Question Agents for with 3 dimensional observers are too expensive to teach?

Hello, I noticed that majority of projects/examples of ml agents implementation are 2d. Even if it resembles a shooter genre, it comes down to doom-like one dimensional aiming. So I wonder, is it that hard/expensive to teach agent "lift his head up" from time to time? Do anybody have some info in "steps" on how it prolongs teaching process? Thank you in advance!

 

It's totally doable. I made a small demo a few months ago (not publicly available), where the agent could move its head up and down, and take its observation field with it.

You probably want a ray detector that scans not just sideways, but up/down too.

I have an example of such a sensor, though not built for mlagents, but you could easily adapt the concept, at https://github.com/hughperkins/peaceful-pie/blob/main/PeacefulPie/RayCasts.cs

 

Thank you! I will definitely have a look.

 

FYI. ML agents has built in 3D Ray perception sensor. Highly configurable.
I suggest to look at official examples https://github.com/Unity-Technologies/ml-agents/tree/develop/Project/Assets/ML-Agents/Examples . Mostly 3D.

 

Even more I suggest to start with Getting started section from doc, not outdated 3rd party tutorials. Here https://unity-technologies.github.io/ml-agents/Getting-Started/

 

Thank you for the insights!