Vision-Based Control Using Object Detection and Depth

Estimation for Robotic Pick and Place Tasks in

Construction Applications

by

Sushilkumar Muralikumar

A Thesis Presented in Partial Fulfillment

of the Requirements for the Degree
Master of Science

Approved August 2022 by the

Graduate Supervisory Committee:

Spring Berman, Chair

Hamid Marvi
Hyunglae Lee

ARIZONA STATE UNIVERSITY

December 2022
 ABSTRACT

The construction industry holds great promise for improvement through the use of

robotic technologies in its workflow. Although this industry was an early adopter of such

technologies, growth in construction robotics research and its integration into current

construction projects is progressing slowly. Some significant factors that have contributed

to the slow pace are high capital costs, low return on investments, and decreasing public

infrastructure budgets. Consequently, there is a clear need to reduce the overall costs

associated with new construction robotics technologies, which would enable greater

dissemination.

One solution is to use a swarm robotics approach, in which a large group of relatively

low-cost agents are employed to produce a target collective behavior. Given the

development of deep learning algorithms for object detection and depth estimation, and

novel technologies such as edge computing and augmented reality, it is becoming feasible

to engineer low-cost swarm robotic systems that use a vision-only control approach.

Toward this end, this thesis develops a vision-based controller for a mobile manipulator

robot that relies only on visual feedback from a monocular camera and does not require

prior information about the environment. The controller uses deep-learning based methods

for object detection and depth estimation to accomplish material retrieval and deposition

tasks. The controller is demonstrated in the Gazebo robot simulator for scenarios in which

a mobile manipulator must autonomously identify, pick up, transport, and deposit

individual blocks with specific colors and shapes. The thesis concludes with a discussion

of possible future extensions to the proposed solution, including its scalability to swarm

robotic systems.

i
 ACKNOWLEDGMENTS

Words cannot express my gratitude to my advisor, Dr. Spring Berman, for allowing me

to be part of the Autonomous and Collective Systems (ACS) lab and for her relentless

patience and invaluable feedback over the years. She has been an incredible source of

support and motivation, without which this thesis would not have come to fruition. I would

also like to express my gratitude to Dr. Hamid Marvi and Dr. Sze Zheng Yong for their

incredible classes on Robotics and Control and their constant support over the years.

I am incredibly grateful to members of the ACS Lab, especially Dr. Aniket Ravindra

Shirsat, Dr. Zahi Kakish, Dr. Amir Salimi Lafmejani, Dr. Azadeh Doruduchi, Shenbagaraj

Kannapiran, and Rakshith Subramanyam, for their invaluable support and feedback at

various points throughout my journey.

This endeavor would not have been possible without the support of my family and

friends, who have been a constant source of support through thick and thin.

ii
 TABLE OF CONTENTS

Page

LIST OF TABLES ...................................................................................................................vi

LIST OF FIGURES ............................................................................................................... vii

CHAPTER

1 INTRODUCTION .................................................................................................. 1

1.1 Literature Survey ............................................................................................... 3

1.2 Contribution of the Thesis ................................................................................. 6

2 MATHEMATICAL MODELING OF A MOBILE MANIPULATOR ............... 8

2.1 Introduction to Mobile Manipulators (MMs) ................................................... 8

2.2 Homogenous Transformations .......................................................................... 9

2.3 Denavit-Hartenberg (DH) Parameters ............................................................ 10

2.4 Kinematics and Dynamics of a Differential Drive Wheeled Mobile Robot.. 11

2.5 Kinematics and Dynamics of a Serial Manipulator ........................................ 14

2.6 Kinematics and Dynamics of a Mobile Manipulator ..................................... 15

3 BASIC CONCEPTS IN COMPUTER VISION .................................................. 17

3.1 What is Computer Vision? .............................................................................. 17

3.2 Object Detection .............................................................................................. 17

3.3 Depth Estimation ............................................................................................. 19

3.4 Neural Networks and Deep Learning ............................................................. 21

3.5 Depth Estimation and Object Detection Using Deep Learning ..................... 23

3.6 What is Transfer Learning............................................................................... 24

3.7 Yolo v5 ............................................................................................................ 24

iii
CHAPTER Page

3.8 MiDaS .............................................................................................................. 28

4 CONTROL ARCHITECTURE AND SIMULATION SETUP .......................... 29

4.1 Base Control .................................................................................................... 29

4.2 Arm Control ..................................................................................................... 31

4.3 Complete Control Architecture ....................................................................... 32

4.4 Simulation Setup – Assumptions .................................................................... 35

4.5 Simulation Setup – Description ...................................................................... 37

4.6 Distributed Computing .................................................................................... 39

4.7 Experiment 1 – Simulation Using Object Detection Algorithms and RGB-D

Camera ............................................................................................................. 40

4.8 Experiment 2 – Simulation Using Object Detection and Depth Estimation

Algortihms ...................................................................................................... 40

5 RESULTS AND DISCUSSION........................................................................... 41

5.1 Object Detection Algorithm ............................................................................ 41

5.2 Simulation 1 – Using RGB-D Camera ........................................................... 44

5.2.1 Failure Analysis ..................................................................................... 46

5.3 Simulation 2 – Using Monocular Camera ...................................................... 50

5.3.1 Failure Analysis ..................................................................................... 52

6 CONCLUSION AND FUTURE WORK............................................................. 55

6.1 Conclusion ....................................................................................................... 55

6.2 Future Work ..................................................................................................... 55

REFERENCES ....................................................................................................................... 57

iv
 Page

APPENDIX

A YOLO V5 SYSTEM ARCHITECTURE ................................................................ 63

B TRAINING RESULTS – OBJECT DETECTION ALGORITHM ........................ 65

v
 LIST OF TABLES

Table Page

1. Dimensions of the Blocks Used in Simulation ...................................................... 38

2. Specifications of the Systems Used to Run the Simulations................................. 40

3. Training and Dataset Parameters ........................................................................... 41

vi
 LIST OF FIGURES

Figure Page

1. Point 𝑷 in Space Defined With Respect to a Local and Fixed Frame .................... 9

2. Denavit-Hartenberg Convention ............................................................................ 11

3. Differential Drive WMR ........................................................................................ 13

4. Object Detection Examples .................................................................................... 18

5. Example of Depth Estimation Using Stereo Images ............................................. 19

6. Basic Architecture of a Perceptron ........................................................................ 22

7. Example of a Feedforward Neural Network ......................................................... 22

8. Example of Depth Estimation Using Deep Learning ............................................ 23

9. Architecture of YOLO v1 ...................................................................................... 25

10. YOLO v4 FPS vs. AP50 Plots for Different GPU Architectures ......................... 27

11. Inference Speed vs. AP of YOLO v5 Pretrained Models ..................................... 28

12. Turtlebot3 Waffle Pi with OpenMANIPULATOR X........................................... 29

13. Finite-State Machine Describing the Robot’s Controller...................................... 33

14. Robot Controller for Each of the Following States in Figure 4.2 ......................... 34

15. Position of the Camera on the Simulated Robot (Green Box) and on the

Physical Robot (Yellow Box) ................................................................................ 36

16. Top-Down View of the Testbed in Gazebo ........................................................... 38

17. Closer View of the Repository Containing Blocks in the Simulated Testbed ..... 39

18. Confusion Matrix ................................................................................................... 42

19. Precision-Recall Curve........................................................................................... 43

20. F1 Curve ................................................................................................................. 43

vii
Figure Page

21. Sample Predictions of the Model for Images From the Test Dataset ................... 44

22. Outcomes for 50 Trials for Simulation Using RGB-D Camera ........................... 45

23. Instances of Successful Block Pickup in Simulation 1 ......................................... 46

24. Instances of Failed Block Pickup in Simulation 1 ................................................ 47

25. Illustration for Failure Mode 1 ............................................................................... 48

26. Sample Robot Paths for Failure Mode 1................................................................ 49

27. Sample Erroneous Outputs of the Object Detection Algorithm ........................... 50

28. Outcomes of 50 Trials for the Simulation Using Monocular Camera.................. 51

29. Instances of Successful Block Pickup in Simulation 2 ......................................... 51

30. Instances of Failed Block Pickup in Simulation 2 ................................................ 52

31. Depth Map for Different Runs Before Block Pickup ............................................ 53

32. Typical Failure Envelope around a block in Simulation 2 .................................... 54

33. Yolo v5s Model Architecture................................................................................. 64

34. Mean Average Precision (mAP) vs. Number of Epochs for Intersection over

Union (IOU) = 0.5, and IOU = 0.5-0.95 ................................................................ 66

35. Precision vs. Number of Epochs and Recall vs. Number of Epochs .................... 66

36. Precision Curve ...................................................................................................... 67

37. Recall Curve ........................................................................................................... 67

38. Training: Box Loss, Class Loss, and Object Loss vs. Number of Epochs ........... 68

39. Validation: Box Loss, Class Loss, and Object Loss vs. Number of Epochs ........ 69

viii
 CHAPTER 1

INTRODUCTION

Worldwide, the percentage of people living in cities is expected to grow from the current

estimate of 54% to 66% by 2050 (Petersen, 2019). Based on historical data, even the recent

COVID-19 pandemic is unlikely to negatively influence this trend (Florida, 2021). This

has led to an increase in the need for infrastructure in urban areas. However, owing to a

shortage of workers, the construction industry has been unable to keep up with the demand,

dropping the industry’s productivity by 50% compared to the 1960s (Cilia, 2019). Robotics

and autonomous systems possess immense potential to revolutionize the construction

industry (Delgado, 2019). However, the adoption of robotics in the construction industry

has been relatively low owing to several factors, such as high initial capital costs, unskilled

workforce, and low R&D budgets, with high initial capital costs being the prime factor

(Delgado, 2019). Thus, there is a clear need to reduce the high initial capital costs

associated with adopting robotic technologies in the construction industry.

Robots deployed in construction applications must operate reliably in the presence of

the inherent uncertainties associated with a construction environment and work safely

alongside humans, thus, giving rise to the need for semantically driven decision-making

(Garg, 2020). According to Garg et al. (2020), semantics in a robotics context is defined as

the ability of the robot to understand the meaning of places, objects, or other entities in its

surroundings, which is essential for developing intelligent behaviors and safe human-robot

collaboration. Such decision-making abilities can be realized by coupling visual sensing

with learning-based methods, which is a rapidly growing field of research and has seen

increased research output in recent times (Garg, 2020).

1
 Monocular cameras are one of the most widely-used visual sensing solutions for mobile

robots owing to their small form factor, low cost, and suitability for agents with space and

power constraints. Depth estimation using a monocular camera is a research problem that

has gained significant traction in recent years owing to advances in computer vision

(Mertan, 2022). With improvements in techniques for depth estimation using monocular

cameras, the need for more expensive depth sensors such as Lidar and other time-of-flight

sensors could potentially be eliminated.

