Aleksei Petrenko

That's fair. I remember values between 0.001 and 0.005 were the most robust for discrete-action tasks. Obviously, this whole thing does not apply to continuous-action tasks, and for them, you...

@bycn as for LSTMs in RLLIB, I never really got it to work. Sample Factory uses an implementation which handles episode resets in the middle of the trajectory quite differently....

From the learning point of view, it is fine. But it might mess up some logic that relies on counting episodes. Namely: 1) calculating per-episode rewards and other metrics (this...

Hi! Is it possible to get the whole experiment output (it should be in sf_log.txt in the experiment folder). Also, are you using M1 mac or x86?

Sorry you experienced this. I don't have a Mac at hand, but I'm 100% confident several people used SF on macOS before. I vaguely remember there was something specific about...

I am developing a new version now that should use "spawn" method of process creation uniformly on all platforms. While I believe this should fix the issue I don't have...

Looks like there are merge conflicts with your previous commit @wmFrank Can you please resolve this?

Consider also removing megaverse from CI coverage analysis because we can't run tests for this on github actions even if we had them.

Hi! Thank you for reporting this. First of all, please make sure that you executed the `setup_env` script in Vulkan SDK, i.e. following the instruction: ``` $ cd vulkansdk/1.2.162.0 $...

Oh, this is interesting! Thank you for letting us know! Looks like some package conflict then. We probably also shouldn't have precise versions of all packages specified in the conda...