NavOL: Navigation Policy with Online Imitation Learning

arXiv · Models and data · 中文说明

🏆 Accepted at ICML 2026

Project page and videos

NavOL trains an embodied point-goal navigation policy through online imitation learning in Isaac Lab. A privileged NavMesh expert supplies trajectory-level supervision during simulation training; deployment uses only RGB-D observations and the goal, without a map or expert planner.

✨ Highlights

• Online imitation learning: alternate simulator rollouts and updates on states visited by the current policy.
• Safety-aware diffusion policy: sample multiple waypoint trajectories and use a shared-backbone critic to rank candidates before execution.
• Massively parallel training: the canonical recipe uses 256 environments across eight GPUs and 50 processed 3D-FRONT scenes.
• Portable release: public launchers contain no personal cluster paths; large models, robot assets, and datasets are versioned separately on Hugging Face.
• Reproducibility tools: dependency-light command generation, archive validation, smoke tests, and CI can run without starting Isaac Sim.

🧭 Method at a glance

The NavOL pipeline alternates between simulator rollout and policy update phases. During rollout, the policy maps RGB-D observations and the goal to a trajectory tracked by MPC and the low-level controller. A privileged global planner supplies expert trajectories and safety scores for the subsequent trajectory-generation and critic updates.

The expert-planning branch is used only for simulator training. The released policy does not require a NavMesh at deployment time.

Runtime status. The lightweight tests and command generation can run without Isaac Sim. Training and benchmark evaluation require a Linux workstation with a CUDA GPU, Isaac Sim 4.5, Isaac Lab 2.1, and Habitat-Sim. Those simulator workflows are not exercised by GitHub Actions.

中文使用说明

🚀 Quick start

1. Prepare the simulator environment

Install Isaac Sim 4.5 and Isaac Lab 2.1 by following the Isaac Lab installation guide. Activate the resulting Python environment, then clone NavOL and install its three local packages:

git clone https://github.com/WAboutMe/NavOL.git
cd NavOL

python -m pip install -e source/rsl_rl
python -m pip install -e source/torchinterp1d
python -m pip install -e source/navol

NavOL also needs Habitat-Sim for NavMesh planning. Install a Habitat-Sim build compatible with your CUDA/Python environment before starting Isaac Sim. The NavOL package declares ordinary Python dependencies, while Isaac Sim, Isaac Lab, PyTorch, and Habitat-Sim must be installed using their upstream instructions.

Verify the dependency-light package import:

python -c "import navol; print(navol.__version__)"

Importing navol.tasks initializes Isaac Lab task registration and must be done only inside the configured Isaac environment.

2. Choose an asset directory

Models and datasets are not stored in Git. The default asset directory is ./assets; on a shared machine, point NavOL at another directory:

export NAVOL_ASSET_ROOT=/absolute/path/to/navol-assets
mkdir -p "$NAVOL_ASSET_ROOT"

Path precedence is: an explicit CLI argument, then NAVOL_ASSET_ROOT, then the repository-local assets/ directory.

3. Download released artifacts

Install the Hugging Face CLI. Authenticate while the dataset repository is private or gated; authentication is not required after it becomes public:

python -m pip install -U huggingface_hub
hf auth login

If hf is not found after a --user installation, open a new terminal or add the Python user scripts directory to PATH. On Linux it is usually $(python -m site --user-base)/bin; on Windows it is usually the Scripts directory below the path printed by python -m site --user-base.

Released artifact	Hugging Face path	Local path below NAVOL_ASSET_ROOT	Required for
NavDP initialization	models/navdp-cross-modal.ckpt	models/navdp-cross-modal.ckpt	Training
Canonical NavOL policy	models/checkpoints/navol-mpc-iter1000.pt	models/checkpoints/navol-mpc-iter1000.pt	Evaluation
Additional NavOL policies	models/checkpoints/*.pt	models/checkpoints/*.pt	Ablations/inspection
Dingo robot	robots/dingo.usd	robots/dingo.usd	Training and evaluation
Processed 50-scene set	datasets/train/3d_front_scene_50/	datasets/train/3d_front_scene_50/	Training
Processed benchmark ZIPs	data/benchmarks/processed/	Extract to datasets/benchmarks/	Evaluation
Raw benchmark ZIPs	data/benchmarks/raw/	User-selected workspace	Optional pipeline research

