Procgen

We use procgen==0.10.7 as the codebase. See https://github.com/openai/procgen/tree/0.10.7

Note

Linux manylinux_2_28 wheels bundle Qt 6.11.1 QtCore/QtGui, built from source in the release image. Procgen does not require a system Qt installation.

macOS wheels require Homebrew Qt 6 at runtime: run brew install qtbase before importing EnvPool. Windows wheels bundle their Qt runtime DLLs.

Render Compare

Representative first-frame compares for Procgen tasks that support rendering. In each panel, EnvPool is on the left and the exact channel-last observation oracle used by the test suite is on the right.

../_images/procgen_oracle_compare.png

Options

  • task_id (str): see available tasks below;

  • num_envs (int): how many environments you would like to create;

  • batch_size (int): the expected batch size for return result, default to num_envs;

  • num_threads (int): the maximum thread number for executing the actual env.step, default to batch_size;

  • seed (int | Sequence[int]): the environment seed. When a sequence is provided, it must contain exactly one seed per environment. Default to 42;

  • max_episode_steps (int): the maximum number of steps for one episode, each procgen game has different timeout value;

  • channel_first (bool): whether to transpose the observation image to (3, 64, 64), default to True;

  • env_name (str): one of 16 procgen env name;

  • num_levels (int): default to 0;

  • start_level (int): default to 0;

  • use_sequential_levels (bool): default to False;

  • center_agent (bool): default to True;

  • use_backgrounds (bool): default to True;

  • use_monochrome_assets (bool): default to False;

  • restrict_themes (bool): default to False;

  • use_generated_assets (bool): default to False;

  • paint_vel_info (bool): default to False;

  • use_easy_jump (bool): default to False;

  • distribution_mode (int): one of (0, 1, 2, 10, 20); 0 stands for easy mode, 1 stands for hard mode, 2 stands for extreme mode, 10 stands for memory mode, and 20 stands for exploration mode. Exploration mode follows Procgen’s Python wrapper behavior: it uses hard mode levels, overrides num_levels to 1, and overrides start_level to a game-specific exploration seed. The default value is determined by task_id.

Note: arguments after env_name are provided by procgen environment itself. We keep the default value as-is. We haven’t tested the setting of use_sequential_levels == True, and have no promise it is aligned with the original version of procgen (PRs for fixing this issue are highly welcome).

Observation Space

The observation image shape is (3, 64, 64) when channel_first is True (default), (64, 64, 3) when channel_first is False.

Action Space

15 action buttons in total, ranging from 0 to 14.

Available Tasks

  • BigfishEasy-v0

  • BigfishHard-v0

  • BossfightEasy-v0

  • BossfightHard-v0

  • CaveflyerEasy-v0

  • CaveflyerHard-v0

  • CaveflyerMemory-v0

  • CaveflyerExploration-v0

  • ChaserEasy-v0

  • ChaserHard-v0

  • ChaserExtreme-v0

  • ClimberEasy-v0

  • ClimberHard-v0

  • ClimberExploration-v0

  • CoinrunEasy-v0

  • CoinrunHard-v0

  • CoinrunExploration-v0

  • DodgeballEasy-v0

  • DodgeballHard-v0

  • DodgeballExtreme-v0

  • DodgeballMemory-v0

  • FruitbotEasy-v0

  • FruitbotHard-v0

  • HeistEasy-v0

  • HeistHard-v0

  • HeistMemory-v0

  • HeistExploration-v0

  • JumperEasy-v0

  • JumperHard-v0

  • JumperMemory-v0

  • JumperExploration-v0

  • LeaperEasy-v0

  • LeaperHard-v0

  • LeaperExtreme-v0

  • LeaperExploration-v0

  • MazeEasy-v0

  • MazeHard-v0

  • MazeMemory-v0

  • MazeExploration-v0

  • MinerEasy-v0

  • MinerHard-v0

  • MinerMemory-v0

  • NinjaEasy-v0

  • NinjaHard-v0

  • NinjaExploration-v0

  • PlunderEasy-v0

  • PlunderHard-v0

  • StarpilotEasy-v0

  • StarpilotHard-v0

  • StarpilotExtreme-v0