AIDino

Overview

Features

• Chrome DevTools Protocol integration for direct browser communication
• Efficient screenshot capture and processing using MSS
• Custom OpenAI Gym environment for reinforcement learning
• PPO model implementation via Stable Baselines3
• Image preprocessing with Sobel edge detection
• Progressive model checkpoints for tracking training progress

Technical Architecture

Chrome Integration

• Programmatic control of the Dino game
• Precise input simulation (keyboard events)
• Real-time game state monitoring
• Screenshot capture of the game area

Custom Gym Environment

• Defines a discrete action space (jump, duck, do nothing)
• Processes screenshots into suitable observations
• Provides appropriate rewards based on game progress
• Handles game reset and initialization

Reinforcement Learning

• Proximal Policy Optimization (PPO) algorithm
• MLP policy network architecture
• Reward function that encourages longer survival
• Periodic model checkpoints to track training progress

Model Checkpoints

• ppo_dino_100k.zip through ppo_dino_1000k.zip
• Each checkpoint represents increased training (100k to 1M timesteps)

Getting Started

Environment Setup

1. Create and activate a virtual environment:
   bash

   1# Create virtual environment 2python -m venv venv 3 4# Activate on Linux/macOS 5source venv/bin/activate 6 7# Activate on Windows 8# venv\Scripts\activate
2. Install dependencies:
   pip install -r requirements.txt

Running the Project

1. Launch Chrome with remote debugging:
   bash

   1# Linux 2google-chrome --remote-debugging-port=1234 3 4# Windows 5# chrome.exe --remote-debugging-port=1234
2. Train the model:
   python rf.py
3. Use a pre-trained model by modifying the load path in rf.py:
   python

   1# Uncomment and modify this line in rf.py 2# model = PPO.load("models/ppo_dino_1000k", env=env)

How It Works

1. The environment connects to Chrome via DevTools Protocol
2. The game is initialized and the initial state is captured
3. For each timestep:
   • The current game state is captured through screenshots
   • Images are processed using edge detection
   • The model selects an action (jump, duck, nothing)
   • The action is executed through the Chrome connection
   • Rewards are calculated based on survival and game progress
4. Training continues until the model achieves optimal performance

Performance

Contributing

• Improve the reinforcement learning model or try different algorithms
• Enhance the image processing for better feature detection
• Optimize the performance for faster training
• Report bugs or suggest features by opening an issue