This paper introduces SC2LE (StarCraft II Learning Environment), a reinforcement learning environment based on the game StarCraft II. The environment presents a new grand challenge for reinforcement learning, involving multi-agent interaction, imperfect information, a large action space, a large state space, and delayed credit assignment. The environment provides an open-source Python-based interface for interacting with the game engine, along with a suite of mini-games and a dataset of game replay data from human expert players. The paper also presents initial baseline results for neural networks trained on this data to predict game outcomes and player actions, as well as results for canonical deep reinforcement learning agents applied to the StarCraft II domain. These agents learn to achieve a level of play comparable to a novice player on mini-games but struggle to make significant progress on the main game. The paper also describes the environment's observation, action, and reward specifications, as well as the architecture of the StarCraft II API and PySC2, an open-source environment optimized for RL agents. The environment includes a variety of mini-games that focus on different aspects of StarCraft II gameplay, and the paper provides results for these mini-games. The paper also discusses the performance of different RL algorithms on the environment, including A3C, Atari-net, FullyConv, and FullyConv LSTM agents. The results show that while these agents perform well on mini-games, they struggle to learn effective strategies for the full game. The paper also discusses the use of supervised learning from replays to train value functions and policies for StarCraft II, and presents results for these approaches. Overall, the paper highlights the challenges of reinforcement learning in the StarCraft II environment and the potential of SC2LE as a benchmark for deep reinforcement learning research.This paper introduces SC2LE (StarCraft II Learning Environment), a reinforcement learning environment based on the game StarCraft II. The environment presents a new grand challenge for reinforcement learning, involving multi-agent interaction, imperfect information, a large action space, a large state space, and delayed credit assignment. The environment provides an open-source Python-based interface for interacting with the game engine, along with a suite of mini-games and a dataset of game replay data from human expert players. The paper also presents initial baseline results for neural networks trained on this data to predict game outcomes and player actions, as well as results for canonical deep reinforcement learning agents applied to the StarCraft II domain. These agents learn to achieve a level of play comparable to a novice player on mini-games but struggle to make significant progress on the main game. The paper also describes the environment's observation, action, and reward specifications, as well as the architecture of the StarCraft II API and PySC2, an open-source environment optimized for RL agents. The environment includes a variety of mini-games that focus on different aspects of StarCraft II gameplay, and the paper provides results for these mini-games. The paper also discusses the performance of different RL algorithms on the environment, including A3C, Atari-net, FullyConv, and FullyConv LSTM agents. The results show that while these agents perform well on mini-games, they struggle to learn effective strategies for the full game. The paper also discusses the use of supervised learning from replays to train value functions and policies for StarCraft II, and presents results for these approaches. Overall, the paper highlights the challenges of reinforcement learning in the StarCraft II environment and the potential of SC2LE as a benchmark for deep reinforcement learning research.