RoboJEPA's Uncomfortable Number: 50% Real-Robot Success After All That Scaling (2610.10515)

RoboJEPA scales a robot world model to an 8B-parameter predictor and reports clean scaling laws, but its largest model succeeds on 50% of real-robot Object Lift trials, 67% on Grasp and 27% on Pick and Place (authors' numbers, Table 3). The simple mean across those three tasks is 48% (arithmetic on the authors' table, not a figure from the paper). The scaling curves are the strong result; the robot numbers are the honest reminder of how far there is to go.

What is RoboJEPA?

RoboJEPA is an action-conditioned latent world model: given past frames, actions and robot state, it predicts the next frame's latent representation in a single forward pass. The paper, submitted 7 October 2026 by 12 authors at FAIR at Meta, Mila and Chandar Research Lab, builds on the Joint Embedding Predictive Architecture family (arXiv abstract).

The authors claim it is the first work to establish scaling laws for multi-embodiment robotic world models trained on real robot data, and that at 8B parameters it is the largest JEPA predictor trained to date. Both are the authors' claims.

How does it work?

The predictor operates on the latents of a frozen V-JEPA 2.1-G encoder and is trained with a plain l1 loss in that latent space (paper HTML). Tokens are interleaved per timestep: visual tokens for up to several camera views, then one action token and one proprioceptive-state token. The blocks use pre-norm, RoPE, RMSNorm, GELU MLPs and QK-norm, with per-embodiment linear projections for actions and state.

To act, the model plans. A planner samples action sequences, rolls them through the world model, and picks those whose final imagined latent is closest to the latent of a single goal image. The authors note that actions are sampled uniformly, with no learned policy proposal.

Source: authors' Table 6, selected rows.

What data was it trained on?

The authors report 23 public datasets across 12 platforms, about 2.87 million episodes, 15,022 video hours and 6,692 action hours (paper HTML, Table 1).

More than half the video hours come from one dataset. The real-robot evaluation uses the DROID platform, which contributes under 3% of the video hours. That is worth knowing before reading the deployment numbers.

What is the scaling law?

The authors fit imagination error, the error of the model's latent rollouts, as L(C) = A * C^(alpha - gamma ln C) + E, where C is training compute and E is irreducible loss. They call this a second-order power law and report that it holds on a real-robot DROID holdout and a RoboCasa simulated holdout (paper HTML).

The abstract says this lets them predict model quality well beyond the scale at which the law was fit, and that planning performance improves predictably with compute and correlates strongly with imagination error. If that holds up, a lab could estimate a planner's quality from a cheap offline error metric. For the broader logic of such curves, see Scaling Laws Explained and the later correction in Chinchilla vs Kaplan.

The paper also reports capability thresholds. Only models trained with more than 10^22 FLOPs (2B, 4B and 8B) solve the RoboCasa object-interaction tasks through non-greedy planning, and only models with 2B parameters or more keep the gripper closed after grasping. Smaller ones often open it and drop the object. The authors report more than 50,000 evaluation episodes across two robot platforms.

What are the real-robot results?

On a single-arm Franka with an external camera and a wrist camera, the authors evaluated three tasks using one final goal image, with the same planning hyperparameters across tasks and sizes. Progress is the mean task score; success is reaching full completion.

Values in percent, from the authors' Table 3 (paper HTML). Mean success across the three tasks, by this article's arithmetic: 8B, 48%; 4B, 37%.

Going from 4B to 8B, mean success rises from 37% to 48%. That is a real trend, and the project page says success improves with model size across the three tasks. It is also a long way from reliable.

Does the evidence hold up?

Several points limit how far to read these numbers.

• Granularity. Each cell is a success rate over a modest number of physical episodes, with counts given in the paper's appendix rather than beside the table. A handful of episodes shifts a cell by several points, so the 4B-to-8B gaps on single tasks deserve caution.
• The VLA comparison is not like for like. The authors write that goal specification differs (text versus a single image) and training differs (DROID fine-tuning for the VLAs versus pretraining on a mixture that includes DROID for RoboJEPA), so the VLA results are "contextual references rather than directly comparable baselines." RoboJEPA's goal image is also made by a predefined motion plus manual object placement, which is information a text prompt does not carry.
• The ordering flips on Pick and Place. pi0.5 reaches 53% success there but 0% on Object Lift. The authors hypothesize that the verb "lift" is underrepresented in DROID, which is plausible but untested in the paper's main text.
• Fits near saturation. The authors say that because they train multiple epochs on a fixed corpus, their fits "already sit close to data saturation."
• Frozen encoder. The laws describe a dynamics model on a fixed representation, not encoder and predictor jointly.- Replication. Everything here is authors' reporting until someone else reruns it. The released code and checkpoints make that possible.

How does it compare with related work?

The closest contrast for a working researcher is with the text-conditioned VLA policies the authors use as reference points.

The Object Lift result, where pi0.5 scores 0% and RoboJEPA-8B scores 50%, is the starkest contrast, and it comes from three tasks on one platform.

What does a working researcher take from this?

Treat imagination error as a candidate cheap proxy, not a settled one. The abstract calls it "a reliable proxy for real-robot evaluation," and that is worth testing on your own embodiment before using it to pick checkpoints.

Also note the compute thresholds. If object-interaction planning only emerges above about 10^22 FLOPs, small-scale ablations may miss behaviours that appear at scale. The released code and checkpoints let you check this directly on GitHub.

What's still unclear?

• Does the proxy hold up? Correlation between imagination error and planning quality is shown on the authors' platforms. Whether it transfers to other robots, tasks and goal formats is open.
• How much is data, how much is size? The authors say further progress will likely need more diverse interaction data rather than parameters alone.
• Can planning compete head to head? A comparison with matched goal specification and matched fine-tuning has not been reported.
• Is 50% the ceiling or the start of the curve? The trend is upward, but three tasks and two largest sizes are thin support for extrapolating to production reliability.

Basis: the abstract and main text of the arXiv HTML version, including Table 3 and the limitations; appendices only partly, and not the PDF.