RoboJEPA: Scaling Robotic Latent World Models

RoboJEPA. Scaling Robotic Latent World Models. By Artem Zholus and eleven others, from FAIR at Meta, the Chandar Research Lab, and two other institutions. Published on the seventh of October.

Latent world models have shown they can predict future states and plan in the real world. What has been missing, the authors argue, is a principled way to estimate how those capabilities scale with model size, with data, and with compute. Without one, nobody can say in advance what a bigger model or a bigger dataset will buy, and that open problem slows the field down.

The claim is that such a law can be found for robotic world models. The quantity it governs is what the authors call imagination error: the error of the model's latent rollouts, its predictions of what will happen next carried forward in its own latent space. For their model, imagination error follows a second-order power law in compute. And the law is not only descriptive. It lets the authors predict model quality well beyond the scale at which the law was fit.

The model behind the claim is RoboJEPA, a world model built on the Joint Embedding Predictive Architecture, or JEPA, and trained on a large-scale dataset spanning twelve robotic embodiments. At eight billion parameters, it is described as the largest JEPA predictor model trained to date.

Two further results tie the law to what matters on a robot. Downstream robotic planning performance also improves predictably with compute. And imagination error is strongly correlated with that planning performance, which makes it a reliable proxy for real-robot evaluation. A number that can be read off the model stands in for an experiment that would otherwise need hardware.

Finally, latent world models can be deployed zero-shot as robotic agents. Given a single goal image, RoboJEPA plans toward it and solves tasks requiring long-horizon planning on real hardware.

The scope of the claim is specific. It concerns multi-embodiment robotic world models trained on real robot data, and to the authors' knowledge this is the first work to establish scaling laws for that setting. The law is stated in compute, and the quantity it predicts directly is imagination error. Planning performance enters through its correlation with that error, and through its own predictable improvement with compute, rather than as the subject of the power law itself.

All model checkpoints are being released, together with the training and robot deployment code. Taken as a whole, the result offers a way to reason about robotic world models before building them: fit the curve on smaller runs, read off the expected imagination error at a larger scale, and use that error as a stand-in for how well the robot will plan.