See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265825340

Quadrotor Dynamics and Control

Article

CITATIONS READS

100 2,139

1 author:

Randal W. Beard
Brigham Young University - Provo Main Campus
224 PUBLICATIONS   18,655 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Aerial Recovery View project

All content following this page was uploaded by Randal W. Beard on 01 June 2015.

The user has requested enhancement of the downloaded file.

 Quadrotor Dynamics and Control
Randal W. Beard
Brigham Young University
February 19, 2008

1 Reference Frames
This section describes the various reference frames and coordinate systems that
are used to describe the position of orientation of aircraft, and the transformation
between these coordinate systems. It is necessary to use several different coordi-
nate systems for the following reasons:

• Newton’s equations of motion are given the coordinate frame attached to

the quadrotor.

• Aerodynamics forces and torques are applied in the body frame.

• On-board sensors like accelerometers and rate gyros measure information

with respect to the body frame. Alternatively, GPS measures position,
ground speed, and course angle with respect to the inertial frame.

• Most mission requirements like loiter points and flight trajectories, are spec-
ified in the inertial frame. In addition, map information is also given in an
inertial frame.

One coordinate frame is transformed into another through two basic opera-
tions: rotations and translations. Section 1.1 develops describes rotation matrices
and their use in transforming between coordinate frames. Section 1.2 describes
the specific coordinate frames used for micro air vehicle systems. In Section 1.3
we derive the Coriolis formula which is the basis for transformations between
between between translating and rotating frames.

1
1.1 Rotation Matrices
We begin by considering the two coordinate systems shown in Figure 1. The

Figure 1: Rotation in 2D

vector p can be expressed in both the F 0 frame (specified by (î0 , ĵ 0 , k̂ 0 )) and in

the F 1 frame (specified by (î1 , ĵ 1 , k̂ 1 )). In the F 0 frame we have
p = p0x î0 + p0y ĵ 0 + p0z k̂ 0 .
Alternatively in the F 1 frame we have
p = p1x î1 + p1y ĵ 1 + p1z k̂ 1 .
Setting these two expressions equal to each other gives
p1x î1 + p1y ĵ 1 + p1z k̂ 1 = p0x î0 + p0y ĵ 0 + p0z k̂ 0 .

Taking the dot product of both sides with î1 , ĵ 1 , and k̂ 1 respectively, and stacking
the result into matrix form gives
 1  1 0 1 0 1 0   0
px î · î î · ĵ î · k̂ px
1 4  1   
1 0 1
p = py = ĵ · î ĵ · ĵ ĵ · k̂ 0 1 0 p0y  .
1
pz k̂ 1 · î0 k̂ 1 · ĵ 0 k̂ 1 · k̂ 0 p0z

From the geometry of Figure 1 we get

p1 = R01 p0 , (1)

2
where  
cos(θ) sin(θ) 0
4
R01 = − sin(θ) cos(θ) 0 .
0 0 1
The notation R01 is used to denote a rotation matrix from coordinate frame F 0 to
coordinate frame F 1 .
Proceeding in a similar way, a right-handed rotation of the coordinate system
about the y-axis gives
 
cos(θ) 0 − sin(θ)
4
R01 =  0 1 0 ,
sin(θ) 0 cos(θ)

and a right-handed rotation of the coordinate system about the x-axis resultes in
 
1 0 0
4
R01 = 0 cos(θ) sin(θ)  .
0 − sin(θ) cos(θ)

As pointed out in [1], the negative sign on the sin term appears above the line with
only ones and zeros.
The matrix R01 in the above equations are examples of a more general class of
rotation matrices that have the following properties:

P.1. (Rab )−1 = (Rab )T = Rba .

P.2. Rbc Rab = Rac .

P.3. det Rab = 1.

In the derivation of Equation (1) note that the vector p remains constant and
the new coordinate frame F 1 was obtained by rotating F 0 through a righted-
handed rotation of angle θ.
We will now derive a formula, called the rotation formula that performs a
left-handed rotation of a vector p about another vector n̂ by an angle of µ. Our
derivation follows that given in [1]. Consider Figure 2 which is similar to Figure
1.2-2 in [1]. The vector p is rotated, in a left-handed sense, about a unit vector n̂
by an angle of µ to produce the vector q. The angle between p and n̂ is φ. By
geometry we have that
q = ON ~ + N~W + W~Q. (2)

3
Figure 2: Left-handed rotation of a vector p about the unit vector n̂ by an angle
of µ to obtain the vector q.

The vector ON~ can be found by taking the projection of p on the unit vector n̂ in
the direction of n̂:
~ = (p · n̂) n̂.
ON
The vector N~W is in the direction of p − ON
~ with a length of N Q cos µ. Noting
° °
° ~ °
that the length N Q equals the length N P which is equal to °p − ON ° we get
that
p − (p · n̂) n̂
N~W = N Q cos µ
kp − (p · n̂) n̂k
= (p − (p · n̂) n̂) cos µ.

The vector W~Q is perpendicular to both p and n̂ and has length N Q sin µ. Noting
that N Q = kpk sin φ we get
p × n̂
W~Q = N Q sin µ
kpk sin φ
= −n̂ × p sin µ.
Therefore Equation (2) becomes
q = (1 − cos µ) (p · n̂) n̂ + cos µp − sin µ (n̂ × p) , (3)

4
 Figure 3: Rotation of p about the z-axis.

which is called the rotation formula.

As an example of the application of Equation (3) consider a left handed rota-
tion of a vector p0 in frame F 0 about the z-axis as shown in Figure 3.
Using the rotation formula we get

q0 = (1 − cos θ)(p · n̂)n̂ + cos φp − sin φn̂ × p

   0  0
0 px −py
0   0
= (1 − cos φ)pz 0 + cos φ py − sin φ p0x  
1 p0z 0
 
cos φ sin φ 0
= − sin φ cos φ 0 p0
0 0 1
= R01 p0 .

Note that the rotation matrix R01 can be interpreted in two different ways. The
first interpretation is that it transforms the fixed vector p from an expression in
frame F 0 to an expression in frame F 1 where F 1 has been obtained from F 0
by a right-handed rotation. The second interpretation is that it rotates a vector
p though a left-handed rotation to a new vector q in the same reference frame.
Right-handed rotations of vectors are obtained by using (R01 )T .

5
1.2 Quadrotor Coordinate Frames
For quadrotors there are several coordinate systems that are of interest. In this
section we will define and describe the following coordinate frames: the inertial
frame, the vehicle frame, the vehicle-1 frame, the vehicle-2 frame, and the body
frame. Throughout the book we assume a flat, non-rotating earth: a valid assump-
tion for quadrotors.

1.2.1 The inertial frame F i .

The inertial coordinate system is an earth fixed coordinate system with origin at
the defined home location. As shown in Figure 4, the unit vector îi is directed
North, ĵ i is directed East, and k̂ i is directed toward the center of the earth.

Figure 4: The inertial coordinate frame. The x-axis points North, the y-axis points
East, and the z-axis points into the Earth.

1.2.2 The vehicle frame F v .

The origin of the vehicle frame is at the center of mass of the quadrotor. However,
the axes of F v are aligned with the axis of the inertial frame F i . In other words,
the unit vector îv points North, ĵ v points East, and k̂ v points toward the center of
the earth, as shown in Figure 5.

1.2.3 The vehicle-1 frame F v1 .

The origin of the vehicle-1 frame is identical to the vehicle frame, i.e, the the
center of gravity. However, F v1 is positively rotated about k̂ v by the yaw angle ψ
so that if the airframe is not rolling or pitching, then îv1 would point out the nose

6
Figure 5: The vehicle coordinate frame. The x-axis points North, the y-axis points
East, and the z-axis points into the Earth.

