# Rotations and translations with Eigen and scipy.spatial

November 12, 2023

In this blog post I am going to demonstrate how homogeneus transformations are handled in C++ with Eigen and in Python with scipy.spatial. Let’s start with the latter.

The Scipy module scipy.spatial.transform provides a versatile collection of types to represent rotations in different forms. A widely used way of representing a rotation in 3D is roll-pitch-yaw (RPY) angles. The convention of what RPY means may differ, depending on the discipline (like robotics or aerospace engineering), but one way of defining RPY is the following:

1. Rotate the coordinate frame around its X axis (yaw)
2. The, rotate the frame around its Y axis (pitch)
3. Finally, rotate the frame around its Z axis (roll)

This representaion is natural when we model e.g. a robot gripper, with the Z axis “sticking out” of it.

If you thing in terms of multiplication of rotation matrices, the sequence above can be described as such:

$$R_{x}(yaw) R_{y}(pitch) R_{z}(roll)$$

Let’s model this sequence of rotations with NumPy + SciPy. First, import the libraries:

import numpy as np
import scipy.spatial.transform as st

np.set_printoptions(precision=3)


We will play around with actual values for the angles: $yaw = \pi/4$, $pitch = \pi/3$, and $roll = -\pi/8$.

angle_x = np.pi / 4
angle_y = np.pi / 3
angle_z = -np.pi / 8

input_angles = np.array([angle_x, angle_y, angle_z])


The sequence of X, then Y, then Z rotation are technically Euler angles. We construct a Rotation object using the from_euler initializer.

R = st.Rotation.from_euler('XYZ', input_angles)


Take a notice that we specified the sequence of axes using capital letters. SciPy interprets such formatting as intrinsic rotations, i.e. the angles are considered with respect to the frame being rotated.

Let’s now inspect the actual $SO(3)$ matrix corresponding to the rotation of interest:

print(R.as_matrix())

[[ 0.462  0.191  0.866]
[ 0.295  0.888 -0.354]
[-0.836  0.419  0.354]]


If a homogeneous trasnformation is of interest, when we combine rotation and translation, we can construct one directly using NumPy indexing:

def create_transform(rotation, translation):

transform = np.eye(4)
transform[:3, :3] = rotation.as_matrix()
transform[:3, 3] = translation

return transform


An example is a transformation with the rotation above and translation of $[0.5, 0.4, 0.3]$:

T = create_transform(R, np.array([0.5, 0.4, 0.3]))

print(T)

[[ 0.462  0.191  0.866  0.5  ]
[ 0.295  0.888 -0.354  0.4  ]
[-0.836  0.419  0.354  0.3  ]
[ 0.     0.     0.     1.   ]]


You may see that the original angles we used to construct the Rotation object match the Euler angles extracted from the obejct using as_euler method:

print(input_angles)
print(R.as_euler('XYZ'))

[ 0.785  1.047 -0.393]
[ 0.785  1.047 -0.393]


Let’s now play around with Eigen to create the same rotation, translation, and the homogeneous transformation.

One of the supported way of specifying rotations in Eigen is angle-axis, where you provide an arbitrary axis vector to rotate around, together with an angle of rotation. To model the XYZ sequence, we can construct three rotation objects and then multiply them. We will start with helper functions constructing Eigen::AngleAxisf objects:

#include <Eigen/Dense>
#include <Eigen/Geometry>

Eigen::AngleAxisf rotate_x(float angle)
{
return {angle, Eigen::Vector3f::UnitX()};
}

Eigen::AngleAxisf rotate_y(float angle)
{
return {angle, Eigen::Vector3f::UnitY()};
}

Eigen::AngleAxisf rotate_z(float angle)
{
return {angle, Eigen::Vector3f::UnitZ()};
}


For the actual example, we first construct a translation object:

Eigen::Translation<float, 3> t{0.5f, 0.4f, 0.3f};


To create the rotation object as in the Python example above, we use the previosly defined helper functions:

const float angle_x = M_PI / 4;
const float angle_y = M_PI / 3;
const float angle_z = -M_PI / 8;
const auto R = rotate_x(angle_x) * rotate_y(angle_y) * rotate_z(angle_z);


To get the transformation combining rotation and trasnlation, we perform multiplication corresponding to first translating a coordinate frame, and then rotating it:

const auto transform = t * R;


When we print the contents of the objecs above, we can see that the values are the same as in the Python example:

#include <iostream>

void print_array(const auto& arr, const char* prefix)
{
std::cout << prefix << ":\n" << arr << "\n\n";
}

print_array(t.vector(), "Translation");
print_array(R.matrix(), "Rotation");
print_array(transform.matrix(), "Transformation");


The output:

Translation:
0.5
0.4
0.3

Rotation:
0.46194  0.191342  0.866025
0.29516  0.887626 -0.353553
-0.836356  0.418937  0.353553

Transformation:
0.46194  0.191342  0.866025       0.5
0.29516  0.887626 -0.353553       0.4
-0.836356  0.418937  0.353553       0.3
0         0         0         1


See more Eigen examples in the Space Transformations tutorial.