ETH Library

Nonlinear Quadrocopter Attitude

Control
Technical Report

Report

Author(s):
Brescianini, Dario; Hehn, Markus; D'Andrea, Raffaello

Publication date:
2013

Permanent link:
https://doi.org/10.3929/ethz-a-009970340

Rights / license:
In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection.
For more information, please consult the Terms of use.
 Nonlinear Quadrocopter Attitude Control

Technical Report

Dario Brescianini, Markus Hehn and Raffaello D’Andrea

{bdario,hehnm,rdandrea}@ethz.ch

Institute for Dynamic Systems and Control (IDSC)

ETH Zürich

October 10, 2013

Abstract
This report addresses the issue of nonlinear attitude control for quadrocopters. A globally
stable attitude controller based on unit quaternions is proposed and its stability and robustness
is proven. In addition, a method to prioritize the control of the crucial pitch and roll angles over
yaw is introduced. Finally, experiments with the proposed nonlinear controller are conducted
which demonstrate its performance.
1 Introduction
Quadrocopters are a popular research subject and a vast variety of literature on quadrocopter
attitude control exists. Often, the attitude is described using Euler angles, which is a very
natural way of describing orientation, especially for fixed-wing aircrafts. Controllers are then
derived from linearizations of the nonlinear attitude control problem around hover conditions.
This yields reasonable results for small pitch and roll angles. More recently, the research has
been directed towards high maneuverability. For aggressive maneuvers, i.e. extreme changes in
attitude, however, these locally stable controllers have significant deficiencies which yield poor
performance or even worse, failure to stabilize the vehicle.
In order to exploit the full potential of quadrocopters, globally stable attitude controllers should
be applied. Furthermore, Euler angles give rise to singularities which have to be avoided. This
limits the set of all possible motions for purely mathematical reasons which have nothing to
do with the physical constraints of the quadrocopter and thus Euler angles are not appropriate
for global attitude control. It is crucial for global attitude control that a parameterization is
chosen which is able to represent attitudes globally and is neither geometrically nor kinematically
singular. In [2], a survey of attitude descriptions with their suitability for rigid body attitude
control is given. Commonly, globally stable attitude controllers are based on unit quaternions
[14, 4, 13, 12], but there exist also some controllers representing the attitude using rotation
matrices [2, 9, 10].
This report is the result of a semester project carried out in fall 2011, with the objective of
developing and implementing a globally stable attitude controller for the quadrocopters used in
the Flying Machine Arena1 . It is now published as a technical report to serve as a reference for
the applied attitude controller in the Flying Machine Arena.
The remainder of this report is organized as follows: In Section 2, background material on
attitude representation and the quadrocopter setup is provided. A globally asymptotically stable
control law is then proposed in Section 3. Finally, experimental results for the proposed controller
are shown in Section 4 and the report is concluded in Section 5.

2 Preliminaries
2.1 Attitude Representation
The attitude of a rigid body can be described by a rotation between a reference coordinate frame
and a body-fixed coordinate frame. There exist several rotation parametrizations such as Euler
angles (3 parameters), unit quaternions (4 parameters) and rotation matrices (9 parameters).
All three-parameter representations suffer from singularities and all k-parameter representations
(k > 3) have k − 3 constraints [11].

2.1.1 Rotation Matrix

b R3 to the linear space
A rotation RBI is a linear transformation which maps the linear space I =
B=b R3 while preserving length and right-handedness. From these properties, it follows directly
that RBI ∈ SO(3). Consider the orthogonal bases ~ēI ∈ I and ~ēB ∈ B, where ~ēI is defined as


~ēI := ~eIx , ~eIy , ~eIz (1)
and a vector ~r expressed in the coordinate frames I and B, i.e. I ~r and B ~r, respectively. Then
the linear transformation RBI that maps a vector I ~r to B ~r
B~
r = RBI I ~r, (2)
1
http://www.flyingmachinearena.org

2
is given by a 3 × 3 matrix

RBI = eIx , B ~eIy , B ~eIz
B~ . (3)
The column vectors of RBI are the three orthogonal unit basis vectors of frame I represented in
frame B. Similarly, the three row vectors correspond to the basis vectors of frame B represented
in frame I.

2.1.2 Euler Angles

Euler angles are based on the fact that any rotation matrix can be constructed by consecutively
applying three elementary rotation matrices. The nine entries of the rotation matrix can then be
parametrized by the rotation angles of the three elementary rotations. A very common rotation
sequence for aircraft applications is the ZYX-sequence (Fig. 1):
1. Rotation about the initial z-axis by yaw angle ψ (I → K).
2. Rotation about the new (rotated) y-axis by pitch angle θ (K → L).
3. Rotation about the new (rotated) x-axis by roll angle φ (L → B).
RBI can be written as

RBI = RBL RLK RKI (4)

   
1 0 0 cθ 0 −sθ cψ sψ 0
= 0 cφ sφ   0 1 0  −sψ cψ 0 (5)
0 −sφ cφ sθ 0 cθ 0 0 1
 
cθ cψ cθ sψ −sθ
=  −cφ sψ + sφ sθ cψ cφ cψ + sφ sθ sψ sφ cθ  (6)
sφ sψ + cφ sθ cψ −sφ cψ + cφ sθ sψ cφ cθ
where c(·) and s(·) are abbreviations for cos(·) and sin(·) respectively. The inverse mapping from
rotation matrix to Euler angles is then

φ = arctan2 (R23 , R33 ) (7)

θ = − arcsin (R13 ) (8)
ψ = arctan2 (R12 , R11 ) (9)

with Rij the entry in the i-th row and j-th column of RBI .

