In many real-world deployment settings, demonstrations of negatives (failures) are scarce or entirely unavailable. This project asks: what is the best way of learning from only positive (successful) demonstrations without limiting the learning to behavior cloning?