Real-time obstacle avoidance

for micro aerial vehicles in

2D and 3D narrow spaces.

Wim N. Pilkes
Layout: typeset by the author using LATEX.
Cover illustration: Unknown artist
 Real-time obstacle avoidance for
micro aerial vehicles in 2D and 3D
narrow spaces.

Wim N. Pilkes
11044632

Bachelor thesis
Credits: 18 EC

Bachelor Kunstmatige Intelligentie

University of Amsterdam
Faculty of Science
Science Park 904
1098 XH Amsterdam

Supervisor
Dr. A. Visser
Informatics Institute
Faculty of Science
University of Amsterdam
Science Park 904
1098 XH Amsterdam
Semester 2, 2022
Abstract
Modern Unmanned aerial vehicles (UAVs) have an increasingly wider range of ca-
pabilities. These drones are able to perform tasks that humans cannot do. For
instance, in the field of exploration of dangerous and hard to reach areas. As
a result of more advanced and compact hardware in addition to continuous ad-
vancements in automation, this field of research is expanding rapidly. Even though
completely autonomously flying drones are already utilized for an assortment of
applications, widespread use of true autonomous drone navigation has not yet been
adopted. True autonomy is even harder to achieve when the functioning of a UAV
is limited. This is a focus of research in micro aerial vehicles (MAVs). In this field,
drones are restricted in processing power and the amount of information that can
be observed about its environment. Therefore, efficient path planning algorithms
are necessary. This thesis aims to provide upon real-time obstacle avoidance in in-
door cluttered environments. This is done by comparing four different lightweight
local path planning algorithms. These are both 2D and 3D implementations of the
Virtual Force Field algorithm and the Vector Field Histogram algorithm. These
methods are tested and compared with a dataset of increasingly more difficult 2D
and 3D simulations.
Contents
Abstract
1 Introduction 1

2 Background 5
2.1 Static and Dynamic Enviroments . . . . . . . . . . . . . . . . . . . 5
2.2 Global and Local Path Planning . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Global path planning . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Local path planning . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Potential Field Methods . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Virtual Force Fields . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 Vector Field Histograms . . . . . . . . . . . . . . . . . . . . 9
2.3.3 Path Finding in 3D . . . . . . . . . . . . . . . . . . . . . . . 10

3 Method 12
3.1 Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1 The Crazyflie 2.1 Nano Quadcopter . . . . . . . . . . . . . . 13
3.1.2 AI-deck 1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.3 The Cyberzoo . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 The Simulated Environment . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 Dataset Generation . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2 Evaluation Methods . . . . . . . . . . . . . . . . . . . . . . 17

4 Results 18
4.1 Success Rate Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Distance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Time Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Further Observations . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5 Discussion 23
5.1 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

References 27
Chapter 1

Introduction
Autonomous robot navigation is a popular point of discussion. Recent advance-
ments in this field make it an intriguing area of research. This is shown through a
variety of novel applications in unmanned aerial vehicle (UAV) technology.

Fer example, drones are used to inspect locations that are inaccessible for hu-
mans. This has been applied in the Chernobyl nuclear power plant, where there
have been increasing concerns of nuclear waste leaking from one of the reactors
since its meltdown. Leaking nuclear waste would have devastating consequences
for the containment and clean-up efforts in and around the power plant. Because
of the lethal levels of nuclear radiation, humans cannot enter the reactor and asses
this damage. As a result, the first successful flight of a quadcopter was done in the
reactor in 20201 . This drone was able to record footage within the reactor core,
and it was determined that there was no leaking nuclear waste.

Another example of the usefulness of quadcopter drones is the inspection of

subway tunnel systems. This is already routinely being done in Tokyo2 and ex-
periments for a similar system are being done in Paris3 . It saves in manpower and
resources, such as the setup of scaffolding. Even more so, after a collapse of a
parking garage above a subway line in Boston4 , a drone was utilized to determine
the risk of collapse before inspectors were sent in. This eliminated the risk of
injury for these inspectors.

1
https://www.commercialdroneprofessional.com/inside-chernobyl-elios-2-drone-sent-into-
reactor-on-inspection-mission/
2
https://dronedj.com/2020/02/27/drones-inspect-tokyo-subway-tunnels/
3
https://dronedj.com/2022/06/07/paris-metro-looks-to-modernize-with-drone-inspections/
4
https://gcn.com/public-safety/2022/04/drone-inspects-subway-tunnels-after-nearby-
building-collapse/364049/

1
 Flying drones are also used for science. For example, the European Space
Agency (ESA) is using a drone to explore cave systems that humans cannot reach.
These caves are either too narrow or deemed to dangerous for humans. The ESA
aims to use this technology to do autonomous exploration and research in the fu-
ture on foreign planets5 .

