Short answer
Facility navigation data pairs written or spoken route instructions with floor plans, site maps and landmark names. It can support deployment planning, simulation context, and evaluation of language-guided navigation. It does not contain task demonstrations or robot sensor traces, and the maps often carry security-sensitive detail that a facility may need to remove or restrict before it is willing to share them.
What problem this data solves
A humanoid that works in a hospital, plant or warehouse has to get to the task before it can do the task. People describe routes in their own terms: "past the second fire door, left at the nurses' station, third bay on the right." Those instructions refer to landmarks that may not appear on any CAD drawing and may differ from the official room numbers.
Facility operators hold the two halves of this problem: the maps (floor plans, evacuation diagrams, site plans, rack layouts) and the wayfinding language (new-hire orientation sheets, delivery instructions, visitor directions, shift handover notes). A buyer who gets only one half has either geometry without language or language without geometry.
What published navigation work indicates
Language-guided navigation became a benchmark problem with the Room-to-Room dataset, introduced with the Matterport3D Simulator as the first benchmark for visually grounded natural-language navigation in real buildings (Anderson et al., 2018). The setup it established, an instruction plus a scene and a target route, is the pattern that workplace navigation data can extend. The Room-to-Room scenes are residential interiors, though, not plants, warehouses or clinics, so a buyer cannot assume transfer to workplace layouts.
Types of facility navigation assets
| Asset | What it holds | Typical use | Common problem |
|---|---|---|---|
| Floor plans (PDF, CAD, BIM export) | Walls, doors, rooms, labels | Map building, simulation scenes | Out of date; scale and units vary |
| Evacuation and fire plans | Exits, assembly points, extinguishers | Emergency behavior, landmark lists | Security and life-safety review |
| Site plans | Buildings, roads, docks, yards | Campus-scale routing | Rarely show interiors |
| Rack and zone maps | Aisles, bays, location codes | Warehouse navigation | Location code format differs per system |
| Written directions | Route instructions in natural language | Instruction grounding, evaluation | Informal, inconsistent landmarks |
| Signage inventories | Sign text and positions | Landmark detection | Needs photos to be useful |
| Restricted-area lists | Where access is limited | Constraints for planners | Often confidential by nature |
What this data can support and what it cannot
It can support:
- Building a route-planning test set where instructions and ground-truth paths are known.
- Creating simulation environments from plans, as a base for synthetic scenes.
- Checking that a robot's landmark vocabulary matches the site's, in line with a workplace glossary.
- Scoping a deployment: counting doors, elevators, thresholds and narrow passages the platform must handle.
It cannot support:
- Task demonstrations. Knowing the route to the pharmacy says nothing about how to restock a shelf there.
- Perception training. A plan has no appearance data. Real corridors have clutter, glare, wet floors and moved furniture. Pair maps with signage and safety label images or purpose-recorded walkthroughs.
- Control. Gait, stair handling and door opening need robot data. Plans give dimensions, not dynamics.
- Current layout. Plans diverge from the physical site over time. Treat them as a prior and check them against the building.
Security and access are the main gate
Floor plans can show the location of cash rooms, server racks, controlled substances, security cameras and utility shutoffs. Facilities typically treat these as restricted. The supply-line preparation work is therefore as much about removal as formatting: strip security-sensitive layers, generalize labels, and confirm what the operator is willing to share beyond its own staff.
Some plans are also owned by someone else: an architect, a landlord, a campus authority or a code-compliance consultant. Ask who holds the copyright and whether the plan carries an access restriction. Redaction can reduce what is exposed but does not by itself establish that the operator may license a document.
Illustrative navigation instruction record
The record below is Illustrative; names, IDs and values are synthetic.
# Illustrative only - not an available dataset
site_id: site_017
floor: 2
instruction_id: nav_0093
text: "From the supply room, go past the break room, turn left at the double doors, and stop at the third bay on the right."
landmarks:
- {name: "break room", plan_ref: "R2-114", visible_from_corridor: true}
- {name: "double doors", plan_ref: "D2-031", sign_text: "Staff Only"}
route_ground_truth: [R2-102, C2-A, D2-031, C2-B, bay_3]
plan_version: "2025-rev3"
restricted_layers_removed: [camera_positions, cash_room, shutoffs]
The fields that matter most are the link between each spoken landmark and a plan element, the plan revision and a log of what was removed.
Acceptance criteria you can set
- Landmark resolution. What share of instruction landmarks resolve to a plan element? Set your own threshold.
- Plan currency. Is there a revision date and a note on known changes?
- Coordinate frame. Are units, scale and origin stated?
- Route validity. Do ground-truth routes pass through real, open doors?
- Removal log. Is it clear which layers were stripped?
Evaluation design with navigation instructions
Navigation data is easy to leak between training and evaluation. If the same corridor appears in training instructions and in the test set, a system can score well by memorizing the building. Decide what the claim needs: generalization to new instructions in a known building, to a new floor, or to a new site. Then hold out by that unit, and keep the held-out floors or sites out of any simulation environment built for training.
Instruction style matters as well. Operators write directions in different registers: terse labels ("Dock 4, left"), full sentences, or relative phrases that depend on where the walker started. A mixed set tests robustness; a single-register set tests only that register. Record who wrote each instruction, since a facilities manager and a new hire describe the same route differently. These details also tell you whether to ask for more than one writer per route.
Questions to ask suppliers
- Who maintains the plans, and when were they last verified against the building?
- Do you have written wayfinding instructions in addition to drawings?
- Which layers or labels are security-sensitive, and who decides what is removed?
- Is the plan owned by the operator or by an architect or landlord?
- Can plans be delivered as vector or BIM exports rather than scans?
- Are door types, thresholds, elevators and stairs recorded?
- Do location codes follow a documented scheme?
- Are there multiple sites or floors with comparable formats?
Where this sits among other procedural assets
Navigation instructions are one procedural modality, next to natural worker instructions and emergency procedures. For evaluating whether a system follows them, see instruction-following evaluation for embodied agents. Held-out evaluation routes should come from sites or floors not used for training. The procedures hub and the humanoid training data guide explain the wider mix.
Where a robot and people share routes, shared human-robot workflow procedures covers the right-of-way and handoff rules that maps cannot show.
Discuss your requirements
If your project needs maps and route instructions for specific kinds of facilities, describe the facility types, formats and permitted uses to SourceX. SourceX reviews the requirement, looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources. A request does not guarantee a match. For records adjacent to this asset type, see the catalog page on construction project records.
Discuss your humanoid training-data requirements with SourceX
Sources
- Vision-and-Language Navigation: Interpreting visually-grounded navigation instructions in real environments — Room-to-Room benchmark for language-guided navigation in real buildings.