Screen-Scraping Survival of the Fittest

What it does

IAMDinosaur is a Node.js experiment that plays Chrome’s offline dinosaur game by watching your screen. It uses RobotJS to scan pixels and measure the distance, length, and speed of upcoming cacti, then feeds that data into a neural network. The network’s output decides when the dino jumps, ducks, or idles, while a genetic algorithm evolves twelve genomes per generation—culling the worst, crossing over the survivors, and mutating the offspring.

The interesting bit

The project treats the browser as an uncooperative black box: it never touches the game’s internal state, relying entirely on screen observation and simulated keypresses. The README even documents a genuine game bug where the T-Rex slowly drifts right until the AI mistakes the dinosaur for a cactus because they share the same color.

Key highlights

• Evolves 12 genomes per generation using fitness scores based on jumped cacti
• Reads game state by ray-tracing pixels rather than accessing internal variables
• Built with Synaptic for neural networks and RobotJS for screen capture and keyboard control
• Terminal UI for saving, loading, and selecting trained genome JSON files

Caveats

• The dino’s slow rightward drift eventually breaks the pixel scanner; the README suggests refreshing the page or injecting a console script to reset the x-position
• Requires the terminal and browser window to sit on the same screen so screen-scraping coordinates align correctly

Verdict

A tangible sandbox for developers curious about genetic algorithms and neural networks. Skip it if you want a polished, resolution-agnostic reinforcement-learning framework that doesn’t need occasional babysitting.