Chrome Dino Game AI Bot

An AI-powered bot that plays the Chrome Dinosaur Game. Includes a perfect deterministic bot that plays indefinitely, plus reinforcement learning (Q-learning) bots for experimentation. Uses Selenium for browser automation.

---

Bots Overview

Bot	File	Approach	Learns?	Performance
Perfect Bot	dino_bot_perfect.py	Physics-based rules	No (doesn't need to)	Near-unbeatable
Basic Bot	dino_bot.py	Adaptive thresholds	Yes (JSON)	Good after tuning
Q-Learning Bot	dino_bot_ml.py	Reinforcement learning	Yes (pickle)	Improves over episodes
Q-Learning Headless	dino_bot_ml_headless.py	RL + headless Chrome	Yes (pickle)	Fastest training

---

Perfect Bot (dino_bot_perfect.py)

The star of the show. Uses precise game-state reading and physics-based calculations to make optimal decisions:

• Speed-adaptive jump timing — calculates exact reaction distance based on current game speed
• Obstacle type awareness — jumps over cacti, ducks under low pterodactyls, runs under high ones
• Multi-obstacle lookahead — checks if a second obstacle is close behind to adjust timing
• Fast-fall mechanic — presses DOWN while airborne to descend faster when needed
• 125 Hz polling — reads game state ~125 times/second for precise reactions

python3 dino_bot_perfect.py              # visible mode
python3 dino_bot_perfect.py --headless   # headless mode

How It Decides

For each game tick (~8ms):
  1. Read exact obstacle positions, types, dino state from Runner.instance_
  2. Nearest obstacle is a cactus?
     → Jump when obstacle is within speed-dependent reaction distance
  3. Nearest obstacle is a low pterodactyl (yPos >= 70)?
     → Duck and hold until it passes
  4. Nearest obstacle is a high pterodactyl (yPos < 60)?
     → Do nothing — run right under it
  5. Two obstacles close together?
     → Adjust jump timing to clear both

---

Basic Bot (dino_bot.py)

Adaptive threshold-based bot that learns from crashes:

• Maintains separate jump and duck thresholds per speed range
• Identifies obstacle types (cactus vs pterodactyl) and reacts accordingly
• Adjusts thresholds based on crash feedback (jumped too early/late)
• Saves progress to learning_data.json

python3 dino_bot.py

---

Q-Learning Bots (dino_bot_ml.py / dino_bot_ml_headless.py)

Reinforcement learning bots with improved state representation:

• 5 speed categories (finer than original 3) for better discrimination
• Dino action state (running/jumping/ducking) prevents invalid actions
• Obstacle height categories encode the correct response (jump/duck/ignore)
• Action masking — can't jump while already jumping
• Streak bonuses — extra reward for consecutive obstacle passes
• Saves Q-table to learning_data.pkl for continuous improvement

python3 dino_bot_ml.py            # visible mode (50 episodes)
python3 dino_bot_ml_headless.py   # headless mode (faster training)

RL Hyperparameters

Parameter	Value	Purpose
Alpha	0.15	Learning rate
Gamma	0.95	Discount factor
Epsilon start	1.0	Initial exploration
Epsilon decay	0.995	Per-episode decay
Epsilon min	0.02	Minimum exploration

---

Requirements

Prerequisites

1. Python 3.8+
2. Google Chrome (latest version)
3. ChromeDriver matching your Chrome version
   • Download from ChromeDriver Downloads
   • Or install via package manager (e.g., brew install chromedriver)

Install Dependencies

git clone https://github.com/yourusername/chrome-dino-ai.git
cd chrome-dino-ai
pip install -r requirements.txt

---

How It Works

All bots interact with the Chrome Dino game through Selenium WebDriver and JavaScript injection:

1. Open the game at https://chromedino.com/ or https://elgoog.im/t-rex/
2. Read game state from Runner.instance_ (obstacles, speed, dino position, crash status)
3. Decide action based on bot logic (rules or Q-learning)
4. Execute via keyboard input (SPACE = jump, ARROW_DOWN = duck)
5. Repeat at high frequency (8-15ms per tick)

Game State Available

Runner.instance_ = {
    crashed: bool,           // collision detected
    currentSpeed: float,     // game speed (increases over time)
    tRex: {
        jumping: bool,       // dino is mid-jump
        ducking: bool,       // dino is ducking
        yPos: float,         // vertical position
        xPos: float          // horizontal position
    },
    horizon: {
        obstacles: [{
            xPos: float,     // distance ahead
            yPos: float,     // height (lower = higher on screen)
            width: float,    // obstacle width
            typeConfig: { type: string }  // CACTUS_SMALL, CACTUS_LARGE, PTERODACTYL
        }]
    }
}

---

Troubleshooting

Runner is not defined

• Ensure the game URL loads correctly
• The bot will automatically retry alternate URLs

selenium.common.exceptions.WebDriverException

• Ensure ChromeDriver version matches your Chrome browser
• Add ChromeDriver to your PATH

Bot crashes frequently

• Use dino_bot_perfect.py for near-unbeatable play
• For Q-learning bots, run more episodes — they improve over time

---

License

MIT License