RukawaY/gs_scenes

A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting

Ziyuan Xia • Jingyi Xu • Chong Cui • Yuanhong Yu • Jiazhao Zhang • Qingsong Yan • Tao Ni
Junbo Chen • Xiaowei Zhou • Hujun Bao • Ruizhen Hu • Sida Peng

🤗 About This Dataset

This is the official GS dataset for Habitat-GS, a high-fidelity embodied navigation simulator built on 3D Gaussian Splatting and dynamic gaussian avatars. The dataset contains 129 indoor/outdoor 3DGS scenes, along with 6 gaussian avatar assets, pre-generated navigation episodes, and VLN trajectory data for StreamVLN and Uni-NaVid — everything needed to train and evaluate embodied navigation agents in high-fidelity Gaussian Splatting environments!

Key statistics:

	Train	Val	Total
Self-reconstructed scenes (scene01–scene65)	55 (scene01–scene55)	10 (scene56–scene65)	65
InteriorGS scenes (interior_*)	55	9	64
All scenes	110	19	129
PointNav episodes	110,000	1,900	111,900
ImageNav episodes	110,000	1,900	111,900
ObjectNav episodes	110,000	1,900	111,900
VLN episodes	22,000	950	22,950

Each self-reconstructed scene (scene01–scene65) comes with a foreground 3DGS render asset (<scene>.gs.ply), an optional background 3DGS asset (background.ply, can use tool script to merge with foreground GS), a collision mesh (<scene>.mesh.ply), and a navigation mesh (<scene>.navmesh). Each InteriorGS scene (interior_*) only ships 3DGS and navmesh — <scene>.gs.ply + <scene>.navmesh. The dataset also includes 6 gaussian avatars exported from AnimatableGaussians, with SMPL/SMPL-X body models for motion driving.

🏛️ Dataset Layout

The dataset is organized into five independent categories that can be downloaded separately:

	Category	Size	Required For
1	GS Scenes (train/, val/)	~27 GB	Everything — core scene assets
2	Gaussian Avatars (avatars/)	~3.1 GB	Dynamic avatar simulation
3	Habitat-Lab Nav Data (configs/, episodes/{pointnav,imagenav,objectnav}/)	~30 MB	PointNav / ImageNav / ObjectNav training & evaluation
4	StreamVLN Data (configs/, episodes/vln/, trajectory_data/vln/)	~40 GB	VLN training & evaluation (StreamVLN)
5	Uni-NaVid Data (configs/, episodes/vln/, trajectory_data/uninavid/)	~25 GB	VLN training & evaluation (Uni-NaVid)

Dataset layout:

.
├── train.scene_dataset_config.json       # Habitat scene dataset config (train)
├── val.scene_dataset_config.json         # Habitat scene dataset config (val)
│
├── train/                                # [Category 1] 110 training GS scenes (~24 GB)
│   ├── scene01/                          #   self-reconstructed (full assets)
│   │   ├── scene01.gs.ply               #     foreground 3DGS render asset
│   │   ├── background.ply               #     background 3DGS asset (sky / distant geometry; optional)
│   │   ├── scene01.mesh.ply             #     collision mesh
│   │   └── scene01.navmesh              #     navigation mesh
│   ├── scene02/ ... scene55/             #   55 self-reconstructed scenes total
│   ├── interior_0007_840137/             #   InteriorGS (only 3DGS and navmesh)
│   │   ├── interior_0007_840137.gs.ply  #     3DGS render asset
│   │   └── interior_0007_840137.navmesh #     navigation mesh
│   └── interior_0022_840117/ ... ×55     #   55 InteriorGS scenes total
│
├── val/                                  # [Category 1] 19 evaluation GS scenes (~3.3 GB)
│   ├── scene56/ ... scene65/             #   10 self-reconstructed val scenes
│   └── interior_0516_840045/ ... ×9      #   9 InteriorGS val scenes
│
├── avatars/                              # [Category 2] Gaussian avatar assets (~3.1 GB)
│   ├── avatar1/                          #   canonical gaussians of gaussian avatars
│   │   └── canonical_gs.npz
│   ├── avatar2/ ... avatar8/
│   ├── smpl/                             #   SMPL body models
│   │   ├── SMPL_FEMALE.pkl
│   │   ├── SMPL_MALE.pkl
│   │   └── SMPL_NEUTRAL.pkl
│   └── smplx/                            #   SMPL-X body models
│       ├── SMPLX_FEMALE.{npz,pkl}
│       ├── SMPLX_MALE.{npz,pkl}
│       └── SMPLX_NEUTRAL.{npz,pkl}
│
├── configs/                              # [Category 3, 4 & 5] Hydra YAML configs (~32 KB)
│   ├── ddppo_pointnav_gs_{train,eval}.yaml
│   ├── ddppo_imagenav_gs_{train,eval}.yaml
│   ├── ddppo_objectnav_gs_{train,eval}.yaml
│   ├── vln_gs_eval.yaml                  #   StreamVLN eval config (hfov=79, turn=15)
│   └── vln_uninavid_gs_eval.yaml         #   Uni-NaVid eval config (hfov=120, turn=30)
│
├── episodes/                             # [Category 3, 4 & 5] Navigation episodes (~80 MB)
│   ├── pointnav/{train,val}/             #   PointNav: 110,000 train + 1,900 val
│   ├── imagenav/{train,val}/             #   ImageNav: 110,000 train + 1,900 val
│   ├── objectnav/{train,val}/            #   ObjectNav: 110,000 train + 1,900 val
│   └── vln/{train,val}/                  #   VLN: 22,000 train + 950 val
│
└── trajectory_data/                      # [Category 4 & 5] VLN trajectory data
    ├── vln/                              #   StreamVLN trajectories (~40 GB)
    │   ├── annotations.json              #     action sequences + instructions (train)
    │   ├── annotations_val.json          #     action sequences + instructions (val)
    │   └── images/                       #     per-scene tar archives (extract before use)
    │       ├── scene01.tar               #       scene01 trajectories
    │       ├── interior_0007_840137.tar  #       interior_0007 trajectories
    │       └── ...                       #       129 per-scene archives, 22,950 trajectories total
    └── uninavid/                         #   Uni-NaVid trajectories (~25 GB)
        ├── nav_gs_train.json             #     conversation-format annotations (train)
        ├── nav_gs_val.json               #     conversation-format annotations (val)
        └── nav_videos/                   #     per-scene tar archives of .mp4 videos
            ├── scene01.tar               #       scene01 videos
            ├── interior_0007_840137.tar  #       interior_0007 videos
            └── ...                       #       129 per-scene archives, 22,950 videos total

