Utility functions

Seeding

gymnasium.utils.seeding.np_random(seed: int | None = None) → tuple[np.random.Generator, int][source]

Returns a NumPy random number generator (RNG) along with seed value from the inputted seed.

If seed is None then a random seed will be generated as the RNG’s initial seed. This randomly selected seed is returned as the second value of the tuple.

This function is called in reset() to reset an environment’s initial RNG.

Parameters:

seed – The seed used to create the generator

Returns:

A NumPy-based Random Number Generator and generator seed

Raises:

Error – Seed must be a non-negative integer

Environment Checking

gymnasium.utils.env_checker.check_env(env: Env, warn: bool | None = None, skip_render_check: bool = False, skip_close_check: bool = False)[source]

Check that an environment follows Gymnasium’s API.

To ensure that an environment is implemented “correctly”, check_env checks that the observation_space and action_space are correct. Furthermore, the function will call the reset(), step() and render() functions with a variety of values.

We highly recommend users calling this function after an environment is constructed and within a projects continuous integration to keep an environment update with Gymnasium’s API.

Parameters:

  • env – The Gym environment that will be checked

  • warn – Ignored, previously silenced particular warnings

  • skip_render_check – Whether to skip the checks for the render method. False by default (useful for the CI)

  • skip_close_check – Whether to skip the checks for the close method. False by default

Visualization

gymnasium.utils.play.play(env: Env, transpose: bool | None = True, fps: int | None = None, zoom: float | None = None, callback: Callable | None = None, keys_to_action: dict[tuple[str | int, ...] | str | int, ActType] | None = None, seed: int | None = None, noop: ActType = 0)[source]

Allows the user to play the environment using a keyboard.

Parameters:

  • env – Environment to use for playing.

  • transpose – If this is True, the output of observation is transposed. Defaults to True.

  • fps – Maximum number of steps of the environment executed every second. If None (the default), env.metadata["render_fps""] (or 30, if the environment does not specify “render_fps”) is used.

  • zoom – Zoom the observation in, zoom amount, should be positive float

  • callback –

    If a callback is provided, it will be executed after every step. It takes the following input:

    • obs_t: observation before performing action

    • obs_tp1: observation after performing action

    • action: action that was executed

    • rew: reward that was received

    • terminated: whether the environment is terminated or not

    • truncated: whether the environment is truncated or not

    • info: debug info

  • keys_to_action –

    Mapping from keys pressed to action performed. Different formats are supported: Key combinations can either be expressed as a tuple of unicode code points of the keys, as a tuple of characters, or as a string where each character of the string represents one key. For example if pressing ‘w’ and space at the same time is supposed to trigger action number 2 then key_to_action dict could look like this:

    >>> key_to_action = {
...    # ...
...    (ord('w'), ord(' ')): 2
...    # ...
... }

    or like this:

    >>> key_to_action = {
...    # ...
...    ("w", " "): 2
...    # ...
... }

    or like this:

    >>> key_to_action = {
...    # ...
...    "w ": 2
...    # ...
... }

    If None, default key_to_action mapping for that environment is used, if provided.

  • seed – Random seed used when resetting the environment. If None, no seed is used.

  • noop – The action used when no key input has been entered, or the entered key combination is unknown.

Example

>>> import gymnasium as gym
>>> import numpy as np
>>> from gymnasium.utils.play import play
>>> play(gym.make("CarRacing-v2", render_mode="rgb_array"),  
...     keys_to_action={
...         "w": np.array([0, 0.7, 0]),
...         "a": np.array([-1, 0, 0]),
...         "s": np.array([0, 0, 1]),
...         "d": np.array([1, 0, 0]),
...         "wa": np.array([-1, 0.7, 0]),
...         "dw": np.array([1, 0.7, 0]),
...         "ds": np.array([1, 0, 1]),
...         "as": np.array([-1, 0, 1]),
...     },
...     noop=np.array([0, 0, 0])
... )

Above code works also if the environment is wrapped, so it’s particularly useful in verifying that the frame-level preprocessing does not render the game unplayable.

If you wish to plot real time statistics as you play, you can use PlayPlot. Here’s a sample code for plotting the reward for last 150 steps.

>>> from gymnasium.utils.play import PlayPlot, play
>>> def callback(obs_t, obs_tp1, action, rew, terminated, truncated, info):
...        return [rew,]
>>> plotter = PlayPlot(callback, 150, ["reward"])             
>>> play(gym.make("CartPole-v1"), callback=plotter.callback)
class gymnasium.utils.play.PlayPlot(callback: Callable, horizon_timesteps: int, plot_names: list[str])[source]

Provides a callback to create live plots of arbitrary metrics when using play().

