High Performance Autonomous Racing

with RC Cars
Designing models and controllers to implement a small scale
system of multiple autonomous cars
Bachelor’s thesis EENX15-18-23

Patrick Engström
Daniel Fredriksson
Oscar Olesen
David Olsson
Sanna Sandberg
Ahmed Hamdy Faramawy Mahmoud Shams El Din

Department of Electrical Engineering

C HALMERS U NIVERSITY OF T ECHNOLOGY
Gothenburg, Sweden 2018-05-14
 Bachelor thesis 2018:05:14

High Performance Autonomous Racing

with RC Cars

Designing models and controllers to implement a small scale

system of multiple autonomous cars

Patrick Engström
Daniel Fredriksson
Oscar Olesen
David Olsson
Sanna Sandberg
Ahmed Hamdy Faramawy Mahmoud Shams El Din

Department of Electrical Engineering

Division of Systems and Control
Chalmers University of Technology
Gothenburg, Sweden 2018-05-14
High Performance Autonomous Racing with RC Cars
Designing models and controllers to implement a small scale system of multiple au-
tonomous cars

Supervisor: Baláz Kulcár - Department of Electrical Engineering

Examiner: Torsten Wik - Department of Electrical Engineering
Co-examiner: Péter Kovács - Department of Electrical Engineering

© Patrick Engström, 2018 - pateng@[Link]

© Daniel Fredriksson, 2018 - danifred@[Link]
© Oscar Olesen, 2018 - oleseno@[Link]
© David Olsson, 2018 - daviols@[Link]
© Sanna Sandberg, 2018 - sannasa@[Link]
© Ahmed Hamdy Faramawy Mahmoud Shams El Din, 2018 - hamdy@[Link].

Cover: Picture of remote controlled car used in the project.

iv
Abstract

The field of autonomous cars has been developing rapidly in recent years. One
of the primary arguments in favour of autonomous cars is the potentially reduced
number of traffic accidents. There are several critical areas that must be thoroughly
researched and investigated. The most fundamental aspect of an autonomous car is
being able to follow the road while maintaining speed. This is the primary focus of
this thesis.

One of the problems with autonomous driving is to make the system react in ade-
quate time, so that control over the vehicle is preserved at high speed. Since racing
inherently demands the car to travel as fast as possible while staying on course and
avoiding obstacles, developing a system that allows this, although in a small scale,
is the purpose of this thesis. The methodology is to first mathematically model the
car using a single-track model. From this model it is possible to design controllers
for the steering angle and speed of the car. This is then simulated, and afterwards
realised indoors with small scale remote controlled cars, using a camera mounted in
ceiling to determine the position and direction of the car. Obstacle avoidance can
then be added, to let multiple cars avoid, and possibly race, each other.

This gives new insight into the difficulties and necessities of driving autonomous
vehicles at high speed. The data collecting and processing cause delays that simple
controllers without prediction cannot handle to satisfaction, and high accuracy in
the vehicle model is crucial to the performance of the controllers. For future work,
more focus on predictive control can therefore make even more contribution to this
area of research.

Keywords: Control, PID controller, Autonomous vehicle, RC Racing

v
Sammanfattning

Forskningsarbetet kring autonoma bilar har utvecklats mycket snabbt under senaste
år. Ett av de främsta argumenten för autonoma bilar är möjligheten att minska
olyckor i trafiken. Det finns flera kritiska områden som måste noggrannt forskas och
undersökas. De mest grundläggande aspekterna av en autonom bil är förmågan att
hålla sig på vägen och samtidigt hålla hastighet. Detta är rapportens fokus.

Ett problem med autonom transport är att utveckla system med tillräckligt korta
reaktionstider, så att fordonet kan kontrolleras även i höga hastigheter. Eftersom
motorsport oundvikligen kräver att bilarna kör så fort som möjligt, samtidigt som
de undviker hinder och håller sig på banan är den här rapportens syfte att skapa
ett småskaligt system som kan hantera detta. Metoden är att först utveckla en
matematisk modell, och utifrån denna model sedan designa regulatorer för styrvinkel
och hastighet. Avslutningsvis simuleras detta, för att sedan förverkligas inomhus
med småskaliga radiostyrda bilar genom att använda en kamera som monteras i
taket för att bestämma position och riktning på bilarna. Undanmanövrering kan då
utvecklas, så att flera bilar kan köra samtidigt utan att krocka.

Arbetet ger nya insikter om svårigheterna med självkörande fordon i höga hastigheter.
Att samla in data och sedan behandla den skapar fördröjningar som icke-prediktiva
regulatorer inte kan hantera tillräckligt väl, samt kräver fordonsmodellerna hög nog-
grannhet för regulatorernas prestanda. Senare arbeten bör därför fokusera på mer
avancerade regleringsstrategier, särskilt för hastigheten.

Keywords: Reglering, Självkörande fordon, radiostyrning, racing.

vii
 Contents

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Problem description and aims . . . . . . . . . . . . . . . . . . 4
1.2 Scope and delimitations . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Theory 7
2.1 Mathematical model of the car . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 The bicycle model . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Motor model . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.3 Tyre model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Linearising the model . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Controller design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.1 PID tuning with heuristic stability margins . . . . . . . . . . . 16

3 Methods 19
3.1 Autonomous system structure . . . . . . . . . . . . . . . . . . . . . . 20
3.1.1 Remote controlled cars . . . . . . . . . . . . . . . . . . . . . . 20
3.1.2 Remote controller . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.3 Camera . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1 Linearisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.2 Parameter identification . . . . . . . . . . . . . . . . . . . . . 23
3.3 Controller design and implementation . . . . . . . . . . . . . . . . . . 26
3.3.1 Steering control . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.2 Speed control . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Indoor positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4.1 Position, direction, and identification . . . . . . . . . . . . . . 32
3.4.2 Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.3 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Creating the track . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.5.1 Obstacle avoidance . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Results 37
4.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

ix
Contents

4.1.1 Simulation based validation of the linear model . . . . . . . . 37

4.1.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Positioning and computer performance . . . . . . . . . . . . . . . . . 44
4.3 Track performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.1 Steering control . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.2 Speed control . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3.3 Real-time implementation . . . . . . . . . . . . . . . . . . . . 48
4.3.4 Obstacle avoidance . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3.5 Areas of improvements and error sources . . . . . . . . . . . . 57

5 Conclusion 59

Bibliography 61

A Appendix 1 I
A.1 Matlab code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I
A.1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I
A.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . III
A.1.3 PID Generator loop . . . . . . . . . . . . . . . . . . . . . . . . III

B Appendix 2 V

x
 Nomenclature

• X - Global X position (m)

• Y - Global Y position (m)

• ψ - Car yaw angle. (rad)

• vx - Car longitudinal velocity (m/s)

• vy - Car lateral velocity (m/s)

• ω - Car angular velocity (rad/s)

• Frx - Car rear longitudinal force (N )

• Ff x - Car frontal longitudinal force (N )

• Fry - Car rear lateral force (N )

• Ff x - Car frontal lateral force (N )

• m - Car mass (kg)

• Iz - Car rotational inertia (kg · m2 )

• δ - Car steering angle (rad)

• D - Car control signal

• lr - Distance from c.g to rear wheel (m)

• lf - Distance from c.g to front wheel (m)

• αr - Rear slip angle (rad)

• αf - Frontal slip angle (rad)

xi
Contents

• Cd - Drag coefficient

• Kd - Derivative gain

• Ki - Integral gain

• Kp - Proportional gain

• Td - Derivative time constant

• Ti - Integral time constant

• τ - Time constant

• ϕm - Phase margin

• ωc - Cross-over frequency

• MS - The sensitivity functions maximal gain

• MT - The complementary sensitivity functions maximal gain

• r(t) - Reference signal

• u(t) - Control signal

• e(t) - Error value

• em (t) - Measured error value

• y(t) - System output

• ym (t) - Measured system output

xii
 1
Introduction

The area of autonomous cars has long been of interest to researchers[1]. Companies
such as Google, Volvo and Tesla are working to develop a truly autonomous car.
Progress towards a fully autonomous vehicle has been made, with some companies
test driving autonomous cars on public roads. Volvo is one example, where the
company has been testing its autonomous system in Gothenburg with ordinary
families [2]. If the automotive industry succeeds with developing fully autonomous
vehicles, traffic accidents could potentially be reduced, as human error is the cause
of most incidents. Traffic efficiency also could be improved, reducing travel time [3].

Despite great progress in autonomous car technology over the past few years, a fully
autonomous vehicle is yet to be introduced to the general public. This is due to
the limitations in the current technology. An example is Tesla’s most advanced
Autopilot, which has caused crashes. In Tesla’s defence however, the Autopilot was
never advertised as fully autonomous [4].

The main challenges to an autonomous car would be lane keeping, speed control and
obstacle avoidance. If a vehicle could handle these challenges at high speed, it is
believed that the same system would be able to operate at lower speeds with higher
stability. It is also reasonable to think that autonomous cars very seldom will drive
as fast as physically possible, but if it is safe at those speeds, it most certainly will
be safe in lower speeds. Therefore, it is relevant to investigate the behaviour of an
autonomous system at high speed.

Today, tests are performed with autonomous racing cars competing against each
other. Several new automobile features like improved aerodynamics and active sus-
pension, were introduced through motor sport [5]. It is therefore reasonable to
suggest that developments made in high speed autonomous racing could be applied
to the commercials car market, which would provide even safer autonomous cars
for consumers. Then crashes caused by human error could be reduced, making the
roads safer, and the relevance of high performance racing important.

Therefore, this bachelor thesis project is aiming towards creating a autonomous

small scale system, similar to a racing situation, where the cars try to drive as fast
as possible. Thus, developing an environment to learn and investigate the difficulties
regarding high speed autonomous racing.

1
1. Introduction

1.1 Background

It is only recently that the connections between autonomous racing and autonomous
vehicles in regular traffic have started to be investigated. As already established,
this kind of research aims to increase safety routines in autonomous vehicles by
learning from race car drivers and autonomous race cars.

Standford University has conducted tests with their autonomous vehicle in racing
conditions. The Lab Director of Stanford University’s dynamic design lab claims
that if they can push the autonomous car to its limits and still and maintain control,
that could help provide insight on how to improve safety in commercial autonomous
cars [6]. The Stanford University’s dynamic design lab has researched high-speed
performance of cars and has published works on different topics regarding problems
occurring when racing with both normal and autonomous cars. For example a
Ph.D. dissertation by [Link] at Stanford Dynamic design lab, regarding
high-speed performance vehicle control, notes that

”Many road accidents are caused by the inability of drivers to control

a vehicle at its friction limits, yet racecars drivers routinely operate a
vehicle at the limits of handling without losing control. If autonomous
vehicles or driver assistance systems had capabilities similar to those of
racecar drivers, many fatal accidents could be avoided” [7].

At the University of California, Berkeley, their Model Predictive Control (MPC)

lab does real-time implementations of constrained predictive model-based control
on vehicles. Among many MPC projects, they have also investigated problems
occurring when predictive control methods are applied to autonomous racing cars.
For instance, they simulate a nonlinear MPC, to improve lap time and increase speed
[8]. Drift control, occurring at high speeds, was also investigated[9]. Some results
from their real-sized autonomous cars doing extreme manoeuvres in icy conditions
are shown in videos on their website [10].

This project is inspired by the Optimal RC Autonomous Racing (ORCA) project

carried out at the Swiss Federal Institute of Technology in Zürich. That project
is aimed to demonstrate predictive control solutions working in real life using au-
tonomous Remote Controlled (RC) cars where the goal was to improve lap-times
while avoiding obstacles [11]. The problem presents many challenges regarding a
multitude of aspects in autonomous control of high speed objects and obstacle avoid-
ance capabilities. The results from these kinds of project are valuable for real life
applications such as autonomous cars.

Previous success with obstacle avoidance for RC cars has been made. For example
at Uppsala and Chalmers[12][13]. ORCA is yet the only project pushing the small
scale test to its maximum velocities using advanced control systems such as MPC.