Figure 6: The vehicle-1 frame. If the roll and pitch angels are zero, then the x-axis
points out the nose of the airframe, the y-axis points out the right wing, and the
z-axis points into the Earth.

of the airframe, ĵ v1 points out the right wing, and k̂ v1 is aligned with k̂ v and points
into the earth. The vehicle-1 frame is shown in Figure 6.
The transformation from F v to F v1 is given by

pv1 = Rvv1 (ψ)pv ,

where  
cos ψ sin ψ 0
Rvv1 (ψ) = − sin ψ cos ψ 0 .
0 0 1

7
1.2.4 The vehicle-2 frame F v2 .
The origin of the vehicle-2 frame is again the center of gravity and is obtained by
rotating the vehicle-1 frame in a right-handed rotation about the ĵ v1 axis by the
pitch angle θ. If the roll angle is zero, then îv2 points out the nose of the airframe,
ĵ v2 points out the right wing, and k̂ v2 points out the belly, a shown in Figure 7.

Figure 7: The vehicle-2 frame. If the roll angel is zero, then the x-axis points out
the nose of the airframe, the y-axis points out the right wing, and the z-axis points
out the belly.

The transformation from F v1 to F v2 is given by

pv2 = Rv1
v2
(θ)pv1 ,

where  
cos θ 0 − sin θ
v2
Rv1 (θ) =  0 1 0 .
sin θ 0 cos θ

1.2.5 The body frame F b .

The body frame is obtained by rotating the vehicle-2 frame in a right handed
rotation about îv2 by the roll angle φ. Therefore, the origin is the center-of-gravity,
îb points out the nose of the airframe, ĵ b points out the right wing, and k̂ b points
out the belly. The body frame is shown in Figure 8.
The transformation from F v2 to F b is given by

pb = Rv2
b
(φ)pv2 ,

8
Figure 8: The body frame. The x-axis points out the nose of the airframe, the
y-axis points out the right wing, and the z-axis points out the belly.

where  
1 0 0
b
Rv2 (φ) = 0 cos φ sin φ  .
0 − sin φ cos φ
The transformation from the vehicle frame to the body frame is given by

Rvb (φ, θ, ψ) = Rv2

b v2
(φ)Rv1 (θ)Rvv1 (ψ)
   
1 0 0 cos θ 0 − sin θ cos ψ sin ψ 0
= 0 cos φ sin φ   0 1 0  − sin ψ cos ψ 0
0 − sin φ cos φ sin θ 0 cos θ 0 0 1
 
cθcψ cθsψ −sθ
= sφsθcψ − cφsψ sφsθsψ + cφcψ sφcθ ,

cφsθcψ + sφsψ cφsθsψ − sφcψ cφcθ
4 4
where cφ = cos φ and sφ = sin φ.

1.3 Equation of Coriolis

In this section we provide a simple derivation of the famous equation of Coriolis.
We will again follow the derivation given in [1].
Suppose that we are given two coordinate frames F i and F b as shown in Fig-
ure 9. For example, F i might represent the inertial frame and F b might represent
the body frame of a quadrotor. Suppose that the vector p is moving in F b and that
F b is rotating and translating with respect to F i . Our objective is to find the time
derivative of p as seen from frame F i .
We will derive the appropriate equation through two steps. Assume first that
F is not rotating with respect to F i . Denoting the time derivative of p in frame
b

9
 Figure 9: Derivation of the equation of Coriolis.

d
F i as dti
p we get
d d
p= p. (4)
dti dtb
On the other hand, assume that p is fixed in F b but that F b is rotating with respect
to F i , and let ŝ be the instantaneous axis of rotation and δφ the (right-handed)
rotation angle. Then the rotation formula (3) gives

p + δp = (1 − cos(−δφ))ŝ(ŝ · p) + cos(−δφ)p − sin(−δφ)ŝ × p.

Using a small angle approximation and dividing both sides by δt gives

δp δφ
≈ ŝ × p.
δt δt
Taking the limit as δt → 0 and defining the angular velocity of F b with respect to
4
F i as ωb/i = ŝφ̇ we get
d
p = ωb/i × p. (5)
dti
Since differentiation is a linear operator we can combine Equations (4) and (5) to
obtain
d d
p= p + ωb/i × p, (6)
dti dtb
which is the equation of Coriolis.

10
2 Kinematics and Dynamics
In this chapter we derive the expressions for the kinematics and the dynamics of a
rigid body. While the expressions derived in this chapter are general to any rigid
body, we will use notation and coordinate frames that are typical in the aeronautics
literature. In particular, in Section 2.1 we define the notation that will be used for
the state variables of a quadrotor. In Section 2.2 we derive the expressions for the
kinematics, and in Section 2.3 we derive the dynamics.

2.1 Quadrotor State Variables

The state variables of the quadrotor are the following twelve quantities

pn = the inertial (north) position of the quadrotor along îi in F i ,

pe = the inertial (east) position of the quadrotor along ĵ i in F i ,
h = the altitude of the aircraft measured along −k̂ i in F i ,
u = the body frame velocity measured along îb in F b ,
v = the body frame velocity measured along ĵ b in F b ,
w= the body frame velocity measured along k̂ b in F b ,
φ= the roll angle defined with respect to F v2 ,
θ= the pitch angle defined with respect to F v1 ,
ψ= the yaw angle defined with respect to F v ,
p = the roll rate measured along îb in F b ,
q = the pitch rate measured along ĵ b in F b ,
r = the yaw rate measured along k̂ b in F b .

The state variables are shown schematically in Figure 10. The position (pn , pe , h)
of the quadrotor is given in the inertial frame, with positive h defined along the
negative Z axis in the inertial frame. The velocity (u, v, w) and the angular veloc-
ity (p, q, r) of the quadrotor are given with respect to the body frame. The Euler
angles (roll φ, pitch θ, and yaw χ) are given with respect to the vehicle 2-frame,
the vehicle 1-frame, and the vehicle frame respectively.

11
 Figure 10: Definition of Axes

2.2 Quadrotor Kinematics

The state variables pn , pe , and −h are inertial frame quantities, whereas the ve-
locities u, v, and w are body frame quantities. Therefore the relationship between
position and velocities is given by
   
pn u
d   v 
p e = Rb v
dt
−h w
 
u
b T  
= (Rv ) v
w
  
cθcψ sφsθcψ − cφsψ cφsθcψ + sφsψ u
= cθsψ sφsθsψ + cφcψ cφsθsψ − sφcψ   v  .
−sθ sφcθ cφcθ w

The relationship between absolute angles φ, θ, and ψ, and the angular rates p,
q, and r is also complicated by the fact that these quantities are defined in different
coordinate frames. The angular rates are defined in the body frame F b , whereas
the roll angle φ is defined in F v2 , the pitch angle θ is defined in Fv1 , and the yaw
angle ψ is defined in the vehicle frame F v .
We need to relate p, q, and r to φ̇, θ̇, and ψ̇. Since φ̇, θ̇, ψ̇ are small and noting
that
b v2
Rv2 (φ̇) = Rv1 (θ̇) = Rvv1 (ψ̇) = I,

12
we get
       
p φ̇ 0 0
q  = Rv2
b
(φ̇)  0  + Rv2
b
(φ)Rv1v2
(θ̇) θ̇ + Rv2
b v2
(φ)Rv1 (θ)Rv→v1 (ψ̇)  0 
r 0 0 ψ̇
        