~eB
z
~eL
z
~eIz =~eK
z

~eB
y
~eIx
~eL eB
x =~ x φ
θ

~eK ψ ~eK eL
x y =~ y

~eIy

Figure 1: ZYX-Euler angles.

3
2.1.3 Unit quaternion
Every rotation can be parametrized by a single rotation about a fixed axis described by the unit
vector ~k and a rotation angle α. This is known as eigenaxis rotation and embodies the shortest
rotation between two orientations. Based on this, a unit quaternion is defined as
   
 ⊤ q0 cos( α2 )
q = q0 q1 q2 q3 = = ~ . (10)
q1:3 k sin( α2 )

The adjoint, norm and inverse of a quaternion q are

 
q0
q̄ = , (11)
−q1:3
q
kqk = q02 + q12 + q22 + q32 , (12)
q̄
q−1 = . (13)
kqk

The multiplication of two quaternions q and p is then defined2 by

q · p := Q(q)p (14)

where  
q0 −q1 −q2 −q3
q1 q0 −q3 q2 
Q(q) := 
q2 q3
. (15)
q0 −q1 
q3 −q2 q1 q0
Note that the quaternion, which corresponds to a rotation matrix I is given by
 
1
0
qI =  
0 . (16)
0

The rotation of a vector ~r by a quaternion q is computed by

p(~rr ) = q · p(~r) · q̄ (17)

where ~rr is the rotated vector and p(·) is the quaternion representation of a vector:
 
0
p(~r) = . (18)
~r

The mapping from Euler angles to unit quaternions for the ZYX-sequence is
 
cφ cθ cψ + sφ sθ sψ
 2 2 2 2 2 2

−c φ s θ s ψ + c θ c ψ s φ 

q (ψ, θ, φ) =  2 2 2 2 2 2 . (19)

 c φ2 c ψ2 s 2θ + s φ2 c ψ2 s ψ2 
cφ cθ sψ − sφ cψ sθ
2 2 2 2 2 2

2
Note: For ease of interpretation, the quaternion multiplication used in this report differs from the one defined
in [3]. With the definition in (14), the quaternion multiplication q · p corresponds to first a rotation of q and
afterwards a rotation of p in the new coordinate system.

4
2.1.4 Remarks on attitude representation
Both, rotation matrices and unit quaternions, offer singularity-free representations of attitude.
Note however, that the space of unit quaternions S3 double covers the space of physical at-
titudes SO(3). Hence, unit quaternions are not unique. In fact, each pair of antipodal unit
quaternions ±q ∈ S3 corresponds to the same physical attitude [2]. For attitude control tasks
in S3 , this implies that a controller has to be designed which stabilizes a disconnected set of
equilibrium points. Nevertheless, unit quaternions are often the chosen parametrization in prac-
tise for attitude control (including this report) since they are a minimal globally non-singular
parametrization. Furthermore, unit quaternions offer an easy insight on what is happening on
a geometrical level.

2.2 Quadrocopter
In the following, we consider a quadrocopter with a body-fixed frame B and an inertial frame
I as shown in Fig. 2. The inputs to the quadrocopter are the four motor thrust forces. The

~eB
x
~eB
z

~eB
y
~eIz

~eIy
I
~eIx

Figure 2: Quadrocopter with the body-fixed coordinate frame B and a reference frame I.

outputs are the position, velocity and acceleration as well as the attitude and angular body rate.
As the model has more outputs than inputs, it is clear that not all outputs can be controlled
independently. In the FMA, a cascaded control structure is used to control the quadrocopters,
with a position controller as the outermost control loop (Fig. 3). This control approach is
Feedforward
~vff ~aff ~ ff
Ω
~
rcmd ~acmd ~ cmd
Ω
Desired Position Attitude Onboard F ~cmd
Quadrocopter
Trajectory Controller Controller Controller
ψcmd collcmd

~
r, ~v , ~a q ~
Ω

State feedback (position, velocity, acceleration, attitude, angular body rate)

Figure 3: Cascaded control architecture.

known as time-scale separation and is valid as long as the inner loops are significantly faster
than the outer loops. In this report, only the attitude control loop is taken into consideration.
It is assumed that the dynamics of the onboard control loop and the quadrocopter are much
faster such that the system to be controlled can be modelled as a rigid body with the desired

