metadata

metadata

Configurations for metadata types.

dtype across all configs refers to zarr v3 storage data type. zarr v3 allowed dtypes are defined here https://zarr-specs.readthedocs.io/en/latest/v3/data-types/index.html

BaseObservationMetadataConfig dataclass

BaseObservationMetadataConfig(_target_=MISSING, raw_data_column_keys=MISSING, dimension=MISSING, dtype=MISSING, needs_normalization=MISSING, slice_start=None, slice_end=None)

Fields shared by every observation metadata config.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
raw_data_column_keys	list[str]	List of keys in the raw dataset corresponding to the observation.
dimension	int	Dimension that will be used to store the observation in the zarr store.
dtype	str	Numpy dtype the values are stored as.
needs_normalization	bool	Whether the observation is normalized by the fitted normalizer.
slice_start	int | None	Optional starting index for slicing a larger stored observation vector.
slice_end	int | None	Optional ending index (exclusive) for slicing a larger stored observation vector.

ObservationMetadataConfig dataclass

ObservationMetadataConfig(_target_='versatil.data.metadata.ObservationMetadata', raw_data_column_keys=MISSING, dimension=MISSING, dtype=MISSING, needs_normalization=MISSING, slice_start=None, slice_end=None, is_numerical=MISSING)

Bases: BaseObservationMetadataConfig

Config for ObservationMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
is_numerical	bool	Whether the observation is numerical rather than text. The specialized observation configs omit this field because their runtime classes are numerical by definition.

PositionObservationMetadataConfig dataclass

PositionObservationMetadataConfig(_target_='versatil.data.metadata.PositionObservationMetadata', raw_data_column_keys=MISSING, dimension=MISSING, dtype=MISSING, needs_normalization=MISSING, slice_start=None, slice_end=None, frame=value)

Bases: BaseObservationMetadataConfig

Config for PositionObservationMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
frame	str	Coordinate frame of the position observation.

OrientationObservationMetadataConfig dataclass

OrientationObservationMetadataConfig(_target_='versatil.data.metadata.OrientationObservationMetadata', raw_data_column_keys=MISSING, dimension=MISSING, dtype=MISSING, needs_normalization=MISSING, slice_start=None, slice_end=None, frame=value, orientation_representation=value)

Bases: BaseObservationMetadataConfig

Config for OrientationObservationMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
frame	str	Coordinate frame of the orientation observation.
orientation_representation	str	Representation of the orientation.

GripperObservationMetadataConfig dataclass

GripperObservationMetadataConfig(_target_='versatil.data.metadata.GripperObservationMetadata', raw_data_column_keys=MISSING, dimension=MISSING, dtype=MISSING, needs_normalization=MISSING, slice_start=None, slice_end=None, gripper_type=value, binary_gripper_range=value)

Bases: BaseObservationMetadataConfig

Config for GripperObservationMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
gripper_type	str	Type of gripper ('binary' or 'continuous').
binary_gripper_range	str	Range for binary gripper ('zero_one' or 'minus_one_one').

CameraMetadataConfig dataclass

CameraMetadataConfig(_target_='versatil.data.metadata.CameraMetadata', camera_key=MISSING, dtype=MISSING, channels=MISSING, image_width=None, image_height=None, max_pixel_value=None)

Config for CameraMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
camera_key	str	Camera identifier within the dataset.
dtype	str	Numpy dtype the images are stored as.
channels	int	Number of image channels.
image_width	int | None	Target image width.
image_height	int | None	Target image height.
max_pixel_value	float | None	Optional value used to scale image tensors after resizing and channel reordering.

RGBCameraMetadataConfig dataclass

RGBCameraMetadataConfig(_target_='versatil.data.metadata.RGBCameraMetadata', camera_key=MISSING, dtype=MISSING, image_width=None, image_height=None, max_pixel_value=255.0)

Config for RGBCameraMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
camera_key	str	Key in the raw dataset corresponding to the RGB camera.
dtype	str	Zarr storage dtype for RGB values.
image_width	int | None	Target image width.
image_height	int | None	Target image height.
max_pixel_value	float | None	Value used to scale RGB image tensors after resizing and channel reordering.

