Weaknesses: Generally speaking, I have several doubts in the current paper. In the paper the authors explored the problem of imitation learning, based on a very restrictive assumption that the causal diagram is given. This is highly impossible in most real world applications, which renders the proposed approach not that useful on the practical side. Even if we assume that we know the causal models on the observed variables, how could we know how many hidden confounders exist among them and how hidden confounders influence among them? Apparently, hidden confounders play an extremely essential role in the proposed method in this paper.