Prepare Before You Act: Learning From Humans to Rearrange Initial States