Which lib has better fit for research (PhD/Msc Thesis)?

Hello my fellows. I’m doing my research on embedded system, and I want to use RL for a routing algorithm (I have a very specific scenario where using it is interesting).

I’d like to know which lib should I use considering:

• It is a MAS (multi agent system)
• I will first try with regular DNQ and them shrinking it to fit in a embedded system
• Has a good future as a lib so I can put some effort to learn.
• I want some flexibility to use in different scenarios (since I’m a researcher)

I was taking a look on PettingZoo and TorchRL, the first seems to be the standard, but the second is in its early stages, what you guys recommend? What are your opnions? Any comments and contributions are welcome