Learning Plannable Representations with Causal InfoGAN [pdf] (arxiv.org)

This approach is very similar to a AAAI18 paper [1] that also finds a binary representation ("plannable representation") using VAEs and performs planning algorithms, but the difference is that they included some experiments for continuous abstract states. I personally prefer VAEs because GANs are hard to train.

[1] arxiv.org/abs/1705.00154