These examples demonstrate the importance of quadcopter technology. It does

not only show how business processes and science can be improved, but also how
this technology helps to reduce risk and possibly save lives. However, these drones
are more often than not still being operated by a human. Completely autonomous
systems are being developed and tested, but are currently rarely used in full scale
commercial operations. Furthermore, these examples demonstrate that there is a
need for completely autonomously navigating drones to enter locations that hu-
mans cannot reach.

Given this wide range of application of drone technology, the field of robotics
has a large assortment of different navigation algorithms for numerous purposes.
This is due to a number of factors, for example, different types of drone design,
different working environments and different drone observation methods (LaValle,
2006, Chapter 1). Other constrains, such as limits in processing power also influ-
ence how path navigation is implemented. Nevertheless, robot navigation can be
generalized as follows: a robot needs to move from its current position to a goal
position as efficient as possible, without hitting any obstacles in its path. To do
so, the robot makes observations about its environment and runs this through an
algorithm. This results into one or more paths, of which the path that is consid-
ered the best is chosen and executed.

With the steady reduction in chip size and the advancements in robotics and
navigation algorithms in the last few years, it is possible to create increasingly
smaller and more capable self-processing robots. The International Micro Air Ve-
hicle Conference and Competition (IMAV)6 is a yearly event that aims to advance
micro air vehicles (MAVs). This conference encompasses both new drone designs
and improvements in drone control and path planning. The scientific conference is
combined with a set of competitions to test new technologies. One of these com-
petitions is the nanocopter AI challenge 7 . This challenge achieves to find solutions
for more efficient autonomously flying drones and better path planning and obsta-
cle avoidance in small indoor environments. In the challenge, a nano quadcopter
5
https://www.esa.int/Science_Exploration/Human_and_Robotic_Exploration/CAVES_and_Pangaea/Explorin
6
https://2022.imavs.org/
7
https://2022.imavs.org/index.php/competition/nanocopter-ai-challenge/

2
drone has to navigate an obstacle zone and fly trough a goal as fast as possible,
completely autonomously. To do so, the competition sets out a framework for
the competitors. This consists of a small indoor area, cluttered with a number
of differently shaped obstacles and a square goal. Furthermore, a state of the art
open source MAV by Bitcraze8 is selected, which is able to do real-time image
processing, path planning and decision making. A programmable simulation of
this framework is also provided. This framework is not only suitable for research
in UAVs in indoor enviroments, it enables researchers to compare their implemen-
tations9 .

Another focus of this challenge is to determine capabilities of autonomous nav-

igation under limited circumstances. First of all, the drone is limited to on-board
processing. Even though the drone has a state of the art processor designed specif-
ically to do complex machine learning, a path navigation algorithm is required to
be lightweight enough to compute in real-time. Furthermore, the drone will be
limited in what observations it can do. The only inputs that the drone will be
given is a monocular monochrome front-facing camera attached to the drone and
the data from its internal gyroscope. No other inputs, such as a GPS system or
external cameras are allowed. Therefore a perfect mapping of the drones surround-
ings is not possible. This limits the range of path navigation implementations to
lightweight local path planning algorithms (Zhang et al., 2018).

Considering these limitations, two lightweight local path planning algorithms

are implemented in the IMAV simulation. These are called the virtual force field
(Borenstein and Koren, 1989) and the vector field histogram (Borenstein and Ko-
ren, 1990). Given the real world environment of the IMAV framework, novel
three-dimensional variations of the VFF and the VFH algorithms will be devel-
oped and tested. This thesis aims to improve upon low computing power path
planning capabilities for nanocopter drones in a real world environment. It does
this by answering the following research question: "How well do 2D and 3D im-
plementations of the VFF and VFH algorithms perform compare to each other in
a narrow cluttered environment?"

The structure of the paper is organized as follows. At first, current capabilities

will be determined and relevant work will be summarized. The two different algo-
rithms will be implemented in a simulation. These will then be transformed into
two novel algorithms that operate in three dimensions. Furthermore, a dataset
8
https://www.bitcraze.io/
9
The UvA drone team will participate in this competition in September 2022 in Delft, the
Netherlands. This thesis is part of the preliminary research.

3
of increasingly difficult simulations will be set out to conduct experiments. This
will result in a quantification of the efficiency of the different algorithms, which
will be compared to draw conclusions on the performance and limitations of the
algorithms.

4
Chapter 2

Background
In this chapter a theoretical overview will be given of relevant literature. This
will serve as the foundation for the research that is presented. First, general the-
ory about path planning is summarized. Secondly, fundamental theory about path
planning algorithms is explained. Furthermore, research about relevant algorithms
will be described. Finally, two new implementations of algorithms will be theo-
rized, which are able to operate in three dimensions.

