Meta researchers found that RoboJEPA’s prediction error fell as they scaled its action model, and that the error tracked planning performance in simulation. In physical tests on a Franka robot, its 8-billion-parameter model completed 27% of pick-and-place episodes.
RoboJEPA predicts the future visual state for candidate robot actions. A planner compares those imagined outcomes with a goal image, acts, then observes the robot again and replans.
The team scaled the action predictor from 22 million to 8 billion parameters while keeping its V-JEPA 2.1 visual encoder fixed. Training used 23 public manipulation datasets spanning 12 robot embodiments. As compute increased, prediction error fell along a second-order scaling curve. Across held-out robot and simulation data, the researchers found that lower prediction error correlated with stronger performance on simulated planning tasks.
The real-robot evaluation used a single-arm Franka setup with two cameras and one final image as the goal. In 30 episodes for each model and task, without task-specific fine-tuning, the 8B model recorded success rates of 67% for grasping, 50% for object lifting and 27% for pick-and-place. The paper defines success as full task progress.
Only the predictor was scaled; the visual encoder remained fixed, and training used a fixed dataset. The authors say the scaling fits are already close to data saturation, so more diverse robot-interaction data may be needed for further gains. They released the model checkpoints, training code and deployment code.
Sources
- RoboJEPA: Scaling Robotic Latent World Models (arXiv, October 7, 2026)
- RoboJEPA checkpoints and code
- Image: authors’ visual abstract from the paper, licensed under CC BY 4.0.