The last two tech review sessions with my students featured AI-powered robot kickboxing and EagleEye - AI-powered mixed reality headsets, a tilt toward physical reasoning of AI. Reflecting on those AI gadgets has revealed that the big tech companies are bringing us closer to the era of advanced machine intelligence (AMI). This is what Meta AI's latest V-JEPA 2 model features. This incredible new model, designed to help AI think and act like humans by learning from videos, is poised to power smarter robots and unlock a world of possibilities.
Meta AI's V-JEPA 2
Meta AI's V-JEPA 2 is a state-of-the-art world model designed to enable AI agents, including robots, to better understand and interact with the physical world. Announced on June 11, 2025, this innovative model, trained on video data, is a leap toward advanced machine intelligence (AMI) by equipping AI with the ability to think before acting. Alongside this release, Meta’s Fundamental AI Research team is sharing three new benchmarks to help the global research community evaluate and advance AI’s reasoning capabilities, marking a collaborative push to accelerate progress in the field.
V-JEPA 2 is a model inspired by human intuition. V-JEPA 2 builds on the foundation laid by its predecessor, V-JEPA, which was introduced last year as Meta’s first video-trained model. The new iteration enhances AI’s understanding and predictive abilities, allowing robots and other agents to navigate unfamiliar environments and interact with objects in a manner reminiscent of human physical intuition. V-JEPA 2 mimics the human intelligence of predicting that the world will evolve based on our actions, enabling AI to anticipate outcomes and plan accordingly.
This capability is rooted in the concept of world models, which empower AI with three critical functions: understanding, predicting, and planning. For instance, just as a hockey player skates to where the puck will be rather than where it currently is, V-JEPA 2 helps AI agents make informed decisions by simulating potential scenarios. The model was trained using vast amounts of video data, allowing it to learn patterns such as how objects move, how people interact with them, and how these interactions influence the physical environment.
V-JEPA 2 will enhance Robotic and AI Interactions
One of the standout features of V-JEPA 2 is its ability to enable robots to complete tasks involving unfamiliar objects and settings. The model can analyze video inputs, showing that it can develop a nuanced understanding of physical dynamics, such as gravity’s effect on a tossed tennis ball or the momentum of a moving object. This advancement is particularly promising for applications in robotics, where precise planning and adaptability are essential. Meta envisions V-JEPA 2 as a tool that could revolutionize industries ranging from manufacturing to healthcare, where robots might assist with complex tasks requiring real-time decision-making.
The model’s training process involved ingesting and interpreting extensive video datasets, a method that has proven effective in teaching AI to recognize and predict physical interactions. This approach contrasts with traditional AI models that rely heavily on static data, offering a dynamic learning framework that mirrors human observational learning.
To complement the release of V-JEPA 2, Meta is introducing three new benchmarks designed to assess how well existing AI models reason about the physical world using video. These benchmarks aim to provide researchers with standardized tools to measure progress and identify areas for improvement. The benchmarks focus on evaluating AI’s ability to interpret video data and apply that understanding to practical scenarios, such as predicting the trajectory of moving objects or planning a robot’s path through a cluttered space. This initiative reflects Meta’s commitment to open research, as the company makes both the model and the evaluation tools available to the global scientific community. Researchers are encouraged to build on V-JEPA 2’s capabilities, fostering a collaborative environment to push the boundaries of AI technology.
The release of V-JEPA 2 and its accompanying benchmarks shows Meta’s ambition to lead in the next generation of AI research. Equipping AI with the ability to think before acting helps Meta AI to lay the groundwork for systems that can operate more safely and effectively in the physical world.