Thus, to accelerate the adoption of robotic technologies in the construction industry,

there is a clear need to develop robot control strategies that use measurements from low-

cost sensors, rely on little to no prior knowledge of the environment, and can make

decisions based on context. Such controllers could potentially reduce the cost of the robots

used and enable them to work in dynamic construction environments.

This thesis aims to leverage advances in the field of deep learning to develop a vision-

based controller where the only sensory input to the controller is obtained from a single

monocular camera. The controller uses no other inputs, including any prior information

about the environment. The controller is tested on a small mobile robotic manipulator in

simulation for a material retrieval task in which the robot needs to repeatedly search for a

block in a repository zone, pick up the block, and transport it to a deposition zone. The goal

of this thesis is to demonstrate the effectiveness of the controller at this material retrieval

task. Such an approach, in which the controller works with sensor information from low-

cost sensors such as monocular cameras, could potentially reduce the cost of robots used

while facilitating semantically-driven decisions.

2
 1.1 Literature Survey

Although the earliest works in robotic construction, in the form of robotic bricklaying,

appeared about 25 years ago, the field has seen limited commercial usage and adaptations

(Lussi, 2018). As mentioned in the previous section, high capital costs, aversion to change,

and an unskilled workforce are some of the biggest challenges facing the adoption of

robotics technologies in the construction industry (Delgado, 2019). However, an increased

volume of research in the area lately points to a growing interest in the field among

researchers. Competitions such as the Mohammed Bin Zayed International Robotics

Challenge (MBZIRC) (MBZIRC, 2022) have fueled this growth in the interest of

researchers. Mobile manipulators are suitable for construction applications owing to their

large configuration spaces and ability to maneuver in highly dynamic and unstructured

environments. Some of the works in the field of robotics for construction are summarized

in the following paragraphs.

This paragraph discusses some of the works in the field of on-site autonomous

construction using robots. Jung et al. (2014) demonstrated the use of a humanoid robot for

autonomous on-site floor tiling of mosaics. The tiles were identified using an object

recognition algorithm that performed edge detection and color detection by calculating the

position of the tile in the coordinate frame of the robot’s camera. The positions of the

robot’s joints, relative to the camera frame of reference, were calculated using the robot’s

kinematics. The robot was able to pick up the tiles using a suction gripper. Feng et al.

(2014) utilized a 7 Degree-of-Freedom (DOF) robot manipulator for autonomous robotic

assembly in unstructured construction sites. The manipulator was localized with respect to

the material repository and deposition zones by utilizing AprilTags (Wang, 2016)

3
visualized through a monocular camera. The resulting pose estimates were used for

planning the motion of the manipulator. Feng et al. (2015) expanded their previous work

to include a 3D camera on the manipulator to generate an as-built Building Information

Model (BIM) using the point clouds generated by the camera. Lussi et al. (2018) utilized a

mobile manipulator to construct a full-scale, load-bearing, steel-reinforced concrete wall.

The robot was manually repositioned using a joystick for every meter of wall constructed.

Localization of the robot and the end-effector were achieved autonomously using

AprilTags, which were detected by a wide-baseline stereo camera mounted on the arm. The

system also included an on-the-fly building planner, which used images from the stereo

camera mounted on the arm to compensate for material deflections during the welding

process.

Basri et al. (2021) and Štibinger et al. (2021) developed a mobile manipulator to compete

in the MBZIRC 2020, which involved building a predefined structure in an outdoor

environment. The robots were required to autonomously navigate, identify the materials,

and pick up and place them at different locations to build the predetermined structure.

Three Unmanned Aerial Vehicles (UAVs) were used to identify and communicate the

position of the building materials to the ground robots. For robot localization and waypoint

navigation, Basiri et al. (2021) combined the IMU data, wheel odometry, and GPS in

combination with an Extended Kalman Filter (EKF) to obtain the estimated robot pose. In

contrast, Štibinger et al. (2021) used the Adaptive Monte Carlo Localization (AMCL)

method coupled with LiDAR data for robot localization and waypoint navigation. In both

works, the manipulator uses an eye-in-hand approach to detect and pick up the blocks.

4
 In recent years, various vision-based control approaches have been developed for mobile

robots used in construction tasks. Asadi et al. (2018) developed a vision-based controller

for the real-time application of Unmanned Vehicles (UVs) in construction environments.

The controller utilized a monocular camera-based localization and context awareness

system. The former was achieved using ORB-SLAM (Mur-Artal, 2015), and the latter was

achieved using a lightweight neural network, ENet (Paszke, 2016). The two units worked

in parallel to determine the feasible paths for the UVs. Asadi et al. (2021) built on their

previous work by integrating a manipulator on the UV. The monocular camera was

replaced with a set of two stereo cameras, one for context awareness and localization and

the other for visualizing and estimating the block's position. ORB-SLAM was replaced

with RTAB-MAP (Labbé, 2019), and the neural network architecture was changed to

LNSNet (Asadi, 2019). Tsai et al. (2020) developed a deep convolutional neural network-

based control architecture using imitation learning for controlling an omnidirectional

mobile manipulator. The controller uses visual data from a stereo camera to drive toward

and pick up objects of interest and achieves an average success rate of 78.2% for picking

up the objects.

Some of the other approaches used for controlling mobile manipulators outside of

applications in construction, as outlined in the literature, are as follows. Sandakalum et al.

(2022), in their review of motion planning techniques used for mobile manipulators,

identify two primary approaches used for controlling the system: Decoupled and Unified.

In the Decoupled approach, the mobile base and the manipulator control are considered

independently, and the task execution is achieved sequentially. First, the base is driven to

the goal position, and then the arm is used for object manipulation. Some common

5
algorithms used in this approach for both the base and the arm are A*, RRT, RRT-Connect,

genetic algorithms, or their variants. As the name suggests, the Unified approach considers

a coupled model of both the base and arm of a mobile manipulator and is computationally

more expensive. Some of the algorithms used for controlling robots that are modeled using

the Unified approach include CHOMP (Ratliff, 2009) and STOMP (Kalakrishnan, 2011).

Iriondo et al. (2019), in their review of Deep Reinforcement Learning (DRL) for pick-and-

place operations in logistics, successfully implemented algorithms such as Proximal Policy

Optimization (PPO) and Deep Deterministic Policy Gradient (DDPG) on mobile

manipulators for learning pick-and-place tasks. Similarly, Wang et al. (2020) demonstrated

the effectiveness of PPO for mobile manipulation by implementing it on an experimental

platform.

Collective robot construction using swarms of low-cost robots, which can be deployed

autonomously with minimal or no human oversight to build structures far larger than

themselves, is a promising technique for catering to the increased demand of the

construction industry (Werfel, 2012) (Petersen, 2017) (Petersen, 2019). In addition, swarm

foraging is an application that has been researched for a long time (Lu, 2020) and could

potentially be employed in construction applications for material retrieval and

transportation tasks.

1.2 Contribution of the Thesis

In this thesis, we develop a novel vision-based controller for a nonholonomic

differential-drive mobile robot equipped with a manipulator arm, which can be used for

material retrieval tasks in construction applications. This controller differs from existing in
6
that it uses a monocular camera as the primary sensor and does not require prior

information about the environment. The controller integrates existing deep-learning-based

object detection and depth estimation algorithms, using images obtained from a single

monocular camera as the only sensor input. The controller can be deployed on different

types of mobile manipulator platforms with minimal modifications. We test the controller

in the Gazebo 3D robot simulator (Koenig, 2004) for scenarios in which a robot must

autonomously search for, identify, pick up, transport, and deposit individual blocks that

represent construction material. The controller was simulated for 50 trials, and the success

rate of the controller was 46%. This thesis presents a thorough failure analysis for the

controller and proposes solutions to address the different errors.

7
 CHAPTER 2

MATHEMATICAL MODELING OF A MOBILE MANIPULATOR

2.1 Introduction to Mobile Manipulators (MMs)

Mobile manipulators (MMs) are robotic systems consisting of articulated arms

(manipulators) mounted on holonomic or non-holonomic mobile platforms (Tzafestas,

2014). A mobile manipulator combines a manipulator's dexterity with a mobile robot's

mobility, thus creating a platform with far superior capabilities than either one considered

individually (Tzafestas, 2014). Depending on the number of Degrees of Freedom (DOF)

of the mobile base and the manipulator, a mobile manipulator could be kinematically

redundant. Typically, mobile manipulators are kinematically redundant, i.e., they have

more than 6 DOF (Sandakalum, 2022). A kinematically redundant mobile manipulator can

realize the same end-effector pose in more than one configuration, thus providing more

flexibility in terms of the configurations that can be achieved. Mobile manipulators are

used in a wide array of applications, including but not limited to medicine, military,

construction, and space exploration (Sandakalum, 2022).

In this thesis, a decoupled mathematical model of a mobile manipulator, in which the

mathematical models of the manipulator and the base are defined independently, is used to

control the mobile manipulator. This decoupled model suffices for our control objectives

since we do not require complex robot configurations, and the manipulator is controlled in

an open-loop manner; moreover, a coupled (i.e., unified) model would be more

computationally expensive to simulate. The following section discusses the kinematics and

dynamics of the mobile manipulator.

8
 2.2 Homogeneous Transformations

The following derivation of the homogeneous transformation of an arbitrary point 𝑷 in

space (Figure 1) is adopted from (Siciliano, 2012).

Figure 2.1: Point 𝑷 in Space Defined With Respect to a Local and Fixed Frame
(Siciliano, 2012).

Consider the following notations,

𝒑𝟎 is the position [𝒙𝟎 , 𝒚𝟎 , 𝒛𝟎 ]𝑇 vector of the point 𝑷 with respect to a fixed reference

frame, 𝑶𝟎 .

𝒐𝟎𝟏 is the position vector describing the origin of Frame 1 with respect to Frame 0.

𝒑𝟏 is the position vector [𝒙𝟏 , 𝒚𝟏 , 𝒛𝟏 ]𝑇 of the point 𝑷 with respect to a fixed reference frame,

𝑶𝟏 .

𝑹𝟎𝟏 is the rotation matrix describing the orientation of Frame 1 with respect to Frame 0

Thus, the position of the point with respect to the base frame can be defined as follows:

𝒑𝟎 = 𝒐𝟎𝟏 + 𝑹𝟎𝟏 𝒑𝟏 (2.1)

We know that any point in a three-dimensional space is represented as a (3 × 1) vector,

with the components representing the Cartesian coordinates of that point in space. To
9
achieve a compact representation of the same point defined in two different frames as in

equation 2.1, we adopt the homogeneous representation of a vector. By adding a fourth unit

component to a generic vector 𝒑, we get 𝒑′ which is defined as follows:

𝒑 (2.2)
𝒑′ = [ ]
1
Thus, by adopting the above notation for 𝒑𝟎 and 𝒑𝟏 the coordinate transformation can be

written as,

𝑹𝟎𝟏 𝒐𝟎𝟏 (2.3)

𝑨𝟎𝟏 = [ ]
𝟎𝑻 1

where, 𝟎𝐓 is a (1 × 3) vector of zeros, 𝐨𝟎𝟏 is the (3 × 1) translation vector of Frame 1 with

respect to Frame 0 and 𝐑𝟎𝟏 is as defined earlier.

