Review for NeurIPS paper: The MAGICAL Benchmark for Robust Imitation

It is not clear to me whether the proposed benchmarks are evaluating imitation learning (IL) or robust imitation learning (robust IL). The difference is the standard IL assumes that the expert data and is obtained from an MDP with exactly the same dynamics and the test MDP. Robust IL assumes that we will get a perturbed MDP at test time (where the definition of the perturbation changes depending on the meaning of "robust"). Currently, the paper seems to argue that it is testing imitation learning but is actually testing robust imitation learning. This has consequences in the experiments section.