Troubleshooting¶

uv run eval ... or python -c 'import torch' fails inside nix develop with:

ImportError: libstdc++.so.6: cannot open shared object file: No such file or directory

Cause A previous version of the Nix dev shell only prepended CUDA paths to LD_LIBRARY_PATH. Fresh shells could then miss the rest of the Nix userspace runtime, including libstdc++.so.6, Python shared libraries, FFmpeg, and related native dependencies.

Fix Update to the current dev shell and reload it:
```
cd ~/SIMPLE
direnv reload
```
Verify the full runtime is present:
```
uv run python - <<'PY'
import torch
print(torch.__version__)
PY
```
If you need to run nix from inside an already-loaded dev shell, prefer:
```
env -u LD_LIBRARY_PATH nix --extra-experimental-features "nix-command flakes" develop
```
This avoids making nix itself load the project runtime libraries.

If this is a brand new host, run:
```
./scripts/nix/prereq-check.sh
```
before debugging further.
nix develop itself fails before entering the shell with:

nix: …/libstdc++.so.6: version `CXXABI_1.3.15’ not found

Cause This is usually not a project dependency problem. It means the outer shell is exporting a polluted LD_LIBRARY_PATH, so the nix binary is loading the wrong libstdc++.so.6 before it can even enter the dev shell.

Fix Run nix with LD_LIBRARY_PATH cleared:
```
env -u LD_LIBRARY_PATH nix --extra-experimental-features "nix-command flakes" develop
```
For single-command entry:
```
env -u LD_LIBRARY_PATH nix --extra-experimental-features "nix-command flakes" develop -c <command>
```
To confirm the diagnosis:
```
echo "$LD_LIBRARY_PATH"
env -u LD_LIBRARY_PATH nix --version
```
If clearing LD_LIBRARY_PATH fixes it, open a fresh shell or remove the global export from your shell startup files before re-entering the repo.
Isaac Sim fails to initialize the GPU or vulkaninfo shows:

vkCreateInstance failed with ERROR_INCOMPATIBLE_DRIVER

Failed to create any GPU devices

libGLX_nvidia.so.0: cannot open shared object file

Cause The host Vulkan ICD JSON points to NVIDIA driver libraries such as libGLX_nvidia.so.0, but those driver libraries are not visible inside the dev shell runtime.

Fix The dev shell now stages required host NVIDIA driver libraries into .runtime-state/host-libcuda and prepends only that shim directory to LD_LIBRARY_PATH. Reload the shell:
```
cd ~/SIMPLE
direnv reload
```
Verify Vulkan inside the dev shell:
```
vulkaninfo --summary
```
Expected result:
- Vulkan instance creation succeeds
- the NVIDIA GPU appears in the device list
If this still fails after reload, check that the host actually provides the NVIDIA driver libraries:
```
ls -l /usr/lib/x86_64-linux-gnu/libGLX_nvidia.so.0
ls -l /usr/lib/x86_64-linux-gnu/libnvcuvid.so.1
```
uv sync failed with error:

× Failed to download isaacsim-cortex==4.5.0.0 ├─▶ Failed to extract archive: isaacsim_cortex-4.5.0.0-cp310-none-manylinux_2_34_x86_64.whl ├─▶ I/O operation failed during extraction ╰─▶ Failed to download distribution due to network timeout. Try increasing UV_HTTP_TIMEOUT (current value:30s). help: isaacsim-cortex (v4.5.0.0) was included because simple (v0.1.0) depends on isaacsim[all] (v4.5.0.0) which depends on `isaacsim-cortex

Solution set uv timeout and sync again.
```
export UV_HTTP_TIMEOUT=300
uv sync
```
The dev shell exits early with:

[simple-dev] error: LD_LIBRARY_PATH contains non-runtime paths: …

[bootstrap] PYTHONPATH must be unset inside the Nix runtime

Cause The Nix runtime now enforces a strict boundary. Only /nix/store userspace and the required host NVIDIA driver paths are allowed. Inherited host Python overrides and injected loader state are rejected because they cause non-reproducible host-specific breakage.

Fix Remove the leaking variables from your outer shell and reload:
```
unset PYTHONPATH PYTHONHOME LD_PRELOAD
direnv reload
```
To inspect the bad runtime entries before reload:
```
printf '%s\n' "${LD_LIBRARY_PATH:-}" | tr ':' '\n'
```
curobo installtion error:

RuntimeError: The detected CUDA version (11.8) mismatches the version that was used to compile PyTorch (12.8). Please make sure to use the same CUDA versions.

Solution Install cuda verison 12.8.
```
wget https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_570.86.10_linux.run
sudo sh cuda_12.8.0_570.86.10_linux.run
```
curobo installation error:

RuntimeError: CUDA error: misaligned address CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1

Solution try install curobo again.
```
TORCH_CUDA_ARCH_LIST=8.9+PTX uv pip install --no-build-isolation -e ".[isaacsim]"
```
uv pip install torch==2.7.0 --extra-index-url https://pypi.nvidia.com

❯ UV_HTTP_CONCURRENCY=1 uv pip install torch==2.7.0 –extra-index-url https://pypi.nvidia.com Resolved 25 packages in 808ms × Failed to download torch==2.7.0 ├─▶ Failed to write to the distribution cache ├─▶ error decoding response body ├─▶ request or response body error ├─▶ error reading a body from connection ╰─▶ connection reset

Solution
```
rm -rf ~/.cache/uv
# install again
```
Errors starts IsaacSim

2025-08-28 23:35:04 [8,352ms] [Error] [omni.ext.plugin] Could not load the dynamic library from /home/songlin/workspace/SIMPLE/.venv/lib/python3.10/site-packages/isaacsim/extscache/omni.iray.libs-0.0.0+d02c707b.lx64.r/bin/iray/libneuray.so. Error: libGLU.so.1: cannot open shared object file: No such file or directory (note that this may be caused by a dependent library)

Solution Install libglu
```
sudo apt-get install libglu1-mesa
```
GLFWError: (65537) b’The GLFW library is not initialized’

Usually, this happens on a headless server when running mujoco

Solution Use EGL
```
export MUJOCO_GL=egl
```
Fatal Python error: Segmentation fault when calling world.reset()

It happens when you run IsaacSim on a headless Server with headless set to False.

Solution
```
# make sure headless is True.
```
Erorr install CuRobo

fatal error: Python.h: No such file or directory 12 | #include <Python.h> | ^~~~~~~~~~

Solution
```
sudo apt-get install python3-dev
```
ImportError: libgmpxx.so.4: cannot open shared object file: No such file or directory

Solution
```
sudo apt install libgmp-dev libgmp10
```
× No solution found when resolving dependencies for split (markers: python_full_version == ‘3.10.*’ and sys_platform == ‘darwin’): ╰─▶ Because only cmake==3.25.0 is available and lerobot==0.3.3 depends on cmake>=3.29.0.1, we can conclude that lerobot==0.3.3 cannot be used. And because simple:lerobot depends on lerobot==0.3.3 and your project requires simple:lerobot, we can conclude that your project’s requirements are unsatisfiable.

Solution
```
uv sync --group lerobot --index-strategy unsafe-best-match
```
2025-10-18 22:38:03 [33,607ms] [Error] [omni.kit.app._impl] [py stderr]: TypeError: array.init() got an unexpected keyword argument ‘owner’

Solution Downgrade warp-lang
```
uv pip install "warp-lang==1.7.0" --index-strategy unsafe-best-match
```