V-JEPA 2

Building on the ability to understand and predict, V-JEPA 2 can be used for zero-shot robot planning to interact with unfamiliar objects in new environments.

We train V-JEPA 2 on 62 hours of robot data from the Droid dataset, then deploy it on a robot arm in new environments. By specifying tasks as goal images, the model accomplishes tasks like reaching, grasping, and pick-and-place. Being task-agnostic, it can be trained without extensive robot data or task-specific demonstrations.