RoboJEPA: Scaling Robotic Latent World Models

TL;DR

RoboJEPA is a Meta FAIR + Stanford + Mila effort that scales the Joint Embedding Predictive Architecture into the robotics domain, pretraining on a 12-embodiment real-robot corpus up to 8B parameters — the largest JEPA predictor model trained to date. The headline methodological claim is a second-order power law that predicts latent-rollout “imagination error” from compute, with the error correlating strongly with downstream real-robot planning success and thus serving as a cheap proxy for expensive real-robot evaluation. The resulting world model is deployed zero-shot as a robotic agent, planning toward a single goal image on real hardware for long-horizon tasks. All model checkpoints, training code, and deployment code are released.

Key claims

Latent-rollout imagination error follows a second-order power law in compute and extrapolates predictably well beyond the fit range, allowing model-quality forecasting at new scales from cheap small-scale runs [Abstract].
Downstream robotic planning performance improves predictably with compute, matching the trend of imagination error [Abstract].
Imagination error is strongly correlated with downstream real-robot planning performance, making it a reliable offline proxy for real-robot evaluation [Abstract].
RoboJEPA checkpoints can be deployed zero-shot as robotic agents, planning toward a single goal image to solve tasks requiring long-horizon planning on real hardware [Abstract].
Pretraining spans 12 robotic embodiments of real-robot data; the 8B-parameter configuration is the largest JEPA predictor ever trained [Abstract].

Method

RoboJEPA builds on the Joint Embedding Predictive Architecture (JEPA) family (V-JEPA / V-JEPA 2) and extends it to the robotics regime by pretraining on a large multi-embodiment real-robot dataset. The predictor operates in latent space — rolling forward the embedding of future robot observations conditioned on actions — rather than reconstructing pixels. A single predictor is trained across all 12 embodiments to produce an embodiment-agnostic latent dynamics model.

The scaling study fits imagination error (prediction error of latent rollouts against teacher latents) as a function of compute and reports a second-order power-law form; the fit is used to project model quality to larger scales. For deployment, the trained model is used as a zero-shot planner against a single goal image — no task-specific finetuning on target hardware is required.

Results

Scale. 8B parameters — the largest JEPA predictor trained to date; pretraining data spans 12 real-robot embodiments [Abstract].
Predictability. Second-order power law in compute allows extrapolating model quality well beyond the fit range [Abstract].
Proxy quality. Imagination error strongly correlates with downstream robotic planning performance (specific coefficients not disclosed in abstract) [Abstract].
Deployment. Zero-shot long-horizon planning on real hardware from a single goal image [Abstract].

(Full numerical tables, per-embodiment breakdowns, and benchmark comparisons live in the paper PDF and are not reproduced from the abstract.)

Why it’s interesting

RoboJEPA is the first filed scaling-law entry for latent-predictive world models in the robotics regime, closing a specific gap that Neural Scaling Laws in Robotics documented empirically: only a handful of robotics papers run controlled $\geq 10\times$ parameter sweeps and essentially none at the >1B scale, so an 8B JEPA sweep with a projection-grade fit is a direct response. It extends the V-JEPA 2 line (V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning, V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning) from action-conditioned manipulation on single embodiments into cross-embodiment scaling, and gives the latent-predictive JEPA pole of the World Foundation Models debate its sharpest robotics-scale datapoint — contrasting with the generative-rollout WFM-as-policy entries like Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2 and Rho-1: Collapsing the Multimodal Stack Rho-1 that bet on video prediction rather than JEPA-style latent prediction. The “imagination error as cheap proxy for real-robot success” claim parallels Rhoda AI — Web-video pretraining scaling laws for robot policy learning (DVA)‘s DINO FD proxy result for video-pretrained policies and the pretraining-loss-vs-real-rig-success debate flagged by Empirical results fine-tuning π0.5 on a real manufacturing task.