// Indoor environments · robotics · embodied AI · sim-to-real

Unlimited training
environments.
Architecturally valid.

Six synchronized data layers. Zero manual annotation.
Drop into any robot gym, sim framework, or training pipeline.

Purpose-built for teams training robots in indoor spaces. Every environment is constraint-generated — not stitched from scans — so room connectivity, obstacle volumes, and surface heights are exact and queryable by design. No reconstruction. No cleaning. No wait.

Navigation graph plan_4341973 · hover nodes

drag to rotate · scroll to zoom

65.6m² area
5rooms
4graph edges
96%arch. score

environments per parameter set — without repetition

One configuration. Unlimited structurally distinct buildings. Each with all six data layers ready — no post-processing, no annotation step.

6Data layers per building — geometry, semantics, topology, navigation, furniture, and USD — generated in one pass
0Hours of manual annotation — every label is generated alongside the geometry, not added afterward
100%Architecturally valid layouts — 1M+ construction constraints enforced at generation time, not post-processed

Generative engine

Same parameters.
Different valid building. Every time.

The constraint engine runs over 1,000,000 validated architectural rules — room adjacency, circulation logic, door placement, structural limits. Not random: deterministic diversity. Every output is unique and valid.


Full 3D output

A complete 3D scene —
not a mesh with metadata bolted on.

Walls, roof, furniture, semantic labels, and topology are all written from a single JSON knowledge model. USD, IFC, GLB, and JSON are generated natively — not converted from each other.


Layout diversity

What infinite generation
looks like.

8 configurations. 8 structurally distinct buildings. Unique room arrangements, door placements, and window counts — all within the same parameter set. All architecturally valid.

Bedroom
Bathroom
Kitchen
Living
Entryway
Other

// in production

What robots have already done
with this data.

Working proofs-of-concept — not renders. Each uses the topology graph, semantic layer, and furniture bounding boxes directly from the dataset. No preprocessing required.

POC · Room navigation
01
Room-to-room navigation
Robot navigates between any two rooms using the topology graph. Door openings as waypoints, furniture bboxes as obstacles — no collisions.
topology graphobstacle avoidance
POC · Window cleaning
02
Surface targeting — window
Robot navigates to a target room, approaches the window, lifts to floor-relative height from the geometry layer, and executes a cleaning path.
surface heightsemantic target
POC · Table cleaning
03
Surface targeting — countertop
Robot identifies a countertop from the furniture layer, navigates to it, and cleans the surface using the assumed object height from semantic obstacle data.
furniture layerobject height

// live generator

Build your configuration.
See the dataset it produces.

Adjust rooms and common areas. The constraint engine resolves adjacency, circulation, and door logic in real time. Every configuration produces a unique, valid building — not a random one.

Configuration
Bedrooms2
Bathrooms1
Common areas
Kitchen
Living room
Dining
Need more rooms? Request access →
This config produces
Navigation graph — live
hover nodes to inspect
Floor plan preview
each config = unique valid building · constraint-driven · not random

Knowledge Layer

Six layers.
One generation pass.

Click any layer to inspect real output from building plan_4341973. Geometry, semantics, topology, navigation, furniture, and output formats are all derived from a single JSON knowledge model.

Traditional pipeline
drawing model export cleanup simulation
Manual preparation at every step
Semantic labels added as a separate pass
Topology has to be reconstructed by hand
Indoor Datasets approach
generation structured environment simulation-ready
All six layers generated in a single pass
Topology and semantics embedded at source — not post-processed
Every output passes 1M+ architectural constraint checks
01 // .json
Semantic layer
Room types, attributes
02 // .json .usd
Geometry layer
Walls, roof, heights
03 // .json
Topology layer
Graph, connectivity
04 // derived
Navigation layer
Path planning data
05 // .json .usd
Furniture layer
Objects, obstacles
06 // all formats
Output formats
USD, JSON, IFC, GLB
Semantic layer plan_4341973.json · real data
5 rooms4 doors4 windows21 furniture
Floor plan — real bbox · hover rooms
Semantic attributes per element
"type": "living",
"area": 18.75,
"isObstacle": false,
"isPassable": true,
"hasCollision": false,
"category": "space",
"bbox": { xs, xe, ys, ye }

Comparison

Built for robot training.
Not retrofitted for it.

Scan-based and existing procedural datasets were designed for visual navigation benchmarks. They produce geometry and images — not the structured spatial knowledge required for task execution. Here's where that gap shows.

Capability Scan-based Procedural Indoor Datasets

Most existing indoor datasets were built for visual navigation benchmarks — they provide geometry and images, but not the structured spatial knowledge that robot task execution requires. Indoor Datasets is purpose-built for structured robot training: every environment ships with explicit connectivity, semantic labels, surface heights, obstacle volumes, and multi-format export — all generated without manual annotation.


// get started

Get a sample matched to
what you're building.

Tell us your robot training setup. We'll prepare a complete 6-layer dataset sample matched to your use case — and walk you through it.

Complete 6-layer sample — geometry, semantics, topology, navigation, furniture, USD
Configured to your specific training scenario
30-min technical session with our team
Scale discussion — API, bulk generation, custom parameters
No contract. No cost.

Request received

We'll reply within 1–2 business days with a sample dataset and next steps.

We respond within 1–2 business days