Hierarchical controller for highly dynamic

locomotion utilizing pattern modulation and

impedance control: implementation on the MIT
ASACTHM~?fi
Cheetah robot
OF TECHNOLOY
by

Jongwoo Lee NOV 12 2013

B.S., Seoul National University (2011) L BRAR IES
Submitted to the Department of Mechanical Engineering
in partial fulfillment of the requirements for the degree of

Master of Science in Mechanical Engineering

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

September 2013

@ Massachusetts Institute of Technology 2013. All rights reserved.

A u th or ........ ....
....
....
....
..... .. ........ ..... ....
........
...
Dep/men of Mechanical Engineeri6 L
August 23 201-3

C ertified by .......................... .
Sangbfe Kim
Esther & Harold E. Edgerton Assistant Profpssor of Mechmailcal gineering
-Xhesis Supervisor

Accepted by .............. ......

David E. Hardt
Ralph E. & Eloise F. Cross Professor of Mechanical Engineering
Chairman, Committee on Graduate Students
 Hierarchical controller for highly dynamic locomotion
utilizing pattern modulation and impedance control:
implementation on the MIT Cheetah robot

by
Jongwoo Lee

Submitted to the Department of Mechanical Engineering

on August 23, 2013, in partial fulfillment of the
requirements for the degree of
Master of Science in Mechanical Engineering

Abstract
This thesis presents a hierarchical control algorithm for quadrupedal locomotion. We
address three challenges in developing a controller for high-speed running: locomotion
stability, control of ground reaction force, and coordination of four limbs. To tackle
these challenges, the proposed algorithm employs three strategies. Leg impedance
control provides programmable virtual compliance of each leg which achieve self-
stability in locomotion. The four legs exert forces to the ground using equilibrium-
point hypothesis. A gait pattern modulator imposes a desired footfall sequence. The
control algorithm is verified in a dynamic simulator constructed using MATLAB and
then in the subsequent experiments on the MIT Cheetah robot. The experiments on
the MIT Cheetah robot demonstrates high speed trot running reaching up to the speed
of 6 m/s on a treadmill. This speed corresponds to a Froude number (Fr = 7.34),
which is comparatively higher than other existing quadrupedal robots.

Thesis Supervisor: Sangbae Kim

Title: Esther & Harold E. Edgerton Assistant Professor of Mechanical Engineering

3
4
Acknowledgments
I would like to begin by thanking my advisor, Prof. Sangbae Kim for his advice,

patience, and understanding. I thank the members of the MIT Biomimetic Robotics

Laboratory for their help. I am especially pleased to have met Dr.Dong Jin Hyun

in this lab. Without his support and encouragement, I would not have achieved

anything. I also give many thanks to Sangok Seok for his advice and efforts to

operate the MIT Cheetah robot.

It has been a pleasure to meet all of my friends in Boston, especially the members of

MIT KGSA. It was memorable to hang out with Albert Wang and Meng Yee (Michael)

Chuah. Also, I give thanks to Yoobin (Anna) Seo for correcting grammatical errors.
Additionally, I appreciate Samsung Foundation of Culture for their unconditional

financial support for the last two years.

Finally, and most importantly, I would like to give thanks to my family; Kyuheon

Lee, my father, Namhee Jung, my mother, and Jongmin Lee, my Wangjongmin.

Without their love and support, I would not have finished my study.

The project is supported by the Defense Advanced Research Projects Agency

(DARPA) Maximum Mobility and Manipulation (M3) program.

5
6
Contents

1 Introduction 19
1.1 Motivation and objectives . . . . . . . . . . . . . . . . . 19
1.2 Current state of research . . . . . . . . . . . . . . . . . . 20
1.3 Our approaches and accomplishments . . . . . . . . . . . 24
1.4 Structure of this thesis . . . . . . . . . . . . . . . . . . . 27

2 Modeling of the MIT Cheetah robot 29

2.1 Hybrid dynamic systems . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.1 Hybrid dynamic systems in legged systems . . . . . . . . . 31
2.1.2 hybrid dynamic systems realized in MATLAB . . . . . . . 32
2.2 Constrained equations of motion . . . . . . . . . . . . . . . . . . . 34
2.2.1 Multi-link rigid body equations of motion . . . . . . . . . 34
2.2.2 Numerical error inhibition . . . . . . . . . . . . . . . . . . 35
2.3 Friction model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Impact map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Quadrupedal Locomotive Control Framework 45

3.1 Gait pattern modulator with proprioceptive sensory feedback ..... 46
3.1.1 Phase signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.1.2 Synchronization by proprioceptive sensory feedback . . . . . . 48
3.1.3 Imposing gait patterns . . . . . . . . . . . . . . . . . . . . . . 49
3.1.4 Phase signal generation according to TD elapsed time and phase
differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

7
 3.2 Leg trajectory generator . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.2.1 Design of the trajectory in the swing phase . . . . . . . . . . . 53

3.2.2 Design of the trajectory in the stance phase . . . . . . . . . . 56
3.2.3 Scaling and translation of the designed trajectory . . . . . . . 59
3.3 Low level individual leg controller . . . . . . . . . . . . . . . . . . . . 59

4 Simulation result 63
4.1 Trot running with constant speed . . . . . . . . . . . . . . . . . . . . 64
4.2 Trot running with acceleration and gait transition to gallop . . . . . . 66

5 Experiments on the MIT Cheetah robot 71

5.1 Experimental setup . . . . . . . . . . . . . . . 71
5.2 Control parameters for the experiment . . . . 73

5.3 Experimental results . . . . . . . . . . . . . . 75

5.3.1 Estimated locomotion speed . . . . . . 75
5.3.2 Measured leg trajectories . . . . . . . . 76

5.3.3 Estimation of generated leg-force based on applied currents . 78

5.3.4 Gait transition . . . . . . . . . . . . . 79

6 Discussion 81
6.1 Estimated pitch variation . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Fr and COT............................. . ....... 83

6.3 Evidences for the MIT Cheetah running . . . . . . . . . . . . . . . . 84

6.4 Estimated force exerted on foot end-effectors . . . . . . . . . . . . . . 85

6.5 Force analysis on the front legs . . . . . . . . . . . . . . . . . . . . . 86

6.6 Contribution of virtual stiffness and virtual damping to leg compliance 88

6.7 Force analysis on the back legs . . . . . . . . . . . . . . . . . . . . . . 89

7 Concluding Remarks and Future Research Topics 91

A Model Parameters 95

B Back spine coordination 97

8
C Leg compliance 99

D InverseKinematics 101

E Supplementary Video 103

9
10
List of Figures

1-1 MIT Cheetah robot in the experiment, running with trot gait pattern

on the treadmill while constrained on its sagittal plane. . . . . . . . . 27

2-1 The planar-rigid model of the MIT Cheetah robot. A planar model is

constructed since the robot is constrained in its sagittal plane for the

current experimental setup. Leg indices are also listed. FR: front and

right side. FL: front and left side. BR: back and right side. BL: back
and left side . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2-2 Typical schematic diagram for finite state machine . . . . . . . . . . . 31

2-3 Finite state machine to model dynamics of legged mechanism . . . . . 32

3-1 Schematic overall structure of the proposed hierarchical locomotive

control framework. Each block represents each step and corresponding

control parameters. In this schematic diagram, the reference leg is

front right leg, FR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3-2 Sawtooth phase signals generated by phase modulator for trot gait

and its corresponding foot fall pattern diagram. The phase difference

is represented in the diagram. . . . . . . . . . . . . . . . . . . . . . . 51

3-3 Sawtooth phase signals generated by phase modulator for gallop gait

and its corresponding foot fall pattern diagram. The phase difference

is represented in the diagram . . . . . . . . . . . . . . . . . . . . . . 51

11
3-4 The reference desired foot end trajectory with control parameters. 12
control points of B6zier curve are shown which define the swing phase

trajectory(black solid line). Multiple overlapped points are denoted as

'xn'. Stance phase trajectory as sinusoidal wave is also described in

the figure(red solid line). . . . . . . . . . . . . . . . . . . . . . . . . . 52

3-5 separation of 12 control points of B6zier curve for swing phase trajec-

tory into vertical direction and horizontal direction . . . . . . . . . . 54

3-6 Stance trajectory design with equilibrium-point hypothesis: Along with

vertical axis, the tracking error between the actual foot-end position

and the designed foot-end trajectory is converted into the vertical force

incorporated with the leg compliance. The time-integral of the force

during the stance period is the vertical impulse. . . . . . . . . . . . . 58

3-7 Front leg with the impedance control: virtual leg compliance realiza-

tion with stiffness and damping for radial/angular directions ..... 60

3-8 Block diagram for leg control; implementation of the impedance control

to create virtual leg compliance. . . . . . . . . . . . . . . . . . . . . . 61

4-1 Animation interface constructed using MATLAB. 3-D animation of

planar simulator, ground reaction forces, velocity, pitch variation, and

leg phase(stance: 1 / swing: 0) are shown in the animation. . . . . . 63

4-2 Leg trajectory designed for stable trot running. Both front and back leg

trajectories are presented with respect to local shoulder/hip coordinates. 64

4-3 Trot running at 3.4 m/s. Velocity (top) and pitch of the front body

(middle) are plotted over time. Height variation over distance trav-

eled is also plotted (bottom). Pitch variation and height variation are

damped out as the robot approaches to the steady state running. . . 65

12
4-4 Measured ground reaction forces (top) and estimated ground reaction

forces at each leg (bottom) between time t = 12 s to t = 14 s. Verti-

cal ground reaction forces (solid lines) and horizontal ground reaction

forces (dotted lines) are plotted together. Solid-dotted horizontal line

indicates the weight, mg, of the robot, where m = 29.443kg in the

simulation. Input joint torques are coordinate transformed to estimate

ground reaction forces. . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4-5 Snapshots of trot running at 3.4 m/s in the simulation . . . . . . . . 66

4-6 Leg trajectory designed for gallop running. Both front and back leg

trajectories are presented with respect to local shoulder/hip coordinates. 68

4-7 The state of all four legs from t = 6 s to t = 9 s. Gait transition

happened at t = 8 s. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

4-8 Trot acceleration and gait transition to gallop running. Velocity (top)

and pitch of the front body (middle) are plotted over time. Height

variation over distance traveled is also plotted (bottom). From t = 7

s to t = 8 s, the robot accelerates and at t = 8 s, transits its gait from

trot to gallop. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4-9 Measured ground reaction forces between time t = 6 s to t = 9 s. Gait

transition happened at t = 8 s. Solid lines are vertical forces, and

dotted lines are horizontal forces. . . . . . . . . . . . . . . . . . . . . 70

4-10 Snapshots of gait transition from trot to gallop by varying gait pattern

parameters AS in linear over time . . . . . . . . . . . . . . . . . . . . 70

5-1 Experimental setup: the MIT Cheetah robot runs on the speed-controlled

treadmill. Minimal friction sliding guides constrain the robot on its

sagittal plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5-2 Gait trajectories implemented on the experiment with leg workspaces 73

5-3 Snapshot for the trot running gait of the MIT Cheetah robot with each

footfall diagram: the existence of the aerial phase is observed between

each paired legs' step in trotting gait . . . . . . . . . . . . . . . . . . 74

13
5-4 Estimated forward velocity (top) of the MIT Cheetah robot: ±0.15 m/s
bounded difference was observed between the estimated speed and the
treadmill speed measured by a rotary encoder (CUI Inc. AMT 102-V)
directly attached to the DC motor main axis. Froude number (middle,
blue solid line) and cost of transport(middle, red solid line) of the robot
are plotted together. . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5-5 Graph set of generated desired trajectories by the leg trajectory gen-
erator and measured actual gait trajectories during trotting gait from
1 to 6 m/s for a) front right leg, b) front left leg, c) back right leg, d)
back left leg. In each leg graph, top left: desired trajectories, top right:
actual trajectories with respect to speed, bottom left: comparison be-
tween desired/actual trajectories, bottom right: comparison between
actual trajectories at low/high speeds, respectively . . . . . . . . . . . 76

5-6 Commanded currents applied to each EM motor for shoulder/hip and

knee joints at leg i E {FL, FR, BR, BL}: The motor driver supplies
the commanded current to the EM motor which proportionally gener-
ate torque by the torque constant Kr. . . . . . . . . . . . . . . . . . . 78

5-7 Sawtooth phase signals generated by phase modulator for gait transi-
tion from trot to gallop. The phase signal converges from one to an-
other, linearly with time. Gait transition through linear convergence
from the trot parameters to the gallop parameters . . . . . . . . . . . 79

5-8 Snapshots of gait transition from trot to gallop recorded by a high

speed video captured at 500 fps depicting 150 ms . . . . . . . . . . . 80

6-1 Actual measured trajectories of four legs with respect to each should/hip
joint, where all the trajectories for four legs are overlapped with re-
spect to each shoulder/hip joint. The data allow us to approximate
the height/pitch variation of the robot during trot experiment. . . . . 82

14
6-2 Estimated shoulder height trajectories: convex shape of shoulder height

trajectory as a typical representation of running is observed after 2.2

m /s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6-3 Estimated duty factor Det for the front right leg: measured kinematic

data for virtual leg is used to estimate duration of the stance/swing

phases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6-4 Estimation of generated forces in Cartesian coordinate at foot-end of

front legs. Stance phase(high vertical force region) and swing phase(low

vertical force region) can be approximated from the graph. Increase in

peak force value is observed with decrease in duty factor. . . . . . . . 86

6-5 Estimation of generated forces in Cartesian coordinate at foot-end of

back legs. Increase in peak force value is observed with decrease in

duty factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

6-6 Desired/actual vertical virtual leg lengths L . . . . . . . . . . . . . . . 87

6-7 Virtual stiffness and damping contribution to create vertical force at

the foot-end . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

C-1 Measured virtual leg compliance in the polar coordinate (r, 0): a)

longitudinal stiffness b) angular stiffness . . . . . . . . . . . . . . . . 99

D-1 Inverse kinematics of three-segment leg. The proximal and the distal

segment are parallel due to pantographic leg design. . . . . . . . . . . 101

15
16
List of Tables

2.1 Possible states of a leg after impact and corresponding necessary con-

ditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.1 Nomenclature of the set of control parameters . . . . . . . . . . . . . 47

3.2 12 control points for Bdzier curve for designing swing phase trajectory. 55

4.1 Set of control parameters designed for the trot running simulation. . 64

4.2 Set of control parameters designed for the gallop running simulation. 69

5.1 Set of control parameters designed for the experiment. . . . . . . . . 73

A.1 Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

E.1 Supplementary videos and urls are listed in the table . . . . . . . . . 103

17
18
Chapter 1

Introduction

1.1 Motivation and objectives

Developing dynamic legged machines has been an active field of research in robotics

to exploit potential advantages of the legged locomotion. It is widely believed that

legged machines can outperform wheeled machines on rough, irregular and discon-

tinuous terrains in terms of mobility, versatility and energy efficiency [1]. This belief

is supported by locomotion skills that animals exhibit on natural terrains which are

unreachable for wheeled vehicles [2]. These advantages have increased investigations

of legged robots for a variety of tasks such as future transportation, disaster response

operations, and military purposes. Despite many robots already proved their ability

for stable walking, the performance of legged robots has not reached our expecta-

tion; few robots show both stable and agile performance in a rough terrain on which

animals can traverse easily.

In this thesis, we mainly aim to develop a controller to accomplish a stable high-

speed quadrupedal locomotion. To develop such a controller, we try to exploit in-

sights from nature as well as the previous approaches performed by other roboticists.

Therefore, a study on legged locomotion in general such as biological findings, simpli-

fied mathematical models for quadrupeds, and controllers developed for quadrupedal

robots is conducted. In order to verify the controller, we develop a dynamic simulator

using MATLAB, and mathematical derivation of the simulator is also presented. Fi-

19
nally, the controller is implemented on the MIT Cheetah robot, and the experimental
results and analyses of the data are presented.

1.2 Current state of research

Biological findings

Quadrupedal animals have various kinds of gait patterns (walk, trot, pronk, bound,
half-bound, rotary gallop, transverse gallop, etc.). Interestingly, Alexander revealed
dynamic similarity between different sized quadrupeds, by defining a non-dimensional
speed, Froude number Fr, which is strongly correlated with gait patterns [3]. The
mechanism and reason of gait transition from one to another is still in question, but
it was observed that quadrupeds use the same gait pattern when their corresponding
Fr are the same.
The studies of quadrupeds discovered the relation between speed of locomotion,
gait pattern and spatial/temporal limb coordination such as stride length and stride
frequency. Dogs and horses tend to use different gait patterns over different speeds. It
has been revealed that the way animals modulate their speed varies by gait pattern:
dogs and horses modulate their stride frequency when they use trot, whereas stride
length is modulated when they use gallop. In both gait patterns, the duration of leg
swing is maintained constant for various speeds [4], [5], [6] and [7]. Metabolic energy
efficiency is often suggested to explain the reason why dogs and horses change their
gait from trot for mid-speed running to gallop for high-speed running [5], [7], while
other researchers speculate that horses transit gait pattern to reduce peak forces for
each limb [8].

