---
language:
- en
license: other
license_name: scalenav-data-component-terms
license_link: https://huggingface.co/datasets/Yangyihui/ScaleNav-Data/blob/main/README.md#sources-and-license
task_categories:
- robotics
size_categories:
- 10K<n<100K
tags:
- arxiv:2610.25690
- navigation
- embodied-ai
- video
viewer: false
---

# ScaleNav Data

Prepared Internet navigation videos, instructions, recorded simulation
trajectories and evaluation selections for **ScaleNav: Scalable Internet Video
Pretraining for Language-Conditioned Local Navigation**.

[Paper](https://arxiv.org/abs/2610.25690) ·
[Code](https://github.com/MachEmbodied/ScaleNav) ·
[Model](https://huggingface.co/Yangyihui/ScaleNav) ·
[Project page](https://zju3dv.github.io/ScaleNav/) · [Collection](https://huggingface.co/collections/Yangyihui/scalenav-6ad591e8ff62bb75ec75f645)

## Choose a dataset

[Training data](#training-data) · [Evaluation data](#evaluation-data) · [Access and preparation](#access-and-preparation) · [Sources and license](#sources-and-license)

To try navigation, start with the [Habitat-GS-LN test subset](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/EVALUATION.md#test-subset).
To check the training pipeline, start with the [small training subset](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/DATA.md#small-training-subset).
Each linked guide covers the complete download and preparation process.

### Training data

| Selection and guide | Contents | Download | Prepared disk |
| --- | --- | ---: | ---: |
| [Small training subset](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/DATA.md#small-training-subset) | 17 Internet segments / 4 simulated episodes | 55.60 MB | 239.54 MB |
| [Full training data](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/DATA.md#full-training-data) | 18,249 Internet segments / 2,600 simulated episodes, including validation | 36.10 GB | 153.04 GB |

### Evaluation data

| Selection and guide | Contents | Download | Prepared disk |
| --- | --- | ---: | ---: |
| [Habitat-GS-LN test subset](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/EVALUATION.md#test-subset) | 16 tasks / 13 scenes | 991.32 MB | 2.96 GB |
| [Full Habitat-GS-LN](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/EVALUATION.md#full-habitat-gs-ln) | 157 tasks / 45 scenes | 3.17 GB | 10.36 GB |
| [Short-Horizon OVON test subset](#ovon-test-subset) | 7 tasks / 7 scenes | 642.85 MB | 642.90 MB |
| [Full Short-Horizon OVON](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/EVALUATION.md#full-short-horizon-ovon) | 524 tasks / 36 scenes | 5.84 GB | 3.67 GB |

Test subsets contain selected successful ScaleNav cases for trying the pipeline.
Use the full evaluation sets for benchmark scores.

**Download** includes all required sources; the HF sidebar counts only files hosted
in this repository. **Prepared disk** includes prepared data and retained downloads
after the guide's cleanup. Sizes use decimal MB/GB.
Full OVON needs **9.34 GB** before archive cleanup.
Model weights, environments, latent caches and run outputs are separate.
[Exact size breakdowns](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/DATA.md#choose-a-download).

## Access and preparation

### OVON test subset

This ScaleNav subset contains 7 selected Short-Horizon OVON tasks in 7 HM3D
scenes. Accept the Matterport academic-use agreement on the
[subset access page](https://huggingface.co/datasets/Yangyihui/ScaleNav-OVON-Subset).
Access is granted automatically after submitting the form. After completing
the [installation](https://github.com/MachEmbodied/ScaleNav#installation), run
from the ScaleNav repository root:

```bash
hf auth login
python scripts/prepare_data.py ovon --subset test_subset \
  --source data/upstream/evaluation/ovon \
  --scenes data/upstream/evaluation/hm3d \
  --output data/evaluation/ovon_test_subset --download
```

The command downloads only the required scene files, original navmeshes,
semantic labels, task metadata and official ABotN starting poses, verifies their
SHA-256 hashes and applies our instructions. No full HM3D archives are needed.
Continue with [OVON evaluation](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/EVALUATION.md#run-short-horizon-ovon-evaluation)
using `EVAL_SET=ovon_test_subset`.

The files are hosted in a separate access-controlled repository for the
Matterport agreement. This page is the shared entry point for all ScaleNav data.

### InteriorGS scenes

InteriorGS supports individual scene and file downloads. After obtaining
[official access](https://huggingface.co/datasets/spatialverse/InteriorGS), the
preparation commands fetch only the files needed by the selected data split.
The small training preview downloads one compressed scene
(`0003_839989`, 43,943,514 bytes), alongside the 11,660,372-byte Internet pack
and index. Habitat-GS-LN preparation likewise fetches only the required scene
assets and labels. Neither requires downloading the entire InteriorGS dataset.

Scene files remain at their official sources under the InteriorGS terms.
Our instructions, recorded trajectories and permitted Habitat-GS navmeshes
are supplied separately and combined automatically during preparation.

