RoboJEPA scales robot world models to eight billion parameters
RoboJEPA reports a fitted relationship between training compute, latent prediction error and planning performance in its evaluated tasks; its training data spans twelve robot embodiments.
RoboJEPA studies action-conditioned latent world models, with a largest predictor of 8B parameters trained on data spanning twelve robot embodiments. Its planner searches for actions whose predicted latent state approaches a goal image. The research question is whether offline world-model error can help choose a model for a planning task.
The paper fits a second-order compute relationship to prediction error on held-out DROID and RoboCasa data. It also evaluates planning in RoboCasa and on a Franka robot. A fixed V-JEPA 2.1 encoder underlies the predictor sweep, so the reported relationship is conditional on that representation and training setup. It is a measured relationship, not a guarantee of equal progress on every robot.
The real-robot tests cover Grasp, Object Lift, and Pick and Place. Appendix G.2 specifies thirty RoboJEPA episodes per model–task combination and fifty per VLA baseline. RoboJEPA receives a goal image; those baselines receive text instructions. The tested planner controls translation and the gripper while holding orientation fixed. Those distinctions matter when choosing a comparison or a new deployment.
I would start by reproducing one documented task with the released training and deployment code. Check camera inputs, action representation and goal specification before moving to different hardware. Then measure task success, failures and planning latency together with prediction error. The proxy is useful only if its relationship to the task survives that test.