Robot policy reaches 97.9% on LIBERO by matching a frozen world model's features
single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC
What this means for you
If you train manipulation policies, the useful part is the recipe, not a model: add one feature-alignment term against a frozen world model, run the teacher once over your frames, cache it, then discard it. The deployed policy is unchanged, so there is no new inference cost. No code release is mentioned.
A robot-control method trains a vision-language-action policy (a model that turns camera images and an instruction into motor commands) to match the internal features of a frozen world model — a network that predicts how a scene will change and is too slow for the control loop. The world model runs once over the training frames and is cached; the deployed policy is unchanged from the undistilled baseline — 32ms and 1.86GB on a consumer RTX 5090. A 0.8B student reaches 97.9% on LIBERO, and RoboCasa-GR1 humanoid manipulation improves from 48.2% to 50.5%. Gains held across student scales, backbones and teachers. The paper is a preprint with no evaluation date.
Key facts
- ·A 0.8B vision-language-action policy reaches 97.9% on the LIBERO benchmark after feature alignment with a frozen world model. source
- ·RoboCasa-GR1 humanoid manipulation improves from 48.2% to 50.5% with the same training objective. source
- ·The deployed policy runs in 32ms and uses 1.86GB on a consumer RTX 5090, identical to the undistilled baseline. source
- ·The frozen world model is run once over the training frames and cached, no teacher is loaded during training, and the projector is discarded afterwards. source
- ·The gain held across changes of student scale, backbone, alignment layer and teacher, and carried over to single-arm and bimanual real hardware. source
- ·The reported scores carry no evaluation date, and the work is an arXiv preprint (2609.24682) rather than a peer-reviewed publication. source
What the sources say
- Hugging Face Daily Papers (research) — Abstract-level report of a training trick that adds world-model feature alignment to robot policies.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersThink Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies2026-09-20