Study of quadrupedal locomotion with reduced-order models

Researchers found reduced-order models that represent the dynamic behaviors of

legged locomotion, which provided better understanding of their characteristics. The
Spring-Loaded Inverted Pendulum (SLIP) model is the most well known model which

20
well predicts the center of mass behavior of running animals with symmetric gait

pattern [9]. In this model, the center of mass behavior is represented as the motion

of a point mass which is attached to an ideal spring. Farley et al. examined that the

effective stiffness of the spring is independent from locomotion speed but determined

by body mass [6].

Ruina et al. [10] developed a collisional model to study the mechanism of redirect-

ing vertical body motion from downward to upward during running, in the perspective

of collisional energy loss. The collisional loss due to impact can be reduced by dis-

tributing impacts through multiple foot strikes. Aligned with this model, Bertram et

al. [11] tried to explain fundamental difference in mechanics between hind-limb initi-

ated gallop and front-limb initiated gallop, while explaining why galloping is effective

at high speed in terms of metabolic cost.

Such simple models exhibit fundamental behaviors of lumped dynamics of legged

locomotion, whereas more elaborate models with specific controllers explain the de-

tails of quadrupedal locomotion with various gait patterns.

A planar model consists of a rigid body and four mass-less springy legs described

pronk and bound of quadrupeds [12] with Energy-pumping feedback controller. Sim-

ilar models were derived in [13] and [14] to compare trot, bound and gallop and

to investigate coordination of four limbs. Herr and McMahon developed a sim-

ilar planar model with additional degrees of freedom in body and legs [15], [16].

Feedforward/feedback control were applied to impose the sequence of four legs with

self-stability property. Stride length and stride frequency were related with forward

running speed to predict animal behavior.

The concept of Central Pattern Generator (CPG) has been introduced as a way

to generate gait patterns. Evidences in biology suggested that a neural network

mechanism can generate gait patterns without any sensory input, and the rhythmic

pattern of limbs and body dynamics are coordinated with sensory feedback [17], [18].

CPGs are often modeled as a set of nonlinear oscillators of which phases are coupled

with each other, to mimic neural networks [19]. Smooth gait transition was possible

by employing existence of the limit-cycle and bifurcation property of such oscillators

21
[20]. The stability and synchronization properties have been studied with contraction
analysis in [21], [22], [23]. An open research topic in controllers using CPGs is to

coordinate the robot body dynamics and the CPG signals with sensory feedback.

Controllers for quadrupeds

Animal studies provided the roboticists with a better understanding in locomotive

mechanism and guides to possible control strategies. A controller based on the varia-

tion of the SLIP model was implemented on RHex robot to achieve stable locomotion

on rough terrain [24]. The stability of the SLIP model also inspired hybrid controllers

which combined the passive dynamics predicted by the SLIP model and higher level

controller [25], [26]. Nonlinear oscillators that mimic biological CPG include Fukuoka

model [27], Van-der-pol oscillator, [28], and Hopf-oscillator [29]. The Tekken [27] and
the Kotetsu [30] achieved walking on a rough terrain, with sensory feedback from

ground reaction forces and body attitude. The Cheeetah-cub [31] showed a successful

trot running using an open-loop CPG, depending on the passive stability given from

compliance of the legs.

On the other hand, Raibert and his company Boston Dynamics, the pioneers of the

legged robotics community, have developed control algorithms which are not directly

inspired by biological findings [32], [33]. The controllers for their recent work are

not published, although they have achieved remarkable performances. Yoneda et al.

presented a stable walking platform which implemented a ZMP-based controller [34].

Trot and pronk gait was realized by a leg thrust regulation controller on the KOLT

robot [35]. A combination of intuitive approaches for a quadrupedal robot controller

achieved various performances in simulation, with a global state feedback [36].

Optimization techniques have been also used to design control algorithms for

legged locomotion. Optimal motion planning algorithms were implemented on Lit-

tleDog platform to achieve a walking and a bounding gait [37], [38], with a machine

learning technique [39], for slow locomotion on rough terrain where careful motion

planning is required. Genetic algorithm (GA) has been also proposed for trajectory

optimization problems for legged machines, because it is a suitable multi-objective

22
optimization tool for dynamical systems with a large number of degrees of freedom.

These properties make GA fits to legged locomotion problem [40]. For example, foot

trajectories were optimized for the metabolic cost of locomotion using GA [41].

In order to directly control body dynamics of the robot, there have been efforts

to optimize forces exerted by four legs. Valenzuela and Kim showed the possibility of

pronk and bound gait for planar quadrupeds by optimizing hip torques at each stride

[42]. The StarlETH realized walk and trot by computing joint torques to generate

optimal forces and torques for controlling posture of the body [43].

Dynamic simulation techniques

Dynamic simulation techniques for a rigid body with kinematic constraints have been

developed in two different ways. One approach directly calculates forces to impose the

constraints, and the other, penalty method, indirectly applies the forces through dis-

placements. Barzel et al. [44] and Witkins et al. [45] proposed the former method. Fric-

tional dynamics on non-penetrating rigid body was also simulated with the Coulomb

friction model [46].

The constraints enforced by the former method are violated due to numerical

errors accumulated during integration. Various methods have been extensively devel-

oped to eliminate this drift on the constraints, either in state-space formulation or

projection methods [47], [48], [49], [50], [51].

Among various methods, Baumgarte's stabilization method, which creates a second-

order dynamics for the constraints to dissipate numerical errors, has been widely used

due to its simplicity and intuitive approach [52]. Optimal Baumgarte parameters for

particular integration solvers with fixed time step size were investigated in [53], [54].

Collision has been another issue in modeling legged dynamics, since collisional

event happens whenever one of the legs hits the ground. It causes the system states

undergo an abrupt change due to large impulsive forces during a short period of

time. Two different approaches have been studied to solve this problem: algebraic

formulations and differential formulations [55]. Differential formulations separate the

impact period into compression and restitution period. They define relationships

23
between normal impulses at contact point of the two objects before/after the impact

according to the Poisson's hypothesis, and integrate the system equations of motion

over the short compression and restitution period to compute states after the impact.

On the other hand, algebraic formulations deal with impact as an instantaneous event.

They define kinematic relation between normal velocities of the two objects at contact

point before/after impact, and solve for the generalized velocities after the impact

using momentum equation. Hurmuzlu et al. [56] proposed a systemic procedure to

solve multi-link rigid body collision problems using both formulations, especially for

the case when impact between a system and the other occurs at one point while they

are already having contact with each other at multiple points. Advanced algorithms

for the algebraic collision law were invented afterwards [55], [57], but these are not

deeply studied in this thesis.

Dynamic simulations for bipeds and quadrupeds have been built based on the

techniques above. Often, they are accompanied with constrained optimization to find

a periodic solution while avoiding slip on the ground. Optimal controller exploit-

ing hybrid-zero dynamics was studied with bipedal robot simulator in [58]. Direct

collocation technique for legged locomotion to find optimal trajectories of the robot

and optimal ground reaction forces were studied with bipedal and quadrupedal robot

simulator in [59], [2], using MATLAB.

1.3 Our approaches and accomplishments

Approaches

Although many robots already proved their ability for stable walking, only a few

robots show both stable and agile performance. To enable quadrupedal robots to

achieve high-speed locomotion, our controller is designed to solve three challenges in

quadrupedal locomotion controller: 1) stabilization of the robot, 2) control of ground

reaction forces, and 3) modulation of gait pattern.

The control algorithm is developed based on three strategies to deal with each of

24
the three issues. Programmable leg compliance is achieved in order to exploit self-

stabilizing property of passive dynamics. We modulate the ground reaction forces us-

ing equilibrium-pointhypothesis by designing trajectories for four legs. A gait pattern

modulator is developed which has proprioceptive touchdown detection as a sensory

feedback. These three strategies are integrated in the hierarchical structure of the

controller.

First, we hypothesized that a quadrupedal robot can self-stabilize without active

posture stabilization using high bandwidth attitude feedback such as body pitch and

pitch rate. The previous studies on the SLIP model proved the existence of the state

space region where the self-stability in the sense of limit-cycle is passively achieved

[9], [60]. It was also shown that this stable region could be enlarged with additional

active actuation [24], and this approach was implemented on the RHex robot [61]. The

self-stabilizing property have also been proposed for quadrupedal legged locomotion

[62], [63], [16]. Inspired by the SLIP model, previous researches installed mechanical

springs on the legged machines to exploit the self-stability [64], [31]. For the MIT
Cheetah robot, we achieve the programmable virtual leg compliance by the impedance

control [65] using our proprioceptive actuation system [66].

Second, we exploit equilibrium-point hypothesis to control the ground reaction

forces. The leg compliance created by the impedance control enables modulation

of the ground reaction force by adjusting its equilibrium position. This interaction

force control mechanism was observed in animals [67] and successfully demonstrated

using a planar manipulator [68]. Aligned with this hypothesis, we propose to design

desired leg trajectories for stance phase to penetrate into the ground. The penetration

depth of the trajectory is controllable to modulate the ground reaction force. The

larger impulse is required, the deeper the trajectory is designed to be. Besides, this

hypothesis provides an unified approach to control both motion and interaction force,

thus no switching between stance/swing control for each leg is required.

Lastly, we propose a gait pattern modulator which coordinates four legs per stride,

combined with a leg touch down event detection feedback. The stride-to-stridemod-

ulation is proposed to coordinate the robot body dynamics and the rhythmic motion

25
of the four legs. The touch down event of the reference leg triggers the pattern mod-

ulator to generate phase signals for four legs. The phases of the other three legs

are coupled with that of the reference leg to accomplish the desired gait pattern.

The touch down event is detected without any additional sensor on the foot. We

call this proprioceptive touchdown detection, which is allowed by the leg actuation

paradigm proposed in [66]. In the previous controllers using CPG, CPG signals are

directly transformed to joint-level trajectories for each joint position controller [27],

[30], [31]. This is distinguished from our approach of which stance phase trajectories

are designed to be reference trajectory for compliance control.

Accomplishments

The controller was verified in simulation experiments with properly chosen control

parameters. The running robot using trot gait reached the constant speed of 3.4 m/s

in the simulation, without any global posture feedback. With a simple accelerating

algorithm proposed for trot gait in this thesis, it also achieved acceleration, without

explicitly changing any control parameter but the target speed. Furthermore, the gait

transition algorithm proposed in this paper was accomplished in the simulation. To

achieve stable strides after transition from trot to gallop, we had to design different

reference trajectories for gallop from trot. This implies the necessity of a further

research on gallop.

The controller was then verified on the MIT Cheetah robot. The robot equipped

with the controller showed a remarkable performance in terms of both speed and

energy efficiency. The Frounde number (Pr) is one of indices to describe dynamic

locomotion which was studied by Alexander [3]. He examined that quadrupeds change

their gait from walk to run at Fr of 1, and transit their gait to asymmetric running

at Fr of 2 or 3. While only a few quadrupedal robots have reached Fr of 1, the MIT

Cheetah accomplished stable trot running up to the speed of 6 m/s, which corresponds

to Fr of 7.34, outperforming the other quadrupedal robots according to [31]. Also,

it achieved highly efficient trot gait rivaling real cheetah, though this result is not

addressed in details in this thesis.

26
Figure 1-1: MIT Cheetah robot in the experiment, running with trot gait pattern on
the treadmill while constrained on its sagittal plane.

1.4 Structure of this thesis

This introduction is followed in Chapter 2 by mathematics and algorithms in detail to
construct the simulator. Hybrid dynamic system in legged locomotion is explained.
MATLAB toolboxes to construct the simulator are briefly introduced as well.
The hierarchical controller is introduced in Chapter 3. Chapter 3 presents the
structure of the controller and the set of control variables to parameterize quadrupedal
locomotion. The underlying ideas and considerations for each level of controller are
explained in detail.
Chapter 4 presents notable simulation results which convinced the authors in
validity of the controller.
The controller is implemented on the MIT Cheetah robot, and the experimental
setup and important raw data of the experiment are exhibited in Chapter 5. The
data are post-processed, analyzed and discussed in Chapter 6 with description of
remarkable performances of the MIT Cheetah robot.
The thesis is concluded in Chapter 7, where summary of the work and future
research directions are presented.

27
28
Chapter 2

Modeling of the MIT Cheetah

robot

A simulator is constructed using MATLAB in order to verify the controller before

implement it on the MIT Cheetah robot. A planar model of the MIT Cheetah robot
running on the flat, rigid ground is derived in the simulator.

BL BR FR FL

Figure 2-1: The planar-rigid model of the MIT Cheetah robot. A planar model
is constructed since the robot is constrained in its sagittal plane for the current
experimental setup. Leg indices are also listed. FR: front and right side. FL: front
and left side. BR: back and right side. BL: back and left side.

The robot and its planar model is depicted in Fig.2-1 with generalized coordinates

29
and leg index i E {FR, FL, BR, BL}. The proximal segment and the distal
segment of each leg are parallel according to the pantographic leg design so that
configuration of three-segmented-leg can be determined by two joint coordinates.
Three independent coordinates (x, y, qpit) define the body coordinate with respect
to the inertial frame. The flexible joints of the spine are designed to be dependent
on hip angles of hind legs (ql,BR, ql,BL) through differential gears. Therefore, any

additional coordinate for describing the motion of the spine is not necessary. The
robot model has 11 degrees of freedom (DoF). Inertial and geometrical parameters are
listed in Table A.1 in APPENDIX.A. The kinematic relationship between each joint of
the segmented spine and hip joints of the hind legs is approximated in APPENDIX.B.

The distinct property of legged locomotion compared to wheeled locomotion is

that it involves 'Touch Down (TD)' and 'Lift Off (LO)' events of all the legs. Be-
fore/after these boundary events, system dynamics changes and the recurrence of
different dynamics on the system is called hybrid dynamics. Note that a floating-
based model is derived in the simulator with appropriate constraint equations to
describe all different dynamics.

MIT Cheetah robot is modeled using Lagrangian formulation with following ad-
ditional assumptions.

" Each segment of the robot is perfectly rigid.

" The legs of the robot interact with the ground as point feet.

" The ground is perfectly rigid half-space.

" The interaction between each foot and the ground follows the Coulomb friction
model.

" Impact between the foot and the ground is perfectly inelastic.

" Each joint is frictionless.

30
2.1 Hybrid dynamic systems

2.1.1 Hybrid dynamic systems in legged systems

Hybrid dynamic system exhibits both continuous and discrete dynamics. The system
consists of multiple continuous dynamic systems, and when a certain condition is
satisfied the system states jump from one to the other continuous dynamic system[58].
The discrete transition of the hybrid system can be modeled as a Finite state machine
(FSM) as drawn in FIG.2-2. [69].

guard 1/action a

State A State B

Initial state guard 2/

Figure 2-2: Typical schematic diagram for finite state machine

If 'guard 1' is satisfied, transition from state A to state B happens, and action a
is executed. If 'guard 2' is satisfied in dynamics in state B, transition from state B
to state A happens without executing any action.
Discrete footfall pattern of legged locomotion inevitably involves hybrid dynamics
[2]. One should notice that the dynamics of the machine interacting with the ground
(stance phase dynamics) is different from that of the machine in the air (flight phase
dynamics). The transition happens whenever any leg of the machine touches down or
lifts off the ground. In particular, when a leg touches down the ground, not only the
dynamics changes from the moment but also instantaneous change in system states
will occur due to impact.
As described in 2-3, flight phase, stance phase, leg touch down, leg lift off and
instantaneousjump in states due to impact map correspond to state A, state B, #1 , 02
and action a, respectively. Typically, touch down event is detected with foot height
and lift off event is detected with normal component of ground reaction force at the
leg [70]. Mathematical description of the guard functions are as follows, for a leg

31
 Touchdown(Oi)/Im pact Map

§ ght hase Stance Phasje

initial state Liftof(,)

Figure 2-3: Finite state machine to model dynamics of legged mechanism

which is labeled as i:

01 : hi(x) = 0 A hi(X) < 0 (2.1)

q2 : Fie,i(x) = 0 A P'gt,j (X) < 0, (2.2)

where x is the state of the dynamic system, hi (x) is height of leg i and Fn ,i(x) is the
normal component of ground reaction force of leg i.

