EchoWalk: On-Device Guidance for Low-Vision Users
EchoWalk runs 4 AI models entirely on the Snapdragon NPU. A semantic safety radar warns you about obstacles and finds things around you through spatial audio; a tap-to-describe vision-language path speaks what’s in the room.
Core trick
Runs four ExecuTorch models on the Snapdragon NPU and renders obstacle alerts as spatial audio and haptics.
Description
$2a
README excerpt
# EchoWalk On-device indoor navigation assistant for blind / low-vision users, running entirely on a Samsung Galaxy S25 Ultra (Snapdragon 8 Elite NPU) via **PyTorch ExecuTorch + Qualcomm QNN**. No cloud. No latency. The camera feed and your personal maps never leave the device. ## What it does | Mode | Team | Trigger | Output | | --- | --- | --- | --- | | **Safety Radar** (continuous) | A | always-on | spatial audio + haptics for nearby obstacles, distinguishing a wall you can trail from a trip hazard | | **Scene Description** (on-demand) | B | volume key / tap | a spoken description of the room from a VLM | | **Familiar Places** (learned) | C | when in an enrolled space | localizes you and guides you to saved destinations | See the full plan in [`.cursor/plans/on-device_blind_navigation_69ca1444.plan.md`](.cursor/plans/on-device_blind_navigation_69ca1444.plan.md). ## Repository layout ``` hackathon/ ├── android/ # Multi-module Gradle project. Each team owns a module. │ ├── shared/ # :shared — foundation: camera, ExecuTorch helper, mode + audio (no team deps) │ ├── teama/ # :teama — Team A: Safety Radar (depends on :shared only) │ ├── teamb/ # :teamb — Team B: Scene Description (depends on :shared only) │ ├── teamc/ # :teamc — Team C: Familiar Places (depends on :shared only) │ └── app/ # :app — integration: MainActivity + ModeManager (depends on all) ├── ml/ # Model export to