On April 13, 2019, OpenAI Five defeated Team OG, the reigning Dota 2 world champions, becoming the first AI system to beat the world champions at an esports game. OpenAI documented the work in its paper “Dota 2 with Large Scale Deep Reinforcement Learning” (arXiv 1912.06680), whose abstract states plainly: “On April 13th, 2019, OpenAI Five became the first AI system to defeat the world champions at an esports game.”
Dota 2 is far harder for a machine than board games like chess or Go. It is a five-on-five team game played in real time, with hidden information, thousands of possible actions each step, and matches that stretch over long time horizons where the consequences of early decisions only appear much later. OpenAI Five learned to coordinate as a team almost entirely through self-play reinforcement learning, training continuously for about ten months and processing roughly two million game frames every two seconds on a large distributed system.
The result extended the lineage of self-play reinforcement learning, from TD-Gammon and the Deep Q-Network through AlphaGo, into the messy, real-time, partially observable world of modern video games. It demonstrated that the same broad approach could scale to one of the most complex competitive games humans play.
Note on sourcing: the OpenAI announcement page (openai.com) blocks automated fetching, so its content was corroborated through the OpenAI Five paper on arXiv (which states the April 13, 2019 victory directly) and through web search. The canonical announcement URL is cited above.