Meta released V-JEPA 2 and physical-reasoning benchmarks for video world-model research

The June event joined a video world model to three physical-reasoning benchmarks^{[source 1][source 2]}

On June 11, 2025, Meta introduced V-JEPA 2 as a video-trained world model and said it was releasing three benchmarks for physical reasoning. The launch describes the research goal as learning from observations, predicting how a scene may evolve and using those predictions to plan actions. It also says code and model checkpoints were available for commercial and research applications. That is a public research-artifact event. It does not show that a model has a complete understanding of a physical environment or that a planned action will be safe in a new setting.

Meta describes the system as an encoder that maps raw video to semantic embeddings and a predictor that produces expected embeddings from context. The distinction matters because the release is built around learned representations and predicted consequences rather than a direct image-to-command guarantee. A benchmark result may show progress on a defined understanding, prediction or planning task, while leaving sensors, object properties, control hardware, latency and human oversight outside the result. The June record supports a technical direction, not an operational promise.

The research record separated action-free learning from later action-conditioned work^{[source 1][source 2]}

Meta said V-JEPA 2 first used self-supervised, action-free pre-training on more than one million hours of video and one million images. It then added action-conditioned training with robot data so the predictor could account for candidate actions. This two-stage structure is important for interpreting the claim. Broad video learning can support motion and scene representations, while a later control-oriented stage gives the system information about actions. Neither stage alone turns a video model into a generally capable robot operator.

The publication says the action-conditioned version, V-JEPA 2-AC, used less than 62 hours of unlabeled robot video from the DROID dataset. It reports that the model was deployed zero-shot on Franka arms in two labs for image-goal picking and placing without collecting data in those environments. Those are bounded laboratory experiments with named hardware, tasks and goal formats. They should not be widened into a finding about homes, factories, public spaces, people nearby or systems that must recover safely from unseen failures.

Reported scores belonged to their datasets, readouts and test conditions^{[source 1][source 2]}

Meta reported motion-understanding, action-anticipation and video-question-answering results, including measurements on Something-Something v2, Epic-Kitchens-100, PerceptionTest and TempCompass. The research page provides the numerical scores and says some video-question-answering results followed alignment with a language model. Those facts are useful for locating the work in a published evaluation record. They are not portable rankings for every video task, because data distribution, readout method, prompt design and comparison system all affect the measured result.

World-model research did not amount to a certified physical agent^{[source 1][source 2]}

This article does not project later robot trials, model updates or field deployments back into June 2025. It treats Meta’s results as vendor-published research under the methods described in the source records. The historical conclusion is narrower: Meta released V-JEPA 2 with a staged video-and-action learning design, reported controlled robot-planning experiments and supplied benchmarks for further physical-reasoning evaluation. Those facts support scrutiny and replication, not a claim that physical agents were ready for unrestricted operation.