2
 1. Introduction

1.1.1 Purpose

The purpose of this project is to develop a small scale system that allows high speed
autonomous driving and obstacle avoidance. Resembling the system architecture of
the ORCA project it will be implemented at an indoor track with two small RC
cars and an overhead camera tracking the cars. Therefore, the purpose is to
create a system architecture which can handle all necessary calculations
to autonomously drive the small RC cars as fast as possible. This system
will involve control theory, indoor positioning and car modifications.

This project will give an understanding of the problems and limitations that such a
system could have and to what extent it could be applied in modern day automotive
applications. This subject is prevalent in today’s society where the automotive
industry is racing to develop a truly autonomous car.

3
1. Introduction

1.1.2 Problem description and aims

The main challenge with the project would be developing a system able to handle
high speed performance. To achieve this, a similar system architecture to the one
used in the ORCA project was chosen, as shown in figure 1. The system consists of
two miniature RC cars fitted with IR-LEDs. A camera, overlooking the virtual track,
provides data, such as velocity and orientation, and sends it to the controller. Using
a predetermined virtual path, the controller utilises the provided data to control the
paths and velocities of the car. Developing this system involves the points below:

• Mathematical modelling of the small RC cars.

• Creating a virtual path.

• Controller design for path keeping and velocity maximisation.

• Indoor positioning with image analysis tools.

• Modifying the cars and their remote controllers to communicate with a com-
puter.

Figure 1: An overall description of a system architecture inspired by the ORCA

project.

4
 1. Introduction

Aims

The aim, as partially described in the purpose, is to develop the system described
as above, and drive the small RC cars as fast as possible around a virtual track,
while avoiding obstacles. This involves linearising a nonlinear vehicle model, and
constructing a linear control system, inspired by established areas of research within
controller design. Moreover an accurate tracking system needs to be developed to
provide the controller with required data. Finally a remote connection needs to be
established, to send controller outputs to the cars. Thus the goal of creating a small
scale learning environment for autonomous racing with cars is achieved.

1.2 Scope and delimitations

Since the small cars will be tested to their limits, high demands will be put on
making an accurate mathematical model of the car, and a precise indoor tracking
system. This will be put into priority. Since several vehicle models exists, a simple
bicycle vehicle model will be used, for its simplicity and ability to simulate the
dynamics with satisfying accuracy.

There are enough aspects regarding just the controller to make it a project on its own.
Due to time constraints the controller type will be limited to a Proportional, Integral
and Derivative gain (PID) controller, as opposed to model predictive control types.
The robustness of the controller will also be considered of lesser importance. The
reason for this is that high-speed performance is a strong primary focus, which means
the system will operate close to stability failure, particularly within the testing
environment.

Complexity and versatility in obstacle avoidance will be limited. The project aims to
deliver a system that can avoid static obstacles at high speed. Since previous project
has been conducted when RC cars avoid obstacles, it will not be of highest priority.
If there is time, the system can be further developed to avoid moving obstacles or
overtaking of other cars.

5
1. Introduction

6
 2
Theory

Creating the system architecture in figure 1 involves mathematical modelling and

controller design. This chapter presents the basic structure of the mathematical
model used and how an optimum controller is adapted to the model and purpose.

The mathematical model of the car will be a nonlinear system of differential equa-
tions. To enable the use of more simple control methods, the model is linearised
about certain work points. The linear approximation of the car will be a Multiple
Input Multiple Output (MIMO) system. To use linear control methods, the MIMO
system will be converted to a Single Input Single Output (SISO) system. This SISO
system is then used to develop linear PID controllers.

The methods concerning indoor positioning are described in Chapter 3, since it is

not developed from any special mathematical or control theory.

2.1 Mathematical model of the car

To create a working control system, a sufficiently precise model of the car is needed
so that the controllers can be designed and simulated. The vehicle model is based
on a single-track type of model referred to as a bicycle model.

2.1.1 The bicycle model

To describe the dynamics of the car, a bicycle-vehicle model is used. This is a simple
vehicle model that describes the motion of a symmetric vehicle by simplifying it as a
a single track vehicle, such as a bicycle [14]. A bicycle model is sufficiently accurate
under the assumptions that load transfer is negligible and that vehicle width is by
comparison smaller than turning radius. Negligible load transfer can be assumed
because the RC car has very low centre of gravity and only some minor suspension.
The assumption that turning radius is larger than the car width is made because
the RC car has a width of 6.9 cm, and testing shows that the RC car has a no-slip
minimal turning radius of roughly 23 cm. By examining the different forces acting

7
2. Theory

upon the vehicle, the equations for the dynamics of the car can be obtained. Figure
2 illustrates the dynamics of the bicycle model.

Ffx
αf
Ffy

vy
vx
Fry
αr
Y
lf

Frx lr

Figure 2: A visual representation of the bicycle model used in the project.

Equations of motion

From the physical quantities illustrated in figure 2, equations describing the motion
of the vehicle are defined as the following.

1
v̇x = m
(Frx − Ff y sin(δ) + mvy ω)
1
v̇y = m (Fry − Ff y cos(δ) − mvy ω) (1)
ω̇ = I1z (Ff y lf cos(δ) − Fry lr )

In the equations of motion (1), the x and y subscripts denote local coordinates.
To be able to determine the position of the car, a conversion to global X and
Y coordinates is needed. These are particularly important for simulation. This
conversion achieved by expressing the global velocities Ẋ, Ẏ and yaw rate ψ̇ in
terms of the local velocities vx , vy and w as shown in 2.

Ẋ = vx cos(ψ) − vy sin(ψ)
Ẏ = vx sin(ψ) + vy sin(ψ) (2)
ψ̇ = ω

8
 2. Theory

Nonlinear model

By combining the equations in (1) and (2) a state-space model of the car is acquired.




 Ẋ = vx cos(ψ) − vy sin(ψ)







 Ẏ = vx sin(ψ) + vy sin(ψ)



ψ̇ = ω



1
(3)




v̇x = m
(Frx − Ff y sin(δ) + mvy ω)


1
− Ff y cos(δ) − mvy ω)




 v̇y = m
(Fry



1
ω̇ = (Ff y lf cos(δ) − Fry lr )



Iz

The state-space model describes the nonlinear behaviour of the vehicle. The states
Y and X are the global coordinate positions, vx and vy are the longitudinal and
lateral velocities, ψ and ω are the yaw angle and yaw rate. The variable δ is a
control signal that represents the steering angle of the wheels. The second control
signal which the forces are dependant on is the PWM duty cycle percentage D for
the motor, which ranges from 0-1 and is a part of the motor force Frx as shown in
section 2.1.2. The state space will be linearised around specific operating points and
used to design the controllers. Since the goal of the project is to maximise track
performance, it is essential that the vehicle model is accurate to achieve the desired
result.

Since some parameters are specific to the car in question, they will be determined
by test driving the car in different driving scenarios. With an accurate model, the
dynamics of the car can be simulated and measured, which is essential to the design
of the controller.

Acting forces

The bicycle model describes all tyres connected to one axle as one single, meaning
that the four tyres of the car could be simplified to one frontal and one rear. The
forces acting on the tyres can then be described as the following four:

• Ff x - Longitudinal forces acting upon the front wheel, only rolling resistance
due to rear-wheel drive.

• Ff y - Lateral forces acting upon the front wheel caused by steering, cornering,
slipping, and other lateral movement that depends on the tyres.

• Frx - Longitudinal force acting upon the rear wheel, net force from the drive
line.

9
2. Theory

• Fry - Lateral force acting upon the rear wheel caused by cornering, slipping,
and other lateral movement that depends on the tyres.

Because the car has rear-wheel-drive the only longitudinal force contribution to the
front wheel is rolling resistance caused by friction. The rolling resistance is hard to
estimate due to small scale. Therefore Ff x is regarded as zero, but its effect will
indirectly be included in the motor equation. Another force necessary to include is
drag. Drag force is the resistive force that acts upon a body moving through a fluid
such as air, in other words air resistance. The drag force is dependant on the fluid
density, surface area, velocity, and a coefficient determined by shape factors [15].
The drag force affects the whole body but will be included in the rear longitudinal
force equation for the sake of compactness.

2.1.2 Motor model

As previously stated, the driving force that makes the car move is caused by the
motor Frx acting upon the rear wheels. This force can be calculated using a model
of an electric motor.

Usually when modelling an electric motor, measurements and different motor pa-
rameters are needed. In this case however, since sufficient data about the motor used
was not available, a simpler parameterised model was used to describe the driving
force Frx . [16]
ρvx2
Frx = Cm0 D − C0 − C1 vx − CD A (4)
2
Here, Cm0 is the force parameter of the electric motor. C0 is the constant resistive
parameter for the driveline, which contains contributions from rolling resistance and
motor inertia. C1 is the resistive force in the drive line depending on vehicle speed.
2
The final term CD A ρv2x is a contribution from the drag force included for the sake
of keeping the equations compact.

Because a separate force for friction is included but instead indirectly included in
the term C0 the motor model’s accuracy will be limited to the specific surface used
for the project.

2.1.3 Tyre model

The tyres of the car play an important role in its behaviour. All lateral movement,
slipping, and rotation will depend upon them. There exists plenty of in-depth work
and studies on the subject, but some of it is excessive for the simple design of the
tyres for an RC car. The chosen model is a simplified variant of the Pacejka magic
tyre formula [17] [18]. Figure 3 shows how the formula works. The lateral force
is about linear for small slip angles and for higher angles reaches a peak and later

10
 2. Theory

saturates. n h io
R(k) = d · sin c · arctan b(1 − e)k + e · arctan(bk) (5)

Figure 3: Illustration of the magic tyre formula and how the lateral force depends
on slip angle

The Pacejka magic tyre formula is a model based off semi-empirical experiments. It
is commonly used because of its accuracy compared to its relative simplicity. The
simplified formula of (5) sets the term e to zero.
 
Ff y = df sin cf (bf αf )
 
Fry = dr sin cr (br αr )
  (6)
ωlf +vy
αf = δ − arctan vx
 
ωlr −vy
αr = arctan vx

The parameters d, c, and b with corresponding subscript are the peak factor, shape
factor, and stiffness factors in respective order. αf and αr are the front and rear slip
angles. It is assumed that the parameters for both front and rear tyres are equal,
due to the simplicity of the tyres, thus the subscripts are from now on omitted. The
model possess a weakness at very low or zero velocity due to the division by zero.
This is an accepted inaccuracy because the model only intends to describe the car
during motion, starting from standing still is not of interest.

11
2. Theory

2.2 Linearising the model

To linearise a nonlinear model is the process of formulating a linear approximation

of a nonlinear system around a specific point. In situations where a nonlinear model
is hard to work with and it is assumed to operate at or close to certain points,
in other words variable values, a linear approximation at those points can be used
instead. [19]
 
∂f (z0 , u0 )  ∂f (z0 , u0 ) 
f (z, u) ≈ f (z0 , u0 ) + (z − z ) + (u − u0 ) (7)
 
 0 
∂z 
z0 ,u0
∂u 
z0 ,u0

To linearise a nonlinear model a suitable equilibrium work point has to be deter-

mined.
z0 = [X0 , Y0 , ψ0 , vx0 , vy0 , ω0 ]
Equilibrium implies that the car is driving in steady-state conditions, with no
changes to acceleration in any direction. Since a linear model is specific to one
work point, deviation from said work point will decrease the accuracy. Therefore
a variety of work points for different driving conditions are chosen because no ab-
solute state of equilibrium will [Link] driving conditions that are of interest are
steady-state cornering at different speeds and driving straight.

By choosing multiple work points z0 and linearising the model accordingly, the
overall accuracy of the system can be improved. Multiple linear models require
multiple controllers, created according to each model. Thus the system, though
linear, will be able to handle many different scenarios. A linear state space model
can then be constructed for analysis and controller design.
 