5
 ~ as direct input. The quadrocopter dynamics can then be modelled as
angular body rate Ω
  " #
q˙0 1   − 1 ⊤ ~
q Ω
q̇ = ~ =
= q·p Ω 2 1:3 (20)
q̇1:3 2 1 ~
2 (S (q1:3 ) + q0 I) Ω

with S(q1:3 ) denoting the skew-matrix of q1:3

 
0 −q3 q2
S(q1:3 ) =  q3 0 −q1  . (21)
−q2 q1 0

3 Attitude Control Design

As a first step, the commanded acceleration from the position loop and the commanded yaw
angle have to be converted into a desired attitude qcmd (Section 3.2). The objective is then
to design a feedback law which stabilizes the quadrocopter at any desired physical attitude.
Since any physical attitude in SO(3) corresponds to two antipodal quaternions in S3 , this can
be achieved by stabilizing the attitude q at ±qcmd . When neglecting this fact, quaternion-based
controllers can cause undesirable phenomena such as unwinding, where the rigid body rotates
unnecessarily through a full rotation [7, 2]. To solve the problem of unwinding, the controller
must satisfy
~ cmd (q) = Ω
Ω ~ cmd (−q). (22)
This can easily be seen, since q and −q represent the same physical attitude, the control output
should be the same as well. This issue can be solved by changing the sign of q whenever q0 < 0.
Changing the sign however results in a discontinuous controller. Interestingly, this is not a
drawback since it can be shown that all continuous state-feedback control laws are at most
almost globally stabilizing.

3.1 Control Law

Consider the following control law:
(
~ cmd (q) = 2 sgn(qe,0 )qe,1:3 , 1, qe,0 ≥ 0
Ω sgn(qe,0 ) = (23)
τ −1, qe,0 < 0

where

ˆ τ = first-order system time constant [s],

ˆ qe := q−1 · qcmd error measure, representing the rotation from q to qcmd .

Then, ±qcmd is a globally asymptotically stable equilibrium point of (20).

Proof. Without loss of generality, set qcmd = qI . Define an autonomous hybrid automaton [6]
H = (Z, Q, f, Init, Dom, E, G, R) with

ˆ Z = {z1 , z2 }, two discrete states, corresponding to the upper and lower hemisphere of S3 .

ˆ Q = S3 ⊂ R4 , continuous state q, which lies in the three-sphere and represents the current
attitude.

6
 ˆ A vector field   
1q · p Ω ~ cmd (z, q) , if z = z1
2  
f (z, q) = (24)
− 1 q · p Ω ~ cmd (z, q) , if z = z2
2

with (
2 2
~ cmd (z, q) =
Ω τ qe,1:3 = − τ q1:3 , if z = z1
(25)
− τ2 qe,1:3 = τ2 q1:3 , if z = z2
describing the evolution of q in time when the quaternion is in the upper hemisphere z1
or in the lower hemisphere z2 respectively.
ˆ Domains Dom(z1 ) = Dom(z2 ) = S3 .
ˆ Edges E = {(z1 , z2 ), (z2 , z1 )}, representing the possibility to go from z1 to z2 and vice
versa.
ˆ Guards G(z1 , z2 ) = {q ∈ S3 |q0 < 0}, and G(z2 , z1 ) = {q ∈ S3 |q0 < 0}
ˆ Reset maps R(z1 , z2 , q) = R(z2 , z1 , q) = −q, since a rotation of 180◦ about an axis is equal
to a rotation of −180◦ about the same axis.
Now, consider the following Lyapunov candidate function:
V (z, q) = q⊤
1:3 q1:3 + (q0 − 1)
2
∀z ∈ Z (26)
Differentiating (26) with respect to time and inserting (24) and (20) yields
∂V (z, q)
V̇ (z, q) = f (z, q) (27)
∂q
 1 ⊤ 2

 
= 2 (q0 − 1) 2q1:3 ⊤ 2 q1:3 τ q1:3 (28)
− 12 (S (q1:3 ) + q0 I) τ2 q1:3
2 2 ⊤
= (q0 − 1) q⊤
1:3 q1:3 − q1:3 (S(q1:3 ) + q0 I) q1:3 (29)
τ τ
2
= − q⊤ (S(q1:3 ) + I)q1:3 (30)
τ 1:3
2
= − q⊤ q1:3 ∀z ∈ Z (31)
τ 1:3
By Lyapunovs stability theorem for hybrid systems [6], the attitude qI is a stable equilibrium
because:
1. V (z, ·) = 0 if and only if q = qI ,
2. V (z, q) > 0 ∀q, (z, q) ∈ Dom(z)\{qI }, and
∂V (z,q)
3. V̇ (z, q) = ∂q f (z, q) ≤ 0 ∀q, (z, q) ∈ Dom(z).

Furthermore, V̇ (z, q) < 0 ∀z ∈ Z, ∀q ∈ Dom(z)\{qI } and V̇ (z, qI ) = 0 ∀z ∈ Z. Note that

V (z, q) does not jump when z changes its state, therefore V (z, q) is strictly decreasing, meaning
that qI is a globally asymptotically stable equilibrium point.

Remark 1. The error measure qe represents the rotation from the current attitude q to the
desired attitude qcmd . By taking the sign of qe , it is ensured that the rotation angle is always
less or equal to 180◦ . This means that the controller always rotates the rigid body in such a way
that the rotation angle is minimal. Note however, a minimal rotation angle does in general not
imply a time-optimal rotation maneuver [1, 5]. This can easily be seen if the angular body rates
are limited, i.e. |Ωi | ≤ Ωi,max , i ∈ {x, y, z} and Ωi,max 6= Ωj,max for i 6= j.

