OpenAI Five
Five neural networks learning to play Dota 2 as a team.
The project
OpenAI Five is a team of reinforcement-learning agents built to play Dota 2. The project demonstrated coordinated play in a complex game with long matches, partial information, and many possible actions.
Where AI comes in
Agents learned through large-scale self-play and reinforcement learning. This is a game-playing AI research case, rather than an LLM-based dialogue system or a game-generation tool.
For game makers
A landmark example of training coordinated behavior in a game environment. It helps frame questions about rewards, teamwork, simulation throughput, and the difference between research performance and production NPC design.
Go to the source
OpenAI Five · Official articleResearch paper · arXivThe official page restricts automated access. See the source notes for the supporting reference.
Keep exploring
Explore
Relight: Lantern Expanse
Fly between worlds, explore their surfaces, and relight the way home.
Hollow Crown
A browser-playable pixel adventure, with AI helping build and demonstrate the game.
Wildbrush
A family-made adventure with paintbrush combat and floating islands.