2.1.2 hybrid dynamic systems realized in MATLAB

To simulate dynamics in MATLAB one should define equations of motion (EoM) of

the system. We choose to derive EoMs of the MIT Cheetah robot by floating-based
model with appropriate constraint equations in MATLAB.
If the number of different dynamic systems is not large, for example hopping
machine, it is desirable to derive EoMs for each state because it gives accurate dy-
namics satisfying the constraints. However, as the MIT Cheetah robot has four legs,
it becomes much complicated and inefficient to derive equations of motion for all 256
dynamic states, combination of each leg condition:

case 1. not in contact with the ground

case 2. in contact with the ground and fixed to the contact position

case 3. in contact with the ground and slipping in one direction

32
case 4. in contact with the ground and slipping in the other direction.

In this case it is more beneficial to derive one set of floating-body-based equations of

motion and impose appropriate constraints according to each dynamic states. Fur-

thermore, floating-based model can directly computes constraint forces such as ground

reaction forces, which is required to detect 'lift-off' event. Therefore, it is common to

choose floating based model in the legged robot dynamic simulation as in [58] or [21

and the author also follows the convention. However, it naturally induces accumula-

tion of numerical error on constraints and therefore an effort to inhibit this drift is

additionally required.

To realize the hybrid dynamics in the simulator one should evaluate the guard

functions and provide proper action in transition while integrating the dynamics

over time. The author uses MATLAB ode45 solver with odeevent option. The

event function looks for either foot height for aerial phase legs or normal component

of ground reaction forces for stance phase legs. When one of these values crosses

zero from positive to negative, the ode solver stops integration, the set of constraint

equations is changed, and ode45 restarts integrating with updated initial conditions,

if necessary.

Note that, using the standard ode toolbox of MATLAB has advantages and dis-

advantages. Because it is a "black box" with varying step size, it turned out that

first, it is difficult to choose proper parameters for numerical error inhibition, second,

it is hard to have access and print dependent variables which are not defined as state

of the dynamics, such as ground reaction forces or input efforts, and third, sometimes

the event is not detected precisely without a clear reason. Nonetheless, MATLAB's

ode solvers has varying step size, which enables them to detect events more accurately

than fixed time step solvers. The powerful aspect of this accurate event detection is

desirable, thus we build the simulator on MATLAB ode45.

33
2.2 Constrained equations of motion

2.2.1 Multi-link rigid body equations of motion

Dynamics of the 11-DoF MIT Cheetah is governed by

D(q)ij+ C(q,4)4 + G(q) = Bu + Jc(q)TFest (2.3)

where q :[qitch, ql,FR, q2,FR, -'' , ql,BL, q2,BL, x, yJT E R"? is the generalized
coordinates of the model. The front body is chosen as reference floating body with
three coordinates [qitch, x, y]T to describe its posture and position. D(q), C(q, 4)4,

G(q) and B are the inertial matrix, Coriolis and centrifugal terms, gravitational

torque vector, and the input matrix, respectively. The matrices are calculated through

Lagrange formulation,

d OL(q,4) _L(q,) -(
dt~ aziai (2.4)

where L(q, 4) = T(q, ) -V(q) is Lagrangian of the system which is difference between

kinetic energy (T) and potential energy (V), and Ei is generalized forces. u E R 8 is

leg joint actuation torques which are transformed from geared BLDC motor outputs.

Jc(q) = ) is a Jacobian matrix of pc(q) E 3 2Ncx1, a stack of position vectors

a9q

of each ground-contact foot with respect to the inertial frame. Nc is the number of

ground-contact feet. Jc(q)TFe,, is the ground reaction forces(GRFs) transformed into

the joint space, which is obtained by principle of virtual work.

To solve for 4 and GRFs, holonomic equality constraints on positions of the

ground-contact legs have to be considered:

cIh(q) = Ohx1, 'h(q,4) = Ohx1, (2.5)

where h is the number of constraints. Note that, since the holonomic constraint D is

34
function of q only,

'Ih(q, 4) = Jh', 'h(q, 4, 4) = jh4 ± Jh4, (2.6)

where Jh(q) = The partial derivative terms with respect to time, h , are
ignored, since constraints are not explicitly dependent on time in this thesis. Ib = 0,
( = 0 and (5 = 0 are named legal position, legal velocity, and legal accelerationand
the state satisfying these constraints is named legal state in [45]. Equality constraints
for each leg are applied from the moment when the leg makes contact with the flat
ground.

2.2.2 Numerical error inhibition

If the initial condition of the dynamics satisfies Eqn.(2.5), and the dynamics is accu-
rately integrated, the states will remain in the manifold which suffices the constraint
equations, if we solve Eqn.(2.3) with N = 0. In practice, however, feedback terms are
required to bring the states back to the desired manifold from the numerical error,
or, drift, which is accumulated during integration or initiated by impact map [45].

Various methods have been extensively studied to inhibit drift on the constraint
either in state-space formulation or projection methods[47]. Particularly, the Baum-
garte's stabilization method [52] is an intuitive and therefore widely studied method
including this work. It creates a virtual damper and a virtual spring around the legal
state so that even though small error occurs, the state of the equations converges to
= 0 and e = 0. Baumgarte's stabilization method is formulated as

4(q, 4) + c4(q, 4) + /3((q) = 0, (2.7)

where a and # are called Baumgarte parameters. Ideally, any set of Baumgarte
parameters will eventually make the system states converge to legal state. In practice,
however, Baumgarte parameters should be carefully chosen since the stability of the
discrete numerical integration is affected by choice of the parameters. Choosing proper

35
Baumgarte parameters have been studied for a particular type of numerical integrator
with fixed time step [53], [54]. However, a rigorous parametric study on varying time

step does not exist.

In our case, it was observed that Baumgarte's parameter values do not affect that

much on the stability of the system. It is because constraints of a running legged

system is frequently changing and therefore duration for any error to accumulate

is very short. Therefore, we arbitrarily choose parameters to create an critically-

damped 2da order dynamics to inhibit oscillation which may induce inaccurate event

detection and to employ the fastest convergence rate of the critically damped system:

a = 2(w, # = w, while ( = 1, w,, = 15rad/s.

As a result, two bundle of equations Eqns.(2.3), (2.7) are solved for 4 and F,,t

simultaneously with the Coulomb friction model and the impact map.

2.3 Friction model

Designing a controller which is able to prevent a robot leg from slipping is a very

challenging task, even many studies have built optimal controllers which satisfy the

no-slip assumption [58], [59]. The task itself is difficult even in the simulation, and

moreover, because of inaccurate ground contact model, even if the controllers are

implemented on a real robot, they cannot guarantee the robot will not fall off at all.

Therefore, in the author's opinion, a simulator should allow a robot to slip with ap-

propriate friction model, to provide more realistic environment to assess performance

of a controller designed for a robot.

We decide to use the Coulomb friction model which is widely used in legged lo-

comotion community, as [2], [38], and [58]. The Coulomb friction model defines a

relation between tangential and normal component of the GRF, F and F, respec-

tively, namely friction cone: Ft <; pFI. Here p is the friction coefficient between a

foot and the ground. We set the same value for both static and kinetic friction coeffi-

cient, P, = Pk = 1. Computed GRFs, Fext, from Eqns (2.3), (2.7) should obey friction
cone, and are highly related to the construction of equality constraints, Eqn. (2.5). Es-

36
pecially, if a foot slips, the magnitude of tangential force is determined as jFt = pl F,,

and the number of unknown variables decreases.

Consequently, the construction of Jacobian J, at each time step highly affects

construction of a set of equations and solution of the problem. We classify a 'non-slip

ground-contact foot (NS)' and a 'slip ground-contact foot (S)' and to distinguish legs

having different states, indez vectors, are created: IdxNs is a 4 x 1 vector for non-

slipping legs and IdxS is a 4 x 1 vector for slipping legs. Sum of these two vectors

yields IdxTD, a 4 x 1 vector for ground-contact legs. Each boolean element of these

vectors represents FR, FL, BR, and BL, and if the value of the element is 1 then

corresponding leg is a element of the set. For example, if IdxNS = [0, 0, 1, 1 JT and
IdXS = [1, 0, 1, 1 ]T, then FR leg is slipping, BR and BL legs are fixed to the ground,
and FL leg is not contacting with the ground. Though these discrete states are not

explained in detail in the following algorithm, they play an important role to write

code in MATLAB.

In the simulator developed in this work, the frictional multi-contact problem is

solved in following seven steps.

Step 1. Solve with non-slip assumption and check friction cone condition

As Baraff mentioned in [71], a good guess to begin iteration with is to assume

each contact is maintained during integration . Additionally, we can assume

each contact is maintaining non-slip contact, since this is most likely happen

during legged locomotion with a well designed controller, and this assumption

reduces computational burden. In that case, unknown variables 4 and Ft are

simply solved for. Let H = C4+G -Bu and R = 6h-Jhq = -)h-f3 - Jhq.

Then, Eqns. (2.3), (2.7) can be rewritten as,

D4 + H = JeFext (2.8)

JO = R (2.9)

37
 and then solutions are

4 = D-'(-H + JcFext) (2.10)

Fext = (JhD~'Jc)-1 (R + JhD- H). (2.11)

Calculate ratio of tangential component to normal component of the ground

reaction force for each leg i and check whether the ratio is not larger than

friction coefficient, pi = i < p. If no force vector disobeys friction cone, the

iteration is terminated.

Step 2. Sort out a sliding leg for the next iteration

The leg of which pi is maximum, among those whose values are larger than

friction coefficient, is sorted as a leg sliding on the surface for the next itera-

tion. Update slip index vector IdxS and non-slip index vector IdxNS for the

next iteration. Here, the direction of slip is determined to be opposite of the

tangential force computed in the previous iteration.

Step 3. Reconstruct new constraints set and corresponding Jacobians

The holonomic equality constraints are applied to the position of each ground-

contact foot. Non-slip ground-contact foot position should be fixed during in-

tegration, whereas the constraint in tangential direction for slip ground-contact

foot should be vanished to allow it to accelerate along the tangential direction.

-D~q [cS
()
NS
) PNS(qO)
psq)
02NNsxl
ON~1 (2.12)
[pc,.(q)J [p (o)j ONsX 1 j
S[pNSq NS(q) . 02NNSx1
h(q)= - = q , (2.13)
c'.
p(q) J.S (q) ONS x 1

where pcjS(q) is a set of position vectors of non-slip contact feet, pc,(q) is a

set of normal component of position vectors of slip contact feet, and qO is the

generalized coordinates at 'TD' event. NNS is the number of non-slip ground-

contact foot, and NS is the number of slip ground-contact foot. Since the

38
 constraint sets are modified, the Jacobian of holonomic constraint equations in
Eqn.(2.9) also should be modified.

jNS
Jh = (2.14)

Step 4. Reconstruct necessary Jacobians for EoMs

For tangential component of GRF at the foot which slips, its direction is pre-
defined from the previous iteration and its magnitude is function of normal
component of the GRF (IF 3 l = IFil). Therefore, Ft is no longer an indepen-
dent variable and the number of unknown variables should be reduced from #
and Fezt = [FtNS, F , F, F]T to 4 and Fext = [FtNS, Fn, Fl]T. Conse-
quently, Eqn.(2.8) should be reorganized as

D(q)4 + C(q, 4)4 + G(q) =Bu + JNSFNS + (J ± J)F +---

jNS F NS
=Bu +
JZ + MJt F
=Bu + JeFe.t, (2.15)

kp. .-- 0
where M = '. contains information of the direction of tangential

0 ---. k
force at corresponding leg.

Step 6. Solve for constraint forces and accelerations

With the updated J, and JA from Eqns.(2.14), (2.15), solutions 4 and Fext =
[FtNs, F,N3 , Fl]T can be attained directly from Eqns.(2.10), (2.11). Fts is
calculated as F 3 = ±LF.

Step 7. Terminate Iteration if all tangential forces are bounded.

Iteration from Step 2. through Step 6. is terminated when the ratios pi are all
bounded by [.

39
 Lastly, the outputs of the iteration are observed by the solver at each time step.

Once any non-contact foot position becomes penetrating the ground, or any normal

component of ground reaction force of contact foot becomes negative, the integrator

is terminated and re-initiated with updated IdxTD.

2.4 Impact map

When system dynamics make a transition from one to another with 'TD' event(s), the

system states undergo an abrupt change due to large impulsive forces during a short

period time of impact. Two different approaches have been studied to resolve this

collision problem: algebraic formulations and differential formulations [56]. Differ-

ential formulations separate collision into compression and restitution period, define

relation between normal impulses of two object at contact point before/after impact

with Poisson's hypothesis, and integrate the system EoM over the short duration to

compute states after the impact. On the other hand, algebraic formulations deal with

impact as an instantaneous event, and define kinematic relation between two obejcts

using normal velocities at contact point before/after impact, and solve for generalized

velocities after impact using momentum equation.

In this work, we use algebraic formulation to update the state of the system at

impact. Mason and Wang [72] observed contradiction in using the algebraic formu-

lation, such that kinetic energy of the system increased after the impact for certain

cases. However, if perfect plastic collision was assumed, such problem did not occur.

The perfect plastic impact is normally presumed in dynamic simulation of legged

robots [58], thus we use algebraic formulation which is more stable, faster, and easier

to implement than differential formulation.

Compliant ground model can be employed in order to avoid difficulties in rigid

ground models such as inconsistency in impact law. It also provides continuity for

gradient-based optimization, as well as simulation environment for both soft and rigid

grounds [73]. Nonetheless, it requires additional states which increases computational

cost and induces stiff problem due to high stiffness and damping of the ground model

40
[59]. Besides, we want to evaluate the performance of the controller under the ex-
istence of the discrete jump in the states, which occasionally happens in the real
system.
Generalized positions are invariable before/after impact (q+ = q-). The general-
ized velocities after impact, 4+, can be solved via the algebraic impact law, Eqn. (2.17),
with appropriate equality constraints, Eqn.(2.18).

q+ = q- (2.16)

D(q)4+ - D(q)4- = Je(q)T j Fextdt (2.17)

~h(q) = J4

= Ohxl (2.18)

The momentum equation is constructed by integrating the governing EoMs over a

short instant (At = t+- t-) before and after impact. Coriolis and centrifugal terms,
gravitational terms and input terms are negligible compared to impulsive forces, thus
ignored in the equation [58].
Multi-link collision problem for the legged locomotion is a complicate problem to
solve. When collision occurs at a leg, some of the other legs are in unconstrained
contact with the ground, and whether these legs rebound or slip after impact will
strongly affect the result. Hence a systematic procedure is required to solve the
problem, and we slightly modified that of [56].

Step 1. Similar to the multi-link frictional contact problem, we start the iteration
with the simplest conditions. Solve with no-slip, no-bound assumptions for each
ground-contact leg after impact, and check whether conditions are satisfactory
with the assumptions. Let us define 'mp = ft Fextdt, impulse. Then, solutions
for Eqns. (2.17), (2.18) are

4+ = 4- + D1 JImp (2.19)
Imp = (JhD- 1 JT)~1 Jh4-. (2.20)

41
 In general, a leg having contact with the ground before impact can have either
one of four different states after impact, and solution with each state should

satisfy a certain condition. Possible leg states after impact and correspond-

ing necessary conditions are listed in 2.1. If the solutions from Eqn. (2.19),
Eqn.(2.20) satisfy all the conditions, terminate the iteration. Otherwise, pro-

ceed to the next step.

State at t+, after impact Conditions to satisfy

Detaches without interaction V+ > 0
Imp = 0

Sticks to the ground Imp,n > 0 and |MP't|

Imp,nl
: = v+ = 0
Slides on the ground in the posi- Imp,n > 0 and sign(Imp,t) $ sign(vf)
tive tangential direction
:+ = 0 and 'mp,t
mp,n
Slides on the ground in the nega- Imp,n > 0 and sign(Imp,t) # sign(vf)
tive tangential direction
V+ = 0 and Imp,n
mp,t

Table 2.1: Possible states of a leg after impact and corresponding necessary conditions

Step 2. Update IdxFS, and IdXBS.

Obtain set of combinations of leg states by reorganizing state vectors, IdxBS

and IdxFS, those which slide in the positive and negative tangential direction,

respectively. Construct M, Jh and J, accordingly as in Eqn.(2.14), (2.15). If

the solutions from Eqn. (2.19), (2.20) satisfy all the conditions, terminate the

iteration. Otherwise, proceed to the next step.

Step 3. Update IdxNs, and Idxs.

Obtain set of combinations of leg states by reorganizing state vectors, IdxNS

and Idxs, those which stick to the ground and slide on the ground, respectively.

Construct M, Jh and J, accordingly as in Eqn.(2.14), (2.15). If the solutions from

Eqn. (2.19), (2.20) satisfy all the conditions, terminate the iteration. Otherwise,

proceed to the next step.

42
Step 4. Update IdxTD.

Obtain set of combinations of leg states by reorganizing state vectors, IdXTD

those which maintain contact after impact. Some of leg contacting with the
ground may rebound after impact. Construct M, Jh and J accordingly as in
Eqn.(2.14), (2.15). If the solutions from Eqn. (2.19), (2.20) satisfy all the
conditions, terminate the iteration. Otherwise, proceed to the next step.

