Resource not found (ValueError), but only on consecutive runs.

[hubernikus]

When running mujoco.MjModel.from_xml_path multiple times in a row in an active python session, the loop at one stage stumbles raises aResource not found error (see console output below).

This is unexpected, as calling the function worked multiple times, with all variables being local. Only at a certain iteration did the loading of the model fail.

The error issue pops up earlier, when the model becomes more complex, e.g., having more of the same robot, or having robots with more and more complex meshes.

While it can be resolved by just restarting the console and reloading everything, in some cases this is not desired, e.g., during unit tests.

System Information: Microsoft Windows 10 Enterprise Python 3.10 / Mujoco 3.1.5 / dm_control: 1.0.19

Console Output:

Did it again 0
Did it again 1
... 
Did it again 24
Traceback (most recent call last):
  File "c:\Code\mujoco_example\example_multi_collision.py", line 29, in <module>

    example_mujoco()
  File "c:\Code\mujoco_example\example_multi_collision.py", line 22, in example_mujoco
    model = mujoco.MjModel.from_xml_path(posix_path)                            mujoco
ValueError: Error: resource not found via provider or OS filesystem: '.mujoco_cac
[mujoco_example.zip](https://github.com/google-deepmind/mujoco/files/15400644/mujoco_example.zip)
he/cube-2722594e0d5761a6def692d31d8bac29b96eef8b.obj'                          che/cube-2722594e0d5761a6def692d31d8bac29b96eef8b.obj'
PS C:\Code\mujoco_example>

Minimal example: mujoco_example.zip

[kbayes]

Thanks for reporting this. I have a very good idea where the issue is coming from, and it should hopefully be fixed soon.

[hubernikus]

Thanks so much for the fast reply!

Just an update, I realized that the issue also pops up if I add more than 255 meshes at once, i.e., in the file attached, 256 (or more) robots are added at once.

[kbayes]

Are you able to reproduce the error with the provided mujoco_example.zip? I ran the given code with the same versions of dm_control and mujoco multiple times without triggering an error, including with your suggestion of attaching 256 robots.

[yuvaltassa]

I suspect this might be a Windows-specific issue...

[hubernikus]

Yes, the error occurs with the attached files (without modification).

Could Windows specific, yes.

I'll try to find time to investigate it further.

[z-yf17]

I have the same problem. It doesn't occur every time. I wrote a for loop and I read a .xml file in every loop. It works fine at the beginning and the error occures after some loops. I don't know how to solve it.

[yuvaltassa]

@z-yf17 are you using Windows?

[z-yf17]

@z-yf17 are you using Windows?

Yes.

There is a workable but cumbersome method, which is to change all the paths, including the include paths, to absolute paths. This is troublesome, but it won't result in errors.

[kbayes]

@z-yf17 Thanks for the extra info. This is useful.

[hubernikus]

True, passing the absolute path delays the issue. I wrote a xml-post-processing to try that.

But for me that just delays the issue and the main xml cannot be found anymore, and the function call fails at now 520 runs.

Below the error and minimal example.

Traceback (most recent call last):
Attempting it= 1
Succeded it= 1
Attempting it= 2
..
Succeded it= 520
Attempting it=521
  File "c:\Code\mujoco_example\example_multi_collision.py", line 49, in <module>
    example_mujoco()
  File "c:\Code\mujoco_example\example_multi_collision.py", line 43, in example_mujoco
    model = mujoco.MjModel.from_xml_path(infile.absolute().as_posix())
ValueError: mjParseXML: resource not found via provider or OS filesystem: 'C:/Code/mujoco_example/.mujoco_cache/Scene.xml'

mujoco_example.zip

[skarbeli]

i can also reproduce the OP's example on a debian-based system after 50 iterations:

[...]
Did it again 50
Traceback (most recent call last):
  File "/home/luda/mujoco_example/example_multi_collision.py", line 27, in <module>
  File "/home/luda/mujoco_example/example_multi_collision.py", line 21, in example_mujoco
ValueError: Error: resource not found via provider or OS filesystem: '.mujoco_cache/cube-2722594e0d5761a6def692d31d8bac29b96eef8b.obj'

[hubernikus]

I tried to run the script on an Ubuntu 22.04 machine, and I surprisingly obtained inconsistent results depending on where it is run. (Same machine, same virtual environment.)

I get an error after 5 runs if I call it in the bash terminal:

The script crashes with a different error after 260 runs if I use vscode, with the run button:

The script runs without error (for large numbers of iterations), when executed via script in the vscode bash-terminal (python main.py):

[rihat99]

Faced the same issue on Ubuntu, a maximum of 32 iterations for me. As noticed by @hubernikus, the issue not appears with vscode terminal. After some time I figured out that the reason is the limit in the number of open file descriptors. By default it's higher in vscode. You can change it any terminal session with ulimit -n 10000 (Change to some high number you need). There is a way to make it permanent but it requires more steps and depends on OS. Hope it helps

[kbayes]

Before, I was never able to reproduce the issue on my system. However, I was able to find the missing fclose. Thanks for the pointer! :)