Can we fake robot training data without faking the physics?

Real robot learning requires expensive demonstrations collected via teleoperation. RoboDream generates photorealistic training data by anchoring to physically valid robot trajectories while separately synthesizing novel scenes and objects around them. This decoupling enables two tricks: reusing old robot motions in new environments, and "prop-free" teleoperation where operators gesture at empty air and the model adds objects later. Real-world experiments show synthetic data meaningfully improves downstream task performance and cuts demand for expensive real-world collection.