2.3 Denavit-Hartenberg (DH) Parameters

The Denavit-Hartenberg parameters are used to define the relative position of two

consecutive links in a serial manipulator (Siciliano, 2012). To achieve this, the convention

lays out a set of basic rules. The following rules and nomenclature have been adopted from

(Siciliano, 2012). Consider the following assumption:

𝒊 is the axis of the joint connecting link 𝒊 − 𝟏 to link 𝒊, as shown in Figure 2.2.

Then, the Denavit-Hartenberg convention is defined as follows:

• Choose axis 𝒛𝒊 along the axis of Joint 𝒊 − 𝟏.

• Locate the origin 𝑶𝒊 at the intersection of the axis 𝒛𝒊 with the common normal to

axes 𝒛𝒊−𝟏 and 𝒛𝒊 . Locate 𝑶𝒊′ at the intersection of the common normal with axis

𝒛𝒊−𝟏 .

10
 • Choose axis 𝒙𝒊 along the common normal to axes 𝒛𝒊−𝟏 and 𝒛𝒊 with direction from

Joint 𝒊 to 𝒊 + 𝟏.

• Choose axis 𝒚𝒊 to complete a right-handed frame of reference.

Using the DH Parameters, it is possible to calculate the homogeneous transformation

matrix between links 𝒊 − 𝟏 and 𝒊 as in equation (2.3), which can be repeated for all links

to calculate the forward kinematics of the robot arm.

Figure 2.2: Denavit-Hartenberg Convention (Siciliano, 2012).

However, the convention provides a non-unique definition of the link frames in certain

cases and configurations of a manipulator (Siciliano, 2012). These have not been discussed

in this thesis. The reader is encouraged to refer to the cited resource for more information.

2.4 Kinematics and Dynamics of a Differential Drive Wheeled Mobile Robot (WMR)

The following derivation has been adopted from (Tzafestas, 2014). Consider a

differential drive WMR as shown in Figure 2.3 with the following nomenclature:

11
𝑸 is the midpoint of the line connecting the two driven wheels and 𝒗𝑸 is its velocity.

𝒃 is the distance of 𝑸 from the center of gravity 𝑮 of the robot.

𝟐𝒂 is the width of the robot.

(𝒙𝑸 , 𝒚𝑸 ) and (𝒙𝑸̇ , 𝒚𝑸̇ ) are the position and velocity coordinates of point 𝑄, respectively.

𝝓 is the heading angle of the robot.

𝒗𝒓 and 𝒗𝒍 are the linear velocities of the right and left wheels, respectively.

(𝜽𝒓 , 𝜽𝒍 ) and (𝜽𝒓̇ , 𝜽̇𝒍 ) are the angular positions and speeds of the right and left wheels,

respectively.

𝒓 is the radius of the wheels.

The following assumptions are made:

• The wheels roll without any slippage

• The steering axis is perpendicular to the 𝒙 − 𝒚 plane

• The point 𝑸 coincides with the center of gravity 𝑮

Then, the kinematic model of the WMR can be described by the following equations,

𝒓 (2.4)
𝒙𝑸̇ = (𝜽 ̇ 𝒄𝒐𝒔 𝝓 + 𝜽̇𝒍 𝒄𝒐𝒔 𝝓)
𝟐 𝒓

𝒓 (2.5)
𝒚𝑸̇ = (𝜽 ̇ 𝒔𝒊𝒏 𝝓 + 𝜽̇𝒍 𝒔𝒊𝒏 𝝓)
𝟐 𝒓

𝒓 (2.6)
𝝓̇ = (𝜽 ̇ − 𝜽̇𝒍 )
𝟐𝒂 𝒓

with the non-holonomic constraints given as follows:

𝑴𝒑̇ = 𝟎 (2.7)

where,
12
 𝑴 = [− 𝒔𝒊𝒏 𝝓 𝒄𝒐𝒔 𝝓 𝟎] (2.8)

and,

(𝒓⁄𝟐) 𝒄𝒐𝒔 𝝓 (𝒓⁄𝟐) 𝒄𝒐𝒔 𝝓 (2.9)

𝒑̇ = [ (𝒓⁄𝟐) 𝒔𝒊𝒏 𝝓] 𝜽𝒓 + [ (𝒓⁄𝟐) 𝒔𝒊𝒏 𝝓 ] 𝜽̇𝒍
̇
𝒓⁄𝟐𝒂 − 𝒓⁄𝟐𝒂

Figure 2.3: Differential Drive WMR (Tzafestas, 2014).

The dynamics of differential drive WMR is given by (Tzafestas, 2014):

𝟏 (2.10)
𝒗̇ = (𝝉 + 𝝉𝒍 )
𝒎𝒓 𝒓

𝟐𝒂 (2.11)
𝝎̇ = (𝝉 − 𝝉𝒍 )
𝑰𝒓 𝒓

Where, 𝛕𝐫 and 𝛕𝐥 are the torques on the right and left wheels due to the corresponding forces

𝐅𝐫 and 𝐅𝐥 , respectively; 𝐈 is the inertia matrix; 𝐦 is the mass of the robot; 𝐫 and 𝐚 are as

defined earlier. The reader is encouraged to refer to the cited reference for the derivation

of equations (2.10) and (2.11).

13
 2.5 Kinematics and Dynamics of a Serial Manipulator

The kinematic analysis of a Serial Manipulator consists of three parts: Forward

Kinematics, Inverse Kinematics, and Differential Kinematics. Forward Kinematics

involves the task of finding the end-effector position, given a certain joint configuration of

the robot arm. Consider the following nomenclature:

Let 𝒃 and 𝒆 denote the base frame and the end-effector frame of the robot respectively.

Let 0 … 𝒏 denote the intermediate frames corresponding to the joints between 𝒃 and 𝒆.

Thus, using equation (2.3) and the DH convention, the direct kinematics function of the

serial manipulator which computes the end-effector pose with respect to the base frame can

be defined as follows (Siciliano, 2012):

𝑨𝒃𝒆 = 𝑨𝒃𝟎 𝑨𝟎𝒏 𝑨𝒏𝒆 (2.12)

In Differential Kinematics the goal is to find the relationship between the joint velocities

and the end-effector linear and angular velocities (Siciliano, 2012).

Consider the following nomenclature (Siciliano, 2012):

𝒒 = [𝑞1 𝑞2 ⋯ 𝑞𝑛−1 𝑞𝑛 ]𝑇 be the vector of joint variables.

𝐩̇ 𝒆 and 𝝎𝒆 be the linear and angular velocity of the end-effector respectively.

𝑱𝑷 (𝒒) and 𝑱𝑶 (𝒒) be the (3 × 𝑛) Jacobian matrices relating the contribution of the joint

velocities to the end-effector linear and angular velocities respectively.

Thus, the following definitions can be used to determine the relationship between 𝐩̇ 𝒆 and

𝝎𝒆 (Siciliano, 2012):

𝒑̇ 𝒆 = 𝑱𝑷 (𝒒)𝒒̇ (2.13)

14
 𝝎𝒆 = 𝑱𝑶 (𝒒)𝒒̇ (2.14)

In Inverse Kinematics the goal is to find the configuration 𝒒 of the robot arm given an

end-effector pose (Tzafestas, 2014). The solution to the inverse kinematics problem is not

unique; and depending on the number of degrees of freedom of the system, numerous

solutions are possible (Siciliano, 2012). Both analytical and numerical methods have been

developed for calculating the inverse kinematics of a serial manipulator. The reader is

encouraged to refer to the cited works for a comprehensive treatment of the topic.

The dynamics of a serial manipulator are typically derived using the Lagrange method and

the reader is encouraged to refer to the cited works for a thorough treatment of the topic.

2.6 Kinematics and Dynamics of a Mobile Manipulator

The following equations and nomenclature for the kinematics and dynamics of the

mobile manipulator have been adopted from (Tzafestas, 2014).

Consider the following four coordinate frames:

𝑶𝒘 𝒙𝒘 𝒚𝒘 𝒛𝒘 is the world coordinate frame.

𝑶𝒑 𝒙𝒑 𝒚𝒑 𝒛𝒑 is the platform coordinate frame.

𝑶𝒃 𝒙𝒃 𝒚𝒃 𝒛𝒃 is the coordinate frame of the manipulator’s base.

𝑶𝒆 𝒙𝒆 𝒚𝒆 𝒛𝒆 is the coordinate frame of the manipulator’s end-effector.

𝒙𝒘
𝒆 is the end-effector pose in the world coordinate frame.

𝑇
𝒖(𝒕) = [𝑢𝑝𝑇 (𝑡), 𝑢𝑚
𝑇
(𝑡)] is the vector of control commands to the platform, and the end-

effector respectively.

𝒑 = [𝑥 𝑦 𝜙]𝑇 represents the platform configuration.

15
𝜽 = [𝜃1 𝜃2 𝜃3 ⋯ 𝜃𝑛 ]𝑇 represents the manipulator configuration.

𝒒 = [𝒑𝑻 , 𝜽𝑻 ]𝑻 denotes the combined configuration of the WMR and manipulator.

𝑱(𝒒) be the combined Jacobian of the WMR and manipulator.

So, the transformation from the world coordinate frame to the end-effector coordinate

frame is given as follows using equation (2.3):

𝒑
𝑨𝒘 𝒘 𝒃
𝒆 = 𝑨𝒑 𝑨𝒃 𝑨𝒆 (2.15)

And the overall kinematic model of the mobile manipulator is given as follows:

̇
𝒙𝒘
𝒆 = 𝑱(𝒒)𝒖(𝒕) (2.16)

subject to the nonholonomic constraint,

𝑴(𝒑)𝒑̇ = 0, 𝑴(𝒑) = [𝑠𝑖𝑛̇ 𝜙 𝑐𝑜𝑠 𝜙 0 ⋯ 0] (2.17)

The dynamics of the mobile manipulator are derived using the Lagrange method and the

reduced (unconstrained) model that describes the dynamic evolution of 𝒒(𝒕) in terms of

the dynamic evolution of 𝒗(𝒕), is as follows:

̅ (𝒒)𝒗 +
𝑫 ̅ (𝒒, 𝒒̇ )𝒗 + 𝒈
̇ 𝑪 ̅𝝉
̅ (𝒒) = 𝑬 (2.18)

̅, 𝑪
For the definitions of the matrices 𝑫 ̅, 𝒈 ̅ , the reader is encouraged to refer to
̅ and 𝑬

references (Tzafestas, 2014, Siciliano, 2012 and Lynch, 2019).

