Mechanical Design and Simulation Studies of A Quadruped Robot Motion Control

System

by

Xiang Sheng
B.Sc., Harbin Institute of Technology, 2013

A Thesis Submitted in Partial Fulfillment of the

Requirements for the Degree of

MASTER OF APPLIED SCIENCE

in the Department of Mechanical Engineering

c Xiang Sheng, 2018

University of Victoria

All rights reserved. This thesis may not be reproduced in whole or in part, by
photocopying or other means, without the permission of the author.
 ii

Mechanical Design and Simulation Studies of A Quadruped Robot Motion Control

System

by

Xiang Sheng
B.Sc., Harbin Institute of Technology, 2013

Supervisory Committee

Dr. Yang Shi, Supervisor

(Department of Mechanical Engineering)

Dr. Ben Nadler, Departmental Member

(Department of Mechanical Engineering)
 iii

Supervisory Committee

Dr. Yang Shi, Supervisor

(Department of Mechanical Engineering)

Dr. Ben Nadler, Departmental Member

(Department of Mechanical Engineering)

ABSTRACT

This thesis focuses on mechanical design and simulation studies of a quadruped robot
motion control system, targeting at designing an autonomous legged robot. The
designed quadruped robot with “X”-configuration is developed for traversing rocky
and sloped terrain with a static walking gait.
The mechanical design of the quadruped robot is illustrated in Chapter 2, includ-
ing the main body design, leg design and component selection. In the design process,
appropriate mechanical structures are utilized to minimize the energy consumption.
To improve energy efficiency, a set of principles is proposed. Corresponding imple-
mentations are also concretely introduced in this chapter. In addition, to simplify the
mechanical structure of the quadruped robot, the mass is symmetrically distributed
about the frontal and lateral planes. To improve the leg agility, the leg mass is min-
imized. On the one hand, the lightweight design is implemented by optimizing the
mass distribution of the leg mechanism. On the other hand, the key components are
assembled in the body part instead of the legs as many as possible. A sufficient leg
length is also selected not only to allow the robot to step on or over obstacles, but also
to avoid the leg getting caught by objects. Particularly, the leg structure is demon-
strated, including the hip joint, thigh part, knee joint and limb part with a telescoping
joint. When the robot sustains extensive payload, the deformed shape in joints may
lead to structural failures, thereby influencing the quadrupedal locomotion. Finite
element analysis (FEA) is performed when designing the structural components in
reasonable structures. The design processes of the shoulder part and brass rod are
demonstrated as examples. Based on the setup of loads and fixtures, the maximum
 iv

deformed shape of these structural components are analyzed. From FEA simulation
results, the yield strength is two orders of magnitude larger than the maximum of
von Mises stress. Hence, these components are suitable to be incorporated in the
quadruped robot.
Based on the designed mechanical structure, simulation studies of the quadruped
robot motion control system are analyzed in Chapter 3, including the modeling for
a robotic leg and animated simulation. Since the quadrupedal locomotion is exe-
cuted by manipulating the postures of four legs, the leg model is significant to the
motion control system, thereby being analyzed mathematically. Two links kinematic
conversion is implemented between the foot-end trajectory and joint angles. The dy-
namic model of the leg is also computed to discovery the relationship between the
actuating torques and joint angles. To animate the quadrupedal locomotion, a CAD
robot model is converted into MATLAB. Following the predefined footfall pattern,
four legs move in sequence to execute the creeping gait. The segment of the desired
trajectory of the foot-end fits a fifth order polynomial and does not include the set
of singular configurations. Then, the PD control is utilized to adjust the leg pos-
ture to track the desired path. Furthermore, the actual joint angles are calculated in
the MATLAB/SimMechanics quadruped robot by using Euler-Lagrange equations.
Lastly, simulation results are presented to analyze the tracking performance in the
joint angles and foot-ends.
Finally, conclusions of the thesis are summarized, and future work is presented in
Chapter 4.
 v

Contents

Supervisory Committee ii

Abstract iii

Table of Contents v

List of Tables vii

List of Figures viii

Acknowledgements xi

Nomenclature xii

1 Introduction 1
1.1 Configurations of the Quadruped Robot . . . . . . . . . . . . . . . . 3
1.1.1 “M”-configuration . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 “X”-configuration . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.3 “O”-configuration . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Leg Mechanism of the Quadruped Robot . . . . . . . . . . . . . . . . 7
1.2.1 Robotic Legs with Compressed Air . . . . . . . . . . . . . . . 8
1.2.2 Compliant Legs with Steel Coil Springs . . . . . . . . . . . . . 10
1.3 Some Distinctive Mechanisms . . . . . . . . . . . . . . . . . . . . . . 12
1.3.1 Robots with One DOF for Each Leg . . . . . . . . . . . . . . 12
1.3.2 Robots with Flexor Reflex . . . . . . . . . . . . . . . . . . . . 13
1.4 Some Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Mechanical Design of a Quadruped Robot 17

2.1 Design Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
 vi

2.2 The Main Body Structure . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.1 Aluminum Plates . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.2 Motor Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.3 Timing Belts . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.4 Potentiometers . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.5 Gears . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 The Leg Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.1 Hip Joint Design . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.2 Thigh Part Design . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.3 Knee Joint Design . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.4 Limb Part Design . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 Finite Element Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.1 FEA for the Shoulder Part Design . . . . . . . . . . . . . . . . 32
2.4.2 FEA for the Brass Rod Design . . . . . . . . . . . . . . . . . . 35
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3 Simulation Studies of a Quadruped Robot 39

3.1 Modeling for a Robotic Leg . . . . . . . . . . . . . . . . . . . . . . . 40
3.1.1 Forward Kinematics . . . . . . . . . . . . . . . . . . . . . . . 40
3.1.2 Inverse Kinematics . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.3 Jacobian Matrix and the Singularities . . . . . . . . . . . . . . 45
3.1.4 The Euler-Lagrange Equations . . . . . . . . . . . . . . . . . . 47
3.2 Animated Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.1 Gait Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.2 PD Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.3 Generating Smooth Trajectories . . . . . . . . . . . . . . . . . 54
3.2.4 MATLAB/SimMechanics Model . . . . . . . . . . . . . . . . . 56
3.2.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4 Conclusions and Future Work 63

4.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

Bibliography 66
 vii

List of Tables

Table 1.1 Performance of representative quadruped robots developed in his-

tory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Table 2.1 Design parameters for the quadruped robot. . . . . . . . . . . . 19

Table 2.2 Key elements and their influence on desired characteristics. . . . 25

Table 3.1 Link parameters for the two-link planar leg. . . . . . . . . . . . 42

Table 3.2 Some parameters of the designed quadruped robot. . . . . . . . 57
 viii

List of Figures

Figure 1.1 Different configurations shown by “StarlETH” [1]. . . . . . . . . 4

Figure 1.2 Illustrations of “Cheetah” series with the “M”-configuration: (a)
“Cheetah I” [2] and (b) “Cheetah II” [3]. . . . . . . . . . . . . . 5
Figure 1.3 Illustrations of quadruped robots with the “M”-configuration by
Hiroshi Kimura et al.: (a) “Patrush I” [4] and (b) “Tekken I” [5]. 5
Figure 1.4 Illustrations of quadruped robots by Boston Dynamics Inc: (a)
“BigDog ”, (b) “LittleDog” and (c) “LS3” [2]. . . . . . . . . . . 7
Figure 1.5 (a) 3D one-legged hopping machine and (b) its pneumatic circuit
[6]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Figure 1.6 (a) The biped “Lucy” on treadmill, (b) its physical pendulum
and (c) 3 contraction levels of the PPAM [7]. . . . . . . . . . . 9
Figure 1.7 Illustration of “Mowgli”: (a) the jumping posture and (b) land-
ing posture [8]. . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Figure 1.8 (a) The ARL “Monopod” and (b) the schematic of its important
variables [9]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Figure 1.9 (a) The “KOLT” quadruped robot, (b) its electro-pneumatic leg
and (c) the schematic of its thrusting system [10]. . . . . . . . . 11
Figure 1.10Illustration of “Scout” series: (a) “Scout I” [11] and (b) “Scout
II” [12]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Figure 1.11(a)“Tekken II” and (b) flexor reflex activated on stumbling [13]. 14

Figure 2.1 The CAD model of a quadruped robot. . . . . . . . . . . . . . . 17

Figure 2.2 (a) Energy flow diagram of the quadruped robot. (b) Design
principles to improve energy efficiency. (c) Strategies to imple-
ment design principles. . . . . . . . . . . . . . . . . . . . . . . . 18
Figure 2.3 The CAD model of the body part. . . . . . . . . . . . . . . . . 20
 ix

Figure 2.4 Illustration of three types of aluminum plates in the body part:
(a) the aluminum plate A which is assembled to fix the poten-
tiometer at both ends of the body part, (b) the aluminum plate
B which is machined to clamp the motor and (c) the aluminum
plate C which is mounted in the middle of the body part to clamp
the motor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Figure 2.5 Illustration of two types of bearing housings in the body part:
(a) the bearing housing A to fix the potentiometer at both ends
and (b) the bearing housing B to clamp the ball bearing in the
middle portion. . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Figure 2.6 Illustrations of the motor and its connection: (a) the Swiss Maxon
DC motor and (b) the CAD model to illustrate the connection
of the motor, aluminum plate and bevel pinion. . . . . . . . . . 22
Figure 2.7 Illustrations of the potentiometer and its connection: (a) the
potentiometer–SAKAE 22HP-10, (b) the CAD model of the con-
nection among the motor, two bearings, two types of bearing
housings and the potentiometer and (c) section views of the CAD
model without the motor. . . . . . . . . . . . . . . . . . . . . . 23
Figure 2.8 Illustrations of the structure in the hip joint: (a) the brass rod,
(b) brass tube, (c) shoulder part, (d) pin joint and (e) key parts
of the hip joint. . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Figure 2.9 Illustrations of the thigh part: (a) the CAD model of the thigh
part with the brass rod, (b) the assembly of the thigh part and
(c) Section views of the CAD model of the thigh part with the
brass rod. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Figure 2.10Illustrations of the CAD model of thigh rails: (a) the inside thigh
rail and (b) outside thigh rail. . . . . . . . . . . . . . . . . . . . 29
Figure 2.11Illustrations of two types of bearing housings in the knee joint:
(a) the bearing housing C to fix the potentiometer at the outside
thigh rail and (b) the bearing housing D to clamp the ball bearing
at the inside thigh rail. . . . . . . . . . . . . . . . . . . . . . . . 30
Figure 2.12Illustrations of the knee joint: (a) the CAD model to illustrate
the knee joint and (b) the mechanical structure of the knee joint. 31
 x

Figure 2.13Illustrations of the limb part: (a) the CAD model of the limb
part, (b) the assembly of the limb part and (c) section views of
the CAD model of the limb part. . . . . . . . . . . . . . . . . . 33
Figure 2.14Illustrations of analysis results for the shoulder part: (a) von
Mises stress, (b) resultant displacement and (c) equivalent strain. 34
Figure 2.15Illustration of study results of a brass rod: (a) von Mises stress,
(b) resultant displacement and (c) equivalent strain. . . . . . . 37

Figure 3.1 Diagram of one leg with relative coordinates for each joint. . . . 40
Figure 3.2 Classification of periodic gaits [14]. . . . . . . . . . . . . . . . . 51
Figure 3.3 Internal contact directions of the creeping gait. . . . . . . . . . 51
Figure 3.4 The schematic model of a quadruped system. . . . . . . . . . . 52
Figure 3.5 Sequence of the locomotion cycle simulation in [15]. . . . . . . . 53
Figure 3.6 PD control schematics. . . . . . . . . . . . . . . . . . . . . . . . 55
Figure 3.7 Nominal trajectory of the leg. . . . . . . . . . . . . . . . . . . . 55
Figure 3.8 A posterior extreme set point for one step shown with Cartesian
coordinates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Figure 3.9 The simulated animation for the simplified model. . . . . . . . 57
Figure 3.10The MATLAB/SimMechanics quadruped model. . . . . . . . . 58
Figure 3.11Illustrations of the desired and tracked angle of the hip joint and
knee joint of hind legs: (a) the left-hind hip joint, (b) left-hind
knee joint, (c) right-hind hip joint and (d) right-hind knee joint. 59
Figure 3.12Illustrations of the desired and tracked angle of the hip joint and
knee joint of front legs: (a) the left-front hip joint, (b) left-front
knee joint, (c) right-front hip joint and (d) right-front knee joint. 60
Figure 3.13Illustrations of the desired and tracked angle of four foot-ends:
(a) the left-hind foot-end, (b) left-front foot-end, (c) right-hind
foot-end and (d) right-front foot-end. . . . . . . . . . . . . . . . 61

Figure 4.1 An overview of the control architecture for the quadrupedal lo-
comotion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
 xi

ACKNOWLEDGEMENTS

Firstly, I would like to show my sincerest gratitude to my supervisor, Dr. Yang

Shi, a decent and professional scholar. It is my honor to be a member of his research
group. He provided insightful guidance for my academic research and my personal
life. He provided insightful guidance, not only for my academic research, but also in
the aspects of personal life. He shared many tips and gave me advice in the individual
meetings and group meetings. He also is my role model to teach me how to handle
problems in a professional attitude. When I encountered problems, he encouraged
me with great enthusiasm, impressive kindness and extraordinary patience. I really
appreciate Dr. Yang Shi for providing me the opportunity to initiate my MASc study
in Canada.
I would like also to thank committee members, Dr. Ben Nadler and Dr. Hong-
Chuang Yang, for their constructive comments.
Moreover, I really enjoy to work with my group members in the Applied Control
and Information Processing Lab at the University of Victoria. I want to show my
gratitude to the senior students, Dr. Mingxin Liu, Dr. Bingxian Mu, Dr. Chao
Shen, Dr. Yuanye Chen, Jicheng Chen, and Kunwu Zhang. They not only helped
me countless times to solve academic problems during my research, but also offered
the help in life aspect. They helped me in adapting to the new study environment.
In addition, I want to thank current students, Qi Sun, Qian Zhang, Henglai Wei,
Tianyu Tan, Huaiyuan Sheng, Haoqiang Ji, Zhang Zhang, Chonghan Ma, Chen Ma,
Zhuo Li, Tingting Yu, Changxin Liu and Bo Cai for the happy and joyful time that
we have spent on Vancouver Island.
Last but not least, I would like to thank my parents for their help and support.
 xii

Acronyms

PC personal computer
DOF degrees of freedom
PAM pneumatic artificial muscle
PPAM pleated pneumatic artificial muscle
WSM wide stability margin
CPG central pattern generator
SLIP spring loaded inverted pendulum
IMU inertial measurement unit
IEPF iterative-end-point-fit
MPC model-predictive control
CAD computer-aided design
PCB printed circuit board
PD proportional-derivative
RISC reduced instruction set computing
ARM advanced RISC machine
FPGA field-programmable gate array
RPM revolutions per minute
FEA finite element analysis
LH left-hind
LF left-front
RH right-hind
RF right-front
COM center of mass
ESC electronic speed control
Chapter 1

Introduction

Mobile robots have been developed to replace human beings to work in some complex
and dangerous environments, such as humanitarian de-mining, disaster sites, and bat-
tlefields. For example, since increasing number of landmines have been deployed, the
detection and removal of antipersonnel landmines draw a wide range of concerns. To
clear the mines in an area efficiently and safely, the robot is required to be equipped
with sensitive sensors, efficient manipulators and advanced mobility functions [16].
The integrated robotic system is built based on a legged robot combining with a sen-
sor for the landmine detection, a manipulator to adjust the position of the sensor, a
global-position system to identify the landmine location. To accomplish these tasks,
legged robots exhibit potential advantages as follows:

• With solid footholds, legged robots reduce the possibility of stepping on an

antipersonnel mine.

