FlowVLA: Visual Chain of Thought-Based Motion Reasoning for Vision-Language-Action Models
IEEE Transactions on Robotics · 2026 · 42 · 3765-3780
DOI: 10.1109/tro.2026.3732824 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Robotics, 2026. Verfasst von Zhide Zhong und 10 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 52 Literaturangaben. Gefördert von: Natural Science Foundation of China, New Generation Artificial Intelligence-National Science and Technology Major Project, National Key Research and Development Program of China.
Abstract (Originalsprache)
Many Vision-Language-Action (VLA) models are built upon an internal world model trained via next-frame prediction “ $V_{t} \rightarrow V_{t+1}$ .” However, this paradigm attempts to predict the future frame's appearance directly, without explicitly reasoning about the underlying dynamics. This lack of an explicit motion reasoning step often leads to physically implausible visual forecasts and inefficient policy learning. To address this limitation, we introduce the Visual Chain of Thought , a paradigm that compels the model to first reason about motion dynamics before generating the future frame. We instantiate this paradigm by proposing FlowVLA , an autoregressive Transformer that explicitly materializes this reasoning process as “ $V_{t} \rightarrow F_{t} \rightarrow V_{t+1}$ ”, where $F_{t}$ is an intermediate optical flow prediction that inherently encodes motion. By forcing the model to first follow the motion plan encoded by $F_{t}$ , this process aligns the pretraining objective of dynamics prediction with the downstream task of action generation. We conduct experiments on challenging robot manipulation benchmarks, as well as a real-robot platform. Our FlowVLA not only generates more coherent and physically plausible visual predictions , but also achieves state-of-the-art policy performance with substantially improved sample efficiency , pointing toward a more principled foundation for world modeling in VLAs.
Förderung
- Natural Science Foundation of China
- New Generation Artificial Intelligence-National Science and Technology Major Project
- National Key Research and Development Program of China
- Guangdong Basic and Applied Basic Research Foundation
Zitieren
APA
Zhong, Z., Yan, H., Li, J., Liu, X., Gong, X., Zhang, T., Song, W., Chen, J., Zheng, X., Wang, H., & Li, H. (2026). FlowVLA: Visual Chain of Thought-Based Motion Reasoning for Vision-Language-Action Models. IEEE Transactions on Robotics, 42, 3765-3780. https://doi.org/10.1109/tro.2026.3732824
Vancouver
Zhong Z, Yan H, Li J, Liu X, Gong X, Zhang T, et al. FlowVLA: Visual Chain of Thought-Based Motion Reasoning for Vision-Language-Action Models. IEEE Transactions on Robotics. 2026;42:3765-3780. doi:10.1109/tro.2026.3732824
BibTeX
@article{Zhong2026,
author = {Zhong, Zhide and Yan, Haodong and Li, Junfeng and Liu, Xiangchen and Gong, Xin and Zhang, Tianran and Song, Wenxuan and Chen, Jiayi and Zheng, Xinhu and Wang, Hesheng and Li, Haoang},
title = {FlowVLA: Visual Chain of Thought-Based Motion Reasoning for Vision-Language-Action Models},
journal = {IEEE Transactions on Robotics},
year = {2026},
volume = {42},
pages = {3765--3780},
doi = {10.1109/tro.2026.3732824}
}
EUVIMED