V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

https://arxiv.org/abs/2506.09985

Over 1M hours of video and images; V-JEPA 2-AC post-trained on under 62 h of DROID robot video; zero-shot pick-and-place with image goals on Franka arms in two labs; SSv2 77.3, EK-100 R@5 39.7.

Physical-world foundation models