NeurIPS 2020

MultiON: Benchmarking Semantic Map Memory using Multi-Object Navigation

Meta Review

The paper had initially received 4 mixed reviews, with two reviewers recommending acceptance and two reviewers recommending rejection. Three of the four reviews where borderline. The reviewers agreed on most points, in particular the strong experimental nature of the paper, which does not propose a new method for the targeted navigation problem, but tests the impact of ego-centric birds-eye representations and projective mapping. The reviewers also agreed that the performed experiments were interesting and important for the field of embodied vision / robotics. The question was therefore, whether the results of the experiments were strong enough and whether they provided enough insights, i.e. whether the were representative for a sufficiently large class of problems. Minor weaknesses perceived by some of the reviewers were the artifically rescricted scope of the experiments. The author's response could address several remarks and clarify misunderstandings, and the discussion between the reviewers (and the AC) lead to a more favorable assessment of the paper. The reviewers judged that the experimental results are important for the field, the AC concurs.