Skip to content

Troubleshooting

Common issues and solutions for NavArena installation, data generation, and evaluation.

Installation

gsplat Compilation Failed

If gsplat fails to compile, check:

  1. CUDA version matches PyTorch: python -c "import torch; print(torch.version.cuda)"
  2. C++ compiler is installed: gcc --version
  3. Sufficient disk space (compilation requires temporary space)
  4. First import gsplat triggers JIT compilation which may take several minutes

CUDA Version Mismatch

Ensure the installed PyTorch version is compatible with your system CUDA driver:

nvidia-smi              # Check max CUDA version supported by driver
python -c "import torch; print(torch.cuda.is_available()); print(torch.version.cuda)"

The driver's CUDA version must be ≥ the CUDA version PyTorch was compiled with.

open3d Installation Failed

open3d only supports Python 3.8-3.11, and some systems may lack dependencies:

# Ubuntu/Debian
sudo apt-get install libgl1-mesa-glx libglib2.0-0

Node.js / npm install Failed

The Web Viewer frontend requires Node.js. Install via conda:

conda install -c conda-forge nodejs -y

If npm install fails, try clearing the cache:

cd navarena-gen/web/frontend
rm -rf node_modules package-lock.json
npm install

conda env create fails due to network issues

A common cause is GitHub-based dependencies (like CLIP) failing to download. requirements-lock.txt has separated CLIP into an optional post-install step.

For China users, try a pip mirror:

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements-lock.txt
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/

If CLIP installation fails, retry or use a proxy. CLIP is only needed for ObjectNav; PointNav/ImageNav/VLN work without it.

Permission Errors

pip install --user -e .

Data Generation

Config or scene path not found

  • Ensure NAVARENA_DATA_DIR is set: echo $NAVARENA_DATA_DIR
  • Check that scene_path in your config matches the actual directory under $NAVARENA_DATA_DIR/assets/
  • The config scene_path format is {group}/{scene_id} (e.g. x2robot/17dc3367, where group is the scene group name)

No episodes generated / empty output

  • Verify assets have completed preprocessing (manifest.json, aligned.ply, nav_map.pgm exist)
  • Check that the scene has enough navigable area; some scenes may have too few valid spawn points
  • Reduce num_episodes for a quick test (e.g. --num-episodes 5)

Evaluation

Evaluation returns empty results

  • Ensure eval_results_dir in config exists and is writable
  • Check that the dataset path in config points to valid Parquet data
  • Verify save_trajectories: true in evaluator config if you need replay videos

Agent connection failed (RemoteAgent / ViNT)

  • For RemoteAgent: Ensure the remote service is running and the URL in config is correct
  • For ViNT: Verify visualnav-transformer is cloned and diffusion_policy is installed; check that model weights are in $NAVARENA_DATA_DIR/shared/visualnav-transformer/