects: “The extension of our capability also requires reflection on our intentions. In principle, you can reduce the numbers of cars on the street because no one needs a car parked in front of their door anymore. Cars can park in central facilities and be called with an app on demand, which will also be much cheaper because you don’t need to pay the driver. There is some potential in the idea, though there will still be many things that can go wrong.”


The challenges in his research include ensuring reliable perception systems for handling unseen objects. Gottschalk further elaborates: “We first focused on uncertainty quantification for AI-based perception, developing globally recognized methods. We then shifted to forced errors and out-of-distribution detection for unseen street scenarios. Currently, we are working on vision-language models, which offer better stability and improve the robustness for perception algorithms. We’re proud to have the leading street scene recognition network in Berlin, excelling in the cityscapes challenge.”


Vision-language models, which train on both visual and textual data, show promise in enhancing the stability and robustness of perception systems. These models use contrastive learning, drawing pairs of images and corresponding texts from the internet to train algorithms in matching images with texts without needing extensive labeling. Despite the vast data and computational resources required, Gottschalk’s team works with open-source models to advance their research.


Addressing public concerns about AI and automated driving, Hanno Gottschalk acknowledges the importance of being critical and addressing technology’s blind spots. He highlights the vulnerability of neural networks to adversarial attacks, both intentional and natural, and emphasizes the need for robustification through vision-language models. Gradually expanding the domain of automated driving applications and ensuring critical testing are essential for the technology’s safe and reliable development.


Short bio:

Hanno Gottschalk studied mathematics and theoretical physics in Freiburg and Bochum, earning his doctorate in mathematics from Ruhr Universität Bochum in 1999. After a PostDoc at Universität Bonn, where he habilitated in 2003 on models and structures of quantum field theory, he worked at Siemens Energy from 2007 to 2011. From 2011 to 2023, he served as an associate professor  for “Stochastics” at Bergische Universität Wuppertal, researching mechanical component reliability and AI uncertainty. He also led the interdisciplinary center “Machine Learning and Data Analytics” and the Institute for “Mathematical Modeling, Analysis and Computational Mathematics (IMACM).” Since April 2023, he has been professor for “Mathematical Modeling of Industrial Life Cycles” at TU Berlin. His research interests cover safe AI, mathematical modeling and mathematical physics.





  • Hanno Gottschalk is head of two MATH+ projects: