Reference backend for RL literature

[misterguick]

Hi !

I'm doing research on value-based deep RL. I've been enjoying Brax very much lately ! My question is about which of the 4 backends (spring, positional, generalized, mjx) I should consider to run baselines on (i.e. SAC, TD3, DDPG). I'm not interested in comparing with raw numbers from earlier papers as I can re-train the baselines on Brax. I would like to know which backend is the most reasonable in terms of faithfulness to the usual Mujoco and whether some backends should be avoided because for some reason (too strong inaccuracy, instability, unnecessarily expensive, to be deprecated, ...) they wouldn't be suitable for RL training.

Thank you very much in advance !