 
∂f (z0 ,u0 )  ∂f (z0 ,u0 ) 
A= ∂z
 B= ∂u

 
z0 ,u0 z0 ,u0 (8)
 
∂g(z0 ,u0 )  ∂g(z0 ,u0 ) 
C= ∂z
 D= ∂z

 
z0 ,u0 z0 ,u0

∆ż = A∆z(t) + B∆u(t)

(9)
∆u = C∆z(t) + D∆u(t)

Equation 8 show how the state matrices are constructed, and equation 9 how the
linearised state space is represented [20]. The function f is a function of the state
variables z and control signals u that returns the derivative of a state ż. The function
g is a function that simply returns whichever signal being measured. The matrix
D is a feedforward matrix and unused because the control signals do not directly
affect the states.

12
 2. Theory

2.3 Controller design

The previous section has stated that multiple linear models will be used to describe
the car in different situations such as driving straight forward or turning, thus simu-
lating a nonlinear behaviour with linear models. This demands multiple controllers
for each linear model. Furthermore, the linear models are MIMO state space models
with two inputs and six states as outputs, from equation 1. The two, inputs δ rep-
resents the ”steering wheel” of the car, and D represents the ”gas pedal”, hence δ
affects the direction of the car and D affects speed. These signals will be the control
signals to follow a virtual track as fast as possible.

Established research about controlling RC cars, to resemble a small scale autonomous

racing situation, uses nonlinear control methods as stated in section 1.1. This project
is as mentioned in the delimitations 1.2 constrained to linear and basic control meth-
ods, hence a PID controller is the best suited for the intended goal [21].

Using PID controller means no methods to handle a MIMO systems is possible

[21]. Instead one controller for each input signal is the most reasonable solution.
Yet, linear model provides six different outputs. The desired PID controller is only
compatible with a SISO model. Meaning that the six outputs signals from the
model needs to be modified or combined in a way to only supply the controller
with one error value per input and PID without losing to much information. Again
using multiple simpler control methods to mimic a more advanced nonlinear control
system. Intuitively choosing yaw angle, ψ, as the only output from the model to the
PID controlling steering should be sufficient, as further elaborated in section 3.3.1
about steering control.

The multiple PID parameters will be gathered in a look-up table. Using look-up
tables is commonly used to model nonlinearities over a wide range of work points
[22]. This project applies it to linear and basic control methods instead, creating a
look-up table to enable functionally for the cars over a great amount of work points.

An overall control design satisfying all constraints from the complex structures of
velocity dynamics fitted in a simpler system, is described in figure 4. A detailed
course of action to implement that design is described in Chapter 3, however the
two decoupled PID controllers do not need to resemble each other. The controller
steering the cars along the virtual track will contain the look-up table and requires
great accuracy to stay on the line. Speed control could simply be an algorithm
pushing the cars to drive as fast as possible without deviating to much from the
track.

13
2. Theory

Figure 4: Scheme describing an overall control system, with two decoupled control
loops. Both control loops will be further elaborated in section 3.3. The structure of
the algorithm is further elaborated on in section 3.3.1.

The PID controller

A PID controller is a control loop feedback controller, used in a variety of appli-

cations. The controller calculates an output signal using error values, which is the
deviation of the systems current state from the desired operating point. As the name
implies, a PID controller consists of three parts. Proportional gain, integral gain
and derivative gain. These values are obtained by analysing the system’s transfer
functions. This requires a linear SISO model.

Expressed in the time plane the PID controller is as in equation 10. Where the
control signal u depends of the measured error value: em (t) = r(t) − ym . See figure
5.

!
1 Zt dem (t)
u(t) = Kp em (t) + em (τ )dτ + Td (10)
Ti 0 dt

Laplace transformation of the relations in the time plane means: U (s) = FP ID (s)Em (s)
which leads to the transfer function of the PID controller as in equation 11. Intro-
ducing the relation of both Ki = Kp /Ti and Kd = Kp Td the PID is commonly
described as in equation 12.

14
 2. Theory

!
1
FP ID = Kp 1+ + Td s (11)
Ti s
Ki
FP ID = Kp + + Kd s (12)
s

The proportional gain Kp is proportional towards the control signal: u(t) = Kp em (t).
Increased Kp means increased response speed in the loop transfer function since
the output y is the natural frequency of the loop transfer function (see figure 5).
Increased value of Kp does decrease stability margin and raises the integral- and
derivative gain, which could lead to bigger overshoots in y.

The integral gain Ki /s raises the gain at lower frequencies, hence increases the
control signals accuracy in the static responses and enhances the compensation of
low frequency upsets from the load disturbances. However, too great value of Ki
causes bad stability margins.

The derivative gain, Kd · s contributes to stability margins and compensation for

overshoot and inferior stability margins due to the integral gain. Mathematically it
gives an increasing contribution to the loop transfer function, unfortunately which
also increases the high frequency gain. The high frequency gain can increase sensi-
tivity for measurement noise. When the car is following a line a too great value of
Kd could cause the car to oscillate around the line in a wobbly way. This section
(2.3) is based on theory from [Link]’s book: Reglerteknikens grunder [21].

Figure 5: Simple feedback control-loop with PID controller. G(s) represents the
model, and FP ID the PID controller, u is the control signal and em the measured
error value, y is the output and r is the reference signal [21].

15
2. Theory

Look-up table

To address the issue of the RC car and model behaving differently in different regions
of it, a look-up table is applied to the control loop for steering. A look-up table is
an array of values stored before a program is run. During operation, the program
will fetch these values according to a predetermined algorithm. This replaces run
time computing of different parameters. In this project, a look up table is used to
store different PID parameters for the steering control. During racing, the program
will use known data about the car, such as speed, angle and angular velocity, to
determine suitable parameters for controlling the car.

2.3.1 PID tuning with heuristic stability margins

Functional and stable PID controllers will be created using heuristic reasoning about
the sensitivity function equation (13) and the complementary sensitivity function
equation (14). The theory behind this section is according to [23] and [21].

1
S(s) = (13)
1 + L(s)

L(s)
T (s) = 1 − S(s) = (14)
1 + L(s)

The goal is to find a PID controller that corresponds to the maximum values of
the sensitivity and complementary sensitivity function, named: Ms in equation (15)
and Mt in equation (16). Then according to [23] this could create an optimum
controller. Optimum controller defined as having good mid frequency robustness
and good balance between control activity and the accuracy of the output signal.

1
MS = maxω |S(jω)| = (15)
minω |1 + L(jω)|

L(jω)
MT = maxω |T (jω)| = maxω (16)
1 + L(jω)

For designing a PID controller there are several possibilities, here the parameters
will be based upon deciding phase margin (ϕm ) and cross over frequency (ωc ). Phase
margin being the maximum amount of phase shift possible before the loop trans-
fer function turns unstable. Cross over frequency means the frequency where the
magnitude of the model is 0 dB and the phase corresponding to that frequency.
When designing a PID controller it is possible to start by designing the less complex
Proportional and Integral gain (PI) controller to obtain initial values for ωc and ϕm .

16
 2. Theory

The basic idea is to create an amount of PI controllers by using an assumed interval

of crossover frequencies and phase margins. Then choosing the most optimal of
those PI controllers by calculation Ms and Mt , which provides great initial values
to calculate the desired PID controller.

A simple principle of controller design is to specify the loop transfer function L(s) =
G(s)F (s) for a frequency ω0 , if the models transfer function is known for that par-
ticular frequency: G(jω0 ) = |G(jω0 )|∠G(jω0 ). This makes is possible to calculate
the controllers magnitude: |F (jω0 )| = |L(jω0 )|/|G(jω0 )| and phase: ∠F (jω0 ) =
∠L(jω0 ) − ∠G(jω0 ). Since this method is based upon ω0 being know, an application
of this method is to put ω0 as the cross over frequency ωc , which corresponds well
to the desired method of tuning presented above. Then the magnitude and phase of
the loop transfer function equals |L(jωc )| = 1 ∠L(jωc ) = −180◦ + ϕm − ∠G(jωc ).
Then a PI controller can be design using equation 17, where ϕm is the desired phase
margin for the controller, hence both ωc and ϕm are known [21].

1
|F (jωc )| = |G(jωc )|
(17)
∠F (jωc ) = −180◦ + ϕm − ∠G(jωc )

The above presented method of designing a controller is applicable to the system in

question, since the phase margin and amplitude of the model is known and can be
used to find the phase margin and magnitude of the controller. To find a suitable
cross over frequency for the system to start with, the following observation was made
[23]:

ωc = 0.6 · ωG180 (18)

To tune this PI controller an interval of the phase margin (ϕm ) and the cross over
frequency (ωc ) for example a PI controller: 45◦ < ϕm < 60◦ and 0.3ωG180 < ωc
< 0.6ωG180 where ωG180 = ∠G(jωG180 = −180◦ ). Then calculating several PI con-
trollers and their several Ms and Mt . The PI with the highest Ki that also fulfils:

Ms ≤ 1.7
(19)
Mt ≤ 1.3

is the most optimal PI controller within the interval [23]. Normally ϕm is 50◦ for
both PI and PID controllers [23]. The limits in (19) are based of empirical obser-
vations [23], to guarantee a good combination of robustness and speed for a stable
model. It is relevant to note that increasing the value of ωc could increase the speed
of the loop transfer function. Choosing the correct value of ωc is crucial to achieve
the correct speed that correspond to that of the cars.

17
2. Theory

18
 3
Methods

For the development of this system, miniature RC cars were chosen instead of full-
sized real ones. This choice is due to economic limitations and has the advantage
of being much safer for a student project, where the goal is to achieve maximum
speed. It also reduces the complexity of the system since RC cars are easier to handle
compared to full-sized ones. The small RC cars will serve as sufficient models for
real cars since the vehicle model and control algorithms should be rather similar for
small and large vehicles. Using the vehicle model described in state space model
(3), the RC car’s dynamics could be simulated which is then used to design the
controllers.

The system architecture consists of positioning and controlling of the cars, and the
control system. The system also has to be modelled to make it possible to simulate
it and get accurate control parameters. A nonlinear model is created and linearised
around multiple work points. The linear models needs to be precise, since a high
performance system demands precise calculations. The linear model is further used
to design a special type of PID controller which utilises all of the different linear
models. Each linear model represents a situation, for example a sharp turn or
straight forward. This information is merged to one PID controller for each linear
model combined in a Look-Up table and is used to control the steering angle of the
car.

The speed control will be achieved by an algorithm based on the slowest of two sub-
algorithms. The first of which slows the car down if its current direction deviates
from the target direction, otherwise it can increase it’s velocity. The second tracks
if there is an upcoming angle change in the track ahead, if there is then the car must
slow down, otherwise it can increase it’s velocity. This is then used as the reference
value to a PID which controls the velocity of the car.

The indoor positioning system is based on a camera hanging down from the ceiling,
detecting the movements of the cars. The information obtained through the camera
can then be used to control the cars in the direction of choice.

19
3. Methods

3.1 Autonomous system structure

The autonomous system used in this project consists of a miniature RC cars that
are controlled by a computer. All calculations needed for controlling the cars are
done on this computer, which is sends control signals to the cars. The computer
gets all its information from a camera hanging from the ceiling.

3.1.1 Remote controlled cars

The RC cars in the system are two RC model cars of type Kyosho Mini Z. Scaled at
1:27 in relation to their real counterparts, with a length of around 160mm[24]. They
are fitted with LEDs, emitting IR light, to facilitate the positioning by imaging as
shown in figure 6. It is desirable to detect both direction and position of the car,
as such the LEDs are positioned as a non-equilateral triangle with one in the front
and two in the back; essentially forming an arrow. One of the cars also have an
additional LED to allow separation of two cars by counting LEDs. To power the
LED the cars are supplied with an additional battery mounted on the inside of the
chassis.

Figure 6: This picture shows one of the cars used in the project, including the
addition of the LEDs. The RC car measures 16.5x6.9 [cm].

3.1.2 Remote controller

The controllers are disassembled and modified by disconnecting the existing po-
tentiometers for speed and steering, which are replaced by a Digital to Analogue
Converter (DAC) circuit (see figure 7). Using a dual channel DAC connected to an

20
 3. Methods

Arduino over Serial Peripheral Interface (SPI) in the DAC circuit enables digital
control of the cars.

The Arduino is connected to the central computer controlling the system with a
Robot Operating System (ROS) node running on the Arduino. The Arduino then
receives the control signals to change speed and steering angle, configuring the DACs
accordingly.

