[Air-gapped Deployment] App exits due to dependency solver failure (omni.flowusd) despite disabling registry sync in offline L20 cluster

Omniverse Isaac Sim
1

[Air-gapped Deployment] App exits due to dependency solver failure (omni.flowusd) despite disabling registry sync in offline L20 cluster

Isaac Sim Version [x] 4.5.0

Operating System [x] Ubuntu 22.04

GPU Information Model: NVIDIA L20 (8-GPU Cluster)

Topic Description Detailed Description I am trying to run Isaac Sim 4.5.0 (via OmniGibson / BEHAVIOR-1K) in a strictly air-gapped (offline) GPU cluster environment. To achieve this, I pre-downloaded all assets, the extscache folder (approx. 3.5GB), and the appdata registry caches from an online machine, and transferred them to the offline cluster via OSS.

To prevent Isaac Sim from attempting to fetch extension registries online, I modified the omnigibson_4_5_0.kit file to disable all network syncs (registryEnabled = false, syncRegistry = false, etc.). However, during startup, the SimulationApp still attempts to sync with the online registry, fails to resolve the omni.flowusd dependency locally, and immediately exits, ultimately resulting in a ModuleNotFoundError: No module named 'omni.kit.usd'.

I need guidance on how to completely bypass the online dependency solver and force Isaac Sim to strictly use the local extensions provided in the --ext-folder.

Steps to Reproduce

  1. Run Isaac Sim on an online workstation to populate extscache and appdata/local/cache/Kit.

  2. Package and transfer the extscache to the offline cluster (/opt/conda/envs/behavior/lib/python3.10/site-packages/isaacsim/extscache).

  3. Package and transfer the registry indices to the offline cluster (/opt/BEHAVIOR-1K/OmniGibson/appdata/local/...).

  4. Modify omnigibson_4_5_0.kit to disable online registries:

    Ini, TOML

    [settings.exts."omni.kit.registry.nucleus"]
syncRegistry = false
registries = []

[settings.app.extensions]
registryEnabled = false
skipSyncOutdated = true

  5. Launch the application with explicit local extension folders:

    Bash

    python eval.py policy=websocket task.name=turning_on_radio \
--ext-folder /opt/conda/envs/behavior/lib/python3.10/site-packages/isaacsim/extscache \
--ext-folder /opt/conda/envs/behavior/lib/python3.10/site-packages/isaacsim/exts

Error Messages

[00:00:09.374] [INFO] [omni.kit.registry.nucleus.registry_manager] 'trusted' is not set for registry 'kit/default' in '/exts/omni.kit.registry.nucleus/registries'
2026-03-30 11:59:29 [1,152ms] [Error] [omni.kit.registry.nucleus.index_v2.index_sync] Syncing registry index failed for url: 'https://ovextensionsprod.blob.core.windows.net/exts/kit/prod/106/shared/v2'. Exception: Failed to connect to the registry
...
2026-03-30 11:59:32 [4,099ms] [Error] [omni.ext.plugin] Failed to resolve extension dependencies. Failure hints:
        [omnigibson_4_5_0-1.5.0] dependency: 'omni.flowusd' = { version='^' } can't be satisfied. Available versions:
         (none found)
...
2026-03-30 11:59:32 [4,099ms] [Error] [omni.kit.app.plugin] Exiting app because of dependency solver failure...
...
ModuleNotFoundError: No module named 'omni.kit.usd'

Additional Information What I’ve Tried

  1. Explicitly appending multiple --ext-folder arguments in the launch command to point directly to the local extscache and exts directories.

  2. Manually migrating the exact local/cache/Kit and local/data/Kit directories from the online machine to ensure registry map consistency.

  3. Checking the local extscache directory. Interestingly, omni.flowusd does not seem to be packaged inside the downloaded extscache on the online machine either.

Additional Context Is there an official script or standard procedure to package ALL required extensions (including lazy-loaded ones like omni.flowusd and omni.kit.usd) for a completely offline cluster deployment? I am running this for robotic VLA training without any internet access on the node.

3

Can you try to add below to your TOML and retry?

[settings.exts]
registryEnabled = false
4

Closing since there is no response on this thread for more than 2 weeks. Please feel free to re-open or open a new thread if you still need any support. Thanks!