Meta FAIR's RoboJEPA shows robot world models improve on a predictable curve as compute grows

Researchers at Meta's FAIR lab and Mila posted RoboJEPA on 7 October 2026, a family of world models for robot arms described in an arXiv paper. A world model lets a robot imagine what happens if it takes an action, then choose the actions whose imagined result matches a goal. RoboJEPA does that imagining in the feature space of Meta's frozen V-JEPA 2.1 video encoder rather than in pixels. The team trained versions from 22 million to 8 billion parameters on 23 public manipulation datasets covering 12 robot types and 26 action formats, about 15,022 hours of video, of which 6,692 hours are paired with actions.

The main claim is a scaling law. The paper reports that imagination error, how far the model's predicted futures drift from what really happened, falls with training compute along a second-order power law. The team fit the curve on models up to 2B parameters and says it predicted the error of held-out 4B and 8B models. Planning success in RoboCasa simulation also rose with compute, and the authors say imagination error tracks it closely enough to stand in for some robot testing.

On a real Franka arm, the models planned toward a single goal image with no task-specific training. Over 30 episodes per task, the 8B model succeeded 67% of the time at grasping, 50% at lifting and 27% at pick-and-place, up from 30%, 2% and 17% for the 22M model, according to the paper's full text. The text-guided π0.5 policy, run as a reference over 50 episodes, scored 53% on pick-and-place and 0% on lifting.

Key facts

• Primary: arXiv paper 2610.10515, 7 October 2026, and its Hugging Face paper page
• Independent: AI Weekly, 8 October 2026
• Sizes: 22M to 8B parameters; the authors call 8B the largest JEPA predictor trained so far
• Data: 23 public datasets, 12 embodiments, 26 action spaces, 15,022 video hours
• Release: the paper says all checkpoints and the training and robot deployment code are released, but on 9 October the code address it lists returned a not-found page and the project site read "coming soon"

How to read it

This is a research paper, and every number comes from the authors. The scaling law is the main contribution: it gives robotics labs a way to estimate how much a bigger world model will help before paying to train it. The encoder stayed frozen, so the law covers only the prediction model on top of it. The authors say their fits sit close to data saturation, so more varied robot data, not just more parameters, is likely needed. Real-robot results rest on 30 episodes per task, and the comparison with text-guided policies is not like for like because the goal formats differ. AI Weekly's report adds no outside testing.

Independence

Frontier Flux is not affiliated with Meta, Mila, Hugging Face or AI Weekly. Details come from the paper and AI Weekly's report, not from our own testing.