Overview & Result
typesafe-mario by Faadil Shaik demonstrates Jev playing Super Mario Bros. without capturing screenshots or running computer vision models. By reading structured game state directly from emulator RAM (player position, enemy distances, ground tiles), Jev makes instant discrete button-press choices (run, jump, dodge) at high decision frequencies.
How Jev fits in the loop
- Extracts player coordinate, velocity, and obstacle vectors directly from emulator RAM.
- Constructs a structured state representation and sends it to Jev with candidate controller inputs.
- Emits controller button presses to the emulator with near-zero latency.
How to reproduce
- Clone https://github.com/fhshaik/typesafe-mario and install emulator dependencies.
- Launch NES emulator with Super Mario Bros. ROM and attach memory reader.
- Run the Jev decision loop to observe Mario navigating obstacles based on RAM coordinates.
Why this build matters
Proves that real-time game agents can run at interactive frame rates when perceptual state is structured rather than processed as raw video frames.
Limitations
- Requires access to game memory addresses and emulator hooks.
- Tuned specifically for NES Mario RAM layout.