How close are we to training egocentric robotic policies on exocentric video?