Arduino
Remote Controller 1
DAC 1 3.3 V Pin
Vdd Vout A Throttle
USB IN GND
Pin 8 CS AVss
3.3 V Pin
CLK CLK Vout B Steering
GND
MOSI SDI LDAC

Remote Controller 2
DAC 2 3.3 V Pin
Vdd Vout A Throttle
GND
Pin 7 CS AVss
3.3 V Pin
CLK Vout B Steering
GND
SDI LDAC

GND 3.3V Supply

4.7 kOhms

Figure 7: The simplified circuit diagram over how the cars are controller. In this
setup, pin 7 and 8 on the Arduino are used for chip select.

3.1.3 Camera

The positioning system uses a Point Grey Flea3 camera[25]. The camera captures
150 frames per second (FPS) and uses a monochrome sensor and a global shutter.
The high frame rate is necessary for maintaining precision while tracking since it
minimises movement between data points, even at higher speeds. The monochrome
sensor limits the data to a single channel instead of the multiple channel colour
sensor. Since only infrared light is of interest, more channels would only result in
more data to gain the same result. A global shutter is preferred to the alternative
rolling shutter. This since a rolling shutter captures the image row by row, meaning
that each row of pixels will not all be from the same time instance[26]. Therefore, if
there are fast moving objects in the image it will be distorted, making the positions
of the cars at a given time uncertain. A global shutter, on the other hand, captures
all the pixels in the same instant, which results in more reliable data.

21
3. Methods

An additional physical filter that removes light outside the IR spectrum is fitted to
the camera. By limiting input to light of the IR spectrum little noise is permitted to
disturb the light of the LEDs on the cars, compared to allowing the full spectrum.
This results in a high contrast image of the cars LEDs.

This is an easier way of acquiring precision in a confined space in comparison to

radar or triangulation solutions.

Software architecture

The system built utilises the ROS as a development platform. The system is devel-
oped purely in C++. It has, through OpenCV, necessary tools for image process-
ing. The ROS architecture of using nodes for tasks combined with the possibility
to record data between nodes is useful since it enables offline testing and isolated
development and debugging of tasks.

3.2 Modelling

For the computer to be able to control the car, an accurate mathematical model of
the car is needed. This model will be the foundation of almost all the design and
theoretical work regarding the car and therefore has been given a lot of attention.
Since PID controllers are used in this project the model needs to be linearised, when
this is done the different parameters needs to be identified to complete the model of
the car.

3.2.1 Linearisation

From the vehicle model (3), a linearised system of equations can be derived, from
which PID controllers can be designed. Classic PID controller design strategy dic-
tates that the controller is designed from the transfer functions of a system that
is acquired by linearising the system around an equilibrium. Because the system
describes a car in motion with global coordinates, a state of equilibrium cannot be
found for Ẋ, Ẏ or ψ̇. Thus the chosen strategy is to linearise the system around
multiple states of motion for the car, meaning a work point will be defined only by
vx0 , vy0 and ω0 .

Linearising the equations in (3) poses a few problems. The latter three states can
be linearised around where they equate to zero because it only implies a state of no
accelerations, which it is fair to design the controller around. The first three states
bring some issues. The global velocities Ẋ and Ẏ do not posses any values where a
linearisation is reasonable because the car can move at various speed in the whole

22
 3. Methods

coordinate system. The yaw state ψ̇ should not be zero either. Because the yaw
derivative is equal to the angular velocity ω, there is no point in linearising it. It is
already equal to ω0 .

Because of these factors, as well as the global velocity equations being the only two
that depend on ψ, the chosen strategy is to focus only on the latter three equations of
(3). Doing so, a system of three equations with five variables, the three states vx vy ω
and the two control signals δ and D, is obtained. This system is overdetermined. As
previously mentioned one of the states which are of particular interest in choosing
work points is ω0 . If a value for ω0 is already chosen, the number of variables
is reduced to four. If vx0 is also predetermined as per choice of work points, the
system becomes solvable and the corresponding control signals and resulting lateral
velocity vy0 can be calculated. To accomplish a full linearisation of the whole system
of equations, work points for ψ0 would also necessary. Because the linearisations of
Ẋ and Ẏ are not of any particular use, ψ0 can be set to zero or otherwise ignored.

Thus with the three states ψ0 , vx0 and ω0 defining the work points for the system,
and the global velocities Ẋ and Ẏ being allowed to take whatever values they may
as a result of the other equations, the linearisation can be accomplished. Due to
the design strategy of using multiple work points, the calculations are done using
Matlab.

By producing several linear systems around different combinations of the aforemen-

tioned states, transfer functions that can describe most of the physical states the car
could be in are collected. From these, it is possible to design a space of PID param-
eters to form a look-up table from which the controller will estimate and choose its
parameters depending on the input signals from the camera and tracking software.

Simulation

Simulating the model is a vital step. It enables the verification of model validity as
well as comparison between nonlinear and linear models by studying their signals and
state values. A simulation also helps with tuning physical parameters to suit the real
car and testing control structure. The most suitable software for this is Simulink. In
the software the equations of motion (3) are written into function blocks that simply
returns the state derivatives for corresponding inputs, integrating these returns the
state values.

3.2.2 Parameter identification

Many of the parameters in the equations that model the RC car are unknown at the
point in time when the equations are derived. Much of the theoretical framework
can be laid out with mostly symbolic parameters or placeholder values. Before
applicable results can be obtained from the model, all of the parameters have to

23
3. Methods

be determined. These include mass, inertia, dimensions, centre of gravity, drag

coefficient, tyre parameters, and motor parameters. Most of these properties are
the same or close for both the RC cars. For the sake of simplicity and time the
parameters are only calculated for one of the cars, and used for both.

Mass is measured using a scale with error margin of 2 grams. Length and width is
measured using a ruler. Centre of gravity from the front is estimated with sufficient
accuracy through repeated tests trying to balance the car on a thin surface. The
moment of inertia around the car’s vertical axis is calculated by separating the
it into its three main components. Chassis, body, and batteries. The respective
moments of inertia for the components is calculated by simplifying each into a
standardised shapes with available formulas. They can then be summed by moving
them to the assembled car’s centre of gravity using Steiner’s theorem [27]. The
2
drag force CD A ρv2x is estimated by measuring the car’s frontal area A and using
standard table values for the air density ρ [15]. The drag coefficient CD is assumed
to be comparable to that of a real car with a similar body, because the coefficient
is primarily dependant on shape factors. The Ferrari model is made to resemble a
Ferrari 360 GTC and therefore it is the model the coefficient should be chosen after.
Ferrari does not publish this specific data for their models, so a less credible source
had to be used and for a slightly different model, Ferrari 360 Modena [28]. The
provided coefficient is within expected values for a sports car however and can be
expected to suffice [29].

The tyre and motor parameters are difficult to determine exactly. Thus they are
estimated by fitting the model behaviour to the real car with minimal deviation. The
basic method for the both the tyres and motor are the same. Tests are performed
with the RC car in situations where the specific parameters produce data that is easy
to quantify and compare with the model. An exact solution is not very probable so
instead a set of parameters is produced. The identification process is then repeated
but with the identified sets, and the deviation from the test data is computed using
mean squared errors.

N 
X 2
Ei = Yn − f (xn , yn , Pi ) (20)
n=1

By summarising the mean of the errors squared according to (20) for N amount of
tests for i amount of parameter sets, where f is a function of input signals x and
output states y that depends on the parameters Pi in question resulting in a value
that can be compared to tested property Yn . By selecting the parameter set with
the smallest Ei it is possible to determine a good estimation that takes sign into
account and emphasises larger errors.

24
 3. Methods

Motor parameters

With the motor modelled as in (4) there are three parameters that have to be
identified. To find suitable parameters, the model is tuned to match the behaviour
of the RC car when driving in a straight line, to achieve similar acceleration and
velocity. The car is run at different input signals and its velocity is measured with
the tracking software. To determine the parameters, the steady state speed and
momentary accelerations ax with corresponding momentary velocity are used to
construct linear equation systems. These equations are then solved which returns a
set of parameters. The parameters with the least mean square error is to be used.

By setting up a system of linear equations, an exact set of parameters for the specific
test sample can be determined.
 
0
  
D0 −1 0 Cm0
2
CD A ρv2x
 
 D −1 −vx   C0  =  (21)
    
2

D −1 −vx C1 ax · m · CD A ρv2x

The variable D0 is the input signal for which the RC car just starts rolling, therefore
the velocity and acceleration is zero. The second row equation has velocity but no
acceleration, it represents the signal and velocity for which the RC car has stopped
accelerating. By calculating the momentary acceleration and velocity for several
points during the acceleration phase the third row can be calculated. Solving the
equation system 21 for all tested signals D with corresponding measurement data
vx and ax .

Tyre Parameters

The tyres are modelled using a simplified Pacejka’s tyre model (6). The equipment
normally used to determine the parameters in the tyre formula is expensive and
unavailable, so empirical tuning methods are used instead. The simplest behaviour
where the tyre model will come into effect is circular motion as it affects lateral
slipping and rotation.

Tests are performed with the RC car being given a constant signal to motor and
steering angle. Its speed and positional data is measured and processed using the
camera and tracking software. The linear equation method used for the motor is
harder to apply in this case because the behaviour is dependant on more than one
equation, all which are ordinary differential equations. Brute force iteration is used
in Matlab to find parameters that produce a reasonably similar behaviour by using
the ODE45 function for a large combination of parameters. The ODE45 function
is applied on the system of equations formed by the equations of motion in (1). It
returns the state values over a specified time span. With the formula vx /ω = r a
steady state circle radius can be expressed. The brute force loop runs the script for
a set of input signals D and δ that were used during tests and saves the parameters

25
3. Methods

if they produce a radius and velocity that are close enough to the test data. For the
accepted sets of parameters the set with the least mean square error is used.

Ideally the parameters would be fitted using measurements of lateral force and slip
angles, but since measuring these is not possible due to equipment limitations, the
above described method is instead chosen.

3.3 Controller design and implementation

With the mathematical model in place it was possible to design a control system
for the car. According to the theory chapter the control structure will consist of
two separate loops, one to control steering to keep the car one the line and one to
control the velocity. The steering controller is based on the linearisation cases and
the reference value is the desired direction. The reference speed is calculated using
an algorithm that is based on the deviation between the current global angel ψ and
the target angle, and how sharp the curve of upcoming segments of the virtual track
is.

3.3.1 Steering control

By splitting the control structure in two separate control loops (see figure 4) it
provides a natural solution to the problem of using a MIMO state space model with
a PID controller (SISO). The loop controlling steering is built around the yaw angle
ψ, thus the C matrix in the state space model can be adjusted accordingly so that
the only output will be ψ with respect to δ. One advantage with choosing ψ as a
control signal is that it will allow for drift control if the back wheels lose traction.
In the third equation of the nonlinear state space model (3), ψ is only dependant
on the integration of ω, and ω in turn is independent of D. Thus no information
is lost if the steering control is based entirely upon the state of ψ, since it is only
dependent of the input δ, see ω̇ = I1z (Ff y lf cos(δ) − Fry lr ). The independence of ψ
from D is further shown in table 1.

26
 3. Methods

Table 1: State space model, the D-matrix contains only zeros

A= B=
x1 x2 x3 x4 x5 x6 u1 u2
x1 0 0 -0.1084 0.995 -0.09983 0 x1 0 0
x2 0 0 0.5921 0.09983 0.995 0 x2 0 0
x3 0 0 0 0 0 1 x3 0 0
x4 0 0 0 -2.437 6.757 0.5998 x4 -3.935 9.052
x5 0 0 0 4.357 -56.57 -1.018 x5 17.07 0
x6 0 0 0 1373 -1041 -966 x6 3935 0

C=
x1 x2 x3 x4 x5 x6
y1 1 0 0 0 0 0
y2 0 1 0 0 0 0
y3 0 0 1 0 0 0
y4 0 0 0 1 0 0
y5 0 0 0 0 1 0
y6 0 0 0 0 0 1

Since the goal is to be able to generate a great amount of PID parameters, a method
for looping through the controller designing steps is necessary. This will perhaps
take away the possibility of tuning every PID to optimum, but the hypothesis is that
using a look-up table will compensate for small tuning effects. Matlab was used to
create this PID generating loop.

