Coarse2Fineロバスト模倣学習

抄録

This paper proposes a robust imitation learning framework with a Coarse2Fine policy to achieve repetitive long-horizon robot tasks. We formulate the framework that employs policy robustification by disturbance injection and the Coarse2Fine policy, dividing a policy into coarse policies and fine policies to reduce the amount of training data and the subsequent demonstration burden. We verify the effectiveness of the proposed method with repetitive long-horizon tasks, and the proposed method achieves the best performance to the conventional methods.