16
 CHAPTER 3

BASIC CONCEPTS IN COMPUTER VISION

3.1 What is computer vision?

Computer vision is the enterprise of automating and integrating a wide range of

processes and representations used for visual perception and is the inverse problem of

computer graphics (Szeliski, 2011) (Ballard, 1981). The task is to describe the world we

see in one or more images and reconstruct its properties, such as shape, illumination, and

color distributions (Szeliski, 2011). From the engineering point of view, the goal is to build

autonomous systems which can perform or outperform some of the tasks that the human

visual system can do (Huang, 1996). Computer vision is widely used today in a wide

variety of real-world applications such as photogrammetry, automotive applications,

medicine, motion capture, and 3D modeling (Szeliski, 2011). It is an umbrella term that

encompasses multiple tasks such as image processing, object detection, and pattern

classification (Ballard, 1981). Some of the topics relevant to this thesis have been discussed

briefly in the subsequent sections.

3.2 Object Detection

The object detection problem involves two major tasks: (1) to determine the location of

objects in each image, also called object localization, and (2) to determine which category

each object belongs to, also called object classification (Zhao, 2019). An example of an

object detection task has been illustrated in Figure 3.1. An object detection pipeline has

three major stages: Informative region selection, Feature extraction, and Classification

(Zhao, 2019). A brief description of the tasks involved in the three stages are as follows:

17
(1) Informative region selection: An image can contain several objects located at different

positions, which could be of different sizes or aspect ratios (Zhao, 2019). The goal of

informative region selection is to use a multi-scale sliding window to extract information

about the positions of different objects in the image (Zhao, 2019). Although this method

converges, i.e., it can extract the positional information of all objects in the image, it is

computationally expensive and produces too many redundant windows (Zhao, 2019). (2)

Feature extraction: The next step is to extract information about an image’s visual

features, providing semantics and robust representations, which are then used to identify

the different objects within an image (Zhao, 2019). Scale Invariant Feature Transform

(SIFT) is one of the methods used in feature extraction (Lowe, 2004). (3) Classification:

The last task in object detection is the task of classification, where the goal, as the name

suggests, is to classify the image or the objects in the image into one or more predefined

categories (Zhao, 2019). This step makes the representations of the objects in the image

more hierarchical, semantic, and informative for visual recognition (Zhao, 2019).

Figure 3.1: Object Detection Examples (He, 2017)

18
 3.3 Depth Estimation

Estimating the depth information from 2D images is one of the most critical tasks in

computer vision (Zhao, 2020). It involves evaluating a dense depth map for a given RGB

image, i.e., one must calculate a metric depth value for each pixel in the given RGB image

(Mertan, 2022). It is an essential component of understanding the geometric relationships

within a scene and is used in many applications such as SLAM, navigation, and object

detection (Eigen, 2014) (Zhao, 2020). An example of depth estimation from a pair of stereo

images is shown in Figure 3.2, where the shading in the image becomes darker as the

distance from the camera increases, and the unknown regions are represented in black

(Scharstein, 2003).

Figure 3.2: Example of Depth Estimation Using Stereo Images, Left Camera Image
(Left), Right Camera Image (Center), and Final Disparity Map (Right) (Scharstein, 2003)

There are three primary methods for estimating depth from RGB images: Geometry-

based methods, Sensor-based methods, and Deep learning-based methods (Zhao, 2020).

(1) Geometry-based methods are widely popular within the research community and have

been studied for the past forty years (Zhao, 2020). One of the most widely used techniques

is to estimate depth using a pair of stereo images (Zhao, 2020). A stereo image is a pair of

images captured using two cameras (left and right) separated by a fixed distance. Each
19
camera receives a slightly different view of the world, and this disparity between the

images is exploited to extract information about the image (Saxena, 2007). These

disparities vary inversely with the object’s distance from the camera, and small changes in

disparities result in significant errors in the depth estimates (Saxena, 2007). Another

example of a geometry-based method is Structure from Motion (SfM) (Ullman, 1979),

where feature correspondences and geometric constraints between image sequences are

used to obtain the depth of sparse features (Zhao, 2020). (2) Sensor-based methods, as the

name suggests, utilize sensors such as RGB-D cameras and LiDAR to directly get the

corresponding images' depth information without using geometric estimation techniques

(Zhao, 2020). Although RGB-D cameras can provide pixel-level depth maps, their range

of operation is limited and they are not robust to intense lighting conditions (Zhao, 2020),

while LiDAR can only generate sparse 3D maps. These sensors have a few drawbacks in

terms of cost, weight, and size, thus making them unsuitable for power and size-constrained

platforms such as small drones (Zhao, 2020). Such issues have led to increased research in

depth estimation from monocular cameras using deep learning methods (Zhao, 2020). (3)

Deep learning-based methods have shown incredible potential and breakthrough

performance on various computer vision tasks, ranging from image classification to depth

estimation (Zhao, 2020). These methods employ neural networks with varying

architectures, depending on the task and the accuracy required. The techniques hold great

promise for use on systems with limited sensing capabilities, since increasing research in

the field could potentially eliminate the need for large, expensive sensors.

20
 3.4 Neural Networks and Deep Learning

According to Mitchell (1997), the definition of machine learning is as follows: “A

computer program is said to learn from experience E with respect to some task T and some

performance measure P, if its performance on T, as measured by P, improves with

experience E.” Artificial Neural Networks are popular techniques that simulate the

mechanisms of biological organisms and are widely used in machine learning applications

(Aggarwal, 2018). The biological neurons are connected using axons and dendrites, and

the region between the connections are called synapses (Aggarwal, 2018). An artificial

neural network consists of mathematical neurons connected in either a fully connected or

sparsely connected manner, and they are designed to mimic the functions of a biological

neuron. The perceptron is the simplest neural network and consists of a single input layer

and an output node, and the illustration of its architecture is shown in Figure 3.3 (Aggarwal,

2018). The simple mathematical description of a perceptron is as follows (Aggarwal,

2018):

Let 𝒅 be the number of nodes that transmit the 𝒅 features.

̅ = [𝑥1 ⋯ 𝑥𝑑 ] be the feature vector and 𝑾

𝑿 ̅̅̅ = [𝑤1 ⋯ 𝑤𝑑 ] be the corresponding edge

weights.

Therefore, the prediction 𝑦̂ at the output layer is given as follows:

𝑑 (3.1)
̂ = 𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛_𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛{∑ 𝑤𝑗 𝑥𝑗 }
𝒚
𝑗=1

Where the 𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛_𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛 is chosen based on the application, e.g., sign and ReLU,

a bias value can be selected to influence the output value as well.

21
Figure 3.3: Basic Architecture of a Perceptron, Without Bias (Left) and With Bias (Right)
(Aggarwal, 2018)

Unlike the perceptron, which contains a single input and output layer as seen earlier,

multi-layer neural networks include more than one computational layer (Aggarwal, 2018).

A multi-layer neural network has additional layers between the input and output layers; as

mentioned earlier, these middle layers are called hidden layers. In multi-layer neural

networks, the output from each successive layer feeds into the next layer in the forward

direction and are hence also referred to as feed-forward networks (Aggarwal, 2018). Such

networks can either be fully or partially connected. However, fully connected architectures

perform better in many settings and thus are much more common (Aggarwal, 2018). An

illustration of a fully connected feed-forward neural network is shown in Figure 3.4.

Figure 3.4: Example of a Feedforward Neural Network, Without Bias (Left) and With
Bias (Right) (Aggarwal, 2018)

22
 3.5 Depth Estimation and Object detection using Deep Learning

As discussed in earlier sections, the power of deep learning methods to learn features

from images using pixel-level relationships and other visual cues has made them suitable

for tasks in computer vision, such as depth estimation and deep learning (Zhao, 2020).

Several well-known neural network architectures, such as Convolutional Neural Networks

(CNNs), Recurrent Neural Networks (RNNs), Region-based Convolutional Neural

Networks (R-CNNs), and Generative Adversarial Networks (GANs), have been used for

the tasks of depth estimation and object detection (Zhao, 2019) (Zhao, 2020). Some newer

techniques, such as Mask R-CNN, build on top of Faster R-CNNs by adding a binary mask

for each region of Interest (RoI), thus significantly enhancing the performance metrics of

the model (He, 2017). An example of depth estimation using a deep neural network-based

approach is shown in Figure 3.5. Brighter regions in the output image correspond to parts

of the scene that are closer to the camera, and darker regions represent parts that are farther

away.

Figure 3.5: Example of Depth Estimation Using Deep Learning (Monocular Cameras),
Input Image (Left), Output Image (Right) (Ranftl, 2022)

23
 3.6 What is Transfer Learning

According to Pan (2010), Transfer Learning is defined as, “Given a source domain DS

and learning task Ts , a target domain DT and learning task TT , transfer learning aims to help

improve the learning of the target predictive function fT (. ) in DT using the knowledge in

DS and TS , where DS ≠ DT , or TS ≠ TT ”. There are three main tasks in transfer learning,

(1) What to transfer deals with determining which part of the knowledge can be transferred,

(2) How to transfer deals with the question of how the existing method would be used for

the new application, and depends on the task at hand, and (3) When to transfer deals with

determining the situations and tasks in which transferring should be done (Pan, 2010). An

example of the transfer learning task described in Pan (2010) is as follows. Consider a

sentiment classifier that classifies product reviews. To build this classifier, we would need

to collect a large volume of data, build a model, train it, and tune it to obtain desired output

and accuracy. If our data is limited, or if there are errors in data annotations, model’s

performance would not be as desired. In such cases, transfer learning can improve the

model’s performance and reduce the overall effort in building the model in general.

3.7 YOLO v5

You Only Look Once (YOLO) is a family of state-of-the-art object detection algorithms,

which was initially proposed in 2016 (Redmon, 2016). The model is currently in its fifth

iteration (v5), which is a PyTorch implementation of its fourth version (v4) (Jocher, 2020).

The model is widely used within the computer vision research community for many real-

time computer vision applications. The reason for the model’s popularity is its ability to

24
produce high mean average precision (mAP) while maintaining fast runtimes, or frames

per second (FPS). Different versions of the model and their properties are as follows:

YOLO v1 utilizes a single neural network to predict bounding boxes and class

probabilities directly from full images in one pass and thus, could be optimized directly for

detection performance (Redmon, 2016). The model formulates the object detection

problem as a regression problem and maps the image pixels directly to bounding box

coordinates and probabilities (Redmon, 2016). The model performed extremely well

compared to other state-of-the-art models at the time, with 9-40% better Average Precision

(AP), depending on the dataset on which it was evaluated (Redmon, 2016). The

architecture of YOLO v1 is shown in Figure 3.6.