To create a PID controller from the state space model in table 1 a transfer function
from δ to ψ is computed in Matlab using the command:
1 [ NUM_Delta , DEN_Delta]= s s 2 t f ( s . s y s 2 . A, s . s y s 2 . B, s . s y s 2 . C, s .
s y s 2 . D, 1 )

which creates one transfer function per state in the A matrix regarding the input
δ. Then by saving the transfer function for ψ, a SISO model is created to be used
when designing PID controllers.
1 %SISO p l a n t t o be used i n s t e e r i n g c o n t r o l
2 p s i _ s i s o=t f (NUM_Delta ( 3 , : ) , DEN_Delta ) ;

To create several PID controllers in a loop the Matlab function:

1 [C i n f o ] = p i d t u n e ( p s i _ s i s o , ’ PID ’ , wc_opt , o p t s ) ;

is utilised. As arguments it takes a desired crossover frequency, phase margin, a

SISO model and a string saying which type is desired. To be able to, in a loop,
enter appropriate values off both cross over frequency and phase margin some ap-
proximation needs to be done before using the function.

27
3. Methods

Following the theory detailed in section 2.3.1 it is possible to design and create a
number of PI regulators and choose the most optimal controller by using the limits
in equation (19). An initial value of both the crossover frequency and phase margin
can be chosen, and later used as input to the Matlab function ”pidtune();”. Code to
generate multiple PI controllers and choosing the optimum is available in Appendix
1. The code is from the book Reglerteknikens grunder [21]. Those optimisation
functions also require an initial interval of frequencies. In section 2.3.1 the reasoning
stated that ωc = 0.6 · ωG180 , but after studying how the RC cars behave another
approach is possible.

For the RC car to receive a control signal and for the camera to observe any change
approximately 0.8 seconds goes by. Although the car has a real reaction time closer
to 0.2 seconds, it appears as many uncertainties collaborate, for example the servos
taking some time stabilise. Therefore, the delay time was approximated to be higher.

Accordingly the response time of the loop transfer function does need to be any
faster then that of the reaction time. When running the code in Appendix 1 to
create optimal PI controllers by using the suggested interval of frequencies, the loop
transfer function’s time to reach steady state is too fast. Meaning that the car
will not be able to react to those signals before a new one is received. Instead of
creating PI controllers with the tuning strategy from section 2.3.1, instead using the
crossover frequency of the model does create more reasonable response times. See
figure 8.

28
 3. Methods

(b) Bode plot of L(s) after using the opti-

(a) Crossover frequency of the one model. misation algorithms from section 2.3.1 and
code in appendix 1.

(c) Step response of L(s) from figure 8b

Figure 8: Different plots of one SISO model with δ as input and ψ as output. This
model is linearised around the work point: [X=0 Y=0 ψ=0.3 vx =0.3 vy =0 ω=0.8]

When an interval of crossover frequency is decided, an interval for the phase mar-
gins is approximated according to [23], the optimisation code will generate values
for phase margin and cross over frequencies to input in the pidtune(); command
in Matlab. The Matlab code which is available in appendix 1, uses both the opti-
misation code from [21] and Matlab ” pidtune();” it generates PID controllers for
each linear state space model. This example generates five PID controllers for five
different work points.

Because the controller uses integral gain there is a need for anti-windup [30]. In
systems where the actuator has a physical limit, such as a car with a maximum
steering angle, a problem arises when the actuator is saturated. If the error is
nonzero when the actuator is saturated the system cannot physically increase the
control signal, even though the controller keeps integrating the error. This means
that when the output signal reaches its reference value there is still built-up signal
from the integral gain that the controller has to overcome until the output signal

29
3. Methods

decreases, which results in overshoots. To handle this a simple clamping algorithm

is implemented that constricts the integral gain to −0.3 < I < 0.3 when the steering
output δ is saturated.

Neighbouring Look-up table

A look-up table typically uses interpolation to generate control parameters. The

table used here does not employ interpolation due to computation limits. Instead
the closest values for any work point is used. This variation will still be referred to as
a "look-up table" even though it’s more a neighbouring look-up table. The look-up
table is computed using different values for vx and ω. These values represent work
points for the car. During operation, the car might be operating at work points
not included in the look-up table. In these cases an approximation is made to the
nearest work point. It is therefore important to have linearised the model about
many work points. This should improve steering control and overall performance.
The overall view of the steering control is shown in figure 9.

Figure 9: Scheme describing the overall steering control loop

3.3.2 Speed control

The goal of simply driving as fast as possible provides necessary constrains for the
speed controller. To drive a lap as fast as possible includes staying on the line and
while maintaining high speed. Controlling the speed is then simply an algorithm
saying that the speed can be increased if the present global angle does not departs
to much from the virtual track or decrease speed if there is an upcoming curve.

The velocity of the car is also controlled using a PID controller, see figure 4. The
reference value for the PID controller is calculated using the lowest value of two
separate algorithms.

30
 3. Methods

One algorithm adjusts the speed based on the error in angle between where the car
is actually going and where it should be going. Alas if the car is pointing against
the target point, the speed will be maximum.
(ψr − ψ)
Smψ = 1 − · Wψ (22)
π
Where Wψ is a constant weight.

The other algorithm adjusts the speed based on the upcoming curvature of the track.
It looks at a distance in front of the car, determine the same way as the target point,
two equally long consecutive track segments as straight lines and determines their
angular difference. If they do not differ the speed will be maximum. See figure 10
for a visualisation.

|∠1 − ∠2 |
SmCurve = 1 − · Wc (23)
π

Where Wc is a constant weight. ∠1 and ∠2 are the angles of the two consecutive
track segments starting from a distance before the car.

Figure 10: The two track segments are marked red and green. Their global angle
corresponds to ∠1 and ∠2

The results of these algorithms are then chosen by which is the smallest, with the
minimum value being chosen is 0.25, and then multiplied with the wanted speed.

Sr = S · max(min(SmCurve , Smψ ), 0.25) (24)

31
3. Methods

3.4 Indoor positioning

An important part in controlling the car is being able to get the position, direction
and velocity of the car. To do this the information from the camera are fed into
the computer and with the help of the LEDs mounted to the car it is possible to
determine the position, direction and velocity

3.4.1 Position, direction, and identification

To establish the position, direction and identity of the cars, multiple operations has
to be done. Initially the incoming camera frame is analysed. The frame is searched
at every other row and column for a 3 by 3 squares of pixels where the middle pixel
has a brightness over the set threshold and the surrounding pixels have a brightness
of more than half the threshold. This method speeds up the analysis immensely since
only half the data is analysed. The only downside is that the LEDs are required to
be larger than 3 by 3 pixels on the camera frame.

All found squares are then iterated to merge closely located squares into larger
circles where the middle of the circle is the average middle position of the squares.
At this stage, if the parameters are properly set, circles originating from the centre
of each LED should have appeared. This is then analysed in the same matter, but
without merging and using a larger circle, to see which points belong to which cars.
See figure 11 for visualisation. If a fourth identification dot is found within the car
circle the car is identified as number two otherwise it is identified as number one.

Figure 11: A car found by the image processing, with red circles around LEDs,
orange circle and arrow for centre and direction of the car, and a green rectangle
around it all, symbolising the car.

32
 3. Methods

3.4.2 Velocity

The velocity is calculated by measuring time between two points at a distance of

5 centimetres or if the time passed between measurements is over 100 milliseconds.
Only measuring between two well distanced points allows for greater precision since
small noise in movement is mitigated by the larger movement. The fallback of
a maximum time between measurements ensures that some velocity always gets
reported, even at standstill.

3.4.3 Calibration

To utilise the images from the camera system in a relative matter (difference in
distance). The system needs to be calibrated to ensure a difference in distance is
the same physical distance at different parts of the image.

Each image is first rectified so that it is not bent at the edges. This ensure that
distances can be calculated correctly. This is done using built-in ROS calibration
tools.

To integrate the measured data with the model the physical distance a pixel repre-
sents is calculated. It is done through a calibration process with two IR LEDs at a
known distance in meters. The calibration is done two times, one for the x-axis and
one for the y-axis. The conversion factor is the distance in meters divided by the
distance in pixels, for each axis respectively.

3.5 Creating the track

For the system to be able to know where the car should drive a Matlab script was
created for generating a path. As one of the goals for this project was to be able to
avoid obstacles an obstacle avoidance algorithm was also developed.

To generate a track for the car to follow a Matlab script was created. The user
inputs a few points where they want car to drive, the script then calculates piece
wise polynomials between these points forming a spline. The order of the points
are taken into account so that the user can form almost any shape of track as they
want. See figure 12.

33
3. Methods

2.5

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3

Figure 12: The program used to make a track for the car to follow

3.5.1 Obstacle avoidance

The obstacles being used in this project is other cars, therefore the positioning
algorithms are used for detecting obstacles. When the position of an obstacle is
known the edges of the obstacles are marked, then half the width of a car is added
to the obstacle in every direction, the reason for this will be explained later in this
section.

After the obstacle is identified and located the system checks if the planned path
intersects the obstacle, if this is the case three point are added on the side closest
to the planned path. Two point are added at the edges of the obstacle and a third
in the middle of the first two points, see figure 13.

1.05
0.95 Planned path Planned path
Obstacle 1 Obstacle
Added Data Added Data
0.9
0.95
0.85 0.9
0.8 0.85
0.75 0.8

0.7 0.75

0.65 0.7
1.6 1.7 1.8 1.9 1.6 1.7 1.8 1.9 2

Figure 13: An example of how the system adds points for obstacle avoidance.

When these three points are marked, the two outermost point are connected to
the planned path and to the middle point with straight lines, essentially forming
a triangle around the obstacle, see figure 14. Even though the path is jagged and

34
 3. Methods

might seem to make the car drive in a jagged way it actually turns out that the car
drives in a rather smooth line around the obstacle. This is because the car cannot
react as fast as it would require to drive in a triangular way, but also because the
target point will have changed before the car arrives at the last target point resulting
in a smooth path.

1.1
Planned path
Obstacle
1
New Path

0.9

0.8

0.7

0.6
1.2 1.3 1.4 1.5 1.6 1.7 1.8

Figure 14: The planned path around an obstacle shown in red.

A problem with this method for obstacle avoidance is that it does not take the size
of the car into consideration when planning a new path. So, even if the car drives
along the new path it would still crash into the obstacle. This is solved by adding
half the width of the car in every direction to an obstacle when feeding the position
and size to the system. Then when planning the new path it is not necessary to
take the size of the car into consideration, simplifying the programming.

35
3. Methods

Special case

There is a special case where only one of the corners of the obstacle is in the planned
path, see figure 15. If the obstacle avoidance would work the same for these kinds
of cases as it normally does it would result in an odd path.

The way the obstacle avoidance works in this case is checking where the planned
path enters and exist the obstructed area, if the entrance and exit is on two sides
that border each other, as it does in figure 15 then only one point on the obstacle
is used for the new path resulting in a path like the one in figure 15.

Planned path Planned path

1.5 Obstacle 1.5 Obstacle
Added Data New Path

1.4 1.4

1.3 1.3

1.2 1.2

1.1 1.1
1.3 1.4 1.5 1.6 1.7 1.8 1.3 1.4 1.5 1.6 1.7 1.8

Figure 15: A situation that the obstacle avoidance needs to handle differently than
most cases.

36
 4
Results

Following sections contains both results and discussion, as well a reasoning about
improvements and areas for further development. The results concern both real
tested results and simulations, as well as comparisons between real and simulated.

4.1 Model

Numerous results from the model are obtained. Various parameters and how well
they perform affect the final result. The controllers are designed from the linearised
model, so it is important to check its accuracy. It is also of high interest to see how
the simulation performs and how reliable it is.

