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

Hugging Face Daily Papers Papers
world-model vla robotics distillation representation-learning vision-language-action

Summary

The paper introduces THAW-VLA, a method that distills world-model representations into Vision-Language-Action models for robotics, enhancing robustness and performance on simulation and real hardware without additional inference overhead.

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 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 RTX~5090, so every gain is attributable to the representation rather than to added capacity or test-time compute. A 0.8B 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: https://thaw-vla.trung-dt.com/.
Original Article
View Cached Full Text

Cached at: 09/22/26, 07:28 AM

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

Source: https://huggingface.co/papers/2609.24682

Abstract

Vision-Language-Action(VLA)modelsmapobservationstoactionswithnoobjectivethataccountsforhowtheworldresponds,sotheirrobustnessisboundedprimarilybydatacoverage.Worldmodelscarrypreciselythatmissingobjectiveandarebettergroundedforit,yetrollingthefutureforwardcostssecondsperdecisionandrulesthemoutofthecontrolloop.Weshowthetwocanbeseparated.Whataworldmodelknowsaboutphysicalsceneslivesinitsinternalfeatures;generatingthefutureismerelytheobjectivethatproducedthem,sothegroundingcanbeinheritedwhilethegenerativemachineryisleftbehind.Weaddonefeature-alignmenttermtoordinaryVLAtraining:afrozenworldmodelisrunoverthetrainingframesonceandcached,andthestudentlearnstoagreewiththatcache.Noteacherisloadedduringtraining,theprojectorisdiscardedafterit,andthedeployedpolicyisidenticaltotheundistilledbaseline,runningin32~msand1.86~GBonaconsumerRTX~5090,soeverygainisattributabletotherepresentationratherthantoaddedcapacityortest-timecompute.A0.8Bstudentreaches97.9%onLIBERO,improvesfrom48.2%to50.5%onRoboCasa-GR1humanoidmanipulation,andthesameobjectivecarriesovertorealhardware,onbothasingle-armandabimanualplatform.Thegainsurviveschangesofstudentscale,backbone,alignmentlayer,andteacher,indicatingabroadrepresentationalpriorratherthanafragilealignmentbetweentwoparticularnetworks.Projectpage:https://thaw-vla.trung-dt.com/.

View arXiv pageView PDFProject pageGitHub0Add to collection

Get this paper in your agent:

hf papers read 2609\.24682

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.24682 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.24682 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.24682 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Robots Need More than VLA and World Models

Hugging Face Daily Papers

This position paper argues that advancing robot intelligence requires integrating unstructured behavioral data through specialized interfaces for labeling, embodiment mapping, world modeling, and reward inference, rather than relying solely on scaling Vision-Language-Action (VLA) models and world models.