φ̇ 1 0 0 0 1 0 0 cos θ 0 − sin θ 0
=  0  + 0 cos φ sin φ  θ̇  + 0 cos φ sin φ   0 1 0  0
0 0 − sin φ cos φ 0 0 − sin φ cos φ sin θ 0 cos θ ψ̇
  
1 0 −sθ φ̇

= 0 cφ sφcθ   θ̇  . (7)
0 −sφ cφcθ ψ̇
Inverting we get
    
φ̇ 1 sin(φ) tan(θ) cos(φ) tan(θ) p
 θ̇  = 0 cos(φ) − sin(φ)   q . (8)
ψ̇ 0 sin(φ) sec(θ) cos(φ) sec(θ) r

2.3 Rigid Body Dynamics

Let v be the velocity vector of the quadrotor. Newton’s laws only hold in inertial
frames, therefore Newton’s law applied to the translational motion is
dv
m = f,
dti
where m is the mass of the quadrotor, f is the total applied to the quadrotor, and
d
dti
is the time derivative in the inertial frame. From the equation of Coriolis we
have µ ¶
dv dv
m =m + ω b/i × v = f , (9)
dti dtb
where ω b/i is the angular velocity of the airframe with respect to the inertial frame.
Since the control force is computed and applied in the body coordinate system,
and since ω is measured in body coordinates, we will express Eq (9) in body
4 4
coordinates, where vb = (u, v, w)T , and ωb/i b
= (p, q, r)T . Therefore, in body
coordinates, Eq. (9) becomes
     
u̇ rv − qw fx
 v̇  = pw − ru + 1 fy  , (10)
m
ẇ qu − pv fz

13
 4
where f b = (fx , fy , fz )T .
For rotational motion, Newton’s second law states that

dhb
= m,
dti
where h is the angular momentum and m is the applied torque. Using the equation
of Coriolis we have
dh dh
= + ω b/i × h = m. (11)
dti dtb
Again, Eq. (11) is most easily resolved in body coordinates where hb = Jωb/i
b

where J is the constant inertia matrix given by

R 2 R R 
(y R+ z 2 ) dm R − xy dm − R xz dm
J =  − R xy dm (x2R+ z 2 ) dm R − yz dm 
− xz dm − yz dm (x2 + y 2 ) dm
 
Jx −Jxy −Jxz
4
= −Jxy Jy −Jyz  .

−Jxz −Jyz Jz

Figure 11: The moments of inertia for the quadrotor are calculated assuming a
spherical dense center with mass M and radius R, and point masses of mass m
located at a distance of ` from the center.
As shown in Figure 11, the quadrotor is essentially symmetric about all three
axes, therefore Jxy = Jxz = Jyz = 0 which implies that
 
Jx 0 0
J =  0 Jy 0  .
0 0 Jz

14
Therefore
 1 
Jx
0 0
J−1 =  0 1
Jy
0 .
1
0 0 Jz

The inertia for a solid sphere is given by J = 2M R2 /5[2]. Therefore

2M R2
Jx = + 2`2 m
5
2M R2
Jy = + 2`2 m
5
2M R2
Jz = + 4`2 m.
5
4
Defining mb = (τφ , τθ , τψ )T we can write Eq. (11) in body coordinates as
  1        
ṗ Jx
0 0 0 r −q Jx 0 0 p τφ
q̇  =  0 J1 0  −r 0 p   0 Jy 0  q  +  τθ 
y
ṙ 0 0 J1z q −p 0 0 0 Jz r τψ
 J −J   
1
y
Jx
z
qr τ
Jx φ
 pr 
 +  Jy τθ  .
1
=  JzJ−J
y
x

Jx −Jy 1
Jz
pq τ
Jz ψ

The six degree of freedom model for the quadrotor kinematics and dynamics
can be summarized as follows:
    
ṗn cθcψ sφsθcψ − cφsψ cφsθcψ + sφsψ u
 ṗe  = cθsψ sφsθsψ + cφcψ cφsθsψ − sφcψ   v  (12)
ḣ sθ −sφcθ −cφcθ w
     
u̇ rv − qw fx
 v̇  = pw − ru + 1 fy  , (13)
m
ẇ qu − pv fz
    
φ̇ 1 sin φ tan θ cos φ tan θ p
 θ̇  = 0 cos φ − sin φ   q (14)
sin φ cos φ
ψ̇ 0 cos θ cos θ
r
   Jy −Jz   1 
ṗ Jx
qr τ
Jx φ
q̇  =  Jz −Jx 
 Jy pr  + Jy τθ  .
 1
(15)
ṙ J −J 1
x
J
y
pq τ
Jz ψ
z

15
3 Forces and Moments
The objective of this section is to describe the forces and torques that act on the
quadrotor. Since there are no aerodynamic lifting surfaces, we will assume that
the aerodynamic forces and moments are negligible.
The forces and moments are primarily due to gravity and the four propellers.

front

left right

back

Figure 12: The top view of the quadrotor. Each motor produces an upward force
F and a torque τ . The front and back motors spin clockwise and the right and left
motors spin counterclockwise.

Figure 13: Definition of the forces and torques acting on the quadrotor.

Figure 12 shows a top view of the quadrotor systems. As shown in Figure 13,
each motor produces a force F and a torque τ . The total force acting on the

16
quadrotor is given by
F = Ff + Fr + Fb + Fl .
The rolling torque is produced by the forces of the right and left motors as

τφ = `(Fl − Fr ).

Similarly, the pitching torque is produced by the forces of the front and back
motors as
τθ = `(Ff − Fb ).
Due to Newton’s third law, the drag of the propellers produces a yawing torque
on the body of the quadrotor. The direction of the torque will be in the oppositive
direction of the motion of the propeller. Therefore the total yawing torque is given
by
τψ = τr + τl − τf − τb .
The lift and drag produced by the propellers is proportional to the square of the
angular velocity. We will assume that the angular velocity is directly proportional
to the pulse width modulation commend sent to the motor. Therefore, the force
and torque of each motor can be expressed as

F∗ = k1 δ∗
τ∗ = k2 δ∗ ,

where k1 and k2 are constants that need to be determined experimentally, δ∗ is the

motor command signal, and ∗ represents f , r, b, and l.
Therefore, the forces and torques on the quadrotor can be written in matrix
form as
      
F k1 k1 k1 k1 δf δf
 τφ   0 −`k1 0 `k1   δr  4  δr 
 =   = M .
 τθ   `k1 0 `k1 0   δb   δb 
τψ −k2 k2 −k2 k2 δl δl
The control strategies derived in subsequent sections will specify forces and
torques. The actual motors commands can be found as
   
δf F
 δr   
  = M−1  τφ  .
 δb   τθ 
δl τψ

17
Note that the pulse width modulation commands are required to be between zero
and one.
In addition to the force exerted by the motor, gravity also exerts a force on the
quadrotor. In the vehicle frame F v , the gravity force acting on the center of mass
is given by  
0
fgv =  0  .
mg
However, since v in Equation (13) is expressed in F b , we must transform to the
body frame to give
 
0
b
b
fg = Rv 0 
mg
 
−mg sin θ
=  mg cos θ sin φ  .
mg cos θ cos φ
Therefore, equations (12)–(15) become
    
ṗn cθcψ sφsθcψ − cφsψ cφsθcψ + sφsψ u
 ṗe  = cθsψ sφsθsψ + cφcψ cφsθsψ − sφcψ   v  (16)
ḣ sθ −sφcθ −cφcθ w
       
u̇ rv − qw −g sin θ 0
 v̇  = pw − ru +  g cos θ sin φ  + 1  0  , (17)
m
ẇ qu − pv g cos θ cos φ −F
    
φ̇ 1 sin φ tan θ cos φ tan θ p
 θ̇  = 0 cos φ − sin φ   q , (18)
sin φ cos φ
ψ̇ 0 cos θ cos θ
r
   Jy −Jz   1 
ṗ Jx
qr τ
Jx φ
q̇  =  Jz −Jx 
 Jy pr  + Jy τθ  .
 1
(19)
ṙ Jx −Jy 1
J
pq τ
Jz ψ
z

4 Simplified Models
Equations (16)–(19) are the equations of motion to be used in our six degree-
of-freedom simulator. However, they are not appropriate for control design for

18
several reasons. The first reason is that they are too complicated to gain significant
insight into the motion. The second reason is that the position and orientation
are relative to the inertial world fixed frame, whereas camera measurements will
measure position and orientation of the target with respect to the camera frame.

