google-deepmind/pysc2

Features

• Reinforcement Learning Environments - Wraps StarCraft II as a Python reinforcement learning environment where agents receive observations, send actions, and train on mini-games and full matches.
• Agent Action Spaces - Defines how a game is presented to a learning agent, with typed function-call actions and per-frame lists of currently valid actions.
• Agent-Environment Interaction Loops - Drives one or two players through repeated observe-and-act cycles in the same game session, stepping a configurable number of frames each time.
• Feature Layer Observations - Exposes about twenty-five structured feature layers for the screen and minimap, giving agents meaningful grids instead of raw pixels.
• Game Environment Configuration - Lets users choose the map, players and difficulty, observation resolution, action space, and frames skipped per step when setting up a game environment.
• Game Observation and Action Converters - Convert raw game protocol messages into neural-network-friendly dictionaries of numeric arrays with matching specs, so learning agents can read observations and send actions.
• Observation and Action Converters - Converts raw game protocol messages into dictionaries of numeric arrays with matching specs so neural networks can consume observations and emit actions.
• Spatial Feature Layers - Breaks the screen and minimap into stacked grids of unit attributes so agents read meaningful values instead of raw pixels.
• Real-Time Strategy Game Agents - Supports training and evaluating agents that learn to play StarCraft II through observe-and-act loops against built-in opponents or other agents.
• Game Simulation Environments - Sets up configurable simulated game worlds with chosen maps, opponents, difficulty, and scoring so researchers can run controlled machine learning experiments.
• Agent Map Launchers - Launches a random, scripted, or custom agent on a selected map with a visual interface, or pits two agents against each other.
• Mini-Game Training Scenarios - Offers single-player, fixed-length mini-game scenarios that each exercise one skill and expose a score or reward separating poor from strong play.
• Multi-Modal Game Observations - Offers structured feature layers, rendered RGB images, and a raw unit-data interface so agents can choose how to see the game.
• Replay Processing Pipelines - Plays back recorded games in parallel and converts them into observation and action arrays for building offline training datasets.
• Pixel-Based Game Agent Training - Trains agents on small focused scenarios that each exercise one skill, using score-based rewards to build toward full-length matches.
• RGB Pixel Observations - Provides the main screen and minimap as color images at a chosen resolution, matching a human player's view without menus.
• Spatial Observation Planes - Represents most observations as spatial planes of unit attributes for screen and minimap, and accepts point-and-click actions like a human player.
• Visual Observation Interfaces - Presents game state to agents as stacked feature layers and RGB images of the screen and minimap, plus non-visual tensors.
• Game Replay Parsers - Runs recorded replays through a converter to produce the same array-based data format as live environments, enabling offline training on past games.
• Custom Map Definitions - Defines custom maps by subclassing a base class that sets file location, episode length, player count, and scoring, then registers them on import.

Star history