Download the NavDP initialization checkpoint, four released NavOL policy checkpoints, and Dingo robot asset directly into the asset layout:

hf download WAboutme/NavOL \
  --repo-type dataset \
  --include "models/navdp-cross-modal.ckpt" \
  --include "models/checkpoints/*" \
  --include "robots/dingo.usd" \
  --local-dir "$NAVOL_ASSET_ROOT"

Download the processed 50-scene training asset directly into the same layout:

hf download WAboutme/NavOL \
  --repo-type dataset \
  --include "datasets/train/3d_front_scene_50/**" \
  --local-dir "$NAVOL_ASSET_ROOT"

Download the processed benchmark archives to a temporary download directory:

hf download WAboutme/NavOL \
  --repo-type dataset \
  --include "data/benchmarks/processed/*" \
  --local-dir downloads/navol

mkdir -p "$NAVOL_ASSET_ROOT/datasets/benchmarks"
python -m zipfile -e \
  downloads/navol/data/benchmarks/processed/navol_benchmark_in_domain.zip \
  "$NAVOL_ASSET_ROOT/datasets/benchmarks"
python -m zipfile -e \
  downloads/navol/data/benchmarks/processed/navol_benchmark_out_domain.zip \
  "$NAVOL_ASSET_ROOT/datasets/benchmarks"

After extraction, the required layout is:

navol-assets/
├── models/
│   ├── navdp-cross-modal.ckpt
│   └── checkpoints/
│       ├── navol-mpc-iter1000.pt
│       ├── navol-mpc-iter500.pt
│       ├── navol-nompc-iter200.pt
│       └── navol-rollout128-iter100.pt
├── robots/
│   └── dingo.usd
└── datasets/
    ├── train/
    │   └── 3d_front_scene_50/
    │       ├── index.json
    │       ├── selected.json
    │       ├── scene.glb
    │       ├── usd/
    │       │   ├── config.yaml
    │       │   ├── scene.usd
    │       │   └── textures/
    │       └── navmesh_scenes/scene_*.glb
    └── benchmarks/
        ├── in_domain/scene_000 ... scene_007/
        └── out_domain/scene_000 ... scene_007/

Every benchmark scene directory contains scene.glb, scene.usd, navmesh_scene.glb, sample_100.npy, and textures/. The training workflow additionally requires the 50-scene training asset shown above and models/navdp-cross-modal.ckpt. See the data guide for raw archives, validation, and USD reconstruction.

The Hugging Face repository provides the NavDP initialization checkpoint, the four .pt NavOL policy checkpoints, robots/dingo.usd, the processed 50-scene training asset, and the processed/raw benchmark archives. The commands above place these files in the paths resolved by the public launchers.

The released training asset records the canonical 50-scene selection in selected.json. It intentionally does not include sample_100.npy: canonical random training uses sample_from_npy=False. A fixed reset array is required only when that option is explicitly enabled.

4. Inspect commands before using the simulator

The dry runs print complete commands without starting Isaac Sim:

python scripts/train/train_navol.py --dry-run
python scripts/eval/evaluate_benchmark.py --dry-run

Check the repository's public release surface without reading models or data:

python scripts/check_release.py

🏋️ Training

Start the canonical eight-GPU job from the activated Isaac environment:

python scripts/train/train_navol.py

Use fewer local processes only when adapting the run to available hardware:

python scripts/train/train_navol.py --num-processes 1 --run-name navol_local

The canonical launcher explicitly supplies 8 processes, 32 environments per process, 128 rollout steps, 10 learning epochs, 16 mini-batches, a global mini-batch size of 2048, and 1000 iterations. Camera height is randomized in (0.25, 1.25) metres and pitch in (-30, 0) degrees. MPC and camera randomization are enabled. Training logs and checkpoints are written below logs/rsl_rl/dingo_pointgoal_distillation/.

See scripts/train/README.md for the complete asset contract and configuration details.

📊 Benchmark evaluation

Evaluate both public benchmark splits with the default models/checkpoints/navol-mpc-iter1000.pt checkpoint:

python scripts/eval/evaluate_benchmark.py

Evaluate one split or use an alternate output directory:

python scripts/eval/evaluate_benchmark.py \
  --split in_domain \
  --output-root results/in_domain

The launcher covers eight scenes per split with one environment and 100 episodes per scene. Results are written under the selected output root. The low-level metric implementation is in scripts/rsl_rl/eval_navdp.py; the public launcher does not reinterpret episode termination as success.

