A naukri.com initiative
Marktechpost
4w
271
Meta AI Releases V-JEPA 2: Open-Source Self-Supervised World Models for Understanding, Prediction, and Planning
- Meta AI has introduced V-JEPA 2, an open-source self-supervised world model for visual understanding, prediction, and planning.
- V-JEPA 2 is pretrained on 1 million hours of internet-scale video and 1 million images using a visual mask denoising objective.
- The model uses data scaling, model scaling, training schedule, and spatial-temporal augmentation to achieve an 88.2% average accuracy on benchmark tasks.
- It demonstrates strong motion and appearance understanding capabilities and transferable visual features.
- V-JEPA 2 encoder shows competence in temporal reasoning tasks without language supervision during pretraining.
- V-JEPA 2-AC is an action-conditioned variant fine-tuned on robot video data, enabling zero-shot planning through model-predictive control.
- The model outperforms baselines in planning efficiency, achieving a 100% success rate on reach tasks and excelling in grasp and manipulation tasks.
- Operating with a monocular RGB camera, V-JEPA 2-AC showcases generalization capabilities for real-world applications.
- Meta's V-JEPA 2 signifies progress in self-supervised learning for physical intelligence, showcasing the potential of visual representations for perception and control.
- The research paper, models on Hugging Face, and GitHub page are available for further exploration.
- Meta AI encourages engagement through Twitter, Reddit, and their Newsletter.
Read Full Article
16 Likes
For uninterrupted reading, download the app