[Feature Request] Reproducibility using SB3 - SubProcVecEnv

[douglas-machado-st]

🚀 Feature

Hello,

First of all, congratulations on the amazing job developing and maintaining this work.

I'm trying to achieve reproducible training using multiprocessing for a custom environment. While attempting to do so, I noticed that for SB3, it is an open question on how to accomplish this. According to the Gymnasium documentation for custom environments, when discussing reproducibility, they state:

To get reproducible sampling of actions, a seed can be set with env.action_space.seed(123).

So far, I haven't managed to find where this is set in SB3, and I'm trying my best to find a solution.

Motivation

Get reproducibility in training

Pitch

One simple flow to set seeds through the different levels of the training

Alternatives

No response

Additional context

Maybe provide in the docs all the sources of randomness so we can try to find where to seed properly

Checklist

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

[araffin]

hello,
what have you tried so far?
have you checked the sb3 doc and related issues?