• The inherent omnidirectionality of legged robots can steer the direction of move-
ment without forward/backward manoeuvers.

• Legged robots can step over loose and sandy terrain. Equipped with appropriate
tactile sensors, legged robots can identify the inclination angle of the contact
surface to prevent slippage.

• Legged robot can optimize the posture for the scanning manipulator to approach
landmines without changing its footholds.

Based on a legged robot with a hexapod configuration, the “DYLEMA” system is

developed to detect and locate landmines [16]. DYLEMA is a Spanish acronym which
 2

means efficient detection and location of antipersonnel landmines. This hexapod

platform can realize the locomotion in a relatively high speed. Meanwhile, the robot
can maintain stability when three feet alternately provide supports. In addition, a
remote operator controls the scanning manipulator through the teleoperation and
collaborative control strategies. By adjusting the position of the sensor head and
the posture of the scan manipulator, “DYLEMA” can approach the landmines in an
appropriate trajectory over terrain irregularities.
Normally, mobile robots can be classified into air robots [17], underwater robots
[18] and ground robots [19]. Generally, ground robots can be subdivided into tracked
robots, wheeled robots and legged robots [20]. As illustrated in [21], half of land
surface cannot be exploited by tracked robots and wheeled robots, but most of the
terrain is accessible to legged robots. Additionally, since the path of legged robots
is decoupled from the sequence of the footholds, the trajectory pattern is produced
[22]. Legged locomotion shows the outstanding performance in various kinds of soil
conditions [23] and obstacle avoidance [21]. Compared with wheeled and tracked
robots, legged robots present high efficiency in maintaining static balance with a
range of choices for footholds [24].
Lots of legged robots have been developed, inspired by the biological feature of
wild animals. By applying a reasonable mechanical structure and an appropriate
control strategy, legged robots are capable of moving on irregular terrain. Since most
of wild animals have the symmetric structure, legged robots usually have an even
number of legs [25]. This characteristic can help legged robots to realize efficient
gaits and stable locomotion [26]. Based on the number of legs [27], legged robots can
be classified into three categories:

• Biped robots with two legs [28],

• Quadruped robots with four legs [29],

• Insectoid robots with more than four legs [30].

Among these categories, quadruped robots are superior to biped robots on as-
pects of the static stability. Also, compared with insectoid robots, the design and
manufacture of quadruped robots are relatively easier due to the simple mechanical
structure. To find a compromise leg number for legged robots, following factors are
taken into consideration. On the one hand, reducing the leg number can simplify the
mechanical structure and control strategy. On the other hand, the stability of the
 3

alternative stepping motion is generally maintained by keeping the projected center

of mass (COM) within a polygon, which is formed by the stance legs. The legged
robot can keep the balance easier in the locomotion, when using more stance legs
alternatively to provide the support. Hence, it is a reasonable compromise to apply
four legs in one robot. By choosing alternatively three legs on the ground as the
support pattern, the quadruped robot can be a stable and active platform to fulfill
highly dynamic tasks, such as mild running turns and jumping over obstacles.

1.1 Configurations of the Quadruped Robot

A reasonable mechanical structure plays a significant role to accomplish various tasks.
Robots with different structures are adept at a diverse set of gaits, such as trotting,
bounding, and galloping. Since there are few mature quadruped products, it is hard
to define a standard mechanical structure of a versatile quadruped robot. Despite the
fact that lots of novel mechanisms have been designed, most of them are developed
to suit the designated condition. For example, the structure of the flexor reflex is not
suitable for the de-mining robot. When the flexor reflex is activated by obstacles, the
movement of the foot-end may detonate the mine and lead to an explosion. To avoid
the accident, the foot-end of the de-mining robot is designed as a narrow base. Hence,
the narrow connecting area can significantly reduce the possibility of the detonation,
when the foot-end strikes the ground.
In addition, apart from wild animals, the designers can also absorb the inspiration
from typical quadruped robots. As one of the pioneers in robotics research, Boston
Dynamics Inc. has developed advanced robots with remarkable performances, in
terms of mobility, agility, dexterity and speed. This section introduces several prod-
ucts which are developed by this group, such as “Cheetah I”, “BigDog”, “LittleDog”
and “LS3”.
Following the posture classification of “StarlETH” in [1], quadruped robots are
divided into the “X”-configuration, “O”-configuration and “M”-configuration. As
illustrated in Figure 1.1, when the orientations of the knee joints are towards inside
and outside, the postures correspond to the “X”-configuration and “O”-configuration,
respectively. When the orientations of the knee joints are towards the same direction,
the posture is named as the “M”-configuration. Quadruped robots with different
configurations are appropriate for various kinds of tasks. Quadruped robots with
the “X”-configuration are adept at static gaits on rough terrain locomotion, such as
 4

“LittleDog” [31]. Generally, quadruped robots with the “O”-configuration specialize

in carrying considerable payload, such as “LS3” [32]. Quadruped robots with the “M”-
configuration can perform highly dynamic tasks like running and jumping, such as
“Cheetah II” [7]. These configurations are extensively implemented in the mechanical
design of quadruped robots. The features of different configurations are illustrated
by comparing typical quadruped robots below.

Figure 1.1: Different configurations shown by “StarlETH” [1].

1.1.1 “M”-configuration
“Cheetah” series are typical sample robots with the “M”-configuration. As shown in
Figure 1.2a. “Cheetah I” (or called “WildCat”) is the fastest free running quadruped
robot in the world. The robot can run up to 9 m/s while maintaining its balance [2].
The outstanding running speed largely attributes to its articulated back which can
flex back and forth on each step. Since the elastic structure is implemented in the
body part like some mammals, its running speed can be considerably increased by
extending the stride length. Moreover, by upgrading the running controller of the first
generation, “Cheetah II” (see Figure 1.2b) can execute mild running turns and jumps
over obstacles autonomously [3]. Besides, by minimizing the vertical movements of
the body part, “Cheetah II” can also balance objects on its back even running on
rough terrain.
Hiroshi Kimura et al. from Tokyo Institute of Technology developed “Patrush
II” in 1994, as shown in Figure 1.3a. Each leg is actuated by three DC motors with
neural oscillation central pattern generator (CPG). By applying the biological-type
control, “Patruch II” can perform dynamic walking and present adaptive ability on
irregular terrain.
 5

(a) (b)

Figure 1.2: Illustrations of “Cheetah” series with the “M”-configuration: (a) “Chee-
tah I” [2] and (b) “Cheetah II” [3].

In addition, “Tekken I” with the “M”-configuration was developed around 2000

by Hiroshi Kimura et al.. “Tekken I” (see Figure 1.3b) is controlled by a personal
computer (PC) and is actuated by Switzerland Maxon DC motors. Each joint is
equipped with a photoelectric encoder, a gyroscope, an inclinometer and a tactile
sensor. The leg model is controlled by the neural oscillation CPG and a reflection
mechanism control system. Particularly, CPG generates the rhythmic motion, while
reflection mechanism gains the feedback from the sensor to change the step cycle and
phases of CPG. With the feedback information, this robot can walk not only on the
ground with pebbles and grasses, but also on hollows and slippery surfaces [13, 33].

(a) (b)

Figure 1.3: Illustrations of quadruped robots with the “M”-configuration by Hiroshi

Kimura et al.: (a) “Patrush I” [4] and (b) “Tekken I” [5].

1.1.2 “X”-configuration
Quadruped robots with the “X”-configuration can efficiently walk on rough terrain,
and even are capable of climbing in some circumstances. As two typical robots with
the “X”-configuration, “BigDog” [34] and “LittleDog”, which are developed by Boston
 6

Dynamics Inc. in 2005 and 2013 respectively, perform excellent locomotion on rough
terrain. “BigDog” is appropriate for a relatively smooth surface with dynamic gaits,
and “LittleDog” presents outstanding performance on rough terrain with static gaits.
As shown in Figure 1.4a., “BigDog” is a self-contained quadruped robot, and can
execute dynamic walking in the outdoor environment while carrying heavy loads. Four
articulated legs are incorporated into “BigDog” to absorb shock and to recycle energy.
The four-DOF leg has three active joints and one passive joint, which are built by
hydraulic cylinders and a pneumatic spring, respectively. This robotic platform has
successfully performed different kinds of locomotion gaits, such as walking, trotting,
and bounding. Equipped with a gasoline engine and hydraulic actuators, the robot
can carry up to 154 kg payload and last for around 2.5 hours. The power supply is
provided by an internal combustion engine to drive a hydraulic pump, which delivers
the hydraulic oil to the actuators.
Also, Boston Dynamics Inc. developed “LittleDog” which can traverse a variety
of rough terrain quickly and robustly [31], as shown in Figure 1.4b. “LittleDog” is
designed for research on learning locomotion by applying motor learning, dynamic
control and terrain estimation. For each leg of “LittleDog”, two motors are used
to actuate the hip joint and one motor is utilized to drive the knee joint with an
85:1 ratio. The mass of the body is much larger than the compliant leg, which are
1.8 kg and 0.25 kg, respectively. “LittleDog” is equipped with a variety of sensors
including an IMU, 3-axis force sensors, a high-resolution camera, and so on. Besides,
six degrees of freedom (DOFs) position information is provided by a Vicon motion
capture environment, which is operated for the quadrupedal locomotion. With the
help of a sub-millimeter accuracy terrain map, desired trajectories are computed by
an embedded 1 kHz microprocessor, in which an internal proportional-derivative (PD)
controller is programmed. In addition, an external computer transmits other control
commands through a wireless connection.

1.1.3 “O”-configuration
As a typical robot with the “O”-configuration, Legged Squad Support System (“LS3”)
quadruped robot (see Figure 1.4c) has potential capability in delivering extensive
loads.
“LS3” is designed as a rough-terrain robot to help marines and soldiers to carry
their load in the battlefield [32]. An “LS3” can carry up to 181 kg of load for a mission
 7

(a) (b) (c)

Figure 1.4: Illustrations of quadruped robots by Boston Dynamics Inc: (a) “BigDog
”, (b) “LittleDog” and (c) “LS3” [2].

covering a 32 km and lasting for 24 hours. By obtaining terrain mapping from stereo
and LIDAR sensors, “LS3” can perform autonomous navigation without a dedicated
operator, by using terrain sensors and GPS to identify designated locations. “LS3”
can keep its paces in uneven and slippery surfaces and can stand upright in spite of
some external disturbances.

1.2 Leg Mechanism of the Quadruped Robot

Since the quadrupedal locomotion is executed by controlling the movement of four
robotic legs, the leg mechanism plays an essential part of the whole quadruped struc-
ture. A multi-degree-of-freedom robotic leg is supposed to incorporate a light-weight
construction, low impedance transmission and high torque density motors. A force
sensor is usually assembled in the foot-end to measure the direction of contact. Also,
by using cameras and various sensors, the locomotion can be controlled by the be-
haviour generation module, behaviour control module, and low-level joint controller.
The leg should be carefully designed in several perspectives, such as increasing the
height of center of gravity of the leg and narrowing the volume of the foot-end. The
design objectives of a quadruped robot are to perform highly dynamic tasks and to
execute rough terrain locomotion.
Generally, the robotic leg implements three types of leg mechanisms, including the
compressed air, steel coil springs and compliant mechanisms in [35]. Since the latest
one is not implemented in the designed quadruped robot, this mechanism is briefly
introduced at the end of this subsection. Furthermore, the leg mechanism can also
be classified into linear joints and articulated joints, according the leg configuration.
The features of leg mechanisms are illustrated below for each category.
 8

1.2.1 Robotic Legs with Compressed Air

To increase the energy transformation ratio, the inherent compliant behavior of pneu-
matic actuators can be used to realize mutual transformation of kinetic energy and
potential energy. A considerable energy capacity can be generated by compressing
air in a container which is made of steel, fiberglass and rubber in tension. The high
specific strength can be generated by two common compressed air structures, which
are the pneumatic cylinder and pneumatic artificial muscles (PAMs). From the bio-
logical perspective, this mechanisms play a role of the muscle which is composed of
elastic tendons.

3D One-legged Hopping Machine (linear joints) As a hopping and running

robot with a pneumatic cylinder [22], the three dimensional (3D) one-legged hopping
machine is designed by Raibert et al. in 1986. This robot incorporates a linear tele-
scoping joint to adjust the leg length with hydraulic actuators and air springs. The
3D one-legged hopping machine consists of a body and a compliant leg, as shown
in Figure 1.5a [6]. The body comprises actuators, valves and other electronic equip-
ments. The compliant leg is actuated by a pneumatic cylinder. For the pneumatic
cylinder, a rubber cushion and linear potentiometer are assembled at the bottom and
top, respectively. The compliant leg can minimize the energy consumption in contin-
uous hopping by applying the internal energetic conversion. In addition, the control
system regulates the pressure of the lower chamber by changing the air capacity. The
upper chamber acts as a passive spring, and connects a pressure regulator with a check
valve, as shown in Figure 1.5b [6]. The hopping motion is triggered by controlling the
air capacity in the lower chamber. In the stance phase, the low chamber exhausts the
air to provide a thrust. The thrust is maximized by activating quick-exhaust values.
The spring-mass oscillator is controlled by charging and exhausting the air to adjust
the pressure. Without external support, the robot can hop at a speed up to 2.2 m/s.
By controlling the air capacity, the pneumatic cylinder can adjust the leg stiffness
in the swing phase. From lots of experimental results, the high pressure can increase
the stiffness of spring and reduce the duration of the stance phase. In this way, the
robot can hop in a relatively high speed.

Biped “Lucy” and “Mowgli” (articulated joints) As two typical types of

PAMs, pleated pneumatic artificial muscles (PPAMs) and Mckibben PAMs are uti-
lized in the biped “Lucy” and “Mowgli”, respectively.
 9

(a) (b)

Figure 1.5: (a) 3D one-legged hopping machine and (b) its pneumatic circuit [6].

The biped “Lucy” has been developed as a six-DOF biped robot which is actuated
by twelve PPAMs. As shown in Figure 1.6a, the upper body and two compliant legs
are connected by two one-dimensional pin joints [7], which can avoid legs turning over
in the frontal plane. Since the PPAM can only be pulled, each articulated joint of the
biped “Lucy” is actuated by a pair of PPAMs (see Figure 1.6b) as a bidirectionally
working revolute joint. Unlike the pneumatic cylinder, the joint angles and the leg
stiffness are controlled independently, and are measured by different gauge pressures
separately. By switching the valve, the control strategy is implemented by a one-DOF
pendulum, as illustrated in Figure 1.6c.

(a) (b) (c)

Figure 1.6: (a) The biped “Lucy” on treadmill, (b) its physical pendulum and (c) 3
contraction levels of the PPAM [7].

Besides, a jumping robot named “Mowgli” has been constructed with the McK-
ibben pneumatic muscle, as shown in Figure 1.7. Six McKibben PAMs are incorpo-
rated into two articulated legs to actuate hip, knee and ankle joints. However, unlike
 10

the biped “Lucy”, the artificial muscles of “Mowgli” are composed of rubber tubes.
For the electro-pneumatic system, “Mowgli” has the potentiometer, pressure sensor
and touch switch, which are assembled on the joint, muscle and foot-end, respectively.
With 3 kg weight, “Mowgli” can jump up to 0.5 m with soft landings, which is more
than 50% of its body height (0.9 m with legs extended).

(a) (b)

Figure 1.7: Illustration of “Mowgli”: (a) the jumping posture and (b) landing posture
[8].

1.2.2 Compliant Legs with Steel Coil Springs

As a popular mechanical structure, steel coil springs are widely incorporated in com-
pliant legs. Several legged robots apply this mechanism at hip joints, such as the
ARL “Monopod” and “KOLT”.

