FloodGuard

Android flood-rescue indoor navigation using ARCore VIO for positioning and on-device Gemma 4 E2B (via LiteRT-LM) for scene understanding. Gives rescue teams spatial memory, loop detection, and hands-free Indonesian voice control — entirely offline.

Requirements

• Android device, arm64-v8a, Android 8.0+ (API 26+)
• ARCore-supported device (optional — falls back to sensor PDR if unavailable)
• ~3 GB free storage for the Gemma 4 E2B model
• Android Studio Hedgehog or newer
• JDK 17 (Automatically configured via Gradle Toolchain)

Installation

1. Clone the repo

git clone https://github.com/nii0708/floodguard.git
cd floodguard

2. Open in Android Studio

File → Open → select the floodguard folder. Wait for Gradle sync to finish.

3. Download the Gemma 4 E2B model

The model is not bundled. You can download it in-app on first launch, or manually via ADB for faster setup.

1. Download the gemma-4-E2B-it.litertlm file from Hugging Face.
2. Connect your device and push the model:

adb push gemma-4-E2B-it.litertlm /sdcard/Android/data/com.floodguard.rescue/files/models/

The app validates the file on launch via a magic-header check. If the path or header is wrong, it will redirect to the model setup screen.

4. Build and install

# Debug build
./gradlew installDebug

# Or from Android Studio: Run → Run 'app'

5. Grant permissions

On first launch, accept the prompts for:

• Camera — required for ARCore
• Activity Recognition — required for step detector (PDR fallback)
• Microphone — required for voice commands

Build variants

Command	Output
./gradlew assembleDebug	Debug APK
./gradlew assembleRelease	Release APK (ProGuard enabled)
./gradlew compileDebugKotlin	Compile check only, no APK
./gradlew clean assembleDebug	Clean build

How it works

1. Start mission — 5-second countdown, then ARCore begins tracking.
2. Voice commands — press the on-screen mic button (or volume button) and speak in Indonesian. Gemma classifies the command:
   • "Sudah pernah ke sini?" → checks trajectory for revisits and speaks the result.
   • "Tandai lokasi ini" → captures the current ARCore frame, sends it to Gemma for a scene description, and saves a landmark.
3. Revisit alert — when you return within 1.8 m of a prior position, the app speaks an alert and shows a banner.
4. Stop mission → summary screen with landmark list, 2D map, and JSON export to Downloads/floodguard/.

Credits

Built with Claude Code and Gemini (Android Studio Agents).