Sample-efficient Unsupervised Policy Cloning from Ensemble Self-supervised Labeled Videos