See scripts/eval/README.md for checkpoint selection, scene layout, output files, and common failures.

🧩 Data preparation

Most users should use the processed benchmark archives. They contain portable scene-relative USD texture references and fixed (100, 7) start-goal arrays.

flowchart LR
    R["Raw 3D-FRONT GLB + metadata"] --> N["Normalize axes, scale, and scene geometry"]
    N --> B["Blender material-preserving USD export"]
    B --> I["Isaac Lab collision USD conversion"]
    N --> H["Habitat-Sim NavMesh GLB"]
    H --> S["Fixed start-goal sampling"]
    I --> A["Portable processed archive"]
    H --> A
    S --> A

Loading

The public pipeline deliberately separates reusable processing primitives from dataset-specific source metadata. Processed assets are the supported path for training and evaluation; raw archives are provided for inspection and pipeline research rather than as drop-in runtime data.

If a supplied USD is incompatible with your Isaac Sim/OpenUSD build, recreate only the extracted USD while keeping the downloaded ZIP unchanged:

python scripts/data/prepare_benchmark.py convert-usd \
  "$NAVOL_ASSET_ROOT/datasets/benchmarks/in_domain/scene_000" \
  --dry-run

# Remove --dry-run after inspecting the two generated commands.

The raw in-domain and out-of-domain ZIP files are available for research and inspection. Rebuilding every processing stage from raw scenes requires Blender, Isaac Lab, Habitat-Sim, and dataset-specific source metadata; it is not required for reproducing the released benchmark evaluation. See scripts/data/README.md.

🧪 Tests

These checks do not initialize Isaac Sim:

python -m unittest discover -s tests/unit -p 'test_*.py' -v
python -m unittest discover -s tests/smoke -p 'test_*.py' -v
python -m unittest discover -s tests/data -p 'test_*.py' -v
python -m compileall -q source/navol/navol source/rsl_rl/rsl_rl scripts tests

Data tests skip automatically when benchmark assets are not installed. End-to-end simulator validation must be run separately on the target Isaac Lab/CUDA machine.

✅ Reproducibility scope

Workflow	Dependency-light CI	Isaac/CUDA machine
Package metadata and imports	✓	✓
Asset-path and command generation	✓	✓
ZIP validation and portable-path checks	✓	✓
Canonical training rollout	—	required
Full benchmark simulation	—	required
Habitat-Sim expert planning	—	required

The public launchers encode the canonical paper-scale configuration, but GPU training time and simulator results must be reproduced in a compatible Linux Isaac Lab environment.

🗂️ Repository structure

NavOL/
├── assets/                 # local models, datasets, and robot assets
├── scripts/
│   ├── data/               # portable archive build/validation/repair
│   ├── train/              # canonical training launcher
│   ├── eval/               # canonical benchmark launcher
│   ├── rsl_rl/             # low-level Isaac Lab/RSL-RL runtime
│   ├── preprocess/         # reusable scene-processing primitives
│   └── experiments/        # non-canonical research/ablation recipes
├── source/navol/           # Isaac Lab extension and `navol` package
├── source/rsl_rl/          # NavOL-compatible RSL-RL fork
├── source/torchinterp1d/   # vendored interpolation dependency
└── tests/                  # unit, smoke, and optional data tests

📝 Citation

If NavOL is useful in your research, please cite:

@inproceedings{wei2026navol,
  title     = {Nav{OL}: Navigation Policy with Online Imitation Learning},
  author    = {Xiaofei Wei and Chun Gu and Li Zhang},
  booktitle = {Forty-third International Conference on Machine Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.11762}
}

🙏 Acknowledgements

NavOL builds on Isaac Lab, Isaac Sim, Habitat-Sim, RSL-RL, NavDP, Depth Anything V2, and 3D-FRONT. We thank their authors and maintainers. Vendored code and redistributed assets retain their original terms; see the third-party notice for details.

📄 License

NavOL code is released under the BSD 3-Clause License. Vendored components retain their original licenses; see THIRD_PARTY_NOTICES.md. Model, robot, and dataset files may carry separate terms in the Hugging Face dataset card or their upstream sources. The code license does not override those asset licenses.