ARL “Monopod” (linear joints) The Ambulatory Robotics Lab (ARL) “Mono-
pod” is developed by Buehler et al., which is incorporated with the mechanical struc-
ture of the pneumatic cylinder and steel coil springs. As shown in Figure 1.8a, the
ARL “Monopod” consists of a body and a prismatic leg [9]. The motor torque is
converted into an axial force by a timing belt and pulley, as shown in Figure 1.8b.
The pitch motion of the body is determined by the leg swing. The 55 Nm hip torque
is actuated by an 80 W motor and is transmitted to the joint by a 30:1 gear ratio.
From experimental results, the ARL “Monopod” can move at a speed of 1.2 m/s.
By including the leg spring in the leg mechanism, the hip swing motion minimizes
the energy consumption with the design. The ARL “Monopod” performs the stable
locomotion, including the leg swing phase and vertical hopping motion.
 11

(a) (b)

Figure 1.8: (a) The ARL “Monopod” and (b) the schematic of its important variables
[9].

“KOLT” (articulated joints) The Kinetically Ordered Locomotion Test (“KOLT”)

robot (see Figure 1.9a) is developed to perform the galloping gait by Ohio State Uni-
versity and Stanford University [36].
Inspired by the biological feature of the goat leg, the electro-pneumatic leg of
“KOLT” comprises the body, thigh, shank and coil springs [10]. To decrease the
impact force at collision, the compliant leg applies the pneumatic cylinder with low
inertia, as shown in Figure 1.9b.

(a) (b) (c)

Figure 1.9: (a) The “KOLT” quadruped robot, (b) its electro-pneumatic leg and (c)
the schematic of its thrusting system [10].

For the electro-pneumatic leg, the hip motion and knee motion are set to be
decoupled with the help of a cable and pulleys, as shown in Figure 1.9c. A pneumatic
spring is assembled in the thigh part to store elastic energy. The knee joint determines
the leg stiffness, which is 15.9 kN/m with the leg fully extended. To connect the
pneumatic spring and air cylinder, a check valve is used to retain the pressurized air
 12

in the pneumatic cylinder.

For the compliant leg configuration, except for compressed air and steel coil
springs, compliant mechanisms are also a potential option for the robotic leg de-
sign. Thanks to the development in manufacturing technologies and new materials,
novel leg structures, in which every portion of each leg is elastic, have been designed.
This configuration is widely applied in a variety of robots, such as “Bow Leg Hop-
per” [37] and “RHex” [38]. Using compliant material in the compliant legs is also an
option to obtain leg compliance. Specifically, the “MABEL” robot incorporates the
fiberglass into the transmission system [39]. Several robots apply flexure joints, such
as “DASH” [40] and “HAMR3” [41]. The Sprawl robots incorporate flexural elements
in hip joints to execute the rotation in both stroke and pitch directions [42].

1.3 Some Distinctive Mechanisms

Generally, there are three actuated DOFs in each leg for the quadruped robot. The
foot-end is designed as a ball to minimize ground contact area. Since most of terrains
provide few available footholds, the small contact area can provide better balance
in the quadrupedal locomotion. However, in the robotic history, some distinctive
structures have been developed to accomplish tasks in specific terrains.

1.3.1 Robots with One DOF for Each Leg

Usually, for a quadruped robot, each robotic leg has three actuated DOFs. There are
two actuated DOFs in the hip to swing the thigh link in the frontal plane and lateral
plane. An active rotational joint is assembled in the knee to swing the limb link in the
lateral plane. The roll movement in the hip helps the robot not only to realize walking
on an inclined slope in an upright stance but also to take turns during locomotion
[13]. However, a quadruped robot has been developed by McGill University, whose
leg has one actuated DOF. This mechanism can minimize the energy consumption
and simplify the control strategy.
This special mechanism is developed by McGill University, which has been incor-
porated in two generations of four-legged robots “Scout I” (see Figure 1.10a) and
“Scout II” (see Figure 1.10b). The “Scout” series are designed to perform the bound-
ing gait[11]. In the locomotion, the robot never entirely left the ground. Each leg has
only one actuated DOF, which is powered by an actuator in the hip [24]. Although
 13

having a relatively simple mechanical design, the robot in the series can successfully
achieve dynamic stability. “Scout II” is a self-propelled running robot, which has only
one actuator for each hip as well [12]. Besides the hip actuator, a compliant prismatic
joint is also incorporated in the leg. The controller can determine the locomotion of
the robot by adjusting angular positions and torques of the front and hind legs [43].
The simplification for the leg structure can minimize the energy consumption and
enhance mechanical reliability. With simple stiff legs, these quadruped robots can
realize walking and stair climbing in bounding.

(a) (b)

Figure 1.10: Illustration of “Scout” series: (a) “Scout I” [11] and (b) “Scout II” [12].

Experimental results illustrate that “Scout II” can run up to 1.5 m/s in bound-
ing. This robot is highly dominated by the passive dynamics to realize a high efficient
energy transformation. Overall, “Scout” series set an example of using a relatively
simple control strategy and a simple mechanical structure. However, the mechan-
ical structure cannot actuate a compliant leg with a relatively long length. This
disadvantage may restrict the available footholds on rough terrain.

1.3.2 Robots with Flexor Reflex

Normally, a small contact area at toes is widely used to increase the adaptability
on irregular terrain, such as “LittleDog” from Boston Dynamics Inc. and “SQ3U”
from Lynxmotion Inc. However, along with the development of the robot mechanism,
novel structures have been designed for navigating and traversing rough terrain.
As a self-contained quadruped robot, “Tekken II” has been developed to incor-
porate novel mechanisms, which can be actuated on stumbling [13]. Similar to the
configuration of “Tekken I”, “Tekken II” is approximately 25% larger than the previ-
 14

ous generation in both size and weight. As illustrated in Figure 1.11a, there are four
joints on each leg of “Tekken II”. DC motors (23 W) are used to actuate the hip joint
around the pitch axis, while the motion around the yaw axis is activated by motors
(8 W). A spring-lock mechanism is incorporated in the ankle joint. By controlling hip
yaw joints, the locomotion direction can be adjusted. Rate gyros and inclinometers
are applied to measure the pitch and roll angles of the body part.
The flexor reflex is designed to tackle the collision avoidance with obstacles. Dur-
ing the flexion phase of the step cycle, a stimulus on the paw dorsum produces an
enhanced flexion in order to avoid obstacles. Inspired by a spinal cat, an ankle joint
with a spring and a lock are designed in “Tekken I”. This mechanism can realize
obstacle avoidance by adjusting the posture of the foot-end during the first half of a
swing phase. The sequence of flexor reflex is illustrated in Figure 1.11b [13]. In the
first step (also called the flexor neuron active phase), the knee joint is flexed after
the robot detects the stumbling. Then, the flexor reflex is lifted in the second step.
After the leg escapes the stumbling condition, the ankle joint will move back to the
initial situation. The flexor reflex can be triggered by enlarging the desired angle in
a certain range. The knee joint angle can be measured at the moment of the flexor
neuron active phase. Under the effect of the rubber elasticity, the ankle joint can
return back to the initial angle, when the leg gets rid of the influence from obstacles.
However, since lots of small bumps and pebbles present considerable friction on rough
terrain, the passive ankle mechanism do not perform well in the outdoor environment.
Furthermore, “Tekken II” also applies other excellent structures to improve the
locomotion performance, such as low leg inertia moment and powerful actuators. In
addition, a small gear reduction ratio is applied to reinforce the compliance of joints.

Step 1 Step 3

Step 2

(a) (b)

Figure 1.11: (a)“Tekken II” and (b) flexor reflex activated on stumbling [13].
 15

Table 1.1: Performance of representative quadruped robots developed in history.

Mass Payload Vmax DOF

Name Year Weight Gait
[kg] [m/s] (per leg)
“KUMO-I” [44] 1976 14 — static 0.1 3
“PV-II” [45] 1979 10 — static 0.1 3
“TITAN VI” [46] 1993 195 0.06 static 0.2 4
“JROB 1” [47] 1994 31.5 — static 0.2 3
“Scout II” [12] 1998 2.3 0.02 dynamic 1.5 1
“RHex” [38] 1999 7 — dynamic 0.4 1
“Tekken” [13] 2003 4.3 0.2 dynamic 0.9 4
“BigDog” [34] 2008 75 1.41 dynamic 3.1 3
“LittleDog” [31] 2011 2.8 — static 0.1 3
“Cheetah II” [48] 2014 33 — dynamic 6.4 3
“ERS-7” [49] 2017 1.5 — static 0.3 3

1.4 Some Comments

Although quadruped robots have some distinct advantages comparing to other mo-
bile robots, they have not been widely used in practical engineering field due to some
limitations. For the traditional walking mechanism, such as four telescoping massless
legs, the desired trajectory of the foot-end is limited by the set of singular configura-
tions. Hence, quadruped robots have to adjust the step length to deal with obstacle
avoidance with path planning. Besides, a comprehensive controller is implemented to
manipulate four legs for rough terrain locomotion.
In addition, the payload is restricted by the strength of each joints. To minimize
energy consumption, the quadruped robot should be designed in the lightweight con-
struction. The injuries in the joints should also be avoided which might be caused by
excessive musculoskeletal forces at collision.
Furthermore, noise from quadruped robots becomes one crucial problem, such
as “LS3” from Boston Dynamics Inc. This limitation of the robot itself prevents
the potential possibility of the battlefield, since a loud robot may give away their
position. After all, it is relatively difficult to find a reasonable compromise in both
decent payload and high-speed locomotion with low noise.
Quadruped robots can only perform tasks in some specific occasions to achieve
limited goals. Moreover, the mechanical structure of the quadruped robot is also
 16

determined by the working condition. Some exposed components will be corroded

inevitably in hostile environments. For example, equipped with a pressure sensor and
a signal processing system, the foot-end compromises the shape memory alloy, which
may be damaged by friction and corrosion. Table 1.1 lists some quadruped robots
developed in the last half century with main characteristics.

1.5 Thesis Organization

The thesis is organized as follows. Chapter 1 starts with reviewing the fundamentals
and development of quadruped robots. The motivation of quadruped robots is high-
lighted in this chapter. The feature and performance of several typical quadruped
robots are presented and compared.
The mechanical structure of the designed quadruped robot is described in Chap-
ter 2, such as the body part, hip joint, thigh part, knee joint and limb part. Key
components are also introduced, such as bevel gears, motors, potentiometers, ball
bearings and timing belts. In addition, since the deformed shape may lead to struc-
tural failures, thereby influencing the quadrupedal locomotion. The finite element
analysis (FEA) is performed when designing some structural components in reason-
able structures. The design processes of the shoulder part and brass rod are described
as examples.
Chapter 3 illustrates the modeling for a robotic leg and animated simulation of the
designed quadruped robot in MATLAB. Kinematic models and Euler-Lagrange equa-
tions are formulated for the robotic leg. An animated simulation of the quadrupedal
locomotion is emulated with the creeping gait. The desired trajectory of the foot-
end is also determined to avoid the set of singular configurations. Euler-Lagrange
equations are calculated in the MATLAB/SimMechanics model to compute the ac-
tual joint angles. PD control is implemented to track the desired trajectory. The
evaluation of simulation results is also presented in Chapter 3.
Chapter 4 summarizes the work in this thesis. The future work is also presented
in this part.
 17

Chapter 2

Mechanical Design of a Quadruped

Robot

This chapter focuses on the mechanical design of the designed quadruped robot which
is composed of a body part and four compliant legs. To simplify the mechanical
structure of the designed quadruped robot, the mass distribution is symmetrical about
the frontal and lateral planes. The mechanical structure is built in a computer-aided
design (CAD) model (see Figure 2.1). Particularly, the leg mechanism includes a hip
joint, a thigh part, a knee joint and a limb part with a telescoping joint. In the sequel,
we will introduce the mechanical structure of each part.

Body

Limb Hip

Foot-end
Thigh

Knee

Figure 2.1: The CAD model of a quadruped robot.

 18

2.1 Design Principles

Inspired by the “LittleDog” design in [31] from Boston Dynamics, Inc., a quadruped
robot is designed for research purposes, such as legged locomotion and autonomous
navigation. The system energy flow of a typical quadruped robot is illustrated in
Figure 2.2a. The Joule heating loss mainly occurs at the rotating motor side and the
printed circuit board (PCB). Besides, some mechanical energy is consumed due to
the mechanical transmission friction and interaction friction. To minimize the energy
consumption, some principles are illustrated in Figure 2.2b, including high torque
density motors, reverse current braking, low impedance transmission, low friction in
rotational joints, low inertia legs and internal energetic conversion [50]. To realize
these design principles for energy efficiency improvement, several mechanical struc-
tures and appropriate components have been designed, as illustrated in Figure 2.2c.
These principles are implemented in the mechanical design of the quadruped robot.

System Energy Flow Principles Implementations

High Torque Large Gap
Energy Source
Density Motors Radius Motors
(Battery)
Reverse Current Two Motors
Actuator Braking Drive a Hip Joint
𝑬𝒋 Joule Heating
(DC Motor)
Low Impedance High Gear Ratio Bevel
Transmission Gears in Knee Joints
Mechanical
𝑬𝒇 Friction Low Friction in Pin Joints and Miter
Transmission
Rotational Joints Gears in Hip Joints
Positive work Negative work
(𝑾𝒑𝒐𝒔 ) (𝑾𝒏𝒆𝒈 )
Low Inertia Legs Lightweight Design of Legs
Mechanical
𝑬𝒊 Interaction Energy Internal Energetic
(𝑬𝒌 + 𝑬𝒑 ) Linear Telescoping Joints
Conversion

(a) (b) (c)

Figure 2.2: (a) Energy flow diagram of the quadruped robot. (b) Design principles
to improve energy efficiency. (c) Strategies to implement design principles.

The designed quadruped robot comprises the body part and four compliant legs.
The mass of the body is 4 kg, and the mass of each leg is 1 kg. The whole robot is
actuated by twelve motors to perform passive dynamic walking. The rotation angle
of each motor is measured by a potentiometer.
 19

Table 2.1: Design parameters for the quadruped robot.

Parameters Value
Body length 0.521 m
Body width 0.146 m
Body thickness 0.095 m
Leg length 0.152 m
Ground clearance 0.254 m
Body mass 4 kg
Individual leg mass 1 kg
Total mass 8 kg
Leg spring compliance 5250 N/m
Swiss Maxon DC motor
DC servo motor
41.022.022
SAKAE 22HP-10
Potentiometer
20 KΩ

2.2 The Main Body Structure

To reduce the mass of the platform, the components with small sizes and mass are
selected to realize a lightweight design. The designed quadruped robot is mainly
constructed by aluminum alloys 6061, which is a high strength material with a great
range of mechanical properties. The key components are assembled in the main body
part as many as possible, as shown in Figure 2.3. In this way, the leg weight is reduced
as much as possible to improve the leg agility. Besides, the lightweight design of leg
links can ensure low rotational inertia and reduce the payload for actuators. Hence,
the body part is designed to comprise eight motors, eight potentiometers, sixteen
bearings and eight timing belts. These components are symmetrically distributed in
the body part. Particularly, a timing belt, which transfers the power from the motor
side to the hip joint side, is chosen to connect two synchronizing pulleys.

2.2.1 Aluminum Plates

The body part is designed as a symmetrical structure. To assemble key components in
the body part, three types of aluminum plates are machined, as shown in Figure 2.4.
Two pieces of each type are placed in the front and back of the body part, respectively.
Two pieces of the aluminum plate A, as shown in Figure 2.4a, are placed at two ends
 20

Figure 2.3: The CAD model of the body part.

of the body part. As shown in Figure 2.4b and Figure 2.4c, the other two types of
aluminum plates are assembled in the body part, respectively. These six plates are
used to fix two types of brass shafts via sixteen ball bearings.

