Pixels2Play: One model, any game

Watch our foundation model play a variety of 3D games.

This blog post accompanies our first academic paper, published at the Conference on Games 2025.

We outlined our work building an initial version of our a high-speed vision to action model Pixels2Play-0.1 (P2P0.1). Please see the paper for more details about our approach.

Below we demonstrate our AI playing a number of different games. All recordings are from the first recording (no cherrypicking) with no interventions. The same model is used to play all games. The game play is certainly not perfect, but we can see the model is beginning to understand how to participate in a variety of games. There is lots of room from improvement.

Feel free to suggest games you’d like to see our AI play in future in the comments below.

Roblox / Be a shark

Roblox / Be a snake

MS-DOS / Need for speed

Doom

Quake

Simple FPS (internal test game)

Hovercraft (internal test game)

Need for Speed: Hot Pursuit

Goat Simulator 3

Need for Speed: Rivals (note: this game was not seen during model training and while the model performs fairly poorly you can see that it can generalize enough from games seen during training to complete the lap).