4.1 Model for estimation

For the quadrotor, we are not able to estimate the inertial position or the heading
angle ψ. Rather, we will be interested in the relative position and heading of the
quadrotor with respect to a ground target. The relative position of the quadrotor
will be measured in the vehicle-1 frame, i.e., the vehicle frame after it has been
rotated by the heading vector ψ. The vehicle-1 frame is convenient since x, y,
and z positions are still measured relative to a flat earth, but they are vehicle
centered quantities as opposed to inertial quantities. Let px , py , and pz denote
the relative position vector between the target and the vehicle resolved in the v1
frame. Therefore Eq (16) becomes
    
ṗx cθ sφsθ cφsθ u
ṗy  =  0 cφ −sφ   v. (20)
ṗz −sθ sφcθ cφcθ w

4.2 Model for control design

Assuming that φ and θ are small, Equation (18) can be simplified as
   
φ̇ p
 θ̇  = q  . (21)
ψ̇ r
Similarly, Equation (19) is simplified by assuming that the Coriolis terms qr, pr,
and pq, are small to obtain
  1 
ṗ Jx φ
τ
q̇  =  J τθ  .
1
(22)
y
ṙ 1
Jz ψ
τ
Combining Eq. (21) and (22) we get
  1 
φ̈ Jx φ
τ
 θ̈  =  J1 τθ  . (23)
y
1
ψ̈ Jz ψ
τ

19
 Differentiating Eq. (16) and neglecting Ṙbv gives
    
p̈n cθcψ sφsθcψ − cφsψ cφsθcψ + sφsψ u̇
 p̈e  = cθsψ sφsθsψ + cφcψ cφsθsψ − sφcψ   v̇  . (24)
p̈d −sθ sφcθ cφcθ ẇ

Neglecting the Coriolis terms and plugging Eq. (17) into Eq. (24) and simplifying
gives      
p̈n 0 −cφsθcψ − sφsψ
 p̈e  = 0 + −cφsθsψ + sφcψ  F . (25)
m
p̈d g −cφcθ
Therefore, the simplified inertial model is given by
F
p̈n = (− cos φ sin θ cos ψ − sin φ sin ψ) (26)
m
F
p̈e = (− cos φ sin θ sin ψ + sin φ cos ψ) (27)
m
F
p̈d = g − (cos φ cos θ) (28)
m
1
φ̈ = τφ (29)
Jx
1
θ̈ = τθ (30)
Jy
1
ψ̈ = τψ . (31)
Jz
(32)

The dynamics given in Equations (26)–(31) are expressed in the inertial frame.
This is necessary for the simulator. However, we will be controlling position,
altitude, and heading using camera frame measurements of a target position. In
this setting heading is irrelevant. Therefore, instead of expressing the translational
dynamics in the inertial frame, we will express them in the vehicle-1 frame F v1 ,
which is equivalent to the inertial frame after rotating by the heading angle.
Differentiating Eq. (20) and neglecting Ṙbv1 gives
    
p̈x cθ sφsθ cφsθ u̇
p̈y  =  0 cφ −sφ   v̇  . (33)
p̈z −sθ sφcθ cφcθ ẇ

20
Neglecting the Coriolis terms and plugging Eq. (17) into Eq. (33) and simplifying
gives      
p̈x 0 −cφsθ
p̈y  = 0 +  sφ  F . (34)
m
p̈z g −cφcθ
Therefore, the simplified model in the vehicle-1 frame is given by
F
p̈x = − cos φ sin θ (35)
m
F
p̈y = sin φ (36)
m
F
p̈z = g − cos φ cos θ (37)
m
1
φ̈ = τφ (38)
Jx
1
θ̈ = τθ (39)
Jy
1
ψ̈ = τψ . (40)
Jz
(41)

5 Sensors
5.1 Rate Gyros
A MEMS rate gyro contains a small vibrating lever. When the lever undergoes
an angular rotation, Coriolis effects change the frequency of the vibration, thus
detecting the rotation. A brief description of the physics of rate gyros can be
found at RateSensorAppNote.pdf.
The output of the rate gyro is given by

ygyro = kgyro Ω + βgyro + ηgyro ,

where ygyro is in Volts, kgyro is a gain, Ω is the angular rate in radians per second,
βgyro is a bias term, and ηgyro is zero mean white noise. The gain kgyro should be
given on the spec sheet of the sensor. However, due to variations in manufacturing
it is imprecisely known. The bias term βgyro is strongly dependent on temperature
and should be calibrated prior to each flight.

21
 If three rate gyros are aligned along the x, y, and z axes of the quadrotor, then
the rate gyros measure the angular body rates p, q, and r as follows:
ygyro,x = kgyro,x p + βgyro,x + ηgyro,x
ygyro,y = kgyro,y q + βgyro,y + ηgyro,y
ygyro,z = kgyro,z r + βgyro,z + ηgyro,z .
MEMS gyros are analog devices that are sampled by the on-board processer.
We will assume that the sample rate is given by Ts . The Kestrel autopilot samples
the gyros at approximately 120 Hz.

5.2 Accelerometers
A MEMS accelerometer contains a small plate attached to torsion levers. The
plate rotates under acceleration and changes the capacitance between the plate
and the surrounding walls [3].
The output of the accelerometers is given by
yacc = kacc A + βacc + ηacc ,
where yacc is in Volts, kacc is a gain, A is the acceleration in meters per second, βacc
is a bias term, and ηacc is zero mean white noise. The gain kacc should be given
on the spec sheet of the sensor. However, due to variations in manufacturing it is
imprecisely known. The bias term βacc is strongly dependent on temperature and
should be calibrated prior to each flight.
Accelerometers measure the specific force in the body frame of the vehicle. A
physically intuitive explanation is given in [1, p. 13-15]. Additional explanation
is given in [4, p. 27]. Mathematically we have
 
ax
ay  = 1 (F − Fgravity )
m
az
1
= v̇ + ω bb/i × v − Fgravity .
m
In component form we have
ax = u̇ + qw − rv + g sin θ
ay = v̇ + ru − pw − g cos θ sin φ
az = ẇ + pv − qu − g cos θ cos φ.

22
Using Eq (17) we get

ax = 0
ay = 0
az = −F/m,

where F is the total thrust produced by the four motors. It is important to note
that for the quadrotor, the output of the accelerometers is independent of angle.
This is in contrast to the unaccelerated flight for fixed wing vehicles where the
accelerometers are used to measure the gravity vector and thereby extract roll and
pitch angles.
MEMS accelerometers are analog devices that are sampled by the on-board
processer. We will assume that the sample rate is given by Ts . The Kestrel autopi-
lot samples the accelerometers at approximately 120 Hz.

