A team of researchers from Meta AI, Google DeepMind, and academic institutions has published RoboJEPA, the first work to establish scaling laws for multi-embodiment robotic world models. The paper, released on arXiv, demonstrates that latent world models can predict future states and plan in the real world with predictable quality improvements as model size, data, and compute increase.
Scaling Laws for Robotic Intelligence
The Joint Embedding Predictive Architecture (JEPA) based model was trained on a massive dataset spanning 12 different robotic embodiments. The researchers discovered that RoboJEPA’s “imagination error”—the error of its latent rollouts—follows a second-order power law in compute. This allows them to accurately predict model quality well beyond the scale at which the law is fit.
“This is the first work to establish scaling laws for multi-embodiment robotic world models trained on real robot data,” the authors note. “At 8 billion parameters, RoboJEPA is the largest JEPA predictor model trained to date.”
From Simulation to Real-World Deployment
The research demonstrates that downstream robotic planning performance improves predictably with compute, and that imagination error strongly correlates with real-robot evaluation results. This makes latent world model quality a reliable proxy for actual deployment performance.
Perhaps most significantly, the researchers show that latent world models can be deployed zero-shot as robotic agents. By planning toward a single goal image, RoboJEPA can solve tasks requiring long-horizon planning on real hardware without any fine-tuning.
Implications for Robotics and AI
The ability to predict how model capabilities will scale before training offers tremendous practical value for the field. Researchers and engineers can now make informed decisions about compute budgets and model sizes for robotic applications, potentially accelerating development cycles significantly.
The team has released all model checkpoints together with their training and robot deployment code, enabling the broader research community to build on this foundation.