Spatial descriptions and

transformations

Sami Haddadin

Robotics and Mechatronics Center

German Aerospace Center (DLR), Germany

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Transform equation

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Transform equation

Two ways to express {D} as products of transformations:

U U
TD = TA A TD (1)
U U B C
TD = TB TC TD (2)

We may set these to descriptions of U TD equal to form a

transform equation
U
TA A TD = U TB B TC C TD (3)

This can be used to solve for transforms in the case of n unknown

transforms and n transform equations.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Cayley’s formula

so far: rotation only by means of 3 × 3 matrix

all columns are mutually orthogonal and have unit magnitude
we will see: proper RMs have det = 1
we call them proper orthonormal matrices
nonpoper ones have det = −1
General question:
Can we describe orientation with fewer than nine numbers?

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Cayley’s formula

Result from linear algebra: for any orthonormal matrix R, there

exists a skew symmetric matrix, S, such that

R = (I3 − S)−1 (I3 + S) (4)

Now a skew symmetric matrix (i.e. S = −S T ) of dimension 3 is

specified by three paramters (sx , sy , sz ) as
⎡ ⎤
0 −sz sy
S = ⎣ sz 0 −sx ⎦ (5)
−sy sx 0
→ any 3 × 3 rotation matrix can be specified by three parameters.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Cayley’s formula
Clearly, the nine elements of a rotation matrix are not all
independent. Six dependencies between the elements are easily
obtained:  
R = X̂ Ŷ Ẑ (6)
Six constraints:
|X̂ | = 1
|Ŷ | = 1
|Ẑ | = 1
X̂ · Ŷ = 0 (7)
X̂ · Ẑ = 0
Ŷ · Ẑ = 0
Please note that rotations do not generally commute. That is that
AR B R is not the same as B RC A RB .
B C
Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik
Roll-pitch-yaw

One method of describing the orientation of a frame {B} is as

follows:
Start with the frame coincident with a known reference frame
{A}. First rotate {B} about X̂A by an angle γ, then rotate
about ŶA by an angle β, and then rotate about ẐA by an
angle α.
Each rotation takes place about an axis in the fixed reference
fame, {A}. This convention is called X-Y-Z fixed angles or roll,
pitch, yaw.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Roll-pitch-yaw

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Roll-pitch-yaw

The derivation of the equivalent rotation matrix A RB (γ, β, α) is

straight forward because all rotations occur about axes of the
reference frame.

A
RB (γ, β, α) = RZ (α)RY (β)RX (γ) =
⎡ ⎤⎡ ⎤⎡ ⎤
cα −sα 0 cβ 0 sβ 1 0 0
⎣ sα cα 0 ⎦ ⎣ 0 1 0 ⎦ ⎣ 0 cγ −sγ ⎦ (8)
0 0 1 −sβ 0 cβ 0 sγ cγ

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Roll-pitch-yaw

Thinking in terms of rotations as operators:

⎡ ⎤
cαcβ cαsβsγ − sαcγ cαsβcγ + sαsγ
A
RB (γ, β, α) = ⎣ sαcβ sαsβsγ + cαcγ sαsβcγ − cαsγ ⎦ (9)
−sβ cβsγ cβcγ

Keep in mind that the definition given here specifies the order of
the three rotations.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Inverse problem

Extracting equivalent X − Y − Z fixed angles from a rotation

matrix is often of interest. The solution depends on solving a set
of transcendental equations: there are nine equations and three
unknowns. Among the nine equations are six dependencies, so
essentially we have three equations and three unknowns.
⎡ ⎤
r11 r12 r13
A
RB (γ, β, α) = ⎣ r21 r22 r23 ⎦ (10)
r31 r32 r33

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Inverse problem


β = arctan 2(−r31 , r11
2 + r2 )
21
α = arctan 2(r21 /cβ, r11 /cβ) (11)
γ = arctan 2(r32 /cβ, r33 /cβ),

where arctan 2(y , x) is a two argument arc tangent function (uses

both signs of x and y ).
convention: use this solution for β for which −90o ≤ β ≤ 90o
if β = ±90o (so that cβ = 0), the solution of (??) degrades
in those cases, only the sum or the difference of α and γ may
be computed
possible convention: α = 0o

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Inverse problem

If β = 90o , then the solution may be calculated as

β = 90o
α = 00 (12)
γ = arctan 2(r12 , r22 ).

