Plate 00 / Frontispiece

Field atlas of interior geometries.

Crowdsourced LiDAR scans from any phone, reconstructed into robotics-ready datasets.

iOS / Android ROS / PLY / COLMAP / NeRF on-device privacy
Live Edition
Plate 00 / The World, surveyedorthographic
graticule 30°
active device / Tokyo
35.7°N 139.7°E
0
scans / 12 cities
Fig. 00 / Devices in the field, live● recording
Plate I / The Collection

Four specimens from a single walk through one room.

Recorded / 14 May 2026
Location / 47.6°N 122.3°W
Spec. 001 / Depth

Depth maps.

Per-pixel distance to the world, registered against device pose. Fused from on-device LiDAR and photogrammetric stereo.

range0.3 to 5.0 msourcelidar / tof / mlcompresszlib / LZ4rate60 Hz
contour 0.6 m
Spec. 002 / Layout

Spatial layouts.

Watertight room geometry, walls, doors, openings, with semantic room labels and per-room area in SI units.

formatPLY + JSONclasses12 room typesaccuracy± 2.0 cmtopologymanifold
scale 1 : 80
Spec. 003 / Environment

Indoor scenes.

Furniture, surfaces, and objects, dense-reconstructed and labelled in 3D, anchored to a single right-handed room frame.

formatmesh + NeRFpipelineCOLMAP / Open3Ddensity1.2 M triframeZ-up, SI
iso / 1 : 40
Spec. 004 / Motion

Movement trajectories.

6-DOF pose at 60 Hz. Every step, turn, and gesture, calibrated against the scene with IMU sensor fusion and loop closure.

formatROS / nav_msgsrate60 Hz, IMU-fusedaccuracy± 0.4 cmduration32.4 s
SE(3) / 60 Hz
Plate II / The Instrument

From a phone in a pocket to a tensor in a training run.

Workflow / three stages
Avg. cycle / 2 to 5 min/scene
I.

Walk the space.

Press record and walk the room. On-device ARKit / ARCore runs lightweight SLAM. Only keyframes, sparse cloud, and pose graph are kept. Faces and plates are blurred before anything leaves the phone.

ClientSpatialOS.app
II.

Reconstruct in the cloud.

Chunks arrive and dense-reconstruct with COLMAP + Open3D, anchored to scale by LiDAR depth. Multi-session fusion merges every scan of the same place via ICP into one scene. On T4 GPUs, 2 to 5 minutes each.

GPU workerscolmap / open3d
III.

Export to your stack.

Pull the unified scene as a ROS bag, PLY/PCD, or NeRF-ready COLMAP folder, each with a quality-scored JSON sidecar. Served over REST and as live MCP resources for agents.

Protocolrest / mcp/spatial
Plate III / The Export

One scan, every format your models speak.

Standard / MCP 1.0
REST / v0.4.2

Real rooms, exported natively into the robotics and reconstruction stacks.

Every scene is reconstructed once and exported into the formats the ecosystem already runs on. No custom adapters, no lossy conversions. Pull it over a plain REST call, or subscribe to live phones as spatial:// resources over MCP.

  • .bagROS bagsensor_msgs/PointCloud2 + nav_msgs/Path, ready to replay.robot nav / sim
  • .ply/.pcdOpen3DDense, voxel-normalised clouds with estimated normals.processing / viz
  • /sparseCOLMAPSparse + dense folders, LiDAR-anchored scale.nerfstudio / ngp
  • .jsonMetadataDevice, depth_source, 0 to 100 quality score per scene.indexing / QA
export.sh / reference clientrest python mcp
1# 1 / create a session, stream keyframe chunks
2POST /sessions                      # to id
3POST /sessions/{id}/chunks         # gzip, 5 MB
4POST /sessions/{id}/complete       # trigger fuse
5
6# 2 / poll until the scene is reconstructed
7GET  /sessions/{id}/status
8# score 87 / accepted / 218,440 pts / lidar
9
10# 3 / export in the format you need
11GET  /sessions/{id}/export/ros
12GET  /sessions/{id}/export/ply
13GET  /sessions/{id}/export/colmap   # NeRF-ready
14
15# or, live, over MCP:
16subscribe spatial://hub.local/scn_4a7c
Plate IV / Quality & Provenance

Every scene is scored before it ships.

Scoring / 0 to 100
Weighted / four axes
0
score
✓ training-ready
living_room / 3 merged sessions / lidar
Coverage 30%
92
Depth quality 25%
88
Trajectory 25%
84
Point density 20%
81
< 40 / reject
Auto-rejected. The contributor is nudged to re-scan the gaps.
40 to 70 / flagged
Accepted but marked low-quality; held back from robotics training.
> 70 / accept
Full accept, training-ready and published to the corpus.
We collect
  • Point-cloud geometry & depth maps
  • Pose trajectory
  • Device model & depth_source
× Never leaves the phone
  • Human faces, blurred on-device
  • License plates, blurred on-device
  • Precise GPS (city-level only) & personal info
Plate V / The Register

Scan more. Unlock more. Keep the credit.

Tiers / four
Attribution / perpetual
i

Bronze

on sign-up
Read-only API access, rate limited.
ii

Silver

10+ sessions / avg > 60
Full read & export of the dataset.
iii

Gold

50+ sessions / avg > 75
Early access to new dataset drops.

Verified

team reviewed
Named attribution credit in the data.
Regional register / this weeklive
i
depthwalker
Tokyo, JP / Gold
214 sessions
48.6 km²
ii
lidar_lena
Berlin, DE / Gold
187 sessions
41.2 km²
iii
mapmaker_q
Austin, US / Silver
163 sessions
37.9 km²
iv
voxel.vu
Lisbon, PT / Silver
142 sessions
33.4 km²
Plate VI / Applications

For everything that has to move through space.

Domains / 04 surveyed
Adjacent fields / many
Robotics training

Pretrain manipulation policies on the rooms they will actually work in.

Crowd-captured scans as a domain-randomization source for imitation learning, behavior cloning, and sim-to-real bridges. Real rooms, not synthetic ones.

328 scansacross 12 cities
home / office / café / warehouse
Embodied AI

Ground a model in the room it is standing in.

Stream the active scan into an agent's context via MCP. “Where is the coffee mug?” becomes a question the model can actually answer.

spatial://addressable
resources
Navigation

Indoor maps without GPS.

Trajectory and layout data train routing policies for delivery bots, vacuums, and AR wayfinding through real homes and offices.

Spatial models

A foundation model for space itself.

Geometry-native pretraining at scale, the way language models treated the open web, but for the volumetric structure of every human environment.

Marginalia / Project Thesis

Every smartphone, a robot training device.

spatial-os / manifesto.md/open / perpetual / public

Open spatial data
belongs to everyone.

SpatialOS is Apache-2.0, MCP-native, and built in public. Bring a phone. Bring a robot. Help map the world for the machines that will live in it.

$pip install spatial-os

About

An open platform turning the phones in our pockets into the world's largest spatial-data corpus for robots and embodied AI.

Community

SpatialOS / Field Atlas / Vol. I / Edition 04 / © 2026328 active scans / all systems normal