MemVLN: Streaming Episodic Memory for Vision-Language Navigation
Paper • 2610.35066 • Published • 1
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This archive contains Matterport3D-derived observations. Read and accept https://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf before accessing the data. Use is limited to non-commercial academic purposes under those terms.
Log in or Sign Up to review the conditions and access this dataset content.
MemVLN data
Processed R2R/RxR training data, processed FGR2R sub-trajectory annotations,
VLN-CE evaluation files, and real-world demonstration recordings used by
MemVLN. The supplied vlnce_data.zip is preserved byte for byte.
Paper: arXiv:2610.35066.
Contents
The archive is 77,060,946,100 bytes (71.77 GiB), with 847,315 files and
16,974 directory entries. Extraction produces one top-level vln/ directory.
| Directory | Contents | Counts |
|---|---|---|
vln/r2r/train/ |
RGB trajectories, GT, and train_sub.json with processed FGR2R sub-instructions/sub-trajectories |
10,819 training annotations |
vln/rxr/train_v2/ |
RGB trajectories and annotations.json |
19,996 training annotations |
vln/realworld/episode_*/ |
RGB frames and trajectory_log.txt JSONL files |
24 episodes, 2,171 logged steps |
vln/r2r/val_seen/, val_unseen/ |
VLN-CE episodes and GT | 778 / 1,839 episodes |
vln/rxr/val_seen/, val_unseen/ |
Guide/follower episodes and GT; English Val-Unseen guide file | 3,669 English Val-Unseen guide episodes |
The complete annotation inventory and file counts are in archive_inventory.json. The archive contains simulation RGB observations, including Matterport3D-derived images. It does not contain the scene meshes required to run Habitat.
The processed FGR2R fields are embedded in vln/r2r/train/train_sub.json.
train_v2 is the on-disk RxR cache directory name. Preserve it when extracting.
Download and extract
Before downloading, accept the access form on the
dataset page,
then log in with hf auth login.
hf download Yannnnnnnn/MemVLN-data vlnce_data.zip SHA256SUMS \
--repo-type dataset \
--revision 2a97637c79aeef17526e8dde96334f44c4adbc65 \
--local-dir ./downloads/MemVLN-data
(cd ./downloads/MemVLN-data && shasum -a 256 -c SHA256SUMS)
mkdir -p ./datasets
unzip ./downloads/MemVLN-data/vlnce_data.zip -d ./datasets
Allow about 80 GiB free for extraction, in addition to the downloaded ZIP. The existing image files are already compressed, so the extracted payload is about the same size as the archive.
For MemVLN training:
data.annotation_base=./datasets
data.vln.realworld.root_dirs=[./datasets/vln/realworld]
For Habitat evaluation:
export MEMVLN_SCENES_DIR=/path/to/scene_datasets
export MEMVLN_R2R_DATA_ROOT="$PWD/datasets/vln/r2r"
export MEMVLN_RXR_DATA_ROOT="$PWD/datasets/vln/rxr"
export MEMVLN_RXR_GT_ROOT="$PWD/datasets/vln/rxr"
The release recipe evaluates val_unseen; its RxR episode file is
val_unseen_guide_en.json.gz, and its GT file is
val_unseen_guide_gt.json.gz. The guide GT includes all guide languages;
evaluation selects entries by episode ID. Val-Seen English conversion is not
included in this archive.
Sources and terms
See LICENSE.md for component-specific terms. Retain the source attributions when using the processed annotations.
Access and recipient information
Access requests require affirmative acceptance of the Matterport3D agreement. Hugging Face records the request and its answers. The form requests a name, academic institution, email (through the Hugging Face account), and consent to the collection and sharing described in that agreement. This information is used to administer dataset access and meet the agreement's recipient information requirements, including sharing it with Matterport as required. It is retained with the access records while access is administered or as required by the agreement. Contact the repository owner through Hugging Face for questions about an access record.
Access is automatically granted after completing and accepting the access form. File downloads require a Hugging Face account with approved access.
Related model: MemVLN.
Citation
@misc{wang2026memvln,
title = {MemVLN: Streaming Episodic Memory for Vision-Language Navigation},
author = {Wang, Yan and Huang, Jiangyong and Huang, Siyuan and Jia, Baoxiong},
year = {2026},
eprint = {2610.35066},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.35066}
}
- Downloads last month
- 10