NeurIPS 2020

Learning About Objects by Learning to Interact with Them

Meta Review

There was substantial discussion about this paper, and the reviewers are mixed. The major point of discussion was whether the model would generalize to real-world robot situations. All the reviewers agreed the learning component was very interesting and the technical approach is solid, although obscure due to the fact that the supplemental material contains key details. The main concern is that the approach makes assumptions that might only hold during simulation, and the value of this approach is minimal during a simulation because all the attributes are known to the simulation engine. The AC closely examined the paper, and agrees with the reviewers that the approach is very interesting, although perhaps unrealistic for physical robots. However, it does represent significant progress, and the conference should be accepting of bold work. Please improve the paper for the camera ready by incorporating details from the key details in supplemental material into the paper, as outlined by reviewers.