RoboJEPA 8B Model Ties Compute Scaling to Robot Planning

RoboJEPA, an 8B JEPA world model on 12 robot embodiments, reports compute scaling laws and zero-shot goal-image planning.

Researchers present RoboJEPA, a Joint Embedding Predictive Architecture world model trained on real-robot data spanning 12 embodiments. Both reports say its latent imagination (rollout) error follows a second-order power law in compute and correlates strongly with downstream planning on real robots, offering a proxy for hardware evaluation. At 8 billion parameters it is described as the largest JEPA predictor trained to date and can plan zero-shot from a single goal image; the later report adds that this covers long-horizon tasks and that planning quality improves predictably with compute. Checkpoints, training code, and robot deployment code are released. The sources agree on the core claims; the arXiv note (2026-10-07) restates the Hugging Face daily-papers item (2026-10-06) with slightly more specific wording on latent rollout error and long-horizon deployment, without contradicting it.