Distilling Internet-Scale Vision-Language Models into Embodied Agents