4.1.1 Simulation based validation of the linear model

Before using a linearised model to develop a controller, the behaviour of the linear
model is compared to the nonlinear one, as to verify the reliability of the obtained
values. To test this, output data from a linearised model is compared to output
data from the nonlinear model, operating at the same work point. To further prove
the validity of the data, a different input value is introduced after a certain time in
the simulation. This shows how the model behaves close to set work point. This
is an open loop test, meaning there is no feed back from the system. The several
of the parameters used are placeholder values. The test is therefore not indicative
of anything with concerns to the nonlinear model or its behaviour. As long as the
parameters are the same for linear and nonlinear, the comparison is still valid.

The figures bellow are the results from the comparison. For this test, the model is
linearised about the following work point:

vx 2
vy 0
ω 0

37
4. Results

In practical terms, this work point describes the situation where the car is driving
at a constant speed, in a straight line. The input values to the model are shown in
figure 16.

(a) InputsignalD (b) Inputsignalδ

Figure 16: Input values for D and δ

Initially, the input values are D = 0.39 and δ = 0, which means the car is driving
straight at a constant speed. At t = 2.5s, D is incremented by 0.1, which results
in an increase in vx . At t = 5s, the input signal δ in incremented by 0.01, which
results in a change in direction.

(a) State vx (b) State vy

Figure 17: vx and vy comparison

Initially, as shown in figure 17a the longitudinal speed vx is constant in both models.
At t = 2.5s, when D is incremented, vx increases in both models at the same rate,
which is on par with the expected behaviour. Figure 17b shows that vy = 0, which
means the car has no lateral speed, which is the case when the car is driving straight.
At t = 5s, when δ is incremented, the car starts cornering. Until this point in time,
both the linear and nonlinear models behave identically.

38
 4. Results

When the car starts cornering, there is a decrease in nonlinear vx . There is also a
change in both nonlinear vy and linear vy . The nonlinear behaviour is accurate to
how a car drives in reality, where longitudinal speed vx decreases, and lateral speed
vy becomes non-zero. A negative vy means that the car is performing a left turn.

The linear behaviour is no longer identical during this manoeuvre. Linear vx is

unchanged during cornering, while linear vy is changed to become non zero. This is
due to the linear model being linearized about a certain work point, where in this
case vx = 2m/s, which means the model will behave similarly, but not identically,
to the nonlinear model when deviating from said work point. A decreased accuracy
is therefore expected from the linear model.

This is further proved when comparing the cars angular velocity ω and global angle
ψ positions, as is shown in figure 18.

(a) State ω (b) Global state ψ

Figure 18: ω and ψ comparison

Here, the same pattern as in figure 17 reappears, where both models produce iden-
tical results until the car starts cornering at t = 5s.

39
4. Results

(a) Global state X (b) Global state Y

Figure 19: Global X and Y comparison

Here, the linear X and Y are not accurate any time during this test. Even though Y
seems to be accurate during the first 5 seconds, it is only so because the simulation
was initiated with ψ = 0, meaning the car is driving straight in the X direction.
It was believed that the linear global positions would be accurate while the linear
model is operating about its work point. As shown in figure 19, however, points
towards the contrary.

The reason for this large deviation is uncertain, but a possible cause could be a
problem in the simulation algorithm. Due to the restricted time frame of this project,
this issue was not further investigated. It has, however, been deemed unimportant
for the intended goal, since the global positions are not used in the PID controller.

To conclude, the four states vx , vy , ω and ψ, behave similarly, but not identically,
in both the linear and nonlinear models. This proves the need for multiple linear
model to describe the nonlinear one with sufficient accuracy.

4.1.2 Parameters

Most of the physical parameters in the model are simply measured or calculated as
described in the methods chapter 3.2.2.

Mass and geometry can be assumed to have about as much accuracy by-hand mea-
surement with ordinary tools can achieve. The drag coefficient has some degree of
unreliability. As described in section 3.2.2 the drag coefficient was assumed to be
about the same as that of a full-scale Ferrari GTC 360, same model as the RC car is
made to look like. This assumption has somewhat weak support because the shape
is not guaranteed to be the exact same, and the material is vastly different. Ferrari
also does not provide drag coefficient in the data specifications for this model, so the
reference used is not the most reliable. Neither was a full aerodynamic model used,

40
 4. Results

m Mass 183 g
L Length 16.5 cm
W Width 6.9 cm
cg Centre of gravity 9.25 cm
lf Distance from front to cg 9.25 cm
lr Distance from read to cg 7.25 cm
Iz Moment of inertia 7.3526 ·10−5 kg · m2
A Frontal cross sectional area 0.2135 m2
Cd Drag coefficient 0.335

Table 2: RC car physical parameters

no lift forces or skin friction were taken into account. These weakness is accepted
because the motor parameters can compensate to some degree.

Proceeding as detailed in section 3.2.2 the parameters for the motor are found to be
the following:

Cm0 1.6584
C0 0.2226
C1 0.1829

Table 3: Motor parameter values

By running the simulation for the duty cycle percentage D corresponding to the
different voltage inputs in the tests, the motor’s behaviour can be compared in the
following graphs

41
4. Results

(a) Test with motor signal D=0.1515 (b) Test for motor signal D=0.2727

(c) Test with motor signal D=0.3939 (d) Test for motor signal D=0.5152

(e) Test for motor signal D=0.6364

Figure 20: Motor tests

A number of observations can be made. The result for the slowest test as seen in
figure 20a is poor. The middle tests presented in figures 20b, 20c, and 20d appear
to be quite accurate. The fifth test as shown in 20e is decently accurate for the first
second, the simulation continues to accelerate whereas the test data dips because it
reached the end of the testing course.

42
 4. Results

The results for the motor parameters are sufficient, but not completely accurate.
There are a number of errors that could be the cause for this. The most obvious
cause is the erratic behaviour of the raw test data. Attempting to tune the motor
parameters after the raw data would be unreasonable. Therefore the filtered data
has to be used. The laboratory in which the tests where performed had space
limitations, as can be noticed in how the tests time span decreases as the duty cycle
increases. This meant that the higher speeds could not be tested for. The method
for identifying the parameters also lacked any sort of time dependency. Motor
temperature, and battery wear are also possible factors that could affect results but
were not taken in to account.

The tyre parameters were determined as described in section 3.2.2 and found to be
the following values:

d 1.16
c 1.96
b 1.44

Table 4: Tyre parameter values

(a) Test 1 (b) Test 2

Figure 21: Comparison between circular trajectory for simulation and RC car at
different speed and radius

Figures 21a and 21b shows the trajectory for the simulation and RC car during a
test. The RC car is set to follow a circular path with an empirical P-controller.
The simulation runs in a circle with the mean control signals. For the first test the
results are very accurate with some deviation from slipping. In the second test the
results are decently accurate, but the trajectory of the car is wavy. The wavy path
is a result of poor control, but the circle has about the same radius. The results
indicate that the simulation’s behaviour is accurate at least for circular trajectories.
No tests for other slip and lateral movement were conducted because they are very
hard to compare and repeat with the available tools.

43
4. Results

4.2 Positioning and computer performance

The positioning and computer performance are critical for testing the systems in
reality. If either one performs badly the data will be skewed and controllers will
fail. In this project the systems performed adequately but not error-free and not all
cases could be tested.

Positioning

The positioning of a car worked well, both the position and the direction could be
established at 150Hz, although some calibration to match the light of the diodes had
to be done. The calibration had to be redone as the battery of the LEDs discharged
since the light intensity depend on the battery voltage.

When having two cars on the track, the positioning was not as satisfying. As the
intensity of light on both cars have to be similar to be able to calibrate properly.
The second car have one diode more and therefore the battery discharge faster and
thus changing intensity faster than the other car. Thus the intensity is rarely the
same, and it is therefore hard to detect both of the cars at the same time. In
addition, the extra diode on the second was placed close to the others, putting even
higher demands on the calibration for proper separation of dots. These problems
made driving with two (or more) cars unfeasible since the second car would falsely
identify as the first and thus skew data. As the focus of the project was elsewhere,
these issues were never resolved.

Computations

As the computations for each frame delivered from the camera, including calculating
new control signals, was done in sufficient pace, the system could in general keep
up with the frame rate of 150 fps. As the behaviour of the car depended on the
computer used, it seems reasonable that the delay between a frame arriving to new
control signals sent out depends on the computational speed of the computer. This is
reasonable as the system fully depends on reacting to the inputs it gets and a longer
reaction time will impact the performance. Although this can be compensated for,
it can not be removed unless some predictive control is used.

44
 4. Results

4.3 Track performance

The goal of the project was to develop stable controllers, for steering and speed, able
to drive the car around the track at relatively high speed. The mathematical model
of the car was used to make multiple PID controllers for steering. Speed control
is achieved using an improvised algorithm that uses the curvature of the oncoming
segments of the track to determine the controller output.

4.3.1 Steering control

With several linear models, multiple PID controllers were developed using the meth-
ods described in section 3.3.1. These controllers are only for steering. The obtained
parameters were saved in a look-up table, and tested in the simulation. Figure 22
shows the simulation results of the car driving around a track with a constant duty
cycle of D = 0.4.

Figure 22: A simulation of the car driving with the acquired PID parameters

From this simulation, it is apparent that the car can steer around the track with
a constant D. This result proves that the constructed PID controller, using the
look-up table, is in fact capable of steering the car to follow the track with good
stability.

To test the effect of the look-up table, an other test was conducted with one single
PID controller. The simulation results are shown figure 23

45
4. Results

Figure 23: A simulation of the car driving with one single PID controller

As shown in figure 23, the single PID controller was able to steer the car around the
track. However, the trajectory is not as smooth or close to the track as in Figure
22. These tests, as expected, prove the need for a look-up table to improve track
performance.

4.3.2 Speed control

The simulation results showed that the steering controller works in theory. To
improve overall track performance, however, the need for speed control is essential.
This would allow the car to accelerate on straights and brake before corners, which
improves lap times. This was implemented according to the methods in section
3.3.2.

As the project mainly focused on steering control, the PID controller for speed was
tuned manually. The most effective laptimes measure occurred using Kp = 0.52,
Ki = 0.37, Kd = 0. The distance from the cars position to the curve algorithms
measuring points was manually set at 0.3m with 0.225m long segments. The weights
were set at Wψ = 1.0 and Wcurve = 3.2.

Notably, the Kd parameter for the PID was set to zero. This was a result of noisy
speed measurement data, causing false actions with an added derivative gain.

An observation here was that the curve algorithm’s, Scurve , distance needed to be
manually adjusted during actual racing to be fully tuned, as it was speed depen-
dent. This could probably have improved speed handling significantly if tuned and

46
 4. Results

adjusted to the actual velocity of the car.

The Sψ algorithm, which based speed on the angular error of the car, was mostly
used to resume control over the car when slipping, and did so effectively.

The figure 24 clearly show the need to maintain high speeds when arriving at straight
segments, and slowing down before entering curves.

Figure 24: A comparison of the curve segment dependent adjustment of speed

along the map. The speed adjustment value is to be read as a percentage of target
speed in that instance.

47
4. Results

4.3.3 Real-time implementation

Simulating the car driving around the track by using a look-up table based control
method for steering control was a success. However the results of the real-time
implementation for the same look-up table is somewhat inconclusive and contents
several sources of error. These results are shown in figure 25 and figure 26.

Figure 25: Recorded results of one car following a virtual line with the look-up
table based PID controller. Observe that the car does manage to complete several
laps.

Table 5: Results of speed performance for the big look-up table

Mean velocity 1.021 m/s

Maximum velocity 1.527 m/s

48
 4. Results

(a) PID parameters during one lap.

(b) vx and control signal in [m/s].

(c) Control signal δ and error value of eψ (d) The observed lap in (a)-(c)

Figure 26: The different control signals and system responses during one lap when
controlled by a big look-up table. PID output speed is the signal sent from the PID
controller to the car measured in m/s.