Note that in the numerical computation, the conditions in 2.1 should have thresh-

old values, for example, vt > le - 10. The problem often issued about this algebraic

approach is that sometimes increase in kinetic energy after is observed. However,

with the perfect plastic assumption in this simulation, the problem has never been

observed.

43
44
Chapter 3

Quadrupedal Locomotive Control

Framework

This chapter presents a new locomotive control framework for quadrupedal robots

developed in this thesis. The control algorithm is developed under guidance and lots

of discussion with Dr. Dong Jin Hyun. We address three challenges in designing

quadrupedal locomotion controller: 1) stabilization of the robot, 2) control of ground

reaction forces, and 3) modulation of gait pattern.

The control algorithm is developed based on three strategies to deal with each

of the three issues. Programmable leg compliance is achieved in order to exploit

self-stabilizing property. We modulate the ground reaction forces using equilibrium-

point hypothesis by designing trajectories for four legs. A gait pattern modulator

is developed which has proprioceptive touchdown detection as a sensory feedback.

These three strategies are integrated in the hierarchical structure of the controller.

The controller is governed by a set of control parameters, such as target speed, target

gait pattern, and the spatial properties of the leg trajectories.

The overall structure of the proposed controller is schematically represented in

Fig.3-1. An operator specifies 1) a desired velocity vd and 2) a target gait pattern

expressed by AS. The gait pattern modulator and the leg trajectory generator con-

stitute the high-level controller which manages and plans the motion and role of four

legs. Proprioceptive touch-down detection on a reference leg provides the stride initi-

45
 TD Event Detection
High-level Controller

Gait Pattern, Modulator Leg Irajectory Generator

Desired Foot-end Trajectory

----------- -------------- --------

Desired Speed Target Gait Pattern Leg Cotroller

ASPRPL K,, Kd,, -

A$= ASPRBR K K
ASPRBL__
....... ----..
.... -..-.--.. -- - - - - - - _ - - -
Operator Low-level Controller Encoder
Signals
Currents

MIT Cheetah Robot

& --
Envirornent

Figure 3-1: Schematic overall structure of the proposed hierarchical locomotive con-
trol framework. Each block represents each step and corresponding control parame-
ters. In this schematic diagram, the reference leg is front right leg, FR.

ation timing. The phase signals generated by the gait pattern modulator are mapped

to the desired trajectories of four legs. The low level individual leg controller controls

each leg as planned. Table 3.1 shows the control variables to be managed in the

proposed controller.

3.1 Gait pattern modulator with proprioceptive

sensory feedback

The gait pattern modulator achieves a desired velocity and a target gait pattern

by coordinating the robot's four limbs with the phase signals, S. The phase signal

assigned to each leg is a fraction of the desired swing and stance period, T,, and

Tt, respectively, which are determined in accordance with the commanded desired

velocity.

Gait pattern is one of characteristics of quadrupedal locomotion. Animals tend

to have desirable gait pattern according to their speed of locomotion. Energetic

46
 Terminology Definition
Vd Desired speed
ASFR,i the phase lag between FR leg and leg i
T~t Desired stance phase period
t. Desired swing phase period
C = [CF, CB] Bezier control points for swing phase trajectory
L.,a. = (L.pan,F, L.pan,B)T half of the stroke length
6 = (SF, B)T penetration depth for stance phase trajectory
PO = (P,F, POB)T reference point of the trajectory
Kp,r radial stiffness of each leg
Kd,r radial damping of each leg
K,,o angular stiffness of each leg
Kd,O angular damping of each leg

Table 3.1: Nomenclature of the set of control parameters

advantages of choosing different gait patterns for different speeds have been observed
and therefore it is believed that robots should also be capable of choosing different
gait patterns. To impose a specific target gait pattern such as walk, trot, canter, and
gallop, the gait pattern modulator assigns the phase lag between four legs, As, so
that the temporal limb coordination is achieved.

3.1.1 Phase signals

Generated phase signals are referred as parameters for designing stance/swing phase
trajectories by the leg trajectory generator. Each signal Sj E [0, 1] is time-normalized
by T8 t and T,,, where i E {FR, FL, BR, BL} is leg index and j E {st, sw} is leg state
index(stance, swing).

Maes et al. [4] illuminated that the duration of leg swing remains steady over
various speeds. Respecting the natural behavior of biological systems, the desired
swing phase period, Ts, is prescribed as constant, 0.25 sec.

While maintaining the swing duration, animals decrease stride frequency and con-
tact time as the locomotion speed increases [5], [7]. Therefore the desired stance phase
period, T8 t, is determined according to the desired speed Vd as:

2Lsa
'st = 2L 1" (3.1)
Vd

47
where L.,an is half of the stroke length as depicted in Fig.3-4. The stroke length
approximates a distance traveled of the shoulder/hip joint during stance phase. This
implies we only modulate stride frequency to accelerate, which is similar to [74].

Once the gait pattern modulator is triggered, it assigns stance phase Si" and then
swing phase S" to each leg, which completes a whole stride.

3.1.2 Synchronization by proprioceptive sensory feedback

The phase signals generated by the gait pattern modulator have to be synchronized
with the environment such that stance phase or swing phase is properly commanded
to a leg when it touches down or lifts off the ground, respectively. Therefore, detection
of 'TD' event and 'LO' event become crucial. In lots of legged robot researches, a force
sensor is attached to the foot for the event detection, and furthermore, for ground
reaction force feedback control and for stance/swing switching control.

The MIT Cheetah robot leg with low mechanical impedance enables the propri-
oceptive 'TD' detection without additional force sensor, by sensing abrupt change
in the commanded force computed by the impedance control scheme. When a leg
touches down the ground, radial force in outward direction is commanded by the
impedance controller in order to impede the leg's sudden flexion, which is induced by
the ground impact. The value is comparatively higher than one commanded during
the swing phase; therefore, the 'TD' event can be recognized in a way like a limit
switch, by setting a threshold value on radial force, F,.

Note that the leg may lose contact for a short instant during the stance phase due
to uneven terrain, which leads to the controller's mis-detection for 'stance to swing
transition'. To prevent this failure mode, we activate proprioceptive 'TD' detection
after completion of 90% of the swing phase. 10% margin is given for the leg's agile
response to the early occurrence of the 'TD' event'.

'The touch down sensor thus acts similar to limit switch. This type of sensing might be vulnerable
to chattering of leg but note that the chattering after 90% of the swing phase is mainly in the radially
outward direction by the centrifugal force.

48
3.1.3 Imposing gait patterns

In order to impose a desired gait pattern, phase coupling between four legs is required.

If all the leg phases are initiated by each 'TD' event, however, it is challenging to

maintain the consistency of the specific gait pattern. A novel yet simple way is
developed to solve this problem: TD event-based stride to stride pattern modulation.

Detection of 'TD' event initiates generation of the phase signals of the reference leg,

and the other three legs' phases are coupled with that of the reference leg having

predefined phase lag, AS. The proposed method can be versatilely applied to any

symmetric/asymmetric gait pattern.

This 'stride to stride' method generates phase signals just for one stride and waits

for the next trigger. Therefore, after a completion of one stride, all the legs become

stationary until next 'TD' event on the reference leg is detected. This undesired
stationary period is the result of minimal adaptation to the environment while main-

taining a specific gait pattern, which is necessary for locomotion on rough terrain

with unexpected external disturbances. Ideally, the undesired stationary period can

be vanished if the pre-defined swing phase period, 'T,, is exactly matching to the

actual value, or feasibly estimated during stance phase in priori.

3.1.4 Phase signal generation according to TD elapsed time

and phase differences

A mathematical formulation of the phase signal generation is presented here. The

phase signals are created referring to clocks of each leg. The reference leg clock is the

time elapsed after 'TD' event, and the other clocks have time delay with respect to

the reference leg clock.

The proprioceptive 'TD' detection is represented as a boolean variable 'TD'. It

changes from FALSE to TRUE when the event is detected, and one stride cycles with

the elapsed time after the event as following:

49
 t - tTD
telapse - ref (3.2)
refia
refid if telapse>Ttra
ref >stride

re = t if (S' > 0.9) A ('TD') (3.3)

ti= t!las --refASrefistride,
~ i (3.4)

where tlpe is the elapsed time after tTI which is the moment of the TD event on
the reference leg. ti is the clock for each leg, where i E {FR, FL, BR, BL}. ASrefi is
the phase lag of the leg i with respect to the reference leg, represented as a fraction
of the desired stride period. Note that ASrefref is zero by definition. tTj is updated

discretely when the next 'TD' event is detected, while 90% of swing phase of the
reference leg(Sre") is completed.

Si increases from zero to one during ist and subsequently Sj" repeats from zero
to one during ,,. Each leg phase signal follows the reference time having time delay,
or, phase lag with respect to the reference leg.

_ti-

Si = t-i 0 < ti < tst (3.5)

Is ti$t" if -iTm < ti < 0

S:" = t~ (3.6)
iT"
Taw
if Tst < ti < Tstride

Here we present phase signals generated by proposed phase modulator and cor-
responding foot fall pattern diagrams for trot and gallop gait pattern used in the
experiment. A front right leg (FR) is chosen as the reference leg in the examples. To
accomplish a specific gait pattern, appropriate phase lags are prescribed with respect
to the FR leg. Trot gait is a symmetric gait pattern, of which phase for contralateral
legs are the same.

ASFR,FL 0.5
AStrot = ASFR,BR 0.5 (3.7)

ASFR,FL trot

50
 FL *A.;
BR

Figure 3-2: Sawtooth phase signals generated by phase modulator for trot gait and
its corresponding foot fall pattern diagram. The phase difference is represented in
the diagram.

-_.'75k,*'" jp
BL~ASRF

The phase difference between each leg for a gallop, which is an asymmetric gait
pattern, is not well defined but varies by literature. Here, following phase difference
is chosen for gallop.

[ASFRFL [0.2]
Agallop = ASFRBR] =allop5=(3.8)

. .... go p .ed

FR
FL
BR-D

Figure 3-3: Sawtooth phase signals generated by phase modulator for gallop gait and
its corresponding foot fall pattern diagram. The phase difference is represented in
the diagram

51
3.2 Leg trajectory generator

The leg trajectory generator transforms the phase signals from the phase modulator to
the desired trajectories for each foot-end which are designed with variable parameters.
The leg trajectory generator plays a key role in our scheme; the robot is capable
of managing its stability with adjustable trajectories while following a specific gait
pattern commanded by the gait pattern modulator.
Swing phase and stance phase trajectories are designed individually in different
perspectives while guaranteeing continuous and smooth transitions: pure position
control and compliance force control, respectively.
Swing phase trajectories are designed by Bezier curve defined by a set of 12 control
points, C = {ck}, and stance phase trajectories are spatially designed as sinusoidal
wave of which period is twice of stroke length. Also, the trajectory is designed about
the reference point, PO with respect to the local shoulder/hip coordinate system.
Fig.3-4 shows the designed reference gait trajectory.

0: control points for Bezrturve

200 -

300 -

Ctx2.
400- rotr on
IWW-throujgh
7 n

stance

7001 1
-200 -100 0 100 200 3
x-axis (mm)

Figure 3-4: The reference desired foot end trajectory with control parameters. 12
control points of Bezier curve are shown which define the swing phase trajectory (black
solid line). Multiple overlapped points are denoted as 'xn'. Stance phase trajectory
as sinusoidal wave is also described in the figure(red solid line).

52
3.2.1 Design of the trajectory in the swing phase

The primal objective of designing the swing phase trajectory is to protract a leg

with sufficient ground clearance to avoid obstacles and with sufficient leg retraction

rate. Also, feasible swing phase trajectory, pi"(t), should require relatively low energy

consumption for overall energy efficiency of the locomotion.

The three-segmented pantographic design of the MIT Cheetah leg enables protrac-

tion with low energy usage by exploiting its natural dynamics. Swing phase trajectory

is designed by two dimensional Bezier curve constructed with 12 control points while

approximating the protraction trajectory to around the natural behavior of the legs

as well as satisfying geometrical requirements (ground clearance). It is parameterized

by corresponding swing phase signal, S,"w E [0, 1] as:

n
piW(t) = pi"(Sfw(t)) = cB (Ssw(t)) (3.9)
k=O
S dpi dSr" 1 dpi
= dSfw dt dS(3.10)

where Bn(Sm (t)) is the Bernstein polynomial of degree n, (n+1) is the number of

control points, Ck E R2 is a k-th control point and k E {0, ... , 11}.

Note that a r-th derivative of a Bezier curve is as below 2 and therefore it has the

following useful properties:

p(rs) ! A(r)ckB -(s) (3.11)

(n -k=1

pwis =0 = cO V|S5w= = nAco/Ts

8

* P1Istw=1 = c vi"|sw=1 = nAcn/s

* Double-overlapped control points generate zero velocity.

* Tripple-overlapped control points generate zero acceleration.

Horizontal (x) and vertical (y) positions of control points are designed indepen-
2
A, is r-th forward difference operator: Ar ck = A- 1
Ck+1 - Ar-1ck

53
dently based on these properties, in the parasagittal plane with respect to hip/shoulder

joint. In vertical direction,

1. double overlapped control points, co,y and ci,.,, are used for zero velocity at 'LO'.

2. For transition from 'follow-through' to 'protraction', the acceleration direction

is changed by triple overlapped control points.

3. During 'protraction', double overlapped control points change the direction of

trajectory.

4. For comfortable 'TD' with 'swing leg retraction' in the y-direction, the acceler-

ation direction have to be changed by triple overlapped control points.

5. double overlapped control points, cio,, and cii,,, are used for zero velocity at

'TD'.

The total required 12 control points are shown in Fig.3-5 as red circles.

y-axIs

. 0.a v.

Figure 3-5: separation of 12 control points of Bezier curve for swing phase trajectory
into vertical direction and horizontal direction

In horizontal direction, the leg initially has 'follow-through', changes the direction,
has 'protraction', changes direction again and has 'swing-leg retraction' sequentially.
Therefore,

1. The horizontal velocity at LO is defined as the difference between co, and ci,.

54
 2. To change the direction after 'follow-through', double overlapped control points

are positioned.

3. During protraction, acceleration direction is changed. Thus, triple overlapped

control points are positioned.

4. Changing direction to have 'swing-leg retraction' requires double overlapped

control points.

5. The velocity of 'swing-leg retraction' is defined by the difference in the x-

direction between cio,x and cax

Fig.3-5 shows 11 required control points as blue circles. In this thesis, 12 control
points are used for the horizontal direction as well for 'single instruction multiple

data streams (SIMD)', a fast and efficient computation algorithm, by coping with
B6zier control points as 12 x 2 array.
The 12 control points for the designed gait trajectory are listed in Table. 3.2.
Satisfying the requirements described, the swing phase leg-trajectory, of front right
leg for example, is designed to create ground clearance of 0.15 m in the middle of
'protraction', swing leg retraction speed of 4.0 m/s with prescribed '' = 0.25 ms,
Lp,,= 200 mm, and Po = (0 mm, 500 mm) as Fig.3-4.

x (mm) y (mm)
co -200.0 500.0
ci -280.5 500.0
c2 -300.0 361.1
c3 -300.0 361.1
c4 -300.0 361.1
c5 0.0 361.1
c 0.0 361.1
c7 0.0 321.4
c8 303.2 321.4
cq 303.2 321.4
cIO 282.6 500.0
c1 1 200.0 500.0

Table 3.2: 12 control points for B6zier curve for designing swing phase trajectory.

55
3.2.2 Design of the trajectory in the stance phase

The stance phase control of each leg directly affect the performance on quadruped

locomotion via interaction with the ground. Therefore, the stance phase trajectory

has to be designed in a different standpoint from a pure trajectory control as for

swing phase. Rather, it should be understood as a reference trajectory to generate

force with impedance controller as compliance force control scheme. A trajectory

in horizontal direction is related to forward locomotion, and a trajectory in vertical

direction is related to the vertical GRF applied to the robot, respectively.

The stance phase trajectory is proposed to simply be a sinusoidal wave with two

parameters depicted in Fig.3-4: 1) half of the stroke length, L,,, and 2) amplitude

variable, 6. Based on the phase signals, Sit, the reference foot-end trajectory of each

leg is generated during it 8 . Stance phase trajectories for each leg are parameterized

by corresponding phase signal, Sit E [0, 1] as

px (t) = - 2Si'(t)) + P 0 ,x
1an(I (3.12)

pi (t) = 6 cos( 2 L ,(t)) + P, (3.13)

dS' dt T

vi (t) = , ' = 6ir sin( pi, (t)) (3.15)

dpi, dt TS 2Lspan

The stance phase trajectory in horizontal direction is linearly designed under the

assumptions that the foot does not slip and the relative velocity of the shoulder/hip

to the foot-end is constant as the desired velocity during Tst. Lspan in stance phase

means a half of the distance traveled of a hip/shoulder during each stance phase. To

guarantee continuity at transition between stance and swing phases, Lspan and P are

identically set to the values of the swing phase trajectory.

