Assistive Navigation

See what's ahead. Hear it instantly.

RoomSighter is a real-time assistive-navigation app. It detects 60 common indoor obstacles — from people and furniture to appliances — from a live camera feed, estimates how far away each one is, and speaks a clear, spoken cue — entirely on-device, in real time.

A person walking outdoors with a guide dog, representing independent, assisted mobility
0.63
mAP@0.5, held-out test set
60
obstacle classes detected live
100%
on-device detection, no video upload
~0.4m
depth-model error, near range
What it detects

60 classes, tuned for indoor obstacles

Fine-tuned from a COCO-pretrained checkpoint onto the objects that matter most for everyday indoor navigation — every indoor-plausible COCO class, plus door, stairs, and window sourced and verified outside COCO.

Person

Person

Detects people ahead and announces how far away and which direction they are.

Chair

Chair

Flags seating obstacles before you walk into them, with a proximity warning up close.

Table

Table

Picks out tables and desks — common waist-height hazards that canes can miss.

Door

Door

Locates doorways so a room transition is never a surprise.

Also detected
StairsBenchCatDogBackpackUmbrellaHandbagTieSuitcaseFrisbeeSports ballBaseball batBaseball gloveSkateboardTennis racketBottleWine glassCupForkKnifeSpoonBowlBananaAppleSandwichOrangeBroccoliCarrotHot dogPizzaDonutCakeCouchPotted plantBedToiletTVLaptopMouseRemoteKeyboardCell phoneMicrowaveOvenToasterSinkRefrigeratorBookClockVaseScissorsTeddy bearHair drierToothbrushWindowLamp
How it works

One pipeline, three responsibilities

Detection, distance estimation, and speech are each handled by the component best suited for the job — nothing is reimplemented twice across platforms.

01

Detect, on-device

Two detection models, RT-DETR-L and YOLO26-l, run on your device — in the browser on WebGPU (onnxruntime-web) or on iOS via CoreML — and a small learned scorer merges their results. Your camera video stays on the device.

02

Estimate distance

A monocular depth model gives each object a distance and direction: on-device on iOS, and on our distance service on the website (one downscaled frame per cycle, used in memory and discarded).

03

Speak, instantly

The native speech engine (Web Speech API / AVSpeechSynthesizer) reads the most urgent obstacle aloud — important classes, close and straight ahead first — free, no cloud TTS.

Ready to try it live?

Grant camera access in your browser and hear real-time obstacle announcements.

Launch live demo →