“The DreamZero paper demonstrated training a joint model of state transitions and actions by initializing it with an open-source video diffusion model and fine-tuning on 500 hours of teleoperated data.”