Apple's REFACTOR-VLA: Robots Learn Reusable Libraries of Typed Motor Programs

Apple ML Research · rss · 2026-09-02

Apple ML Research presents REFACTOR-VLA. Current VLA models (OpenVLA, π0, RT-2, RDT-1B) are monolithic—they emit raw motor commands without reusable behavioral abstractions, performing poorly on long-horizon tasks and resisting interpretation. The work learns typed motor program libraries unsupervised, tackling the core problem existing approaches like AtomicVLA and AtomSkill avoid: deciding when two action sequences are behaviorally equivalent.

Original post →

More from Embodied

Embodied channel →