This class is instantiated with a function that accepts information about a single environment transition:

  • obs_t: observation before performing action

  • obs_tp1: observation after performing action

  • action: action that was executed

  • rew: reward that was received

  • terminated: whether the environment is terminated or not

  • truncated: whether the environment is truncated or not

  • info: debug info

It should return a list of metrics that are computed from this data. For instance, the function may look like this:

>>> def compute_metrics(obs_t, obs_tp, action, reward, terminated, truncated, info):
...     return [reward, info["cumulative_reward"], np.linalg.norm(action)]

PlayPlot provides the method callback() which will pass its arguments along to that function and uses the returned values to update live plots of the metrics.

Typically, this callback() will be used in conjunction with play() to see how the metrics evolve as you play:

>>> plotter = PlayPlot(compute_metrics, horizon_timesteps=200,                               
...                    plot_names=["Immediate Rew.", "Cumulative Rew.", "Action Magnitude"])
>>> play(your_env, callback=plotter.callback)
Parameters:

  • callback – Function that computes metrics from environment transitions

  • horizon_timesteps – The time horizon used for the live plots

  • plot_names – List of plot titles

Raises:

DependencyNotInstalled – If matplotlib is not installed

callback(obs_t: ObsType, obs_tp1: ObsType, action: ActType, rew: float, terminated: bool, truncated: bool, info: dict)[source]

The callback that calls the provided data callback and adds the data to the plots.

Parameters:

  • obs_t – The observation at time step t

  • obs_tp1 – The observation at time step t+1

  • action – The action

  • rew – The reward

  • terminated – If the environment is terminated

  • truncated – If the environment is truncated

  • info – The information from the environment

class gymnasium.utils.play.PlayableGame(env: Env, keys_to_action: dict[tuple[int, ...], int] | None = None, zoom: float | None = None)[source]

Wraps an environment allowing keyboard inputs to interact with the environment.

Parameters:

  • env – The environment to play

  • keys_to_action – The dictionary of keyboard tuples and action value

  • zoom – If to zoom in on the environment render

process_event(event: Event)[source]

Processes a PyGame event.

In particular, this function is used to keep track of which buttons are currently pressed and to exit the play() function when the PyGame window is closed.

Parameters:

event – The event to process

Environment pickling

class gymnasium.utils.ezpickle.EzPickle(*args: Any, **kwargs: Any)[source]

Objects that are pickled and unpickled via their constructor arguments.

Example

>>> class Animal: pass
>>> class Dog(Animal, EzPickle):
...    def __init__(self, furcolor, tailkind="bushy"):
...        Animal.__init__(self)
...        EzPickle.__init__(self, furcolor, tailkind)

When this object is unpickled, a new Dog will be constructed by passing the provided furcolor and tailkind into the constructor. However, philosophers are still not sure whether it is still the same dog.

This is generally needed only for environments which wrap C/C++ code, such as MuJoCo and Atari.

Uses the args and kwargs from the object’s constructor for pickling.

Save Rendering Videos

gymnasium.utils.save_video.save_video(frames: list, video_folder: str, episode_trigger: Callable[[int], bool] = None, step_trigger: Callable[[int], bool] = None, video_length: int | None = None, name_prefix: str = 'rl-video', episode_index: int = 0, step_starting_index: int = 0, save_logger: str | None = None, **kwargs)[source]

Save videos from rendering frames.

This function extract video from a list of render frame episodes.

Parameters:

  • frames (List[RenderFrame]) – A list of frames to compose the video.

  • video_folder (str) – The folder where the recordings will be stored

  • episode_trigger – Function that accepts an integer and returns True iff a recording should be started at this episode

  • step_trigger – Function that accepts an integer and returns True iff a recording should be started at this step

  • video_length (int) – The length of recorded episodes. If it isn’t specified, the entire episode is recorded. Otherwise, snippets of the specified length are captured.

  • name_prefix (str) – Will be prepended to the filename of the recordings.

  • episode_index (int) – The index of the current episode.

  • step_starting_index (int) – The step index of the first frame.

  • save_logger – If to log the video saving progress, helpful for long videos that take a while, use “bar” to enable.

  • **kwargs – The kwargs that will be passed to moviepy’s ImageSequenceClip. You need to specify either fps or duration.

Example

>>> import gymnasium as gym
>>> from gymnasium.utils.save_video import save_video
>>> env = gym.make("FrozenLake-v1", render_mode="rgb_array_list")
>>> _ = env.reset()
>>> step_starting_index = 0
>>> episode_index = 0
>>> for step_index in range(199): 
...    action = env.action_space.sample()
...    _, _, terminated, truncated, _ = env.step(action)
...
...    if terminated or truncated:
...       save_video(
...          frames=env.render(),
...          video_folder="videos",
...          fps=env.metadata["render_fps"],
...          step_starting_index=step_starting_index,
...          episode_index=episode_index
...       )
...       step_starting_index = step_index + 1
...       episode_index += 1
...       env.reset()
>>> env.close()
gymnasium.utils.save_video.capped_cubic_video_schedule(episode_id: int) → bool[source]

The default episode trigger.

