Show notes
vla0 introduces a novel approach to converting vision-language models into vision-language-action models using normalized integer actions, ensemble prediction, and masked action augmentation on an untouched Qwen-VL-2.5-3B backbone, achieving 94.7% average success on LIBERO and strong real-robot performance without heavy pretraining. This lightweight method lowers the barrier to building capable VLAs.
Website: share.transistor.fm

