The new model from Google DeepMind is a huge step toward robots that can generalize. What will really matter in the long run?