5.3 Camera
The control objective is to hold the position of the quadrotor over a ground based
target that is detected using the vision sensor. In this section we will briefly de-
scribe how to estimate px and py in the vehicle 1-frame.
We will assume that the camera is mounted so that the optical axis of the
camera is aligned with the body frame z-axis and so that the x-axis of the camera
points out the right of the quadrotor and the y-axis of the camera points to the
back of the quadrotor.
The camera model is shown in Figure 14. The position of the target in the
vehicle-1 frame is (px , py , pz ). The pixel location of the target in the image is
(²x , ²y ).
The geometry for py is shown in Figure 15. From the geometry shown in
Figure 15, we can see that
µ ¶
η
py = pz tan φ − ²x , (42)
My
where η is the camera field-of-view, and My is the number of pixels along the
camera y-axis. In Figure 15, both py and ²x are negative. Positive values are
toward the right rotor. A similar equation can be derived for px as
µ ¶
η
px = −pz tan θ − ²y . (43)
My

23
 Figure 14: Camera model for the quadrotor.

6 State Estimation
The objective of this section is to describe techniques for estimating the state of the
quadrotor from sensor measurements. We need to estimate the following states:
px , py , pz , u, v, w, φ, θ, ψ, p, q, r.
The angular rates p, q, and r can be obtained by low pass filtering the rate
gyros. The remain states require a Kalman filter. Both are discussed below.

6.1 Low Pass Filters

The Laplace transforms representation of a simple low-pass filter is given by
a
Y (s) = U (s),
s+a

24
 Optical
Axis

Target

Figure 15: The geometry introduced by the vision system. The height above
ground is given by −pz , the lateral position error is py , the roll angle is φ, the
field-of-view of the camera is η, the lateral pixel location of the target in the image
is ²x , and the total number of pixels along the lateral axis of the camera is Mx .

were u(t) is the input of the filter and y(t) is the output. Inverse Laplace trans-
forming we get
ẏ = −ay + au. (44)
Using a zeroth order approximation of the derivative we get
y(t + T ) − y(t)
= −ay(t) + au(t),
T
where T is the sample rate. Solving for y(t + T ) we get
y(t + T ) = (1 − aT )y(t) + aT u(t).
For the zeroth order approximation to be valid we need aT ¿ 1. If we let α = aT
then we get the simple form
y(t + T ) = (1 − α)y(t) + αu(t).

25
Note that this equation has a nice physical interpretation: the new value of y
(filtered value) is a weighted average of the old value of y and u (unfiltered value).
If u is noisy, then α ∈ [0, 1] should be set to a small value. However, if u is
relatively noise free, then α should be close to unity.
In the derivation of the discrete-time implementation of the low-pass filter, it is
possible to be more precise. In particular, returning to Equation (44), from linear
systems theory, it is well known that the solution is given by
Z T
−aT
y(t + T ) = e y(t) + a e−a(T −τ ) u(τ ) dτ.
0

Assuming that u(t) is constant between sample periods results in the expression
Z T
−aT
y(t + T ) = e y(t) + a e−a(T −τ ) dτ u(t)y(t + T )
0
= e−aT y(t) + (1 − e−aT )u(t). (45)
2
Note that since ex = 1 + x + x2! + . . . we have that e−aT ≈ 1 − aT and 1 − e−aT ≈
aT .
Therefore, if the sample rate is not fixed (common for micro controllers), and
it is desired to have a fixed cut-off frequency, then Equation (45) is the preferable
way to implement a low-pass filter in digital hardware.
We will use the notation LP F (·) to represent the low-pass filter operator.
Therefore x̂ = LP F (x) is the low-pass filtered version of x.

6.2 Angular Rates p, q, and r.

The angular rates p, q, and r can be estimated by low-pass filtering the rate gyro
signals:

p̂ = LP F (ygyro,x ) (46)
q̂ = LP F (ygyro,y ) (47)
r̂ = LP F (ygyro,z ). (48)

6.3 Dynamic Observer Theory

The objective of this section is to briefly review observer theory.

26
 Suppose that we have a linear time-invariant system modeled by the equations
ẋ = Ax + Bu
y = Cx.
A continuous-time observer for this system is given by the equation
x̂˙ = |Ax̂ {z
+ Bu} + L (y − C x̂), (49)
| {z }
copy of the model correction due to sensor reading
where x̂ is the estimated value of x. Letting x̃ = x − x̂ we observer that
x̃˙ = (A − LC)x̃
which implies that the observation error decays exponentially to zero if L is cho-
sen such that the eigenvalues of A − LC are in the open left half of the complex
plane.
In practice, the sensors are usually sampled and processed in digital hardware
at some sample rate Ts . How do we modify the observer equation shown in Equa-
tion (49) to account for sampled sensor readings?
The typical approach is to propagate the system model between samples using
the equation
x̂˙ = Ax̂ + Bu (50)
and then to update the estimate when a measurement is received using the equation
x̂+ = x̂− + L(y(tk ) − C x̂− ),
where tk is the instant in time that the measurement is received and x̂− is the
state estimate produced by Equation (50) at time tk . Equation (50) is then re-
instantiated with initial conditions given by x̂+ . The continuous-discrete observer
is summarized in Table 6.3.
The observation process is shown graphically in Figure 16. Note that it is not
necessary to have a fixed sample rate. The continuous-discrete observer can be
implemented using Algorithm 1.
Note that we did not use the fact that the process was linear. Suppose instead
that we have a nonlinear system of the form
ẋ = f (x, u) (51)
y = c(x) (52)
, then the continuous discrete observer is given in table 6.3.
The real question is how to pick the observer gain L.

27
 Figure 16: sdf

Algorithm 1 Continuous-Discrete Observer

1: Initialize: x̂ = 0.
2: Pick an output sample rate Tout which is much less than the sample rates of
the sensors.
3: At each sample time Tout :
4: for i = 1 to¡N do ¢ {Prediction: Propagate the state equation.}
Tout
5: x̂ = x̂ + N (Ax̂ + Bu)
6: end for
7: if A measurement has been received from sensor i then {Correction: Mea-
surement Update}
8: x̂ = x̂ + Li (yi − Ci x̂)
9: end if

28
 System model:
ẋ = Ax + Bu
y(tk ) = Cx(tk )
Initial Condition x(0).
Assumptions:
Knowledge of A, B, C, u(t).
No measurement noise.
Prediction: In between measurements (t ∈ [tk−1 , tk )):
Propagate x̂˙ = Ax̂ + Bu.
Initial condition is x̂+ (tk−1 ).
Label the estimate at time tk as x̂− (tk ).
Correction: At sensor measurement (t = tk ):
x̂+ (tk ) = x̂− (tk ) + L (y(tk ) − C x̂− (tk )) .

Table 1: Continuous-Discrete Linear Observer.

6.4 Essentials from Probability Theory

Let X = (x1 , . . . , xn )T be a random vector whose elements are random variables.
The mean, or expected value of X is denoted by
   
µ1 E{x1 }
   
µ =  ...  =  ...  = E{X},
µn E{xn }
where Z
E{xi } = ξfi (ξ) dξ,

and f (·) is the probability density function for xi . Given any pair of components
xi and xj of X, we denote their covariance as
cov(xi , xj ) = Σij = E{(xi − µi )(xj − µj )}.
The covariance of any component with itself is the variance, i.e.,
var(xi ) = cov(xi , xi ) = Σii = E{(xi − µi )(ξ − µi )}.
The standard deviation of xi is the square root of the variance:
p
stdev(xi ) = σi = Σii .

