Source code for hermespy.core.transformation
# -*- coding: utf-8 -*-
from __future__ import annotations
from abc import ABCMeta, abstractmethod
from functools import cached_property
from typing import overload, Set, Type
import numpy as np
from numba import jit
from ruamel.yaml import SafeConstructor, SafeRepresenter, MappingNode, Node
from .factory import Serializable
__author__ = "Jan Adler"
__copyright__ = "Copyright 2024, Barkhausen Institut gGmbH"
__credits__ = ["Jan Adler"]
__license__ = "AGPLv3"
__version__ = "1.4.0"
__maintainer__ = "Jan Adler"
__email__ = "jan.adler@barkhauseninstitut.org"
__status__ = "Prototype"
[docs]
class Direction(np.ndarray):
"""A cartesian unit norm vector pointing towards a direction."""
@staticmethod
@jit(nopython=True)
def __from_spherical(angles: np.ndarray) -> np.ndarray: # pragma: no cover
"""Generate a cartesian unit vector from spherical coordinates.
Args:
angles (numpy.ndarray):
Azimuth and zenith in radians.
Returns: The unit vector.
"""
cos = np.cos(angles)
sin = np.sin(angles)
unit_vector = np.array([sin[1] * cos[0], sin[1] * sin[0], cos[1]], dtype=np.float64)
return unit_vector
[docs]
@classmethod
def From_Spherical(cls: Type[Direction], azimuth: float, zenith: float) -> Direction:
"""Initialize a direction from spherical parameters.
Args:
azimuth (float):
Azimuth angle in radians.
zenith (float):
Zenith angle in radians.
Returns: The initialized direction.
"""
direction = cls.__from_spherical(np.array([azimuth, zenith], dtype=float)).view(cls)
return direction
@jit(nopython=True)
def __to_spherical(unit_vector: np.ndarray) -> np.ndarray: # pragma: no cover
"""Transform a unit vector to spherical coordinates.
Args:
unit_vector (numpy.ndarray):
Cartesian numpy vector.
Returns: An array representing azimuth and zenith angles in radians.
"""
# Equation 7.1-8 of ETSI TR 138901 v17
azimuth = np.arctan2(unit_vector[1], unit_vector[0])
# azimuth = np.angle(unit_vector[0] + 1j * unit_vector[1])
# Equation 7.1-7 of ETSI TR 138901 v17
# Only valid if the direction has been normalized!!!!
zenith = np.arccos(unit_vector[2])
return np.array([azimuth, zenith], dtype=np.float64)
[docs]
def to_spherical(self) -> np.ndarray:
"""Represent the direction as spherical coordinates.
Returns: An array representing azimuth and zenith angles in radians.
"""
return self.__to_spherical().view(np.ndarray)
[docs]
@classmethod
def From_Cartesian(
cls: Type[Direction], vector: np.ndarray, normalize: bool = False
) -> Direction:
"""Initialize a direction from a cartesian vector.
Raises:
ValueError: If `vector` does not represent a valid cartesian vector.
Returns: The initialized direction.
"""
vector = vector.flatten().astype(np.float64)
ndmin = vector.size
if ndmin > 3:
raise ValueError("Vector is not a valid cartesian vector")
if normalize:
norm = np.linalg.norm(vector)
if norm == 0:
raise ValueError("Zero-vectors cannot be normalized")
vector /= norm
unit_vector = np.zeros(3, dtype=np.float64)
unit_vector[:ndmin] = vector
return unit_vector.view(Direction)
[docs]
class Transformation(np.ndarray, Serializable):
"""Coordinate system transformation."""
yaml_tag = "Transformation"
@property
def translation(self) -> np.ndarray:
return self[:3, 3].view(np.ndarray)
@translation.setter
def translation(self, value: np.ndarray) -> None:
value = value.flatten()
if len(value) != 3:
raise ValueError("Translations must be three-dimensional cartesian vectors")
self[:3, 3] = value
@property
def rotation_rpy(self) -> np.ndarray:
"""Orientation in Roll, Pitch and Yaw Angles.
Returns: Roll, Pitch and Yaw in Radians.
"""
c = (self[0, 0] ** 2 + self[1, 0] ** 2) ** 0.5
if c != 0:
rpy = np.arctan2(
[self[2, 1] / c, -self[2, 0], self[1, 0] / c], [self[2, 2] / c, c, self[0, 0] / c]
)
else:
rpy = np.array([np.arctan2(self[0, 1], self[1, 1]), 0.5 * np.pi, 0], dtype=np.float64)
return rpy.view(np.ndarray)
@rotation_rpy.setter
def rotation_rpy(self, value: np.ndarray) -> None:
self[:3, :3] = self._rotation_from_rpy(value)
@staticmethod
@jit(nopython=True)
def _rotation_from_quaternion(q: np.ndarray, normalize: bool) -> np.ndarray: # pragma: no cover
"""Calculate a rotation matrix from a quaternion.