This function will trigger recordings at the episode indices \(\{0, 1, 4, 8, 27, ..., k^3, ..., 729, 1000, 2000, 3000, ...\}\)

Parameters:

episode_id – The episode number

Returns:

If to apply a video schedule number

Old to New Step API Compatibility

gymnasium.utils.step_api_compatibility.step_api_compatibility(step_returns: TerminatedTruncatedStepType | DoneStepType, output_truncation_bool: bool = True, is_vector_env: bool = False) → TerminatedTruncatedStepType | DoneStepType[source]

Function to transform step returns to the API specified by output_truncation_bool.

Done (old) step API refers to step() method returning (observation, reward, done, info) Terminated Truncated (new) step API refers to step() method returning (observation, reward, terminated, truncated, info) (Refer to docs for details on the API change)

Parameters:

  • step_returns (tuple) – Items returned by step(). Can be (obs, rew, done, info) or (obs, rew, terminated, truncated, info)

  • output_truncation_bool (bool) – Whether the output should return two booleans (new API) or one (old) (True by default)

  • is_vector_env (bool) – Whether the step_returns are from a vector environment

Returns:

step_returns (tuple) – Depending on output_truncation_bool, it can return (obs, rew, done, info) or (obs, rew, terminated, truncated, info)

Example

This function can be used to ensure compatibility in step interfaces with conflicting API. E.g. if env is written in old API, wrapper is written in new API, and the final step output is desired to be in old API.

>>> import gymnasium as gym
>>> env = gym.make("CartPole-v0")
>>> _, _ = env.reset()
>>> obs, reward, done, info = step_api_compatibility(env.step(0), output_truncation_bool=False)
>>> obs, reward, terminated, truncated, info = step_api_compatibility(env.step(0), output_truncation_bool=True)

>>> vec_env = gym.make_vec("CartPole-v0", vectorization_mode="sync")
>>> _, _ = vec_env.reset()
>>> obs, rewards, dones, infos = step_api_compatibility(vec_env.step([0]), is_vector_env=True, output_truncation_bool=False)
>>> obs, rewards, terminations, truncations, infos = step_api_compatibility(vec_env.step([0]), is_vector_env=True, output_truncation_bool=True)
gymnasium.utils.step_api_compatibility.convert_to_terminated_truncated_step_api(step_returns: DoneStepType | TerminatedTruncatedStepType, is_vector_env=False) → TerminatedTruncatedStepType[source]

Function to transform step returns to new step API irrespective of input API.

Parameters:

  • step_returns (tuple) – Items returned by step(). Can be (obs, rew, done, info) or (obs, rew, terminated, truncated, info)

  • is_vector_env (bool) – Whether the step_returns are from a vector environment

gymnasium.utils.step_api_compatibility.convert_to_done_step_api(step_returns: TerminatedTruncatedStepType | DoneStepType, is_vector_env: bool = False) → DoneStepType[source]

Function to transform step returns to old step API irrespective of input API.

Parameters:

  • step_returns (tuple) – Items returned by step(). Can be (obs, rew, done, info) or (obs, rew, terminated, truncated, info)

  • is_vector_env (bool) – Whether the step_returns are from a vector environment

Runtime Performance benchmark

Sometimes is neccary to measure your environment’s runtime performance, and ensure no performance regressions take place. These tests require manual inspection of its outputs:

gymnasium.utils.performance.benchmark_step(env: Env, target_duration: int = 5, seed=None) → float[source]

A benchmark to measure the runtime performance of step for an environment.

example usage:

`py env_old = ... old_throughput = benchmark_step(env_old) env_new = ... new_throughput = benchmark_step(env_old) slowdown = old_throughput / new_throughput `

Parameters:

  • env – the environment to benchmarked.

  • target_duration – the duration of the benchmark in seconds (note: it will go slightly over it).

  • seed – seeds the environment and action sampled.

Returns: the average steps per second.

gymnasium.utils.performance.benchmark_init(env_lambda: Callable[[], Env], target_duration: int = 5, seed=None) → float[source]

A benchmark to measure the initialization time and first reset.

Parameters:

  • env_lambda – the function to initialize the environment.

  • target_duration – the duration of the benchmark in seconds (note: it will go slightly over it).

  • seed – seeds the first reset of the environment.

gymnasium.utils.performance.benchmark_render(env: Env, target_duration: int = 5) → float[source]

A benchmark to measure the time of render().

Note: does not work with render_mode=’human’ :param env: the environment to benchmarked (Note: must be renderable). :param target_duration: the duration of the benchmark in seconds (note: it will go slightly over it).