Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies

原文：Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies

作者：Trung Dao, Sankalp Yamsani, Jaden Park, Joohyung Kim, Yong Jae Lee

来源：arXiv cs.CV（计算机视觉）

正文

Computer Science > Robotics

arXiv:2609.24682v1 (cs)

[Submitted on 21 Sep 2026 (this version), latest version 9 Oct 2026 (v3)]

Title:Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies

Authors:Trung Dao, Sankalp Yamsani, Jaden Park, Joohyung Kim, Yong Jae Lee

View a PDF of the paper titled Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies, by Trung Dao and 4 other authors

View PDF

HTML (experimental)

Abstract:Vision-Language-Action (VLA) models map observations to actions with no objective that accounts for how the world responds, so their robustness is bounded primarily by data coverage. World models carry precisely that missing objective and are better grounded for it, yet rolling the future forward costs seconds per decision and rules them out of the control loop. We show the two can be separated. What a world model knows about physical scenes lives in its \emph{internal features}; generating the future is merely the objective that produced them, so the grounding can be inherited while the generative machinery is left behind. We add one feature-alignment term to ordinary VLA training: a frozen world model is run over the training frames once and cached, and the student learns to agree with that cache. No teacher is loaded during training, the projector is discarded after it, and the deployed policy is identical to the undistilled baseline, running in $32$~ms and $1.86$GB on a consumer RTX5090, so every gain is attributable to the representation rather than to added capacity or test-time compute. A $0.8$B student reaches $97.9%$ on LIBERO, improves from $48.2%$ to $50.5%$ on RoboCasa-GR1 humanoid manipulation, and the same objective carries over to real hardware, on both a single-arm and a bimanual platform. The gain survives changes of student scale, backbone, alignment layer, and teacher, indicating a broad representational prior rather than a fragile alignment between two particular networks. Project page: this https URL.

Subjects:

Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)

Cite as:

arXiv:2609.24682 [cs.RO]

(or

arXiv:2609.24682v1 [cs.RO] for this version)

https://doi.org/10.48550/arXiv.2609.24682

Focus to learn more

arXiv-issued DOI via DataCite

主题

机器人 · 计算机视觉

由「前沿雷达」于 2026-09-21 采集。正文取自原文页面，已保留出处链接。标题与正文版权归原作者所有。