7
 q⊤
1:3 q1:3
1
~ cmd =
Ω 2
sgn (qe,0 ) qe,1:3
τ

~ cmd =
Ω 2
q
τ e,1:3

−1 1 q0

Figure 4: The proposed control law stabilizes the attitude to q0 = 1 or q0 = −1. However, there
is a discontinuity at q0 = 0. If the sign of qe,0 is not taken into consideration, the controller
would only be almost globally asymptotically stable with q0 = −1 as an unstable equilibrium
point [8].

Remark 2. A visual interpretation of the control law (23) for qcmd = qI is shown in Fig. 4.
Remark 3. By definition, kqk = 1 ∀q ∈ S3 , and therefore kqe,1:3 k ≤ 1. The maximum possible
~ cmd k ≤ 2 . This is especially an advantage for systems with
control output is thus limited to kΩ τ
saturations.
Remark 4. Assume, without loss of generality, that qcmd = qI . Inserting the control law (23)
into the state equation (20) yields
   2 
q̇0 q1 + q22 + q32
q̇1  1  −q0 q1 
q̇ =   
q̇2  = τ  −q0 q2  .
 (32)
q̇3 −q0 q3
For small deviations from the equilibrium qI (q0 ≈ 1, q1:3 ≪ 1), (32) simplifies to
   
q̇0 0
q̇1  1 −q1 
q̇ =    
q̇2  = τ −q2  , (33)
q̇3 −q3
which is an uncoupled first-order system with time-constant τ 3 .

3.1.1 Robustness
Although the proposed controller achieves global asymptotic stability of any desired attitude, it
can be shown that the controller is not robust to arbitrarily small measurement noise. Following
closely the proof in [8], a noise signal qnoise can be constructed such that when starting arbitrarily
close to the discontinuity, then q stays close to the discontinuity for all time.
Let the current attitude q be 

cos α2
q= ~
. (34)
k sin α2
Define qnoise to be   
β(q)
cos
qnoise (q) :=   2  , (35)
~k sin β(q)
2

3 1
Note that τ represents the time when the quaternion error decays to e
and not when the error in terms of
Euler angles decays to 1e .

8
where ~k is eigenaxis of the attitude q and


 β0 , if 0 < α ≤ π


−β , if π < α < 2π
0
β (q) = , with 0 < β0 < π. (36)

 −β 0 , if 0 ≥ α > −π


β , if − π ≥ α > −2π
0

e is the current attitude rotated by the noise signal, i.e.

If the measured attitude q
  
α+β(q)
cos
e = q · qnoise (q) = 
q  2  , (37)
~k sin α+β(q)
2

then, the kinematic equation for qcmd = qI is

(
1
q · p (sgn(q0 )e
q1:3 ) , if |π − α| ≤ β0
q̇ = τ 1 (38)
− τ q · p (sgn(q0 )e
q1:3 ) , if |π − α| > β0 .

Define the set of all discontinuities M := {q ∈ S3 : q0 = 0}, which corresponds to all attitudes
that are 180◦ rotation from qcmd . Now consider the Lyapunov candidate function VM (q) = q02 ,
with VM (S3 \M) > 0 and VM (M) = 0. Then, for |π − α| ≤ β0 ,

V̇M (q) = 2q0 q̇0 (39)

1
= − |q0 | q⊤ e1:3
1:3 q (40)
τ  
1 α α + β(q)
~ ⊤~
= − |q0 | k k sin sin (41)
τ 2 2
≤ 0. (42)

Thus, starting from any attitude with |π − α| ≤ β0 , the system does not converge to the desired
attitude.
In [8], a discontinuous controller with a hysteric memory state is proposed to avoid this unwanted
equilibrium. The hysteric memory state defines the rotation direction and is always chosen such
that the body rotates in direction of the shortest transition. If |π − α| ≤ β0 , the hysteric memory
state does not get updated and the rotation direction remains constant, thus leaving the region
where an arbitrarily small noise signal can destabilize the controller. In practice, a hysteric
memory state is not necessary if the control law is implemented on a discrete-time controller.
Since the output of a discrete-time controller is constant between two updates, the rotation
direction does not get changed during that time period and hence, a discrete-time controller
behaves like a continuous-time controller with hysteresis.

3.2 Desired Attitude

Given the quadrocopter setup in Fig. 2 with the body-fixed frame B, it is clear that the
quadrocopter can only accelerate in direction of ~eB z . Therefore, the commanded acceleration
~acmd has to be transformed into a target orientation qcmd such that the corresponding z-axis is
aligned with the desired acceleration. As with every pointing application, the rotation about the
pointing direction is irrelevant, i.e. the rotation about the thrust direction ~eB
z has no influence
on the translational behaviour of the quadrocopter. It makes therefore sense to split up the
control task into two parts:

9
 ˆ Reduced attitude control: Only the crucial pointing direction of the thrust is con-
trolled. The yaw angle is not controlled directly, but qcmd is always chosen such that no
rotation about the yaw axis is induced.

ˆ Full attitude control: The pointing direction of the thrust vector as well as the yaw
angle is controlled. qcmd is chosen such that the corresponding z-axis is aligned with ~eB
cmd,z
and ψ = ψcmd .

