RSL_RL training performance worse than jax

[MankaranSingh]

Hi! the brax based training works really well for given locomotion examples. But on training with rsl_rl, the robots mostly just learn to output vey fast actions and fall over.
Is rsl rl based training tested or the focus is more on brax? Would love if rsl_rl examples work because they contain some nice RNN based actor critic implementations.

[btaba]

Hi @MankaranSingh , I would refer to the whitepaper and the RL config, where we show RSL-RL examples for a limited set of envs.

The RSL-RL impl is there to show that we are somewhat agnostic to the RL lib. but many environments (reward config, hypers) were not really tuned specifically to RSL-RL; it was easier for us to sweep over brax hypers. If you'd like to submit better hypers for RSL-RL, please feel free!

[HaoLiuHust]

meet the same issue