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.
More from Embodied
- Designing clothes for humanoid robots: fabric serves the mechanism, not the body — yongqianme · 2026-09-22
- Inside robot data collection: First Contact adds mocap to its setup — Scobleizer · 2026-09-22
- Robot arm grasps cube 71 minutes in: model writes its own kinematics, no training — Scobleizer · 2026-09-22
- NUS's Grounded Action Model Tops Robot Manipulation Benchmarks with 3D Grounding — NationalUniversityofSingapore · 2026-09-22
- HIRO Industries Unveils Origin: A 17-DOF Dual-Arm Robot for Packing and Kitting Workstations — Scobleizer · 2026-09-22
- Stealth team unveils new robot Origin; Scobleizer bets it would sell at Home Depot — Scobleizer · 2026-09-22