The controllers used in figure 25 and figure 26 is the speed controller from results
in section 4.3.2, and the steering controller is the same as in figure 22, which is a
look-up table consisted of 744 sets of PID parameters. When using a great amount
of PID-parameters in this real-time implementation the result is not as successful.
They differ greatly in the ability to follow a line. The real-time implementation of
the large look-up table does manage to complete several laps, but the car oscillates
significantly around the reference line. Due to overshoots in almost every attempt
to turn. Since the speed controller only accelerates the car if the error value for
direction (eψ ) is moderate, it never succeeds to drive the RC car at any high velocity.
Hence several sources of error are present in the real-time implementation causing
the simulation and implementation results to differ.

Constructing another look-up table enabled satisfying results and the possibility to
further investigate the potential sources of error. It was made by identifying error

49
4. Results

sources when observing the systems behaviour with different PID controllers, thus
providing unique constrains to create a new look-up table. Following are the ob-
servations. The system had a high sensibility towards the derivative gain, which
approximately should not exceed 0.03. To minimise overshoot when steering, the
ratio between the proportional- and integral gain could not be greater than approx-
imately 1:2.5. The system was slower than earlier anticipated and the big look-up
table containing small incremented steps of vx and ω was unmanageable for the
system. In figure 26a it is clear that the PID parameters change very many times
during one lap. Thus, creating a delay affecting the reaction time and making the
system unable to control the momentary scenario. When the car starts to react
towards the new PID parameters it has already moved to a complete different sce-
nario. Combined with to many PID parameters for similar vx and ω the steering
control is unable to keep the car on the path.

Given this, the work points was chosen to make sure it corresponded to the observed
constrains for integral- and derivative gain. Resulting in 12 different sets of work
points representing 12 linear models with great stability. Also representing less
scenarios to prohibit the system from switching PID parameters before leaving the
corresponding work points. To represent a slower system 0.3 radians is subtracted
from the crossover frequency as input to the PID generating code in appendix 1,
decreasing the raise time to reach a steady state output. As well, the speed controller
was set to decelerate more before every turn. After these adjustments the results was
a great success and the validity of the assumed error sources was their by strengthen.
The smaller look-up table and the work points can be seen in appendix 2 and the
result in figure 27 and figure 28 .

50
 4. Results

Figure 27: Recorded results of one car following a virtual line, with a smaller look-
up table of only 12 PID controllers, by fine tuning which particular linear models
are suitable for driving the RC car.

Table 6: Results of speed performance for the smaller look-up table.

Mean velocity 1.219 m/s

Maximum velocity 1.581 m/s

51
4. Results

(a) PID parameters during one lap.

(b) vx and control signal in m/s.

(c) Control signal δ and error value eψ (d) The observed lap in (a)-(c)

Figure 28: The different control signals and system responses during one lap when
the system is controlled by a small Look-up table of 12 PID controllers.

The most obvious reason for this great improvement in steering control is best vi-
sualised when comparing figure 28a to figure 26a and figure 28b to figure 26b. Now
the system is given the chance to operate according to the momentary situation,
causing less oscillation. If switching PID parameters causes one of the most signif-
icant sources of error, it is possible to assume that using only one PID controller
could improve the performance even more. However, this is not correct. Similar to
the results when simulating in Simulink, one single PID controller could not operate
the system better than a look-up table. See results from using a single set of PID
parameters in figure 29. It is able to steer the RC car around the virtual track but
the trajectory is much worse compared to figure 27.

52
 4. Results

Figure 29: Real-time test result using only one single PID controller with param-
eters Kp = 0.5912 Ki = 1.119 Kd = 0.00713

53
4. Results

(a) PID parameters during one lap.

(b) vx and control signal in m/s.

(c) Unsaturated control signal δ and error

(d) The observed lap in (a)-(c)
value eψ

Figure 30: The different control signals and system responses during one lap when
the system is controlled by one PID controller with the same parameters as in figure
29.

It is very clear from 7 that the small look-up table had much better performance than
the big look-up table and the single PID controller. The lap times are roughly 2-3
seconds faster consistently. The single PID controller does reach a higher maximum
velocity than the small look-up table, but a lower mean. The higher maximum does
not mean much in terms of race performance since the overshoots it causes results
in slower lap times.

54
 4. Results

Big look-up table Small look-up table Single PID

Lap 1 [s]: 10.7961 7.4290 9.6511
Lap 2 [s]: 10.8713 7.5631 9.7738
Lap 3 [s]: 10.2807 7.4396 9.0610
Max velocity [m/s]: 1.527 1.581 1.62
Mean velocity [m/s]: 1.021 1.219 1.075

Table 7: Performance data for the tests

In the introduction it was stated that this project could contribute to understanding
difficulties in ongoing research about autonomous racing cars. Since the RC cars do
not reach any high speed, the amount of correlations between this project and real
scale autonomous racing are not many. However one of this system’s sources of error
could be counteracted with help of a more predictive control method. The controller
seem to improve their accuracy when given the chance to look ahead. Which could
be the case even in real scale. When an autonomous racing car drives progressively
faster it would need to look increasingly further ahead. which could be problematic
since the further is hard to predict. This is most likely an area in need of research
to develop secure traffic flows in the further autonomous traffic.

4.3.4 Obstacle avoidance

Unfortunately the obstacle avoidance was not tested due to the tracking system hav-
ing problems separating the cars when they got close to each other. The algorithm
was however tested in simulations were made using Simulink, giving an indication
of how well it would work in reality. Results from the simulations are shown in
figure 31. The car comes from the bottom of the picture and drives in the positive
Y-direction.

55
4. Results

2.8
Car Trajectory
2.6 Planned Path
Target Points
2.4 obstacle
car
2.2
Y [m]

1.8

1.6

4 4.5 5 5.5
X [m]

Figure 31: A simulation of the obstacle avoidance. The red box marked as the car
is not to scale.

When looking at figure 31, it does look like the car crashes right into the obstacle.
But this is not necessarily the case since half of the width of the car is added in
every direction of an obstacle, as explained in section 3.5.1. Even though trajectory
of the car is a bit close to the obstacle in the first corner, it probably wont crash
but even if it doesn’t, the margin will be very small. This can be solved by adding
more safety margin to the area that is marked as obstructed.

As shown in figure 31 the track around the obstacle seems to be shifted a bit behind
the obstacle. Possible improvements to this could be to shift the obstacle avoidance
track back, before the obstacle, resulting in a better obstacle avoidance. The method
used could also be improved from the ground up since this method is a reworked
version of an earlier idea, partly to save computing power.

56
 4. Results

4.3.5 Areas of improvements and error sources

Identified error sources:

• Incorrect ratios between Ki and Kp .

• The big look-up table contained PID parameters constructed from unrealistic
work points.

• The slow reaction time of the system causing the controller to not use the right
PID parameters for the right situation.

• No anti-windup in the PID implementation

• Not using a MIMO supported controller method is most likely causing loss of
information that could have tuned the controllers further.

Whose effects could be decreased by following improvements:

The linear state space model is not a good enough representation of the RC
car. Since the linear controllers are created after the insufficient model, they could
not handle the real situation accurately. As seen in the simulation results in figure
22 and results of the real time implementation in figure 25, the controllers are well
adapted to the model but not to the RC cars. Nevertheless, it was possible to create
a good look-up table for the RC car by observing the error sources of the system.
Several of them, for example sometime unrealistic ratio between PID parameters
are a result of an insufficient linear model. It is also important to note that using
work points representing only stable models and realistic scenarios complicates the
routine of generating more than 700 PID parameters. Overlooking the work points
is necessary with this method.

Reaction time of the system As mentioned, the system switches PID parameters
to quick compared to the reaction time of the car.

The image sent out from the camera is a matrix of pixels and the position of each
LED is calculated simply by using the brightest pixel that is showing the LED.
The calculated position of each LED is therefore rarely matching the actual exact
position. This impacts the readings of the velocity, as the calculated position jumps
between discrete values. The speed reading may therefore oscillate, as the system
detects movements that are either smaller or larger than they actually are. Since
this happens for each LED, and position of the car is calculated by using at least
three, this is made even worse. Therefore, the velocity data was highly unreliable.

Anti-windup design, or rather lack of, is a weakness in the implementation that

is important to discuss. Anti-windup is necessary in any system with integral gain
and a limited actuator, to avoid signal build-up when the actuator is saturated. A
simple clamping algorithm was implemented to constrict the integral gain when the
steering was saturated, and other options were not considered.

57
4. Results

Predictive control methods could improve how the system chooses PID parame-
ters, if the systems speed could not be increased. Obviously many of those problems
could be solved by using more advanced control methods, like MPC which also
can handle MIMO systems. To manage it with linear methods, a more advanced
predictive method could be develop to predict curves and

Material, different RC cars and track surface could improve the stability of the sys-
tem. Using bigger RC cars with better motor parameters could simplify the linear
model. The track surface of this project was a rubber carpet but with a rather slip-
pery surface whenever dust accumulated. Using fabric or a more roughened rubber
carpet could help the RC cars drive faster, since it often slipped when accelerating.

Obstacle Avoidance, there is some problems here, the first and most obvious, it
was never implemented and therefore never properly tested. To implement this the
positioning algorithm needs improvement so that it does not lose track of the cars.
The margin to the area marked as obstructed is a bit small, this can be fixed by
adjusting the method or just by adding more safety margin to the obstruced area.

Summary of track performance

After considering the result from simulation in Matlab and the real-time implemen-
tation, it is obvious that a look-up table is needed to control a RC car when using
linear control methods. Since the car is never operating in a single steady state
when following a path, several linear models is necessary to mimic this nonlinear
scenario. However, the look-up table failed to make the RC car follow the line well
and since only one PID controller also failed to do so, the best compromise in this
project was to create a smaller look-up table by specifically choosing linear models
corresponding to observed sources of error.

Because of problems implementing obstacle avoidance, no tests were conducted, but

it was simulated and the RC car did avoid an obstacle which shows that the concept
works. However, it needs more work.

There are room for improvements of the system since the mean velocity is only
24.3% of the cars modelled maximum on the track used in this project. The velocity
is track dependant, and since this track is rather difficult with many sharp turns a
velocity of this sort is not unexpected without methods of for drifting or better grip.
However, creating a learning environment for autonomous RC racing is definitely
achieved. From this platform a few hypotheses can be made about the future need
of predictability in traffic.

58
 5
Conclusion

Developing a system to autonomously drive two small RC cars around a virtual

track, thus creating a platform for learning about autonomous racing, was the main
goal. This was partially achieved.

It is possible to drive one car around the virtual track with satisfying accuracy
when following a line but with somewhat inconclusive results regarding speed. The
steering control was an attempt to simulate a nonlinear system with multiple linear
ones and the simulation results was a great success. The performance of a large look-
up table over a wide span of work points was very poor compared to the significantly
better results when using a smaller table designed from choice work points, shown in
the lap times and speed table 7. Since the linear model does not represent the real
RC car well enough, the implementation of the controllers needed extra constraints
to fulfil the goals.

The system is not able to avoid obstacles or drive two cars simultaneous in the
real-time implementation, but the basis for development is done in simulation with
promising results.

This project was a great learning platform for autonomous cars in general, and
autonomous racing specifically. It has provided many insights on the limitations
and the possibilities on racing with autonomous cars. To further develop this system
another linear model is necessary, as well a more advanced predictive algorithm is
necessary to help the linear controllers predict the next scenario on the track, thus
increasing reaction time of the system.

Suggestions for future research

Due to difficulties grasping all the challenges when planning the project, the speed
controllers were not optimised. It also became apparent that the speed control
affected the car performance more than first estimated. This resulted in a control
system that while working, could be improved by implementing MIMO control.
Problems with implementing the look-up table and the accuracy of the linear model
also affected lap times negatively.

59
5. Conclusion

Since a nonlinear system is being controlled with linear control strategies, a lot more
emphasis has to be put on developing methods for state estimation, in other words
looking ahead to see what is most likely to happen next. How the next curve will
look and what will happen after that curve, this would likely improve the speed
control to a greater degree than the steering control.

