Particle.news
Download on the App Store
6 ARTICLES
·
last wk.

Meta Debuts V-JEPA2 to Teach Robots to Predict Physical Interactions

A video-trained world model aims to let robots simulate real-world physics before executing tasks.

From delivery robots to self-driving cars, Meta’s V-JEPA 2 aims to power a new era of intelligent machines that think before they act.
Image
Meta V-JEPA 2

Overview

  • Meta released V-JEPA2 on June 11, 2025, an AI world model trained solely on raw video to learn physical interactions without labeled data.
  • The model enhances action prediction capabilities and runs up to 30× faster than Nvidia’s Cosmos model in benchmark evaluations.
  • Lab-based robots powered by V-JEPA2 have performed tasks such as picking up unfamiliar objects and placing them in new locations.
  • Meta published three open benchmarks to help researchers assess AI agents’ ability to reason and learn from video sequences.
  • The launch underscores Meta’s broader push into advanced machine intelligence and follows parallel efforts by DeepMind and World Labs.