Challenge description clarification

It is a three class classification problem, in train set (train_labels.csv) two classes are given : NRG and RG. While the test set will contain all three classes NRG, RG and ungradable. Is it the correct interpretation? Moreover, can I use GAN (pre-trained on ImageNet or to be trained on the given train set) to cater the train set imbalance problem?

That is indeed the correct interpretation; the train set contains RG and NRG cases, while the test set contains RG, NRG and U (ungradable) cases.

We do not have any rules against using GANs, as long as the used training data for it follows the rule stated here.