REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs

Researchers have introduced REFACTOR-VLA, a framework designed to improve vision-language-action models by converting raw motor outputs into structured, reusable library programs. By moving away from monolithic architectures that struggle with long-term tasks, this method allows models to organize behaviors into abstract, typed motor sequences. This development addresses a core limitation in robotics where current systems fail to retain learned skills across complex, multi-step actions.