Hi all,
The question I have is just to re-confirm my findings.
So if I use an environment (say from the examples folders given in the MLAgents repo) that uses ray casts as their observations, seems to be necessary to have the rendering ON (no_graphics parameters for the UnityEnvironment instantiation)?
In other words, if I turn off rendering, the received observations are all garbage or what are they? I’m running on the cloud thus I have the rendering off and I intuitively think that no rendering will make the environment less computationally expensive (is that true also?).
You CAN turn off graphics (using --no-graphics) when using raycasts (very useful for server training!). However, if you’re using Camera or RenderTexture observations, turning off rendering will cause an error.
Following what you are saying, then using Camera or RenderTextures with --no-graphics will cause an explicit runtime error (environment to stop running) or the environment will run normally and we would have implicit problems with the learning?