Metas V-JEPA 2 Model Enhances AIs Understanding of the Physical World

On Wednesday, Meta unveiled its latest AI model, V-JEPA 2, which serves as a «world model» designed to help AI agents grasp their environment better.

V-JEPA 2 builds upon the previous model, V-JEPA, launched by Meta last year, which was trained on over a million hours of video content. This extensive data is expected to enable robots and other AI agents to navigate the physical world by understanding and predicting how concepts like gravity affect sequential events.

The model aims to establish the same kind of commonsense connections that young children and animals develop as their brains mature. For instance, when playing fetch with a dog, the pet (ideally) understands that if you throw a ball into the air, it will bounce back down, prompting it to chase after where the ball is likely to land rather than its current position.

Meta provides scenarios where a robot, holding a plate and a spatula while approaching a stovetop with cooked eggs, can anticipate that its next action will probably involve using the spatula to transfer the eggs to the plate.

According to Meta, V-JEPA 2 performs 30 times faster than Nvidia’s Cosmos model, which also aims to enhance intelligence related to the physical realm. However, it seems that Meta might be evaluating its models based on different criteria compared to Nvidia.

“We believe that world models will usher in a new era of robotics, enabling real AI agents to assist in daily tasks and perform physical activities without requiring massive amounts of training data,” explained Meta’s Chief AI Scientist, Yann LeCun, in a video presentation.

Additionally, I would like to recommend BotHub—a platform where you can test all popular models freely. Accessing the service does not require a VPN, and Russian cards can be utilized. You can obtain 100,000 free tokens for your initial tasks by following this link and start working right away!

*Please note that Meta and its products (Instagram, Facebook) are banned in the territory of the Russian Federation.

Source