Vision-Language Models as a Source of Rewards