3.2.1 Reduced Attitude Control

Given any desired acceleration ~acmd , the desired thrust and direction can be computed by

I~
acmd
eB
I~cmd,z = , (43)
kI ~acmd k
collcmd = kI ~acmd k. (44)

The reduced error quaternion qe,red (Fig. 5), which rotates the quadrocopter from the current
attitude to the desired attitude is then




α  cos α2
cos 2  
qe,red = ~
=
I ~eB eB
z ×I ~
, (45)
k sin α2 sin α2 k ~B
e × ~e
cmd,z
B
k I z I cmd,z

where I eB
z is the corresponding z-axis of the current attitude q and α is the angle between the
current thrust direction and the desired thrust direction:
 ⊤ B

α = arccos I ~eB
z I~ ecmd,z . (46)

Finally, the desired attitude can be produced by

qcmd,red = q · qe,red . (47)

Note that last entry of qe,red is always zero because ~k ⊥ ~eB

z , which implies that Ωcmd,z = 0.
Due to the structure of the Euler angles, this does not mean that the yaw angle ψ is constant.

~eB
cmd,z
~eIz ~eB
z

~eB
x
~eIx
~k

~eB
y
~eIy

Figure 5: Geometric interpretation of qe,red .

10
3.2.2 Full Attitude Control
A desired yaw angle ψcmd and thrust direction ~eB cmd,z fully define the attitude. Pitch and roll
angles, θcmd and φcmd , can thus be recovered from ψcmd and ~eB eB
cmd,z using projections of ~cmd,z
onto the intermediate coordinate frames K and L.
First, the desired pointing direction ~eB
cmd,z is expressed in the K-frame, which is rotated by ψcmd
about the initial z-axis:
 
cos (ψcmd ) sin (ψcmd ) 0
K~eB eB
cmd,z = RKI I ~

cmd,z = − sin (ψcmd ) cos (ψcmd ) 0 I ~
 eB
cmd,z . (48)
0 0 1

The pitch angle θcmd can then easily be reconstructed by projecting the pointing vector onto
the ~eK eK
x ,~z -plane (Fig. 6): !
eB
K~ cmd,z,1
θcmd = arctan . (49)
eB
K~ cmd,z,3

Expressing ~eB
cmd,z in the L-frame, which is rotated about ~eK
y by the newly computed computed
pitch angle θcmd yields
 
cos (θcmd ) 0 − sin (θcmd )
L~eB eB
cmd,z = RLK K ~ cmd,z =
 0 1 0  K ~eB
cmd,z . (50)
sin (θcmd ) 0 cos (θcmd )

Again, projecting ~eB eL

cmd,z onto the ~ eL
y ,~z -plane, it can be seen from Fig. 7 that the roll angle
φcmd is given by  
φcmd = arctan2 −L~eB eB
cmd,z,2 , L~ cmd,z,3 . (51)

Finally, the desired attitude can be constructed using (19):

qcmd,full = q (ψcmd , θcmd , φcmd ) . (52)

~eB
cmd,z
~eL
z
~eIz =~eK
z
θ

~eIx

~eK ψ ~eK eL
x y =~ y

~eIy

Figure 6: Projection onto the ~eK eK

x ,~z -plane.

11
 ~eB
cmd,z
~eL
z
~eIz =~eK
z
φ

~eB
y
~eIx
~eL eB
x =~ x
θ

~eK ψ ~eK eL
x y =~ y

~eIy

Figure 7: Projection onto the ~eL eL

y ,~z -plane.

3.2.3 Mixing Full and Reduced Attitude Control

Although it is not necessary that yaw is controlled in order to follow any desired trajectory, it is
nevertheless sometimes desirable. In practice, using full attitude control (qcmd = qcmd,full ) does
not yield good results. The dynamics of Ωz are much slower than for Ωx and Ωy . The reason
for this is that Ωz can not be controlled by applying differential thrust but only by exploiting
the air drag of the propellers. When controlling all angular rates with the same gain τ1 , either
the controller is very slow for the pitch and roll angles or there is a lot of overshoot in yaw.
Furthermore, since with the control structure kΩ~ cmd k is limited, a lot of control effort might be
wasted to control yaw although it is not crucial for the quadrocopter motion. This can become
important for very aggressive maneuvers. One possibility to overcome these issues is to mix
reduced and full attitude control. Define qmix := q−1
cmd,red · qcmd,full to be the rotation between
qcmd,red and qcmd,full . By definition, qmix always has the form


cos αmix
2
 0 
qmix =  , (53)
0

sin αmix
2

where αmix is the rotation angle between qcmd,red and qcmd,full about B ~eB
z . Choosing qcmd to
lie anywhere in between the two attitudes qcmd,red and qcmd,full , i.e.


cos pα2mix
 0 
qcmd = qcmd,red · 

,
 with p ∈ [0, 1] , (54)
0

sin pαmix
2

then it is guaranteed that the resulting rotation yields correct pitch and roll angles. Although
the controller does only correct p-fraction of the actual yaw error, it nevertheless converges to
the desired yaw angle as t → ∞. Ωcmd,z is now limited to


2 sin pπ
2
|Ωcmd,z | ≤ . (55)
τ

12
3.2.4 Limiting Maximum Tilt Angle
For some applications, e.g. carrying loads, it might be wanted that the quadrocopter can only
tilt up to some maximum tilt angle αmax . This can be implemented right after the step in (43).
First, the tilt angle αtilt is defined as the angle between ~eIz and ~eB
cmd,z :
 ⊤

