Stealth Thief Escape RL

A Reinforcement Learning project where an agent (Thief) learns to navigate a hazardous 10x10 city grid to reach a getaway vehicle while avoiding police officers, hardcoded walls, and a central police station's surveillance zone.

Key Features

• Environment: Custom Gymnasium-compatible environment (StealthThiefEnv) with:
  • 10x10 grid with specified coordinate mapping (01 to 100).
  • Randomly spawned Police officers.
  • Central Police Station (44, 45, 54, 55) with a 1-grid "Busted" search radius.
  • Static Wall obstacles.
• Agent: Advanced Dueling Double DQN architecture for stable and efficient learning.
• Persistence:
  • Automatic saving and loading of model weights (models/stealth_thief_latest.pth).
  • Persistent Replay Buffer (data/stealth_replay_buffer.pkl) to preserve experience across training sessions.
• Continuous Training: Training script runs in a loop, resetting episodes automatically but allowing manual termination via the 'Q' key in the game window, which triggers a final persistent save.

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Assets: Ensure the assets/ directory contains:
   • agent.png (Thief)
   • police.png (Police)
   • wall.png (Walls)
   • car.png (Getaway Vehicle)

How to Run

Train the Agent

Watch the agent explore the environment and learn optimal escape routes:

python scripts/train.py

Press 'Q' while focused on the game window to save progress and exit.

Evaluate the Agent

Watch the best-trained model attempt the escape:

python scripts/play.py

Disclaimer

This project is developed for game simulation and research purposes. Please refer to DISCLAIMER.md for more information.