The joint research paper by Jens U. Brandt (THK-AI Research Cluster, first author), Noah C. Pütz (THK-AI Research Cluster), and Prof. Dr. Thomas Bartz-Beielstein (Director THK-AI Research Cluster), Marcus Greiff, Thomas Lew, and John Subosits (Toyota Research Institute), as well as Marc Hilbert (Toyota Gazoo Racing Europe), was accepted at the renowned NeurIPS 2025 conference in San Diego – one of the most important forums for advances in artificial intelligence and machine learning.
Titled “From Faults to Features: Pretraining to Learn Robust Representations against Sensor Failures,” the work addresses a central challenge in safety-critical applications such as autonomous driving and driver assistance systems: How can AI models function reliably when sensor signals are faulty or have failed?
To this end, the team proposes a novel pretraining approach that specifically simulates sensor failures and trains the model to reconstruct missing information. The results are impressive: on a vehicle dynamics dataset, a significant increase in robustness against known and unknown failures is demonstrated – without significant losses in nominal performance. Particularly practical: the approach was tested on a modified Lexus LC 500, where the model enabled stable driving performance in an autonomous racing environment, even under critical sensor failures.
This is an exciting work that is not only scientifically convincing but also paves the way for robust, reliable AI systems in real-world applications and sends a strong signal from TH Köln and its partners to the international research community.
The research project is funded by the EU research project ShapeFuture. Its implementation is also made possible by the powerful resources of the THK-AI Research Cluster.