29
 System model:
ẋ = f (x, u)
y(tk ) = c(x(tk ))
Initial Condition x(0).
Assumptions:
Knowledge of f , c, u(t).
No measurement noise.
Prediction: In between measurements (t ∈ [tk−1 , tk )):
Propagate x̂˙ = f (x̂, u).
Initial condition is x̂+ (tk−1 ).
Label the estimate at time tk as x̂− (tk ).
Correction: At sensor measurement (t = tk ):
x̂+ (tk ) = x̂− (tk ) + L (y(tk ) − c(x̂− (tk ))) .

Table 2: Continuous-Discrete Nonlinear Observer.

The covariances associated with a random vector X can be grouped into a matrix
known as the covariance matrix:
 
Σ11 Σ12 · · · Σ1n
 Σ21 Σ22 · · · Σ2n 
  T T T
Σ =  .. . . ..  = E{(X − µ)(X − µ) } = E{XX } − µµ .
 . . . 
Σn1 Σn2 · · · Σnn

Note that Σ = ΣT so that Σ is both symmetric and positive semi-definite, which

implies that its eigenvalues are real and nonnegative.
The probability density function for a Gaussian random variable is given by
(x−µx )2
1 −
fx (x) = √ e σx2 ,
2πσx
where µx is the mean of x and σx is the standard deviation. The vector equivalent
is given by
· ¸
1 1 T −1
fX (X) = √ exp − (X − µ) Σ (X − µ) ,
2π det Σ 2
in which case we write
X ∼ N (µ, Σ) ,

30
 Figure 17: Level curves for the pdf of a 2D Gaussian random variable.

and say that X is normally distributed with mean µ and covariance Σ.

Figure 17 shows the level curves for a 2D Gaussian random variable with
different covariance matrices.

6.5 Derivation of the Kalman Filter

In this section we assume the following state model:

ẋ = Ax + Bu + Gξ
yk = Cxk + ηk ,

where yk = y(tk ) is the k th sample of y, xk = x(tk ) is the k th sample of x, ηk is

the measurement noise at time tk , ξ is a zero-mean Gaussian random process with
covariance Q, and ηk is a zero-mean Gaussian random variable with covariance
R. Note that the sample rate does not need to be be fixed.
The observer will therefore have the following form:

31
 Prediction: In between measurements: (t ∈ [tk−1 , tk ])
Propagate x̂˙ = Ax̂ + Bu.
Initial condition is x̂+ (tk−1 ).
Label the estimate at time tk as x̂− (tk ).
Correction: At sensor measurement (t = tk ):
x̂+ (tk ) = x̂− (tk ) + L (y(tk ) − C x̂− (tk )) .
Our objective is to pick L to minimize tr(P (t)).

6.5.1 Between Measurements.

Differentiating x̃ we get

x̃˙ = ẋ − x̂˙
= Ax + Bu + Gξ − Ax̂ − Bu
= Ax̃ + Gξ.

Therefore we have that

Z t
At
x̃(t) = e x̃0 + eA(t−τ ) Gξ(τ ) dτ.
0

We can therefore compute the evolution for P as

d
Ṗ = E{x̃x̃T }
dt
˙ T + x̃x̃˙ T }
= E{x̃x̃
© ª
= E Ax̃x̃T + Gξ x̃T + x̃x̃T AT + x̃ξ T GT
= AP + P AT + GE{ξ x̃T }T + E{x̃ξ T }GT .

As in the previous section we get

½ Z t ¾
T AT t T T AT (t−τ )
E{ξ x̃ } = E ξ(t)x̃0 e + ξ(t)ξ (τ )G e dτ
0
1
= QGT ,
2
which implies that
Ṗ = AP + P AT + GQGT .

32
6.5.2 At Measurements.
At a measurement we have that
x̃+ = x − x̂+
¡ ¢
= x − x̂− − L Cx + η − C x̂−
= x̃− − LC x̃− − Lη.
Therefore
P + = E{x̃+ x̃+T }
n¡ ¢¡ ¢T o
= E x̃− − LC x̃− − Lη x̃− − LC x̃− − Lη
©
= E x̃− x̃−T − x̃− x̃−T C T LT − x̃− η T LT
− LC x̃− x̃−T + LC x̃− x̃−T C T LT + LC x̃− η T LT
ª
= −Lηx̃−T + Lηx̃−T C T LT + Lηη T LT
= P − − P − C T LT − LCP − + LCP − C T LT + LRLT . (53)
Our objective is to pick L to minimize tr(P + ). A necessary condition is
∂
tr(P + ) = −P − C T − P − C T + 2LCP − C T + 2LR = 0
∂L
=⇒ 2L(R + CP − C T ) = 2P − C T
=⇒ L = P − C T (R + CP − C T )−1 .
Plugging back into Equation (53) give
P + = P − + P − C T (R + CP − C T )−1 CP − − P − C T (R + CP − C T )−1 CP −
+ P − C T (R + CP − C T )−1 (CP − C T + R)(R + CP − C T )−1 CP −
= P − − P − C T (R + CP − C T )−1 CP −
= (I − P − C T (R + CP − C T )−1 C)P −
= (I − LC)P − .
For linear systems, the continuous-discrete Kalman filter is summarized in
Table 6.5.2
If the system is nonlinear, then the Kalman filter can still be applied but we
need to linearize the nonlinear equations in order to compute the error covariance
matrix P and the Kalman gain L. The extended Kalman filter (EKF) is given in
Table 6.5.2, and an algorithm to implement the EKF is given in Algorithm 2.

33
 System model:
ẋ = Ax + Bu + ξ
yi (tk ) = Ci x(tk ) + ηk
Initial Condition x(0).
Assumptions:
Knowledge of A, B, Ci , u(t).
Process noise satisfies ξ ∼ N (0, Q).
Measurement noise satisfies ηk ∼ N (0, R).
Prediction: In between measurements (t ∈ [tk−1 , tk )):
Propagate x̂˙ = Ax̂ + Bu.
Propagate Ṗ = AP + P AT + Q.

Correction: At the ith sensor measurement (t = tk ):

Li = P − CiT (Ri + Ci P CiT )−1 ,
P + = (I − Li Ci )P − ,
x̂+ (tk ) = x̂− (tk ) + L (y(tk ) − C x̂− (tk )) .

Table 3: Continuous-Discrete Kalman Filter.

6.6 Application to the quadrotor

In this section we will discuss the application of Algorithm 2 to the quadrotor.
We would like to estimate the state
 
p̂x
p̂y 
 
p̂z 
 
p̂˙x 
 
 
x̂ = p̂˙y  ,
˙ 
p̂z 
 
 φ̂ 
 
 θ̂ 
ψ̂

where the rate gyros and accelerometers will be used to drive the prediction step,
and an ultrasonic altimeter and camera will be used in the correction step.

34
 System model:
ẋ = f (x, u) + ξ
yi (tk ) = ci (x(tk )) + ηk
Initial Condition x(0).
Assumptions:
Knowledge of f , ci , u(t).
Process noise satisfies ξ ∼ N (0, Q).
Measurement noise satisfies ηk ∼ N (0, R).
Prediction: In between measurements (t ∈ [tk−1 , tk )):
Propagate x̂˙ = f (x̂,¯ u),
∂f ¯
Compute A = ∂x x=x̂(t) ,
Propagate Ṗ = AP + P AT + Q.

th
Correction: At ¯ the i sensor measurement (t = tk ):
∂ci ¯
Ci = ∂x x=x̂−
Li = P − CiT (Ri + Ci P CiT )−1 ,
P + = (I − Li Ci )P − ,
x̂+ (tk ) = x̂− (tk ) + L (y(tk ) − ci (x̂− (tk ))) .