Args:
q (numpy.ndarray):
Quaternion in w, x, y, z representation.
normalize (bool):
Normalize the quaternion before computing the rotation matrix.
Returns: A :math:`3 \\times 3` numpy matrix representing the rotation.
"""
q = q / np.linalg.norm(q) if normalize else q
return np.array(
[
[
1 - 2 * (q[2] ** 2 + q[3] ** 2),
2 * (q[1] * q[2] - q[0] * q[3]),
2 * (q[0] * q[2] + q[1] * q[3]),
],
[
2 * (q[1] * q[2] + q[0] * q[3]),
1 - 2 * (q[1] ** 2 + q[3] ** 2),
2 * (q[2] * q[3] - q[0] * q[1]),
],
[
2 * (q[1] * q[3] - q[0] * q[2]),
2 * (q[0] * q[1] + q[2] * q[3]),
1 - 2 * (q[1] ** 2 + q[2] ** 2),
],
],
dtype=np.float64,
)
@property
def rotation_quaternion(self) -> np.ndarray: # pragma: no cover
"""Orientation in Quaternion representation.
A numpy vector representing w, x, y, z.
"""
tr = np.trace(self[:3, :3])
if tr > 0:
S = 2 * (tr + 1.0) ** 0.5
return np.array(
[
0.25 / S,
(self[2, 1] - self[1, 2]) / S,
(self[0, 2] - self[2, 0]) / S,
(self[1, 0] - self[0, 1]) / S,
],
dtype=np.float64,
)
if self[0, 0] > self[1, 1] and self[0, 0] > self[2, 2]:
S = 2 * (1 + self[0, 0] - self[1, 1] - self[2, 2]) ** 0.5
return np.array(
[
(self[2, 1] - self[1, 2]) / S,
0.25 / S,
(self[0, 1] + self[1, 0]) / S,
(self[0, 2] + self[2, 0]) / S,
],
dtype=np.float64,
)
if self[1, 1] > self[2, 2]:
S = 2 * (1 + self[1, 1] - self[0, 0] - self[2, 2]) ** 0.5
return np.array(
[
(self[0, 2] - self[2, 0]) / S,
(self[0, 1] + self[1, 0]) / S,
0.25 / S,
(self[1, 2] + self[2, 1]) / S,
],
dtype=np.float64,
)
S = 2 * (1 + self[2, 2] - self[0, 0] - self[1, 1]) ** 0.5
return np.array(
[
(self[1, 0] - self[0, 1]) / S,
(self[0, 2] + self[2, 0]) / S,
(self[1, 2] + self[2, 1]) / S,
0.25 / S,
],
dtype=np.float64,
)
@rotation_quaternion.setter
def rotation_quaternion(self, value: np.ndarray) -> None:
self[:3, :3] = self._rotation_from_quaternion(value, False)
[docs]
@classmethod
def No(cls: Type[Transformation]) -> Transformation:
return np.eye(4, 4, dtype=float).view(cls)
@staticmethod
@jit(nopython=True)
def _rotation_from_rpy(rpy: np.ndarray) -> np.ndarray: # pragma: no cover
"""Calculate a rotation matrix from roll pitch yaw angles.
Args:
rpy (numpy.ndarray):
Numpy vector of length 3 representing roll pitch and yaw in radians.
Returns:
A :math:`3 \\times 3` numpy matrix representing the rotation.
"""
# Compute rotational transformation portion
cos = np.cos(rpy)
sin = np.sin(rpy)
rotation = np.array(
[
[
cos[2] * cos[1],
cos[2] * sin[1] * sin[0] - sin[2] * cos[0],
cos[2] * sin[1] * cos[0] + sin[2] * sin[0],
],
[
sin[2] * cos[1],
sin[2] * sin[1] * sin[0] + cos[2] * cos[0],
sin[2] * sin[1] * cos[0] - cos[2] * sin[0],
],
[-sin[1], cos[1] * sin[0], cos[1] * cos[0]],
],
dtype=np.float64,
)
return rotation
[docs]
@classmethod
def From_Quaternion(
cls: Type[Transformation],
quaternion: np.ndarray,
pos: np.ndarray,
normalize_quaternion: bool = True,
) -> Transformation:
"""Initialize a transformation from a quaternion.
