RoboJEPA: Scaling Robotic Latent World Models

TL;DR

RoboJEPA is a large-scale latent world model trained on diverse real robotic data across 12 robot types, demonstrating predictable scaling of imagination accuracy with compute via a second-order power law. It enables zero-shot robotic planning on real hardware using goal images and shows that imagination error reliably predicts downstream task performance.

Why It Matters

Establishing scaling laws for robotic world models allows researchers to forecast performance gains from larger models, data, or compute, accelerating development of autonomous robots. The ability to deploy world models zero-shot on real hardware reduces reliance on task-specific training, advancing practical robotics.

Key Findings

• Imagination error in RoboJEPA scales predictably with compute following a second-order power law, enabling extrapolation beyond observed training scales.
• Downstream robotic planning performance increases consistently with compute, and imagination error strongly correlates with task success rate.
• RoboJEPA can be used as a zero-shot agent, planning toward a single goal image to complete long-horizon tasks on real robotic hardware.