Publications | Jan Tauberschmidt

arXiv

Behavior-Constrained Reinforcement Learning with Receding-Horizon Credit Assignment for High-Performance Control

Siwei Ju, Jan Tauberschmidt, Oleg Arenz, Peter van Vliet, Jan Peters

We propose a behavior-constrained reinforcement learning framework for high-performance control that improves beyond demonstrations while explicitly limiting deviations from expert behavior, using receding-horizon trajectory prediction for look-ahead credit assignment in professional-driver race car simulation.

arXiv DOI

BibTeX

@misc{ju2026behaviorconstrained,
  title         = {Behavior-Constrained Reinforcement Learning with Receding-Horizon Credit Assignment for High-Performance Control},
  author        = {Ju, Siwei and Tauberschmidt, Jan and Arenz, Oleg and van Vliet, Peter and Peters, Jan},
  year          = {2026},
  eprint        = {2604.03023},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48550/arXiv.2604.03023},
  url           = {https://arxiv.org/abs/2604.03023}
}

ICML 2026 · International Conference on Machine Learning

Heavy-tailed Physics-Informed Neural Networks

Jephte Abijuru, Mayank Nagda, Phil Ostheimer, Jan Tauberschmidt, Sebastian J. Vollmer, Stephan Mandt, Marius Kloft, Sophie Fellenz

We study heterogeneous, heavy-tailed residuals in physics-informed neural network training and use this perspective to develop more robust physics-informed learning objectives for scientific machine learning.

Poster

BibTeX

@inproceedings{abijuru2026heavytailed,
  title     = {Heavy-tailed Physics-Informed Neural Networks},
  author    = {Abijuru, Jephte and Nagda, Mayank and Ostheimer, Phil and Tauberschmidt, Jan and Vollmer, Sebastian J. and Mandt, Stephan and Kloft, Marius and Fellenz, Sophie},
  booktitle = {International Conference on Machine Learning},
  year      = {2026}
}

ICLR 2026 · International Conference on Learning Representations

Physics-Constrained Fine-Tuning of Flow-Matching Models for Generation and Inverse Problems

Jan Tauberschmidt, Sophie Fellenz, Sebastian J. Vollmer, Andrew B. Duncan

We present a framework for fine-tuning flow-matching generative models using weak-form physical residuals, with applications to generation and inverse problems in scientific systems.

OpenReview arXiv Code DFKI

BibTeX

@inproceedings{tauberschmidt2026physicsconstrained,
  title={Physics-Constrained Fine-Tuning of Flow-Matching Models for Generation and Inverse Problems},
  author={Jan Tauberschmidt and Sophie Fellenz and Sebastian Josef Vollmer and Andrew B. Duncan},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=khBHJz2wcV}
}