Args:
quaternion (numpy.ndarray):
Quaternion in w, x, y, z representation.
pos (numpy.ndarray):
Cartesian position in m.
normalize (bool, optional):
Normalize the quaternion before computing the rotation matrix.
Enabled by default.
Returns: The initialized transformation.
"""
# Generate empty transformation matrix
transformation = np.empty((4, 4), dtype=float)
# Compute rotational transformation portion
transformation[:3, :3] = Transformation._rotation_from_quaternion(
quaternion, normalize_quaternion
)
# Copy translational transformation portion
transformation[0:3, 3] = pos
# Fill in the static portion
transformation[3, :] = (0, 0, 0, 1)
# Return by view
return transformation.view(cls)
[docs]
@classmethod
def From_RPY(cls: Type[Transformation], rpy: np.ndarray, pos: np.ndarray) -> Transformation:
"""Initialize a transformation from roll pitch yaw angles.
Args:
rpy (numpy.ndarray):
Roll, pitch and yaw angles in radians.
pos (numpy.ndarray):
Cartesian position in m.
Returns: The initialized transformation.
"""
# Generate empty transformation matrix
transformation = np.empty((4, 4), dtype=float)
# Compute rotational transformation portion
transformation[:3, :3] = cls._rotation_from_rpy(rpy)
# Copy translational transformation portion
transformation[0:3, 3] = pos
# Fill in the static portion
transformation[3, :] = (0, 0, 0, 1)
# Return by view
return transformation.view(cls)
[docs]
@classmethod
def From_Translation(cls: Type[Transformation], translation: np.ndarray) -> Transformation:
translation = translation.flatten()
ndim = len(translation)
if ndim > 3:
raise ValueError("Translation must be a valid three-dimensional cartesian vector")
# Generate non-rotating transformation matrix
transformation = np.eye(4, 4, dtype=float)
# Copy translational transformation portion
transformation[0:ndim, 3] = translation
# Return by view
return transformation.view(cls)
[docs]
def transform_position(self, position: np.ndarray) -> np.ndarray:
"""Transform a cartesian position.
Args:
position (numpy.ndarray):
Numpy array representing the cartesian position.
Returns: The transformed position.
Raises:
ValueError: If `position` is invalid.
"""
position = position.flatten()
ndmin = len(position)
if ndmin > 3:
raise ValueError("Position must be a valid cartesian numpy vector")
factor = np.array([0.0, 0.0, 0.0, 1.0], dtype=float)
factor[:ndmin] = position
transformed_position = self.view(np.ndarray) @ factor
return transformed_position[:ndmin]
[docs]
def rotate_direction(self, direction: np.ndarray) -> Direction:
"""Rotate a cartesian direction.
Args:
direction (numpy.ndarray):
A directional vector.
Returns: The transformed direction.
Raises:
ValueError: If `direction` is invalid.
"""
direction = direction.flatten()
ndim = len(direction)
if ndim > 3:
raise ValueError("Direction must be a valid cartesian numpy vector")
rotated_direction = np.tensordot(self[:ndim, :ndim], direction, axes=(1, 0))
return Direction.From_Cartesian(rotated_direction)
[docs]
def transform_direction(self, direction: np.ndarray, normalize: bool = False) -> Direction:
"""Transform a direction.
Args:
direction (numpy.ndarray):
Direction to be transformed.
normalize (bool, optional):
Normalize the resulting transformed direction to a unit norm vector.
Disabled by default.
Returns: The transformed direction.
"""
return Direction.From_Cartesian(self.rotate_direction(direction), normalize=normalize)
[docs]
def invert(self) -> Transformation:
"""Invert the transformation.
Returns: The inverted transformation.
"""
return np.linalg.inv(self).view(Transformation)
[docs]
def lookat(
self,
target: np.ndarray = np.array([0.0, 0.0, 0.0], float),
up: np.ndarray = np.array([0.0, 1.0, 0.0], float),
) -> Transformation:
"""Rotate and loook at the given coordinates. Modifies `orientation` property.
Args:
target (numpy.ndarray):
Cartesean coordinates to look at.
Defaults to np.array([0., 0., 0.], float)
up (numpy.ndarray):
Global catesean sky vector.
Defines the upward direction of the local viewport.
Defaults to np.array([0., 1., 0.], float)
Returns:
self (Transformation): This modified Transformation.
