Meta V-JEPA 2

2025-06-11 · Publisher : Meta · Domain : Robotics

Meta releases V-JEPA 2, a self-supervised video world model pretrained on over one million hours of video, whose V-JEPA 2-AC variant plans robot motions without data from the target environment.

### Background On June 11, 2025, Meta released V-JEPA 2, a video model based on the JEPA (joint-embedding predictive architecture) championed by Yann LeCun. The model learns to understand and predict the world by watching unlabeled video, then adapts to robot motion planning with very little data. ### Specifications - **Pretraining**: self-supervised, on a video and image dataset including over one million hours of internet video. - **Understanding**: 77.3 top-1 accuracy on Something-Something v2 and 39.7 recall-at-5 on Epic-Kitchens-100 action anticipation (scores published by the authors). - **Robotics**: the action-conditioned V-JEPA 2-AC variant is post-trained on under 62 hours of unlabeled robot video from the DROID dataset. - **Demo**: zero-shot pick-and-place on Franka arms in two labs, planning toward a goal image, with no data collected on site and no reward. - **Openness**: code and weights released, mostly under the MIT license. ### Why it matters V-JEPA 2 argues for an alternative to generative models: predicting in representation space rather than pixel by pixel. It shows that a web-trained world model can drive a robot with very little robot-specific data. ### Key takeaway Meta joined the robotics world-model race with an open, self-supervised approach; V-JEPA 2.1 followed in March 2026.

Tags : Modèle du monde, Apprentissage auto-supervisé, Meta, Open Source