A.R.C. — Autonomous Rescue Cluster

Gemma 4 on LiteRT · offline disaster response

Post-disaster heterogeneous rescue fleet (UAV + UGV + aerostat) coordinated by Decision Hubs with Gemma 4 reasoning at the edge (LiteRT E4B) and optional cloud planning. Built for the Gemma 4 Good Hackathon (Impact: Global Resilience · Technology: LiteRT).

Marketing homepage (/) — project overview and entry to the live demo.

Mission Command (/simulation) — tactical map, fleet FPV, and Gemma 4 Decision Hub.

Live demo	Precomputed playback
Mission Command — real Gemma 4 inference	/demo-player — timeline JSON replay
Requires local LiteRT weights + bridge	Works without GPU / model file

---

Judge Quick Start (LiteRT only)

# 1) Download Gemma 4 E4B LiteRT weights → models/gemma-4-E4B-it.litertlm
#    https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm

pip install -r requirements.txt
pnpm install

# Terminal A — LiteRT OpenAI bridge (default :8787)
python scripts/litert_openai_server.py
# or: pnpm litert:server

# Terminal B — copy env and start Next.js
cp .env.example .env.local
# Edit .env.local — keep only:
#   LITERT_OPENAI_BASE_URL=http://127.0.0.1:8787/v1

pnpm dev

Open http://localhost:3000/simulation?ai=gemma

1. Click GEMMA4 (not MOCK).
2. Wait for header ● LIVE Gemma 4.
3. Confirm the metrics panel (MODE · BACKEND · LATENCY · …).
4. Press RUN.

Health check

curl http://localhost:3000/api/gemma-chat
# → {"ok":true,"backend":"litert","model":"gemma-4-E4B-it-litertlm",...}

Apple Silicon (M1/M2/M3/M4): .env.example enables LITERT_BACKEND=gpu / LITERT_VISION_BACKEND=gpu by default — Metal is dramatically faster than CPU. Comment those lines out on Intel/AMD machines without a supported GPU build.

Measure latency: bash scripts/bench_gemma.sh runs three prompts per agent and reports min/median/max from the X-Arc-Latency-Ms header. The browser console also logs [gemma] agent=… latency=…ms tokens=… during a live round.

---

What this repo contains

arc_core/          Python package — agents, GemmaPerceiver, simulation, tests
app/               Next.js 15 App Router
public/simulation/ Mission Command (static UI + /api/gemma-chat)
scripts/           litert_openai_server.py — OpenAI-compatible LiteRT bridge
models/            Place gemma-4-E4B-it.litertlm here (not in git)
Writeup.md         Kaggle submission narrative

Inference path (submission): browser → POST /api/gemma-chat → LITERT_OPENAI_BASE_URL → scripts/litert_openai_server.py → Gemma 4 E4B (multimodal FPV supported).

---

Web routes

URL	Description
/	Marketing site (Three.js hero)
/simulation	Mission Command — tactical map, FPV, Decision Hub, fleet dialogue
/simulation?ai=gemma	Default live AI mode
/simulation?ai=mock	Rule-based / template mode (labeled honestly in UI)
/lite	2D lite sim (public/lite/scenario_canvas_lite.json)
/demo-player	MapLibre + canvas timeline playback
/whitepaper	System design document

---

Mission Command (/simulation)

• GEMMA4 / MOCK toggle — use GEMMA4 for hackathon screenshots and video.
• AI metrics panel (screenshot-friendly): MODE · BACKEND · LATENCY · TOKENS · AGENT · ROUND.
• Phase label syncs with mode: CLOSED LOOP · GEMMA-4 (LiteRT) vs SIMULATION · RULE-BASED.
• Footer uplink/plan — measured latency or (simulated) in MOCK mode (no fake fixed ms).
• Agents: Drone_Alpha (vision), Track_Beta, Relay_Gamma, Orchestrator — via streaming/non-streaming chat.