αtilt = arccos eIz
I~ eB
I~ cmd,z (56)

If |αtilt | > αmax , the angle can be limited using (17). First, the rotation axis is computed,

eIz
I~ × I ~eB
cmd,z
~k = , (57)
⊤
kI ~eIz eB
I~ cmd,z k

and a tilt rotation is defined: 

αtilt 
cos 2
.
qtilt = ~ αtilt (58)
k sin 2
The new target thrust direction, which lies within a cone with an aperture of 2αmax , is then
   
p I ~eB
new cmd,z = qtilt · p I ~
e B
cmd,z · q̄tilt . (59)

It is worth mentioning that this does not prevent the quadrocopter from flipping over. Although
the reachable set of attitudes is now constrained, the quadrocopter might get rotated by some
external momentum to an extreme attitude. The shortest angular rotation to recover to the
commanded attitude can then still involve performing a flip.

3.3 Heuristic
The proposed control law (23) always stabilizes the quadrocopter to the desired orientation with
the shortest possible rotation. This is a very desirable result, however, it was assumed that Ω ~
can be controlled directly and infinitely fast, which is in practise not true. Imagine the case
where the quadrocopter has a yaw error of almost 180◦ . It does not make a big difference in
time if the quadrocopter rotates about ~k or −~k. However, if |Ωz | is large, it is obviously faster
to rotate in the same direction as Ωz than to accelerate in the opposite direction. The question,
whether it is faster to undertake the longer rotation or not, is not trivial. In fact, the rise time
~ depends on many parameters, e.g. thrust magnitude and the value of Ω
of Ω ~ itself and no
algebraic expression exists to answer this question. To keep the computational effort small, a
simple heuristic is used.
Here, only the heuristic for errors in yaw is introduced, however, it works the same way for pitch
and roll angles. Assume that the quadrocopter has an angular body rate of Ω. ~ The rotation
~ ⊤ ~ ⊤
axis of qmix is either k = (0, 0, 1) or k = (0, 0, −1) . If the quadrocopter already rotates in
the same direction as commanded, i.e. ~k⊤ Ω ~ ≥ 0, then obviously nothing has to be changed.
Otherwise, if ~k⊤ Ω
~ < 0, it might be faster, depending on the error in yaw, to undertake the
longer rotation. The decision criterion is
 
~ < 0 AND ~k⊤ Ω
~k⊤ Ω ~  ≥ threshold (αmix ) . (60)

For errors in yaw of almost 180◦ , the threshold should be close to 0 and should strictly increase
for smaller error angles. A good threshold was found experimentally to be
π − αmix
threshold (αmix ) = Ωz,min . (61)
π

13
  
 ~
Whenever ~k⊤ Ω is above the threshold, it is better to continue in this direction. To gain as
much time as possible, the rotation until an error angle of 180◦ should be conducted with full
speed, i.e. 

cos pπ
2
qe,full = qe,red · ~
. (62)
−k sin pπ
2
Note that it has to be ensured that once the rotation direction is decided, it does not get changed
anymore. Therefore, the commanded rotation rate should be well above the threshold which
implies that

2 sin pπ
2
Ωz,min < . (63)
τ
Because unit quaternions are normalized and large errors in pitch and roll can decrease the
Ωcmd,z , this is not ensured when Ωz,min = 2 sin (pπ/2) /τ .

Remark 1. The above heuristic also works for Ωx,y . However, Ωx,y can be controlled much
faster and therefore the time earning is not that large anymore. Experiments have shown that
the heuristic is only beneficial for extremely large angular rates in pitch and roll, which are
almost out of the measurement range of the gyroscopic sensors. Furthermore, the heuristic is
designed for a single rotation about a principle axis of the quadrocopter. It is not totally clear
what happens in the 3D case. It is therefore recommended to only apply the heuristic to the yaw
axis, which can be controlled only slowly and therefore the time earnings are the highest.

Remark 2. Instead of the heuristic (60), it was also attempted to solve the problem by simulating
the state equation (20) forward in time, once with a rotation in the shorter direction and once
in opposite direction. However, the results were not any better than with the above heuristic.
Moreover, the simulation was computationally very expensive which made it useless for real-time
uses.

Remark 3. If using the above heuristic, then stability and robustness of the controller would
have to be proven again. Although the results are quite intuitive, the proof is not done here.

4 Results
4.1 Time constant
The proposed controller (23) is based upon the assumption that Ω ~ can be controlled directly.
This assumption certainly holds for large τ . In order to determine how small τ can be chosen
such that the above assumption still holds, a series of step responses in attitude are analyzed.

4.1.1 Step in Pitch Angle

Starting from hovering, a step of 45◦ in pitch angle φ is commanded. As the problem is symmetric
in pitch and roll, only the step response for pitch is analyzed. Fig. 8 shows the result for various
τ and p such that τyaw = 0.02s (64).
For τ ≥ 0.1s, q2 behaves like a first order system. For τ < 0.1s, the assumption that Ω ~ can be
controlled directly does not hold anymore. At time t = 0.2s, the commanded angular pitch rate
Ωcmd,y is equal to zero, but Ωy (t = 0.2s) ≫ 0rad/s and hence, the system exhibits even overshoot.
Therefore, fast responses with almost no overshoot can be achieved with τ ∈ [0.08s, 0.15s].

