About

I am Yu Deng, a Ph.D. candidate in the AI/ML Lab at TU Darmstadt, supervised by Prof. Dr. Kristian Kersting. I am also affiliated with hessian.AI and the JUPITER AI Factory.

My long-term goal is embodied robotic reasoning: building robots that can form explicit, inspectable models of the physical world and their own behavior, test those models through interaction, and use what they learn to act more reliably in changing environments.

Email Google Scholar GitHub

Publications

2026

STORM: Segment, Track, and Object Re-Localization from a Single Image
Yu Deng*, Teng Cao*, Hikaru Shindo, Quentin Delfosse, Jiahong Xue, Kristian Kersting
International Conference on Machine Learning (ICML 2026)Paper Code
Tracks 6D object pose from a reference image, detects drift, and re-localizes after occlusion or viewpoint changes.
Robot-DIFT: Correspondence-Sensitive Diffusion Features for Contact-Rich Robot Manipulation
Yu Deng*, Yufeng Jin*, Xiaogang Jia, Jiahong Xue, Gerhard Neumann, Georgia Chalvatzaki
Conference on Robot Learning (CoRL 2026, accepted)Paper
Distills diffusion features into a fast visual backbone that preserves geometric correspondences needed for contact-rich control.
Nautilus: From One Prompt to Plug-and-Play Robot Learning
Yufeng Jin*, Jianfei Guo*, Xiaogang Jia, Yu Deng, Zechu Li, Han Liu, Weiran Liao, Vignesh Prasad, Mathias Franzius, Gerhard Neumann, Georgia Chalvatzaki
Conference on Neural Information Processing Systems (NeurIPS 2026, Poster)Paper
Turns a single prompt into validated workflows for reproducing, evaluating, fine-tuning, and deploying robot learning methods.

Preprints

ALDER: Discovering the Laws of a World by Acting in It
Teng Cao*, Yu Deng*, Quentin Delfosse, Kristian Kersting
arXiv 2026Paper
Discovers and revises explicit world-model equations through targeted experiments, then uses validated laws for goal-directed control.
Kintsugi: Learning Policies by Repairing Executable Knowledge Bases
Teng Cao*, Yu Deng*, Hikaru Shindo*, Quentin Delfosse, Lanxi Wen, Suli Wang, Jannis Blüml, Christopher Tauchmann, Kristian Kersting
arXiv 2026Paper
Repairs explicit policy knowledge from rollout failures and admits each edit only after execution and regression checks.
Learning Explicit Behavioral Models with Adaptive Questions and World-Model Probes
Hikaru Shindo, Yu Deng, Teng Cao, Quentin Delfosse, Christopher Tauchmann, Jannis Blüml, Gopika Sudhakaran, Kristian Kersting
arXiv 2026Paper
Learns inspectable behavioral models through task feedback, adaptive questions, and executable world-model probes.
CogniFold: Always-On Proactive Memory via Cognitive Folding
Suli Wang*, Yiqun Duan*, Yu Deng*, Rundong Zhao, Dai Shi, Minghua Deng, Chen Chen, Yiqi Wang, Xinliang Zhou
arXiv 2026Paper Code
Folds incoming events into evolving memory structures that can surface concepts and intentions for proactive agents.