RoboJEPA: Scaling Robotic Latent World Models
RoboJEPA ties robot-planning gains to a measurable drop in latent “imagination” error.
The paper says RoboJEPA was trained across 12 robotic embodiments and reaches 8B parameters. Its latent rollout error follows a second-order power law in compute, which the authors use to predict model quality beyond the fitted scale. They report that the same error tracks downstream planning performance, making it a proxy for real-robot evaluation. The model is also shown zero-shot on real hardware, planning from a single goal image for long-horizon tasks. ArXiv · AI/CL/LG's note
The paper says RoboJEPA was trained across 12 robotic embodiments and reaches 8B parameters. Its latent rollout error follows a second-order power law in compute, which the authors use to predict model quality beyond the fitted scale. They report that the same error tracks downstream planning performance, making it a proxy for real-robot evaluation. The model is also shown zero-shot on real hardware, planning from a single goal image for long-horizon tasks. ArXiv · AI/CL/LG's note
score 6