We develop an inverse reinforcement learning approach that avoid undesirable demonstrations