## Contents

- `internet/train-*.tar`, `internet/val.tar`: selected SpatialVID-derived videos
  at 512×288 / 4 Hz. `internet/train-smoke.tar` is independently downloadable.
- Each TAR contains `videos/<clip-id>.mp4`, `annotations.jsonl`, `LICENSE` and
  `SOURCE.txt`. Annotations include both canonical and paraphrased instructions,
  fixed splits, upstream archive paths and zero-based, end-exclusive frame bounds.
- `internet/videos.jsonl.gz`: per-video hashes, sizes, frame counts, source
  identifiers and full-release archive membership.
- `internet/manifest.json`: pack hashes and the associated annotation checksum.
- `annotations/internet.jsonl.gz`: 18,249 accepted Internet navigation segments.
- `annotations/simulation.jsonl.gz`, `simulation.npz`: 2,600 trajectories and
  their instructions, actions, poses and camera settings. No pickle is required.
- `annotations/evaluation/`: separate Habitat-GS-LN and Short-Horizon OVON tasks.
- `evaluation/habitatgs_ln/navmeshes/`: 45 original Habitat-GS navigation meshes
  (980,920 bytes), with [source attribution](evaluation/habitatgs_ln/SOURCE.txt).
- `annotations/prompts/`: extraction, review, refinement and paraphrasing prompts.

Internet training and validation are disjoint by upstream clip ID. Simulation
uses 618 training and 32 validation scenes, with no scene overlap; its 650 scenes
are disjoint from the 45 Habitat-GS-LN evaluation scenes. The included
`annotations/manifest.json` records fixed sample IDs and upstream revisions.

## Simulation and evaluation

Download InteriorGS scenes from the official source and replay our recorded
poses to render training video. Simulation scenes, rendered simulation videos
and latent caches are not included in these Internet video packs.

The [evaluation guide](https://github.com/MachEmbodied/ScaleNav/blob/main/docs/EVALUATION.md) provides preparation and
evaluation commands for both benchmarks. The test subsets contain selected
successful ScaleNav cases, with paper examples first; use the full sets for
benchmark scores.

The preparation command downloads the required original Habitat-GS navmeshes
from this repository and verifies their SHA-256 hashes. Full Habitat-GS-LN
reconstructs 39 scene visuals from official InteriorGS files; the pinned
Habitat-GS collection supplies the remaining six scene visuals. Short-Horizon
OVON uses unchanged original HM3D navmeshes, served in the compact subset or
obtained from the official archives, and prepares target footprints and reference
paths locally.
These are closed-loop evaluations; observations are rendered during navigation.

## Sources and license

The SpatialVID-derived videos and ScaleNav annotations use
[CC BY-NC-SA 4.0](LICENSE): attribute the original source and ScaleNav, retain
license notices, identify changes, use the data noncommercially and share
adaptations under the same license. Code uses a separate Apache-2.0 license.

- **SpatialVID-HQ**, Jiahao Wang, Yufeng Yuan and collaborators (CVPR 2026):
  [dataset](https://huggingface.co/datasets/FelixYuan-YF/SpatialVID-HQ),
  [paper](https://arxiv.org/abs/2509.09676),
  [code and license declaration](https://github.com/NJU-3DV/SpatialVID#license).
  ScaleNav selected clips, resampled them to 512×288 / 4 Hz, and
  added navigation intervals, instructions, paraphrases and fixed splits.
  See [the source and modification notice](internet/SOURCE.txt).
- **Habitat-GS**: the 45 original evaluation navmeshes are included with the
  authors' permission and retain their upstream terms. The ScaleNav annotation
  license does not relicense these files; see [source attribution](evaluation/habitatgs_ln/SOURCE.txt).
- **InteriorGS**, SpatialVerse Research Team / Manycore Tech Inc.:
  obtain [official access](https://huggingface.co/datasets/spatialverse/InteriorGS)
  and follow its [terms](https://kloudsim-usa-cos.kujiale.com/InteriorGS/InteriorGS_Terms_of_Use.pdf).
  Downloaded scene data must not be redistributed under those terms.
- **HM3D**, [HM3D-OVON](https://huggingface.co/datasets/nyokoyama/hm3d_ovon), and
  [ABot Short-Horizon OVON](https://huggingface.co/datasets/acvlab/ABotN-Short-Horizon-OVON):
  obtain source assets under their respective terms, including the
  [Matterport academic-use agreement](https://matterport.com/legal/matterport-end-user-license-agreement-academic-use-model-data).

InteriorGS scenes and geometry annotations are obtained from their original
sources. HM3D assets are hosted separately in the gated OVON subset or obtained
from the official archives, under their original terms.
The Internet videos are an independently prepared subset, not an official
SpatialVID release or endorsement. Copyright in the original footage remains
with its respective holders; the license grants only rights its licensors have
authority to grant. For attribution corrections or removal requests, provide the
affected clip IDs in this repository's Discussions or the linked code repository's
Issues. The full license includes the disclaimer of warranties.