Ju Dong 1,2,3, Liding Zhang 2, Lei Zhang 1,2,3 *†, Yu Fu 2, Kaixin Bai 1,3,
Zoltán-Csaba Márton 3, Zhenshan Bing 2, Zhaopeng Chen 3, Alois Christian Knoll 2, Jianwei Zhang 1
1.TAMS(Technical Aspects of Multimodal Systems), Universität Hamburg, Germany. 2. Technical University of Munich, Germany.
3. Agile Robots SE, Munich, Germany.
† Corresponding author: lei.zhang-1@studium.uni-hamburg.de
* The authors contributed equally to this work
Bringing your policy into real-time: 125Hz high-frequency control, 43× acceleration, delivering precision and power even under dynamic disruption.🎉
Overview of the distribution-level distillation framework and data diversity.
Network Structure
Real-World Experiments
🥳Successful Task:
Dynamic tasks performed excellently!
Facing disruptions with consummate skill!
Complex tasks completed swiftly!
😈Failed Task:
Cannot complete tasks in dynamic scenes!
Cannot complete tasks under disturbance!
Cannot complete tasks with 1-step denoising!
In case this work has provided some assistance in your research endeavors, your citation would be greatly appreciated.
@misc{dong2026flowsteprealtimemultimodal,
title={From Flow to One Step: Real-Time Multi-Modal Trajectory Policies via Implicit Maximum Likelihood Estimation-based Distribution Distillation},
author={Ju Dong and Liding Zhang and Lei Zhang and Yu Fu and Kaixin Bai and Zoltan-Csaba Marton and Zhenshan Bing and Zhaopeng Chen and Alois Christian Knoll and Jianwei Zhang},
year={2026},
eprint={2603.09415},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2603.09415},
}