A scalable VLA pretraining pipeline that converts unstructured egocentric human videos into robot training data, trains a dexterous hand VLA, and fine-tunes on robot data, achieving strong zero-shot generalization and real-robot dexterous manipulation.