Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies
A 0.8B robot policy distilled from cached world-model features hits 97.9% on LIBERO without extra test-time compute.
The paper adds a feature-alignment term so a Vision-Language-Action policy matches cached features from a frozen world model without generating futures at decision time. No teacher is loaded in training after caching, the projector is discarded, and the deployed policy matches the undistilled baseline. A 0.8B student reaches 97.9 percent on LIBERO and improves RoboCasa-GR1 from 48.2 percent to 50.5 percent, with transfer to single-arm and bimanual real robots. It runs in about 32 ms and 1.86 GB on an RTX 5090, and gains persist across scale, backbone, alignment layer, and teacher.
- One feature-alignment loss uses features cached once from a frozen world model.
- The deployed policy is identical and loads no teacher at inference.
- A 0.8B student reaches 97.9 percent on LIBERO.
- RoboCasa-GR1 rises from 48.2 percent to 50.5 percent, with real-robot results.
- Inference is about 32 ms and 1.86 GB on an RTX 5090.
Full article239 words · extracted from huggingface.co · click to collapse
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/.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.24682