MarioStudent (Core ML)

A 9.0M-parameter pixel policy that plays Super Mario Bros. (NES) on the Apple Neural Engine.

It is a MobileNetV3-Large that looks at the screen around Mario (now and 4 frames earlier) and picks one of 11 eight-frame moves: run, walk, hold, jump, hop, up, left, leftjump, wait, swim, down. It was trained by imitation on a save-state beam-search solver that clears all 32 levels, plus DAgger rounds, with auxiliary tile and enemy grid targets. It was not distilled from JEV.

MarioStudent.mlpackage: fp16, 17 MB, all 185 ops on the Neural Engine, about 0.5 ms per decision (M5 Pro). Its move choice matches PyTorch on CPU, GPU and ANE.
mario-student.pt: the PyTorch checkpoint.
config.json: inputs, outputs and move order.

Results

32 levels × 3 start offsets:

	Flags	Levels cleared at least once	Cleared every time
Model alone	53/96	23	12
Model + lookahead	78/96	30	21

Lookahead: before each move, candidate moves are simulated on an emulator save state, each followed by the model's next 6 moves, and the first that survives and gains ground is played. With lookahead the result is the model plus emulator simulation, not the network alone. The maze castles 4-4 and 8-4 are not cleared.

Input / output

Input crop: [1, 6, 120, 96] fp16 in [0, 1]. RGB now, then RGB 4 frames earlier, covering 32 px behind Mario to 160 px ahead, every 2nd pixel of the 256×240 NES frame.
Output logits: 11 move logits (order in config.json), then 11 survival logits.
Moves are 8-frame button macros. The last input is held while the game has control, e.g. during a pipe entry.
Mario's on-screen x, which places the crop, is read from emulator RAM.

Demo

FluidUse Sources/JevMarioDemo is a SwiftUI app that runs this model on the Neural Engine, with or without lookahead, next to JEV-9B on Core ML.

You need your own copy of the game. No ROM, frames or gameplay video are included here.

Downloads last month: 20

Video Preview

Reinforcement Learning