60
 Bibliography

[1] K. Bimbraw. Autonomous cars: Past, present and future a review of the de-
velopments in the last century, the present scenario and the expected future
of autonomous vehicle technology. In 2015 12th International Conference on
Informatics in Control, Automation and Robotics (ICINCO), volume 01, pages
191–198, July 2015.
[2] Volvo cars. Drive me trial in Gothenburg. [Online] Available: [Link]
[Link]/intl/buy/explore/intellisafe/autonomous-driving,
[Accessed: 2018-04-16].
[3] P. Fernandes and U. Nunes. Platooning with ivc-enabled autonomous vehicles:
Strategies to mitigate communication delays, improve safety and traffic flow.
IEEE Transactions on Intelligent Transportation Systems, 13(1):91–106, March
2012.
[4] David Shepardson. Tesla says crashed vehicle had been on autopilot prior to
accident. [Online] Available: [Link]
crash/tesla-says-crashed-vehicle-had-been-on-autopilot-prior-to-
accident-idUSKBN1H7023, [Accessed: 2018-04-20].
[5] Hyundai. How racing car technology influenced the future of commercial vehi-
cles. [Online] Available: [Link]
racing-car-technology-influenced-the-future-of-commercial-
vehicles/, [Accessed: 2018-04-16].
[6] Stanford University dynamic design lab. Shelley, stanford’s autonomous car,
at thunderhill raceway. [Video: Online] Available: [Link]
edu/news/2012/august/videos/[Link], [Accessed: 2018-04-26].
[7] [Link]. Autonomous vehicle control at the limits of handling. Ph.d
dissertation, Stanford University, June, 2012.
[8] U. Rosolia, A. Carvalho, and F. Borrelli. Autonomous racing using learning
model predictive control. In Proceedings 2017 IFAC World Congress, 2017.
[9] F. Zhang, J. Gonzales, K. Li, and F. Borrelli. Autonomous drift cornering with
mixed open-loop and closed-loop control. In Proceedings IFAC World Congress,
2017.

61
Bibliography

[10] MPC Lab @ UC Berkely. Automotive driving. [Online] Available: http:

//[Link]/research/advanced-vehicle-dynamic-control,
[Accessed: 2018-04-26].
[11] Swiss Federal Institute of Technology Zürich (ETH). Autonomous RC Racing.
[Online] Available: [Link] [Accessed: 2018-
01-05].
[12] Uppsala University. Camera-based autonomous racing system. [Online] Avail-
able: [Link] [Accessed: 2018-04-20].
[13] J. Ek et a. Advanced vehicle control systems. Chalmers University of Technol-
ogy, 2015.
[14] E. Velenis, E. Frazzoli, and P. Tsiotras. On steady-state cornering equilibria
for wheeled vehicles with drift. In Proceedings of the 48h IEEE Conference
on Decision and Control (CDC) held jointly with 2009 28th Chinese Control
Conference, pages 3545–3550, Dec 2009.
[15] F. White. Fluid Mechanics, chapter 7. McGraw Hill, 7 edition, 2011.
[16] J. Liljestrand. State estimation of RC cars for the purpose of drift control.
Master’s thesis, Linköping University: Faculty of Science and Engineering,
Linköping, 2011.
[17] H. Pacejka. Tire and Vehicle Dynamics, chapter 3&4. Elsevier Science, 2012.
[18] Alexander Liniger, Alexander Domahidi, and Manfred Morari. Optimization-
based autonomous racing of 1:43 scale RC cars. Optimal Control Applications
and Methods, 36(5):628–647, 9 2015.
[19] Massachusetts Institute of Technology Open Courseware. Feedback control sys-
tems. [Online] Available: [Link]
astronautics/16-30-feedback-control-systems-fall-2010/lecture-
notes/MIT16_30F10_lec05.pdf, Accessed : 2018-05-11.
[20] K. J. Åström and R. M. Murray. Feedback Systems: An Introduction for Sci-
entists and Engineers. Princeton University Press, 2008. Available online at
[Link]
[21] [Link]. Reglerteknikens grunder, chapter 7&8. Studentlitteratur, 4:9
edition, 2002.
[22] Magnus Nilsson Björn Fridholm, Torsten Wik. Kalman filter for adaptive learn-
ing of look-up tables with application to automotive battery resistance estima-
tion. Control Engineering Practice, 48:78–86, 2016.
[23] [Link] [Link]. Robust tuning of pi and pid controllers: using
derivative action despite sensor noise. IEEE Control Systems, 26(1):55–69, 2
2006.
[24] Kyosho. Kyosho Mini Z MR03 Sports 2 Ferrari 360 GTC. [Online] Avail-

62
 Bibliography

able: [Link]
sports-2-ferrari-360-gtc, [Accessed: 2018-02-02].
[25] FLIR (EU). Flea3 1.3 mp mono usb3 vision (vita 1300). [On-
line] Available: [Link]
vita-1300-2-eu, [Accessed: 2018-02-02].
[26] M. Sakakibara et al. A back-illuminated global-shutter cmos image sensor with
pixel-parallel 14b subthreshold adc. In 2018 IEEE International Solid - State
Circuits Conference - (ISSCC), pages 80–82, Feb 2018.
[27] P. Jansson R. Grahn. Mekanik, pages 505–509,517–521. Studentlitteratur, 3
edition, 2013.
[28] Heriot Watt MSc Mechanical Engineer Educator Dimitrios Christoloukas. Rac-
ing Cars Technology - Aerodynamics. [Online] Available: [Link]
[Link]/aerodynamics, [Accessed: 2018-05-14].
[29] J. Katz. Race Car Aerodynamic - Designing for Speed, chapter Appendix.
Bentley Publishers, 2 edition, 1995.
[30] Youbin Peng, D. Vrancic, and R. Hanus. Anti-windup, bumpless, and condi-
tioned transfer techniques for pid controllers. IEEE Control Systems, 16(4),
Aug 1996.

63
Bibliography

64
 A
Appendix 1

A.1 Matlab code

A.1.1

These two following functions are from the book Reglerteknikens grunder by Bengt
Lennartsson. They contain minor changes to be compatible with present linear
model.

P Iopt

1 f u n c t i o n [ F , wc , phm , Ki ,MS,MT] = pi_opt (G, wc_min , wc_max ,

phm_min , phm_max) ;
2 %S e a r c h i n g f o r o p t i m a l PI r e g u l a t o r
3

4 Nwc=20; Nphm=10;
5 wc=[wc_min : ( wc_max−wc_min ) / (Nwc−1) : wc_max ] ;
6 phm=[phm_min : ( phm_max−phm_min) / (Nphm−1) : phm_max ] ;
7 [ F , Ki , Ti ] = pi_wc_phm(G, wc , phm) ;
8 MS=norm ( f e e d b a c k ( 1 ,G∗F) , i n f , 1 e −3) ;
9 MT=norm ( f e e d b a c k (G∗F , 1 ) , i n f , 1 e −3) ;
10 s t a b=norm ( f e e d b a c k (G∗F , 1 ) ) ;
11 [ Ki , i o p t ]=max( Ki ’ . ∗ ( MS<=1.7) . ∗ (MT<=1.3) . ∗ ( stab<i n f ) ) ; %MT
=1.3
12 F=F ( : , : , i o p t ) ; MS=MS( i o p t ) ; MT=MT( i o p t ) ;
13 [Am, phm , wpi , wc]= margin (G∗F) ;
14 end

P Iwc−phm

1 f u n c t i o n [ F , Ki , Ti ] = pi_wc_phm(G, wc , phm) ;
2 %C r e a t i n g m u l t i p l e p id r e g u l a t o r s

I
 A. Appendix 1

4 Nwc=l e n g t h ( wc ) ; Nphm=l e n g t h (phm) ;

5 f o r k=1:Nphm
6 f o r i =1:Nwc
7 j=i +(k−1)∗Nwc ;
8 [ gain_G , phase_G]= bode (G, wc ( i ) ) ;
9 %[ gain_G , phase_G]= bode (NUM,DEN, wc ( i ) ) ;
10 %gain_G = [ 1 1 0 0 0 0 ] ∗ gain_G . ’ ;
11 %phase_G = [ 1 1 0 0 0 0 ] ∗ phase_G . ’ ;
12 %[ ~ , indx ] = min ( phase_G−wc ( i ) ) ;
13 phase_F=−180+phm( k )−phase_G ;
14 Ti ( j )=c o t (−phase_F∗ p i /180) /wc ( i ) ;
15 Ki ( j )=wc ( i ) / ( gain_G∗ s q r t (1+( Ti ( j ) ∗wc ( i ) ) ^2) ) ;
16 F ( : , : , j )=t f ( [ Ki ( j ) ∗ Ti ( j ) Ki ( j ) ] , [ 1 0 ] ) ;
17 %c r e a t e f o r PIDparameters . . .
18 %.
19 %.
20 end
21 end

[21]

II
 A. Appendix 1

A.1.2

A.1.3 PID Generator loop

psi_siso is the name of the SISO-linear state space model used in this report.
1 f o r i =1:5
2 nr=s t r c a t ( ’ s y s ’ , num2str ( i ) ) ; %P s i _ s i s o
3

4 [ NUM_Delta , DEN_Delta]= s s 2 t f ( s . ( nr ) . A, s . ( nr ) . B, s . ( nr ) . C, s
. ( nr ) . D, 1 ) ; %Input D e l t a
5 p s i _ s i s o=t f (NUM_Delta ( 3 , : ) , DEN_Delta ) ;
6

7 [ Gpsi , Pm_psi , Wcg_psi , Wcp_psi]= margin ( p s i _ s i s o ) ;

8 PmF= [ 3 0 7 0 ] ;
9

10 [ F , wc , phm , Ki ,MS,MT]= pi_opt ( p s i _ s i s o , 0 . 7 ∗ Wcp_psi , 1 . 3 ∗

Wcp_psi ,PmF( 1 ) ,PmF( 2 ) ) ;
11 % Thinner i n t e r v a l l
12 [ F_opt , wc_opt , phm_opt , Ki_opt , MS_opt , MT_opt]= pi_opt (
p s i _ s i s o , 0 . 9 ∗ wc , 1 . 1 ∗ wc , phm−2,phm+2) ;
13

14 [Gm, Pm_feedback , Wcg, Wcp ] = margin ( p s i _ s i s o ∗F_opt ) ;

15 i f ( Ki==0)
16 Pm_feedback =50;
17 end
18

19 o p t s = p i d t u n e O p t i o n s ( ’ PhaseMargin ’ , Pm_feedback ) ; %’
DesignFocus ’ , ’ r e f e r e n c e −t r a c k i n g ’ ) ;
20 [C i n f o ] = p i d t u n e ( p s i _ s i s o , ’ PID ’ , wc_opt , o p t s ) ;
21 [ Kp Ki_pid Kd ] = p i d d a t a (C) ;
22

23

24 PID . ( nr )= [ Kp Ki_pid Kd wp( i , 4 ) wp( i , 6 ) ] ;

25 end

III
A. Appendix 1

IV
 B
Appendix 2

Here are the work points and PID parameters presented which are used to produce
the results in figure 27 and figure 28. They produce a very successful result for the
RC cars when following a virtual track.

Work points Look-up table

Work points PID parameters

States: X Y ψ vx vy ω Kp Ki Kd
wp1 0 0 0.1 0.1 0 0.3 0.00013009 2.6601e-09 0
wp2 0 0 0.1 0.6 0 0.8 1.2266 1.212 0.020638
wp3 0 0 0.1 0.9 0 0.3 1.2892 1.9909 0.0013805
wp4 0 0 0.1 0.9 0 0.8 0.53992 0.37601 0.013611
wp5 0 0 0.1 1 0 0.8 0.54906 0.43472 0.012175
wp6 0 0 0.1 1.2 0 0.8 0.56309 0.55411 0.0099844
wp7 0 0 0.1 1.3 0 3 0.59208 0.77443 0.010641
wp8 0 0 0.1 1.3 0 0.8 1.3148 3.0652 0
wp9 0 0 0.1 1.5 0 0.8 0.57026 0.66539 0
wp10 0 0 0.8 1.5 0 0.3 0.570356 0.64082 0
wp11 0 0 0.8 1.7 0 0.8 0.57444 0.73399 0
wp12 0 0 0.1 1.7 0 3 0.59122 1.1186 0.0071305