Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models
Paper • 2609.34677 • Published
Accept the Matterport3D Terms of Use to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is rendered from Matterport3D scenes and is Matterport3D "Derived Information". It is provided for non-commercial academic research only, under the Matterport3D Terms of Use (https://kaldir.vc.cit.tum.de/matterport/MP_TOS.pdf). By requesting access you agree to those terms, including that you will not use the data for commercial purposes and will not redistribute it outside those terms.
Log in or Sign Up to review the conditions and access this dataset content.
FAR datasets (SoundSpaces)
Pre-computed latent corpora of audio-visual navigation episodes rendered with SoundSpaces 2.0 in Matterport3D scenes, used to train and evaluate FAR, a latent-diffusion world model with a learned, action-conditioned retrieval memory.
- Paper: Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models
- Code: https://github.com/sony/far
- Checkpoints: https://huggingface.co/1202kbs/FAR-Checkpoints
- AI2-THOR corpora: https://huggingface.co/datasets/1202kbs/FAR-Datasets
- Project page: https://1202kbs.github.io/FAR-Project-Page/
Access
Access is granted automatically once you accept the Matterport3D Terms of Use above. Then log in so the download tools can use your account:
hf auth login
Layout
Everything is stored at the path the code expects, relative to the repository root:
<corpus>/demo.tar, test-NNN.tar, train-NNN.tar # tar shards; members are datasets/<corpus>/latent/<episode>/...
datasets/<corpus>/latent/audio_feat_params.json # parameters of the audio features
results/manifests/<corpus>/*.json # dataset manifest
results/indices/<corpus>/*.json # train / test indices
datasets.json # every shard and file with size, SHA-256 and episode count
<corpus> is soundspaces (v1) or soundspaces_v2. Shards are plain uncompressed tar
files: tar -xf <shard> -C <repo> puts the episodes 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):
python scripts/download_release.py --no-models --data demo --corpus soundspaces_v2 # the rollout-demo clips
python scripts/download_release.py --no-models --data test --corpus soundspaces_v2 # the test split
python scripts/download_release.py --no-models --data train --corpus soundspaces_v2 # test + train
python scripts/download_release.py --no-models --data test --corpus soundspaces_v1 # the same for v1
Corpora
Both corpora share the same 14,425 trajectories over 81 Matterport3D scenes (13,258 train and 1,167 test; the split follows the official SoundSpaces scene splits, with its val scenes in test). An agent walks out-and-back loops (A→B→A or A→B→C→A, leg-length tiers of 5, 10, 20 and 40 m) while 2 to 4 sound sources, placed at annotated objects of the scene, play category- matched clips; it hears them binaurally. v2 adds 360° observation scans along the outbound walk (at room changes, stair landings and every 8 m), so the context holds more views of the places the return path revisits; the routes and sources are identical to v1.
| Corpus | Tier | Episodes | Shards | Size |
|---|---|---|---|---|
| SoundSpaces v2 | demo (the 120 rollout-demo / figure clips (clip_lists)) |
120 | soundspaces_v2/demo.tar (1) |
1.4 GB |
| SoundSpaces v2 | test (test split) |
1,167 | soundspaces_v2/test-000.tar (1) |
17.6 GB |
| SoundSpaces v2 | train (train split) |
13,258 | soundspaces_v2/train-000.tar, soundspaces_v2/train-001.tar .. (10) |
194.2 GB |
| SoundSpaces v1 | demo (the 120 rollout-demo / figure clips (clip_lists)) |
120 | soundspaces/demo.tar (1) |
1.1 GB |
| SoundSpaces v1 | test (test split) |
1,167 | soundspaces/test-000.tar (1) |
14.9 GB |
| SoundSpaces v1 | train (train split) |
13,258 | soundspaces/train-000.tar, soundspaces/train-001.tar .. (9) |
164.6 GB |
Per episode (datasets/<corpus>/latent/<episode>/):
| File | Shape | Contents |
|---|---|---|
latents.npy |
(T, 4, 32, 32) float16 | SDXL-VAE latents of the 256x256 RGB frames (madebyollin/sdxl-vae-fp16-fix, posterior mode, at the VAE's 0.13025 scale) |
audio.wav |
16 kHz, 2 channels, 16 bit | the binaural audio the agent hears, frame-aligned at 10 fps |
audio_feat.npy |
(4, 64, T·10+1) float16 | log-mel left / right and cos / sin interaural phase difference (audio_feat_params.json) |
latents.keys_audio.npy |
(T, 256) float32 | cached audio-retriever keys |
poses.npz |
pos (T, 3), quat (T, 4), heading, t, region (T,) |
agent pose per frame (habitat frame; the loader converts to the z-up convention) and room id |
actions.npz |
action, dpos (T, 3), dyaw |
per-frame action and motion |
legs.npz |
leg_index, phase (T,) |
which leg / phase each frame belongs to |
closures.npz |
i, j, pos_dist, heading_diff |
loop-closure frame pairs |
meta.json |
scene, seed, route, waypoints, legs, sound sources, scans, generator |
License
The frames, audio and all derived arrays are rendered from Matterport3D scenes and are Matterport3D Derived Information: use is limited to non-commercial academic research under the Matterport3D Terms of Use. The episodes were rendered with SoundSpaces 2.0 (habitat-sim with RLR audio propagation) using its semantic sound clips; please also cite SoundSpaces and Matterport3D when you use this data. Our own contributions (trajectory design, manifests, indices) are released under CC BY-NC 4.0, like the code.
Citation
@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}
}
- Downloads last month
- -