Figure 3.6: Architecture of YOLO v1 (Redmon, 2016)

YOLO v2, also called YOLO9000, is the second iteration in the YOLO family of

models. As the name suggests, it could detect over 9000 object categories and used a multi-

scale training method, thus enabling it to run at varying sizes, offering a trade-off between

speed and accuracy (Redmon, 2017). It also was designed to overcome some of the

25
significant drawbacks of YOLO v1, namely localization errors and low recall compared to

region-proposal-based networks (Redmon, 2017). On the VOC 2012 dataset, it achieved a

mAP of 73.4% while running at much faster speeds than competing models (Redmon,

2017).

YOLO v3 was an incremental improvement from the previous version, with a newly

trained classifier network with added connections to the backbone layer (Redmon, 2018).

It had a comparable performance with other state-of-the-art models at the time; as

mentioned earlier was simply an incremental improvement over v2.

YOLO v4 implemented a new CSPDarknet53 architecture for the backbone network

(Bochkovskiy, 2020). It also adds a PANnet for feature aggregation, SPP for increasing the

receptive field, and data augmentation methods (“bag of freebies”) which increases the

performance of the network without sacrificing inference times (Bochkovskiy, 2020). This

resulted in significantly better performance characteristics both in terms of AP and FPS,

and the plots of FPS vs. AP50 for different GPU architectures are shown in Figure 3.7.

26
 Figure 3.7: YOLO v4 FPS vs. AP50 Plots for Different GPU Architectures on the
Microsoft COCO Dataset: Volta (Top), Pascal (Bottom) (Bochkovskiy, 2020)

YOLO v5 is a PyTorch implementation of YOLO v4, is open source, and combines the

power of the YOLO model with the flexibility of PyTorch, making it suitable for real-time

applications in research (Jocher, 2020). The implementation has a set of pre-trained models

with varying sizes with trade-offs in inference speed vs. accuracy. The performance

characteristics of the different implementations are shown in Figure 3.8. Refer to Appendix

A for the architecture of YOLO v5.

27
 Figure 3.8: Inference Speed vs. AP of YOLO v5 Pretrained Models (Jocher, 2020)

3.8 MiDaS

MiDaS, developed by Ranftl et al. (2022), is a robust monocular depth estimation model

that can perform across various environments. The authors trained the model on diverse

datasets, using a novel loss function, making the model invariant to major sources of

incompatibilities between datasets (Ranftl, 2022). The model trained for about six months

on different datasets. The performance was evaluated using the principle of “zero-shot

cross dataset transfer,” i.e., the model was evaluated on datasets it had never seen before

(Ranftl, 2022). The model outperforms other models in the monocular depth estimation

using deep learning space, thus making it state-of-the-art. A sample prediction of the model

has been illustrated in Figure 3.5. The reader is encouraged to refer to the cited work for

details about the model.

28
 CHAPTER 4

CONTROL ARCHITECTURE AND SIMULATION SETUP

This thesis considers a Robotis Turtlebot3 Waffle Pi with an OpenMANIPULATOR-X

for the simulations (Figure 4.1). A sequential, decoupled approach has been adopted for

controlling the mobile manipulator, i.e., the control of the base and the manipulator are

independent of each other. First, the mobile base is controlled using vision-based methods

and driven to specific locations in the simulated test environment. Once the mobile base

reaches its goal position, the arm controller is activated and is used to pick up a block. The

specifics of the individual components of the control architecture, assumptions, and

simulation setup are discussed in the following sections.

Figure 4.1 Turtlebot3 Waffle Pi With OpenMANIPULATOR X (source: Robotis)

4.1 Base Control

The proposed base control method has three major components, (1) tag-based

navigation, (2) object detection, and (3) depth estimation using a Raspberry Pi monocular

camera, and they are explained briefly in the following paragraphs.

29
(1) The tag-based navigation system consists of AprilTags (Wang, 2016) placed in the

Gazebo testbed corresponding to pickup and deposition zones (discussed in later

sections). The onboard monocular camera (Raspberry Pi v2) visualizes the tags,

and the apriltag_ros package (Wang, 2016) (Malyuta, 2017) computes the pose of

the tags with respect to the desired frame of reference on the robot. The robot uses

this information for the initial navigation between different zones in the simulated

test environment.

(2) The object detection algorithm was developed by transfer-learning a pre-trained

YOLOv5 model to identify blocks belonging to one of five classes: Red Cube, Red

Cylinder, Red Sphere, Green Cube, and Green Cylinder. The robot uses its onboard

monocular camera to visualize the world around it. The output from the camera is

resized to 416 × 416 pixels, and the neural network uses this input to run inference

in real-time. The inference results from the model contain data that includes

information about the object's class and bounding boxes. The robot's base controller

calculates the object’s centroid coordinates from the inference results and uses them

to navigate toward a block of interest.

(3) Depth estimation using the monocular camera is carried out by employing the

MiDaS v3 (Ranftl, 2022) network. Like the object detection algorithm, the depth

estimation network runs real-time inference on the resized output images from the

camera. The output is a depth map of the image that consists of relative inverse

depth values corresponding to each pixel in the image. Thus, as mentioned earlier,

combining the outputs from the object detection algorithm and the depth estimation

30
 network, we can identify the depth corresponding to the centroid of the object of

interest.

Controlling the robot’s base to navigate the different zones in the test environment is

achieved using a sequential combination of the three components. The robot uses

AprilTags for initial navigation toward the Pickup and deposition zones. Once the robot

reaches a predetermined distance threshold relative to the tags, it switches over to the deep-

learning-based methods for navigating closer to a block. The object detection algorithm

returns the centroid (𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑_𝑥 and 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑_𝑦) of the block, which is used to obtain the

corresponding depth value and is further used to calculate the angular velocity command

for the mobile base using the following equation:

𝜔𝑧 = (213 − 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑_𝑥)/45 (4.1)

where 𝜔𝑧 is the angular velocity in rad/s of the mobile base w.r.t the z-axis of the world

frame, pixel value 213 corresponds to the midpoint of the image, and 45 is a scaling factor.

The resulting value of 𝜔𝑧 is checked to ensure that it is within predetermined limits (± 0.05

rad/s) to ensure safe operation of the system. Section 4.3 describes the equations for

calculating the velocity commands in further detail. The velocity commands generated by

the control equations are realized on the robot using a standard PID implementation.

4.2 Arm Control

The arm control for the system uses the MoveIt Motion Planning Framework (Coleman,

2014), which is implemented on the Open Motion Planning Library (OMPL) (Șucan,2012)

backend. The planner chosen for this application is RRT-Connect. The arm is programmed

to reconfigure into predetermined sets of joint angles to pick up and place objects in the
31
test environment. The arm’s control system is an open-loop controller, i.e., the arm does

not have sensing capabilities to ensure block pickup and deposition. Instead, we rely on the

accuracy of the base controller to drive the robot to the required pose (described in Figure

4.3(b)) to ensure that the arm picks up the block. Although this method works well in

simulation, a real-world implementation might require the following modifications: (1) an

additional camera on the arm and a closed-loop controller with vision-based feedback from

this camera to ensure block pickup, or (2) a neural network that is trained to determine

whether the block has been picked up using images from the camera mounted on the mobile

base.

4.3 Complete Control Architecture

The overall control architecture of the robot is set up as a finite-state machine, shown in

Figure 4.2, that incorporates the control approaches discussed in Sections 4.1, 4.2 and 4.4.

The implementation of the state machine is realized using the ROS smach package. The

state machine has the following five states: (1) Navigate towards the Material repository,

(2) Navigate to a block, (3) Pick up a block, (4) Navigate to the Construction zone, and (5)

Place a block.

32
 Figure 4.2 Finite-State Machine Describing the Robot’s Control Architecture

Figure 4.3 illustrates the controller that governs the robot’s actions in each of the states

in Figure 4.2. In Figure 4.3(a), d denotes the distance between the reference frames of the

robot’s camera and the tag, respectively, in the x-direction, and the value of i is 0.3 for

State 1 and 0.2 for State 4. In Figure 4.3(b), 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑𝑥 denotes the pixel coordinate

corresponding to the centroid of the block along the horizontal axis. Since the input image

to the object detection algorithm is resized to 416 × 416 pixels, the output for 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑𝑥

varies from 0 to 416 pixels, with the pixel value 213 representing the center of the image.

Thus, the robot executes the states sequentially until it moves all blocks from the Material

Repository to the Deposition Zone. If the robot is in State 1, 2, or 4 and moves too close to

a wall, it moves backwards and, if it loses sight of the tag/block, it rotates until the tag/block

reappears in its field of view (Figure 4.3(d)).

33
34
 Figure 4.3 Robot controller for each of the following states in Figure 4.2: (a, Top Left)
Move robot base w.r.t AprilTags (State 1 and State 4), (b, Top Right) Move robot base
w.r.t blocks (State 2), (c, Bottom Left) Move arm (State 3 and State 5), and (d, Bottom
Right) Rotate recovery.

4.4 Simulation Setup – Assumptions

To validate the vision-based control approach outlined in this thesis, a test environment

was set up in the Gazebo robot simulator (Koenig, 2004). The test world is a simplified

approximation of a controlled construction environment with two zones: the Material

Repository and the Material Deposition zone. The goal of the robot is to retrieve building

blocks (one of the five classes described earlier, specified by the user) from the Material

Repository, and transport them one at a time to the Material Deposition zone. The next

section describes the Gazebo test world in greater detail.

The following assumptions are used in creating the simulation environment and defining

the information available to the robot:

• The blocks that the robot needs to pick up are located on the ground, and so for the

robot’s onboard camera to be able to detect the blocks, it is placed at a lower

position on the simulated robot than on the physical robot, as illustrated in Figure

4.4.

• To avoid collisions with walls in the simulated environment, the robot uses

estimates of its distance to the walls from the depth estimation algorithm, based on

images from its onboard camera and measurements from its onboard Lidar.

35
Figure 4.4 Position of the Camera on the Simulated Robot (Green Box) and on the
Physical Robot (Yellow Box).

• The robot does not have a global localization mechanism and does not require prior
mapping of the environment for it to execute the controller. Thus, the robot moves

only based on its perception of the environment through images from its on-board

camera, and it uses the AprilTags to help it navigate towards the different zones in

the test environment.

• The controller does not depend on the dimensions of the robot. This makes the

controller agent-agnostic, and it can thus potentially work on any mobile base with

any arm, requiring minimal modifications to the control system. However, this

approach has a downside: the robot could potentially collide with the objects in its

environment, since the controller can only account for objects in the field of view

of the camera. For instance, if the robot performs a 360° rotation close to a wall,

the back part of the robot could potentially collide with the wall since the controller

is unaware of the dimensions of the robot.

