The navigation app from the OFFGRID prototype, as demonstrated for the SIH 2026 idea submission.

Install it, ride, and watch what happens when the satellites stop mattering.

What it does

IDR Nav shows your position on an offline map at 10 Hz. When GNSS drops — or when you tell the app to withhold it on a road it recognises — the marker keeps moving, driven by a 240 KB neural network reading forward speed out of the phone's own vibration, gyro heading, and the geometry of the road you are on. When GNSS comes back, the app draws the truth alongside the dead-reckoned track and tells you how far off it was.

On the recorded demonstration ride bundled with the app: off by 17.3 m after 569 m and 151 seconds without GNSS.

What is in the build

• The full engine in Kotlin, running live at 0.83 ms per 10 Hz tick
• LiteRT inference at 0.074 ms per window, single-threaded on CPU
• Offline campus basemap (PMTiles, 812 KB) with a course-up 3-D camera and optional satellite imagery
• The IDR Logger built in as a second tab — 21 sensor and GNSS streams to CSV at native rates
• A replay harness: push a recorded session to the device and watch the engine run it back
• A field-test screen for on-device conformance checks

Install

adb install -r IDRNav-debug-arm64.apk

Or copy it to the phone and open it. You will need to allow installation from unknown sources.

Requirements: Android 10 (API 29) or newer, arm64-v8a, and location permission. Grant location "all the time" and leave the app battery-unrestricted, or Android will suspend the sensor service when the screen goes off.

This is a debug build, unsigned for release and not on any store. It is a prototype demonstrator, not a shipping navigation app.

Scope

The offline map and the two recognised corridors cover Shiv Nadar University, Greater Noida. Away from those roads the app runs its general 2-D mode and shows position over satellite imagery instead. The speed model was trained on a bicycle with the phone in a bottle cage; it has not been trained or validated on cars.

Also attached

SIH2026_OFFGRID_26168.pdf — the six-slide idea submission.

Where the numbers come from

Everything above is measured, and the measurements are in the repository:

• Held-out engine results — docs/11_ENGINE_RESULTS.md
• On-device conformance, latency and the outage scenario — docs/16_APP_RESULTS.md
• The engine specification — docs/12_ENGINE_SPEC_for_kotlin.md

The 17.3 m figure is a training ride replayed with live online alignment. It shows what the app does; it is not a held-out generalisation claim. The held-out numbers are 1.52 % median drift over 600–850 m outages, 10 of 10 rides under 10 %.