https://ai.meta.com/research/vjepa/

Video Joint Embedding Predictive Architecture (V-JEPA) is a Meta AI initiative pioneering “world models” that learn to understand physical interactions by observing video rather than relying on labeled data. V-JEPA 2 scales this approach to over 1 billion parameters, enhancing AI capabilities in action anticipation and robotic planning. Explore the research at Meta AI. Introducing V-JEPA 2 – Meta AI