Meta Introduces V-JEPA 2, a Video-Based World Model for Physical Reasoning

Meta's V-JEPA 2, a video-based world model trained on over 1 million hours of footage, enhances machine understanding of physical environments. Fine-tuned on robot data, it enables action-conditioned predictions and achieves 65-80% success rates in pick-and-place tasks with novel objects. The model excels in motion recognition and future action prediction. Meta is releasing new benchmarks, the model, code, and datasets for public use and further research.