Class PusherEnvV5
- Namespace
- Gym.Environments.Envs.MuJoCo
- Assembly
- Gym.Environments.dll
Pusher-v5: a 7-joint arm that pushes a cylinder to a goal on a table, driven by 7 joint torques in [-2, 2]. The
reward charges the cylinder's distance to the goal, half the fingertip's distance to the cylinder and the squared
torques; the episode never ends on its own (make's 100-step limit truncates it). Byte for byte what
Gymnasium's Pusher-v5 computes on MuJoCo 4.0.0 (DR-61, DR-64).
[PyName("PusherEnv")]
public class PusherEnvV5 : MujocoEnv, IEnv, IDisposable, IDynamicAttributes
- Inheritance
-
PusherEnvV5
- Implements
- Inherited Members
- Extension Methods
Remarks
Observation (float64, 23 values): the arm's 7 joint angles and velocities, then the positions of the
fingertip, the cylinder and the goal. Reward: -|object - goal| * reward_dist_weight - sum(action**2) *
reward_control_weight - |object - tips_arm| * reward_near_weight, the control term in the action's dtype.
Info: reward_dist, reward_ctrl, reward_near; the reset's info is empty.
Footgun, as in Gymnasium: the reset writes the cylinder's and the goal's start into InitQpos
itself (Gymnasium's qpos = self.init_qpos is no copy), so init_qpos's last four values change on
every reset. The dynamics are unaffected: those four are overwritten before they are used.
Not thread-safe. Reference: gymnasium/envs/mujoco/pusher_v5.py.
Its info reads typed (DR-91): PusherInfoV5 info = step.Info; (PusherInfoV5), with the dict's own values still behind the indexer.
Constructors
PusherEnvV5(PusherEnvV5Options)
Builds the instance from its options record: the properties set, the constructor's defaults for the rest.
public PusherEnvV5(PusherEnvV5Options options)
Parameters
optionsPusherEnvV5OptionsThe options.
Exceptions
- ArgumentNullException
Thrown when
optionsis null.
PusherEnvV5(RenderMode, string, int, IReadOnlyDictionary<string, object?>?, double, double, double, int, int, int?, string?, int, IReadOnlyDictionary<int, bool>?, EnvViewerFactory?)
Builds the env with the render mode given as the RenderMode enum (DR-84, DR-93), as
PusherEnvV5(string, int, IReadOnlyDictionary<string, object?>?, double, double, double, string?, int, int, int?, string?, int, IReadOnlyDictionary<int, bool>?, EnvViewerFactory?)
builds it from the mode's string: new PusherEnvV5(RenderMode.RgbArray) builds what
new PusherEnvV5(renderMode: "rgb_array") builds, the member reaching the env as Gymnasium's string
(Name(RenderMode)), which its RenderMode property then holds.
public PusherEnvV5(RenderMode renderMode, string xmlFile = "pusher_v5.xml", int frameSkip = 5, IReadOnlyDictionary<string, object?>? defaultCameraConfig = null, double rewardNearWeight = 0.5, double rewardDistWeight = 1, double rewardControlWeight = 0.1, int width = 480, int height = 480, int? cameraId = null, string? cameraName = null, int maxGeom = 1000, IReadOnlyDictionary<int, bool>? visualOptions = null, EnvViewerFactory? viewerFactory = null)
Parameters
renderModeRenderModeThe render mode, which the env must list in its
metadata["render_modes"]as for the string; it behaves exactly as its Gymnasium string does in the other constructor.xmlFilestringThe model: a bundled file name (
xml_file, DR-63);"pusher.xml"is v4's.frameSkipintPhysics steps per env step (
frame_skip).defaultCameraConfigIReadOnlyDictionary<string, object>The free camera's settings (
default_camera_config); null is Gymnasium's default (trackbodyid-1, 4 m away), and an empty dictionary leaves MuJoCo's camera as it is.rewardNearWeightdoubleThe weight of the fingertip's distance to the cylinder (
reward_near_weight).rewardDistWeightdoubleThe weight of the cylinder's distance to the goal (
reward_dist_weight; Gymnasium's default is the int 1).rewardControlWeightdoubleThe weight of the squared torques (
reward_control_weight).widthintThe rendered frame's width (
width).heightintThe rendered frame's height (
height).cameraIdint?The camera to render from (
camera_id).cameraNamestringThe camera to render from by name (
camera_name).maxGeomintThe most geoms a frame draws (
max_geom).visualOptionsIReadOnlyDictionary<int, bool>MuJoCo visual flags to switch (
visual_options).viewerFactoryEnvViewerFactoryGym.NET's
viewer_factory(DR-20) forhumanmode.
Remarks
The mode is required and comes first, so new PusherEnvV5() keeps meaning the other constructor, without a
mode; leave the mode out there for no rendering. Every other parameter means what it means there.
Exceptions
- ArgumentOutOfRangeException
Thrown for a value RenderMode does not declare (a cast integer such as
(RenderMode)99).- NotSupportedException
xmlFileis not a bundled model.- OverflowError
A frame skip of 0 (Python's
OverflowErrorderiving the frame rate).- ValueError
Both a camera id and a camera name.