14
 0.5 8

0.4 6

0.3

 rad 
4

s
q2

Ωy
0.2
τ = 0.08s 2
τ = 0.1s
0.1
τ = 0.15s
τ = 0.2s 0
0
0 0.2 0.4 0.6 0 0.2 0.4 0.6
Time [s] Time [s]

Figure 8: Step response of q2 and Ωy . The dotted lines are the corresponding Ωcmd,y .

4.1.2 Step in Yaw Angle

A step input of magnitude 180◦ in yaw is given into the attitude controller at t = 0s, starting
from hovering with ψ = 0◦ . The results for different τ (p = 1) are shown in Fig. 9.
For τ ≥ 0.2s, q3 certainly behaves very similar to a first-order system. For smaller τ , the system
begins to show overshoot since the fundamental assumption is hurt. This can also be seen in
Fig. 9, where Ωz reached saturation for τ = 0.1s even when decelerating. For controlling yaw,
τ should therefore lie in the range of [0.2s, 0.4s].
Remark 1. In Fig. 9, the response of q3 for τ = 0.5s does oscillate slightly. This is due to the
small control gain of this time constant which cannot prevent the quadrocopter from oscillating
about its pitch and roll axis.
Remark 2. The above results strongly depend on the total thrust. For both experiments, τ was
analyzed with a collective thrust that allows to maintain altitude.

1 8

0.8 6

4
0.6
 rad 

2
s
q3

τ = 0.1s
Ωz

0.4 0
τ = 0.2s
τ = 0.3s −2
0.2
τ = 0.4s
−4
τ = 0.5s
0
−6
0 1 2 3 0 1 2 3
Time [s] Time [s]

Figure 9: Step response of q3 and Ωz . The dotted lines are the corresponding Ωcmd,z .

15
4.2 Mixing Full and Reduced Attitude Control
The experiments in Section 4.1 have shown that a good controller gain for the yaw axis is
significantly smaller than for roll and pitch axis. One method to implement this is to mix full
and reduced attitude control and to choose the prioritization factor p of (54) to be
τ
p= , (64)
τyaw

where τyaw is a time constant which attains good responses in yaw and τ is the desired time con-
stant of the overall controller (for pitch and roll). Equation (64) originated from a linearization
of the control law with mixed attitude control for small errors in yaw. If p is chosen according
to (64), then for small errors in yaw, the controller behaves just like one with a time constant
of τyaw instead of τ . For a very aggressive controller with τ = 0.08s and τyaw = 0.2s, p is

0.08s
p= = 0.4. (65)
0.2s
In Fig 10, an experiment with different values for p and τ = 0.08s is shown. Starting from a
position with Euler angles (ψ, θ, φ) = (180◦ , 45◦ , 0◦ ), a step of 180◦ in yaw is commanded at
t = 0s.
As a benchmark, the results with p = 0 are plotted in Fig. 10a. The rotation about the pitch
angle does not change significantly for different values of p and the correct tilt angle is achieved
in all cases shortly after 0.2s. But examining Fig. 10 more carefully, it can be seen that with
increasing p, the pitch angle is corrected slightly slower. This is due to the increasing amount
of control power that is taken up by yaw with larger p. Remarkable is that the larger p is, the
faster yaw is controlled. For p = 0.4, the correct attitude is achieved at t = 0.88s whereas for
p = 0.267 and p = 0.133, the desired attitude is not yet achieved at t = 1s.

Remark 1. For extremely large angular rates about ~eB z , mixed attitude control achieves again
better results than pure reduced attitude control. If |Ωz | is large, than the eigenaxis ~k of qe,red
rotates with a speed of Ωz in the ~eB eB
x ,~y -plane. This implies that the rotational velocity of the
motors increase and decrease constantly. However, there are limits on the motor dynamics and
the motor speed can not be controlled infinitely fast. Further, since Ωcmd,z = 0, a lot of control
effort is used by the onboard controller to break down Ωz , leaving almost no control power to
rotate the quadrocopter about its pitch and roll axis.

16
 ~eIz

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time [s]
a)
~eIz

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
~eIy

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time [s]
b)
~eIz

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
~eIy

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time [s]
c)
~eIz

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
~eIy

~eIx 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time [s]
d)

Figure 10: View of the quadrocopter in the xz-plane and xy-plane respectively for different
prioritization factors: a) p = 0.0, b) p = 0.133, c) p = 0.267 and d) p = 0.4.

17
4.3 Heuristic
An experiment with the heuristic (60) enabled is carried out to show its effectiveness. Starting
from hovering, the quadrocopter rotates about its yaw axis with different angular rates. Once
the yaw angle ψ reaches 90◦ , the quadrocopter is commanded to return to ψcmd = 0◦ . The
control parameters for this experiment are τ = 0.1s, p = 0.3 and Ωz,min = 4 rad
s . The results are
shown in Fig. 11.

