README.md

---
license: other
pretty_name: NavOL
tags:
  - robotics
  - navigation
  - imitation-learning
  - diffusion-policy
  - isaac-lab
---

# 🧭 NavOL artifacts

This dataset repository contains the model checkpoints, Dingo robot asset,
processed 50-scene training asset, and benchmark data used by NavOL. Source
code and executable data tools are maintained at
<https://github.com/WAboutMe/NavOL>.

## 🧠 Checkpoints

| File | Intended use |
|---|---|
| `models/navdp-cross-modal.ckpt` | NavDP initialization checkpoint required to start NavOL training |
| `models/checkpoints/navol-mpc-iter1000.pt` | Default NavOL inference and benchmark evaluation |
| `models/checkpoints/navol-mpc-iter500.pt` | MPC checkpoint retained for experiment reproduction |
| `models/checkpoints/navol-nompc-iter200.pt` | Non-MPC checkpoint retained for experiment reproduction |
| `models/checkpoints/navol-rollout128-iter100.pt` | 128-step rollout checkpoint |

The filenames retain their training mode and iteration and should not be
treated as interchangeable copies of one generic checkpoint.

## 🏋️ Processed training data

The canonical processed training asset is stored at:

```text
datasets/train/3d_front_scene_50/
├── index.json
├── selected.json
├── scene.glb
├── usd/
│   ├── config.yaml
│   ├── scene.usd
│   └── textures/
└── navmesh_scenes/scene_*.glb
```

`selected.json` records the 50-scene selection. The merged `scene.usd` and its
textures are loaded by Isaac Sim, while the 50 per-scene GLBs under
`navmesh_scenes/` are used by Habitat-Sim planning. `scene.glb` is retained as
a portable reconstruction source, and `usd/config.yaml` is an informational,
path-sanitized conversion record rather than a training input.

The training asset intentionally does not include `sample_100.npy`. Canonical
random training uses `sample_from_npy=False`; a fixed reset array is needed
only when that option is explicitly enabled.

## 🤖 Robot asset

`robots/dingo.usd` is the Dingo robot asset loaded by both the training and
evaluation environments. Downloading it with `--local-dir assets` places it at
the path resolved by NavOL without additional conversion.

## 🗺️ Benchmark data

| Split | Processed archive | Raw archive |
|---|---|---|
| in-domain | `data/benchmarks/processed/navol_benchmark_in_domain.zip` | `data/benchmarks/raw/raw_scenes_in_domain.zip` |
| out-of-domain | `data/benchmarks/processed/navol_benchmark_out_domain.zip` | `data/benchmarks/raw/raw_scenes_out_domain.zip` |

Each processed split contains eight scenes. Every scene includes visual GLB,
simulator USD, navigation mesh, textures, and 100 fixed start-goal tasks. Use
the processed archives for evaluation; use the raw archives to inspect or
rebuild the scene-processing pipeline.

See [BENCHMARK.md](BENCHMARK.md) or the
[Chinese guide](BENCHMARK.zh-CN.md) for extraction, validation, evaluation,
and instructions for rebuilding an incompatible USD from its included GLB.

## 📥 Download

Install the Hugging Face CLI and authenticate while this repository is private
or gated. Authentication is not required after it becomes public:

```bash
python -m pip install -U huggingface_hub
hf auth login
```

Download the initialization checkpoint, trained NavOL checkpoints, and Dingo
robot asset into a NavOL checkout:

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

Download the processed training asset into a NavOL asset root:

```bash
hf download WAboutme/NavOL \
  --repo-type dataset \
  --include "datasets/train/3d_front_scene_50/**" \
  --local-dir assets
```

Download the benchmark archives:

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

The default policy, initialization checkpoint, robot, and training data resolve
to:

```text
assets/models/checkpoints/navol-mpc-iter1000.pt
assets/models/navdp-cross-modal.ckpt
assets/robots/dingo.usd
assets/datasets/train/3d_front_scene_50/
```

## 📄 License

NavOL-authored source and configuration files use the BSD 3-Clause License.
Checkpoints, training and benchmark scenes, textures, robot assets, and other
third-party-derived artifacts retain the terms of their respective upstream
sources. The dataset repository therefore uses `license: other`. See
[LICENSE.md](LICENSE.md) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).