2.1 Static and Dynamic Enviroments

Buniyamin et al. (2011) make a distinction between static environments and dy-
namic environments. An environment is considered static when obstacles do not
move. To the contrary, an environment is considered dynamic when the environ-
ment is not fully explored or when obstacles move in an unpredictable manner.
This is an important factor to determine when a path planning algorithm is cho-
sen, considering that a robot should be able to behave efficiently to its environment.

2.2 Global and Local Path Planning

Robot path planning algorithms are categorized in two groups, global path plan-
ning and local path planning (Zhang et al., 2018). The difference between these
two categories is defined by whether the algorithm has complete information about
its environment. Therefore, the amount of observation information given to a robot
will determine what category of path planning is suitable.

5
2.2.1 Global path planning
Global path planning requires a complete mapping of an environment before exe-
cution of the algorithm. This information is loaded into the algorithm and a full
path from start to goal is calculated (Giesbrecht, 2004). Then, the robot is in-
structed to follow this path. If at least one path between a start and a goal exists,
this category of path planning will guarantee to find a successful path.

However, global path planning algorithms cannot function in unknown en-

vironments. Furthermore, these algorithms have difficulty adapting to dynamic
environments. Only when trajectories of obstacles are known beforehand, then
global path planning algorithms can be designed to adapt to this. When this is
not known, these algorithms require a completely new path evaluation for each
change in an environment (Raja and Pugazhenthi, 2012).

2.2.2 Local path planning

Local path planning, or real-time obstacle avoidance, does not require complete
information about the environment. A local path planning algorithm uses sensory
data to make observations about a segment of the environment, localized around
the robot. This is executed in real-time. As a result, local path planning is able to
navigate in unknown and changing environments. Furthermore, it is able to avoid
dynamic obstacles (Khaksar et al., 2015). In general, when no obstacle is detected,
a local path planning attempts to move towards a predetermined goal in a straight
line. It will deviate from this line to avoid a detected obstacle (Buniyamin et al.,
2011).

However, this category of path planning has certain limitations. First of all,
the incompleteness of information can make it difficult to identify a goal. If the
robot cannot observe a goal, the robot cannot determine where to move towards.
Thus failing the planning algorithm. Second of all, the algorithm is prone to local
minima . A local minimum is a situation in which a robot is trapped in a position
(Borenstein and Koren, 1989). This happens when a robot encounters one or more
obstacles that are configured in a manner that no viable path can be found. In
such a situation, any movement forwards or sideways will move the robot too close
to an obstacle, resulting in the robot moving back to the local minimum. Any
movement backwards will result in the robot moving further away from the goal
and free itself from the trap. It will then continue moving forwards towards the
goal and re-enter the trap. Raja and Pugazhenthi (2012) reviewed a number of
different approaches to detect and exit local minima.

6
2.3 Potential Field Methods

Figure 2.1: A fully calculated potential field. Each vector represents the
potential at its corresponding position. These potentials are calculated
by the repulsing force generated by the obstacle (red) and the attracting
force generated by the goal (green). The robot starts at the bottom left.
For each position, the corresponding potential is applied to the robot.
Resulting in a path (blue) towards the goal that avoids the obstacle.

Potential field methods (PFMs) execute path navigation by considering a robot as

a particle that can be influenced by a potential field (Sabudin et al., 2016). This
field is determined by repulsing potentials and an attracting potential. Within this
field, obstacles are emitting repulsing potentials and the goal is emitting an at-
tracting potential. This potential will move the robot. Moreover, the motion and
direction of the drone is calculated by the net force of this potential on any given
moment, see Figure 2.1. This results in a path in which the drone is attracted
towards the goal, whilst being repulsed by obstacles. Khatib (1985) calculated the
artificial potential for each position a robot can have within an environment. This
created a global path planning algorithm for a static environment. This makes it

7
unsuitable for real-time autonomous obstacle avoidance.

However, this method can be transformed in a local path planning algorithm by

only considering obstacles in the direct proximity of a robot and by continuously
recalculating the potential field with only these obstacles and the goal. Therefore,
only calculating the potential at the exact location of the robot. This results in
algorithms such as, the virtual force field and the vector field histogram.

2.3.1 Virtual Force Fields

First used by Borenstein and Koren (1989), a virtual force field (VFF) is a local
path planning algorithm. This method builds upon the concept of attracting and
repulsing forces by reducing the potential field to a small area, centered around
the robot. Because the robot moves, and this field moves with it, the field is con-
stantly updated. This allows the robot to work with incomplete knowledge of the
environment, making it a local path planning algorithm. Furthermore, since the
field is constantly recalculated, the robot can operate in a dynamic environment.

