WallJump example giving bad results when using PPO only

Hello!
So I have been following the article below to learn more about this WallJump example

https://towardsdatascience.com/an-introduction-to-unity-ml-agents-6238452fcf4c

and in this article, they used PPO algorithm to train the agent (no curriculum or gail/bc) with slightly adjusted hyperparameters. Their results were very good.

I have tried to replicate that to see for myself if this will work that well but in my case, the results were pretty bad.


Which brings me to my question - how did they achieve such good results using PPO only? The results for SmallJump model were very similar but the BigJump is a difficult task so my results seem to be more real.
Did I miss something during this training? The article is 1.5 years old so maybe something was changed with this example?

Thank you in advance for any help/tips with this!

 

Hi @andrewcoh_unity

Thank you so much for your response! I have noticed two other things:
1. When I tried to use the models made by Unity, sometimes some of the agents on the scene are acting weird (like they cannot jump during the phase with the big wall). But not all of them, despite all of them having the same models assigned... Not sure what is going on and I don't know how to consistently recreate it.
2. This is probably something I just don't know myself but when I tried to train the agents on my own, models I created look very different, and they have different extensions.

This is the Unity one:


And this is mine (.onnx instead of .nn):


Thanks!

 

Hi there! Do you have any news on this matter?