FlappyAI

Vision-based agents are often limited by slow screen capture and input, which makes real-time control in fast games hard.

Screen frames are captured with mss, processed with OpenCV to extract the bird and obstacle positions, and fed as a small state vector to a NEAT network. Decisions are sent back to the browser through the Chrome DevTools Protocol.

[ Browser game ]
        │ mss frame capture
[ OpenCV state extraction ]
        │ positions, distances, velocity
[ NEAT network ]
        │ flap or glide
[ Playwright / DevTools input ]

• Playwright over the DevTools Protocol for low-latency input instead of Selenium.
• Neuroevolution needs no labelled training data and converges quickly on a task this small.
• Frames stay in memory, so there is no disk I/O in the loop.

• End-to-end loop from capture to action fast enough for real-time play.
• A compact experiment in applying evolutionary algorithms to real-time control.

• Python
• NEAT-Python
• OpenCV
• Playwright
• mss
• NumPy