This method is further improved by transforming the potential field into a cer-
tainty grid (Moravec, 1987). Here, each cell within the grid has a certainty value.
As described by Moravec and Elfes (1985) and by Elfes (1986), for each cell, the
certainty value is defined as the probability that there is an object in the cell.
These probabilities are estimated by the observations of the robot, such as sonar
data. Each cell within the grid is converted into a force vector. This vector is
proportional to the certainty value. Therefore, cells with a high certainty value
result in a stronger force. Furthermore, the vector is inversely proportional to the
distance, cells closer to the robot result in stronger force. This makes the obstacle
avoidance more sensitive to objects close to the robot . The strength and direction
towards the robot of each force are added together. This results in a net force that
acts upon the drone. The force vector F for each cell (i, , j) is defined as follows
(Borenstein and Koren, 1988):

 
Fcr C(i, j) xi − xo yj − yo
F (i, j) = 2 x̂ + ŷ (2.1)
d (i, j) d(i, j) d(i, j)
Here, Fcr is the force constant, a hyper-parameter to adjust the sensitivity of
the repelling force. C(i, j) is the certainty value of the cell. d(i.j) is the euclidean
distance between the robot and the cell. xi and yj are the coordinates of the cell,
and x0 and y0 are the coordinates of the robot. The sum of these forces result in
a single force that acts upon the robot:

8
 X
Fr = F (i.j) (2.2)
i,j

This repulsive force will ensure that the robot will avoid obstacles, but will not
steer the robot towards a desired goal. Thus, an attracting force is necessary. This
is defined as a single cell of which the coordinates are known. The attracting force
Ft can be calculated similarly.
 
xt − xo yt − yo
Ft = Fct x̂ + ŷ (2.3)
d(t) d(t)
Here, Fct is the force constant, dt is the distance between the goal and the
robot, and xt and yt are the coordinates of the robot. The sum of the repulsing
force and the attracting force result in a net force vector R that acts upon the
robot:
R = Ft + Fr (2.4)
For each iteration of the algorithm this net force will be exerted upon the robot.

2.3.2 Vector Field Histograms

The VFF has a high level of data reduction. All the certainty values in the grid are
reduced in one step to a single vector and a lot of features about the environment
are lost (Koren et al., 1991). As a result, a robot using a VFF has issues with de-
tecting narrow passages or navigating through cluttered environments (Oroko and
Nyakoe, 2022). Borenstein and Koren (1990) propose to change the single-stage
data reduction into a two-stage data reduction to solve these problems, creating
the vector field histogram (VFH).

The first stage of the data reduction is similar to the first steps of the VFF.
A grid is defined around the robot and, using the robots observations, certainty
values are determined for each cell in the grid. However, the robot is set to do
rapid range readings of a 1 cell wide range, to reduce the amount of observations
and computations that are necessary. Each individual sample is used to increment
a certainty value corresponding to the position of the range reading. This creates
a one dimensional polar histogram, consisting of certainty based obstacle vectors.
Each vector c∗ (i, j) represents the set of range readings in that direction β(i, j).
Directions in which obstacles are being observed often, have an continuously in-
creasing magnitude m(i, j):

m(i, j) = (c∗ (i, j))2 ∗ [a − b ∗ d(i, j)] (2.5)

9
Here, a and b are positive constants. and d(i, j) is the distance from the cell to
the robot.

Furthermore, this histogram grid is not used to create a single repulsing force
vector. The grid is divided in k evenly spaced sectors around the drone. The
magnitudes of the cells within that sector is summed, creating a polar obstacle
density value for each sector hk :
X
hk = m(i, j) (2.6)
i,j

This results in a force vector for each sector. To reduce noise and improve
efficiency, this histogram grid is smoothed and only values above a predefined
threshold are considered.

The second stage of data reduction determines which sector is the optimal
path for the robot to navigate towards. To do so, it first creates a set of candidate
valleys. Which are sections within the histogram grid that are below the threshold.
These valleys can be one or more sections wide, were valleys with a low number
of sectors is considered as "narrow" and valleys with a high number of sectors is
considers "wide". The VFH algorithms then selects the valley that matches its
direction of travel the closest to the direction of the target the robot is intended
to move towards. Then, the angle of most optimal sector of this valley is the
new direction of the robot. This is the middle sector for both narrow and wide
valleys. However, when a wide valley is chosen that encounters an obstacle, the
robot is instructed to travel along that obstacle from a predetermined distance.
This reduces steering fluctuations (Borenstein et al., 1991).

2.3.3 Path Finding in 3D

The VFF and VFH as described above operate in a two dimensional environment.
However, some research has been done in potential field methods in three dimen-
sional environments. For example, Saravanakumar and Asokan (2013) apply a
three dimensional PFM to an autonomous underwater vehicle. This is done by
adding a height vector ẑ to the potential field.