The trajectory in vertical direction is determined with J which is the amplitude

of the proposed sinusoidal wave. Again, the vertical trajectory should be interpreted

in the compliance control scheme.

The robot body dynamics is resulting from net horizontal impulse and net verti-

56
cal impulse exerted by four legs. To achieve periodic motion of the center of mass

of the body in vertical direction during running, the following vertical momentum

conservation equation must be satisfied.

mgT,t~id, j, /Tst Fenx(t) dt, (3.16)

contact f

where m is the robot mass, Tsatide and T, is the actual stride period and actual
stance phase period, and Fet are external forces, i.e., GRFs. While maintaining
the sum of vertical impulses, redistribution vertical impulses to each leg is necessary

because the combination of net horizontal impulse and net vertical impulse induces

pitch moment to the robot [75]. Therefore, stance phase trajectories of the front and

the rear legs are designed separately to regulate pitch of the robot.

To directly control the robot body dynamics as described, full state sensory feed-

back is required to exactly compute the desired forces for for legs at each moment

considering the body acceleration. However, there is no doubt that measurement of

the exact body states and proper distribution of the forces are challenging.

Nonlinear coupling between variables such as Tt, T,,side and Fet in Eqn.(3.16)

requires well defined model with minimal model uncertainties. Moreover, to compute

and generate desired force vectors at each foot-end correctly, exact measurement of the

robot orientation, or, reliable state estimation, which requires concurrent regulation

for accumulated error, is necessary .

Due to these underlying difficulties, a stance phase control in this thesis is de-

signed in local coordinate only with feasible sensory feedback such as kinematic data

measured by joint encoders. As we mentioned briefly above, we create virtual com-

pliance in four legs to employ possible advantage of intrinsic stability. This virtual

compliance provides an effective way to solve the complex problem of contact force

control, equilibrium point hypothesis [67]. The hypothesis proposes that a method
animals might use to exert proper force is controlling the equilibrium point of the

virtual compliant system of their limb to have a penetration depth into a contact sur-

57
face. Then, instantaneous difference between the actual position and the equilibrium
position generate the requisite torques without complex inverse dynamics problem.
Trajectory tracking error in the stance phase is mainly induced by GRF due to
the interaction with the ground. The impedance control realized by the leg controller
generates forces at the foot-end in the reactive manner according to the displace-
ment/velocity errors at the foot-end. The integral on the generated force by the
contact legs during the stance period is the impulse as expressed in the right-hand-
side of Eqn.(3.16). Especially at high speed, leg dynamics is comparatively faster
than the body dynamics, and the shoulder/hip height ye is almost stationary while
the robot's leg is moving. When this assumption is effective, the shape of trajectory
error in vertical direction follows that of the desired trajectory, so as the vertical
force exerted to the ground according to the virtual impedance. If an implementable
gains for the impedance control is predefined, the vertical impulse exerted on the
ground during the stance period is possible to be adjusted through the design of the
trajectory [76].

r Impulse
generation

Y' a vertical
force Fy
6.
error
position/velocity
Vutual
e. , e. leg compliance
Desired Trajectory

Figure 3-6: Stance trajectory design with equilibrium-point hypothesis: Along with
vertical axis, the tracking error between the actual foot-end position and the de-
signed foot-end trajectory is converted into the vertical force incorporated with the
leg compliance. The time-integral of the force during the stance period is the vertical
impulse.

Therefore, in order to generate the required impulse for running, the leg force
exerted on the ground can be modulated through adjusting penetration depth of the
virtual reference trajectory into the ground as shown in Fig.3-6 where y, and ye are
the vertical component of a desired trajectory and the position of the ground with
respect to the shoulder/hip joint, respectively.

58
 Series of harmonic functions can create various shape of trajectories with a few

parameters which can generate different impulses. In this thesis, only a sinusoidal

stance phase trajectory is generated with the amplitude, 6, as a control variable for

trot gait. We expect higher order series can be used for asymmetric gait pattern.

3.2.3 Scaling and translation of the designed trajectory

Once the reference trajectory is defined, additional 4 DoFs are permissible to the

operator for adjusting the predesigned reference trajectory in order to adapt to the

environment. The four parameters are translation of the reference point P in hori-

zontal and vertical direction, scaling of the entire trajectory in horizontal direction,

and penetration depth 6. The potential usage of Po,, is inspired by [77]. For trot,
the value is set to zero, but further investigation is required for gallop where more

aggressive acceleration and deceleration per step should be realized.

3.3 Low level individual leg controller

The impedance control is introduced to each leg controller to provide virtual com-

pliance for intrinsic stabilitiy as well as to accomplish motion control in swing phase

and interaction control in stance phase without solving computationally expensive

inverse kinematics nor inverse dynamics. Since the impedance controller provides
stable motion control and effective interaction control, we can avoid complex problem

of switching between stance and swing control which involves exact event detection

problem.

The agility of the mechanical and electrical leg system enables the low level leg

controller to impose a virtual impedance despite of rapidly time-varying equilibrium

point. The bandwidth of the mechanical parts of the robot's leg is measured to be 100

Hz with efforts to minimize mechanical impedance of the robot leg. For the legged

robot, its legs inevitably has ground impact which generate wide frequency range of

disturbance. In order to stabilize such a leg system and create virtual compliance, the

fast overall closed loop is highly required and therefore a FPGA/RT based control

59
architecture is constructed to achieve 250 us loop time. Consequently, a virtual
compliant leg system could be realized as the visco-elastic model of a muscle, without

installing any mechanical spring and damper on the leg.

Virtual leg impedance is created in the polar coordinate as in Fig.3-7 when the
'virtual leg' is defined as the straight line from the should/hip joint to the foot-end

of each leg. By imposing this virtual spring and damper, one can adopt results

from biological examinations, simplified dynamic model studies and previous robotic

experiments with mechanical springs installed on legs, such as proper value for leg

stiffness and angle of attack.

Generalized Coordinate - Polar Coordinate

virtual,
e spring &d r r
92.9
.3

ground

Figure 3-7: Front leg with the impedance control: virtual leg compliance realization
with stiffness and damping for radial/angular directions

Fig.3-8 shows the block diagram for individual leg impedance controller imple-

mented on the real robot. Measuring joint positions with the rotary encoders, leg

forward kinematics and Jacobian return a foot-end position and velocity with re-

spect to a shoulder/hip. The Cartesian position and velocity errors in trajectories are

transformed to the polar coordinate and multiplied with the predefined impedance to

calculate torque commands for each motor as:

T Kp,rer + Kd,rr
pKp,oeo + Kd,oe(

60
where er, ,., eo, and do are radial position error, radial velocity error, angular position
error and angular velocity error and Jpa,. is Jacobian from hip to foot end.

Environment
Faxt

PD le leg dynamics |--

Polar
Gait generator transformation

[. I
[ forward kinematics
from body
to foot-end

Figure 3-8: Block diagram for leg control; implementation of the impedance control
to create virtual leg compliance.

61
62
Chapter 4

Simulation result

The control framework is assessed in the simulator. The simulation results are con-
vincing enough for the author to verify the validity of the proposed control framework
to enable fast and stable quadrupedal locomotion. It is also able to change its gait
pattern from trot to gallop, followed by a few strides in the simulation. A still shot
of the video is shown in Fig 4-1. The ground patch is adopted from the work of [78].
The simulator is constructed using MATLAB as developed in section 4. Equations of
motion are integrated with the ode45 solver, with the RelTol of le-6 and the AbsTol
of le-6.

Figure 4-1: Animation interface constructed using MATLAB. 3-D animation of planar
simulator, ground reaction forces, velocity, pitch variation, and leg phase(stance: 1 /
swing: 0) are shown in the animation.

63
4.1 Trot running with constant speed

The robot achieved 3.4 m/s running with trot gait in the simulation experiment.
A set of properly chosen control parameters for trot running are listed in Table.4.1.
Corresponding leg trajectories for front and back legs are drawin with respect to local
shoulder/hip joint of legs, in Fig.4-2

.1(03Back Hip .10. Front Shoulder

-WO0

4000 P0,11
-7001
-N00 -2D -100 0 100 33 30 - 0 -100 0 100 20D 3W0
x-Ws (11M1)

Figure 4-2: Leg trajectory designed for stable trot running. Both front and back leg
trajectories are presented with respect to local shoulder/hip coordinates.

Terminology Value
T., determined according to Eqn.(3.1) with vd

'a, 0.25 s
93F Bezier control points in Table. 3.2
Lspan,F 170 mm
5
F 36 mm
PO,F (0 mm, -500 mm)
93B Bezier control points in Table. 3.2
Lspan,B 170 mm
JB 10 mm
PO,B (0 mm, -500 mm)
Kp,r 5,000 N/m
Kd,, 100 Ns/m
Kp,o 100 Nm/rad
Kd,o 4 Nms/rad

Table 4.1: Set of control parameters designed for the trot running simulation.

Fig. 4-3 shows the transient response of the simulation with the initial perturba-
tion. The robot successfully reach steady-state running of approximately 3.4 m/s with

64
 1...........

-1*1 1 2 4 a 10 12 14 16 18 20
imei (9)

-4

... . . .. . .....-- --- - -

0 102 30 10 8 18 0
Time (n)

Figure 4-3: Trot running at 3.4 m/s. Velocity (top) and pitch of the front body
(middle) are plotted over time. Height variation over distance traveled is also plot-
ted (bottom). Pitch variation and height variation are damped out as the robot
approaches to the steady state running.

stable trot gait; which was a bit faster than the target velocity, 3.0 m/s. Note that

the steady-state running is achieved without global orientation information, resulting

from the robot's intrinsic self-stability.

Fig. 4-4 shows GRFs solved in the simulator and GRFs estimated from torque

command at each joints. The symmetric trot gait organized by the gait pattern
modulator can be observed from the ground reaction force profiles. Integration of

the vertical forces over one stride, i.e., from t = 13.06 s to t = 13.44 s, matches to

the gravitational impulse over the same period, mgT,,,. Relatively flat portion of
the force plot is due to saturation of knee joint torques for front legs. In addition,

the comparison supports the concept of the proprioceptive feedback; the computed

ground reaction forces and the estimated ground reaction forces are deviated from

each other less than an order of magnitude.

Lastly, snapshots from the trot running of the planar MIT Cheeta robot simulation

is shown in Fig.4-5.

65
 12 122 14 12 1 11 12 134 11 13 14

Figure 4-4: Measured ground reaction forces (top) and estimated ground reaction
forces at each leg (bottom) between time t = 12 s to t = 14 s. Vertical ground
reaction forces (solid lines) and horizontal ground reaction forces (dotted lines) are
plotted together. Solid-dotted horizontal line indicates the weight, mg, of the robot,
where m = 29.443kg in the simulation. Input joint torques are coordinate transformed
to estimate ground reaction forces.

Figure 4-5: Snapshots of trot running at 3.4 in/s in the simulation

4.2 Tr~ot running with acceleration and gait tran-

sition to gallop
The robot achieved accelerating trot running and subsequent gait transition to gallop
in the simulation experiment. The acceleration is done by adjusting desired stance
phase period according to Eqn.(3.1), while keeping the same trajectories for both
front and back legs.
A gait transition algorithm compatible with the gait pattern modulator is also
proposed, which is implemented in the experiment as well. Biological principle of gait
transition in animal is unveiled yet, but it is both hypothesized and observed that
animals have desirable gait pattern for different speed and therefore gait transition
is necessary while accelerating or decelerating. To accomplish the gait transition we
propose to change the phase lag vector AS linearly with time within the predefined
number of strides. As a gait pattern is defined with the set of the phase lag vector

66
AS, smooth and continuous gait transition is achieved by linear transition of AS over

a multiple period of Ttride.

We first tried to change gait from trot to gallop, while maintaining the leg tra-

jectory parameters. It is revealed that the leg trajectories for back leg have to be

designed in a different way. The phase lag vector AS and the leg trajectory control

parameters are both changed linearly in time during gait transition. The modified

equation for stance phase trajectory is as follows. Still, the trajectory is combination

of harmonic waves to create desired shape with a few control parameters [3]. The
back leg trajectory is designed considering that back legs provide strong propulsion

forces to the quadrupeds in general. The additional variable GL represents gallopiza-

tion, such that GL = 1 will modulate trajectory from sinusoidal shape for trot gait

pattern.

Also, note that the reference point is moved backwards by 50 mm. This is specu-

lated by the reasoning in [77] so that it can produce nose-down pitch and propulsion.

Bt~t)r 6 (cos( GL. 2wr

P!, = 2 L" (t)) - - sin( ) ± Po,2 (4.1)
z7Y 2span Pix 2 2span )+P'

Vot(t)
S
= '
dpi,x dt
' = --r (sin r p,(t)) + GL cos( )2 (4.2)
T8± 2Lvpan 2
Lspan

A set of properly chosen control parameters for gallop running is listed in Table.4.2.

Corresponding leg trajectories for front and back legs are drawn with respect to local

shoulder/hip joint of legs, in Fig. 4-6

In Fig.4-7, the stance/swing states of all four legs are plotted between time t = 6

s and t = 9 s. The gait transition happened when all four legs are in the swing phase,

or, the robot has aerial phase. A clear distinction between trot gait and gallop gait

is possible. Also, note the decrease in duty factor, a fraction of stance period in a

stride, during acceleration in trot running.

Fig. 4-8 shows the velocity, pitch variation and center of mass height variation

during steady state trot running, accelerating trot running, and gait transition to

gallop. The pitch of the gallop running gait fluctuates much more than that of trot

67
 D1 *.a3~

-WO

*000 P0.8 __O pox

-00 -230
tO -10'0 0 100 O 130 :A0 -100 0 100 200 300O

Figure 4-6: Leg trajectory designed for gallop running. Both front and back leg
trajectories are presented with respect to local shoulder/hip coordinates.

running gait.

The measured ground reaction forces for four legs are also plotted in Fig. 4-9.

It can be seen that the peak value for ground reaction forces increases as the robot

accelerates. Note that rate of increase is more distinguishable in back legs than front

legs. This can be explained by as robot accelerates, a nose-up pitch is induced,

and back leg forces increase to cancel the pitch moment, which corresponds to the

biological examination explained in [75]. Note the clear difference in coordination

of four legs between trot and gallop. During gallop in the simulator, the forces are

distributed almost equally to four legs.

Lastly, Fig.4-10 shows the simulated gait transition snapshots in one stride. In

the simulation, the gait transition induces momentary instability which is desired to

be stabilized in the gallop gait. Even though the stable gallop gait is designed with

the proposed controller on the simulator, it is left for the future work.

68
 Terminology Value
Tst determined according to Eqn.(3.1) with Vd
tsw 0.25 s
93F Bdzier control points in Table. 3.2
Lapan,F 170 mm
6F 36 mm
PO,F (0 mm, -500 mm)
93 Bezier control points in Table. 3.2
L,,an,B 170 mm
6B 25 mm
PO,B (0 mm, -500 mm)
GL 1
Kp,, 5,000 N/m
Kdr 100 Ns/m
K,,o 100 Nm/rad
Kd,o 4 Nms/rad
Table 4.2: Set of control parameters designed for the gallop running simulation.

Transition
Trot Gallop

5 6.5 7 75 85 9

-I-
8 55 5 9

-i
M

5 8.5 7 7.5 9 a 9
. . . . . .. . . . . - -----

--
I - - ----
!.. ...
8 6.5 75 9
ti(s)

Figure 4-7: The state of all four legs from t = 6 s to t = 9 s. Gait transition happened
at t = 8 s.

69
 S 4 ----- -- - -- - --------
------------------- ---------
C,) n II
6 7 7 a 8.5 9
Time (s)

6 6.5 7 75 8 85 9
lime4s)

~0-
20 21 22 23 24 26 26 27 28 29 30 31
comx (M)

Figure 4-8: Trot acceleration and gait transition to gallop running. Velocity (top) and
pitch of the front body (middle) are plotted over time. Height variation over distance
traveled is also plotted (bottom). From t = 7 s to t 8 s, the robot accelerates and
at t = 8 s, transits its gait from trot to gallop.

