Researchers have introduced RoboJEPA, a latent world model for robotics built on the Joint Embedding Predictive Architecture (JEPA) and trained on a large-scale dataset spanning twelve different robotic embodiments. The work addresses a methodological gap that has slowed progress in the field: the absence of a principled way to estimate how these models' capabilities scale with size, data volume, and compute.
The central finding is that RoboJEPA's imagination error, meaning the error in its latent rollouts when simulating future states, follows a second-order power law relative to compute. This relationship allows researchers to predict model quality well beyond the scale at which the law was originally fit, offering a practical tool for planning future compute investment without needing to train ever-larger models just to verify expected gains.
The authors also show that this imagination error correlates strongly with downstream robotic planning performance, making it a reliable proxy that can substitute for costly real-robot evaluations. Another notable result is that the model can be deployed zero-shot as an autonomous robotic agent, planning toward a goal defined by a single image and successfully completing tasks that require long-horizon planning on real hardware.
At 8 billion parameters, RoboJEPA is described as the largest JEPA-based predictor model trained to date. The team is releasing all model checkpoints along with training and robot deployment code, a move that should help the broader community reproduce and extend these findings. According to the authors, this is the first work to establish scaling laws for multi-embodiment robotic world models trained on real robot data, a methodological milestone in a field where the scarcity and cost of physical data usually make such analysis difficult.