kapynResearch

REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs

REFACTOR-VLA is a research effort that learns reusable typed motor programs from vision-language-action models. It tackles monolithic VLA models that output raw motor commands, hindering long-horizon tasks. By unsupervised library learning, it groups behaviorally equivalent action sequences into reusable abstractions, boosting interpretability and multi-step performance.

Apple ML Research·Sep 2, 2026

Opening Kapyn…