"""
# Validate arguments
target_ = np.asarray(target)
if target_.shape != (3,):
raise ValueError(
f"Got target of an unexpected shape (expected (3,), got {target_.shape})"
)
up_ = np.asarray(up)
if up_.shape != (3,):
raise ValueError(f"Got up of an unexpected shape (expected (3,), got {up_.shape})")
up_ /= np.linalg.norm(up_)
# Calculate new orientation
# forward vector
pos = self.translation
f = target_ - pos
f_norm = np.linalg.norm(f)
f = f / f_norm if f_norm != 0.0 else pos # normalize
# side/right vector
s = np.cross(up_, f)
s_norm = np.linalg.norm(s)
s = s / s_norm if s_norm != 0.0 else up_ # normalize
# up vector
u = np.cross(f, s)
# Calcualte the new transformation matrix
self[:3, 0] = s
self[:3, 1] = u
self[:3, 2] = f
self[3, :] = [0.0, 0.0, 0.0, 1.0]
return self
@classmethod
def to_yaml(
cls: Type[Transformation], representer: SafeRepresenter, node: Transformation
) -> MappingNode:
state = {"translation": node.translation, "rotation": node.rotation_rpy}
return representer.represent_mapping(cls.yaml_tag, state)
@classmethod
def from_yaml(
cls: Type[Transformation], constructor: SafeConstructor, node: Node
) -> Transformation:
state = constructor.construct_mapping(node, deep=False)
return cls.From_RPY(state.get("rotation", None), state.get("translation", None))
[docs]
class TransformableLink(metaclass=ABCMeta):
"""Abstract base class of kinetmatic chain links."""
__linked_frames: Set[TransformableLink]
def __init__(self) -> None:
self.__linked_frames = set()
@property
@abstractmethod
def forwards_transformation(self) -> Transformation:
"""Transformation to convert local coordinates to global coordinates.
Updated every time the kinematic chain's parameters change.
Returns: The transformation.
"""
... # pragma no cover
@abstractmethod
def _kinematics_updated(self) -> None:
"""Callback notifying the link that its kinematics have been updated."""
... # pragma no cover
[docs]
@abstractmethod
def set_base(self, base: TransformableLink | None) -> None:
"""Set the relative base coordinate frame of this link.
Args:
base (TransformableLink | None):
The base to be coordinate frame to be set.
`None` to detach the link.
"""
... # pragma no cover
@property
def linked_frames(self) -> Set[TransformableLink]:
return self.__linked_frames
[docs]
def add_link(self, link: Transformable) -> None:
"""Establish a new link to a coordinate frame depending on this frame.
Args:
link (Transformable):
The transformable frame to be registered.
"""
self.__linked_frames.add(link)
link.set_base(self)
[docs]
def remove_link(self, link: Transformable, force_removal: bool = True) -> None:
"""Remove an established link to this coordinate frame.
Args:
link (Transformable):
The coordinate frame to be linked to this frame.
force_removal(bool, optional):
Raise a RuntimeError if the link is not registered
Raises:
RuntimeError:
If the `link` is not currently registerd and `force_removal` is enabled.
"""
if link not in self.__linked_frames:
if force_removal:
raise RuntimeError("Transformable link not registered")
else:
return
self.__linked_frames.remove(link) # Remove the link from the list of links
link.set_base(None) # Notify the link that it's now detached
[docs]
class TransformableBase(TransformableLink):
"""Base of kinematic chains."""
@property # @cached_property
def forwards_transformation(self) -> Transformation:
return Transformation.No()
def _kinematics_updated(self) -> None:
raise RuntimeError("Called base updated routine of a base, this should not be possibel")
[docs]
def set_base(self, base: TransformableLink | None) -> None:
if base is not None:
raise RuntimeError("A base link may not be assigned another link as base")
[docs]
class Transformable(Serializable, TransformableLink):
"""Representation of a Coordinate Frame within a Kinematic Chain."""
property_blacklist = {"pose"}
__base: TransformableLink | None
__pose: Transformation
def __init__(self, pose: Transformation | None = None) -> None:
"""
Args:
pose (Transformation, optional):
Transformation of the transformable with respect to its reference frame.
By default, no transformation is considered, i.e. :meth:`Transformation.No`
"""
# Init base class
TransformableLink.__init__(self)
self.__base = None
self.pose = Transformation.No() if pose is None else pose
@property
def position(self) -> np.ndarray:
"""Position of the Transformable.
Cartesian offset to the reference coordinate frame.
Returns:
Cartesian position x, y, z in m.
Raises:
ValueError: If `position` is not a valid three-dimensional vector.