This method be used to create a 3D VFF. A force vector for a cell, described
in equation (2.1), is defined as follows:
 
Fcr C(i, j, k) xi − xo yj − yo zj − zo
F (i, j, k) = 2 x̂ + ŷ + ẑ (2.7)
d (i, j, k) d(i, j, k) d(i, j, k) d(i, j, k)

10
 The total repulsive force, described in equation (2.2), is defined as follows:
X
Fr = F (i.j, k) (2.8)
i,j,k

When this repulsive force changes the height of the drone, a counteracting
force is needed to bring the drone back to the height of the goal. Therefore, the
attracting force, as described in (2.3) should also be transformed:
 
xt − xo yt − yo zt − zo
Ft = Fct x̂ + ŷ + ẑ (2.9)
d(t) d(t) d(t)

Given that both the repulsing and the attracting forces are three dimensional,
the resulting force, defined by equation (2.4) would also be in 3D.

Furthermore, this conversion into a 3D algorithm can also be done for the VFH.
Adding height to this algorithm would make polar histogram a 2D object. The
magnitude, defined in equation (2.5) and the polar obstacle density, defined in
(2.6) will require a height dimension k:

m(i, j, k) = (c∗ (i, j, k))2 ∗ [a − b ∗ d(i, j, k)] (2.10)

X
hk = m(i, j, k) (2.11)
i,j,k

This chapter has provided an overview of path planning algorithms, with a

focus on local path planning. Two different 2D local path planning algorithms were
described and two novel 3D implementations of these algorithms were theorized.

11
Chapter 3

Method
To be able to test the abilities of the algorithms described in the background, a
relevant environment is necessary. In this chapter, a real-time obstacle avoidance
competition is chosen as environment. This chapter describes this platform and
how the four local path planning algorithms are implement and evaluated.

3.1 Platform
The 2022 edition of the International Micro Air Vehicle Conference and Competi-
tion outlines a set of requirements for the IMAV Nanocopter AI challenge1 . It is
the goal of this challenge to be able to navigate a small obstacle zone as efficient as
possible without hitting the obstacles. Furthermore, this has to be accomplished
with real-time processing on a MAV that is only able to make observations through
an onboard camera. These factors make it a suitable testing environment for the
research presented in this thesis.

This challenge requires all participants to use identical hardware. This con-
sists of the AI deck 1.1 attached to the Crazyflie 2.1 nano quadcopter, both are
produced by Bitcraze. The competition will be held in an arena named the "Cy-
berzoo". An accurate simulation of the nanocopter and the cyberzoo is provided
in the webots 3D robot simulator.

1
https://2022.imavs.org/index.php/competition/nanocopter-ai-challenge/

12
 Figure 3.2: The AI-deck 1.1.
Figure 3.1: The Crazyflie 2.1 nano
quadcopter.

3.1.1 The Crazyflie 2.1 Nano Quadcopter

The Crazyflie 2.1 is an open source flying development platform designed and
produced by Bitcraze, see Figure 3.1. This platform allows for efficient devel-
opment and testing of autonomous drone software. The platform consist of an
control board, a small LiPo battery, 4 DC coreless motors with motor mounts
and propellers. This platform is designed to support expansion decks for further
functionality, such as the AI-deck 1.1.

3.1.2 AI-deck 1.1

Considering that the drone is limited in size and required to run image processing
and path navigation algorithms without off-board processing, a lightweight and
powerful processor is required. Palossi et al. (2019) used the GreenWaves GAP8
IoT application processor 2 for this purpose. The boards 8 cores and 1 micro-
processor produce a processing power of 22.65GOP S at a power consumption of
4.24mW/GOP 3 . Furthermore, the processors architecture is specifically designed
to do efficient machine learning. For example, it contains a Hardware Convolution
Engine (HWCE), a convolutional neural network accelerator. This processor was
mounted on a Crazyflie 2.1. To do so, a shield was created that connects the
processor to the Crazyflie header pins, allowing it to be attached above or under
2
https://greenwaves-technologies.com/gap8m cua i/
3
https://greenwaves-technologies.com/wp-content/uploads/2021/04/Product-Brief-GAP8-
V1_9.pdf

13
the Crazyflie. In addition, a camera was added to the shield.

This prototype resulted in the AI-deck 1.1 4 , a commercially available platform

by Bitcraze, see Figure 3.2. The deck has an 320x320 pixel monochromatic camera,
directly attached to the processor.

3.1.3 The Cyberzoo

To be able to test the nanocopter in a obstacle filled space, an 8 meter by 8 meter
arena is provided by the competition. This obstacle zone contains a yellow square
goal and a set of cylindrical poles and grey rectangular panels. The Crazyflie is
placed within this zone, in which it will fly autonomously towards the goal whilst
avoiding the obstacles.