(a) (b) (c)

Figure 2.4: Illustration of three types of aluminum plates in the body part: (a) the
aluminum plate A which is assembled to fix the potentiometer at both ends of the
body part, (b) the aluminum plate B which is machined to clamp the motor and (c)
the aluminum plate C which is mounted in the middle of the body part to clamp the
motor.

Four circle grooves are machined on the surface of the aluminum plate A as bearing
housings to clamp four ball bearings. Two clearance holes and two circle grooves (play
a role as bearing housing) are manufactured in each piece of the other two types of
plates. Particularly, the clearance holes in Figure 2.4b and Figure 2.4c are machined
for brass shafts and servo motors, respectively. The brass shaft is used to connect
 21

the motor, pulley, ball bearing and potentiometer to transmit the actuating torque.
Hence, the potentiometer can measure the angular position of the motor shaft. The
pulley can transmit the actuating torque to the hip joint. Two types of bearing
housings are designed to be mounted on the body part, as shown in Figure 2.5. The
bearing housing A in Figure 2.5a is assembled in the middle of the body part to clamp
the ball bearing. The bearing housing B in Figure 2.5b is placed on aluminum plates,
which are illustrated in Figure 2.4b and Figure 2.4c. This bearing housing not only
can help the ball bearing to clamp the brass shaft, but also can fix the housing of the
potentiometer.

(a) (b)

Figure 2.5: Illustration of two types of bearing housings in the body part: (a) the
bearing housing A to fix the potentiometer at both ends and (b) the bearing housing
B to clamp the ball bearing in the middle portion.

2.2.2 Motor Selection

Twelve motors are assembled in the quadruped robot to actuate four hip joints and
four knee joints. Eight of them are assembled in the body part, and one motor is
assembled in each thigh part of legs. The motor is appropriately selected to provide
the sufficient torque to the joint without loss of motion, meaning that no skipping
or slipping happens in the quadrupedal locomotion. To provide a sufficient driving
torque to help bevel gears to perform effective rotations, the low-speed/high-torque
motor is selected as the actuator.
The Swiss Maxon DC motor with a dual encoder is selected as the actuator in the
designed robot. The rated speed of the motor is 46 revolutions per minute (RPM),
and the rated power is around 5 W. With the help of bevel gearboxes, actuating forces
are redirected in both hip joints and knee joints. Generally, to assemble a motor in
 22

the robot, three screws are used to fix the motor at the aluminum plate B and C. In
the aluminum plate B and C, small clearance holes are machined to clamp motors.
To actuate the hip joint, motors are horizontally fixed on the aluminum plate B and
C, which are shown in Figure 2.4b and Figure 2.4c. Through brass shafts, motors in
the body part are connected to potentiometers. Meanwhile, the torque is transformed
from the motor side to the drive gear by the timing belt.

(a) (b)

Figure 2.6: Illustrations of the motor and its connection: (a) the Swiss Maxon DC
motor and (b) the CAD model to illustrate the connection of the motor, aluminum
plate and bevel pinion.

2.2.3 Timing Belts

Eight timing belts are used to transfer the actuating torque from the motor shaft
to the steel shoulder shaft. To improve the measurement accuracy, aluminum alloy
synchronous pulleys, with 24 teeth and a diameter of 10 mm, are chosen as the timing
belt. Hence, these two pulleys can rotate at the same speed, and the torque generated
from the motor side can be transmitted to the joint side. By applying this design, the
load can be successfully sustained on the teeth of gears with an acceptable deflection.
To improve the torque transmission efficiency, the distance between two pulleys is
chosen such that the belt is assembled with an appropriate tension, otherwise, the
excessive tension will add extra stress on the belt. A proper belt tension can avoid
slipping in a high load or bearing an extra load. In tests, we measure the load in
different center distances between two synchronous pulleys. Based on measurement
results, the appropriate distance is chosen as 44.5 mm to ensure proper belt tension.
 23

(a) (b)

(c)

Figure 2.7: Illustrations of the potentiometer and its connection: (a) the
potentiometer–SAKAE 22HP-10, (b) the CAD model of the connection among the
motor, two bearings, two types of bearing housings and the potentiometer and (c)
section views of the CAD model without the motor.

2.2.4 Potentiometers
SAKAE 22HP-10 potentiometers, pictured in Figure 2.7a, are chosen because of their
precise ±0.25% linearity to the measurements of the angular positions of the hip and
knee joints. The potentiometer is a three-terminal resistor whose rated power is 2 W
and the resistance is 20 KΩ. To measure the angular position of the hip joint, a brass
shaft is used to connect the shafts of the potentiometer to motor by a spring pin and
a set screw, respectively. By using a nut and a tooth lock washer, the housing of the
potentiometer is clamped to the bearing housing and fixed to the aluminum plate in
 24

the body part. The angular position can be measured by the angle difference between
the shaft and the housing of the potentiometer.
Figure 2.7b illustrates the connection among the motor, pulley, and potentiometer
in the body part. A brass shaft is used to transmit the actuating torque from the
motor side to the joint side. To prevent bending or twisting in the brass shaft,
the shaft length is designed as short as possible while the position of the motor is
appropriately selected. Two bearings are used to fix the brass shaft to two aluminum
plates. The motor is placed at one end of the brass shaft, while the potentiometer is
mounted on the other end. The similar structure is applied to measure the angular
position of the knee joint.

2.2.5 Gears
The bevel gearbox plays a central role to transmit the torque from the motor side to
the joint side and change the operating angle. Bevel gears in the designed quadruped
robot are mounted on shafts, which are capable of supporting high forces and are 90◦
apart. To reduce the risk of skipping in the hip joint and knee joint, large gears with
deep teeth are selected. In addition, couplings are not implemented in the body and
compliant legs, which reduces the possibility of slipping.
For the hip joint, to execute the rotation of the thigh part, miter gears are mounted
on a brass rod and two brass tubes. Unlike the mechanical structure of the hip joint,
the knee joint is actuated through a bevel gear pair with a 5:1 ratio. Either type of
joints transmits the rotational motion at a 90◦ angle. To actuate the joint effectively,
bevel gears should be precisely mounted on the input link and the output link. In
addition, to avoid concentrating the load at the end of the tooth, bevel gears should
be assembled with minor adjustments. This assembly allows the displacement of
bevel gears, which are caused by deflection under operating loads. The incorrect
installation may damage the joint.

2.3 The Leg Structure

Four compliant legs are designed to connect the body part. Each leg consists of a
hip joint, a thigh part, a knee joint and a limb part. To design artificial quadruped
legs, desired characteristics, such as speed, endurance, agility and strength, are incor-
porated into mechanical design. To realize these characteristics, main considerations
 25

are summarized in Table 2.2 and are enumerated as follows:

• The effective leg length will influence the speed of movement and the agility of
the leg. The stride length can be increased by extending the leg length. Hence,
the mobile speed of the quadruped robot can be improved. In addition, extend-
ing the leg length can reduce the energy cost of transport, thereby improving
the endurance.

• Mass distribution affects the speed of movement by influencing the stride fre-
quency. In this way, the posture of the robotic leg would be easily controlled,
when key components are assembled in the body part as many as possible.

• Leg kinematics determines the endurance. The hip joint can rotate in the frontal
plane and lateral plane. The knee joint rotates in the frontal plane. The bevel
gear pairs with high gear ratio is supposed to be used in the joints to reinforce
the endurance. As a prismatic joint, the compressed spring can be assembled in
the limb link to adjust the leg stiffness and to absorb the impact at collision.

• Elastic energy storage in tendons improves agility. By applying internal ener-

getic conversion, the compressed spring can improve the endurance. The tendon
stiffness influences the duration of the stance phase.

• Muscle power is determined by the actuating torque, which is provided by the

DC motor. Hence, the speed and strength of the compliant leg are influenced
by the rated power of the motors.

Table 2.2: Key elements and their influence on desired characteristics.

Speed Endurance Agility Strength

Effective length X X
Mass distribution X X
Kinematics X X
Elasticity X X X
Muscle power X X

In the process of the leg design, above elements are taken into consideration. Table
2.2 demonstrates the relationships between desired characteristics and technological
instantiations.
 26

(a) (b)

(c) (d)

(e)

Figure 2.8: Illustrations of the structure in the hip joint: (a) the brass rod, (b) brass
tube, (c) shoulder part, (d) pin joint and (e) key parts of the hip joint.

2.3.1 Hip Joint Design

Hip joints are supposed to be strong enough to support the body weight including
some attached peripherals in the stance phase. Besides, the hip joint should be able
to sustain the leg weight in the swing phase. The hip joint comprises the shoulder
part, the brass rod, three bevel gears and two pulleys. As shown in Figure 2.8c, the
shoulder part is composed of a steel shaft and two brass tubes. Figure 2.8d illustrates
the connection of the pin joint, which consists of the shoulder part and the brass rod.
 27

To provide the support force in the hip joint, three bevel gears are mounted in
the pin joint. The steel shoulder shaft and brass rod are perpendicularly assembled.
For each hip joint, two bevel gears are mounted on two brass tubes which are placed
in the steel shaft as drive gears. The third one is mounted on the brass rod which
is placed in the thigh part as the drive gear. For each hip joint, two miter gears are
mounted on two brass tubes, and are driven by two groups of timing belts separately.
Since two bevel gears are actuated by two motors, they do not need to synchronize
with each other. When rotating directions of the two gears are the opposite, the
thigh part swings forward or backward. When rotating directions of the two gears
are the same, the thigh part is actuated to move in the frontal plane. By controlling
the angle difference between gears, the direction of the quadrupedal locomotion can
be adjusted. Mounted on two brass tubes, two drive gears are able to rotate in
opposite directions. These two tubes are assembled on the shoulder steel shaft and
are separated by two delrin washers. The steel shaft connects two aluminum plates
of the body part. Two delrin washers can avoid the friction loss between two brass
tubes. With the help of the lubricant, two brass tubes can rotate around the steel
shaft flexibly.
On the leg side, the hip joint of the quadruped robot is designed in a frontal and
vertical sliding mechanism. The pin joint (see Figure 2.8e) prevents the leg from
rotating in the horizontal plane. Additionally, when the drive gear rotates around
the brass rod, the angular offset is transferred to the hip joint to actuate the thigh
part.
Furthermore, the position of the drive gear is determined by the length of the brass
rod. The hip joint can transmit the actuating torque efficiently through the miter
gearbox. Correct assembly of three miter gears can avoid skipping in the mechanical
transmission.

2.3.2 Thigh Part Design

For each compliant leg, a motor is mounted between thigh rails to actuate the knee
joint. Although the traditional serial link robot increases the inertia of the distal
links, the benefit arises in the simplification of the control strategy.
To reduce the weight of the whole robot, one potential solution is to shorten the
leg length. However, thigh rails should be designed to cover the motor and to protect
components from unexpected forces, as shown in Figure 2.9a. In addition, designed
 28

(a) (b)

(c)

Figure 2.9: Illustrations of the thigh part: (a) the CAD model of the thigh part with
the brass rod, (b) the assembly of the thigh part and (c) Section views of the CAD
model of the thigh part with the brass rod.

with sufficient lengths, leg links can help the robot to step onto or over obstacles.
After taking above factors into consideration, the lengths of the thigh link and limb
link are both selected around 0.15 m. The step height can reach up to approximately
0.1 m.
As shown in Figure 2.9c, two partially threaded cap screws are applied at the side
 29

surface to fix the inside thigh rail, aluminum tube and driven gear. On the one hand,
the inside thigh rail and aluminum tube are machined with two clearance holes at
their side surfaces. On the other hand, two threaded holes are machined in the side
surface of the driven gear. For each hole, a partially threaded cap screw is used to fix
the thigh rail, aluminum tube and driven gear successively through these holes at the
side surface. The aluminum tube connects the brass rod through two ball bearings.
Besides, circle grooves are machined in thigh rails and aluminum tubes as bearing
housings, which are used to clamp ball bearings to hold the pin joint, as shown in
Figure 2.10b. Two partially threaded cap screws are applied to fix the outside thigh
rail and aluminum tube through the clearance hole and threaded hole.

(a) (b)

Figure 2.10: Illustrations of the CAD model of thigh rails: (a) the inside thigh rail
and (b) outside thigh rail.

As illustrated in Figure 2.10, the inside and outside thigh rail have the similar
outline. However, the upper portions of these thigh rails are different. Figure 2.10a
shows that two clearance holes are machined in the side surface. A relatively large hole
is designed for the aluminum tube to be assembled. As shown in Figure 2.10b, a cycle
groove is machined in the outside thigh rail as a bearing housing. Two clearance holes
are manufactured for two cap screws to fix the aluminum tube to the outside thigh
rail. For each thigh part, a motor is placed vertically between thigh rails to actuate
the knee joint. The actuating force is redirected by a bevel gearbox. An aluminum
plate is placed horizontally between thigh rails to clamp the motor. Moreover, this
aluminum plate is fixed by four standard rounded head screws to provide a channel
 30

bracing for thigh rails to avoid collapsing. A bevel pinion is mounted on the motor
shaft through a collar.

2.3.3 Knee Joint Design

The knee joint is actuated by a bevel gear pair with a 5:1 ratio. By applying the bevel
gearbox, the actuating force is redirected to drive the limb part. The bevel pinion is
vertically placed between thigh rails. A set screw is applied to fix the bevel pinion
to the motor shaft. The thread size of the set screw is M3, and its length is 3 mm.
The driven gear is mounted on the steel shaft by a steel pin. Two alloy steel flat-tip
set screws are used to fix the driven gear to the outside limb rail. The steel shaft is
fixed to the inside limb rail by a partially threaded screw on the side surface. The
steel shaft is applied to connect the thigh rails through two ball bearings. Overall,
the motor drives a bevel pinion to actuate the limb part via a large couple gear.
Meanwhile, the pair of straight-tooth bevel gears, being mounted on perpendicular
shafts, is used to transmit torques from the motor side to the joint side.

(a) (b)

Figure 2.11: Illustrations of two types of bearing housings in the knee joint: (a)
the bearing housing C to fix the potentiometer at the outside thigh rail and (b) the
bearing housing D to clamp the ball bearing at the inside thigh rail.

Potentiometers are assembled to measure the angular position of knee joints. Ball
bearings are mounted on thigh rails to fix rotating shafts and are clamped by bearing
housings. The stress pattern is press-fit among the brass rod, ball bearing and thigh
rail. Two ball bearings are assembled to fix both ends of the brass rod. Each ball
bearing is clamped by an aluminum plate and a bearing housing. Two types of bearing
housings are used in the body part. As shown in Figure 2.12, the potentiometer is
fixed on the brass shaft by a spring pin. The potentiometer can obtain the angular
position of the knee joint by measuring the deviation between the thigh part and limb
 31

part. On the one hand, the knee brass shaft and bevel gear rotate with the limb and
the shaft of the potentiometer. On the other hand, the housing of the potentiometer
rotates with the thigh part. Hence, the measured information can be transmitted to
the controller in real time.

(a) (b)

Figure 2.12: Illustrations of the knee joint: (a) the CAD model to illustrate the knee
joint and (b) the mechanical structure of the knee joint.