• The test environment is assumed to be free of obstacles; the only static obstacles in

the environment are the walls of the simulated testbed.

36
 • The test environment is assumed to have adequate lighting in order to eliminate

issues pertaining to low lighting conditions such as problems with tag detections.

• The size of each AprilTag is 18 × 18 𝑐𝑚. This ensures that the tags are visible from

either end of the testbed, eliminating issues pertaining to tag detections.

• As mentioned in Section 4.2, the arm does not have sensor feedback, and so the

robot’s success at picking up a block depends on the accuracy of the controller that

drives its mobile base. However, adding feedback to the arm controller would

improve the effectiveness of the robot at picking up blocks (e.g., its robustness to

navigation errors) and would provide the robot with additional capabilities.

4.5 Simulation Setup – Description

The Gazebo simulation consists of the two zones mentioned earlier, the Material

Repository and the Material Deposition zone. The testbed has the following dimensions:

4 × 3 × 0.3 𝑚. Figure 4.5 illustrates a top-down view of the testbed in Gazebo, with two

blocks located in the Repository zone. The red, green, and blue lines correspond to the x-

axis, y-axis, and z-axis respectively, of the global coordinate frame.

37
Figure 4.5 Top-Down View of the Testbed in Gazebo. The Red Blocks are Located in the
Repository.

The repository and the deposition zones are positioned diagonally opposite to each other,

and the robot is tasked with retrieving objects from the repository and transporting them to

the deposition zone. Figure 4.6 depicts a closer view of the repository with different shape

blocks and Table 4.1 outlines the dimensions of the blocks used in the simulation.

Table 4.1

Dimensions of the Blocks Used in Simulation

OBJECT DIMENSIONS

Red and Green cubes 4 x 4 x 4 cm

Red and Green Cylinders Radius: 2.5 cm; Length: 10 cm

Red Sphere Radius: 2.5 cm

38
 Figure 4.6 Closer View of the Repository Containing a Few Blocks in the Simulated
Testbed.

4.6 Distributed Computing

To reduce the computational load, this thesis uses a distributed computing approach to

running the simulations. The simulations leverage the power of the Robot Operating

System (ROS) Melodic architecture to run the deep learning nodes on separate computers.

Let 𝑆𝑦𝑠𝑡𝑒𝑚1 and 𝑆𝑦𝑠𝑡𝑒𝑚2 denote the two systems on which the nodes run. 𝑆𝑦𝑠𝑡𝑒𝑚1 runs

the ROS Master and the object detection node, while 𝑆𝑦𝑠𝑡𝑒𝑚2 runs the depth estimation

node. The systems are connected to each other with wired connections via a common

router, which is in turn connected to the Arizona State University wired network, thus

ensuring a bidirectional connection between the two systems. Table 4.2 highlights the

specifications of both systems used to run the simulations.

39
Table 4.2

Specifications of the Systems Used to Run the Simulations

𝑺𝒚𝒔𝒕𝒆𝒎𝟏 𝑺𝒚𝒔𝒕𝒆𝒎𝟐

Intel Xeon E5-2623 processor Intel i7-9750H processor

32 GB RAM 16 GB RAM

Nvidia Quadro M4000 8 GB GPU Nvidia GTX 1660Ti 6 GB GPU

4.7 Experiment 1 – Simulation using Object Detection Algorithm and RGB-D Camera

To establish the baseline performance of the proposed vision-only control approach, the

monocular camera was replaced by an RGB-D camera to obtain depth values directly,

without estimating them from monocular images. The depth data obtained from the RGB-

D camera was used in conjunction with the object detection algorithm to drive the robot to

positions that facilitate block pickup. The next chapter discusses the results of the

experiment.

4.8 Experiment 2 – Simulation Using Object Detection and Monocular Depth Estimation

Algorithms

The second experiment was carried out using only the input from the robot’s on-board

monocular camera, together with the control approaches described earlier in the chapter.

The next chapter discusses the results of the experiment.

40
 CHAPTER 5

RESULTS AND DISCUSSION

5.1 Object Detection Algorithm

The object detection algorithm for identifying the building blocks was created by

transfer learning on top of the YOLO v5 model. A set of 1087 images were collected from

the Gazebo robot simulator by initially capturing a video of the objects of interest in the

test bed, and individual frames were isolated to create the images required for training. The

generated images were manually annotated using roboflow. Then data augmentation

methods, namely changing the brightness and mirroring, were applied to increase the

robustness of the model to variations in lighting and the orientations of the blocks. The size

of the final dataset generated was 2777 images. The model was trained on the Google Colab

platform on a Tesla P100 GPU. Although the model was trained for 500 epochs, there was

no significant improvement in the model after epoch 294 (best model), and the training was

terminated at epoch 394. The parameters of the dataset and the training routine used for

transfer learning are summarized in Table 5.1.

Table 5.1

Training and dataset parameters

Parameter Value

Number of Images 2777; Training: 2535, Test: 121, Validation: 121

Image Classes 5 (Red – cube, cylinder, sphere; Green – cube, cylinder)

Input Image size (pixels) 416×416

41
 YOLO model YOLOv5s

Batch size 16

Number of Epochs 500

The results of transfer learning are shown in the following plots: Confusion matrix

(Singh 2021) (Figure 5.1), F1 curve (Kulkarni, 2020) (Figure 5.2), and Precision-Recall

curve (Ozenne, 2015) (Figure 5.3).

Figure 5.1 Confusion Matrix

42
Figure 5.2 Precision-Recall Curve

Figure 5.3 F1 Curve

43
 From the plots above, it is evident that the model fits the data very well, since it has a

mean average precision (mAP) of 0.99 and an average F1 score of 0.98 at a confidence

level of 0.7. Some sample predictions from the model on the test dataset are shown in

Figure 5.4. Additional results on the performance of the algorithm are included in

Appendix B, and the reader is encouraged to refer to that section for further information.

Figure 5.4 Sample Predictions of the Model for Images From the Test Dataset. The
Numeric Values Indicate the Confidence Levels for Each Identified Object.

5.2 Simulation 1 – Using RGB-D Camera

The first simulation uses an RGB-D camera (Section 4.7) to set a performance

benchmark for the controller. The robot's goal is to retrieve one block (red cube) at a time

from the Material Repository and transport it to the Deposition Zone. Once the robot picks

up a block from the repository, another block is manually replaced in the Material

Repository at a different location with respect to the previous block. Only one block is

present at a time in the Material Repository to avoid sudden variations in the value of

𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑𝑥 that arise when the object detection algorithm identifies multiple blocks in a

single image. Section 5.2.1 discusses this in further detail.

44
 A Trial is defined as a single iteration in the simulation where the robot navigates to the

Material Repository, identifies a block, picks it up with its manipulator, navigates to the

Deposition Zone, and places the block down. A Trial is successful (Success) if the robot

correctly executes each of the actions in this sequence. If the robot fails to execute any

action in a trial successfully, then the Trial is a failed run (Fail). Since the control of the

arm is open-loop, the controller cannot determine from sensor feedback whether the robot

has indeed picked up a block. Thus, the robot still performs a complete Trial even in the

case of a Fail; however, in such a case the robot does not transport the block to the

Deposition Zone (it arrives there “empty-handed”).

A human observer determines whether the outcome of a Trial is a Success or a Fail. The

controller was tested for 50 Trials, and the outcome of each trial is plotted in the graph

shown in Figure 5.5. The value 1 on the y-axis corresponds to Success and the value 0

corresponds to Fail. This implies that the green regions of the graph correspond to

successful Trials and the white regions correspond to failed attempts.

Figure 5.5 Outcomes of 50 Trials for the Simulation Using RGB-D Camera
45
 Since 34 out of 50 Trials were successful, the controller has a success rate of 68%.

Figure 5.6 shows snapshots of the robot picking up a block during several successful Trials

for Simulation 1.

Figure 5.6 Instances of Successful Block Pickup in Simulation 1

5.2.1 Failure Analysis

It is imperative to mention the different cases under which the controller Fails in order

to understand the overall performance of the system better. Figure 5.7 illustrates a few

examples where the controller Fails. Based on observations of the outcomes for of 25

Trials, three major factors have been identified which cause a Trial to Fail. They are:

Failure Mode 1: Failure during navigation towards Material Repository (State 1, Section

46
4.3), Failure Mode 2: Errors in the object detection algorithm, and Failure Mode 3:

Sensor errors. The subsequent paragraphs discuss the Failure Modes in greater detail.

Figure 5.7 Instances of Failed Block Pickup in Simulation 1

Failure Mode 1: The tags corresponding to the Material Repository and the Deposition

zones are placed almost diagonally opposite to each other. Since the shortest path between

two given points is a straight line, the controller is set up to traverse a straight line while

navigating between the two zones. However, since there are an infinite number of parallel

lines with the same slope, the robot could traverse along any parallel line, which could

cause the controller to potentially Fail.

To understand this further, consider Figure 5.7. The blue region represents the

approximate operating region where the controller executes State 1 and State 4 as intended.

47
The red arrow represents the direction along which the robot and the tags are aligned. The

maroon lines represent the parallel lines along which the robot could potentially traverse.

When the robot loses sight of a tag, it performs a Rotate Recovery behavior (Figure 4.3)

until the robot identifies the tags. At this point, the robot continues to move along another

parallel line with the same slope determined by the robot controllers associated with State

1 and State 4 (Section 4.3).

Figure 5.8 Illustration for Failure Mode 1

Therefore, when the robot picks a line below the blue operating region, it could

potentially traverse along a path like the purple trajectory in Figure 5.9. Similarly, when

the robot picks a line above the blue operating region, it could display a behavior such as

48
the yellow trajectory shown in Figure 5.9. Thus, in the unlikely case that the robot steers

far off from the blue operating region, it could potentially strike the walls of the test bed

and not exhibit the desired behaviors. A different controller for navigating between the tags

could potentially resolve this problem.

Figure 5.9 Sample Robot Paths for Failure Mode 1. Red Arrows Indicate Direction of
Travel.

Failure Mode 2 is caused by erroneous outputs from the object detection algorithm.

Figure 5.10 illustrates some examples of such erroneous outputs. Erroneous outputs cause

the value of 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑𝑥 to fluctuate suddenly, which then causes sudden variations in the

angular velocity commands. Thus, the robot could suddenly steer off course and miss the

49
block pickup, which causes the Trial to Fail. This behavior is especially pronounced when

multiple blocks are in the Repository Zone. Since the object detection algorithm cannot

uniquely identify instances of objects within a given image, the order in which the objects

are detected varies with each frame and thus causes the output value of 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑𝑥 to

fluctuate. Using an instance segmentation algorithm instead of the object detection

algorithm could solve this problem.