300 ----- ---- ----- 4--- -- --- - - -- - ---. -...-. - 4 - - -

200
6 100 -

0 4z

6 6.2 6.4 6.6 6.8 7 72 74 7.6 7.8 8 8.2 8.4 8.6 8.8 9
Tune (s)

Figure 4-9: Measured ground reaction forces between time t = 6 s to t = 9 s. Gait

transition happened at t = 8 s. Solid lines are vertical forces, and dotted lines are
horizontal forces.

trot 1 3 4

5 6 7 8 ga

Figure 4-10: Snapshots of gait transition from trot to gallop by varying gait pattern
parameters AS in linear over time

70
Chapter 5

Experiments on the MIT Cheetah

robot

The experiment demonstrated the performance of the proposed controller imple-

mented on the real robot. During the experiments, the sagittally-constrained MIT

Cheetah robot achieved fast and stable trot running up to the speed about 6 m/s, and

transited to gallop running. The control system designed with National Instrument

devices on the MIT Cheetah robot is largely the work of Sangok Seok. Readers are

referred to Appendix.E for the video of the experiment.

5.1 Experimental setup

The MIT Cheetah robot is constrained on its sagittal plane, running on a modified

commercial treadmill (SOLE TT8) as shown in Fig.5-1. The treadmill motor (3.5 hp

TURDAN DC.PM 90V-25A) is manually controlled by a NI sbRIO-9642 with a motor

driver (Apex Microtechnology MSA260KC) to synchronize its speed with the robot

running speed. The maximum speed of the treadmill is measured to be 6 m/s. A

minimal friction sliding guide is connected to the CoM of the robot through a revolute

joint in order to prevent the roll and yaw motion of the robot. The effective mass

of the guide on the robot, 3 kg, is equivalent to the mass of the four 22.2 V serially-

connected LiPo batteries (465 Whrs) which is in the explosion proof box outside of

71
the robot for safety.

sagittally constrained guide

sikde guide

Cheah

Treadmill

Figure 5-1: Experimental setup: the MIT Cheetah robot runs on the speed-controlled
treadmill. Minimal friction sliding guides constrain the robot on its sagittal plane.

The feasible sensory feedback on the controller are only positional data for each leg
joint given from the 13-bit rotary magnetic encoders installed at dual-coaxial BLDC
motor units of each leg. Other state feedback of the robot are not available in the
experiment. For the 'TD' detection of the reference leg the proposed proprioceptive
feedback is set with the value of the threshold force, 2N, which is obtained in the
preliminary experiment. Four Dynamixel EX-106+ smart motors are used to pre-
vent ab/addution of each leg. Also, the customized motor driver (MCU: Microchip
dsPIC3OF6010) handles the current control of each motor at 20 kHz.

The developed control framework is implemented with 4 kHz overall closed-loop

control speed on the FPGA/RT based compactRIO (cRIO-9082) with the LabVIEW
software. A wireless network (802.11 protocol) with the robot is established and
the controllable parameters can be wirelessly adjusted by the operator through the
LabVIEW front panel. All the data such as measured joint angles and commanded
currents are saved with the sampling rate 250 ps through the RT FIFO memory to
the embedded flash memory during experiment.

72
 0 .......
..
..
............ .......... ..
.. .. .. .. 0

100 ................ i.-Shoulder-Jolni .............. ......

.... 100
~~ ~ ...
. ~~ 11 J.........i...
n..
t...

2DO

3W
.............................

A
2W
3W0
....(
.....:.p j it ..
...H ......... .
.......:
............
4W ...........
1. 5W
400
.
.... ..... ....... ....
.....
..

ODO .............. ........ ..... .....

...... .............. em
-0 -n 0 M 400 O 'N -40 -30 0 -30 400 3

a) Front Lep b) Back Legs

Figure 5-2: Gait trajectories implemented on the experiment with leg workspaces

5.2 Control parameters for the experiment

The predefined gait trajectories for the experiment are shown with respect to the
shoulder joint for the front legs and the hip joint for back legs in Fig.5-2. Shaded
regions are work space of each leg based on each mechanical joint limits. Control
parameters for front and back legs, denoted as F and B respectively, are listed in
Table.5.1.

Terminology Value

t~e Varies according to Eqn.(3.1)

t. 0.25s
SF B6zier control points in Table. 3.2
LIPan,F 170 mm
6F 36 mm
PO,F (0 mm, -500 mm)
ZB Bezier control points in Table. 3.2
L,,an,B 170 mm
6B 0 mm
PO,B (0 mm, -550 mm)
K,, 5,000 N/m
Kd,,. 100 Ns/m

Kp, 100 Nm/rad

Kd,O 4 Nms/rad

Table 5.1: Set of control parameters designed for the experiment.

The swing trajectories for four legs are set to be identical by sharing the same
set of control points for the Bezier curve. During trot-walking at low speed in the

73
 6 6
initial experiment, however, F and B are adjusted separately to stabilize the body

pitch. Due to the coupling effect between the semi-active spine and back legs, the

stance trajectory design of the back legs is conservatively designed just as flat simply
6
by describing B = 0, which is discussed in 6. PO,F and PO,B are designed considering

the stable attitude of the robot in the static posture. The desired swing leg retraction

speed with Lspa=170 mm is determined as 3.4 m/s by the property of the Bezier

curve.

Referring to the dimensionless analysis of 'relative leg stiffness' of running animals

[79], krei,ieg, the virtual radial stiffness of each leg of the robot is set to be 5,000
N/m. The gain values for radial damping (Kd,r), angular stiffness (Kp,O) and angular

damping (Kd,o) are obtained with preliminary tests.

To increase the speed of trot running, the stride frequency is modulated in ac-

cordance with Vd = 2Lpan as described in Chapter.3. , while each leg compliance

Tst
and trajectories are fixed during the experiment for a range of speeds. Biological

observations and previous robotic experiment inspired this algorithm [6]. Farley et

al. observed that leg stiffness(keg) is independent on locomotion speed and exam-

ined that animals such as dogs and horses increase their stride frequency linearly

with speed in their treadmill running experiment. However, the statement itself is

controversial and a further investigation is required.

6 m/s trot running

Footfall Pattern

Figure 5-3: Snapshot for the trot running gait of the MIT Cheetah robot with each
footfall diagram: the existence of the aerial phase is observed between each paired
legs' step in trotting gait

74
 5.3 Experimental results

It is demonstrated that the robot can accelerate up to 6m/s with the trot gait pattern
according to the operator's forward speed command. Stable running in the sense of
the limit cycle is observed over all the range of recorded velocity with the trot footfall
pattern as in Fig.5-3 in the experiment.
At low speed the robot walked with trot gait, but as speed goes up, it started
running, having the aerial phase which all for legs are off the ground.

5.3.1 Estimated locomotion speed

Fig.5-4 shows the estimated forward speed of the MIT Cheetah robot based on the
measured kinematic data during the experiment because the controlled stance phase
period (''s) for each leg is determined according to the desired speed. Comparing the

...
r...ry
s.red.by .. der.(CUI I. . AMT...2-.)
..... .tta.hed
d i ..- t the D C

0 5 10 15 20 25 30 35 40 45 /s
lime (8)

Figure 5-4: Estimated forward velocity (top) of the MIT Cheetah robot: ±0.15 in/s
bounded difference was observed between the estimated speed and the treadmill speed
measured by a rotary encoder (CUL Inc. AMT 102-V) directly attached to the DC mo-
tor main axis. Froude number (middle, blue solid line) and cost of transport (middle,
red solid line) of the robot are plotted together.

observed speed of the treadmill, the difference is smaller than 0.15 in/s. As shown,
the robot accelerated up to approximately 6m/s in 50 seconds. The nondimensional
metrics Froude number (Fr), cost of transport (CoT) shown together in the graph
below. Fr was a record-breaking high number as 7.34 and CoT was a impressively

75
low number as 0.52 to rival legged animals around the robot's maximum estimated
speed, 6 m/s.

5.3.2 Measured leg trajectories

sh~ot~ddr a) Front Right Leg 0hquydr b) Front Left Leg

Desired Taectory 6 Desired 6

-200 . 4 -200

-20
- 0 a
-400 o
-200 20 _x00(mm 0-xs(
m
-200 0 200 ~ m JO-2ws 214
-200-200
6
x-exis (mm) x-xx( s m) x-axs (mm) (mm) (m()
sho
-200 Bkder -200 shoRLder shoulder ShoWldr
0 0 0 - 1
0 - tDesired 0 -Dkwred. -v-t.06 rn/
- Measured -v=.94 un/s -Measured -v--5.94 rn/s
200 --2 -200~"-2DO0 -00
200 x-axis (fvn 20v0~d

-400 "400
-400 -400
-200 0 200 -20 0 200 -200 0 200 -200 0 200
x-axis (mm) x-axis (mm) x-ads (mm) x-axis (mm)
E
c) ack Right Leg hi d) Back Left Leg
0
Desired Tnd-bw/ Desired Tfejector6

-00

-800 -2000
E-200 -40
-200 0 200 ~ m)
-200 ( M)
x-as (mm) x-xx(m (mm) s-axis (mm)s-xs(m ax m)
0 0 0 0
- Desired - v--1.06 un/s - Desired: - M6r/s
- Measured - v-5.94 rn/s -Measured - v5.94 rn/s
200
-. E-0 EE ..

-800 -4001 -400- -400

-200 0 200 -200 0 200 -20 0 200 -20 0 200

s-axws (mm) x-aAdS (mm) s-axds (mm) X-axis (mm)

Figure 5-5: Graph set of generated desired trajectories by the leg trajectory generator
and measured actual gait trajectories during trotting gait from 1 to 6 m/s for a) front
right leg, b) front left leg, c) back right leg, d) back left leg. In each leg graph, top left:
desired trajectories, top right: actual trajectories with respect to speed, bottom left:
comparison between desired/actual trajectories, bottom right: comparison between
actual trajectories at low/high speeds, respectively I

Fig.5-5 shows trajectories of all four legs measured over various speeds during the

experimentation. Trajectories are plotted with respect to each shoulder or hip joint

in the body local coordinate since access to global orientation is precluded in this

experimental setup. Nevertheless, trajectories are consistent with the desired trajec-

76
tory over a large speed variation. The data supports that both the pitch and height

of the frontal body fluctuate a little during the experiment. Slight pitch variation

at the beginning and the end of the swing phase can occur due to effect of the cou-

pling mechanism between hind legs and the spine. The top left subplot shows the

desired trajectory. Again, the shape of the desired trajectory remained the same over

different target speeds. Note the short straight line connecting the end of the swing

phase trajectory and the beginning of the stance phase trajectory. This short-cut

transition in commanded trajectory was due to early TD detection. The top right

subplot expands the actual trajectories with respect to its achieved speeds. As speed

goes up, the vertical width of the trajectory closed-loop decreases such that shoul-

der/hip heights in the stance phase decreases. The shorter stance time requires the

higher vertical force to provide the required vertical impulse for running as explained

in Eqn.(3.16). The bottom left compares the desired trajectory and the measured

trajectory. We can observe the tracking performance of the impedance control for the

swing phase; the trajectory converges within a fourth of duration of the entire swing

phase. On the other hand, for the stance phase, the impedance control exerted high

enough forces to make the robot run, due to the errors in stance phase trajectories.

Stance phase performance will be discussed with force plot. The bottom right plot

compares the measured trajectory at low speed and high speed. The relatively flat

portion of the trajectory indicates ground-contact. It can be concluded that the dis-

tance traveled of the front right shoulder was shorter than we designed as the target

speed increases.

A coupling effect of back legs with the spine through the differential gear was

observed in the measured back leg gait trajectories, which leads to the deviation of

the actual trajectories away from the desired trajectory even during the swing phase

where there was no interaction with the ground. Therefore, in order to improve the

performance of back legs, which are coupled with the spine, understanding of the

back legs-spine system is required.

77
5.3.3 Estimation of generated leg-force based on applied cur-
rents

To investigate interaction between the robot and the environment, forces generated at
foot-end of the robot should be analyzed. However, the experiment was done without
a force plate on/under the treadmill or any force sensor at the foot to measure GRFs.
Nevertheless, the forces exerted at foot-end can be accurately estimated by joint
torques generated from the impedance control law, which has been demonstrated in
the simulation and in [66]. Each joint torques can be estimated by currents measured
at the customized motor driver for the MIT Cheetah robot considering actuation
transmission and pantographic leg design.

S a) FL - Shoulder/Hip - Knee

15.2 15.4 156 15.8 16 162 164 16.6 47. 47.6 48 48.2 48.4 48.6 48.6 49
-- Knee
Db) FR
-Shoulder/Hip
-0

152 154 156 15. is 182 164 166 I47. 47.8 48 482 454 48.6 468 49
- Shoulder/Hip - Knee
5D c) BL
47.6 47.8 48 482 48.4 48.6 48.8 48

5dK
15.2 18.4 156 18 16 15.2 164 16.
d) BR 50 Shoulder/Hip - Knee

tO 15.2 15.4 15.6 18.8 16 18.2 16.4 166 476I 47.8 48 48.2 4&.4 4865 4885 48
Time (a) This()

Figure 5-6: Commanded currents applied to each EM motor for shoulder/hip and
knee joints at leg i E {FL, FR, BR, BL}: The motor driver supplies the commanded
current to the EM motor which proportionally generate torque by the torque constant
Kr.

The currents at each actuator are precisely measured for the entire experiment.
Fig.5-6 shows measured currents provided to the dual-coaxial electric motors at the
robot's shoulder/hip, for 1.5 second around different speeds (4.1 m/s and 5.9 m/s).
These two sections are selected for fair comparison between different speed, because
the robot CoT, an indicator of efficiency, was similar.
Note the difference between currents at low speed and high speed, as well as for

78
front legs and back legs. We can observe contralateral symmetry in the graphs, which

is the characteristic of symmetric gait pattern, trot. Increase in peak current and

decrease in duration of high-current region are observed at higher speed. Currents

measured at back legs are generally much higher than that at front legs. This is
because 1) the back legs are longer than the front legs and therefore higher torques

are required to exert forces at foot end, given similar hip/shoulder joint height, and

2) the currents used in back legs also contribute to rotating spine.

5.3.4 Gait transition

The gait transition method proposed in the Chapter 4 is implemented on the MIT

Cheetah robot. At the maximum speed 6 m/s in the experiment, the robot was com-

manded to change its gait from trot to gallop and it was recorded by the high-speed

camera. The Fig.5-7 shows the phase signals for all legs when the gait pattern mod-

ulator linearly changes its phase lag parameters from trot to gallop in the predefined

period, 6 x '',,.. The desired smooth and continuous gait transition was observed

and the robot could keep running a few strides with the gallop gait pattern. However,

this is the preliminary result and further research will enlighten stable gallop with

smooth gait transition.

A~t~t
-linear pattern parameter transition

Thme~sse) ..- FR -- FL 9-R -BL

Figure 5-7: Sawtooth phase signals generated by phase modulator for gait transition
from trot to gallop. The phase signal converges from one to another, linearly with
time. Gait transition through linear convergence from the trot parameters to the
gallop parameters

79
Figure 5-8: Snapshots of gait transition from trot to gallop recorded by a high speed
video captured at 500 fps depicting 150 ms

80
Chapter 6

Discussion

To the best of the author's knowledge, the MIT Cheetah robot is the fastest and the

most efficient quadruped robot among those which are actuated by electro-magnetic

motors. The FPGA/RT based system architecture enables the fast closed-loop con-

trol including data logging for desired/measured leg joint positions and commanded

currents. Based on these logged data, key features of the experiment are discussed in

this chapter.

6.1 Estimated pitch variation

Effective indices for locomotive stability in the sense of limit-cycle can be body height

and pitch variations. We hypothesizes the stable gait has small repetitive fluctuation

of body height and pitch. These height and pitch fluctuations can be estimated with

measured foot-end trajectory data for four legs as showin in Fig.6-1. The flat regions

of the measured foot-end trajectories is due to interaction with the ground and there-

fore represents the stance phase. The height of each shoulder/hip joint is equivalent

to the distance between shoulder/hip joint and foot-end positions in the stance phase

at instant. Therefore, pitch variation can be deduced from difference between front

and back height data, considering the shoulder-to-hip horizontal distance, 66 cm.

During trotting, the heights of both shoulder/hip joints are observed to remain

around 0.5 m. Maximum estimated height difference between a pair of legs for trot is

81
 shoulder/hip joint
0 I4] _

X FR
T FL
Y BR .2-
100- BL

*0
200- -
U)
%6-
0
E
E -
300

400 aa

50-o

1
Estimated ground region
600'
-300 -200 -100 0 100 200 300
x axis (mm)

Figure 6-1: Actual measured trajectories of four legs with respect to each should/hip
joint, where all the trajectories for four legs are overlapped with respect to each
shoulder/hip joint. The data allow us to approximate the height/pitch variation of
the robot during trot experiment.

roughly around 4 cm at low speed and around 2.5 cm at high speed. The estimated

pitch variation decreases as locomotion speed increases. Maximum estimated pitch

variation is estimated to be 4 deg at low speed and only 2.2 deg at high speed. This

small pitch variation relies on intrinsic stability. Since virtual compliance exerts force

proportional to error, if a large pitch motion occurs, force exerted by front/back legs

naturally adapt to induce pitch moment to cancel out the disturbance.