2.3.4 Limb Part Design

The compression spring is incorporated into the limb structure design. This mech-
anism can not only prevent the impact at collision, but also can change the limb
length in the locomotion. To meet the required stiffness of the limb, a compression
spring with the appropriate compressed rate should be carefully selected. Conse-
quently, the linear telescoping joint is included in the limb part to adjust the leg
length. Mutual transformations between the kinetic energy and potential energy are
realized by utilizing the inherent compliance behavior of the telescoping joint. When
the foot-end sustains an exterior force, the telescoping joint will be compressed. Then
the compressed gas and mechanical spring provide a high force to the compliant leg.
Meanwhile, the high pressure from the foot-end increases the stiffness of the limb part
and reduces the duration of the stance phase.
As shown in Figure 2.13b, the delrin tube and rod are machined to be press
fit. The mechanical spring not only absorbs energy from the external load such as
the unexpected impact from the ground, but also reduces the energy cost during
the continuous movement by applying internal energy conversion. By performing
a relatively high-frequency hopping in a reasonable stride length, the compliant leg
can realize the resonant bounding motion. In this way, the quadruped robot can
 32

perform multi-modal locomotion, such as the trotting gait, pacing gait and bounding
gait. This special mechanism reduces the energy consumption during continuous leg
movements, and the passive prismatic joint can reduce the vibration of the body part
effectively. Moreover, all parts should be machined within an appropriate tolerance
in the machining process. Otherwise, machining errors may cause unsuccessfully
installation, which may damage the mechanical structure.

2.4 Finite Element Analysis

Accurate physical models of the designed quadruped robot is drawn in the CAD
software. Besides, the predefined information of the designed CAD quadruped model
is illustrated in SolidWorks, such as masses and link lengths. Since the deformed
shape may lead to structural failures, the finite element analysis (FEA) is performed
when designing some structural components. After finding out the fragile portion of
the whole robot, we can strengthen the designed mechanical structure by selecting
proper materials and designing reasonable structures.
Generally, bending and twisting may happen in long brass shafts, when they
sustain extensive payloads. For example, in the body part, a brass shaft is used to
connect a motor to a potentiometer. This brass shaft also transfers the motor torque
to actuate the hip joint through timing belts and gearboxes. In this process, the brass
shaft suffers from the compressive stress and tensile stress, and the deformation of the
shaft may be triggered. The deformed shape may lead to structural failures, thereby
influencing the quadrupedal locomotion. By applying FEA to the CAD model, the
extreme pressure-point can be found. Several actions are taken to strengthen the
fragile position by adjusting the shape and material of structural components.

2.4.1 FEA for the Shoulder Part Design

The shoulder part is expected to sustain the leg weight in the swing phase and to
support the body weight in the stance phase. FEA is applied in the process of the
shoulder part design. In the shoulder part, two brass tubes are mounted on the steel
shaft.
 33

(a) (b)

(c)

Figure 2.13: Illustrations of the limb part: (a) the CAD model of the limb part, (b)
the assembly of the limb part and (c) section views of the CAD model of the limb
part.

Loads and Fixtures

In the FEA simulation, the shoulder part sustain the leg, whose mass is 1 kg. Hence,
the load is set as 10 N and distributed in the small cylinder portions of two brass
 34

(a)

(b)

(c)

Figure 2.14: Illustrations of analysis results for the shoulder part: (a) von Mises
stress, (b) resultant displacement and (c) equivalent strain.

tubes. As shown in Figure 2.8e, the steel shaft is fixed to the body part by two
aluminum plates. The fixture type is set as fixed geometry. Based on the above
information, the strain distribution is analyzed.
 35

Results

Figures 2.14 illustrates the compressive and tensile strain distribution of the steel
shoulder shaft with two brass tubes. As mentioned in Section 2.1, a steel shaft
connects two aluminum plates and provides a support force for the compliant leg in
the swing phase. Two brass tubes are mounted on the steel shaft. A gear and a pulley
are mounted on each of brass tubes to actuate the hip joint. Hence, two miter gears
are able to rotate against each other. Figures 2.14a shows the distribution of von
Mises strain in steel shaft. The von Mises strain of the shoulder part is distributed
along lines extending from two ends to the midpoint. The transverse compressive
strain is mainly distributed in the middle position of the steel shoulder shaft and the
area around grooves. As shown in Figure 2.14a, the minimum von Mises stress is
approximately 15.45 N/m2 . The utmost pressure-point is the midpoint of the steel
shaft, which sustains 6.55×106 N/m2 .
Figure 2.14b illustrates that the maximum displacement of the shoulder part is
3.65×10−3 mm. In Figure 2.14c, the equivalent strain of the steel shaft with two brass
tubes presents a range from 2.50×10−10 to 2.36×10−5 . Overall, these analysis results
show the stress condition of the shoulder part. The mechanical structure meets the
expected stress distribution.

2.4.2 FEA for the Brass Rod Design

In the shoulder part, two brass tubes are mounted on the steel shaft. With the help
of the lubricating oil, two tubes can rotate against each other. Two delrin rings are
used to separate two tubes to prevent the frictional loss. The outside diameter of the
delrin ring is smaller than the outside diameter of the brass tube. By adjusting the
size of the delrin ring, the extra frictional loss can be avoided. The brass rod can
provide the support force to the steel shaft, which is the utmost pressure-point in
the shoulder part. In this way, the support force of the hip joint can be enforced to
sustain the leg weight.
For the hip joint, the brass rod sustains the leg weight during the swing phase and
bears the body weight during the stance phase. By applying the FEA, we optimize the
shape of the brass rod. After several tests, the brass rod is determined as Figure 2.6a.
The model information and study result of the brass rod is illustrated in sequence.
The structure of the shoulder part is determined according to the data of the von
Mises stress, resultant displacement and equivalent strain.
 36

Loads and Fixtures

The brass rod sustains a reaction force 10 N through a driven gear and two ball
bearings, as shown in Figure 2.9c. Particularly, the bevel gear is mounted on the
middle cylinder of the brass rod. Two ball bearings are used to fix the small cylinder
of the brass rod to the aluminum rod. Hence, the force is applied on three contact
faces. For the fixture details, the clearance hole in the brass rod is set as the fixture,
whose type is fixed geometry.

Results

Figure 2.15 shows the analysis results of the brass rod, such as the distribution of
compressive and tensile stresses.
In Figure 2.15a, the von Mises stress of the brass rod ranges from 8.71 N/m2 to
2.09×106 N/m2 . The data of von Mises stress illustrates that the utmost pressure
position is the connection area between the small brass cylinder and middle brass
cylinder. In addition, the yield strength of the brass is 2.39×108 N/m2 . Hence, the
maximum pressure on the brass rod is smaller than the yield strength of the material.
The brass rod can successfully sustain the leg weight.
Figure 2.15b shows the resultant displacement, when the load is applied in the
brass rod. As expected, the connection portion between the small cylinder and middle
cylinder sustains the maximum deflection. The maximum resultant displacement is
approximately 9.44×10−3 mm.
Figure 2.15c shows the distribution of equivalent strain in the brass rod. The area
between two cylinders of different diameters sustains the maximum equivalent strain,
which is 1.50×10−5 . The plot indicates the minimum equivalent stain is 5.71×10−10 .
In summary, from the FEA of the shoulder shaft, the middle portion sustains
extensive pressure. A brass rod is designed to provide a support force by wrapping
this fragile position. Hence, the shoulder part and brass rod constitute a pin joint
which is the key part of the hip joint to execute the leg movement. The brass rod
can help the shoulder part to sustain the leg weight in the swing phase and to bear
the body weight in the stance phase. In addition, to reduce the pressure of the fragile
portion of the steel shaft, three miter gears are mounted in the hip joint to bear the
load as well. In this way, three miter gears not only drive the thigh part accurately,
but also to some extent bear the leg weight during the swing phase.
 37

(a)

(b)

(c)

Figure 2.15: Illustration of study results of a brass rod: (a) von Mises stress, (b)
resultant displacement and (c) equivalent strain.

2.5 Conclusions
In this chapter, the mechanical structure of the designed quadruped robot is illus-
trated, aiming to develop a legged robot. The designed symmetrical structure can
realize the bi-direction locomotion and simplify the control problem to some extent.
 38

The mechanical design consists of main body design, compliant leg design, joint de-
sign, and component selection. From the mechanical perspective, the inertia of the
body and leg is reduced by adjusting the positions of components and by changing
the shapes of thigh and limb rails. Besides, to increase the stiffness of the compliant
leg, a linear telescoping joint is incorporated into the limb part. The passive prismatic
joint is included in the leg structure to absorb the impact force at collision and to
prevent unforeseen perturbations in a passive way. With the designed mechanical
structure, the quadruped robot can perform various types of behaviour gaits, such as
trotting, pacing and bounding.
The mechanical structure of the quadruped robot is analyzed and verified by the
FEA simulations. The shoulder part is taken as an example to demonstrate the use
of FEA in the mechanical design. The middle position of the shoulder part is the
utmost pressure-point. To sustain the extensive load, a brass rod is used to provide
an additional support to this fragile portion. Due to the deflection under operating
loads, the maximum displacement of the rod end is approximately 9.44×10−3 mm,
which is acceptable in experiments. For both components, since the yield strength is
two orders of magnitude larger than the maximum of von Mises stress, the design of
structure components is reasonable.
 39

Chapter 3

Simulation Studies of a Quadruped

Robot

Simulation studies of a quadruped robot motion control system are illustrated in this
chapter, including modeling for a robotic leg and the animated simulation. Since the
quadrupedal locomotion is implemented by controlling the postures of four legs, the
leg model plays an essential role in the motion control system. The leg model needs
to be analyzed mathematically, facilitating the controller design to track the desired
trajectory. To solve this tracking problem, kinematic and dynamic models of the leg
are derived in light of the work in [51]. Two links kinematic conversion is calculated
between the foot-end trajectory and joint angles. The dynamic model is implemented
to compute the relationship of the actuating torques and joint angles.
After analyzing the leg model, the collaboration of four robotic legs is developed
for the quadrupedal locomotion. By setting various footfall patterns, different kinds
of gaits can be performed. A four-beat gait has been defined based on the legged
motion analysis in [52]. One step cycle of the designed gait is divided into the swing
and stance phases. In the swing phase, the contact schedule is associated to the
creeping gait and predefined in Section 3.2.1. In the stance phase, the footholds are
maintained while shifting the body part forward.
PD control is applied for each joint to operate the leg posture. The corresponding
actuating torques are computed to manipulate the foot-ends to realize path following.
Moreover, the segment of the desired trajectory of the foot-end in the swing phase
fits a fifth order polynomial (results in relatively smooth accelerations). A CAD
robot model is converted into MATLAB to animate the quadrupedal locomotion.
 40

Simulation results are demonstrated to analyze the tracking performance in the joint
angles and foot-ends.

3.1 Modeling for a Robotic Leg

Since the configurations of four legs are similar, the right front leg model is analyzed
as an example, which is shown in Figure 3.1. The robotic leg consists of a set of
links connected by joints. Kinematic models and Euler-Lagrange equations are for-
mulated for the robotic leg in this section. By using kinematic equations, the foot-end
trajectory and joint angles can be converted to each other. The set of singular con-
figurations is not included in the desired trajectory by calculating the singularities
of the Jacobian matrix. Euler-Lagrange equations are implemented to compute the
actual angular positions based on the joint torques in the MATLAB/SimMechanics
quadruped model.

Leg base 𝜽𝟏
𝒛𝟎 𝒐𝟎
𝒚𝟎 𝒙𝟎
𝒍𝟏
𝜽𝟐
𝒚𝟏 𝒐𝟏

𝒛𝟏 𝒙𝟏

𝒍𝟐
𝒚𝟐 𝒐𝟐
Foot End
𝒛𝟐 𝒙𝟐

Figure 3.1: Diagram of one leg with relative coordinates for each joint.

3.1.1 Forward Kinematics

Forward kinematic equations are applied in the leg model to compute the joint an-
gles based on the foot-end trajectories. A set of conventions has been developed to
 41

provide a systematic way for this analysis. The Denavi-Hartenberg (DH) convention
is implemented in this two-link model, according to [51].
In the leg model, two revolute joints are taken into consideration, and the prismatic
joint in the limb part is not considered. The rotations of the hip and knee joints in
the lateral plane manipulate the quadruped robot to move forward and backward.
The movement direction is maintained which is controlled by rotating the hip joint in
the frontal plane. Since we only consider the rectilinear translation in this thesis, the
rotation of the hip joint in the frontal plane is not taken into consideration. Besides,
joint friction is ignored in the quadruped model. The frame of the foot-end (the
tool frame) is fixed based on DH coordinate frame assumptions [51]. By the DH
convention, there exists a unique homogeneous transformation matrix Ai that takes
the coordinates from the base frame oi−1 xi−1 yi−1 zi−1 to the tool frame oi xi yi zi . This
matrix Ai is depicted as a product of four basic transformation matrices

Ai =Rotz,θi T ransz,di T ransx,li Rotx,αi

    
cθi −sθi 0 0 1 0 0 0 1 0 0 li 1 0 0 0
    
 0 cαi −sαi
sθi cθi 0 0 0 1 0 0  0 1 0 0  0
=
0
  
 0 1 0  0 0 1 d  0
 i  0 1 0  0 sαi cαi
 0
0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1
 
cθ −sθi cαi sθi sαi ai cθi
 i 
sθi cθi cαi −cθi sαi ai sθi 
=
 , (3.1)
 0 s θi c θi d 
i 

0 0 0 1

where
   
cθ −sθi 0 0 1 0 0 0
 i   
sθi cθi 0 0 , T ransz,d = 0 1 0 0

Rotz,θi =
0 i
,
 0 1 0
0
 0 1 di 

0 0 0 1 0 0 0 1
   
1 0 0 0 1 0 0 li
   
0 cαi −sαi 0 , T ransx,l = 0
 1 0 0
Rotx,αi =
0 i
,
 s α i cα i 0
0
 0 1 0 
0 0 0 1 0 0 0 1
 42

cθi = cosθi , sθi = sinθi , cαi = cosαi and sαi = sinαi according to [51]. In addition, θi ,
li , di , αi are parameters associated with the link i and the joint j and denote joint
angle, link length, link offset and link twist, respectively. θi and αi are unique angles
in the model within a multiple of 2π. θi is the joint valuable for a revolute joint, and
di is the joint valuable for a prismatic joint.
The robotic leg is defined using Cartesian coordinates. The coordinate frames of
the leg are depicted in Figure 3.1. The base frame o0 x0 y0 z0 is established at the hip
joint of the leg model, where the origin o0 is located at the middle point of the brass
rod between the thigh rails. The o1 x1 y1 z1 frame is fixed at the knee joint, as shown
by the DH convention. The origin o1 is assigned at the middle point of the brass shaft
between the knee rails. The tool frame o2 x2 y2 z2 is fixed by setting the origin o2 at
the foot-end. l1 and l2 denote the thigh and limb lengths, respectively. The angular
positions of the hip and knee joints are θ1 and θ2 , respectively. The DH parameters
are illustrated in Table 3.1.

Table 3.1: Link parameters for the two-link planar leg.

Link Name a α d θ
Thigh l1 0 0 θ1
Limb l2 0 0 θ2

From (3.1), the A-matrices are given as

 
c1 −s1 0 l1 c1
 
s1 c1 0 l1 s1 
A1 = 
0
,
 0 1 0  
0 0 0 1

where c1 = cos(θ1 ) and s1 = sin(θ1 ), and

 
c2 −s2 0 l2 c2
 
s2 c2 0 l2 s2 
A2 = 
0
,
 0 1 0  
0 0 0 1

where c2 = cos(θ2 ) and s2 = sin(θ2 ).

