video demonstration includes sensing intermediate goals, e.g., a human demonstrator searches for the apple at a few

We appreciate the insightful question. Given such demonstration data, we can train an RPN to predict these sensing goals (e.g., RPN does not assume a global planning setting. Furthermore, RPN works with an incomplete predicate set. We will revise the paper to further clarify these differences and the advantages of RPN. We will update the paper to highlight the relevance.