If β = −90o , then the solution may be calculated as

β = −90o
α = 00 (13)
γ = − arctan 2(r12 , r22 ).

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Z-Y-X Euler angles

Another method of describing the orientation of a frame {B} is as

follows:
Start with the frame coincident with a known reference frame
{A}. First rotate {B} about ẐB by an angle α, then rotate
about ŶB by an angle β, and then rotate about X̂B by an
angle γ.
Each rotation takes place about an axis in the moving reference
fame, {B}. Such a set of rotations is called Euler angles. Note
that each rotation takes place about an axis whose location
depends upon the preceding rotations. This particular set of
rotations is called Z-Y-X Euler angles.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Z-Y-X Euler angles

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Z-Y-X Euler angles

Using intermediate frames {B  } and {B  } we may immediately

write
A  
RB = A RB  B RB  B RB (14)
The final result will be exactly the same as for Roll-Pitch-Yaw. In
total there are 24 angle set conventions. Of these, 12
conventions are for fixed angle sets, and 12 for Euler angle sets.
Because of the duality of fixed and Euler angle sets, there are
really only 12 unique parameterizations of a rotation matrix using
successive rotations about principal axes.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Angle-axis

RX (30) is an example of an equivalent angle-axis representation.

If the axis is a general direction (rather than one of the unit
directions) any orientation may be obtained through proper axis
and angle selection. Consider the following description of frame
{B}:
Start with the frame coincident with a known frame {A}:
Then rotate {B} about the vector A K̂ by an angle θ
We will write A RB (K̂ , θ) or RK (θ).

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Angle-axis

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Angle-axis

If the axis of rotation is a general axis, it can be shown that the

equivalent rotation matrix is
⎡ ⎤
kx2 v θ + cθ kx ky v θ − kz sθ kx kz v θ + ky sθ
RK (θ) = ⎣ kx ky v θ + kz sθ ky2 v θ + cθ ky kz v θ − kx sθ ⎦ ,
kx kz v θ − ky sθ ky kz v θ + kx sθ kz2 v θ + cθ
(15)
where v θ = 1 − cos θ, and A K̂ = [kx ky kz ]T . The signs of θ is
determined by the right-hand rule with the thumb pointing along
the positive sense of A K̂ .

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Angle-axis: inverse problem

r11 + r22 + r33 − 1

θ = arccos (16)
2
⎡ ⎤
r32 − r23
1 ⎣
K̂ = r13 − r31 ⎦ (17)
2 sin θ
r21 − r12

0o ≤ θ ≤ 180o and for any angle axis pair, there exists another
pair with negative signs.
for small angular rotations, the axis becomes ill-defined.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Euler parameters

In terms of the equivalent axis K̂ = [kx ky kz ]T and the equivalent

angle θ, the Euler parameters are given by
θ
1 = kx sin (18)
2
θ
2 = ky sin (19)
2
θ
3 = kz sin (20)
2
θ
4 = cos (21)
2

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Euler parameters

It is then clear that these four quantities are not independent, but
that
21 + 22 + 23 + 24 = 1 (22)
must always hold. Hence, an orientation might be visualized as a
point on a unit hypersphere in four-dimensional space. Euler
parameters are also known as a unit quaternion.

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Euler parameters

The rotation matrix, R , which is equivalent to a set of Euler

parameters is given as
⎡ ⎤
1 − 222 − 223 2(1 2 − 3 4 ) 2(1 3 + 2 4 )
R = ⎣ 2(1 2 + 3 4 ) 1 − 221 − 223 2(2 3 − 1 4 ) ⎦ . (23)
2(1 3 − 2 4 ) 2(2 3 + 1 4 ) 1 − 221 − 222

Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik

Euler parameters
In terms of the equivalent axis K̂ = [kx ky kz ]T and the equivalent
angle θ, the Euler parameters are given by
r32 − r23
1 =
44
r13 − r31
2 =
44
r21 − r12
3 = (24)
44
1√
4 = 1 + r11 + r22 + r33
2
Note that (??) is not useful in a computational sense if the
rotation matrix represents a rotation of 180o about some axis,
since 4 goes to zero. However, it can be shown that in the limit
all the expressions in (??) remain finite even for this case.
Haddadin, Robotics: http://www6.in.tum.de/Main/TeachingWs2012Robotik