Pixel Flappy Bird AI

A retro pixel-art Flappy Bird simulation where a browser-based AI learns to survive pipes with Canvas 2D, TensorFlow.js, and genetic training.

Project Snapshot

Item	Details
App type	Browser game and AI training visualizer
Rendering	HTML Canvas 2D pixel-art scene
AI runtime	TensorFlow.js loaded in the browser
Training style	Genetic population training plus TensorFlow fine-tuning
Local server	Node.js static file server
Package managers	pnpm or npm
Node.js	>=18 recommended, tested with v22.21.1

Demo

Live Demo

The demo screen is split into two main areas:

• Game canvas: renders the bird, pipes, score, ground, clouds, and the currently observed agent.
• Learning dashboard: shows generation progress, current agent, current score, best scores, average generation score, deaths, and simulation speed.
• AI live state: exposes the current training state, raw neural-network output, and whether the AI decided to jump or wait.
• TensorFlow evolution: tracks fine-tuning loss, accuracy, sample count, and the historical evolution chart.
• SPACE key simulation: visually indicates when the AI would press SPACE to flap.
• Controls: starts, pauses, resets, speeds up, normalizes speed, and changes render mode.

Training In Motion

The GIF demonstrates the training loop in action:

• Each agent observes the game state and produces a neural-network output between 0 and 1.
• Outputs greater than or equal to 0.5 trigger a jump; lower values make the bird wait.
• The app records each agent's state/action experiences while it plays.
• When agents die, their fitness is calculated from survival time and pipe score.
• At the end of a generation, the strongest agents influence the next population.
• TensorFlow fine-tuning uses the best collected experiences to improve the mentor model before crossover and mutation create new agents.

Features

• Pixel-art Flappy Bird clone built with Canvas 2D and custom-drawn sprites.
• Neural-network controlled bird using TensorFlow.js directly in the browser.
• Genetic algorithm loop with population scoring, elite retention, crossover, and mutation.
• Fine-tuning step with model.fit() using experiences collected from top agents.
• Live training dashboard with generation, score, deaths, speed, AI decision, and TensorFlow metrics.
• Evolution chart for best score, average score, and TensorFlow loss across generations.
• Single-agent visual mode for easier observation.
• Batch training mode for faster multi-agent simulation with simplified rendering.
• Zero backend framework: Node.js only serves static files locally

How The AI Works

The bird brain receives five normalized inputs from the game:

Input	Meaning
birdY	Bird vertical position inside the playable area
velocityY	Current vertical velocity
distanceX	Horizontal distance to the next pipe
gapTop	Top edge of the next pipe gap
gapBottom	Bottom edge of the next pipe gap

The TensorFlow.js model is a small dense neural network:

5 inputs -> Dense(8, relu) -> Dense(6, relu) -> Dense(1, sigmoid)

The final sigmoid output is interpreted as the flap decision:

output >= 0.5 => jump
output <  0.5 => wait

Current training defaults:

Setting	Value
Population size	30 agents
Elite agents kept	4
Mutation rate	0.12
Mutation amount	0.32
Fitness formula	score * 160 + frames survived
Fine-tune epochs	4
Fine-tune batch size	64
Fine-tune elite sample source	Top 6 agents
Max fine-tune samples	5000
Max simulation speed	12x

Tech Stack

• HTML5 for the page structure.
• CSS3 for the retro dashboard layout and responsive styling.
• JavaScript ES Modules for game, AI, UI, and training logic.
• Canvas 2D API for all rendering.
• TensorFlow.js 4.22.0 from CDN for model prediction, crossover, mutation, and fine-tuning.
• Node.js for the local static server.
• pnpm or npm for script execution.

Requirements

• Node.js 18 or newer.
• A modern browser with ES module support.
• Internet access on first load if TensorFlow.js is not already cached, because it is loaded from:

https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@4.22.0/dist/tf.min.js

Running With pnpm

pnpm install
pnpm start

Then open:

http://localhost:3000

Building Static Assets

pnpm run build

The build script copies index.html, style.css, js/, and assets/ into dist/. The dist/ directory is generated output and is not committed to the repository.

Running With npm

npm install
npm start

Then open:

http://localhost:3000

The project does not require local runtime dependencies today. The install step is still useful to keep the package-manager workflow consistent and to validate the project metadata.

Available Scripts

Command	Description
pnpm run build	Generates the production static files in dist/
pnpm start	Starts the local Node.js static server
pnpm dev	Same as start, useful during development
npm run build	Generates the production static files in dist/
npm start	Starts the local Node.js static server
npm run dev	Same as start, useful during development

The server listens on port 3000 by default. You can override it with:

PORT=4000 npm start

Controls

Control	Behavior
Start training	Starts or resumes AI simulation
Pause training	Pauses simulation without clearing progress
Reset training	Recreates the population and clears historical metrics
Speed up training	Increases simulation speed up to the configured maximum
Normal speed	Restores simulation speed to 1x
Show 1 agent visually	Runs one visible agent at a time for clear observation
Train multiple agents	Runs the population faster with simplified rendering

Project Structure

.
├── api/
│   └── index.js           # Vercel Function entrypoint
├── assets/
│   └── demo/
│       ├── demo.png        # Main README screenshot
│       └── flap.gif        # Training demonstration GIF
├── js/
│   ├── ai.js               # TensorFlow.js model, prediction, clone, crossover, mutation, training
│   ├── bird.js             # Bird physics, bounds, and pixel rendering
│   ├── evolution-chart.js  # Dashboard chart for score and TensorFlow loss history
│   ├── game.js             # Game loop state, pipes, scoring, collisions, and Canvas drawing
│   ├── main.js             # Browser bootstrap and animation frame loop
│   ├── pipe.js             # Pipe geometry, movement, collision, and rendering
│   ├── training.js         # Population lifecycle, fitness, fine-tuning, and generation evolution
│   └── ui.js               # Dashboard updates, controls, render mode, and SPACE simulation
├── index.html              # App shell and TensorFlow.js CDN script
├── package.json            # Scripts and package metadata
├── pnpm-lock.yaml          # pnpm lockfile
├── scripts/
│   └── build-static.js     # Static production build generator
├── server.js               # Static file handler and local server
├── vercel.json             # Vercel build, function, and rewrite configuration
└── style.css               # Retro dashboard styling

Development Notes

• The app is intentionally client-side: training runs in the browser, not in Node.js.
• TensorFlow tensors are wrapped with tf.tidy() or disposed after training datasets are used to reduce memory growth.
• Canvas image smoothing is disabled to preserve the pixel-art look.
• Batch mode prioritizes training speed over visual detail.
• The static server sets Cache-Control: no-store so changes are easier to test while developing.
• In production, api/index.js serves only files generated into dist/; run pnpm run build before testing the serverless handler locally.

Troubleshooting

Problem	Fix
TensorFlow.js does not load	Check internet access and the CDN script in index.html.
Browser shows a blank page	Open DevTools and check for JavaScript module or CDN loading errors.
Port 3000 is already in use	Run with another port, for example PORT=4000 npm start.
pnpm is unavailable	Use npm start, or enable Corepack and prepare pnpm for your Node installation.
Vercel deploy cannot find files	Confirm the build ran and dist/index.html, dist/style.css, and dist/js/main.js exist.
Vercel settings differ from production	Set the framework to Other, use pnpm run build, output dist, and remove the install command override.
Training looks slow at first	Use Speed up training or switch to batch render mode. Early generations are expected to fail quickly.

License

No license file is currently included. Add one before publishing or redistributing this project.