RoboJEPA: Scaling Robotic Latent World Models

RoboJE, a world model based on the Joint Embedding Predictive Architecture (JEPA), is trained on a large-scale dataset spanning 12 robotic embodiments. The paper shows that imagination error follows a second-order power law in compute, enabling prediction of model quality beyond the scale of the law. It also demonstrates that downstream robotic planning performance improves predictably with compute, and that imagination error is strongly correlated with it, making it a reliable proxy for real-robot evaluation. RoboJEPA, at 8B parameters, is the largest JEPA predictor model trained to date.