Recently, reinforcement learning has allowed dexterous manipulation skills with increasing complexity. Nonetheless, learning these skills in simulation still exhibits poor sample-efficiency which stems from the fact these skills are learned from scratch without the benefit of any domai