another amazing research from Google DeepMind: PhysWorld — "robot learning from AI-generated task videos via physical world modeling" key points from the paper: - achieves 82% task success vs 67% for the best video-only baseline - learns from AI-generated videos via physical world reconstruction - uses object-centric residual RL for physics-grounded actions - enables zero-shot manipulation without new data seems a clear move toward practical video-to-robot transfer