Struct StepResult<TObs>
- Namespace
- Gym.Observations
- Assembly
- Gym.dll
One environment transition, (observation, reward, terminated, truncated, info). Deconstructs like a tuple:
var (obs, reward, terminated, truncated, info) = env.Step(action);.
public readonly record struct StepResult<TObs> : INDArrayCarrier, IDisposable, IEquatable<StepResult<TObs>>
Type Parameters
TObsThe observation type, an element of the env's observation space.
- Implements
-
IEquatable<StepResult<TObs>>
- Inherited Members
- Extension Methods
Remarks
Equality is exact and never broadcasts: observations that are arrays must have the same shape, the same dtype and
identical elements (a NaN element equals nothing, not even itself); any other observation compares as Python's
== does (PyEquality), and so does the info. The reward compares with
Equals(double), which, unlike Python's ==, treats NaN as equal to NaN so equality stays
reflexive. Reference: gymnasium/core.py (Env.step).
Constructors
StepResult(TObs, double, bool, bool, Info)
One environment transition, (observation, reward, terminated, truncated, info). Deconstructs like a tuple:
var (obs, reward, terminated, truncated, info) = env.Step(action);.
public StepResult(TObs Observation, double Reward, bool Terminated, bool Truncated, Info Info)
Parameters
ObservationTObsThe observation after the action; owned by the caller (envs return a fresh array).
RewarddoubleThe reward as float64, the precision of Gymnasium's Python floats; a float32 reward would change every accumulated return.
TerminatedboolThe MDP reached a terminal state: the value of what follows is zero, so do not bootstrap past it. Call
Resetbefore stepping again.TruncatedboolSomething outside the MDP ended the episode, typically a
TimeLimit: bootstrapping from the final observation stays valid. CallResetbefore stepping again.InfoInfoDiagnostics with Python dict semantics (insertion order is observable); never null from Gym's own envs.
Remarks
Equality is exact and never broadcasts: observations that are arrays must have the same shape, the same dtype and
identical elements (a NaN element equals nothing, not even itself); any other observation compares as Python's
== does (PyEquality), and so does the info. The reward compares with
Equals(double), which, unlike Python's ==, treats NaN as equal to NaN so equality stays
reflexive. Reference: gymnasium/core.py (Env.step).
Properties
Info
Diagnostics with Python dict semantics (insertion order is observable); never null from Gym's own envs.
public Info Info { get; init; }
Property Value
Information
0.2's name for Info (0.2's Step.Information), kept so 0.2 code compiles until 2.0 (DR-78):
the same dict, never null from Gym's own envs where 0.2's could be.
[Obsolete("0.2's name: read Info (or deconstruct the info). Removed in 2.0.")]
public Info Information { get; }
Property Value
Observation
The observation after the action; owned by the caller (envs return a fresh array).
public TObs Observation { get; init; }
Property Value
- TObs
Reward
The reward as float64, the precision of Gymnasium's Python floats; a float32 reward would change every accumulated return.
public double Reward { get; init; }
Property Value
Terminated
The MDP reached a terminal state: the value of what follows is zero, so do not bootstrap past it. Call
Reset before stepping again.
public bool Terminated { get; init; }
Property Value
Truncated
Something outside the MDP ended the episode, typically a TimeLimit: bootstrapping from the final observation
stays valid. Call Reset before stepping again.
public bool Truncated { get; init; }
Property Value
Methods
Deconstruct(out TObs, out double, out bool, out Info)
0.2's four-element deconstruction, var (obs, reward, done, info) = env.Step(action);, kept so 0.2 code
compiles until 2.0 (DR-58): done is terminated || truncated, as gym 0.21 and
Gym.Legacy report it.
[Obsolete("The pre-0.26 4-tuple: done is terminated || truncated, which loses why the episode ended. Deconstruct into (observation, reward, terminated, truncated, info). Removed in 2.0 with Gym.Legacy.")]
public void Deconstruct(out TObs observation, out double reward, out bool done, out Info info)
Parameters
observationTObsThe observation after the action.
rewarddoubleThe reward, float64 (0.2's was a float, so a 0.2 line declaring
float rewardstill fails).doneboolWhether the episode ended, for either reason.
infoInfoThe info.
Remarks
Footgun: done loses why the episode ended, and bootstrapping from the final observation is
right after a truncation but wrong after a termination. The five-element deconstruction keeps both flags.
Dispose()
Releases the arrays the transition holds, the observation's and the info's, nested ones included, so
using var step = env.Step(action); frees them at the end of the scope: the arrays an env returns belong to
the caller (D19).
public void Dispose()
Remarks
Footgun: copies of this struct share the same arrays, so dispose once, after the last use of the observation
and the info; keep anything needed longer with .copy(). Disposing again does nothing more.
Equals(StepResult<TObs>)
Exact equality (see the type's remarks): observations, rewards, flags and infos all equal. Replaces the
compiler's member-wise equality, which would call NDArray's broadcasting Equals.
public bool Equals(StepResult<TObs> other)
Parameters
otherStepResult<TObs>The other transition.
Returns
GetHashCode()
A hash consistent with Equals(StepResult<TObs>): an array observation contributes its shape only, because GetHashCode() throws (arrays are mutable, hence unhashable in NumPy).
public override int GetHashCode()
Returns
- int
The hash.