"""
return self.__pose.translation
@position.setter
def position(self, value: np.ndarray) -> None:
if value.ndim != 1:
raise ValueError("Position must be a vector")
if len(value) > 3:
raise ValueError("Position may have at most three dimensions")
# Make vector 3D if less dimensions have been supplied
if len(value) < 3:
value = np.append(value, np.zeros(3 - len(value), dtype=float))
self.__pose.translation = value
self._kinematics_updated()
@property
def orientation(self) -> np.ndarray:
"""Orientation of the Transformable.
Returns: The transformation in radians for roll, pitch and yaw.
Raises:
ValueError: If `orientation` is not a three-dimensional numpy vector.
"""
return self.__pose.rotation_rpy
@orientation.setter
def orientation(self, value: np.ndarray) -> None:
if value.ndim != 1 or len(value) != 3:
raise ValueError("Orientation must be a three-dimensional vector")
self.__pose.rotation_rpy = value
self._kinematics_updated()
[docs]
def set_base(self, base: TransformableLink | None) -> None:
if self.__base == base:
return
if self.__base is not None:
self.__base.remove_link(self, force_removal=False)
self.__base = base
self._kinematics_updated()
if self.__base is not None:
self.__base.add_link(self)
@property
def global_position(self) -> np.ndarray:
"""Position of the represented object within the global coordinate system.
Returns: Three-dimensional numpy vector representing the cartesian object coordinates.
"""
return self.forwards_transformation.translation
@property
def global_orientation(self) -> np.ndarray:
"""Orientation of the represented object within the global coordinate system.
Returns: Three-dimensional numpy vector representing roll, pitch and yaw in radians.
"""
return self.forwards_transformation.rotation_rpy
@property
def is_base(self) -> bool:
"""Is this transformable acting as a base frame?
Returns: Boolean indicator.
"""
return self.__base is None
@property
def pose(self) -> Transformation:
"""Pose of the Transformable with respect to its reference link.
Returns: The pose's transformation.
"""
return self.__pose
@pose.setter
def pose(self, value: Transformation) -> None:
self.__pose = value.copy()
self._kinematics_updated()
@cached_property
def forwards_transformation(self) -> Transformation:
if self.is_base:
return self.pose
transformation = self.__base.forwards_transformation @ self.pose
return transformation.view(Transformation)
@cached_property
def backwards_transformation(self) -> Transformation:
"""Transformation to convert global coordinates to local coordinates.
Updated every time the kinematic chain's parameters change.
Returns: The transformation.
"""
return self.forwards_transformation.invert()
@overload
def to_local_coordinates(
self, global_object: Transformable
) -> Transformation: ... # pragma no cover
@overload
def to_local_coordinates(
self, global_object: Transformation
) -> Transformation: ... # pragma no cover
@overload
def to_local_coordinates(
self, position: np.ndarray, orientation: np.ndarray | None = None
) -> Transformation: ... # pragma no cover
[docs]
def to_local_coordinates(self, arg_0: Transformable | Transformation | np.ndarray, arg_1: np.ndarray | None = None) -> Transformation: # type: ignore
if isinstance(arg_0, Transformable):
arg_0 = arg_0.forwards_transformation
elif isinstance(arg_0, Transformation):
... # pragma no cover
elif isinstance(arg_0, np.ndarray):
arg_1 = np.zeros(3, dtype=float) if arg_1 is None else arg_1
arg_0 = Transformation.From_RPY(arg_1, arg_0)
else:
raise ValueError("Unknown type of first argument")
local_transformation = self.backwards_transformation @ arg_0
return local_transformation.view(Transformation)
[docs]
def lookat(
self,
target: np.ndarray = np.array([0.0, 0.0, 0.0], float),
up: np.ndarray = np.array([0.0, 1.0, 0.0], float),
) -> None:
"""Rotate and loook at the given coordinates. Modifies `orientation` property.
Args:
target (numpy.ndarray):
Cartesean coordinates to look at.
Defaults to np.ndarray([0., 0., 0.], float).
up (array of 3 numbers):
Global catesean sky vector.
Defines the upward direction of the local viewport.
Defaults to np.ndarray([0., 1., 0.], float).
"""
self.pose.lookat(target, up)
def _kinematics_updated(self) -> None:
# Clear the cached forwards transformation if the base has been updated
if "forwards_transformation" in self.__dict__:
del self.forwards_transformation
if "backwards_transformation" in self.__dict__:
del self.backwards_transformation
for frame in self.linked_frames:
frame._kinematics_updated()