btaba

Thanks for the bug report! We'll send out a fix shortly

This should be fixed in 0825bcb74b53e36a62d50405a85e540fe2c25a95 , please let us know if that doesn't work! Closing for now

Hi @vyeevani Thanks for the proposal, indeed there is logic leaking into reset from the episode wrapper. If you changed `first_pipeline_state` to `first_state` (which would contain the first `State` object)...

Hi @wangyian-me, indeed when assert_is_replicated fails it's usually because of a NaN in training. So it looks like humanoidstandup trained with APG causes a NaN?

@queenxy are you getting NaNs on humanoidstandup with PPO with the generalized backend (and on which device)? Afaik this was tested on TPU, but would be good to know Thanks...

Hey @imoneoi , would you have a simple example that we can use to reproduce the issue you're seeing, along with a description of the expected behavior and failure mode?

Hi @jinseokbae ! The best documentation we have now are the notebooks in the root directory here: https://github.com/google/brax/tree/main/notebooks We could certainly have better documentation! We plan on releasing a major...

@evlabs-jventajas can you try again and let us know if you face any issues? just pushed some fixes

Contact impulses are not populated in the `State`, although the generalized backend populates qf_constraint. #254 references a previous version of brax RE: ETA, what are you needing the contact forces...

Hi @Kallinteris-Andreas , we have examples [here](https://github.com/google/brax/tree/main/brax/envs) loading a whole host of mujoco envs. I believe the contact is the only missing feature, but it can be added. Are you...