Table of Contents

Class EnvChecker

Namespace
Gym.Utils
Assembly
Gym.dll

Gymnasium's gymnasium.utils.env_checker: the checks an env must pass to follow the API. Today it holds DataEquivalence(object?, object?, bool), which tests and checkers use to compare observations, actions and infos; the env checks themselves arrive with the registry (P3).

public static class EnvChecker
Inheritance
EnvChecker
Inherited Members

Remarks

Methods

CheckEnv(IEnv, bool?, bool, bool)

Checks that an env follows Gymnasium's API (Gymnasium's check_env): its spaces, then reset (return type, seed determinism, options), the passive checks of a reset and a step, step determinism, fresh data on every call, every declared render mode, and a double close. Problems that make the env unusable are assertions; questionable choices are warnings through Logger.

public static void CheckEnv(IEnv env, bool? warn = null, bool skipRenderCheck = false, bool skipCloseCheck = false)

Parameters

env IEnv

The env, ideally unwrapped (a wrapper is checked too, with a warning).

warn bool?

Ignored, as in Gymnasium; giving it warns that it is.

skipRenderCheck bool

Skip the render checks, which build one env per declared render mode.

skipCloseCheck bool

Skip building an env from the spec to close it twice.

Remarks

The render checks build an env for every mode in metadata["render_modes"] through the env's spec, so human needs a viewer (Gymnasium.HumanViewerFactory, DR-20) as Gymnasium's needs pygame; pass skipRenderCheck where there is none. Gymnasium's JAX and PyTorch conversions do not apply: Gym.NET's envs return NumSharp arrays.

The env is reset and stepped many times and left in whatever state the last check put it. The observations the checks collect are left to the garbage collector rather than disposed: an env that wrongly hands out its own buffer, which CheckReturnedDataNotReused(IEnv, long) reports, must not have it disposed under it first.

Reference: gymnasium/utils/env_checker.py.

Exceptions

TypeError

Thrown when env is null, or the render checks find a render_modes that is not iterable.

AttributeError

Thrown when the env has no action or observation space.

AssertionError

Thrown for the first violation found, with Gymnasium's message.

KeyError

Thrown when the render checks run and the metadata has no render_modes.

CheckResetOptions(IEnv)

Checks that the env can be reset with options (Gymnasium's check_reset_options): reset(options={}) must not raise TypeError.

public static void CheckResetOptions(IEnv env)

Parameters

env IEnv

The env.

Remarks

Exceptions

AssertionError

Thrown when the reset raises TypeError, with Gymnasium's message.

CheckResetReturnInfoDeprecation(IEnv)

Warns when the env's reset still takes the pre-0.26 return_info argument (Gymnasium's check_reset_return_info_deprecation): a public Reset overload with a parameter of that Python name.

public static void CheckResetReturnInfoDeprecation(IEnv env)

Parameters

env IEnv

The env.

Remarks

CheckResetReturnType(IEnv)

Checks what an unseeded reset returns (Gymnasium's check_reset_return_type): an observation within the space and an info dict. The 2-tuple form Gymnasium also asserts is a ResetResult<TObs> by type.

public static void CheckResetReturnType(IEnv env)

Parameters

env IEnv

The env.

Remarks

Exceptions

AssertionError

Thrown for an observation outside the space or a null info, with Gymnasium's messages.

CheckResetSeedDeterminism(IEnv)

Checks that seeded resets are deterministic and reseed the env's generator (Gymnasium's check_reset_seed_determinism): two reset(seed=123) give equivalent observations and the same generator state, the unseeded resets after them agree too, and reset(seed=456) moves the generator elsewhere.

public static void CheckResetSeedDeterminism(IEnv env)

Parameters

env IEnv

The env.

Remarks

The determinism of the observations is asserted only for an env whose spec says it is deterministic, as in Gymnasium; the generator checks always run. Gymnasium's checks of the signature have no counterpart: every Reset takes a seed, and C# allows a BigInteger? parameter no default but null, so the warning about a seed default other than None cannot arise. Reference: gymnasium/utils/env_checker.py.

Exceptions

AssertionError

Thrown for an observation outside the space, a reset that leaves no generator or does not reseed it (super().reset(seed=seed) missing), or observations that differ for the same seed, with Gymnasium's messages; and for a TypeError raised by the resets, which Gymnasium turns into one.

CheckReturnedDataNotReused(IEnv, long)

Checks that every reset and step returns new observation and info data (Gymnasium's check_returned_data_not_reused): a reset, up to two steps and another reset, compared pairwise with DataSharesObjects(object?, object?).

public static void CheckReturnedDataNotReused(IEnv env, long seed = 123)

Parameters

env IEnv

The env.

seed long

The seed of the action space and of both resets.

Remarks

Identity, not value, is compared, so even an array an env means to be constant must be a new array in every result: neither the user nor this check can know a later call will not change it. Reference: gymnasium/utils/env_checker.py.

Exceptions

AssertionError

Thrown when two calls returned data sharing an object, naming both calls.

CheckSeedDeprecation(IEnv)

Warns when the env still has the pre-0.26 seed function (Gymnasium's check_seed_deprecation, which tests callable(getattr(env, "seed", None))): a public instance method named Seed, or a seed attribute set at run time to a function. Seeding is Reset(seed: s)'s job.

public static void CheckSeedDeprecation(IEnv env)

Parameters

env IEnv

The env.

Remarks

A run-time seed attribute that is not a function (a number, a list) is not one, as in Python. Reference: gymnasium/utils/env_checker.py.

CheckSpaceLimit(Space, string)

Warns about box bounds that make learning hard (Gymnasium's check_space_limit, run only by CheckEnv(IEnv, bool?, bool, bool)): infinite bounds, and for a 1-d action box any range other than a symmetric or [0, 1] one within [-1, 1]. Tuples and dicts are checked subspace by subspace.

public static void CheckSpaceLimit(Space space, string spaceType)

Parameters

space Space

The space.

spaceType string

observation or action.

Remarks

CheckStepDeterminism(IEnv, long)

Checks that one step after a seeded reset is deterministic (Gymnasium's check_step_determinism): the same seed and action give the same generator state, equivalent observation, reward and info, the same termination, and no truncation. An env whose spec is nondeterministic is skipped.

public static void CheckStepDeterminism(IEnv env, long seed = 123)

Parameters

env IEnv

The env, which must already pass the reset checks.

seed long

The seed of the action space and of both resets.

Remarks

Exceptions

AssertionError

Thrown for the first difference, with Gymnasium's message.

DataEquivalence(object?, object?, bool)

Whether two values are equivalent (Gymnasium's data_equivalence): of the same Python type, dicts with the same keys and equivalent values, tuples and lists of equivalent items, arrays of the same shape and dtype with close (or, when exact, equal) entries, and anything else equal by Python's ==.

public static bool DataEquivalence(object? data1, object? data2, bool exact = false)

Parameters

data1 object

The first value.

data2 object

The second value.

exact bool

Whether arrays must be equal entry for entry rather than within np.allclose(rtol=1e-5, atol=1e-5).

Returns

bool

true when equivalent.

Remarks

The type test is on the values' Python counterparts (PyBuiltins): every .NET integer is an int, an object?[] or ITuple a tuple, a typed array or list a list, a 0-d array the NumPy scalar it stands for (DR-08), so 1 and 1L are equivalent but a tuple never matches a list, as in Gymnasium.

A 0-d array compares with ==, exactly, as Gymnasium compares NumPy scalars (only arrays of one or more dimensions go through allclose); a NaN therefore never equals itself. Reference: gymnasium/utils/env_checker.py (data_equivalence).