🎒 Selective Download

You can download one or more categories using huggingface_hub's allow_patterns / ignore_patterns:

from huggingface_hub import snapshot_download

REPO = "RukawaY/gs_scenes"
LOCAL = "data/scene_datasets/gs_scenes"

# ── Download only GS scenes ──
snapshot_download(REPO, local_dir=LOCAL,
    allow_patterns=["train/**", "val/**", "*.scene_dataset_config.json"])

# ── Download GS scenes + avatars ──
snapshot_download(REPO, local_dir=LOCAL,
    allow_patterns=["train/**", "val/**", "*.scene_dataset_config.json", "avatars/**"])

# ── Download everything for Habitat-Lab navigation tasks ──
snapshot_download(REPO, local_dir=LOCAL,
    ignore_patterns=["trajectory_data/**", "avatars/**", "episodes/vln/**"])

# ── Download everything for StreamVLN ──
snapshot_download(REPO, local_dir=LOCAL,
    ignore_patterns=["avatars/**", "episodes/pointnav/**", "episodes/imagenav/**",
                     "episodes/objectnav/**", "trajectory_data/uninavid/**"])

# ── Download everything for Uni-NaVid ──
snapshot_download(REPO, local_dir=LOCAL,
    ignore_patterns=["avatars/**", "episodes/pointnav/**", "episodes/imagenav/**",
                     "episodes/objectnav/**", "trajectory_data/vln/**"])

# ── Download a few specific scenes' trajectories (StreamVLN) ──
snapshot_download(REPO, local_dir=LOCAL,
    allow_patterns=["trajectory_data/vln/annotations*.json",
                    "trajectory_data/vln/images/scene01.tar",
                    "trajectory_data/vln/images/interior_0007_840137.tar"])

# ── Download a few specific scenes' trajectories (Uni-NaVid) ──
snapshot_download(REPO, local_dir=LOCAL,
    allow_patterns=["trajectory_data/uninavid/nav_gs_*.json",
                    "trajectory_data/uninavid/nav_videos/scene01.tar",
                    "trajectory_data/uninavid/nav_videos/interior_0007_840137.tar"])

# ── Download everything (~95 GB) ──
snapshot_download(REPO, local_dir=LOCAL)

After downloading trajectory archives, extract per-scene trajectories in place:

# StreamVLN trajectories
cd data/scene_datasets/gs_scenes/trajectory_data/vln/images
for f in *.tar; do tar xf "$f" && rm "$f"; done

# Uni-NaVid trajectories
cd data/scene_datasets/gs_scenes/trajectory_data/uninavid/nav_videos
for f in *.tar; do tar xf "$f" && rm "$f"; done

🚖 Placement

Place the downloaded data under habitat-gs/data/scene_datasets/gs_scenes/ so that the directory structure matches the layout above. The Habitat configs and training/evaluation scripts in Habitat-GS expect this exact path. See the Habitat-GS README for full setup and usage instructions.

📙 Citation

If you find Habitat-GS useful in your research, please consider citing:

@misc{xia2026habitatgs,
    title={Habitat-GS: A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting}, 
    author={Ziyuan Xia and Jingyi Xu and Chong Cui and Yuanhong Yu and Jiazhao Zhang and Qingsong Yan and Tao Ni and Junbo Chen and Xiaowei Zhou and Hujun Bao and Ruizhen Hu and Sida Peng},
    year={2026},
    eprint={2604.12626},
    archivePrefix={arXiv},
    primaryClass={cs.RO},
    url={https://arxiv.org/abs/2604.12626}, 
}