Multimodal LLMs are currently the natural research step towards AGI robots that can do mundane non-creative work. I believe this is just the reality of the situation. If you can generate a video of a robot doing the dishes then your model understands the physical world quite well. That should be useful for robot control.