Table 4: Continuous-Discrete Extended Kalman Filter.

The propagation model is obtained from Equations (35)–(37), and (18) as

 
ṗx
 ṗy 
 
 ṗz 
 
 cos φ sin θa 
 z 
f (x, u) = 
 −sinφa z
,

 g + cos φ cos θaz 
 
p + q sin φ tan θ + r cos φ tan θ
 
 q cos φ − r sin φ 
sin φ
q cos θ
+ r cos φ
cos θ

where we have used the fact that the z-axis of the accelerometer measures az =

35
Algorithm 2 Continuous-Discrete Extended Kalman Filter
1: Initialize: x̂ = 0.
2: Pick an output sample rate Tout which is much less than the sample rates of
the sensors.
3: At each sample time Tout :
4: for i = 1 to¡N do ¢ {Prediction: Propagate the equations.}
Tout
5: x̂ = x̂ + N (f (x̂, u))
6: A = ∂f∂x ¡ out ¢ ¡ ¢
7: P = P + TN AP + P AT + GQGT
8: end for
9: if A measurement has been received from sensor i then {Correction: Mea-
surement Update}
10: Ci = ∂c∂x
i

11: Li = P CiT (Ri + Ci P CiT )−1

12: P = (I − Li Ci )P
13: x̂ = x̂ + Li (yi − ci (x̂)).
14: end if

−F/m. Differentiating we obtain

 
0 0 0 1 0 0 0 0 0
0 0 0 0 1 0 0 0 0
 
0 0 0 0 0 1 0 0 0
 
0 0 0 0 0 0 −sφ sθ az cφ cθ az 0
 
A=  0 0 0 0 0 0 −c a
φ z 0 0
0 0 0 0 0 0 −sφ cθ az −cφ sθ az 0
 
0 0 0 0 0 0 qcφtθ − rsφtθ (qsφ + rcφ)/c2 θ 0
 
0 0 0 0 0 0 −qsφ − rcφ 0 0
qcφ−rsφ tθ
0 0 0 0 0 0 cθ
−(qsφ + rcφ) cθ 0

Note that it may be adequate (not sure) to use a small angle approximation in

36
the model resulting in  
ṗx
 ṗy 
 
 ṗz 
 
 θaz 
 
f (x, u) = 
 −φa z
,

g + az 
 
 p 
 
 q 
r
and  
0 0 0 1 0 0 0 0 0
0 0 0 0 1 0 0 0 0
 
0 0 0 0 0 1 0 0 0
 
0 0 0 0 0 0 0 az 0
 
A=
0 0 0 0 0 0 −az 0 0.
0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
If this form works, then the update equation for P can be coded by hand, signif-
icantly reducing the computational burden. Note also that f (x, u) does not take
into account the motion of the target. A feedforward term can be added to f to
account for the target motion as
 
ṗx − ṁx
 ṗy − ṁy 
 
 ṗ 
 z 
 θaz − m̈x 
 
f (x, u) = 
 −φa z − m̈ y
,

 g + az 
 
 p 
 
 q 
r

where (ṁx , ṁy ) is the velocity of the target in the vehicle-1 frame, and (m̈x , m̈y )
is the acceleration of the target in the vehicle-1 frame.

37
 Let the relative position between the quadrotor and the target be denoted by
p = (px , py , pz )T . We can tranform to the camera frame as
v1

pc = Rgc Rbg Rv1

b
pv1 ,
where  
0 1 0
Rgc = 0 0 1
1 0 0
is the rotation matrix from gimbal coordinates to camera coordinates,
 
0 0 1
Rbg =  0 1 0
−1 0 0
transforms body coordinates to gimbal coordinates, and
 
cθcψ sφsθcψ − cφsψ cφsθcψ + sφsψ
b
Rv1 = cθsψ sφsθsψ + cφcψ cφsθsψ − sφcψ 
−sθ sφcθ cφcθ
transforms vehicle-1 coordinates to body coordinates. Using the small angle ap-
proximation for φ and θ we get
 
φθpx + py + φpz
pc =  −px + θpz  .
θpx − φpy + pz
The model for the pixel coordinates are therefore computed as
4 φθpx + py + φpz
²x = c²x (x) = f
θpx − φpy + pz
4 −px + θpz
²2 = c²y (x) = f .
θpx − φpy + pz
We therefore have the following sensors available to correct the state estimate:
4
sensor 1: altimeter y1 = calt (x) = −pz + η1 [k] (54)
sensor 2: vision x-pixel y2 = c²x (x) + η2 [k]. (55)
sensor 3: vision y-pixel y3 = c²y (x) + η3 [k] (56)
4
sensor 4: vision heading y4 = c²ψ (x) = π/2 + ψ + η4 [k]. (57)

38
 The linearization of the first and fourth output functions are given by
¡ ¢
C1 = 0 0 −1 0 0 0 0 0 (58)
¡ ¢
C4 = 0 0 0 0 0 0 0 1 . (59)

The linearization of the expression for the pixel coordinates is messy and can
easily be computed numerically using the approximation
.
∂f (x1 , · · · , xi , . . , xn ) f (x1 , · · · , xi + ∆, · · · , xn ) − f (x1 , · · · , xi , · · · , xn )
≈ ,
∂xi ∆
where ∆ is a small constant.

7 Control Design
The control design will be derived directly from Equations (35)–(40). Equa-
tions (38)–(40) are already linear. To simplify Equations (35)–(37) define

4 F
ux = −cosφ sin θ (60)
m
4 F
uy = sin φ (61)
m
4 F
uz = g − cos φ cos θ , (62)
m
to obtain

p̈x = ux (63)
p̈y = uy (64)
p̈z = uz . (65)

The control design proceeds by developing PID control strategies for ux , uy , and
uz . After ux , uy , and uz have been computed, we can compute the desired force
F , and the commanded roll angle φc and the commanded pitch angle θc from
Equations (60)–(62) as follows. From Equation (62) solve for F/m as

F g − uz
= . (66)
m cos φ cos θ

39
Substituting (66) into (61) gives
g − uz
uy = tan φ.
cos θ
Solving for φ and letting this be the commanded roll angle gives
µ ¶
c −1 uy cos θ
φ = tan . (67)
g − uz
Similarly, we can solve for the commanded pitch angle as
µ ¶
c −1 ux
θ = tan . (68)
uz − g

7.1 Vision Based Altitude Hold

For the altitude hold we need to develop an expression for uz to drive pz to a
desired altitude based on the size of the object in the image. We will assume that
the camera returns the size of the object in the image plane in pixels, which is
denoted by ²s . Figure 18 shows the geometry of the quadrotor hovering over a

Target

Figure 18: The size of the target is S in meters, and the size of the target in the
image plane is denoted by ²s in pixels. The focal length is f , and the height above
the target is −pz .

target of size S. From similar triangles we have that

−pz f
= .
S ²s

40
Solving for pz an differentiating we obtain
f S ²̇s
ṗz = . (69)
²2s
Differentiating again gives
f S²̈s ²̇2s
p̈z = 2 − 2f S 3 .
²s ²s
Substituting uz = p̈z and solving for ²̈s gives
µ 2¶
²s ²̇2
²̈s = uz + 2 s .
fS ²s
Defining µ ¶
4 ²2s ²̇2s
us = uz + 2 , (70)
fS ²s
we get
²̈s = us .
We can now derive a PID controller to drive ²s → ²ds as
Z t
d
us = kps (²s − ²s ) − kds ²̇s + kis (²ds − ²s ) dτ.
0