 43

So, the T-matrices can be calculated as follows:

T10 = A1
 
c1 −s1 0 l1 c1
 
s1 c1 0 l1 s1 
=0

 0 1 0  
0 0 0 1
T20 = A1 A2
  
c1 −s1 0 l1 c1 c2 −s2 0 l2 c2
  
s1 c1 0 l1 s1  s2 c2 0 l2 s2 
=0
 
 0 1 0 0
 0 1 0  
0 0 0 1 0 0 0 1
 
c12 −s12 0 l1 c1 + l2 c12
 
s12 c12 0 l1 s1 + l2 s12 
=0
,
 0 1 0 

0 0 0 1

where c12 = cos(θ1 + θ2 ) and s12 = sin(θ1 + θ2 ). The first two entries of the last
column of T20 are the x and y coordinates of the foot-end in the base frame, and they
are defined as

x = l1 c1 + l2 c12 ,
(3.2)
y = l1 s1 + l2 s12 .

3.1.2 Inverse Kinematics

After obtaining the x and y coordinates in (3.2), we can use inverse kinematics equa-
tions in [51] to calculate joint angular positions θ1 and θ2 . From (3.2), we get

x2 + y 2 = l12 + l22 + 2l1 l2 c2 . (3.3)

Using the Law of Cosines to solve (3.3) for c2 , we obtain

x2 + y 2 − l12 − l22
c2 = ,
2l1 l2
q
s2 = ± 1 − c22 . (3.4)
 44

The multiple solutions in (3.4) correspond to the “elbow-up” position and “elbow-
down” position, respectively. Then, we compute θ2 by using the two-argument arct-
angent routine
θ2 = atan2(s2 , c2 ).

This approach ensures that all solutions are found in the proper quadrant.
After finding θ2 , we solve (3.2) for θ1 . Particularly, (3.2) can be rewritten as

x = k1 c1 − k2 s1 ,
(3.5)
y = k1 s1 + k2 c1 ,

where

k1 = l1 + l2 c2 ,
k2 = l2 s2 .

In order to solve (3.5), r and γ are defined as

q
r= k12 + k22 ,

and
γ = atan2(k2 , k1 ).

Then, we have the following results

k1 = r cos γ,
(3.6)
k2 = r sin γ.

By substituting (3.6) into (3.5), we get

x
= cos γ cos θ1 − sin γsin θ1 ,
r
y
= cos γ sin θ1 + sin γcos θ1 ,
r

or equivalently,

x
cos(γ + θ1 ) = ,
r
y
sin(γ + θ1 ) = .
r
 45

By applying the two-argument arctangent, we get

y x
γ + θ1 = atan2 , = atan2(y, x).
r r

Hence,
θ1 = atan2(y, x) − atan2(k2 , k1 ). (3.7)

Substituting (3.6) into (3.7), we obtain

θ1 = atan2(y, x) − atan2(l2 s2 , l1 + l2 c2 ). (3.8)

Note that in (3.4), the sign of s2 determines the value of k2 , thereby influencing the
value of θ1 according to (3.8). These two solutions correspond to two different pos-
tures of the leg. The angular position of the hip joint (θ1 ) and knee joint (θ2 ) can
be computed based on the related parameters, such as the desired position and link
length. In summary, forward and inverse kinematics can realize mutual transforma-
tion between the foot-end trajectory and joint angular positions.

3.1.3 Jacobian Matrix and the Singularities

To identify a set of singular configurations in the leg movement, the singularities of
the Jacobian matrix are calculated. By avoiding the singular configurations in the
desired trajectory, the movement direction of the foot-end is arbitrary. Hence, the
leg can swing freely to track the desired trajectory of the foot-end.
According to the model of a two-link planar manipulator in [51], the Jacobian
matrix is a 6 × 2 matrix and can be defined as
" #
z0 × (o2 − o0 ) z1 × (o2 − o1 )
J(θ) = , (3.9)
z0 z1

where z0 × (o2 − o0 ) denotes the vector cross product and

       
0 l1 c1 l1 c1 + l2 c12 0
o0 = 0 , o1 = l1 s1  , o2 = l1 c1 + l2 s12  , z0 = z1 = 0 .
       

o o o 1
 46

Then, (3.9) can be rewritten as

"
#
Jv
J(θ) =
Jw
 
−l1 s1 − l2 s12 −l2 s12
 
 l1 c1 + l2 c12 l2 c 12 
 
 0 0 
= , (3.10)
 

 0 0  

 0 0  
1 1
   
−l1 s1 − l2 s12 −l2 s12 0 0
where Jv =  l1 c1 + l2 c12 l2 c12  and Jw = 0 0 . For the two-link planar leg,
   

0 0 1 1
the Jacobian matrix in the frame o0 x0 y0 can be represented as
" #
−l 1 s 1 − l2 s 12 −l 2 s 12
J◦ = .
l1 c1 + l2 c12 l2 c12

Then, we start to calculate kinematic singularities of J ◦ . The set of singular config-

urations of the leg model satisfies det(J ◦ ) = 0. Moreover, the set of singular configu-
rations is computed as θ2 = ±π. The results mean that the end-effector do not have
the velocity along the x1 -axis in singular configurations. Substituting θ2 = ±π into
(3.10) yields the singularities of Jv◦2
 
−l1 s1 + l2 s1 +l2 s1
Jv◦2 =  l1 c1 − l2 c1 −l2 c1 
 

0 0

and  
−l1 s1 − l2 s1 −l2 s1
Jv◦2 =  l1 c1 + l2 c1 +l2 c1  .
 

0 0
Hence, the foot-end can be controlled to move in an arbitrary direction generally.
However, when θ2 = ±π, the foot-end can only move in two specified directions
which are vertical to the limb link.
 47

3.1.4 The Euler-Lagrange Equations

By using the Euler-Lagrangian equations in the MATLAB/SimMechancis quadruped
model, the actual angular positions can be computed based on the actuating torques
of the hip and knee joints. In the leg model, vo1 and vo2 denote the linear velocities
for the origin of o1 x1 y1 z1 frame and for the origin of o2 x2 y2 z2 frame, respectively. The
kinetic energy K can be given as

1 1
K = ml1 voT1 vo1 + ml2 voT2 vo2 ,
6 6

where ml1 and ml2 denote the mass of the thigh and limb links, respectively. Besides,
the potential energy P due to gravity can be represented as

P = ml1 gh1 + ml2 gh2 ,

where h1 = c12 l2 + c1 l21 and h2 = c12 l22 . h1 and h2 denote the heights of the center
of gravity for thigh and limb links, respectively. The center of mass is located in
the geometric center of each link, which means the distance from the origins in the
thigh and knee joints to the center of masses of links are l21 and l22 , respectively. The
Lagrangian of the system, L is given by

L =K − P
1 1
= ml1 voT1 vo1 + ml2 voT2 vo2 − (ml2 gh2 + ml1 gh1 ). (3.11)
6 6

The orientation of the frame o1 x1 y1 z1 with respect to the frame o0 x0 y0 z0 is

 
l1 c1
o0 o1 = l1 s1  ,
 

where o0 and o1 are the origins of the base and tool frames in the thigh and knee
joints, respectively. Then, we take the derivation for (3.11), the velocity vo1 at the
 48

origin o1 relative to the base frame o0 x0 y0 z0 is given as

−l1 s1 θ˙1
 

vo1 =  l1 c1 θ˙1  ,
 

We also get
2
voT1 vo1 = l12 θ˙1 . (3.12)

The orientation of the frame o2 x2 y2 z2 with respect to the frame o0 x0 y0 z0 is

 
l1 c1 + l2 c12
o0 o2 = l1 s1 + l2 s12  .
 

Similarly, the velocity vo2 at the origin o2 relative to the base frame o0 x0 y0 z0 is
depicted as
−l1 s1 θ˙1 − l2 s12 (θ˙1 + θ˙2 )
 

vo2 =  l1 c1 θ˙1 + l2 c12 (θ˙1 + θ˙2 )  .

 

0
We also get

2 2
voT2 vo2 = (l12 + 2l1 l2 c2 + l22 )θ˙1 + (2l1 l2 c2 + 2l22 )θ˙1 θ˙2 + l22 θ˙2 . (3.13)

Then, we substitute (3.12) and (3.13) into (3.11) to obtain

1 1 2 1
L(θ, θ̇) = m1 l12 θ12 + m2 (l12 + 2l1 l2 c2 + l22 )θ˙1 + m2 (l1 l2 c2 + l22 )θ1 θ2
6 6 3 (3.14)
1 2 l1 l2
+ m2 l22 θ˙2 − (m1 g c1 + (m1 gl2 + m2 g )c12 ).
6 2 2

The Euler-Lagrange equation is

d ∂L ∂L
( )− = τ,
dt ∂ θ̇ ∂θ
" # " #
τ1 θ1
where τ = and θ = . τ1 and τ2 denote the torques of the hip and knee
τ2 θ2
 49

joints, respectively. It can also be described as

" # " # " #
d ∂L ∂L
( )
dt ∂ θ˙1 ∂θ1
τ1
d ∂L
− ∂L
= , (3.15)
( )
dt ∂ θ˙2 ∂θ2
τ2
" # " # " #
∂L d ∂L ∂L
∂L ∂ θ˙1 d ∂L
( )
dt ∂ θ˙1 ∂L ∂θ1
where ∂ θ̇
= ∂L
, ( )
dt ∂ θ̇
= d ∂L
and ∂θ
= ∂L
. The elements in (3.15) are
∂ θ˙2
( )
dt ∂ θ˙2 ∂θ2
calculated and presented as

∂L 1 1 1
= m1 l12 θ˙1 + m2 (l12 + 2l1 l2 c2 + l22 )θ˙1 + m2 (l1 l2 c2 + l22 )θ˙2 ,
∂ θ˙1 3 3 3
∂L 1 1
= m2 (l1 l2 c2 + l22 )θ˙1 + m2 l22 θ˙2 ,
∂ θ˙2 3 3
∂L l1 l2
=m1 g s1 + (m1 l2 + m2 )gs12 ,
∂θ1 2 2
∂L l2 1 2 1
=(m1 l2 + m2 )gs12 − m2 l1 l2 θ˙1 s1 − m2 l1 l2 θ˙1 θ˙2 s2 ,
∂θ2 2 3 3
d ∂L 1 1 1
( ) =( m1 l12 + m2 (l12 + 2l1 l2 c2 + l22 ))θ¨1 + m2 (l1 l2 c2 + l22 )θ¨2
dt ∂ θ˙1 3 3 3
2 1 2
− m2 l1 l2 s2 θ˙1 θ˙2 − m2 l2 s2 θ˙2 ,
3 3
d ∂L 1 1 1
( ) = m2 (l1 l2 c2 + l22 ))θ¨1 + m2 l22 θ¨2 − m2 l1 l2 s2 θ˙1 θ˙2 .
dt ∂ θ˙2 3 3 3

Substituting above equations into (3.15) yields

" #" #
1
m l2
3 1 1
+ 13 m2 (l12 + 2l1 l2 c2 + l22 ) 1
m (l l c + l22 )
3 2 1 2 2
θ̈1
1 1
m (l l c + l22 )
3 2 1 2 2
m l2
3 2 2
θ̈2
" #" # " # " #
− 2 m2 l1 l2 s2 θ˙2 − 31 m2 l1 l2 s2 θ˙2 θ˙1 −m1 g l21 s1 − (m1 gl2 + m2 g l22 )s12 τ1
+ 13 + = .
m l l s θ˙
3 2 1 2 1 1
0 θ˙2 l
−(m1 gl2 + m2 g 22 )s12 . τ2
(3.16)

Moreover, (3.16) can be represented as

D(θ)θ̈ + C(θ, θ̇)θ̇ + G(θ) = τ, (3.17)

 50

where
" #
1
+ 31 m2 (l12 + 2l1 l2 c2 + l22 ) 31 m2 (l1 l2 c2 + l22 )
m l2
3 1 1
D(θ) = 1 1
,
m (l l c + l22 )
3 2 1 2 2
m l2
3 2 2
" #
− 23 m2 l1 l2 s2 θ˙2 − 13 m2 l1 l2 s2 θ˙2
C(θ, θ̇) = 1 ,
m l l s θ˙
3 2 1 2 1 1
0
" #
−m1 g l21 s1 − (m1 gl2 + m2 g l22 )s12
G(θ) = ,
−(m1 gl2 + m2 g l22 )s12
" # " # " #
τ1 θ¨1 θ˙1
and τ = , θ̈ = , θ̇ = . D(θ) is the mass-inertia matrix, C(θ, θ̇) is the
τ2 θ¨2 θ˙2
matrix of centripetal and Coriolis terms, and G(θ) is the gravity vector. By utilizing
(3.17), the actual angular positions can be calculated in the MATLAB/SimMechanics
quadruped model based on the joint torques.

3.2 Animated Simulation

An animated simulation is developed to analyze the tracking performance in the
quadrupedal locomotion. First, the movement of four legs are assigned in a sequence
based on a proper gait. Then, PD control is implemented to manipulate the leg
posture to execute the locomotion, according to the designed trajectories. Based on
a simplified CAD model, a SimMechanics quadruped model is established to emulate
the motion of the robot. Relative simulation results are also analyzed.

3.2.1 Gait Selection

An appropriate gait is selected to emulate the quadrupedal locomotion by comparing
different gaits. In [14], quadruped gaits are classified into the two-beat gait and the
four-beat gait, as shown in Figure 3.2. The two-beat gait can be subdivided into the
pacing (see Figure 3.2a) and trotting gaits (see Figure 3.2b). In the two-beat gait,
four feet are assigned into two groups. Two feet in either group are designed to touch
the ground synchronously. When the feet strike the ground, identical forces will act
on hip joints. Theoretically, since no extra torque is transmitted to the other side of
the body part, no vertical vibration is caused in the body.
For each step, the balance footholds are defined as the location on the terrain where
 51

Figure 3.2: Classification of periodic gaits [14].

the feet should be placed to continue balanced steady-state locomotion. Compared

to the two-beat gait, the four-beat gait provides more choices of footholds in rough
terrain. Hence, with respect to the available footholds on the ground, the placement
of the feet can interact with the stability and general behavior of the quadruped
robot. For the single foot lateral gait (see Figure 3.2c) and the single foot diagonal
gait (Figure 3.2d), the foot-ends strike the ground in a symmetric footfall sequence.
In both of these four-beat gaits, triple feet touch the ground alternately at the same
time. To maintain stability in the locomotion, the projected center of gravity lies
within the support polygon, which is defined by the footholds [1].

Figure 3.3: Internal contact directions of the creeping gait.

Several typical kinds of gaits are described in the following. The pacing, trotting
and bounding gaits are two-beat gaits. The creeping and tölting gaits are four-beat
gaits.
 52

• In the pacing gait, lateral pairs of legs move synchronously. The pacing gait is
appropriate for the quadruped robot to execute a high-speed locomotion on flat
ground.

• In the trotting gait, diagonal pairs of legs move at the same time. Its energy
consumption is relatively low at a high-speed motion.

• In the bounding gait, the front pair of legs and hind pair of legs move in se-
quence.

• In the creeping gait, as shown in Figure 3.4, the quadruped robot continuously
executes the footfall pattern LH, LF, RH, and RF in a lateral footfall sequence
[52]. The abbreviations stand for left hind (LH), left fore (LF), right hind
(RH) and right fore (RF). In this static gait, triple feet alternately provide the
support. Since its energy cost is relatively low at a low-speed motion, this gait
is appropriate for traversing rocky and sloped terrain.

• The tölting gait is extracted from the locomotion performance of icelandic horses
in [53]. The tölting gait has the same footfall sequence as the creeping gait.
But the touch-down phase of the tölting gait occupies less percentage in one
step cycle than the creeping gait. Consequently, a single foot and double feet
alternately provide the support.

Main body
Travel

RH LH

RF
LF

Figure 3.4: The schematic model of a quadruped system.