DepthCameraMetadataConfig dataclass

DepthCameraMetadataConfig(_target_='versatil.data.metadata.DepthCameraMetadata', camera_key=MISSING, dtype=MISSING, image_width=None, image_height=None, max_pixel_value=None)

Config for DepthCameraMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
camera_key	str	Key in the raw dataset corresponding to the depth camera.
dtype	str	Zarr storage dtype for depth values.
image_width	int | None	Target image width.
image_height	int | None	Target image height.
max_pixel_value	float | None	Optional value used to scale depth image tensors after resizing and channel reordering.

ActionMetadataConfig dataclass

ActionMetadataConfig(_target_='versatil.data.metadata.ActionMetadata', prediction_dimension=MISSING, is_numerical=MISSING, needs_normalization=MISSING, dtype=MISSING, is_precomputed=MISSING)

Config for ActionMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
prediction_dimension	int	Dimension for model prediction. May differ from storage, e.g., class labels stored as 1 column but predicted as n_classes logits.
is_numerical	bool	Whether the action is numerical rather than text.
needs_normalization	bool	Whether the action is normalized by the fitted normalizer.
dtype	str	Numpy dtype the values are stored as.
is_precomputed	bool	Whether the action is stored in the dataset instead of computed on the fly.

OnTheFlyActionMetadataConfig dataclass

OnTheFlyActionMetadataConfig(_target_='versatil.data.metadata.OnTheFlyActionMetadata', source_metadata=MISSING, computation_method=MISSING)

Config for OnTheFlyActionMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
source_metadata	Any	Metadata of the source observation used to compute the action.
computation_method	str	Method to compute the action, default 'delta' for subtraction between consecutive observations.

BasePrecomputedActionMetadataConfig dataclass

BasePrecomputedActionMetadataConfig(_target_=MISSING, raw_data_column_keys=MISSING, storage_dimension=MISSING, prediction_dimension=MISSING, needs_normalization=MISSING, dtype=MISSING)

Fields shared by every precomputed action metadata config.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
raw_data_column_keys	list[str]	List of keys in the raw dataset corresponding to the action.
storage_dimension	int	Dimension that will be used to store the action in the zarr store.
prediction_dimension	int	Dimension for model prediction. May differ from storage, e.g., class labels stored as 1 column but predicted as n_classes logits.
needs_normalization	bool	Whether the action is normalized by the fitted normalizer.
dtype	str	Numpy dtype the values are stored as.

PrecomputedActionMetadataConfig dataclass

PrecomputedActionMetadataConfig(_target_='versatil.data.metadata.PrecomputedActionMetadata', raw_data_column_keys=MISSING, storage_dimension=MISSING, prediction_dimension=MISSING, needs_normalization=MISSING, dtype=MISSING, is_numerical=MISSING)

Bases: BasePrecomputedActionMetadataConfig

Config for PrecomputedActionMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
is_numerical	bool	Whether the action is numerical rather than text. The specialized action configs omit this field because their runtime classes are numerical by definition.

PositionActionMetadataConfig dataclass

PositionActionMetadataConfig(_target_='versatil.data.metadata.PositionActionMetadata', raw_data_column_keys=MISSING, storage_dimension=MISSING, prediction_dimension=MISSING, needs_normalization=MISSING, dtype=MISSING, frame=MISSING, computation_method=None)

Bases: BasePrecomputedActionMetadataConfig

Config for PositionActionMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
frame	str	Coordinate frame of the position, camera or robot base.
computation_method	str | None	Whether actions are deltas or next-timestep poses.

OrientationActionMetadataConfig dataclass

OrientationActionMetadataConfig(_target_='versatil.data.metadata.OrientationActionMetadata', raw_data_column_keys=MISSING, storage_dimension=MISSING, prediction_dimension=MISSING, needs_normalization=MISSING, dtype=MISSING, frame=MISSING, orientation_representation=MISSING)

Bases: BasePrecomputedActionMetadataConfig

Config for OrientationActionMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
frame	str	Coordinate frame of the orientation, camera or robot base.
orientation_representation	str	Representation of the orientation.

GripperActionMetadataConfig dataclass

GripperActionMetadataConfig(_target_='versatil.data.metadata.GripperActionMetadata', raw_data_column_keys=MISSING, storage_dimension=MISSING, prediction_dimension=MISSING, needs_normalization=MISSING, dtype=MISSING, gripper_type=MISSING, binary_gripper_range=MISSING)

Bases: BasePrecomputedActionMetadataConfig

Config for GripperActionMetadata.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
gripper_type	str	Type of gripper action ('binary' or 'continuous').
binary_gripper_range	str	Range for binary gripper action ('zero_one' or 'minus_one_one').