Demo : https://youtu.be/auf6HfCe0H8

Satellite detects. Gemini reasons. Drone confirms. Robot diagnoses. Farmer approves.

AgriScout is a satellite-guided, agentic field-triage system for multi-cause crop stress (pest / water / nutrient / false-alarm). A Google ADK agent powered by Gemini 2.5 Pro reads satellite NDVI anomalies, dispatches a DAC drone for aerial confirmation, then runs a 4-step embodied-reasoning diagnostic with a LeKiwi + SO101 ground robot — with Gemini Robotics-ER 1.6 providing spatial reasoning, a local Gemma 4 (Ollama) fallback for cross-validated perception, and human-in-the-loop approval gates throughout.

End-to-end: satellite anomaly → confirmed diagnosis → work order, in under 3 minutes.

davis-hack-26/
├── backend/                            # FastAPI + Google ADK agent + Gemini/Gemma VLM ensemble
│   ├── app/
│   │   ├── agent/                      # ADK agent loop, tools, ER policy translator
│   │   ├── api/                        # /execute, /runs, /risk, /plan routes
│   │   ├── domain/                     # risk engine + multi-cause irrigation pack
│   │   ├── sim/                        # drone + robot adapters with safety guards
│   │   └── vision/                     # Gemini ER + Gemma 4 + cross-validated ensemble
│   ├── tests/                          # 69 pytest tests (motion budgets, belief, ER loop)
│   └── scripts/                        # smoke tests for VLMs and sims
└── frontend/                           # React + Vite + Tailwind mobile-first phone app
    └── src/
        ├── pages/today/                # morning brief + alert review
        ├── pages/monitor/              # live agent timeline + field map + belief evolution
        ├── pages/orders/               # work orders + recent runs
        ├── components/monitor/         # timeline, stage cards, belief UI
        ├── lib/                        # api client, useRun hook, countdown hook
        └── data/                       # zone catalog (A/B/C/D ↔ A2/B3/C2/D2)

The DAC sims (drone-sim-main, robotsims-main) are external dependencies — clone and run them separately per the DAC hackathon instructions. They are not committed to this repo.

• Agent harness: Google ADK — LlmAgent, Runner, FunctionTool, before_tool_callback
• Orchestration brain: Gemini 2.5 Pro
• Embodied reasoning: Gemini Robotics-ER 1.6 (preview) — emits target_point [y, x] + status
• Local fallback VLM: Gemma 4 (gemma-4-31b-it) via Ollama — cross-validates every frame
• Backend: FastAPI (Python 3.11+) + Pydantic + sse-starlette
• Frontend: React + TypeScript + Tailwind + Vite (mobile-first)
• Drone sim: Cesium-based 3D simulator (DAC)
• Robot sim: SO101 arm + LeKiwi base simulator (DAC)

cd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env   # add GEMINI_API_KEY, optionally OLLAMA_HOST
uvicorn app.main:app --reload --port 8000

VLM_CLIENT=mock in .env lets you run the full agent loop without LLM API keys — useful for offline dev and demo-day insurance.

cd frontend
npm install
npm run dev    # → http://localhost:5174

Both sims must run alongside the backend for the full pipeline. See backend/README.md for sim setup details.

Open localhost:5174 on your phone (or a phone-sized browser viewport). The Today tab shows the morning brief; tap the alert card on Zone B to launch the agentic pipeline. Watch the timeline, approve the robot dispatch when prompted, then review the generated work order on the Orders tab.

cd backend && python -m pytest tests/ -q
# 69 passed

Includes regression tests for motion budgets (no robot tip-overs), belief-evolution snapshots across the diagnostic loop, and ER policy closed-loop behaviour.

Single-zone NDVI alerts tell you something is wrong. AgriScout tells you exactly what — distinguishing pest / water / nutrient stress with physical evidence from a drone and a ground robot, all in the time it takes to make coffee. Compresses detect → verify → diagnose → act from days to minutes, with a human always in the loop.