RoboJEPA: Scaling Robotic Latent World Models
RoboJEPA reports a compute scaling law for robotic latent rollouts, tied to real planning performance.
The paper introduces a JEPA-based world model trained across 12 robot embodiments. Its latent “imagination error” follows a second-order power law in compute, which the authors say lets them predict quality beyond the fitted scale. They also report that the same error tracks downstream robotic planning, making it a proxy for expensive real-robot tests. The released model reaches 8B parameters, with checkpoints, training code, and deployment code promised. HF Daily Papers' note
The paper introduces a JEPA-based world model trained across 12 robot embodiments. Its latent “imagination error” follows a second-order power law in compute, which the authors say lets them predict quality beyond the fitted scale. They also report that the same error tracks downstream robotic planning, making it a proxy for expensive real-robot tests. The released model reaches 8B parameters, with checkpoints, training code, and deployment code promised. HF Daily Papers' note
score 6