VLA policies bind motor programs to absolute coordinates: injection task succeeds 99.6% on π0.5

Bryce Grant · hf · 2026-10-22

Researchers at Case Western Reserve University (Bryce Grant et al.) released a mechanistic interpretability study of Vision-Language-Action (VLA) policies.
A fine-tuned VLA can put alphabet soup in a basket, but shifting the basket 5cm left makes it drop the soup — learned motor programs are bound to absolute coordinates and largely ignore visual scene changes.
Injecting task A's activations into the action expert while the policy views task B's scene makes π0.5 execute task A's motor trajectory in 99.6% of episodes (n=1,968); the effect drops to 52.1% on SmolVLA (interleaved fusion attenuates it) and 57.0% on GR00T N1.5.
Sparse autoencoder analysis separates instruction encoding from motor generation across architectures: not all features are created equal.

More from Embodied