Datasetpublic

spatialgym/SpatialGym_Environmental

SpatialGym Environmental Scenes Indoor scenes for the two environmental tasks of SpatialGym — Camera Relocalization and Novel Path Discovery — with the navigation grids and task pairs precomputed, so the environments run right after download. License: CC BY-NC-SA 4.0 Overview Source Scenes Camera Relocalization pairs (d1 / d2 / d3) Novel Path Discovery pairs (d1 / d2 / d3) Download Unpacked interioragent 25 34 564 / 23 150 / 44 300 3 723 / 4 042 / 7 767… See the full description on the dataset page: https://huggingface.co/datasets/spatialgym/SpatialGym_Environmental.

sourceHugging Facecc-by-nc-sa-4.0updated 19d agoView on Hugging Face

0likes713downloads

Dataset Card

SpatialGym Environmental Scenes

Indoor scenes for the two environmental tasks of SpatialGym — Camera Relocalization and Novel Path Discovery — with the navigation grids and task pairs precomputed, so the environments run right after download. License: CC BY-NC-SA 4.0

Overview

Source	Scenes	Camera Relocalization pairs (d1 / d2 / d3)	Novel Path Discovery pairs (d1 / d2 / d3)	Download	Unpacked
interioragent	25	34 564 / 23 150 / 44 300	3 723 / 4 042 / 7 767	10.6 GB	14 GB
internscenes-gen	116 (bedroom 35, diningroom 42, livingroom 39)	160 203 / 155 106 / 171 394	18 021 / 21 866 / 14 153	18.0 GB	31 GB
scenesmith	194 (29 house + 165 room)	198 030 / 145 847 / 268 644	29 352 / 29 615 / 39 458	172 GB (5 parts)	224 GB
Total	335	392 797 / 324 103 / 484 338	51 096 / 55 523 / 61 378	~201 GB	~270 GB

Every scene carries a walkable grid and both tasks' pair tables; the per-task manifests list the scenes each task draws from (manifest_*.json for Camera Relocalization, manifest_npd_*.json for Novel Path Discovery).

Files

SpatialGym_Environmental/
├── README.md
├── manifest.json                       # Camera Relocalization, all sources
├── manifest_interioragent.json         #   … per source (runs use one source at a time)
├── manifest_internscenes-gen.json
├── manifest_scenesmith.json
├── manifest_npd.json                   # Novel Path Discovery, all sources
├── manifest_npd_interioragent.json
├── manifest_npd_internscenes-gen.json
├── manifest_npd_scenesmith.json
├── interioragent.tar.gz                # unpacks to interioragent/
├── internscenes-gen.tar.gz             # unpacks to internscenes-gen/
├── scenesmith.tar.gz.part-00 … -04     # one gzip stream in five numbered pieces (50 GB file cap):
│                                       #   cat scenesmith.tar.gz.part-* | tar xz  →  scenesmith/
├── SHA256SUMS                          # checksums of the archives above
└── figures/                            # the images in this README

Unpack the archives next to the manifests; manifest paths are relative to that directory. SpatialGym_IsaacLab/assets/download.sh camera_relocalization in the SpatialGym repository fetches, verifies and unpacks everything (or a subset of the sources).

Per-scene structure

Common to every scene (written by the SpatialGym preprocessing, identical format across sources):

File	Description
`walkable_cells.json`	0.25 m navigation grid: cells, per-room ids, 4-neighbour adjacency, height map
`visibility.json`, `fov_cells.json`	Per-state field-of-view tables (1.6 m camera, 90°×90°)
`pair_candidates.json`	All reachable (init, goal) state pairs
`camera_relocalization/pairs.json`	Camera Relocalization pairs with difficulty tier
`novel_path_discovery/pairs.json`	Novel Path Discovery demo tours, obstacles and tiers
`walkable_grid.png`	Top-down grid visualisation

interioragent — kujiale_XXXX/

File / Folder	Description
`kujiale_XXXX.usda`	Scene USD (with its own lights)
`Meshes/`, `Materials/`	Per-object meshes, PBR materials and textures
`rooms.json`	Room polygons
`*.hdr`	Environment map

internscenes-gen — {bedroom,diningroom,livingroom}/{scene_id}/

File / Folder	Description
`composed_scene_no_doors.usd`	Scene USD with doors removed (`composed_scene.usd` for the one scene that has none, livingroom/6032)
`layout.json`	Object layout (category, model id, bounding box)
`StructureMesh/`	Floor / wall meshes and `boundary_points.json`
`textures/`	Texture assets
`renders/`	Preview renders

scenesmith — {house,room}/scene_XXX/

File / Folder	Description
`mujoco/usd/scene_scene_XXX.usda`	Scene USD (MuJoCo export; ceiling fixtures removed)
`mujoco/meshes/`	Mesh assets
`combined_house/house.dmd.yaml`	Object list
`room_geometry/`	Per-room geometry
`room_*/`	Per-room assets