180 8
Yaw angle ψ [◦ ]

90 4

 rad 
s
0 0
Heuristic on

Ωz
−90 −4
Heuristic off
−180 −8
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time [s] Time [s]
a)
180 8
Yaw angle ψ [◦ ]

90 4
 rad 
s

0 0
Ωz

−90 −4

−180 −8
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time [s] Time [s]
b)
180 8
Yaw angle ψ [◦ ]

90 4
 rad 
s

0 0
Ωz

−90 −4

−180 −8
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time [s] Time [s]
c)

Figure 11: Experiment testing the heuristic for different angular rates: a) Ωz = 1.5 rad
s , b)
rad rad
Ωz = 4 s and c) Ωz = 6.5 s .

Clearly, for slow angular rates, there is no difference if the heuristic is enabled or not (Fig. 11a).
But if Ωz is beyond a certain threshold, it is faster to undertake the longer rotation and hence,
the controller with the heuristic enabled attains better results.

5 Conclusion
A nonlinear attitude controller based on unit quaternions for the quadrocopters in the Flying
Machine Arena has been developed. The control inputs are a desired acceleration and yaw
angle, which are then converted to a commanded target orientation. It was assumed that the

18
onboard controller and quadropcopter dynamics are much faster than the attitude control loop
and hence, the quadrocopter can be controlled directly by Ω. ~ The stability and robustness of
the controller were then proven.
By experiments, the performance of the nonlinear controller was demonstrated. The proposed
attitude controller has many benefits, especially for aggressive maneuvers. Due to the global
asymptotic stability, the proposed controller is able to recover the quadrocopter to hovering
from arbitrary attitudes and can track any feasible trajectory. Furthermore, it was shown that
the control of the crucial pitch and roll angles can be prioritized by mixing full and reduced
attitude control. The proposed heuristic is able improve the time to reach a desired attitude,
however, it only yields good results for rotations about the yaw axis.
Nevertheless, there are still many problems to be solved. Currently, the rise time of Ω ~ depends
on the commanded thrust. This can cause very slow rotation rates if the commanded thrust is
low although the whole rotation would take only a few tenths of a second and thus, a large thrust
would have only minor influence on the translational motion. Furthermore, it would be nice if
the prioritization factor p is adapted dynamically depending on the attitude error. Large errors
in pitch and roll should be controlled first, but if no errors in pitch and roll occur, then p can
be large such that errors in yaw are corrected in a reasonable time. Experiments also revealed
that a yaw Euler angle ψ is not well suited to define a desired rotation about the quadrocopter’s
z-axis. Due to the pitch angle θ, which is only defined in a range of [ −π π
2 , 2 ], jumps in ψ occur if
the pitch angle would exceed its range. If these jumps are not considered in the input trajectory,
the quadrocopter may rotate unintentionally about its yaw axis.
Currently, the proposed controller is model independent and can therefore be used for various
flying vehicles. However, it might be beneficial in terms of time to combine the attitude and
onboard controller and take model parameters such as moments of inertia into consideration.

19
References
[1] K. Bilimoria and B. Wie. Time-optimal three-axis reorientation of a rigid spacecraft. Jour-
nal of Guidance Control and Dynamics, 16(3):446–452, 1993.

[2] N.A. Chaturvedi, A.K. Sanyal, and N.H. McClamroch. Rigid-body attitude control. Control
Systems, IEEE, 31(3):30 –51, june 2011.

[3] J. Diebel. Representing attitude: Euler angles, unit quaternions, and rotation vectors, 2006.

[4] J. Dvorak, M. de Lellis, and Z. Hurak. Advanced control of quadrotor using eigenaxis
rotation. In Control Applications (CCA), 2011 IEEE International Conference on, pages
153 –158, sept. 2011.

[5] T. Lee, M. Leok, and N.H. McClamroch. Time optimal attitude control for a rigid body.
In American Control Conference, 2008, pages 5210 –5215, june 2008.

[6] J. Lygeros, S. Sastry, and C. Tomlin. Hybrid systems: Foundations, advanced topics and
applications.

[7] C.G. Mayhew, R.G. Sanfelice, and A.R. Teel. On quaternion-based attitude control and the
unwinding phenomenon. In American Control Conference (ACC), 2011, pages 299 –304,
29 2011-july 1 2011.

[8] C.G. Mayhew, R.G. Sanfelice, and A.R. Teel. Quaternion-based hybrid control for robust
global attitude tracking. Automatic Control, IEEE Transactions on, 56(11):2555 –2566,
nov. 2011.

[9] D. Mellinger and V. Kumar. Minimum snap trajectory generation and control for quadro-
tors. In ICRA, pages 2520–2525. IEEE, 2011.

[10] K Rudin, M.-D. Hua, G. Ducard, and S. Bouabdallah. A robust attitude controller and its
application to quadrotor helicopters. In 18th IFAC World Congress, 2011.

[11] J. Stuelpnagel. On the parametrization of the three-dimensional rotation group. SIAM

Review, 6(4):pp. 422–430, 1964.

[12] A. Tayebi. Unit quaternion-based output feedback for the attitude tracking problem. Au-
tomatic Control, IEEE Transactions on, 53(6):1516 –1520, july 2008.

[13] A. Tayebi and S. McGilvray. Attitude stabilization of a vtol quadrotor aircraft. Control
Systems Technology, IEEE Transactions on, 14(3):562 – 571, may 2006.

[14] J.T.-Y. Wen and K. Kreutz-Delgado. The attitude control problem. Automatic Control,
IEEE Transactions on, 36(10):1148 –1162, oct 1991.

20