Thrishu/disaster-response-env

🌊 Disaster Response Coordinator Agent

Overview

A flood disaster is unfolding across multiple zones. You are the AI coordinator. You have limited rescue teams, medical kits, food supplies, helicopters, and flood barriers. Roads wash out. Floods intensify. Storms arrive without warning.

Your decisions directly determine how many people survive.

This is a real operational problem — disaster response coordinators face exactly this resource-allocation-under-uncertainty challenge in real emergencies. The environment models the core trade-offs: triage vs. fairness, flood prevention vs. immediate rescue, and helicopter deployment for cut-off zones.

Observation Space

Action Space

All fields are Dict[zone_id, int]. Total allocations across zones must not exceed available resources.

Key constraint: road_blocked and air_only zones receive no rescue, medical, or food unless helicopters are also sent there.

Reward Function

Rewards are shaped across the full trajectory (not just end-of-episode):

• —+0.03 per person rescued via rescue teams
• —+0.015 per person healed via medical kits
• —+0.003–0.006 per food unit (scaled by zone need)
• —+0.02 per flood barrier deployed
• —+0.005 per person evacuated
• —−0.05 to −0.25 per step for high flood levels (≥7 and ≥9)
• —−0.1 to −0.3 per step for high injury ratios (>15% and >30%)
• —−0.15 for over-allocating resources (exceeding available stock)
• —−0.2 for taking no action (idle penalty)

Grader

Final episode score (0.0–1.0) combines:

Tasks

🟢 Easy

• —Zones: 1 (open access)
• —Steps: 10
• —Challenge: Basic allocation, moderate flood
• —Baseline (rule-based): ~0.50–0.60
• —Good agent target: ~0.75–0.85

🟡 Medium

• —Zones: 2 (one road-blocked)
• —Steps: 15
• —Challenge: Must use helicopters to reach blocked zone; heavier injuries
• —Baseline (rule-based): ~0.35–0.45
• —Good agent target: ~0.65–0.75

🔴 Hard

• —Zones: 3 (one road-blocked, one air-only from start)
• —Steps: 20
• —Challenge: Active storm, only 2 helicopters for 2 blocked zones — agent must triage
• —Baseline (rule-based): ~0.20–0.30
• —Good agent target: ~0.55–0.65

API

\POST /reset?task=easy|medium|hard   → initial observation
POST /step                          → next observation, reward, done, info
POST /grade                         → final episode score (0.01–0.99)
GET  /state                         → current observation
GET  /health                        → {"status": "ok"}

Example: Reset and Step

import requests, json

BASE = "http://localhost:7860"

obs = requests.post(f"{BASE}/reset?task=medium").json()

action = {
    "allocate_rescue":    {"A": 2, "B": 1},
    "send_food":          {"A": 20, "B": 15},
    "send_medical":       {"A": 10, "B": 8},
    "deploy_helicopters": {"B": 1},   # B is road_blocked — helicopter required!
    "deploy_barriers":    {"B": 1},
    "evacuate":           {"A": 50}
}

result = requests.post(f"{BASE}/step", json=action).json()
print(result["reward"], result["done"])

Setup

Docker

docker build -t disaster-response-env .
docker run -p 7860:7860 disaster-response-env

Local (Python)

pip install fastapi uvicorn pydantic openai openenv-core
uvicorn server.app:app --host 0.0.0.0 --port 7860

Run Inference

export API_BASE_URL=https://api.openai.com/v1
export MODEL_NAME=gpt-4o-mini
export HF_TOKEN=hf_...   # required — passed as OpenAI client api_key (see submission guidelines)

python inference.py

Project Structure

.
├── Dockerfile
├── pyproject.toml
├── openenv.yaml
├── inference.py          # baseline inference script
├── env.py                # core DisasterEnv
├── models.py             # Observation, Action, Reward pydantic models
├── grader.py             # deterministic episode scorer
├── tasks/
│   ├── easy.py
│   ├── medium.py
│   └── hard.py
└── server/
    └── app.py            # FastAPI server

Author

Thrishank