INFO:unityagents:Cannot use observations with imitation learning

[RossMelbourne]

I received the above message when I tried using Visual Observations with an imitation learning scenario. I understand the overhead of Visual Observations but are there plans to make imitation learning work with Visual Observations?

[RossMelbourne]

It does work! I set the Decision Frequency on each agent to 10 hoping that it would speed up the play during training and it did significantly. The camera attached to each agent is only 160 x 120 but it appears large enough to do imitation learning. I need to play with setting the FPS (not just -1) to see what impact it has on the training.

[Peter-Kocsis]

Hello, I'm facing with the same problem now. Could you help me, please, how did you use visual observations with imitation learning?

[RossMelbourne]

I am restating my comment from #540 here. I was able to make 'visual observations only' work but did have to pass some dummy state data because I received a crash without it. In addition to dummy state data, I was able to get imitation learning to work by using a very small visual observation width and height (120 x 160) and also skipping 10 frames before doing any learning (Decision Frequency on Agent set to 10). I set Quality Level to 1. Also, I ignored the message 'Cannot use observations with imitation learning' after examining the python code and suspecting it would work. I am assuming it is not supported because sending visual observations to python is cpu intensive but I am not sure if that is where the bottle neck might be.

[Peter-Kocsis]

Thanks's a lot! It worked for me too.

[awjuliani]

Hi all, I've made a proper fix for this issue here: #936. We will be releasing it as part of a hotfix to master in the next few days. As such, I will be closing the issue for now.

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.