FPV frames are sent as JPEG base64 to LiteRT (conversation schema).

---

Models (hackathon compliance)

Role	Model	Where
Edge inference (required for live demo)	litert-community/gemma-4-E4B-it-litert-lm	models/gemma-4-E4B-it.litertlm
Cloud planning (optional)	gemma-4-26b-a4b-it	GEMMA_API_MODEL in GemmaPerceiver / timeline

Function-calling tools in arc_core: calculate_survival_score, dispatch_rescue_task.

---

Environment variables

Copy .env.example → .env.local for Next.js.

Variable	Required	Description
LITERT_OPENAI_BASE_URL	Yes (live demo)	e.g. http://127.0.0.1:8787/v1
LITERT_MODEL_PATH	No	Path to .litertlm (default: models/gemma-4-E4B-it.litertlm)
LITERT_BACKEND	No	cpu or gpu (LLM)
LITERT_VISION_BACKEND	No	cpu or gpu (vision tower)
LITERT_SERVER_PORT	No	Bridge port (default 8787)
GEMMA_API_KEY	No	Google AI Studio — timeline / API mode
GEMMA_API_MODEL	No	Default gemma-4-26b-a4b-it
GEMMA_MODE	No	litert | api | mock | auto for Python

---

Install & run (full)

Python

pip install -r requirements.txt

# Skeleton demo
python -m arc_core.runners

# Precompute timeline for demo-player
python -m arc_core.simulation.timeline_generator --steps 200 --output public/demo-player/timeline.json

# Tests
pytest

Timeline uses GemmaPerceiver when GEMMA_API_KEY is set or LiteRT weights exist; otherwise rule/mock trajectory.

Node.js

pnpm install
pnpm dev          # http://localhost:3000
pnpm build        # production build
pnpm litert:server # shortcut for LiteRT bridge

Demo player

# Ensure timeline exists (see Python section above)
pnpm dev
# → http://localhost:3000/demo-player

---

Python architecture (arc_core)

Module	Purpose
perception/gemma_perceiver.py	Gemma 4 backends: LiteRT → API → Mock
agents/decision_hub.py	Cluster “brain”, task allocation
simulation/timeline_generator.py	Offline timeline JSON
scheduler/	Survival scoring, task allocator

python -c "from arc_core.perception.gemma_perceiver import GemmaPerceiver; print(GemmaPerceiver(agent_id='t').stats())"
# Expect mode=litert when weights are present

---

API: /api/gemma-chat

Method	Description
GET	LiteRT health via bridge /health
POST	Chat proxy; body: { agent, message, history?, image_base64?, stream? }

Non-stream response includes meta: backend, latency_ms, model, tokens (may be null on edge).

---

Troubleshooting

Symptom	Fix
LITERT_OPENAI_BASE_URL not configured	Create .env.local from .env.example
Header stays MOCK / offline	Start python scripts/litert_openai_server.py; check curl localhost:8787/health
Empty or slow first reply	Model load on first request; CPU inference can take tens of seconds
Metrics show LATENCY · —	Run one GEMMA4 round after LIVE badge appears
/demo-player empty map	Run timeline_generator; check public/demo-player/timeline.json

Optional GPU: LITERT_BACKEND=gpu LITERT_VISION_BACKEND=gpu python scripts/litert_openai_server.py

---

Submission docs

File	Purpose
Writeup.md	Kaggle writeup — problem, architecture, impact
whitepaper.md	Extended system design

---

Safety

A.R.C. is decision support for research and hackathon demonstration. It does not replace certified incident command or professional rescue teams.

---

Acknowledgements

• LiteRT-LM (Google AI Edge)
• Gemma 4
• Multi-agent UAV–UGV planning and disaster-resilience literature cited in team materials

License

MIT — see LICENSE.

Demo video & cover: see the Kaggle submission (project writeup, gallery, and linked media).