Eliminate the enemy.
The model chooses movement, aiming, and firing actions from the visible engine state.
Recorded runs from the released checkpoint (ec832471…), matching the release report. These are small, structured-state experiments—not general game-playing or navigation benchmarks.
The separate public-data checkpoint has different results. View its report and recordings ↗
Download metrics and route traces ↗The model chooses movement, aiming, and firing actions from the visible engine state.
The mission changes while the environment and evaluation seeds stay the same.
What the model sees: health, ammunition, and visible-enemy coordinates from ViZDoom. It does not infer from video pixels. The recordings play in game time; their duration does not measure inference speed. This training curriculum includes control-shaped tasks.
Game artwork is from Freedoom. Artwork license and attribution ↗
SmallDecide was previously named OpenAuditor. Original recordings and reports use the earlier name.