README.md

---
license: other
license_name: redistributed-with-permission
license_link: https://huggingface.co/datasets/kevinLian/LoopNav
pretty_name: FAR datasets (LoopNav)
tags:
- world-model
- navigation
- minecraft
- latent
- loopnav
---

# FAR datasets (LoopNav)

Pre-computed latents of the **LoopNav** Minecraft loop-navigation benchmark, exactly as used to
train and evaluate **FAR**, a latent-diffusion world model with a learned, action-conditioned
retrieval memory.

> **Please cite LoopNav when you use this data**:
> [LoopNav: Benchmarking Spatial Consistency in World Models](https://arxiv.org/abs/2505.22976)
> (Lian, Cai, Liang and Liu, 2025). BibTeX below.

- LoopNav (original videos): https://huggingface.co/datasets/kevinLian/LoopNav
- FAR paper: [Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models](https://arxiv.org/abs/2609.34677)
- Code: https://github.com/sony/far
- Checkpoints: https://huggingface.co/1202kbs/FAR-Checkpoints
- Other FAR corpora: [AI2-THOR](https://huggingface.co/datasets/1202kbs/FAR-Datasets-AI2THOR), [SoundSpaces](https://huggingface.co/datasets/1202kbs/FAR-Datasets-SoundSpaces) (gated)

## Which LoopNav release this is

These latents were encoded from the LoopNav release as it stood in **May 2026**. On
2026-09-28 the LoopNav repository was refreshed: the ABA trajectories were re-recorded, some
ABCA trajectories were replaced, a TURN subset and official train / val / test splits were
added, and the earlier release is no longer downloadable. The clips here therefore differ from
the current `kevinLian/LoopNav`, and the train / test split below is FAR's own, not the new
official one. This repository preserves the exact corpus behind the FAR paper's LoopNav results.

## Layout

Everything is stored at the path the code expects, relative to the repository root:

```
loopnav/demo.tar, test-NNN.tar, train-NNN.tar   # tar shards; members are datasets/loopnav/latent/<route>/<world>/<radius>/<clip>.*
results/manifests/loopnav/*.json                 # manifest (loopnav_latent.json; *_split.json carries the split field)
results/indices/loopnav/*.json                   # train / test indices
datasets.json                                    # every shard and file with size, SHA-256 and clip count
```

Shards are plain uncompressed tar files: `tar -xf <shard> -C <repo>` puts the clips in place.
`demo` is a subset of `test`; `test` and `train` are disjoint.

## Download

With the code checked out (fetches the shards, verifies them and extracts them once):

```bash
python scripts/download_release.py --no-models --data demo  --corpus loopnav   # the rollout-demo / figure clips
python scripts/download_release.py --no-models --data test  --corpus loopnav   # the test split
python scripts/download_release.py --no-models --data train --corpus loopnav   # test + train
```

## Contents

19,200 first-person Minecraft clips at 20 fps: 9,600 ABA (A→B→A) and 9,600 ABCA (A→B→C→A)
out-and-back routes in village worlds of six biomes (desert, plains, savanna, snowy, taiga, zombie), with navigation-range tiers of 5, 15,
30 and 50 blocks. FAR's split: 15,360 train and 3,840 test clips.

| Tier | Clips | Shards | Size |
|---|---|---|---|
| `demo` (the 782 rollout-demo / figure clips) | 782 | `loopnav/demo.tar` (1) | 25.0 GB |
| `test` (test split) | 3,840 | `loopnav/test-000.tar`, `loopnav/test-001.tar` .. (5) | 94.5 GB |
| `train` (train split) | 15,360 | `loopnav/train-000.tar`, `loopnav/train-001.tar` .. (19) | 379.4 GB |

Per clip (`datasets/loopnav/latent/<route>/<world>/<radius>/<stamp>.*`):

| File | Shape | Contents |
|---|---|---|
| `<stamp>.npy` | (T, 16, 18, 32) float32 | latents from the frozen Oasis 500M ViT-VAE (`models/oasis_500m_vit_vae.pth`), at its tokenizer scale 0.0784 |
| `<stamp>.json` | list of T dicts | LoopNav's per-frame log, unchanged: position `x, y, z`, `yaw`, `pitch`, the `action` record, the `goal`, `frame_count` and `extra_info` (seed, world, route, range) |
| `<stamp>.keys_jepa.npy` | (T, 256) float32 | cached retriever keys used by the FAR arms |
| `<stamp>.keys_longlive.npy` | (T, 256) float32 | cached keys of the LongLive-style retrieval baseline |
| `<stamp>.keys.npy` | (T, 256) float32 | the dataset config's default key set |

## License and attribution

The LoopNav videos and logs belong to their authors; this derived corpus (latents, keys and the
unchanged per-frame logs) is **redistributed with the LoopNav authors' permission**. Please cite
LoopNav whenever you use it, and FAR if you use the latents, keys or split. The content is
Minecraft footage (Minecraft is a trademark of Mojang Synergies AB; this corpus is not affiliated
with or endorsed by Mojang or Microsoft). The latents were encoded with the Oasis 500M ViT-VAE
([open-oasis](https://github.com/etched-ai/open-oasis), MIT license).

## Citation

```bibtex
@article{lian2025loopnav,
  title   = {LoopNav: Benchmarking Spatial Consistency in World Models},
  author  = {Lian, Kewei and Cai, Shaofei and Liang, Yitao and Liu, Anji},
  journal = {arXiv preprint arXiv:2505.22976},
  year    = {2025},
  url     = {https://arxiv.org/abs/2505.22976}
}

@article{kim2026far,
  title   = {Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models},
  author  = {Kim, Beomsu and Lai, Chieh-Hsin and Nguyen, Bac and Bar, Amir and Ye, Jong Chul and Mitsufuji, Yuki},
  journal = {arXiv preprint arXiv:2609.34677},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.34677}
}
```