Distilling World-Model Features into VLAs: 0.8B Policy Hits 97.9% on LIBERO

Trung Dao · hf · 2026-09-22

VLA models map observations to actions with no objective for how the world responds, capping robustness at data coverage; world models carry that objective but cost seconds per roll-out, ruling them out of the control loop. This work shows the two can be decoupled: a world model's scene knowledge lives in its internal features; generating the future is merely the objective that shaped them. The method adds one feature-alignment term to ordinary VLA training — a frozen world model runs over training frames once and is cached, and the student learns to match the cache. No teacher during training, projector discarded after; the deployed policy is identical to the undistilled baseline, running in 32 ms with 1.86 GB on an RTX 5090.

Results: a 0.8B student reaches 97.9% on LIBERO, improves 48.2%→50.5% on RoboCasa-GR1 humanoid manipulation, and transfers to single-arm and bimanual real hardware. Gains survive changes of student scale, backbone, alignment layer, and teacher — a broad representational prior, not a fragile pairwise alignment.

Original post →

More from Embodied

Embodied channel →