Yalpha/AGRITECH-META

🌱 AgriDecisionEnv v3 – Sustainable Farming RL Environment

🚀 Overview

AgriDecisionEnv v3 is a real-world OpenEnv-compatible environment designed to evaluate AI agents on multi-season agricultural decision-making under uncertainty, delayed effects, and resource constraints.

The environment models how farmers must balance:

• —🌾 Crop yield (short-term productivity)
• —🌍 Soil health (long-term sustainability)
• —💧 Water usage (groundwater limits)
• —💰 Budget constraints (economic feasibility)

This environment evaluates agent performance under delayed rewards, stochastic climate conditions, and resource constraints, closely mimicking real-world agricultural decision-making.

🌍 Real-World Motivation

Over 33% of global arable land is degraded due to:

• —monocropping
• —excessive fertilizer use
• —unsustainable irrigation

Existing systems rely on static heuristics and fail under dynamic conditions.

👉 This project provides a benchmark environment for training and evaluating intelligent agents capable of adaptive, long-term agricultural decision-making.

🧠 Environment Design

The environment simulates a farm across multiple seasons. At each timestep, the agent observes the current state and selects actions.

🔁 Interaction Loop

State → Agent → Action → Environment → New State + Reward

📊 Scenario Initialization

🔍 Observation Space

🎯 Action Space

⚙️ Reward Function

reward = clamp( yield_score

• —soilqualitybonus
• —fertilizer_penalty
• —delayedfertilizerpenalty
• —irrigation_penalty
• —monocrop_penalty
• —groundwater_penalty
• —budget_penalty , 0.0, 1.0)

Key Properties

• —Dense reward signal (not sparse)
• —Penalizes unsustainable practices
• —Encourages long-term planning
• —Strictly normalized to [0.0, 1.0]

🧪 Task Design

🟢 Easy (1 step)

• —Objective: maximize immediate yield
• —Deterministic
• —No delayed effects

🟡 Medium (3 steps)

• —Objective: balance yield and soil improvement
• —Introduces temporal dependencies
• —Penalizes overuse

🔴 Hard (5 steps)

• —Includes:
• —stochastic weather
• —delayed fertilizer effects
• —groundwater depletion
• —budget constraints

👉 Tests long-term sustainability strategies

📈 Baseline Results

LLM Agent (via OpenAI client + HF router)

Rule-Based Baselines

👉 Rule-based agent conserves budget but trades off yield; greedy/random overspend

🤖 Inference (OpenAI Client)

The agent uses the OpenAI client interface for all LLM calls.

Required environment variables:

• —APIBASEURL
• —MODEL_NAME
• —HF_TOKEN

Logging Format (STRICT)

[START] [STEP] [END]

✔ Required for evaluation ✔ Fully reproducible

🧱 Project Structure

agriv3/ ├── models.py ├── env.py ├── tasks/ ├── baselineagents.py ├── inference.py ├── openenv.yaml ├── Dockerfile └── README.md

⚙️ Setup & Usage

Local

pip install pydantic openai HFTOKEN=yourtoken python inference.py

Docker

docker build -t agri-env . docker run -e HFTOKEN=yourtoken agri-env

Hugging Face Spaces

• —SDK: Docker
• —Tag: openenv
• —Must respond to /reset

✅ OpenEnv Compliance

• —reset() → Observation
• —step(action) → (Observation, float reward, done, info)
• —state() → Observation

✔ Typed models ✔ Deterministic grading ✔ Reward range [0,1] ✔ Docker-compatible

🧠 Why This Matters

This environment models:

• —real agricultural trade-offs
• —long-term sustainability challenges
• —constrained decision-making under uncertainty

👉 Useful for:

• —RL benchmarking
• —LLM evaluation
• —policy simulation

🏁 Conclusion

AgriDecisionEnv v3 is a realistic and scalable benchmark for evaluating intelligent agents in sustainable agriculture — where short-term gains often conflict with long-term survival.