Skip to main content
Search

Thai Dao

1 related articles

Subscribe to updates
  1. arXiv

    Aligning to Frozen World-Model Features Lifts a 0.8B Robot Policy to 97.9% on LIBERO and +2.3 Points on RoboCasa-GR1 at Zero Inference Cost

    The work introduces a method that distills world-model representations into compact vision-language-action (VLA) policies: ordinary VLA training gains one cosine alignment term against cached features from a frozen world model, the teacher runs only once offline and is never loaded during training, the projector is discarded afterward, and the deployed policy is identical to the undistilled baseline; the 0.8B student reaches 97.9% average success across the four LIBERO suites (2.6 points above the identical undistilled student), improves RoboCasa-GR1 humanoid manipulation from 48.2% to 50.5%, and gains on both single-arm and bimanual real platforms, with the gain surviving changes of student scale, backbone, alignment layer, and teacher.

You’re all caught up