Combine CNN-based camera input with ActorCriticRecurrent

[PanagiotisKotsaris]

I’m currently using the CNN implementation from [ #74 ] to preprocess depth-camera frames. Previously, I was feeding all observations (including images) into the ActorCriticRecurrent module.

I’d like to feed the raw camera observation into the CNN encoder, then concatenate its output embedding with my other (non-image) observations, and finally pass that combined vector through the rest of the ActorCriticRecurrent network. Is this straightforward?

[ClemensSchwarke]

Hi Panagiotis, yes that should be pretty straightforward. We are also currently working on a release that would make this even easier :)

[PanagiotisKotsaris]

Hello Clemens, thanks for the update—that’s very good news! For now I’ll continue using the CNN encoder on its own and look forward to your upcoming release.

[ClemensSchwarke]

Sounds good!