Meta’s V-JEPA 2 model teaches AI to understand its surroundings

On Wednesday, Meta introduced its groundbreaking V-JEPA 2 AI model, designed to enhance the ability of AI agents to comprehend their surroundings. Building on the foundation laid by last year's V-JEPA model, this new iteration has been trained on an impressive dataset of over one million hours of video. The primary goal of V-JEPA 2 is to equip AI systems, including robots, with the skills to navigate and interpret the physical world. This includes grasping fundamental principles like gravity to predict outcomes in various scenarios. Meta likens this capability to the intuitive understanding that children and animals develop as they grow. For instance, when playing fetch, a dog learns to anticipate where a bouncing ball will land rather than simply following its current position. Meta illustrates the practical applications of V-JEPA 2 with scenarios involving robots. For example, if a robot is holding a plate and a spatula while approaching a stove with prepared eggs, the AI can deduce that it is likely to use the spatula to transfer the eggs onto the plate. Notably, V-JEPA 2 boasts a processing speed that is reportedly 30 times faster than Nvidia’s Cosmos model, which also focuses on enhancing AI’s understanding of the physical environment. However, Meta's performance evaluations may differ from those of Nvidia. Yann LeCun, Meta’s chief AI scientist, expressed confidence in the potential of world models like V-JEPA 2, stating that they could mark the beginning of a new era for robotics, enabling AI agents to assist with everyday tasks without the need for extensive training data.