3.2 The Simulated Environment

Figure 3.3: An example of the Webots Cyberzoo simulation. The Crazyflie

nanocopter is showed in the center of the simulated environment. On
the bottom-left the real-time feed from the camera in the nanocopter is
displayed. On the bottom-right the real-time feed from the depth sensor
in the nanocopter is displayed. A horizontal slice from the middle of this
feed is used to create a representation of the depth readings of the different
objects (upper-right).

A complete simulation of the Cyberzoo in the Webots simulation software is pro-

vided by the competition, see Figure 3.3. This simulates the arena in a physics
4
https://store.bitcraze.io/products/ai-deck-1-1

14
engine. The objects and goal are also simulated. Furthermore, the Crazyflie
quadcopter is also simulated, using a built-in PID controller for stabilization. Fur-
thermore, the quadcopter simulates a front facing 320x320 pixel camera, identical
to the real-world Crazyflie.

The simulation is altered to be able to accommodate the path planning al-

gorithms. First of all, the camera is converted into a depth sensor. Instead of
directly observing its environment, the quadcopter observes a distance value for
each pixel. Second of all, the distance and angle between the goal and the drone
is calculated. The algorithms in this thesis require these observations to do path
planning. Implementing a depth estimation and a goal detection algorithm is be-
yond the scope of this thesis.

3.2.1 Dataset Generation

To be able to create a significant sample size, more than one simulated environ-
ment is necessary. Therefore, the simulation is modified into a set of randomly
generated simulations. These simulations have increasingly more obstacles, result-
ing in different difficulty levels. Furthermore, a dataset is generated to simulate a
2D environment and a dataset is generated to simulated a 3D environment. This
allows for comparison between the VFF and VFH and for comparison between the
2D and the 3D implementations of these algorithms.

For each simulation, the Crazyflie is positioned randomly on one end of the
arena, and the goal is positioned randomly on the other side of the simulation.
This avoids simulations in which the goal and drone are placed in close proxim-
ity to each other. Therefore, the drone is forced to avoid obstacles to reach the goal.

15
 Figure 3.4: Objects used in the Cyberzoo simulation: a cylindrical pole
(left), a square gate (middle) and a rectangular grey panel (right).

Next, the obstacles are generated at random positions between the quadcopter
and the goal. As seen in Figure 3.4, two types of obstacles are simulated, a
cylindrical pole and a rectangular panel. A total of 50 simulations are generated
with five difficulty levels. The first set of generated simulations will have two poles
and two grey panels. After every tenth simulation, two more poles and two more
panels will be generated, resulting in the five difficulty levels. Another dataset of
50 simulations will be made with these same requirements, but half the panels are
floating 1.5 meters above the ground. This is to test the capabilities of the 3D
algorithms. Resulting in the following overview:

2D Cyberzoo simulations, 10 for each difficulty

Difficulty Number of Poles Number of Panels Number of Panels

(Normal) (Floating)
1 2 1 1
1 2 2 0
2 4 4 0
3 6 6 0
4 8 8 0
5 10 10 0

16
 3D Cyberzoo simulations, 10 for each difficulty

Difficulty Number of Poles Number of Panels Number of Panels

(Normal) (Floating)
1 2 1 1
2 4 2 2
3 6 3 3
4 8 4 4
5 10 5 5

3.2.2 Evaluation Methods

This dataset is generated to test multiple path planning algorithms. However, a
suitable metric is required to be able to evaluate these algorithms. Therefore the
simulation logs information about each simulation run for each algorithm.

First of all, whether or not the drone reaches the goal is determined. This is
a binary value, 0 means that the path planning failed and 1 means that the goal
is reached. If the drone is trapped in a local minimum, then the simulation is
considered as failed. This is used to determine the reliability of each algorithm.
Furthermore, the simulation is considered as a fail after 120 seconds of runtime.
Second of all, the total distance travelled by the robot is tracked for each simula-
tion. And finally, the total runtime for each simulation is tracked. The distance
and time are used to determine the efficiency of the path planning of each algo-
rithm. Given that failed simulations can generate outliers in the distance and time
values, only successful simulations are used.

The four local path planing algorithms are implemented in python as stated in
the chapter background. Parameter optimization was done empirically.

17
Chapter 4

Results
In this chapter, the results of this thesis are presented. These results are the out-
come of 100 simulations for each of the four different real-time obstacle avoidance
algorithms. 50 of those simulations were done in 2D and 50 simulations were done
in 3D. Both these datasets were further divided in subsets of 10 simulations with
an increasing amount of obstacles. Each simulation resulted in three datapoints,
whether the simulation was successful, how much distance the drone covered, and
how much time it took. For the distance and time values only successful simula-
tions were considered. These three types of results will be evaluated separately.