 53

Generally, various kinds of gaits have their own advantages in different environ-
ments in [54]. The creeping gait is a suitable choice for the quadruped robot in a low
speed motion. Since triple feet alternately provide the support in the locomotion, this
four-beat gait presents high stability. In addition, the energy cost of the creeping gait
is lower than other gaits at a low-speed motion. Hence, the creeping gait is chosen to
be applied in the designed robot. Figure 3.3 shows internal contact directions in the
creeping gait.
As one type of four-beat gaits, the creeping gait is selected to perform the low-
speed locomotion due to its low energy consumption and substantial available footholds.
In the creeping gait, the predefined footfall pattern is (LH→ LF→ RH → RF →
LH→. . . ), and this sequence can be readily generated in MATLAB. A step cycle of
the quadrupedal locomotion can be divided into the swing and stance phases accord-
ing to [52]. Figure 3.5 shows that the sequence of swing and stance phases for a
one-legged robot is separated by a short moment called toe-off in [15].

Figure 3.5: Sequence of the locomotion cycle simulation in [15].

3.2.2 PD Control
To manipulate the leg posture to execute the locomotion, each joint is operated by
the PD controller. Figure 3.6 shows the block diagram of a quadruped robot motion
control system. The desired trajectories of foot-ends are formed based on the footfall
sequence of the creeping gait, and the desired angular positions for the hip and knee
joints can be calculated by using inverse kinematics equations. For each joint, the
control signal is calculated by a PD controller based on the difference between the
desired and actual angular positions. By importing the control signal to the ideal
motor block, the joint torque is generated to manipulate the leg posture, thereby
executing the quadrupedal locomotion. In the MATLAB/SimMechanics model, the
 54

actual angular positions of hip and knee joints are computed based on (3.17). Many
physical effects are ignored in this simplified model.
For each joint, a PD controller is applied to make the response to meet the design
specifications, as shown in the area of the orange dotted lines in Figure 3.4. Based
on the error value of angular positions, the proportional and derivative gains are
implemented to generate the actuating torques. To set up the parameters of the PD
controller, the proportional gain Kp is adjusted firstly. Kp is proportional to the
current value of the error ∆θ = θ − θd , where θ and θd denote the actual joint angle
and the desired joint angle, respectively. Then, to reduce the effect of the error, the
derivative gain Kd is used to exert the control influence, which is generated by the
rate of error change. After several tests, the appropriate Kp and Kd for the hip joint
are chosen as 20 and 2, respectively. Kp and Kd for the knee joint are selected as
10 and 1, respectively. Similarly, by applying the PD controller, the control signal
for the motor in the knee joint is generated to actuate the limb link. Meanwhile, the
PD controller in the hip joint provides the control signal to the ideal motor block to
actuate the whole leg, which is twice the mass of the limb. Hence, it is reasonable
that PD parameters for the hip joint are twice as parameters for the knee joint. Since
the residual error in the system is relatively small and no need to be eliminated, the
integral term is not added in the controller. The details of the trajectory generator
and SimMechanics quadruped model in Figure 3.6 are illustrated in Section 3.2.3 and
3.2.4.

3.2.3 Generating Smooth Trajectories

Based on the above investigation of gait selection, the creeping gait is implemented
for the quadrupedal locomotion. In this gait, triple and quadruple legs alternatively
provide the support to the robot. The quadrupedal locomotion can be divided into
the swing and stance phases. In the swing phase, four legs swing forward in the
predefined sequence while other three legs go through a rest period to provide the
support. In the stance phase, the foot-ends are supposed to connect the ground with
a complete stop, and the legs move synchronously to shift the main body forward.
The foot-ends are modeled as points without geometrical extension. In the stance
phase, the foot-ends are modeled as simple rotational joints, which connect the legs
as fixed points on the flat terrain. The legs move freely around the hip and knee
joints freely, thereby providing a total of three degrees of freedom.
 55

PD control

Trajectory 𝜽𝒅 𝚫𝜽 𝝉 SimMechanics 𝜽
generator
𝑲𝒑 quadruped model

𝑲𝒅 𝑺

Figure 3.6: PD control schematics.

Hip joint
𝒙

Knee joint

Highest

AEP Step Height PEP

Step Length

Figure 3.7: Nominal trajectory of the leg.

The desired trajectory of the foot-end is generated based on the movement of

the foot-end in the predefined pattern. Also, the desired trajectory is limited by the
mechanism of each robotic leg, and the dimensions of one leg are provided in Table
3.2. To move the leg freely without any restrictions, the set of singular configuration
θ2 = ±π is not included in the desired trajectory. As illustrated in Figure 3.7, three
 56

significant points are selected to form the desired trajectory of the foot end in the
swing phase. First of all, the start point of one step cycle is the posterior extreme
position (PEP) where the foot-end lifts off the ground. As shown in Figure 3.8, the
origin of the base frame is set at the hip joint of the leg, and the PEP is at [-0.051 m,
0.254 m, 0 m]. Then, the leg swings to the highest position. This half-way point of
the swing phase is at [0 m, 0.203 m, 0 m]. Finally, the foot-end strikes the ground in
the touchdown phase at the anterior extreme position (AEP), which is at [0.051 m,
0.254 m, 0 m]. In addition, two relative points are also chosen to smooth the path
around the highest position. These relative points are at [0.036 m, 0.218 m, 0 m]
and [0.036 m, 0.218 m, 0 m], respectively. Based on three significant points and two
relative points, a fifth order polynomial can be formulated to form each individual
spline segment of the desired trajectory.

Leg base 𝒛𝟎 𝒐𝟎
at [0 m, 0 m, 0 m] 𝒚𝟎 𝒙𝟎

𝒍𝟏

𝒍𝟐

Extreme anterior set point

at [-0.051 m, 0.254 m, 0 m]

Figure 3.8: A posterior extreme set point for one step shown with Cartesian coordi-
nates.

3.2.4 MATLAB/SimMechanics Model

The designed CAD model in Chapter 2 is too complicated for MATLAB to com-
pute. Hence, a simplified CAD model is developed for an animated simulation, whose
main characteristics of the designed robot are maintained, such as the mass and
 57

Table 3.2: Some parameters of the designed quadruped robot.

Link name Length (m) Width (m) Thickness (m)

Thigh 0.152 0.089 0.089
Limb 0.152 0.076 0.076

link length. The LH, LF, RH and RF legs are painted blue, red, green and yel-
low, as illustrated in Figure 3.9. The simplified CAD model is converted into the
MATLAB/SimMechanics model with the help of MATLAB/SimMechanics packages.
Related MATLAB/Simulink blocks are also set up to model the dynamics of the
quadruped robot. The desired trajectories are generated in MATLAB based on a
fifth polynomial. Environment blocks should be carefully set to provide the global
information about the workspace, such as the gravitational acceleration and global
coordinate system. The gravitational acceleration is set to be 9.81 m/s2 in the simu-
lation.

Figure 3.9: The simulated animation for the simplified model.

To set up the MATLAB/SimMechanics quadruped model, the thigh link connects

the shoulder and limb links through two revolute joint blocks, which represent the
hip and knee joints, respectively. Four shoulder links are rigidly attached to the main
body. Other related information is determined by the simplified CAD model, such as
the link connection, the link mass and the center of mass for each part.
For the leg model, the origin of the base frame is set at the connection of the
shoulder and thigh links. Both the hip and knee joints are designed to rotate about
the z-axis in their corresponding base frames. At the beginning of the simulation,
 58

Figure 3.10: The MATLAB/SimMechanics quadruped model.

relative coordinate systems should be initialized in the hip and knee joints and foot-
ends, as illustrated in Figure 3.1. In other words, as discussed in Section 3.1, we
need to define the transformation matrix from the hip joint to the knee joint, and
furthermore, define the transformation matrix from the knee joint to the foot-end. By
applying transformation matrices, the directions of axes are re-oriented in the base
frames. For example, the knee joint angular position θ2 of the LF should be initialized
as −θ2 to present the rotation angle. The details of the MATLAB/SimMechanics
model are illustrated in Figure 3.10.

3.2.5 Simulation Results

Simulation results are analyzed to demonstrate the effectiveness of the quadruped
robot motion control system. By connecting the three critical points and two relative
points with a fixed step size 0.001s, the desired trajectories of the foot-ends are
generated based on a fifth order polynomial with a shape similar to a sine wave. Then,
the desired joint angular positions can be computed by applying inverse kinematics
 59

(a) (b)
Angle of the Right-hind Hip Joint Angle of the Right-hind Knee Joint
1.4 1.8
Desired Desired
1.3 Actual 1.6 Actual

1.2 1.4

1.1 1.2
Angle(rad)

Angle(rad)
1 1

0.9 0.8

0.8 0.6

0.7 0.4

0.6 0.2
0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
Time(sec) Time(sec)

(c) (d)

Figure 3.11: Illustrations of the desired and tracked angle of the hip joint and knee
joint of hind legs: (a) the left-hind hip joint, (b) left-hind knee joint, (c) right-hind
hip joint and (d) right-hind knee joint.

equations. The actuating torques are computed by using PD control based on the
desired angular positions. By implementing Euler-Lagrange equations, the actual
joint positions are generated, as illustrated in Figure 3.6. The actual trajectories of
the foot-ends are further calculated by using forward kinematics equations. In Figure
3.11, Figure 3.12 and Figure 3.13, the blue curves denote the desired trajectories for
the joint angles or foot-ends, and the red curves denote the actual trajectories for the
joint angles or foot-ends in each case.
We take the left-hind (LH) leg as an example to demonstrate the quadrupedal
locomotion in one step cycle as shown in Figure 3.11a and Figure 3.11b. Each step
cycle lasts for 5.25 s, and the swing phase is from 0.25 s to 1.25 s. Then, the LH leg
goes through a rest period while the other legs are in the swing phase in order. The
LH leg is in a stance phase from 4.5 s to 5.5 s. The foot-end lifts off the ground at
 60

Angle of the Left-front Hip Joint Angle of the Left-front Knee Joint
-1.4 -0.6
Desired Desired
Actual Actual
-1.6 -0.8

-1.8 -1
Angle(rad)

Angle(rad)
-2 -1.2

-2.2 -1.4

-2.4 -1.6

-2.6 -1.8
0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
Time(sec) Time(sec)

(a) (b)

Angle of the Right-front Hip Joint Angle of the Right-front KneeJoint

2.6 1.8
Desired Desired
Actual Actual
2.4 1.6

2.2 1.4
Angle(rad)

Angle(rad)

2 1.2

1.8 1

1.6 0.8

0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
Time(sec) Time(sec)

(c) (d)

Figure 3.12: Illustrations of the desired and tracked angle of the hip joint and knee
joint of front legs: (a) the left-front hip joint, (b) left-front knee joint, (c) right-front
hip joint and (d) right-front knee joint.

the PEP of the step cycle to the half-way point, and then touches down at the AEP.
Figure 3.11 and Figure 3.12 demonstrate the tracking performance of the designed
motion controller for the hind legs and front legs in two step cycles. The desired and
tracked angle trajectories of each joint are illustrated in the subfigures of Figure 3.11
and Figure 3.12. From the simulation results, it can be observed that, at the beginning
of the locomotion, it takes around 0.25 s for each joint to track the desired angular
position. Then, each joint angle is capable of tracking the desired angular position
with an approximately 0.001 rad. The tracking performances around the red circles
are illustrated in the time series plots.
The simulation results of foot-end positions in a sequence of LH, LF, RH, and
RF are shown in Figure 3.13. At the beginning of the simulation, the foot-end is
 61

controlled to track the initial point of the desired trajectory, so the displacement of
the position is relatively large. After around 0.25 s, the foot-end starts to track the
desired path in a relatively good condition.

Position of the Left-hind Foot End Position of the Left-front Foot End
Desired Desired
Actual Actual

-0.2 -0.2

-0.22
-0.22
Y-direction

Y-direction
-0.24
-0.24
-0.26

-0.26
-0.28

-0.3 -0.28
0.1 0.2
0.05 15 15
0.1
0 10 10
0
-0.05 5 5
X-direction -0.1 0 Time(sec) X-direction -0.1 0 Time(sec)

(a) (b)

Position of the Right-hind Foot End Position of the Right-front Foot End
Desired Desired
Actual Actual

-0.2 -0.2

-0.22
-0.22
Y-direction

Y-direction

-0.24
-0.24
-0.26

-0.26
-0.28

-0.3 -0.28
0.1 0.2
0.05 15 15
0.1
0 10 10
0
-0.05 5 5
X-direction -0.1 0 Time(sec) X-direction -0.1 0 Time(sec)

(c) (d)

Figure 3.13: Illustrations of the desired and tracked angle of four foot-ends: (a) the
left-hind foot-end, (b) left-front foot-end, (c) right-hind foot-end and (d) right-front
foot-end.

3.3 Conclusion
In this chapter, we focus on modeling for a robotic leg and the animated simulation of
the designed quadruped robot. Forward and inverse kinematics equations are utilized
in the leg model, and singularities of the Jacobian matrix is calculated to identify the
singular configurations. Then, Euler-Lagrange equations are applied to calculate the
hip and knee joint torques, respectively. Furthermore, an animated simulation of the
 62

quadrupedal locomotion is emulated with the creeping gait. The desired trajectories
of foot-ends are generated based on the predefined footfall sequence. At last, the
MATLAB/SimMechanics model is established in MATLAB. By using PD control to
manipulate the joints, the foot-ends track the desired paths well.
 63

Chapter 4

Conclusions and Future Work

4.1 Conclusions
This thesis mainly introduces the mechanical design and simulation studies of a
quadruped robot motion control system. An animated simulation of the quadrupedal
locomotion has been emulated by combining MATLAB and SolidWorks.
In Chapter 2, the mechanical structure of the quadruped robot is presented. Par-
ticularly, section 2.1 introduces some key components, such as DC motors and poten-
tiometers. Additionally, the leg mechanism is optimized to improve the characteris-
tics of the compliant leg, such as the speed, endurance, agility and strength. Several
methods are applied to design the robot from the aspects of the leg length, mass
distribution, leg kinematics, elastic energy storage and muscle power. To absorb the
impact force at collision, linear telescoping joints are incorporated in the limb links
to realize the internal energetic conversion. In Section 2.4, the structural compo-
nents are optimized to enhance the fragile position by using FEA. Besides, structural
components in the hip joint are demonstrated as examples in the design process.
In Chapter 3, the modeling for the robotic leg and animated simulation are illus-
trated. Kinematic and dynamic parameters of the leg model are specified based on
a CAD model, which is drawn by SolidWorks. Based on the deviation between the
desired and tracked joint angles, the PD control is implemented to generate a PWM
signal to actuate the DC motor. Actuated by the joint torques, the hip and knee
joints are manipulated to adjust the leg postures. By choosing appropriate footholds,
the quadrupedal locomotion is executed in a creeping gait. One step cycle of the
quadruped movement is divided into the stance and swing phases. In the stance
 64

phase, the quadruped robot maintains the foothold position while shifting the body
part forward. In the swing phase, four legs are controlled to execute the quadrupedal
locomotion in the predefined footfall pattern.

