Meta rolled out its new V-JEPA 2 AI model Wednesday. The focus: helping AI agents understand and predict the physical world around them.
V-JEPA 2 builds on last year’s V-JEPA, which trained on over 1 million hours of video. The goal: let robots grasp concepts like gravity to better anticipate actions.
Meta gave an example: a robot holding a plate and spatula walking to a stove with cooked eggs. V-JEPA 2 predicts the logical next step—using the spatula to move eggs to the plate. 
Meta claims V-JEPA 2 runs 30x faster than Nvidia’s Cosmos model, which targets similar physical-world intelligence. But benchmark differences could impact that comparison.
Yann LeCun, Meta’s chief AI scientist, says V-JEPA 2 could cut down on massive robotic training data needs.
“We believe world models will usher a new era for robotics, enabling real-world AI agents to help with chores and physical tasks without needing astronomical amounts of robotic training data,” Yann LeCun stated.
The tech pushes Meta further into AI agents that can “think before they act” in the real world.