Solving (70) for uz we get

µ ¶ Z
fS d fS f S ²̇s fS t d
uz = 2 kps (²s − ²s ) − kds 2 + 2 3 ²̇s + kis 2 (² − ²s ) dτ. (71)
²s ²s ²s ²s 0 s
The downside to this equation is that it requires knowledge of the target size S and
the focal length f . This requirement can be removed by incorporating f S into the
PID gains by defining
4
k̂ps = f Skps
4
k̂is = f Skis
4
k̂ds = f Skds ,
and by noting from Equation (69) that f S²2²̇s = w. Therefore Equation (71) be-
s
comes
à ! Z
(²ds − ²s ) k̂ds w k̂is t d
uz = k̂ps − +2 ²̇s + 2 (² − ²s ) dτ. (72)
²2s ²2s ²s ²s 0 s

41
7.2 Roll Attitude Hold
The equation of motion for roll is given by Eq. (38) as φ̈ = τp /Jx . We will use a
PID controller to regulate the roll attitude as
Z t
c
τp = kpφ (φ − φ) − kdφ p + kiφ (φc − φ) dτ.
0

A block diagram of the control structure is shown in Figure 19.

Figure 19: A block diagram of the roll attitude hold loop.

The gains kpφ , kiφ , and kdφ are selected one at a time using a method called
successive loop closure. To pick kpφ note that if the input command φc is a step of
value A, then at time t = 0, before the integrator has time to begin winding up, τx
is given by
τx (0) = kpφ A.
Therefore, setting kpφ = M/A will just saturate the actuators when a step value
of A is placed on φc .
To select kdφ , let kpφ be fixed and let kiφ = 0, and solve for the characteristic
equation in Evan’s form verses kdφ to obtain
1
Jx
s
1 + kdφ kpφ
= 0.
s2 + Jx

The derivative gain kdφ is selected to achieve a damping ratio of ζ = 0.9.

The characteristic equation including kiφ in Evan’s form verses kiφ is given by
1
Jx
1 + kiφ kdφ kpφ
= 0.
s3 + Jx
s2 + Jx
s

42
The integral gain kiφ can be selected so that damping ratio is not significantly
changed.

7.3 Pitch Attitude Hold

The equation of motion for pitch is given by Eq. (39) as θ̈ = τq /Jy . Similar to the
roll attitude hold, we will use a PID controller to regulate pitch as
Z t
c
τq = kpθ (θ − θ) − kdθ q + kiθ (θc − θ) dτ.
0

7.4 Vision Based Position Tracking

From Equation 64 the lateral dynamics are given by p̈y = uy , where py is the
relative lateral position which we drive to zero using the PID strategy
Z t
uy = −kpy py − kdy v + kiy py dτ.
0

The relative position error py is given by Equation (42).

From Equation 63 the longitudinal dynamics are given by p̈x = ux , where px
is the relative lateral position which we drive to zero using the PID strategy
Z t
ux = −kpx px − kdx u + kix px dτ.
0

The relative position error px is given by Equation (43).

7.5 Relative Heading Hold

The heading dynamics are given in Equation (40) as ψ̈ = τr /Jz . Define ψ d as the
4
inertial heading of the target, and define ψ̃ = ψ − ψ d as the relative heading. The
camera directly measures ψ̃. Assuming that ψ d is constant we get ψ̃¨ = τr /Jz . We
regulate the relative heading with the PID strategy
Z t
τr = −kpψ ψ − kdψ r − kiψ ψ dτ.
0

43
7.6 Feedforward
When the quadrotor is tracking a ground robot, the motion of the robot will cause
tracking errors due to the delayed response of the PID controller. If we can com-
municate with the robot, and we know its intended motion, we should be able to
use that information to help the quadrotor predict where the robot is moving.
To motivate our approach, lets consider a simple example to gain intuition.
Consider the double integrator system

ÿ = u,

where y is position, and u is commanded acceleration. If r is the reference signal

then the standard PD loop is shown in Figure 20.

Figure 20: Standard PD loop

Let e = r − y, then

ë = r̈ − ÿ
= r̈ − u
= r̈ − kp e + kd ẏ
= r̈ − kp e − kd ė + kd ṙ.

In other words we have

ë + kd ė + kp e = r̈ + kd ṙ.
1
Therefore, the signal r̈ + kd ṙ drives the error transfer function s2 +kd s+kp
and if w
is not zero, the tracking error will not go to zero.

44
 From the expression ë = r̈ − u we see that if instead of u = kp e − kd ẏ, we
instead use
u = r̈ + kp e + kd ė,
then we get
ë + kd ė + kp e = 0,
which ensures that e(t) → 0 independent of r(t). The associated block diagram
is shown in Figure 21.

Figure 21: Feedforward term is added to a standard PD loop.

Now consider the lateral motion of the quadrotor in the cage where

ÿ = g tan φ,

where y is the position of the quadrotor, g is the gravity constant, and φ is the roll
4
angle. Let m(t) be the position of the target and define y = m − r as the relative
position. Then

ÿ = m̈ − r̈
= m̈ − g tan φ.

If by some means we know the target velocity ṁ and the target acceleration m̈,
then we should pick φ such that

g tan φ = m̈ + kp y + kd ẏ,

or µ ¶
−1 m̈ + kp y + kd ẏ
φ = tan .
g

45
 Given the decoupling terms described at the beginning of this section, the
motion model for the quadrotor in the vehicle-1 frame is given by
p̈x = ux
p̈y = uy
ψ̈ = uψ .
Let the position of the target in the vehicle-1 frame be given by (mx , my ) and let
it orientation be given by ψt , and define
p̃x = mx − px
p̃y = my − py
ψ̃ = ψt − ψ.
Differentiating twice we get
p̃¨x = m̈x − ux
p̃¨y = m̈y − uy
ψ̃¨ = ψ̈ − u .
t ψ

Therefore, adding a feedforward term we have

ux = m̈x + P IDx
uy = m̈y + P IDy
uψ = ψ̈t + P IDψ
where P ID∗ are the PID controllers derived earlier in this chapter.
The motion model for the target is given by
ṁn = Vt cos ψt
ṁe = Vt sin ψt
V̇t = u1
ψ̇t = u2 ,
where (mn , me ) is the inertial position of the target, Vt is the ground speed of the
target, and u1 and u2 are robot control signals. In the vehicle-1 frame we have
µ ¶ µ ¶µ ¶ µ ¶
ṁx cos ψ − sin ψ Vt cos ψt Vt cos ψ̃
= = ,
ṁy sin ψ cos ψ Vt sin ψt Vt sin ψ̃

46
 View publication stats

where the camera measures the relative heading ψ̃ = ψ − ψt . Differentiating we

obtain µ ¶ µ ¶µ ¶
m̈x cos ψ̃ sin ψ̃ u1
= .
m̈y − sin ψ̃ cos ψ̃ u2 − ψ̇
To implement feedforward on the quadrotors, we must communicate with the
robot to obtain Vt , u1 , and u2 .

References
[1] B. L. Stevens and F. L. Lewis, Aircraft Control and Simulation. Hoboken, New Jersey:
John Wiley & Sons, Inc., 2nd ed., 2003.

[2] D. Halliday and R. Resnick, Fundamentals of Physics. John Wiley & Sons, 3rd ed.,
1988.

[3] http://www.silicondesigns.com/tech.html.

[4] M. Rauw, FDC 1.2 - A SIMULINK Toolbox for Flight Dynamics and Control Analy-
sis, February 1998. Available at http://www.mathworks.com/.

47