For the same reason, as speed increases, the height of the body gets lower as

shown in the trajectory vs speed plot of Fig.5-5. In the experiment, each designed

leg trajectory is identically applied to front/back legs over different speed range. At

higher speed, larger vertical forces are required within shorter stance period to satisfy

a law of momentum conservation in Eqn.(3.16). Therefore, by larger exerted forces

on compliant legs the height of body is observed to become slightly lower as speed

goes up.

82
6.2 Fr and COT

Nondimensional metrics such as Froude number (Fr) [74] [31] and cost of transport
(COT) [80] are often used to appropriately compare different sized legged robots and
animals. Fr is size-independent locomotion speed, defined as below

v2
Fr = -- , (6.1)

where v is a characteristic (forward) speed of the locomotion, g is the gravitational

acceleration, h is a characteristic length [3].1. For legged robots and animals, h is
usually chosen as the shoulder/hip height; in case of the MIT Cheetah robot, it is
assumed to be 0.5 m from Fig.6-1.

The maximum Fr of the MIT Cheetah robot is 7.34. In comparison to other

quadrupedal robots, this is a remarkably high value according to the data found
in [31]. Biological data showed that quadrupedal animals normally transit its gait
from walk to trot at Fr = 1 and trot to gallop at around Fr = 2 or 3 [3], and dogs
are observed to use trot gait up to the speed which corresponds to Fr ~ 5.5 [4].
Interestingly, we could achieve much higher Fr using trot gait pattern with MIT
Cheetah robot. We tentatively speculate that the difference between animals' muscle
actuation system and the robot's electrically powered actuation system may enable
the accomplishment, but we withhold the conclusion.

Indeed, we believe the maximum speed of the robot could be even higher because
1) the recorded maximum speed of the robot was limited by the maximum speed of
the treadmill and 2) the robot's maximum power consumption was not reached yet;
At 6 m/s, the consumed power was 1 kW, whereas the maximum power rating of the
robot is assumed to be 3 kW.

COT is used as a mass-independent energy/power consumption, which measures

energy efficiency of the locomotion. COT is defined as

'Some use different notation,

83
 CO=P
C T=W-, (6.2)

where P is power used, W is weight of the robot and v is locomotion speed [80].
Computed COT of the robot is surprisingly low by virtue of its mechanical and
electrical design. For detailed description of enabling technology applied to the robot
to achieve high energy efficiency, the reader is referred to [81].

6.3 Evidences for the MIT Cheetah running

It is estimated that the MIT Cheetah robot ran around 2.2 m/s, by taking the en-
ergetic criterion of running, which was introduced in [82]. Fig.6-2 shows the vertical
length of the virtual front right leg and the estimated speed in the time interval 2 ~ 10
s. The shoulder height can be estimated by the vertical length of the virtual leg in

corrvex stance trjectory: concave stance trajectory

wallMng running

I...........................

oncave sap ec3 convex shape trajectory

shouderheight shoulder height

Figure 6-2: Estimated shoulder height trajectories: convex shape of shoulder height
trajectory as a typical representation of running is observed after 2.2 m/s.

the stance phase. The convex shape of the vertical length matches to the concave
shape of the shoulder height and vice versa. Therefore, the typical concave/convex
shoulder trajectory for walking/running can be drawn before/after 2.2 m/s which is

84
matched to the Froude number 1 in Fig.5-4. Interestingly, this result is aligned with
the biological studies which have shown that quadrupeds start running at Fr = 1 [3].
The estimated duty factor D.etin Fig.6-3 provides another evidence that the MIT
Cheetah robot could run. Duty factor of a foot is defined as a fraction of the stride

Figure 6-3: Estimated duty factor Dest for the front right leg: measured kinematic
data for virtual leg is used to estimate duration of the stance/swing phases.

for which the foot is in contact with the ground [3]. Therefore, duty factor < 0.5
implies the animal/ robot has aerial phase. Even though the definition of running
is controversial and animals showed running with duty factor > 0.5, [4], [83], [82], it
is clear that having aerial phase is sufficient condition for running. Because of the
absence of force plate in the experiment, we estimated duty factor, De.,t, based on
kinematic data of each leg as we did for pitch estimation. From the duty factor plot,
we can say that the robot had the aerial phase after time t = 4 s, where the speed
estimation was 1.4 in/s.

6.4 Estimated force exerted on foot end-effectors

To analyze the performance and characteristic of the trot running result in Cartesian
coordinate, we estimate the interaction forces between the robot and the ground by
measured current in Fig.5-6. The motor driver is devised to compensate the nonlinear
effects such as cogging torque of the BLDC motor in a feed-forward manner and
therefore guarantees linear relation between motor torque and applied current.
Therefore, torque outputs from BLDC motors are calculated based on measured
currents and measured torque constant, K, = 0.27 Nm/A. Assuming frictionless joints
and rigid leg segments and considering actuator transmission and leg configuration
at each moment, forces exerted by the foot-end are estimated as below:

85
 ;
.F . foot
f = (Jart) ' 2 xl, (6.3)

where JCat is Jacobian from shoulder/hip to foot-end. In the preliminary test, the
bandwidth of the leg system was observed around 400 Hz. Therefore, the estimated
forces are filtered with the cut-off frequency of 400 Hz. The vertical force F. and the
horizontal force F, exerted to the ground by the robot are estimated with respect
to the local coordinate at each shoulder/hip joint of the leg, which is described in
Fig.2-1. Nonetheless, we believe this can be an effective approximation for GRFs
with respect to the inertial frame, since the pitch variation was a little.
Around 4.1 nyu -P, Front Eigh -F. Meent Eight -, Frost Left -P. Front Laft

FR FL FR FL...

15.2 15.4 5Is. 62

M58 IM M66

Around 5.9 rns

"- R FL FR FL..

Figure 6-4: Estimation of generated forces in Cartesian coordinate at foot-end of

front legs. Stance phase(high vertical force region) and swing phase(low vertical force
region) can be approximated from the graph. Increase in peak force value is observed
with decrease in duty factor.

6.5 Force analysis on the front legs

Fig.6-4 shows horizontal and vertical force estimations, F., Fy, at both front legs
for each duration. High vertical force regions represent stance phases of each leg.
Characteristics of symmetric gait pattern, trot, are clearly observed in the graphs.
Identical forces produced by left and right side of legs are rhythmically alternating
each other over time, peak value and frequency of vertical forces increase as desired
speed increases. High propulsion force, expressed as negative F, accompany high

86
 Around 4.1 ms FBa*Right - FBakRight -FBakLdt -F.BackLeft

Figure 6-5: Estimation of generated forces in Cartesian coordinate at foot-end of back

legs. Increase in peak force value is observed with decrease in duty factor

vertical forces during stance phase to compensate impact loss. For swing phase while
protracting legs, forces generated are small compared to that for stance phase, due
to leg design and swing phase trajectory design.

Figure 6-6: Desired/actual vertical virtual leg lengths L,

Fig.6-6 shows the vertical motion of front legs with corresponding desired trajec-
tories at 6 in/s. The desired foot-end vertical trajectories penetrate into the ground in
the short stance period, but the actual foot-end vertical trajectories L, cannot track
the desired trajectory due to the ground contact and remains in the estimated ground
region (around 0.5 in), which enables the distinct separation for the stance and air
phases. Therefore, the hypothesis for stationary height of the shoulder joint during
the short stance period is effective such that the height variation is smaller than 1
cm around 5.9 in/s. When the leg contacts with ground, the virtual leg compliance
generates vertical forces according to kinematic error intentionally induced by the
proposed sinusoidal trajectory for the stance phase and the time-integration of the
forces during the stance period is the vertical impulse required for the trot
running.

87
6.6 Contribution of virtual stiffness and virtual
damping to leg compliance

In order to quantify significance of virtual damping in the stance control, Fig.6-7 plots

contributions of each component of virtual leg compliance (virtual stiffness/damping)

to force exerted by the foot-end at two different speeds. Not only rejecting force ripple

in trajectory tracking control in the swing phase, virtual damping generates significant

compensating force in the stance phase at high speed.

T.W.W&

Figure 6-7: Virtual stiffness and damping contribution to create vertical force at the
foot-end

The two distinctive phenomena observed in the figure are following: 1) virtual
damping creates abrupt high forces at 'TD' event to deal with impact, and 2) the
contribution of virtual damping increases as duration of stance phase decreases, or, as
desired speed increases. Both features can be clearly explained by the fact that force
created by virtual damping is reactive to velocity error of the foot-end. A sudden
impediment due to the ground induces large velocity error which creates a peak force
at impact. As target speed increases, desired stance period decreases and therefore
the commanded velocity becomes faster. Naturally, it results in a huge velocity errors
accompanied with large force exertion. Therefore, virtual stiffness is dominant at low
speed but two compliant elements are equally dominant at high speed. By employing
this effect, we could create very stable fast locomotion with a little pitch fluctuation,
without variable compliance or variable trajectory over speed.

88
6.7 Force analysis on the back legs

Forces exerted at foot-end of the back legs are similarly calculated as shown in Fig.6-
5. As mentioned in chapter 5, however, these values are not good approximation of
GRFs because the back leg's actuating forces are spent for generating GRFs as well
as for bending spine through a differential gear. Nonlinear stiffness of spine is not
accurately modeled yet, thus analysis of GRFs exerted by back legs is complicated.
Still, general trends shown in the front legs are observed in the back legs. similar.
It is possible to observe contralateral symmetry between left and right legs, and legs
exert higher peak forces at higher speeds. To improve tracking/compliance control
performance of the back legs, further consideration of the system identification for
spine is required.

89
90
Chapter 7

Concluding Remarks and Future

Research Topics

A novel hierarchical controller for quadrupedal robots is presented in this thesis. The

controller is based on three key strategies: self-stability, equilibrium point hypothesis,

and gait pattern modulator combined with touch down event detection. Virtual com-

pliance of each leg is created by low level impedance controller, providing self-stability

to the robot. Leg trajectories are designed to control the motions and ground reaction

forces, exploiting the virtual compliance in a manner suggested by the equilibrium-

point hypothesis. Phase coupling between four legs is imposed by the high level gait

pattern modulator to generate specific quadrupedal gait patterns. Stride-to-stride

pattern modulation is proposed to adapt to the environment while maintaining the

target gait pattern.

A dynamic simulator is constructed based on the algebraic impact law and the

Coulomb friction model using MATLAB. In the simulator, the proposed controller

demonstrates capability of achieving a stable trot running, and achieving gait tran-

sition from trot to gallop. Simulation experiments conclude that the desired leg tra-

jectories for gallop should be differently designed from that of trot, while a range of

speeds could be achieved without modulating spatial parameters for the leg trajectory

in trot running.

The proposed controller is implemented on the MIT Cheetah robot running on

91
the treadmill, without any force sensor or attitude measurements. The performance
of the robot is remarkable in terms of both speed and energy efficiency, which are
noticeable from corresponding Fr and CoT.
A set of control parameters is presented in the thesis with additional degrees of
freedom permitted to the operator, but basically the configuration of the set remains
flexible. According to the simulation experiments, the construction of this set of
parameters highly influences the performance of the robot. The hierarchical structure
of the controller and the flexibility in designing relationship between control variables
open various possibilities on further investigation.

1. Achieving stable gallop at even higher speed by using the proposed control
framework remains an ongoing work. Proper design of front/back leg trajecto-
ries and proper temporal coordination of four legs should be investigated.

2. The relationship between control parameters will be further investigated to re-

duce the DoF permitted to the operator. We expect to extract and exploit
suitable principles from biological findings so that the operator ends up with
modulating only the target speed and all the control parameters are automati-
cally adjusted.

3. The ground reaction forces can be more accurately modulated by the force
tracking impedance control. The force tracking impedance control has been
developed for manipulators fixed to the base. This approach can be extended
to legged machines if full-state feedback is feasible.

4. The spine will be modeled so that a force compensation for the effect of the spine
becomes possible in a feed-forward manner. We speculate that this flexible spine
will play an important role in galloping by storing and releasing elastic energy.

5. An intelligent algorithm for active posture control using attitude sensory feed-
back will be investigated for stable running on a rough terrain. Active stabiliz-
ing on top of self-stabilizing will show a better performance on irregular terrain
while maintaining simplicity in control.

92
6. A stabilizing controller for outdoor running using attitude sensory feedback

will be investigated. Roll and yaw stabilization have be to integrated with the

current control framework through additional actuation in ab/adduction of each

leg.

93
94
Appendix A

Model Parameters

Characteristics Value

Mbody,F 10.05 kg

lbody,F , 0.2 m

Mscpua,F, Mhumerus,F, Mradius,F 2.63 kg, 0.16 kg, 0.12 kg

1scqpula,Fi lhumerus,F7 lradius,F 0.16 m, 0.22 m, 0.2222m

mspine 0.225 kg

1spine 0.144 m

mbody,B 7.13 kg

lbody,B 0.187 m

mfemur,B, mtibia,B, mmetatarscds,B 2.70 kg, 0.16kg, 0.085 kg

ifemur,B, ltibia,B, lmetatarsals,B 0.245 m, 0.22 m, 0.165m

Table A.1: Model Parameters

Table A.1 shows the characteristic model parameters of the MIT Cheetah robot.
Note that, mass of the proximal segments(mscapulaF, mfemur,B) include motor rotor
mass.

95
96
Appendix B

Back spine coordination

qs,,ine = Cs(qhumerus,BL ± qhumerus,BR) (B.1)

qbackpitch = qfrontpitdz + 4
qspine (B.2)

Where C. = 3 is the effective differential gear ratio.

97
98
Appendix C

Leg compliance

-FR -FL -BR -BL

40

200
---- ------ --- ---.- ---------- ----- -- - 30 ---- - ................... --.- ---
20 - - - - - -- - -----
40
........... ..to l1OONm/nsd - -- --

..
-.-....-.-.--.-.---
..-.... ...-. . -- .-
200
a -..- .... ---.
... ----..
. - - - - - - -.-
-.30

-100

01 - . 0 4 0. 0 . ....- 0 ..1 0 ............

-000 0 0.00 0.1

0.1 -4.-0. -. 1 0.1 02 0.3
.w 8m ("n)
a) virtual leg length b) virtual leg angle

Figure C-1: Measured virtual leg compliance in the polar coordinate (r, 0): a) longi-
tudinal stiffness b) angular stiffness

99
100
Appendix D

InverseKinematics

The derivation of the inverse kinematics for three-segment leg is adopted from [841.
Typos in the reference is modified and rewritten in this work.

Xr

qr,

Figure D-1: Inverse kinematics of three-segment leg. The proximal and the distal
segment are parallel due to pantographic leg design.

101
0 = arctan(-) (D.1)

r = r (D.2)
2 1+ +21,13)
12+l -r 2
q2 = q3= arccos (( + li l -3)3)
212(11 ± 13)

= arctan (li1+ 3 )sin q2 (D.4)

12 - (l1 + 13) cos q2
q1 = 7r + 0 - q2 - (D.5)
= ryv - rxvy (D.6)
rX2 + r2
=vxrx ± vyr, (D.7)

= r3 (D.8)
12(11 +1 3 )sin q2
(l + 13)(12 cos q2 - (l + 13 ))42 (D.9)
l2 + (1i ± 13)2 - 212(11 + 13) cos q2
1 = - 42 - 4 (D.10)

102
Appendix E

Supplementary Video

A video is provided along with this thesis to show the results of theexperiment.

Experimental demonstration 1 http://www.youtube.com/watch?v=IOXhfp3tWk

Experimental demonstration 2 http://www.youtube.com/watch?v=UBHJqnM8RTU

Table E.1: Supplementary videos and urls are listed in the table

103
104
Bibliography

[1] J. E. Bares and W. L. Whittaker, "Configuration of autonomous walkers for

extreme terrain," The InternationalJournalof Robotics Research, vol. 12, no. 6,
pp. 535-559, 1993.

[2] C. D. Remy, "Optimal exploitation of natural dynamics in legged locomotion,"

Ph.D. dissertation, ETH, 2011.

[3] R. M. Alexander, "The gaits of bipedal and quadrupedal animals," The Inter-
national Journal of Robotics Research, vol. 3, no. 2, pp. 49-59, 1984.

[4] L. D. Maes, M. Herbin, R. hacker, V. L. Bels, and A. Abourachid, "Steady

locomotion in dogs: temporal and associated spatial coordination patterns and
the effect of speed," The Journal of Experimental Biology, vol. 211, pp. 138-149,
2008.

[5] S. J. Wickler, D. F. Hoyt, E. A. Cogger, and G. Myers, "The energetics of the

trot-gallop transition," The Journal of ExperimentalBiology, vol. 206, no. 9, pp.
1557-1564, 2003.

[6] C. T. Farley, J. Glasheen, and T. A. McMahon, "Running springs: speed and

animal size," The Journal of Experimental Biology, vol. 185, no. 1, pp. 71-86,
1993.

