Dota 2 with Large Scale Deep Reinforcement Learning

TL;DR

The OpenAI Five report documents a distributed, continually trained self-play system for a long-horizon, imperfect-information team game.

Why it matters

It remains a useful systems case study: scaling RL required not only an algorithm, but sustained rollout generation, distributed optimization, monitoring, and curriculum management.

Key findings

1. 01

   Self-play can generate a continuously adapting task distribution.

2. 02

   Long-running RL systems need operational tools for stability and iteration.

Scaling dimensions

Models, methods & benchmarks

Models

OpenAI Five

Algorithms

PPO, Self-play

Benchmarks

Dota 2

Topics