Really appreciate the innovative approach you present.

However, I have a question regarding one particular aspect of the model. From my understanding, the model uses past states, current states, and predicted future states for action planning. But when I was examining the code, I noticed that during inference, the model appears to use a posterior distribution inference network, rather than a generative network with an imagination function. This left me somewhat confused.

Specifically, I would like to clarify: during the action planning process, is it based on imagined features? Or is it the case that, when minimizing the KL loss, not only does the generative network learn the posterior distribution of the inference network, but also the inference network is capable of learning the imagination function of the generative network?

I would greatly appreciate any clarification on this matter.