Synthetic data from world models is considered crucial for training general-purpose robots due to a shortage of useful real-world data.