Closed donthomasitos closed 1 year ago
Hi Thomas, thank you for the kind words and bringing this up. The Brax rollout wrapper is still very much "under construction". I am currently battling with the NeurIPS deadline, but will put together a new release with better documentation once I am done with that. I also believe that there might be something wrong with the obs normalization. I compared it with evojax's and on the ant task they start to give different performances after some generations. Will come back to you once I find the time. Best, Rob
Fixed in PR #34 see new brax notebook
Great libary, thank you for your work!
I want to add the HTML output for Brax and stumbled across a problem: For Brax's built in HTML output, it needs a list of
env_state.qp
. This can be collected at test time like:But I can't access the obs_normalizer from the evaluator, as JAX complains that data leaks a JIT'ed function. I wonder if I misunderstand the architecture - do you have a recommendation on how to implement this output? It's no problem if the
normalize_obs
is simply disabled (hence the line stays commented), but I experienced it to be beneficial in many scenarios.