A video prediction paper. Given a series of frames, try to predict the next few ones. It's hard due to multimodality (naive solutions will just average pixels together, leading to blurriness), the super-exponential set of possibilities.
Key idea: combine benefits of VAEs and GANs, while avoiding downsides (blurriness for the former, mode collapse for the latter). Sounds cool, I wonder if there is a nice way to quantify the quality of the video. Of course this is the major challenge of any generative model. And with video prediction we don't know if the results were hand-selected or whatever. Figure 3 shows that there are some interesting differences among the methods where the generator will predict frames that blur out some of the objects touched by the robot arm.
Their method is applied to two datasets, one from the CoRL 2017 paper, and another from an older one which I'm not familiar with.
I have not gone through these other papers in detail, but it would be nice to really understand the difference between this paper and the one from Finn (NIPS 2016) and Ebert (CoRL 2017).