In the linked TRI video they hint they are working on building generic dexterity models. If that proves possible, maybe oneshot/fewshot learning of dexterous behaviors isn't farfetched.