Echo 1
Assistive navigation wearable
A body-worn camera + LiDAR + GPS device that warns of obstacles, drop-offs and steps, answers spoken questions, reads text aloud, and gives clock-direction navigation — fully offline. Safety never depends on the AI model or the network.
A lightweight, body-worn unit
Designed to be unobtrusive and worn all day — here it is in motion.
Awareness, guidance and answers — fully offline
Every capability runs on the device itself. No app to open, no account, no signal required for the features that keep you safe.
Obstacle detection & spoken alerts
YOLOv8n on the Hailo-10H names hazards (person, chair, car…) and speaks the clock direction — one alert per direction, most severe wins.
Drop-off & step detection
A down-facing LiDAR auto-calibrates the floor and warns “Stop. Drop-off ahead.” or “Caution, step up ahead.” with confirm-streak filtering.
Offline voice assistant
Hold the button, ask a question. faster-whisper transcribes on-device, a tool-calling agent answers — 13 tools, zero cloud.
Scene description & reading
SmolVLM2-500M describes what's in front of you and Tesseract OCR reads signs, labels and documents aloud.
Clock-direction navigation
A route state machine gives turn cues as a three-beep earcon then “Turn toward 2 o'clock,” reroutes on deviation, and announces arrival.
Safety that never fails quietly
Hazard logic is pure Python, isolated from the AI. If the assistant, navigation, or dashboard crashes, the safety warnings keep running.
If the AI crashes, you are still warned.
The hazard pipeline is pure, isolated code that never imports the language model or touches the network. It runs as its own supervised process — and a two-layer (supervisor + systemd) restart keeps it alive.
200° Camera
wide-angle vision
YOLOv8n @ Hailo
30 FPS · ~17 ms
Down-facing LiDAR
drop-offs & steps
Hazard Evaluator
clock direction · severity
Alert Manager
spoken, prioritized
Speaker
“Stop. Drop-off ahead.”
Validated hardware
Every sensor was independently bench-tested before integration — measured, not assumed.
| Subsystem | Device | Result |
|---|---|---|
| Compute | Raspberry Pi 5 (16 GB) + Hailo AI HAT+ 40 TOPS | 30 FPS / ~17 ms inference |
| Camera | OV5647, ~200° lens (CSI) | 2592×1944 |
| LiDAR | TF-Luna, down-facing (UART @115200) | drop-off / step detection |
| GPS | NEO-7M (UART @9600) | NMEA, course heading |
| Microphone | INMP441 I2S MEMS | live speech-to-text verified |
| Speaker | USB audio + amplifier | priority-arbitrated playback |
| Button | Push-to-talk (BCM GPIO1) | 50 Hz debounced polling |
8 build phases · ~95% complete · 121 unit tests passing · 14 architecture decision records.
Interested in Echo 1?
We work with clinicians, accessibility organizations and early testers. Tell us how you'd use it.
Get in touch