🚀 Feature

Support vanilla RNN layers (e.g., nn.RNN) in RecurrentPPO, in addition to the current LSTM-based implementation.

Motivation

In neuroscience and related fields, it's important to use simpler recurrent architectures like vanilla RNNs to maintain interpretability and align more closely with biological plausibility.

Pitch

We are using Stable-Baselines' RecurrentPPO within NeuroGym, a growing open-source toolkit widely used in the neuroscience community. Supporting vanilla RNNs would expand compatibility and usability in this domain.

Alternatives

Forking or re-implementing RecurrentPPO to support RNN layers manually, but this duplicates effort and diverges from the mainline codebase.

Additional context

NeuroGym is gaining adoption in neuroscience research, and Stable-Baselines3 is a recommended backend. Enhancing interoperability would benefit a broader user base in scientific machine learning.

Checklist

• I have checked that there is no similar issue in the repo
• If I'm requesting a new feature, I have proposed alternatives