18
4.1 Success Rate Evaluation

Figure 4.1: The succes rates of the different algorithms. With the 2D
VFF algorithm on the top left, the 2D VFH algorithm on the top right,
the 3D VFF algorithm on the bottom left and the 3D VFH algorithm on
the bottom right.

As seen in Figure 4.1, a relation can be observed between the success rate and the
difficulty level for all four the algorithms. An increase in the amount of obstacles
result in a decrease of success rate. However, a noticeable drop-off in performance
is visible in the 2D implementations at difficulty level 4 and 5. Furthermore, the
VFH appear to perform better, maintain a success rate of more than 0.5 for each
set of simulations.

19
4.2 Distance Evaluation

Figure 4.2: The average distance covered by the drone for each successful
simulation of the different algorithms. With the 2D VFF algorithm on the
top left, the 2D VFH algorithm on the top right, the 3D VFF algorithm
on the bottom left and the 3D VFH algorithm on the bottom right.

Figure 4.2 shows that the distance travelled by the different path navigation al-
gorithms are similar. Therefore, it is difficult to evaluate and compare these algo-
rithms.

20
4.3 Time Evaluation

Figure 4.3: The average runtime in seconds for each successful simulation
of the different algorithms. With the 2D VFF algorithm on the top left,
the 2D VFH algorithm on the top right, the 3D VFF algorithm on the
bottom left and the 3D VFH algorithm on the bottom right.

As shown in Figure 4.3, a relation can be observed between the runtime and the
difficulty level for all four the algorithms. In which simulations with more obstacles
have a higher runtime. Particularly, the 2D VFF takes a long time to complete
simulations of difficulty level 5. However, both the VFF algorithms appear to have
a more consistent relation between time and difficulty. Whilst the VFH algorithms
seem to behave more irregular.

21
4.4 Further Observations
During the running of the algorithms some notable observations were made. First
of all, the VFH algorithms often went through narrow passages, whilst the VFF
algorithms would not. Second of all, the 3D algorithms were able to fly over and
under the obstacles in the 3D simulations, creating a easier path to the goal.

Although, the results show differences between the performances of the path
planning algorithms, there is a significant amount of overlap of results. Therefore,
it is not possible to conclusively quantify how the algorithms perform in contrast
to each other.

22
Chapter 5

Discussion
In this chapter, the research done in this thesis will be summarized. The results
will be evaluated to answer the research question. Furthermore, future work will
be discussed.

5.1 Evaluation
This thesis aimed to answer the following research question: "How well do 2D and
3D implementations of the VFF and VFH algorithms perform compare to each
other in a narrow cluttered environment?"

Even though, this thesis was not able to quantify this answer, some notable
observations were done. First of all, the VHF has a more consistent success rate
than the VFF. Second of all, the implementation of 3D path planning algorithms
will allow a drone to more successfully navigate environments with a high amount
of obstacles. Furthermore, as described by Koren et al. (1991) the VFH is able to
successfully navigate through narrow spaces, which the VFF cannot.

5.2 Future work

A number of recommendations can be made to improve upon this thesis. First
of all, the drone simulation can be improved. A real-time local goal detection
algorithm could be implemented. If, for example, the goal would be behind an
obstacle, a method would have to be created to first search for the obstacle. This
would result in an algorithm that can operate in even more circumstances. Ad-
ditionally, the grayscale camera can be used as input for the local path planning
algorithms. To do so, a depth estimation algorithm will have to be implemented,

23
instead of using the perfect depth camera in the simulation. This could be used to
examine how the algorithms would perform with imperfect input. Furthermore,
local goal detection and depth estimation would make the whole system completely
autonomous and would allow the algorithm in real-life situations. Another prob-
lem that could be improved upon, is that local path planning algorithms are prone
to local minima. A variety of different local minimum detection algorithms exist
(Raja and Pugazhenthi, 2012). Implementing such an algorithm could further im-
prove the capabilities of the path planning.

Secondly, a larger dataset can be generated to find more conclusive results.

The results presented in this thesis had a significant amount of overlap. More sim-
ulations under a wider variety of different environments could result in a decisive
evaluation.
Furthermore, it would be interesting how global path planning would compare
with local path planning within this environment. Because this experiment is done
in a simulation, this comparison is possible. This would not only allow conclusions
to be made between different local path planning algorithms under different condi-
tions, but it could also draw conclusions about the efficiency of local path planning
in general.

Finally, the algorithms can be tested outside of the simulation, for instance in
the real-world environment of the IMAVs competition in Delft. This would deter-
mine whether the simulation is applicable in real-life. Furthermore, other narrow
environments, such as metro tunnels or cave systems could be tested on.

