Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies
The paper claims robot policies can inherit a world model’s scene understanding without carrying its slow future-generation machinery.
The method adds a feature-alignment term to standard VLA training, using cached internal features from a frozen world model. At deployment, the extra projector is gone and the policy matches the baseline architecture, reported at 32 ms and 1.86 GB on an RTX 5090. The authors report gains on LIBERO, RoboCasa-GR1, and real single-arm and bimanual robot setups. They argue the effect holds across student sizes, backbones, alignment layers, and teachers. HF Daily Papers' note
The method adds a feature-alignment term to standard VLA training, using cached internal features from a frozen world model. At deployment, the extra projector is gone and the policy matches the baseline architecture, reported at 32 ms and 1.86 GB on an RTX 5090. The authors report gains on LIBERO, RoboCasa-GR1, and real single-arm and bimanual robot setups. They argue the effect holds across student sizes, backbones, alignment layers, and teachers. HF Daily Papers' note
score 5