Dream-1 treats robotic control as a video prediction problem, learning physical tasks from human egocentric video footage shot from a first-person perspective