Figure 5.10 Sample Erroneous Outputs of the Object Detection Algorithm in the
Simulations

Failure Mode 3 occurs due to errors and uncertainty associated with the depth estimates

from the RGB-D camera. These depend upon the type of camera used and its

characteristics. Based on observations, this type of failure was relatively rare and could be

easily eliminated by using a more accurate RGB-D camera.

5.3 Simulation 2 – Using Monocular Camera

The second simulation replaces the RGB-D camera with a monocular camera as

discussed in Chapter 4. The definitions for Trial, Success and Fail are as defined in

Simulation 1 (Section 5.2). The controller was tested for 50 Trial’s and the outcomes have

50
been plotted in the Graph as shown in Figure 5.11. The value 1 on the y-axis corrsponds to

Success and the value 0 corresponds to Fail. This implies that the green regions of the

graph correspond to successful Trial’s and the white regions correspond to failed attempts.

Figure 5.11 Outcomes of 50 Trials for the Simulation Using Monocular Camera

Since 23 out of 50 Trials were successful, the controller has a success rate of 46%. Figure

5.12 shows snapshots of the robot picking up a block during several successful Trials for

Simulation 2.

51
 Figure 5.12 Instances of Successful Block Pickup in Simulation 2

5.3.1 Failure Analysis

The failure analysis for Simulation 2 is similar to that for Simulation 1. Figure 5.13

illustrates a few examples where the controller Fails. The following modes of failure were

observed: Failure Mode 1: Failure during navigation towards Material Repository (State

1, Section 4.3), Failure Mode 2: Errors in the object detection algorithm, and Failure

Mode 3: Failure due to inconsistent depth estimates from the depth estimation algorithm.

Failure Modes 1 and 3 have been discussed in detail in the previous section (Section 5.2.1).

The subsequent paragraphs elaborate on Failure Mode 3.

52
 Figure 5.13 Instances of Failed Block Pickup in Simulation 2

Failure Mode 3 in Simulation 2 occurs due to the variations in the depth estimates from

the depth estimation algorithm. Figure 5.14 illustrates variations in the depth map

generated by the algorithm for an input image during the robot’s transition from State 2 to

State 3 (Section 4.3). Brighter points in the image correspond to the regions closer to the

camera, according to the algorithm, and darker points correspond to regions farther away

from the camera. Such variations in the depth estimates cause the controller to misjudge

the distance of the block from the robot and result in a Failed Trial.

53
 Figure 5.14 Depth Map for Different Runs Before Block Pickup

However, a closer examination of the failed runs in Simulation 2 and Figure 5.14

provides insight into the margin by which the pickup fails in most Trials. Figure 5.15

illustrates the region around the block where the gripper attempts a pickup in the case of a

Fail. This region has an approximate margin of 2 inches on all sides measured with respect

to the vertical faces of the block. Thus, the lower success rate of the controller can largely

be attributed to the slight variations in the estimates of the depth estimation algorithm,

which causes the pickup to fail within a tolerance of approximately ± 2 inches on all sides

with respect to the block, which is a very low margin of error. Transfer learning the depth

estimation model on a custom dataset obtained from the simulations could improve the

stability of the depth estimates, thus increasing the success rate of the controller.

Figure 5.15 Typical Failure Envelope around a block in Simulation 2

54
 CHAPTER 6

CONCLUSION AND FUTURE WORK

6.1 Conclusion

In this thesis, we developed a novel vision-based controller that combines existing

techniques in deep-learning-based object detection and depth estimation using a monocular

camera. The controller is designed to be used for material pick-and-place tasks in

construction applications. Sensor information from a single monocular camera is the only

input that the controller requires. There is no global localization mechanism, and the

controller does not require prior mapping of the environment. Two simulations of 50 trials

each were performed. The first simulation used an RGB-D camera to benchmark the

controller’s performance, and the success rate was calculated to be 68% over the 50 trials.

In the second simulation, the RGB-D camera was replaced with a monocular camera, and

the success rate, in this case, was determined to be 46% over the 50 trials.

To better understand the controller’s performance, a thorough analysis of the cases under

which the controller fails was performed. Based on the analysis, the lower success rate

observed in the second simulation was attributed to the unstable depth estimates from the

depth estimation network. Thus, improving the stability of the depth estimates from the

network could potentially increase the success rate of the robot in Simulation 2.

6.2 Future Work

Some possible extensions of the work in this thesis are as follows:

55
• Train the depth estimation model using a custom dataset derived from the simulation
environment to obtain more stable and consistent depth estimates, which could improve

the success rate of the robot.

• Replace object detection with instance segmentation to uniquely identify different

instances of objects in a single camera frame, such as blocks of different colors and

shapes, which could potentially avoid Failure Mode 2 (Section 5.2.1 and Section 5.3.1).

• Train the object detection model to identify more classes of objects.

• Add force sensing or vision-based feedback to the arm, to enable the use of closed-loop
controllers that can compensate for errors during block pickup and deposition, thus

potentially increasing the success rate of the robot.

• Extend the proposed control approach to develop robot controllers for collective
construction tasks performed by multiple robots, incorporating rules for collision

avoidance and coordination among the robots.

56
 REFERENCES

Siciliano, B., Sciavicco, L., Villani, L., & Oriolo, G. (2012). Robotics: Modelling, Planning
and Control. Springer London.

Lynch, K. M., & Park, F. C. (2019). Modern Robotics: Mechanics, planning, and Control.
Cambridge University Press.

Tzafestas, S. G. (2014). Introduction to Mobile Robot Control. Elsevier.

Sandakalum, T., & Ang, M. H. (2022). Motion Planning for Mobile Manipulators—A
Systematic Review. Machines, 10(2), 97.
https://doi.org/10.3390/machines10020097

Szeliski, R. (2011). Computer vision algorithms and applications. Springer.

Ballard, D. H., & Brown, C. M. (1981). Computer Vision. Prentice Hall.

Huang, T.S. (1996). Computer Vision: Evolution and Promise. CERN School of
Computing, 21-25. DOI: 10.5170/CERN-1996-008.21

Zhao, C., Sun, Q., Zhang, C., Tang, Y., & Qian, F. (2020). Monocular depth estimation
based on deep learning: An overview. Sci. China Technol. Sci. 63, 1612–1627.
https://doi.org/10.1007/s11431-020-1582-8

Eigen, D., Puhrsch, C., & Fergus, R. (2014). Depth map prediction from a single image
using a multi-scale deep network. Proceedings of the 27th International
Conference on Neural Information Processing Systems, (2), 2366-2374.
https://dl.acm.org/doi/10.5555/2969033.2969091

Luo, X., Huang, J.-B., Szeliski, R., Matzen, K., & Kopf, J. (2020). Consistent video depth
estimation. ACM Transactions on Graphics, 39(4), 71:1-71:13
https://dl.acm.org/doi/abs/10.1145/3386569.3392377

Zhao, Z. Q., Zheng, P., Xu, S. T., & Wu, X. (2019). Object Detection with Deep Learning:
A Review, IEEE Transactions on Neural Networks and Learning Systems, 30(11),
3212-3232. doi:10.1109/TNNLS.2018.2876865.

Aggarwal, C. C. (2018). Neural Networks and Deep Learning. Springer Cham.

He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. IEEE International
Conference on Computer Vision (ICCV), 2980-2988.

Cortes, C., & Vapnik, V. (1995). Support vector machine. Machine Learning, 20(3), 273–
297. https://doi.org/10.1007/BF00994018

57
Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints.
International Journal of Computer Vision, 60(2), 91–110.
https://doi.org/10.1023/B:VISI.0000029664.99615.94

Mertan, A., Duff, D. J., & Unal, G. (2022). Single image depth estimation: An overview.
Digital Signal Processing, 123. https://doi.org/10.1016/j.dsp.2022.103441.

Joglekar, A., Joshi, D., Khemani, R., Nair, S., & Sahare, S. (2011). Depth Estimation Using
Monocular Camera. International Journal of Computer Science and Information
Technologies. 2(4), 1758-1763.

Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., & Koltun, V. (2022). Towards Robust
Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset
Transfer. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(3),
1623-1637. Doi: 10.1109/TPAMI.2020.3019967

Scharstein, D., & Szeliski, R. (2003). High-accuracy stereo depth maps using structured
light. 2003 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition. Proceedings., I-I, Doi: 10.1109/CVPR.2003.1211354.

Zhou, B., Krähenbühl, P., & Koltun, V. (2019). Does computer vision matter for
action? Science Robotics, 4(30). Doi: 10.1126/scirobotics.aaw6661

Geiger, A., Lenz, P., & Urtasun, R. (2012). Are we ready for autonomous driving? the kitti
vision benchmark suite. 2012 IEEE Conference on Computer Vision and Pattern
Recognition. 3354–3361. Doi: 10.1109/CVPR.2012.6248074.

Silberman, N., Hoiem, D., Kohli, P., &Fergus, R. (2012). Indoor Segmentation and Support
Inference from RGBD Images. In: Fitzgibbon, A., Lazebnik, S., Perona, P., Sato,
Y., Schmid, C. (eds) Computer Vision – ECCV 2012. ECCV 2012. Lecture Notes
in Computer Science, 7576. Springer, Berlin, Heidelberg.
https://doi.org/10.1007/978-3-642-33715-4_54

Saxena, A., Schulte, J., & Ng, A. Y. (2007). Depth estimation using monocular and stereo
cues. Proceedings of the 20th international joint conference on Artificial
intelligence (IJCAI'07). 2197-2203.

Ullman, S. (1979). The interpretation of structure from motion. Proc. R. Soc. Lond. B.203,
405–426. http://doi.org/10.1098/rspb.1979.0006

Mitchell, T. M. (1997). Machine Learning. McGraw Hill.

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified,
Real-Time Object Detection. 2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 779-788. Doi: 10.1109/CVPR.2016.91.

58
Redmon, J., & Farhadi, A. (2017). YOLO9000: Better, Faster, Stronger. 2017 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 6517-6525.

Redmon, J., & Farhadi, A. (2018). YOLOv3: An Incremental Improvement. ArXiv,

abs/1804.02767.

Bochkovskiy, A., Wang, C., & Liao, H.M. (2020). YOLOv4: Optimal Speed and Accuracy
of Object Detection. ArXiv, abs/2004.10934

Jocher, Glenn, et al. (2020) Ultralytics/Yolov5: V3.1 - Bug Fixes and Performance
Improvements. Zenodo, https://doi.org/10.5281/zenodo.4154370.

Pan, S. J., & Yang, Q. (2010). A Survey on Transfer Learning. IEEE Transactions on
Knowledge and Data Engineering, 22(10), 1345-1359. Doi:
10.1109/TKDE.2009.191.

