title	EcoNav AI
emoji	🌍
colorFrom	green
colorTo	blue
sdk	docker
app_port	7860
pinned	false

🌱 EcoNav AI — Exposure Credit OpenEnv

An OpenEnv-compliant Reinforcement Learning environment built for the Meta AI Environmental Decision Intelligence Hackathon.

🌍 Environment Description & Motivation

EcoNav AI models a realistic urban navigation challenge where an intelligent agent must route through an Indian city network (represented as a dynamic graph). Instead of optimizing solely for distance or time, the agent optimizes to minimize pollution exposure and maximize Exposure Credits—a gamified health currency.

What makes this environment highly functional and unique is its integration with live, real-world data. The environment fetches real-time Air Quality Index (AQI) values from the Open-Meteo API. As the agent navigates segments, edge weights (pollution levels) dynamically shift based on reality. The agent earns credits for traversing clean-air segments (Grade A) and loses credits for heavily polluted segments (Grade F). This fills a real gap in mobility algorithms: exposure-aware routing evaluated as an RL task.

🧠 Action and Observation Spaces

Observation Space

The observation space is a strictly typed JSON dictionary detailing the agent's current location, remaining budget, and available next moves including real-time sensor metrics:

• current_city (str): Current city code (e.g. "A").
• current_city_name (str): Human readable city name (e.g. "Delhi").
• destination (str): Target destination code.
• destination_name (str): Target human readable city name.
• visited (list[str]): City codes already traversed.
• neighbors (list[dict]): Available adjacent cities, featuring:
  • distance: Distance to neighbor (km).
  • aqi: Real-time Air Quality Index.
  • grade: Computed health grade (A to F).
  • credit_delta: Credits that will be won or lost by making this move.
• exposure_credits (int): Current credit balance (starts at 100).
• total_exposure (float): Cumulative normalized pollution exposure metric.
• steps_taken (int): Number of steps used in current episode.
• max_steps (int): Step budget limit for the current task.

Action Space

• city (str): A string specifying the city code to move to next. The chosen code must belong to the list of current neighbors.

📜 Tasks & Expected Difficulty

Four discrete tasks of increasing difficulty force the agent to balance trade-offs between step budget, distance, and real-world pollution grids.

1. easy_route (Easy): Navigate from highly-polluted Delhi to Kolkata. Generous step budget (15 steps) allowing maximum flexibility to seek clean air routes. Baseline passing score: 0.5.
2. medium_route (Medium): Same route, but optimized budget (8 steps). Requires balancing direct paths with pollution spikes. Baseline passing score: 0.6.
3. hard_pollution_dodge (Hard): Start from Agra to Kolkata with a tight budget (6 steps). Requires aggressive credit optimization to survive the early exposure hits of Northern India. Baseline passing score: 0.7.
4. expert_credit_max (Expert): Maximize exposure credits while reaching Kolkata within 10 steps. A rigorous test of the agent's multi-objective reward policy. Baseline passing score: 0.8.

🏆 Baseline Scores

We ran the included inference.py evaluating a greedy baseline heuristic (which strictly prefers neighbors with higher credit deltas and lower AQI values).

Task	Score (0.0 - 1.0)	Grade	Credits Final	Reached Goal
easy_route	0.8161	B	110	✅
medium_route	0.8161	B	110	✅
hard_pollution_dodge	0.8326	B	105	✅
expert_credit_max	0.8161	B	110	✅
Average Score	0.8202

🚀 Setup and Usage Instructions

Run via Docker (Hugging Face Spaces)

The provided Dockerfile is optimized to run on Hugging Face Spaces per the OpenEnv specification (automatically exposing port 7860).

docker build -t econav-openenv .
docker run -p 7860:7860 econav-openenv

Local Development Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt # OR pip install fastapi uvicorn requests pydantic openai numpy pyyaml

# Start the environment
uvicorn apps.backend.main:app --host 0.0.0.0 --port 8000

Running Inference

Run the baseline evaluation or LLM-driven agent via inference.py.

To run the greedy baseline:

export ENV_URL="http://127.0.0.1:8000"
python inference.py

To run the LLM Agent (OpenAI Client format):

export ENV_URL="http://127.0.0.1:8000"
export API_BASE_URL="https://router.huggingface.co/v1"
export MODEL_NAME="meta-llama/Llama-3.1-8B-Instruct"
export HF_TOKEN="your-hf-token"

python inference.py