4.2 Future Work

In the future, we plan to improve the designed quadruped robot from the perspectives
of mechanical design and control strategy. From the aspect of the mechanical design,
a higher gear ratio can be applied to improve the leg stiffness in the hip and knee
joints. As mentioned in Section 1.2, the robotic leg can be composed of compliant
mechanisms, such as the leg of “RHex” [38]. In the limb part, the telescoping joint
can be upgraded to PAM with solenoid valves and check valves, such as the 3D one-
leg hopping machine [22]. Steel coil springs can also be used to avoid assembling the
motor in the thigh part. This mechanism not only reduces the leg weight, but also
improves the agility of the leg.
New type actuators are also expected to be assembled in the robot. For example,
Maxon EC-4pole brushless motors, which is implemented in “StarlETH” [1], can
provide maximal velocity and torque for short-term operation. The ANYdrive in
“ANYmal” can be used to actuate the joint directly [55]. Timing belts and gears are
no longer to be assembled to redirect the actuating torque. Hence, by applying this
type of actuators, the mechanical structure is significantly simplified.
To design an autonomous legged robot, we also plan to work on the control strat-
egy. The overview of the control architecture for the quadrupedal locomotion is illus-
trated in Figure 4.1. The hierarchical control structure has been divided into the high
and low level controls, which are processed by the FPGA and ARM architectures,
respectively.
The FPGA is used to conduct the high level control, such as visual image process-
ing and proper gaits selection [1]. The ARM architecture is applied to control the hip
and knee joints, thereby manipulating the posture of each leg. Force sensors can be
assembled at the foot-ends to measure the slope angle of the connect surface for the
quadruped posture adjustment. Cameras can collect the surrounding information,
thereby helping the robot to choose the appropriate footholds.
For the high level control, the FPGA is applied with cameras for image processing.
Based on the collected data, an updated map can be processed for path planning and
obstacle avoidance. The FPGA is used to set up a vision system for inline processing.
 65

High level control Perception

Behavior generation Terrain estimation

Planning Localization & mapping

Behavior control Robot state estimation

Inverse kinematics &dynamics Sensor fusion

Low level control

Actuator control
Position & force control

Figure 4.1: An overview of the control architecture for the quadrupedal locomotion.

High-speed control operations can be implemented in the FPGA with short processing
time. Proper gaits can also be applied in this architecture according the environmental
conditions. For instance, when the terrain is relatively smooth, dynamic gaits can be
applied for a high speed locomotion. Otherwise, static gaits is the optimal option,
when the robot moves in rough terrain.
For the low level control, the ARM architecture is used to generate the leg tra-
jectory by controlling the joint angles. The stride of one step cycle is adjusted in
this level for the dynamically stable legged robot. By controlling the joint angles,
the act of positioning the foot-ends interacts with the stability of the quadrupedal
locomotion. To balance steady-state mobility, the ARM architecture is applied for
positioning the foot-ends in the appropriate locations. The ARM architecture not
only can operate the legs touching the ground in the desired footholds, but also can
adjust the posture of the robot.
Overall, the hierarchical structure of the path planner is expected to be developed
for the designed quadruped robot in the future.
 66

Bibliography

[1] M. Hutter, C. Gehring, M. A. Höpflinger, M. Blösch, and R. Siegwart, “To-

ward combining speed, efficiency, versatility, and robustness in an autonomous
quadruped,” IEEE Trans. Robot., vol. 30, no. 6, pp. 1427–1440, 2014.

[2] Boston Dynamics Inc. (2017). [Online]. Available:

http://www.bostondynamics.com

[3] S. Seok, A. Wang, M. Y. Chuah, D. Otten, J. Lang, and S. Kim, “Design prin-
ciples for highly efficient quadrupeds and implementation on the MIT Cheetah
robot,” in Proc. IEEE Int. Conf. Robot. Autom., Karlsruhe, Germany, May. 2013,
pp. 3307–3312.

[4] S. Hirose. (2017). [Online]. Available:

http://robotlocomotion.mech.kit.ac.jp/research/Quadruped/photo-movie-
pat2-e.html

[5] ——. (2017). [Online]. Available:

http://robotlocomotion.mech.kit.ac.jp/research/Quadruped/photo-movie-
tekken2-photo.html

[6] M. H. Raibert, H. B. Brown Jr, and M. Chepponis, “Experiments in balance

with a 3D one-legged hopping machine,” Int. J. Robot. Res., vol. 3, no. 2, pp.
75–92, 1984.

[7] B. Vanderborght, B. Verrelst, R. Van Ham, M. Van Damme, P. Beyl, and

D. Lefeber, “Development of a compliance controller to reduce energy consump-
tion for bipedal robots,” Auton. Robots, vol. 24, no. 4, pp. 419–434, 2008.

[8] R. Niiyama, A. Nagakubo, and Y. Kuniyoshi, “Mowgli: A bipedal jumping and

landing robot with an artificial musculoskeletal system,” in Proc. IEEE Int. Conf.
Robot. Autom., Rome, Italy, Apr. 2007, pp. 2546–2551.
 67

[9] M. Ahmadi and M. Buehler, “Preliminary experiments with an actively tuned

passive dynamic running robot,” in Proc. Int. Symp. Exp., Barcelona, Spain,
Jun. 1998, pp. 312–324.

[10] J. Estremera and K. J. Waldron, “Thrust control, stabilization and energetics of

a quadruped running robot,” Int. J. Robot. Res., vol. 27, no. 10, pp. 1135–1151,
2008.

[11] K. S. Yamazaki, “The design and control of Scout I, a simple quadruped robot,”
Ph.D. dissertation, McGill Univ., Montreal, QC, Canada, 1999.

[12] I. Poulakakis, J. A. Smith, and M. Buehler, “Modeling and experiments of un-

tethered quadrupedal running with a bounding gait: The Scout II robot,” Int.
J. Robot. Res., vol. 24, no. 4, pp. 239–256, 2005.

[13] H. Kimura, Y. Fukuoka, and A. H. Cohen, “Adaptive dynamic walking of a

quadruped robot on natural ground based on biological concepts,” Int. J. Robot.
Res., vol. 26, no. 5, pp. 475–490, 2007.

[14] C. D. Remy, K. Buffinton, and R. Siegwart, “Stability analysis of passive dynamic

walking of quadrupeds,” Int. J. Robot. Res., vol. 29, no. 9, pp. 1173–1185, 2010.

[15] E. Garcia, J. C. Arevalo, G. Munoz, and P. Gonzalez-de Santos, “On the

biomimetic design of agile-robot legs,” Sensors, vol. 11, no. 12, pp. 11 305–11 334,
2011.

[16] P. G. de Santos, E. Garcia, J. Cobano, and T. Guardabrazo, “Using walking

robots for humanitarian de-mining tasks,” in Proc. 35th Int. Symp. Robot., Paris,
France, Mar. 2004, CD-ROM.

[17] C. S. Sharp, O. Shakernia, and S. Sastry, “A vision system for landing an un-
manned aerial vehicle,” in Proc. IEEE Int. Conf. Robot. Autom., Seoul, Korea,
May. 2001, pp. 1720–1727.

[18] P. Corke, C. Detweiler, M. Dunbabin, M. Hamilton, D. Rus, and I. Vasilescu,

“Experiments with underwater robot localization and tracking,” in Proc. IEEE
Int. Conf. Robot. Autom., Roma, Italy, Apr. 2007, pp. 4556–4561.
 68

[19] J.-F. Lalonde, N. Vandapel, D. F. Huber, and M. Hebert, “Natural terrain classi-
fication using three-dimensional ladar data for ground robot mobility,” J. Robot.
Syst., vol. 23, no. 10, pp. 839–861, 2006.

[20] C. Ridderström, “Legged locomotion: Balance, control and tools-from equation

to action,” Ph.D. dissertation, Royal Inst. Technol., Stockholm, Sweden, 2003.

[21] K. J. Waldron, V. J. Vohnout, A. Pery, and R. B. McGhee, “Configuration design

of the adaptive suspension vehicle,” Int. J. Robot. Res., vol. 3, no. 2, pp. 37–48,
1984.

[22] M. H. Raibert, Legged Robots that Balance. Cambridge, MA, USA: MIT press,
1986.

[23] M. G. Bekker, Introduction to Terrain-vehicle Systems. Ann Arbor, MI, USA:

University of Michigan Press, 1969.

[24] P. Gregorio, M. Ahmadi, and M. Buehler, “Design, control, and energetics of

an electrically actuated legged robot,” IEEE Trans. Syst., Man Cybern., vol. 27,
no. 4, pp. 626–634, 1997.

[25] Y. Li, B. Li, J. Ruan, and X. Rong, “Research of mammal bionic quadruped
robots: A review,” in Proc. Int. Conf. Robot., Autom. Mechatronics, Qingdao,
China, Sep. 2011, pp. 166–171.

[26] D. J. Todd, Walking Machines: An Introduction to Legged Robots. London,

UK: Kogan Page, 2013.

[27] S. Peng, C. P. Lam, and G. R. Cole, “A biologically inspired four legged walking
robot,” in Proc. IEEE Int. Conf. Robot. Autom., Taipei, Taiwan, Sep. 2003, pp.
2024–2030.

[28] Y. Sakagami, R. Watanabe, C. Aoyama, S. Matsunaga, N. Higaki, and K. Fu-

jimura, “The intelligent ASIMO: System overview and integration,” in Proc.
IEEE/RSJ Int. Conf. Intell. Robots Syst., Lausanne, Switzerland, Oct. 2002,
pp. 2478–2483.

[29] W. Ilg, K. Berns, H. Jedele, J. Albiez, R. Dillmann, M. Fischer, H. Witte,

J. Biltzinger, R. Lehmann, and N. Schilling, “BISAM: From small mammals to
 69

a four legged walking machine,” in Proc. From Animals to Animates: 5th Int.
Conf. Society for Adaptive Behaviour, Zrich, Switzerland, 1998, pp. 400–407.

[30] A. Schneider, J. Paskarbeit, M. Schäffersmann, and J. Schmitz, “Biomechatron-

ics for embodied intelligence of an insectoid robot,” in Proc. 4th Int. Conf. Intell.
Robot. Appl., Aachen, Germany, Dec. 2011, pp. 1–11.

[31] J. Buchli, J. Pratt, N. Roy, M. P. Murphy, A. Saunders, C. Moreira, A. A. Rizzi,

and M. Raibert, “The LittleDog robot,” Int. J. Robot. Res., vol. 30, no. 2, pp.
145–149, 2011.

[32] M. Bajracharya, J. Ma, M. Malchano, A. Perkins, A. A. Rizzi, and L. Matthies,

“High fidelity day/night stereo mapping with vegetation and negative obstacle
detection for vision-in-the-loop walking,” in Proc. IEEE/RSJ Int. Conf. Intell.
Robots Syst., Tokyo, Japan, Nov. 2013, pp. 3663–3670.

[33] H. Kimura, Y. Fukuoka, and A. H. Cohen, “Biologically inspired adaptive walk-

ing of a quadruped robot,” Phil. Trans. R. Soc. A, vol. 365, no. 1850, pp. 153–170,
2007.

[34] M. Raibert, K. Blankespoor, G. Nelson, R. Playter, and T. B. Team, “Bigdog,

the rough-terrain quadruped robot,” in Proc. 17th World Congr., Seoul, Korea,
Jul. 2008, pp. 10 822–10 825.

[35] X. Zhou and S. Bi, “A survey of bio-inspired compliant legged robot designs,”
Bioinspir. Biomim., vol. 7, no. 4, pp. 1–20, 2012.

[36] S. P. Singh and K. J. Waldron, “Attitude estimation for dynamic legged locomo-
tion using range and inertial sensors,” in Proc. IEEE Int. Conf. Robot. Autom.,
Barcelona, Spain, Apr. 2005, pp. 1663–1668.

[37] B. Brown and G. Zeglin, “The bow leg hopping robot,” in Proc. IEEE Int. Conf.
Robot. Autom., Leuven, Belgium, May. 1998, pp. 781–786.

[38] E. Moore, D. Campbell, F. Grimminger, and M. Buehler, “Reliable stair climb-

ing in the simple hexapod ‘RHex’,” in Proc. IEEE Int. Conf. Robot. Autom.,
Washington, USA, May. 2002, pp. 2222–2227.
 70

[39] H.-W. Park, K. Sreenath, A. Ramezani, and J. W. Grizzle, “Switching control

design for accommodating large step-down disturbances in bipedal robot walk-
ing,” in Proc. IEEE Int. Conf. Robot. Autom., Minneapolis, USA, May. 2012,
pp. 45–50.

[40] P. Birkmeyer, K. Peterson, and R. S. Fearing, “Dash: A dynamic 16 g hexapedal

robot,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., St. Louis, USA, Oct.
2009, pp. 2683–2689.

[41] A. T. Baisch, C. Heimlich, M. Karpelson, and R. J. Wood, “Hamr3: An au-

tonomous 1.7 g ambulatory robot,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots
Syst., San Francisco, USA, Sep. 2011, pp. 5073–5079.

[42] A. M. Dollar, C. R. Wagner, and R. D. Howe, “Embedded sensors for biomimetic

robotics via shape deposition manufacturing,” in Proc. 1st IEEE/RAS-EMBS
Int. Conf. Biomed. Robot. Biomechatron., Pisa, Italy, Feb. 2006, pp. 763–768.

[43] M. Buehler, R. Battaglia, A. Cocosco, G. Hawker, J. Sarkis, and K. Yamazaki,

“Scout: A simple quadruped that walks, climbs, and runs,” in Proc. IEEE Int.
Conf. Robot. Autom., Leuven, Belgium, May. 1998, pp. 1707–1712.

[44] S. Hirose, “Some considerations on a feasible walking mechanism as a terrain

vehicle,” in Proc. 3rd RoManSy Symp., Udine, Italy, Sep. 1978, pp. 357–375.

[45] S. Hirose and Y. Umetani, “The basic motion regulation system for a quadruped
walking vehicle,” in Design Engineering Tech. Conf. ASME, Los Angeles, USA,
Sep. 1980, paper 80-DET-34, pp. 1–6.

[46] K. Arikawa and S. Hirose, “Development of quadruped walking robot TITAN-

VIII,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Osaka, Japan, Nov.
1996, pp. 208–214.

[47] S. Kagami, K. Okada, M. Kabasawa, Y. Matsumoto, A. Konno, M. Inaba,

and H. Inoue, “A vision-based legged robot as a research platform,” in Proc.
IEEE/RSJ Int. Conf. Intell. Robots Syst., Victoria, Canada, Oct. 1998, pp. 235–
240.

[48] H.-W. Park, M. Y. Chuah, and S. Kim, “Quadruped bounding control with
variable duty cycle via vertical impulse scaling,” in Proc. IEEE/RSJ Int. Conf.
Intell. Robots Syst., Chicago, USA, Sep. 2014, pp. 3245–3252.
 71

[49] N. Kohl and P. Stone, “Policy gradient reinforcement learning for fast
quadrupedal locomotion,” in Proc. IEEE Int. Conf. Robot. Autom., New Or-
leans, USA, Apr. 2004, pp. 2619–2624.

[50] S. Seok, A. Wang, M. Y. M. Chuah, D. J. Hyun, J. Lee, D. M. Otten, J. H.

Lang, and S. Kim, “Design principles for energy-efficient legged locomotion and
implementation on the MIT Cheetah robot,” IEEE/ASME Trans. Mechatronics,
vol. 20, no. 3, pp. 1117–1129, 2015.

[51] M. W. Spong, S. Hutchinson, and M. Vidyasagar, Robot Modeling and Control.

New York, NY, USA: Wiley, 2006.

[52] R. B. McGhee and A. A. Frank, “On the stability properties of quadruped creep-
ing gaits,” Mathematical Biosciences, vol. 3, no. 3, pp. 331–351, 1968.

[53] M. Hildebrand, “Symmetrical gaits of horses,” Science, vol. 150, no. 3697, pp.
701–708, 1965.

[54] W. Xi, Y. Yesilevskiy, and C. D. Remy, “Selecting gaits for economical locomo-
tion of legged robots,” Int. J. Robot. Res., vol. 35, no. 9, pp. 1140–1154, 2016.

[55] M. Hutter, C. Gehring, D. Jud, A. Lauber, and C. D. Bellicoso, “Anymal–a

highly mobile and dynamic quadrupedal robot,” in Proc. IEEE/RSJ Int. Conf.
Intell. Robots Syst., Daejeon, Korea, Oct. 2016, pp. 38–44.