environment not resetting in build - but does in editor

Hello

* Honestly feel lile mono scripts are all disabled in build ? * Start,Awake,MLAgent Initialize does not work.

So I have this project were it trains in editor but not in build. Both use the same configuration yaml.
I tried enable graphics in the build and it just looks like it does not communicate on EndEpisode.

Very strange , dont reset. What could cause such a disparity between editor and build ?

Version information:
ml-agents: 0.16.1,
ml-agents-envs: 0.16.1,
Communicator API: 1.0.0,
TensorFlow: 2.7.0

Unity Ver. Unity 2021.1.11f1 (64-bit)

// Ofc I use CPU , not gpu.
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Mon_Sep_13_20:11:50_Pacific_Daylight_Time_2021
Cuda compilation tools, release 11.5, V11.5.50
Build cuda_11.5.r11.5/compiler.30411180_0

Only "error" I get is a warning in editor that it received 0 observations - happens when I switch from physical to static. Kinda sucks.

- Tried rebuilding the agent enviroment in diffrent project;
// Surely that must mean the problem is on my side and not unity.

- Tried disabling animator completly;

- I had to see if agents script were disabled / I dont do my failstate like this but I mean
it felt like the script wasent even active. No no it is.

- Tried sending the build exe to my laptop to see it trains diffrently there . Nope same
stuff no episodes.

- Hmm looks like Start() does not run , as I have UI text that is set to state
and its the default text ? ( for debugging )

* Now get this , I cant get it to reset even when just checkin if its incline is below a certain
value in FixedUpdate - its like the script even active !? what the HECK!?

- Forcing the agent script to run via another script in the scene does not help - its
like no mono objects ? are allowed to run ?

- Added an exception to AVG for path and exe

- Update to 2021.2.7f1

- Gonna train with editor for awhile -see what happens Really stumped dont know
why builds are not working.


While training in editor console
2021-12-20 21:47:16 INFO [stats.py:111] EditorWalker: Step: 30000. Time Elapsed: 260.881 s Mean Reward: 0.094. Std of Reward: 0.282. Training.

While in build it says
No episode etc ... which makes sense since enviroment isent reset or ? even starting ?


Edit
nvidia-smi
CUDA Version: 11.5 ? // I thought I had 10 - dont higher version have problems ?

---

Here is a link to the zipped up project
https://ufile.io/p5rq8xz7 ( woah 1MB/s )
if you know any better mirrors please reply.

 

Added image reference

In editor
https://imgur.com/C8L5HNP

In build
https://imgur.com/iTqxhWu

 

Did you try building in debug mode?
Or maybe the MaxStep Field is getting set to 0 after build? That causes infinite episodes.
The project finally downloaded. I'll post back here if with what I find.

 

I will try to build in debug mode once I get home ! MaxStep Field is indeed 0 but I always kept it like that and defined maxsteps in configuration file.

Thank you for taking time out of your day to help me !

 

Hi again TimHogklint.

I tested your environment on my machine and, on my pc at least, the build version does not even connect to the
trainer script. The environment has a critical error as shown in the fig below on the first run.
The Agent Script Fails to initialize properly and just crashes.

The Agent script crashes right at Start()
This maybe why your agent seems to not end episodes. The agent never even runs.

Here is the error:
NullReferenceException: Object reference not set to an instance of an object
at BalanceAgent.Start () [0x0043e] in C:\Users\\Downloads\BipedalBalancer\BalancerAlbatross\Assets\Scripts\BalanceAgent.cs:312

You can see the debug file yourself if you make a Development build.

I'm not sure what causes this. Maybe I'll investigate if I get more time. Hope this helped.

 

I feel like such a dumb person for not testing building in debug - I guess cause I dont know how to use it - I will look it up ! Thank you so much - anything helps !

 

Welcome
Hope you can figure out what the problem is.
Best of luck!

 

It has been resolved ! Thanks to you namuh !

Apparently my sinewave targets became null at sometime during the simulation. I still dont know why - however a quick nullcheck and setting sinepoints to last valid point works

Sinetargets are set in inspector - and they were and correctly and plugged in. But I think the sinewave function might put them at impossible positions that make the transform null ?

To be honest I just dont know - they look fine and do what I want in build and editor now.

Thanks again namuh - Happy holidays and new year !

 

I'm glad I could help.
Merry Christmas!