24
5.3 Conclusion
Local path planning is designed to operate in a variety of real-world implementa-
tions. It can be used on a variety of robots utilizing a range of observation methods
in dynamic environments. Furthermore, creating a system with severe constrains
demonstrate the strength and versatility of local path planning. The limitations
within this thesis were inspired by the nanocopter AI challenge by IMAV. This
challenge was described as follows, a MAV was used with an on-board processor
and a small greyscale camera. This drone had to navigate an small indoor obstacle
space, filled with obstacles. The aim was to fly the drone through the obstacle
zone and reach a goal.

This thesis compared four local path planning algorithms. The aim was to
determine how these compared under different conditions. To do so, a simula-
tion of the real-life framework by IMAV was used. Two datasets of increasingly
more cluttered simulation worlds were created. One dataset that described a 2D
environment and one in 3D. The algorithms were run on these datasets and the
performance was evaluated.

25
References
Borenstein, J. and Koren, Y. (1988), High-speed obstacle avoidance for mobile
robots, in ‘Proceedings of the 3rd International Symposium on Intelligent Con-
trol’, pp. 382–384.

Borenstein, J. and Koren, Y. (1989), ‘Real-time obstacle avoidance for fast mobile
robots’, IEEE Transactions on systems, Man, and Cybernetics 19(5), 1179–
1187.

Borenstein, J. and Koren, Y. (1990), Real-time obstacle avoidance for fast mobile
robots in cluttered environments, in ‘Proceedings., IEEE International Confer-
ence on Robotics and Automation’, IEEE, pp. 572–577.

Borenstein, J., Koren, Y. et al. (1991), ‘The vector field histogram-fast obstacle
avoidance for mobile robots’, IEEE transactions on robotics and automation
7(3), 278–288.

Buniyamin, N., Ngah, W. W., Sariff, N., Mohamad, Z. et al. (2011), ‘A simple local
path planning algorithm for autonomous mobile robots’, International journal
of systems applications, Engineering & development 5(2), 151–159.

Elfes, A. (1986), A sonar-based mapping and navigation system, in ‘Proceed-

ings. 1986 IEEE International Conference on Robotics and Automation’, Vol. 3,
IEEE, pp. 1151–1156.

Giesbrecht, J. (2004), Global path planning for unmanned ground vehicles, Techni-
cal report, DEFENCE RESEARCH AND DEVELOPMENT SUFFIELD (AL-
BERTA).

Khaksar, W., Vivekananthen, S., Saharia, K. S. M., Yousefi, M. and Ismail, F. B.

(2015), A review on mobile robots motion path planning in unknown environ-
ments, in ‘2015 IEEE International Symposium on Robotics and Intelligent Sen-
sors (IRIS)’, IEEE, pp. 295–300.

26
Khatib, O. (1985), Real-time obstacle avoidance for manipulators and mobile
robots, in ‘Proceedings. 1985 IEEE International Conference on Robotics and
Automation’, Vol. 2, IEEE, pp. 500–505.

Koren, Y., Borenstein, J. et al. (1991), Potential field methods and their inherent
limitations for mobile robot navigation., in ‘ICRA’, Vol. 2, pp. 1398–1404.

LaValle, S. M. (2006), Planning algorithms, Cambridge university press.

Moravec, H. and Elfes, A. (1985), High resolution maps from wide angle sonar, in
‘Proceedings. 1985 IEEE international conference on robotics and automation’,
Vol. 2, IEEE, pp. 116–121.

Moravec, H. P. (1987), Certainty grids for mobile robots, in ‘Jet Propulsion Lab.,
California Inst. of Tech., Proceedings of the Workshop on Space Telerobotics,
Volume 1’.

Oroko, J. A. and Nyakoe, G. (2022), Obstacle avoidance and path planning

schemes for autonomous navigation of a mobile robot: a review, in ‘Proceedings
of the Sustainable Research and Innovation Conference’, pp. 314–318.

Palossi, D., Conti, F. and Benini, L. (2019), An open source and open hardware
deep learning-powered visual navigation engine for autonomous nano-uavs, in
‘2019 15th International Conference on Distributed Computing in Sensor Sys-
tems (DCOSS)’, IEEE, pp. 604–611.

Raja, P. and Pugazhenthi, S. (2012), ‘Optimal path planning of mobile robots: A

review’, International journal of physical sciences 7(9), 1314–1320.

Sabudin, E., Omar, R., CKAN, H. and Melor, C. K. (2016), ‘Potential field meth-
ods and their inherent approaches for path planning’, ARPN J. Eng. Appl. Sci
11(18), 10801–10805.

Saravanakumar, S. and Asokan, T. (2013), ‘Multipoint potential field method

for path planning of autonomous underwater vehicles in 3d space’, Intelligent
Service Robotics 6(4), 211–224.

Zhang, H.-y., Lin, W.-m. and Chen, A.-x. (2018), ‘Path planning for the mobile
robot: A review’, Symmetry 10(10), 450.

27