PusherEnvV5(string, int, IReadOnlyDictionary<string, object?>?, double, double, double, string?, int, int, int?, string?, int, IReadOnlyDictionary<int, bool>?, EnvViewerFactory?)
Creates the env (Gymnasium's PusherEnv.init): the model, its spaces and the frame rate the time step gives.
public PusherEnvV5(string xmlFile = "pusher_v5.xml", int frameSkip = 5, IReadOnlyDictionary<string, object?>? defaultCameraConfig = null, double rewardNearWeight = 0.5, double rewardDistWeight = 1, double rewardControlWeight = 0.1, string? renderMode = null, int width = 480, int height = 480, int? cameraId = null, string? cameraName = null, int maxGeom = 1000, IReadOnlyDictionary<int, bool>? visualOptions = null, EnvViewerFactory? viewerFactory = null)
Parameters
xmlFilestringThe model: a bundled file name (
xml_file, DR-63);"pusher.xml"is v4's.frameSkipintPhysics steps per env step (
frame_skip).defaultCameraConfigIReadOnlyDictionary<string, object>The free camera's settings (
default_camera_config); null is Gymnasium's default (trackbodyid-1, 4 m away), and an empty dictionary leaves MuJoCo's camera as it is.rewardNearWeightdoubleThe weight of the fingertip's distance to the cylinder (
reward_near_weight).rewardDistWeightdoubleThe weight of the cylinder's distance to the goal (
reward_dist_weight; Gymnasium's default is the int 1).rewardControlWeightdoubleThe weight of the squared torques (
reward_control_weight).renderModestringGymnasium's
render_mode:human,rgb_array,depth_array,rgbd_tupleor null.widthintThe rendered frame's width (
width).heightintThe rendered frame's height (
height).cameraIdint?The camera to render from (
camera_id).cameraNamestringThe camera to render from by name (
camera_name).maxGeomintThe most geoms a frame draws (
max_geom).visualOptionsIReadOnlyDictionary<int, bool>MuJoCo visual flags to switch (
visual_options).viewerFactoryEnvViewerFactoryGym.NET's
viewer_factory(DR-20) forhumanmode.
Exceptions
- NotSupportedException
xmlFileis not a bundled model.- OverflowError
A frame skip of 0 (Python's
OverflowErrorderiving the frame rate).- ValueError
Both a camera id and a camera name.
Properties
ClassMetadata
Gymnasium's class attribute metadata: the four render modes, the frame rate set per instance. A new dict on every read.
public static Metadata ClassMetadata { get; }
Property Value
CylinderPos
The cylinder's start the last reset drew (cylinder_pos), float64 (2,), or null
before the first reset. Owned by the env, replaced by each reset: do not dispose it.
[NDBorrowed]
public NDArray? CylinderPos { get; }
Property Value
GoalPos
The goal the last reset set (goal_pos), the int64 [0, 0], or null before the first
reset. Owned by the env, replaced by each reset: do not dispose it.
[NDBorrowed]
public NDArray? GoalPos { get; }
Property Value
RewardControlWeight
The weight of the squared torques (_reward_control_weight), read every step.
[PyName("_reward_control_weight")]
public double RewardControlWeight { get; set; }
Property Value
RewardDistWeight
The weight of the cylinder's distance to the goal (_reward_dist_weight), read every step.
[PyName("_reward_dist_weight")]
public double RewardDistWeight { get; set; }
Property Value
RewardNearWeight
The weight of the fingertip's distance to the cylinder (_reward_near_weight), read every step.
[PyName("_reward_near_weight")]
public double RewardNearWeight { get; set; }
Property Value
Methods
Dispose(bool)
Closes, then releases the cylinder's and the goal's last arrays and everything MujocoEnv holds.
protected override void Dispose(bool disposing)
Parameters
ResetModel()
The episode's start (reset_model): the arm at init_qpos; the cylinder drawn as
[uniform(-0.3, 0), uniform(-0.2, 0.2)] until it is farther than 0.17 from the goal at the origin; both
written into init_qpos's last four values (see the class remarks); velocities
init_qvel + uniform(-0.005, 0.005, nv) with the cylinder's and the goal's zeroed.
protected override NDArray ResetModel()
Returns
- NDArray
The first observation.
Step(NDArray)
Applies action for FrameSkip physics steps and reports the transition (step).
public override StepResult<NDArray> Step(NDArray action)
Parameters
actionNDArrayThe 7 torques, float32 (the action space's samples) or any real dtype.
Returns
- StepResult<NDArray>
The observation, the float64 reward,
terminated=False,truncated=Falseand the info.
Exceptions
- ArgumentNullException
actionis null.- ValueError
The action is not of shape
(7,), with Gymnasium's message.- NotSupportedException
An action of a dtype whose cost NumPy keeps in that dtype (float16, complex).
- KeyError
The model lacks one of the bodies
object,tips_arm,goal(MuJoCo's message).