[FEATURE] Goal conditioned reinforcement learning

[kkanarios32]

Please describe the purpose of the feature. Is it related to a problem?

Add support for goal conditioned reinforcement learning algorithms, such as https://arxiv.org/pdf/2206.07568

Describe the solution you'd like

I am willing to implement contrastive reinforcement learning myself, but I am hoping for some guidance from the author on how best to integrate this with the rest of the library. Namely, a standardized way we should be treating goals to allow for the most smooth integration across environments.

Describe alternatives you've considered

I have implemented my own version with a wrapper around the desired environments that returns goals within the extras of a timestep. However, I am currently very unsatisfied with the integration.

How do we know when implementation of this feature is complete?

Checklist:

• Working implementation of monte-carlo contrastive reinforcement learning algorithm

Additional context

The goal is dependent on the environment so a wrapper for each environment seems inevitable. However, there is also a need for new evaluators that take a goal and maybe other considerations.

[EdanToledo]

I think there is probably a way of handling this generically and for future features such as this, maybe there is a way to handle things with environment wrappers and something akin to functional wrappers for the eval but i'd need to think about it more.

[kkanarios32]

Sounds good. Let me know if you ever get around to it; I would be happy to work on it. Big fan of the project!

[adzcai]

Food for thought or discussion: JaxGCRL was released recently, which should provide an interesting source of fast environments and benchmarks.