Meta RoboJEPA: The First Scaling Law for Robot World Models?

RoboJEPA (arXiv 2610.10515, submitted 7 October 2026) trains robot world models from 22M to 8B parameters and reports that their latent rollout error follows a second-order power law in compute. The authors, from FAIR at Meta and Mila, fit the law on models up to 2B parameters and say it predicts the held-out 4B and 8B models. They describe it as the first scaling law for multi-embodiment robotic world models trained on real robot data, and the 8B model as the largest JEPA predictor trained to date. Both are the authors' claims.

We read: the full text via arXiv's HTML version. Appendices on real-robot protocol detail and alternative fits were only partly read.

What problem does this paper solve?

It tries to give robotics teams a way to forecast how good a world model will be before paying to train it. The abstract says there is no principled way to estimate how world-model capability scales with model size, data and compute, and that this slows progress.

Language models have this tool already (see Scaling Laws Explained and Chinchilla vs Kaplan). Robot world models mostly did not. Prior predictors the authors cite are small: DINO-WM has a predictor of about 20M parameters, and a 2026 DROID study used about 229M parameters on about 8,000 trajectories. RoboJEPA goes to 8B parameters and about 2.87 million trajectories.

How does RoboJEPA work?

A feed-forward transformer predicts the next latent state in one pass, in the embedding space of a frozen V-JEPA 2.1-G encoder. It is trained with a plain L1 regression loss on those latents, not on pixels. Inputs are the latent frames, actions and robot state, interleaved per timestep.

Key design choices, per the paper:

• Frozen encoder. Only the predictor is scaled. The authors say that early on V-JEPA 2.1 gave successful and predictable scaling while V-JEPA 2 did not.
• Per-embodiment projections. Each robot gets its own linear action and state projection, padded to 64 dimensions.
• Attention. Joint space, time and view attention with a frame-level causal mask and a local window of 8 frames, plus QK-norm, which the authors call important for stability.
• Loss. Half teacher forcing, half autoregressive rollout. A longer rollout (10 steps) is used in a final cooldown stage.
• Planning. To act, the model searches action sequences with the cross-entropy method (CEM), minimizing the L1 distance between its predicted latent and the latent of one goal image.

Eight sizes were trained in total (22M, 50M, 100M, 300M, 1B, 2B, 4B, 8B). The data is 23 public datasets across 12 platforms: 2.8713 million episodes, 15,022 hours of video and 6,692 hours with action labels, per the paper's Table 1. AgiBot World supplies 8,036 video hours, the nine Open-X Embodiment datasets 4,298, Roboset 1,244 and DROID 414.

What is the scaling law, and does it extrapolate?

The law is L(C) = E + A·C^(α − γ ln C), where C is training compute in FLOPs and E is irreducible loss. Because of the ln C term, ln(L − E) is quadratic in ln C, which is why the authors call it second-order. L is the per-token L1 distance between predicted and true encoder features, averaged over a 9-step latent rollout on held-out scenes.

The test is clean in design: fit on the lowest-error frontier points of the 22M to 2B models, then check against 4B and 8B, which were never in the fit.

The authors report the second-order form is roughly 2 to 3 times better than the standard power law on held-out points. The fitted irreducible error is about 0.2 on DROID and about 0.17 on RoboCasa. They read this as the models being close to saturation. All figures are the authors' own.

Which robot skills appear at which compute?

According to the authors, different manipulation skills switch on at different compute budgets, in a consistent order. Results come from over 50,000 evaluation episodes in RoboCasa simulation.

The paper states that only models trained with more than 10²² FLOPs, at sizes 2B, 4B and 8B, solve object-interaction tasks through non-greedy planning. This pattern resembles capability thresholds reported for language models, which is a reason to treat it as a hypothesis worth testing, not a settled regularity.

How does it do on a real robot?

On a Franka on the DROID platform, success rises with model size across three tasks, and RoboJEPA-8B is strongest on the simplest ones. Only a single goal image is given, with no task-specific finetuning.

Source: Table 3 of the paper, authors' numbers. RoboJEPA loses to both VLAs on Pick and Place. The authors call the VLA results "contextual references rather than directly comparable baselines": the VLAs are finetuned on DROID and take text, while RoboJEPA is pretrained on a mixture that includes DROID and takes an image.

Does the evidence hold up?

Partly. The extrapolation test is a real strength, but several limits narrow what it shows.

• Few held-out points. The law is checked on two held-out sizes, 4B and 8B, on two evaluation sets. That is a short extrapolation, from 2B to 8B parameters, a factor of four.
• Near saturation. The authors say the fits already sit close to data saturation, since training runs multiple epochs on a fixed corpus. More parameters alone will probably not keep paying off without more diverse data.
• Frozen encoder. The laws describe a dynamics model on top of a fixed representation, not encoder and predictor together. The authors also report V-JEPA 2 did not scale predictably, so the result depends on the encoder choice.
• Proxy claim. The "imagination error predicts robot performance" claim rests on a single embodiment family in the real-world test, with three tasks.
• Small real-robot sample. Real-robot results cover only three tasks, so single-task differences may be noisy. A 5-point gap should not be read as a ranking.
• Baselines. As above, the VLA comparison is context, not a controlled test.
• Review status. As of 13 October 2026 this is an arXiv preprint, with no independent replication or published critique known.

How does it compare with related work?

The closest relative is V-JEPA 2-AC from the same group. V-JEPA 2 post-trained an action-conditioned world model on less than 62 hours of unlabeled Droid robot video and deployed it zero-shot on Franka arms in two labs, using image goals. RoboJEPA asks a different question: not "does a little robot data suffice?" but "how does quality scale with a lot of it?" The RoboJEPA authors say the two are not apples-to-apples comparable, and the paper reports no head-to-head numbers.

Against pixel-space generators such as Genie, Cosmos and UniSim, the paper's contrast is that RoboJEPA predicts in frozen-encoder space. Against policy-scaling work (RT-1, RT-2, OpenVLA, π0), it scales a model of dynamics rather than a model that outputs actions. For the policy side of that story, see RT-2 Explained.

What does a working researcher take from this?

Three practical points follow from the paper as written.

1. Fit on small runs, then spend. If your setup resembles theirs, a cheap sweep from 22M to 2B gives a compute forecast. Use a form with a curvature term, since the plain power law had 2 to 3 times the held-out error here.
2. Watch capability thresholds, not only loss. Object pushing showed no success until about 10²² FLOPs, even while loss improved smoothly. Smooth loss can hide the moment a skill appears.
3. Check the encoder first. The authors found one encoder scaled predictably and another did not. Test that before committing a large run.

Code and checkpoints are released, per the authors, so the extrapolation claim can be tested by others.

What's still unclear?

The main open question is whether the law is about robots or about this encoder, corpus and recipe. The authors flag that more and more diverse interaction data, not only parameters, will likely be needed.

Other gaps:

• No text conditioning, a single image goal, and uniform action sampling with no learned proposal policy, all stated limitations.
• No evidence yet that imagination error predicts performance on other robots or on tasks beyond reaching, grasping, lifting, pushing and pick-and-place.
• The Pick and Place deficit against π0.5 is not fully explained. The authors hypothesize that lifting is underrepresented in DROID, and that is a hypothesis.
• Whether the thresholds near 10²⁰ to 10²² FLOPs transfer to other simulators and real scenes.

This post is the site's reading of the paper, not an independent test. Humanbit, which Paper Huddle's author is affiliated with, works in data labelling and evals for physical AI, a market this paper touches.