Wander
Indoor navigation, by ear

The map ends at
the door. We keep going.

Wander turns real places into walkable 3D maps, then guides you through them with a chest-worn phone, live obstacle cues, and a calm voice in your ear, everywhere GPS gives up.

Built on tools teams already trust

NianticARKitOpenAIModalFastAPINext.jsThree.jsSwiftNianticARKitOpenAIModalFastAPINext.jsThree.jsSwift
Powered by OpenAI

A calm voice that understands the space.

Geometry is fast but meaningless; a model understands meaning but is too slow for a safety loop. Wander splits the work: detection on-device, understanding in the model.

Cues you can trust

OpenAI turns route state and detections into short spoken cues, and stays silent when the camera can’t confirm what’s ahead.

Ask about the room

“What’s in front of me?” A vision model answers from the live camera frame, in plain language.

Say where, not which door

A BM25 and dense retriever with rank fusion lets the agent resolve natural destinations to real waypoints.

Codex on the team

We gave Codex the Niantic SDK docs and our conventions; it built the companion app while we wired the phone and backend.

Perception in, grounded action out. The same spatial-agent loop, general enough to drop into any environment we can scan.

Toward a general spatial agent
Live session

See a route the moment it runs.

wander.app/worlds/e7-atrium
123
LocalizedE7 · Atrium
0 m

Localization accuracy

aligned to Niantic VPS

<0 ms

Obstacle alert

detected on-device

0M

Splats per world

walk-through detail

Voice guidance

Hear the path,
not the noise.

Short, spoken cues arrive exactly when they matter. Distances are counted down, landmarks are named, and silence is treated as a feature.

  • Turn-by-turn cues generated from the route, not a script
  • Adaptive verbosity: more detail in busy spaces, less in hallways
  • Works with earbuds or the phone speaker

Spoken cues

  • Turn left in 3 metres
  • Continue for 8 metres
  • Arriving, Room 7302
Obstacle sensing

See what your
phone sees.

A chest-mounted iPhone watches the ground ahead. Benches, doors, and people are detected on-device and folded into the next cue before you reach them.

  • On-device detection with ARKit depth, no video leaves the phone
  • Left, right, and back mounts supported for wider coverage
  • Cues rank by urgency so the important thing is said first
3D maps

Maps you can
walk through.

Every space is a Gaussian splat you can orbit, annotate, and measure. Align it once to Niantic's positioning and the phone knows where it is within a footstep.

  • Import .ply or .splat exports, or capture directly from the app
  • Multi-floor worlds with elevator and stair transitions
  • Waypoints stay anchored as the scan improves
Aligned to Niantic VPS1.2M splats14 waypoints
Grounded in research

We started by reading, not guessing.

Wander is a stack of proven ideas wired together: visual positioning, splat reconstruction, robust pose, and fast routing. Here is what each layer stands on.

SIGIR 2009

Reciprocal Rank Fusion + BM25

Cormack, Clarke, Büttcher

Fuse keyword and semantic search so “take me somewhere quiet” resolves to the right waypoint.

BMVC 2012

Locally Optimized RANSAC (PoseLib)

Lebeda, Matas, Chum

Robust 6-DoF camera pose from 2D to 3D matches, stable even when features are noisy.

1968

A* Search

Hart, Nilsson, Raphael

Optimal shortest-path routing across the waypoint graph behind every spoken turn.

Where it goes

One spatial engine. Many bodies.

The engine that guides a person doesn’t care what’s carrying it. Localize, route, avoid. A robot needs the same three. Wayfinding for humans is the first body, not the last.

Today

People

Blind and low-vision travelers, hands-free, guided by voice.

Next

Warehouse robots

The localize, route, avoid loop an AMR needs, from a phone-grade scan.

Next

Inspection drones

Indoor flight where GPS can’t reach, anchored to a VPS map.

Horizon

AR glasses

The same wayfinding, drawn as a visual overlay for anyone.

Horizon

Emergency response

Pre-scanned buildings guiding responders through smoke and dark.

Horizon

Facilities & venues

Airports, hospitals, campuses. Turn-by-turn, indoors.

Three steps from scan to first walk.

No beacons to install, no floor plans to redraw. If you can walk it, you can map it.

  1. STEP 01

    Capture a space

    Walk the building once with an iPhone. We turn the scan into a Gaussian splat and align it to Niantic's visual positioning system.

  2. STEP 02

    Annotate the route

    Drop waypoints, name destinations, and write the cue for each turn in the web dashboard. Preview exactly what a walker will hear.

  3. STEP 03

    Walk with guidance

    Clip the phone to a chest mount. It localizes, watches for obstacles, and speaks calm turn-by-turn directions the whole way.

Ready to map your first space?

Sign in, create a world, and have a walkable route ready before the coffee cools.