RoboJEPA preprint reports goal-image planning on real robots

Researchers released RoboJEPA model checkpoints and training and deployment code alongside a preprint reporting that a robot world model can plan toward a single goal image on real hardware without task-specific training. The model learns to predict future states from data spanning 12 robotic embodiments. The authors say prediction error follows a second-order power law in compute and correlates strongly with robotic planning performance, offering a proxy for real-robot evaluation. They also report tasks requiring long-horizon planning on physical robots. The released model reaches 8B parameters; the scaling and deployment findings are presented in a preprint.