🚖 TaxiRL: Procedural Content Generation using Reinforcement Learning

A web-based interactive simulator that demonstrates Reinforcement Learning concepts using a custom 10×10 Taxi grid environment, powered by PPO and procedural level generation.

---

🧠 Overview

TaxiRL is an educational and interactive reinforcement learning simulator where an AI agent learns to pick up and drop off passengers in a 10x10 procedurally generated grid world. Trained using Proximal Policy Optimization (PPO) from Stable-Baselines3, it allows both human and agent play modes with visual feedback and reward tracking in real-time.

This project blends AI, Flask web development, and procedural content generation for an intuitive learning experience.

---

🎯 Features

• ✅ Custom 10x10 Grid Environment with level loading from JSON
• 🤖 PPO Agent trained using Stable-Baselines3
• 🧑‍💻 Human vs Agent Play in the same game interface
• 🧩 Procedural Content Generation for dynamic levels
• 📊 Real-time Reward Tracking and agent performance visualization
• 🌐 Flask-based Web Interface with image-based grid rendering
• 📁 Modular Codebase: Easily adaptable and extendable

---

🖥️ Tech Stack

Component	Tools/Technologies
Language	Python 3.8+
RL Framework	Stable-Baselines3
Environment	OpenAI Gym + custom PCG
Web Framework	Flask
Visualization	Matplotlib, Base64 Encoded Images
Deployment Ready	Flask Web App (Localhost, easy to port to cloud)

---

🏁 Getting Started

1. Clone the Repository

git clone https://github.com/ikcod/gym_pcgrl.git
cd gym_pcgrl

2. Create a Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Dependencies

pip install -r requirements.txt

4. Train the PPO Agent (Optional)

If you'd like to retrain the agent:

python train.py

Trained models are stored in /taxi_logs/final_model.zip.

5. Run the Flask Web App

python app.py

Navigate to http://127.0.0.1:5000 in your browser.

---

📂 Project Structure

.
├── app.py                  # Flask web server
├── train.py                # PPO training script
├── pcgrl_env.py            # Custom Taxi-v3 Gym environment
├── static/                 # UI assets (icons, backgrounds)
├── templates/              # HTML templates
├── game_levels.json        # Custom level definitions
├── taxi_logs/              # Trained model files
└── requirements.txt

---

🧪 Results

• PPO Agent converges after ~50k timesteps
• Agent consistently outperforms human gameplay in medium/hard levels
• Intuitive GUI provides easy comparison between agent and player behavior

---

🔮 Future Scope

• 🔁 Integration of advanced RL algorithms (A3C, DQN)
• 🌍 Global leaderboard via Firebase or Supabase
• 🎨 Level Editor with drag-and-drop support
• 📱 Port to Android/iOS with Streamlit or React Native
• 🗣️ Voice-based agent control

---

📎 References

• OpenAI Gym
• Stable Baselines3
• PCGRL Paper – A. Khalifa et al.
• Flask Docs