Štibinger, P., et al., (2021). Mobile Manipulator for Autonomous Localization, Grasping
and Precise Placement of Construction Material in a Semi-Structured Environment.
IEEE Robotics and Automation Letters, 6(2), 2595-2602. Doi:
10.1109/LRA.2021.3061377.

Basiri, M., Gonçalves, J., Rosa, J., Vale, A., & Lima, P. (2021). An autonomous mobile
manipulator to build outdoor structures consisting of heterogeneous brick patterns.
SN Appl. Sci. 3, 558. https://doi.org/10.1007/s42452-021-04506-7

Lussi, M., et al., (2018) Accurate and Adaptive in Situ Fabrication of an Undulated Wall
Using an on-Board Visual Sensing System. 2018 IEEE International Conference
on Robotics and Automation (ICRA). 3532-3539. Doi:
10.1109/ICRA.2018.8460480.

Asadi, K., Haritsa, V.R., Han, K.K., & Ore, J. (2021). Automated Object Manipulation
Using Vision-Based Mobile Robotic System for Construction
Applications. Journal of Computing in Civil Engineering, 35, 04020058.

Wang, C., Zhang, Q., Tian, Q., Li, S., Wang, X., Lane, D., Petillot, Y., & Wang, S. (2020).
Learning Mobile Manipulation through Deep Reinforcement
Learning. Sensors, 20(3), 939. MDPI AG. http://dx.doi.org/10.3390/s20030939

Iriondo, A., Lazkano, E., Susperregi, L., Urain, J., Fernandez, A., & Molina, J. (2019). Pick
and Place Operations in Logistics Using a Mobile Manipulator Controlled with
Deep Reinforcement Learning. Applied Sciences, 9(2), 348. MDPI AG.
http://dx.doi.org/10.3390/app9020348

59
Delgado, J. M. D., Oyedele, L., Ajayi, A., Akanbi, L., Akinade, O., Bilal, M., & Owolabi,
H. (2019). Robotics and automated systems in construction: Understanding
industry-specific challenges for adoption, Journal of Building Engineering, 26.
https://doi.org/10.1016/j.jobe.2019.100868.

Jung, D. H., Park, J., & Schwartz, M. (2014). Towards on-site autonomous robotic floor
tiling of mosaics. 14th International Conference on Control, Automation and
Systems (ICCAS 2014), 59-63. Doi: 10.1109/ICCAS.2014.6987959.

Feng, C., Xiao, Y., Willette, A., McGee, W., & Kamat, V. R. (2014). Towards Autonomous
Robotic In-Situ Assembly on Unstructured Construction Sites Using Monocular
Vision. Proceedings of the 31st International Symposium on Automation and
Robotics in Construction and Mining (ISARC), 163–170.
doi:10.22260/ISARC2014/0022

Feng, C., Xiao, Y., Willette, A., McGee, W., & Kamat, V. R. (2015). Vision guided
autonomous robotic assembly and as-built scanning on unstructured construction
sites. Automation in Construction, 59, 128–138. doi:10.1016/j.autcon.2015.06.002

Asadi, K., Ramshankar, H., Pullagurla, H., Bhandare, A., Shanbhag, S., Mehta, P., Kundu,
S., Han, K., Lobaton, E., & Wu, T. (2018). Vision-based integrated mobile robotic
system for real-time applications in construction. Automation in Construction, 96,
470–482. doi:10.1016/j.autcon.2018.10.009

Cilia, J. (2019). The Construction Labor Shortage: Will Developers Deploy Robotics?
Forbes. https://bit.ly/3QD8aNn

Garg, S., Sunderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q.,
Chin, T., Reid, I.D., Gould, S., Corke, P., & Milford, M. (2020). Semantics for
Robotic Mapping, Perception and Interaction: A Survey. Found. Trends Robotics,
8, 1-224.

Petersen, K. H., Napp, N., Stuart-Smith, R., Rus, D., & Kovac, M. (2019). A review of
collective robotic construction. Science Robotics, 4(28), eaau8479.
doi:10.1126/scirobotics.aau8479

Werfel J., (2012). “Collective Construction with Robot Swarms”, Morphogenetic

Engineering: Toward Programmable Complex Systems, Springer Berlin
Heidelberg, pp. 115-140, ISBN: 978-3-642-33902-8.

Petersen, K., & Nagpal, R. (2017). Complex Design by Simple Robots: A Collective
Embodied Intelligence Approach to Construction. Archit. Design, 87, 44-49.
https://doi.org/10.1002/ad.2194

60
Florida, R., Rodríguez-Pose, A., & Storper, M. (2021). Cities in a post-COVID
world. Urban Studies. https://doi.org/10.1177/00420980211018072

Robotis. (n.d.). Turtlebot3 Emanual. Robotis Co., Ltd. https://bit.ly/3NdEhAj

D. Malyuta. (2017). Navigation, Control and Mission Logic for Quadrotor Full-cycle
Autonomy. Master thesis, Jet Propulsion Laboratory, 4800 Oak Grove Drive,
Pasadena, CA 91109, USA. https://doi.org/10.3929/ethz-b-000248154.

Shirsat, A., Elamvazhuthi, K., & Berman, S. (2020). Multi-Robot Target Search using
Probabilistic Consensus on Discrete Markov Chains. 2020 IEEE International
Symposium on Safety, Security, and Rescue Robotics (SSRR), 108–115.
doi:10.1109/SSRR50563.2020.9292589

Wang, J., & Olson, E., (2016). AprilTag 2: Efficient and robust fiducial detection.
Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS), pp. 4193–4198. https://doi.org/10.1109/IROS.2016.7759617

Coleman, D., Șucan, J. A., Chitta, S., & Correll, N., (2014). Reducing the Barrier to Entry
of Complex Robotic Software: a MoveIt! Case Study. Journal of Software
Engineering for Robotics, 5(1), 3-16. Doi: 10.6092/JOSER_2014_05_01_p3.

Șucan, I. A., Moll, M., & Kavraki, L. E., (2012). The Open Motion Planning Library. IEEE
Robotics & Automation Magazine, 19(4), 72–82. https://ompl.kavrakilab.org

Koenig, N., & Howard, A., (2004). Design and use paradigms for Gazebo, an open-source
multi-robot simulator. 2004 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), 3, pp. 2149-2154. Doi: 10.1109/IROS.2004.1389727.

Tsai, C. Y., Chou, Y. S., Wong, C., C., Lai, Y. C., & Huang, C., C. (2020). Visually Guided
Picking Control of an Omnidirectional Mobile Manipulator Based on End-to-End
Multi-Task Imitation Learning. IEEE Access, 8, pp.1882-1891,
doi:10.1109/ACCESS.2019.2962335.

Lu, Q., Fricke, G.M., Ericksen, J.C. et al. (2020). Swarm Foraging Review: Closing the
Gap Between Proof and Practice. Curr Robot Rep, 1, 215–225.
https://doi.org/10.1007/s43154-020-00018-1

Paszke, A., Chaurasia, A., Kim, S., & Culurciello, E. (2016). ENet: A Deep Neural
Network Architecture for Real-Time Semantic Segmentation. ArXiv,
abs/1606.02147.

Mur-Artal, R., Montiel, J. M. M., & Tardos, J. D., (2015). ORB-SLAM: A Versatile and
Accurate Monocular SLAM System. Trans. Rob. 31(5), 1147–1163.
https://doi.org/10.1109/TRO.2015.2463671

61
Labbé, M., Michaud, F., (2019). RTAB-Map as an open-source lidar and visual
simultaneous localization and mapping library for large-scale and long-term online
operation. J Field Robotics, 35, 416– 446. https://doi.org/10.1002/rob.21831

Asadi, K., Chen. P., Han, K., Wu, T., & Lobaton, E. (2019). LNSNet: Lightweight
Navigable Space Segmentation for Autonomous Robots on Construction
Sites. Data. 4(1), 40. https://doi.org/10.3390/data4010040

Ratliff, N., Zucker, M., Bagnell, J. A., & Srinivasa, S. (2009). "CHOMP: Gradient
optimization techniques for efficient motion planning". IEEE International
Conference on Robotics and Automation, pp. 489-494. Doi:
10.1109/ROBOT.2009.5152817.

Kalakrishnan, M., Chitta, S., Theodorou, E., Pastor, P., & Schaal, S. (2011). STOMP:
Stochastic trajectory optimization for motion planning. IEEE International
Conference on Robotics and Automation, pp. 4569-4574, Doi:
10.1109/ICRA.2011.5980280.

Kulkarni, A., Chong, D., & Batarseh, F. A. (2020). 5 - Foundations of data imbalance and
solutions for a data democracy. Data Democracy, pp. 83–106. doi:10.1016/B978-
0-12-818366-3.00005-8

Singh, P., Singh, N., Singh, K. K., & Singh, A. (2021). Chapter 5 - Diagnosing of disease
using machine learning. Machine Learning and the Internet of Medical Things in
Healthcare, pp. 89–111. doi:10.1016/B978-0-12-821229-5.00003-3

Shirsat, A., Mishra, S., Zhang, W., & Berman, S. (2022). Probabilistic Consensus on
Feature Distribution for Multi-Robot Systems With Markovian Exploration
Dynamics. IEEE Robotics and Automation Letters, 7(3), 6407–6414.
doi:10.1109/LRA.2022.3171905

Ozenne, B., Subtil, F., & Maucort-Boulch, D. (2015). The precision–recall curve overcame
the optimism of the receiver operating characteristic curve in rare diseases. Journal of
Clinical Epidemiology, 68(8), 855–859. doi:10.1016/j.jclinepi.2015.02.010

MBZIRC. (n.d.). Www.mbzirc.com. Retrieved August 20, 2022, from

https://www.mbzirc.com/

62
 APPENDIX A

YOLO v5 SYSTEM ARCHITECTURE

63
Figure A.1 YOLO V5s model architecture, Source: GitHub (Jocher, 2020)

64
 APPENDIX B

TRAINING RESULTS – OBJECT DETECTION ALGORITHM

65
Additional results on the performance of the transfer learning algorithm for object detection

are included in this section.

Figure B.1 Mean Average Precision (mAP) vs. Number of Epochs for (Left) Intersection
over Union (IOU) = 0.5, and (Right) IOU = 0.5-0.95

Figure B.2 Metrics Precision vs. Number of Epochs (Left) and Metrics Recall vs.
Number of Epochs (Right)

66
Figure B.3 Precision Curve

Figure B.4 Recall Curve

67
Figure B.5 Training: Box Loss (Top Left), (b) Class Loss (Top Right), and Object Loss
(Bottom) vs. Number of Epochs

68
Figure B.6 Validation: Box Loss (Top Left), Class Loss (Top Right), and Object Loss
(Bottom) vs. Number of Epochs

69