Move the drone, flip the AI's brain on, and watch it go from wandering to hunting — then see how first-person shooters became one of AI's favorite training grounds.
Dumb AI vs. Smart AI
LIVE SIMULATION
MOVE: Arrow Keys / WASD · blue = you · orange = enemy drone
Enemy Mode
DUMB
Aggression Level
0%
Times It's Spotted You
0
01SENSE
02THINK
03ACT
04LEARN
How FPS Games Train AI
CLICK TO EXPAND
1993–2016
Digital Shooting Ranges ▾
Researchers turned classic shooters like Doom into virtual labs. A research platform called VizDoom let AI agents learn to aim, navigate, and survive using nothing but raw screen pixels — the same way you see the game — instead of being handed the game's internal code.
2016–2023
Beyond Reflexes ▾
Once agents could aim and move, researchers pushed further: teamwork, resource management, and multi-step strategy inside game worlds. Techniques proven in FPS-style environments carried over into robotics and other AI systems that need to act in real time.
2024–2026
Testing Ground for Language Models ▾
Now researchers are dropping today's AI models — including large language models and vision-language models — straight into game worlds using benchmarks like VideoGameBench, to test spatial reasoning and split-second decisions instead of just text answers.
Today
Same Idea, Bigger Stakes ▾
FPS environments now double as an obstacle course for whatever AI system is newest. Success requires perception, fast decisions, and adapting to a changing 3D world — exactly the skills researchers want proven before trusting AI with robots or other real-world jobs.
PLOT TWIST
You'd think a language model that can write an essay on quantum physics could easily beat a video game — but researchers have found the opposite. Today's most advanced language models still lose to simple search algorithms in FPS-style games, because they were trained to understand text, not to see and move through 3D space in real time. Reflexes and spatial awareness turn out to be a completely different skill than writing sentences.