[7] J. A. Vilensky, J. N. Libii, and A. M. Moore, "Trot-gallop gait transitions in

quadrupeds," Physiology & Behavior, vol. 50, pp. 835-842, 1991.

[8] C. T. Farley and C. R. Taylor, "A mechanical trigger for the trot-gallop transition
in horses," Science, vol. 253, no. 5017, pp. 306-308, 1991.

[9] R. Blickhan, "The spring-mass model for running and hopping," Journal of
Biomechanics, vol. 22, no. 1112, pp. 1217 - 1227, 1989.

[10] A. Ruina, J. E. Bertram, and M. Srinivasan, "A collisional model of the energetic
cost of support work qualitatively explains leg sequencing in walking and gal-
loping, pseudo-elastic leg behavior in running and the walk-to-run transition,"
Journal of Theoretical Biology, vol. 237, no. 2, pp. 170-192, 2005.

105
[11] J. E. Bertram and A. Gutmann, "Motions of the running horse and cheetah
revisited: fundamental mechanics of the transverse and rotary gallop," Journal
of The Royal Society Interface, vol. 6, no. 35, pp. 549-559, 2009.

[12] M. D. Berkemeier, "Modeling the dynamics of quadrupedal running," The In-

ternationalJournal of Robotics Research, vol. 17, no. 9, pp. 971-985, 1998.

[13] J. P. Schmiedeler and K. J. Waldron, "The mechanics of quadrupedal gallop-

ing and the future of legged vehicles," The International Journal of Robotics
Research, vol. 18, no. 12, pp. 1224-1234, 1999.

[14] P. Nanua and K. Waldron, "Energy comparison between trot, bound, and gallop
using a simple model," Journal of biomechanical engineering, vol. 117, no. 4, pp.
466-473, 1995.

[15] H. M. Herr and T. A. McMahon, "A trotting horse model," The International
Journalof Robotics Research, vol. 19, no. 6, pp. 566-581, 2000.

[16] , "A galloping horse model," The InternationalJournalof Robotics Research,

vol. 20, no. 1, pp. 26-37, 2001.

[17] S. Grillner and P. Wallen, "Central pattern generators for locomotion, with spe-
cial reference to vertebrates," Annual review of neuroscience, vol. 8, no. 1, pp.
233-261, 1985.

[18] A. J. Ijspeert, "Central pattern generators for locomotion control in animals and
robots: a review," Neural Networks, vol. 21, no. 4, pp. 642-653, 2008.

[19] M. Golubitsky, I. Stewart, P.-L. Buono, and J. Collins, "A modular network for
legged locomotion," Physica D: Nonlinear Phenomena,vol. 115, no. 1, pp. 56-72,
1998.

[20] S. H. Strogatz, "From kuramoto to crawford: exploring the onset of synchroniza-

tion in populations of coupled oscillators," Physica D: NonlinearPhenomena,vol.
143, no. 1, pp. 1-20, 2000.

[21] W. Wang and J.-J. E. Slotine, "On partial contraction analysis for coupled non-
linear oscillators," Biological cybernetics, vol. 92, no. 1, pp. 38-53, 2005.

[22] Q.-C. Pham and J.-J. Slotine, "Stable concurrent synchronization in dynamic
system networks," Neural Networks, vol. 20, no. 1, pp. 62-77, 2007.

[23] K. Seo, S.-J. Chung, and J.-J. E. Slotine, "Cpg-based control of a turtle-like
underwater vehicle," Autonomous Robots, vol. 28, no. 3, pp. 247-269, 2010.

[24] J. Seipel and P. Holmes, "A simple model for clock-actuated legged locomotion,"
Regular and chaotic dynamics, vol. 12, no. 5, pp. 502-520, 2007.

106
[25] I. Poulakakis, E. Papadopoulos, and M. Buehler, "On the stability of the pas-
sive dynamics of quadrupedal running with a bounding gait," The International
Journal of Robotics Research, vol. 25, pp. 669-687, 2006.

[26] L. R. P. III and D. E. Orin, "Force redistribution in a quadruped running trot,"

IEEE International Conference on Robotics and Automation, April 2007.

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

quadruped robot on irregular terrain based on biological concepts," The Inter-
national Journal of Robotics Research, vol. 22, no. 3-4, pp. 187-202, 2003.

[28] C. Liu, Q. Chen, and J. Zhang, "Coupled van der pol oscillators utilised as
central pattern generators for quadruped locomotion," in Control and Decision
Conference, 2009. CCDC'09. Chinese. IEEE, 2009, pp. 3677-3682.

[29] L. Righetti and A. J. Ijspeert, "Pattern generators with sensory feedback for
the control of quadruped locomotion," in Robotics and Automation, 2008. ICRA
2008. IEEE InternationalConference on. IEEE, 2008, pp. 819-824.

[30] C. Maufroy, H. Kimura, and K. Takase, "Integration of posture and rhythmic

motion controls in quadrupedal dynamic walking using phase modulations based
on leg loading/unloading," Autonomous Robots, vol. 28, no. 3, pp. 331-353, 2010.

[31] A. Sprowits, A. Tuleu, M. Vespignani, M. Ajallooeian, E. Badri, and A. J.

Ijspeert, "Towards dynamic trot gait locomotion: Design, control, and experi-
ments with cheetah-cub, a compliant quadruped robot," The InternationalJour-
nal of Robotics Research, vol. 32, pp. 932-950, 2013.

[321 M. Raibert, Legged robots that balance. MIT press Cambridge, MA, 1986, vol. 3.

[33] M. Railbert, K. Blankespoor, G. Nelson, R. Playter, and B. Team, "Bigdog, the

rough-terrain quadruped robot," in Proceedings of the 17th World Congress The
InternationalFederation of Automatic Control (IFAC), 2008.

[34] K. Yoneda, H. Iiyama, and S. Hirose, "Sky-hook suspension control of a

quadruped walking vehicle," in Robotics and Automation, 1994. Proceedings.,
1994 IEEE InternationalConference on. IEEE, 1994, pp. 999-1004.

[35] J. Estremera and K. J. Waldron, "Thrust control, stabilization and energetics

of a quadruped running robot," The internationalJournalof Robotics Research,
vol. 27, no. 10, pp. 1135-1151, 2008.

[36] S. Coros, A. Karpathy, B. Jones, L. Reveret, and M. van de Panne, "Locomotion

skills for simulated quadrupeds," A CM Trans. Graph., vol. 30, no. 4, p. 59, 2011.

[37] J. Z. Kolter, M. P. Rodgers, and A. Y. Ng, "A control architecture for quadruped
locomotion over rough terrain," in Robotics and Automation (ICRA), 2008 IEEE
InternationalConference on. IEEE, 2008, pp. 811-818.

107
[38] A. Shkolnik, M. Levashov, I. R. Manchester, and R. Tedrake, "Bounding on
rough terrain with the littledog robot," The InternationalJournal of Robotics
Research, vol. 30, no. 2, pp. 192-215, 2011.

[39] M. Kalakrishnan, J. Buchli, P. Pastor, M. Mistry, and S. Schaal, "Learning,

planning, and control for quadruped locomotion over challenging terrain," The
InternationalJournal of Robotics Research, vol. 30, no. 2, pp. 236-258, 2011.

[40] D. Gong, J. Yan, and G. Zuo, "A review of gait optimization based on evolu-
tionary computation," Applied ComputationalIntelligence and Soft Computing,
vol. 2010, 2010.

[41] K. Chae and J. Park, "Trajectory optimization with ga and control for quadruped
robots," Journal of Mechanical Science and Technology, vol. 23, no. 1, pp. 114-
123, 2009.

[42] A. K. Valenzuela and S. Kim, "Optimally scaled hip-force planning: A control

approach for quadrupedal running," in Robotics and Automation (ICRA), 2012
IEEE InternationalConference on. IEEE, 2012, pp. 1901-1907.

[43] C. Gehring, S. Coros, M. Hutter, M. Bloesch, M. A. Hoepflinger, and R. Siegwart,

"Control of dynamic gaits for a quadrupedal robot," in Robotics and Automation
(ICRA), 2013 IEEE International Conference on.

[44] R. Barzel and A. H. Barr, "A modeling system based on dynamic constraints,"
SIGGRAPH Comput. Graph., vol. 22, no. 4, pp. 179-188, 1988.

[45] A. Witkin, M. Gleicher, and W. Welch, "Interactive dynamics," SIGGRAPH

Comput. Graph., vol. 24, no. 2, pp. 11-21, 1990.

[46] D. Baraff, "Coping with friction for non-penetrating rigid body simulation,"
SIGGRAPH Comput. Graph., vol. 25, no. 4, pp. 31-41, 1991.

[47] U. M. Ascher, H. Chin, L. R. Petzold, and S. Reich, "Stabilization of constrained

mechanical systems with daes and invariant manifolds," Journal of Structural
Mechanics, vol. 23, no. 2, pp. 135-157, 1995.

[48] K. Park and J. Chiou, "Stabilization of computational procedures for constrained

dynamical systems," Journal of Guidance, Control, and Dynamics, vol. 11, no. 4,
pp. 365-370, 1988.

[49] W. Blajer, "Elimination of constraint violation and accuracy aspects in numerical

simulation of multibody systems," Multibody System Dynamics, vol. 7, no. 3, pp.
265-284, 2002.

[50] M. Cline and D. Pai, "Post-stabilization for rigid body simulation with contact
and constraints," in Robotics and Automation, 2003. Proceedings. ICRA '03.
IEEE InternationalConference on, vol. 3, 2003, pp. 3744-3751.

108
[51] R. Weinstein, J. Teran, and R. Fedkiw, "Pre-stabilization for rigid body articu-
lation with contact and collision," in A CM SIGGRAPH 2005 Sketches. ACM,
2005, p. 79.

[52] J. Baumgarte, "Stabilization of constraints and integrals of motion in dynami-

cal systems," Computer Methods in Applied Mechanics and Engineering,vol. 1,
no. 1, pp. 1-16, 1972.

[53] S.-T. Lin and J.-N. Huang, "Stabilization of baumgartes method using the runge-
kutta approach," Journal of Mechanical Design, vol. 124, p. 633, 2002.

[54] P. Flores, M. F. Machado, E. Seabra, and M. T. Silva, "A parametric study on the
baumgarte stabilization method for forward dynamics of constrained multibody
systems." ASME, 2011.

[55] A. Chatterjee and A. Ruina, "A new algebraic rigid body collision law based
on impulse space considerations," Journal of Applied Mechanics, vol. 65, pp.
939-951, 1998.

[56] Y. Hurmuzlu and D. B. Marghitu, "Rigid body collisions of planar kinematic

chains with multiple contact points," The InternationalJournal of Robotics Re-
search, vol. 13, no. 1, pp. 82-92, 1994.

[57] M. Payr and C. Glocker, "Oblique frictional impact of a bar: Analysis and
comparison of different impact laws," Nonlinear Dynamics, vol. 41, no. 4, pp.
361-383, 2005.

[58] E. R. Westervelt, J. W. Grizzle, C. Chevallereau, J. H. Choi, and B. Morris,

Feedback control of dynamic bipedal robot locomotion. CRC press Boca Raton,
2007.

[59] M. Posa and R. Tedrake, "Direct trajectory optimization of rigid body dynamical
systems through contact," in Algorithmic Foundations of Robotics X. Springer
Berlin Heidelberg, 2013, vol. 86, pp. 527-542.

[60] R. Ghigliazza, R. Altendorfer, P. Holmes, and D. Koditschek, "A simply sta-

bilized running model," SIAM Journal on Applied Dynamical Systems, vol. 2,
no. 2, pp. 187-218, 2003.

[61] J. Seipel and P. Holmes, "Three-dimensional translational dynamics and stability

of multi-legged runners," The InternationalJournalof Robotics Research, vol. 25,
no. 9, pp. 889-902, 2006.

[62] R. Blickhan, A. Seyfarth, H. Geyer, S. Grimmer, H. Wagner, and M. Gunther,

"Intelligence by mechanics," Philosophical Transactions of the Royal Society A:
Mathematical, Physical and Engineering Sciences, vol. 365, no. 1850, pp. 199-
220, 2007.

109
[63] R. Ringrose, "Self-stabilizing running," in Robotics and Automation, 1997. Pro-
ceedings., 1997 IEEE International Conference on, vol. 1, 1997, pp. 487-493
vol.1.

[64] S. Cotton, I. Olaru, M. Bellman, T. van der Ven, J. Godowski, and J. Pratt,
"Fastrunner: A fast, efficient and robust bipedal robot. concept and planar sim-
ulation," in Robotics and Automation (ICRA), 2012 IEEE International Con-
ference on, 2012, pp. 2358-2364.

[65] N. Hogan, "Impedance control-an approach to manipulation. i-theory. ii-

implementation. iii-applications," ASME Transactions Journalof Dynamic Sys-
tems and Measurement Control B, vol. 107, pp. 1-24, 1985.

[66] S. Seok, A. Wang, D. Otten, and S. Kim, "Actuator design for high force pro-
prioceptive control in fast legged locomotion," in Intelligent Robots and Systems
(IROS), 2011 IEEE/RSJ InternationalConference on. IEEE, 2012, pp. 1970-
1975.

[67] E. Bizzi, N. Hogan, F. A. Mussa-Ivaldi, and S. Giszter, "Does the nervous system
use equilibrium-point control to guide single and multiple joint movements?"
Behavioral and Brain Sciences, vol. 15, pp. 603-613, 1992.

[68] N. Hogan, "Stable execution of contact tasks using impedance control," in

Robotics and Automation. Proceedings. 1987 IEEE InternationalConference on,
vol. 4. IEEE, 1987, pp. 1047-1054.

[69] E. A. Lee and S. A. Seshia, Introduction to embedded systems: a cyber-physical

systems approach. Lee & Seshia, 2011.

[70] M. W. Spong, J. K. Holm, and D. Lee, "Passivity-based control of bipedal lo-

comotion," Robotics & Automation Magazine, IEEE, vol. 14, no. 2, pp. 30-40,
2007.

[71] D. Baraff, "Analytical methods for dynamic simulation of non-penetrating rigid

bodies," SIGGRAPH Comput. Graph., vol. 23, no. 3, pp. 223-232, 1989.

[72] M. T. Mason and Y. Wang, "On the inconsistency of rigid-body frictional pla-
nar mechanics," in Robotics and Automation, 1988. Proceedings., 1988 IEEE
International Conference on, vol. 1, 1988, pp. 524-528.

[73] J. B. Keller, "Impact with friction," Journal of applied Mechanics, vol. 53, no. 1,
pp. 1-5, 1986.

[74] S. Kim, J. E. Clark, and M. R. Cutkosky, "isparawl: Design and turning for high-
speed autonomous open-loop running," The International Journal of Robotics
Research, vol. 25, pp. 903-912, 2006.

110
[75] D. V. Lee, J. E. Bertram, and R. J. Todhunter, "Acceleration and balance in
trotting dogs," The Journal of Experimental Biology, vol. 202, no. 24, pp. 3565-
3573, 1999.

[76] S. Jung, T. Hsia, and R. Bonitz, "Force tracking impedance control of robot
manipulators under unknown environment," Control Systems Technology, IEEE
Transactions on, vol. 12, no. 3, pp. 474-483, 2004.

[77] J. K. Hodgins and M. Raibert, "Adjusting step length for rough terrain loco-
motion," Robotics and Automation, IEEE Transactions on, vol. 7, no. 3, pp.
289-298, 1991.

[78] C. D. Remy, K. Buffinton, and R. Siegwart, "A matlab framework for efficient
gait creation," in Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ In-
ternational Conference on. IEEE, 2011, pp. 190-196.

[79] R. Blickhan and R. Full, "Similarity in multi legged locomotion: Bouncing like
a monopde," Journal of Comparative Physiology A, vol. 173, pp. 509-517, 1993.

[80] V. A. Tucker, "The energetic cost of moving about: Walking and running are
extremely inefficient forms of locomotion. much greater efficiency is achieved by
birds, fishand bicyclists," American Scientist, vol. 63, no. 4, pp. 413-419, 1975.

[81] S. Seok, A. Wang, Chuah, M. Y. (Michael), D. Otten, J. Lang, and S. Kim,

"Design principles for highly efficient quadrupeds and implementation on the mit
cheetah robot," in Robotics and Automation (ICRA), 2013 IEEE International
Conference on, 2013, pp. 3292-3297.

[82] T. A. McMahon, G. Valiant, and E. C. Frederick, "Groucho running," Journal

of Applied Physiology, vol. 62, no. 6, pp. 2326-2337, 1987.

[83] T. McGeer, "Passive dynamic walking," The InternationalJournal of Robotics

Research, vol. 9, no. 2, pp. 62-82, 1990.

[84] C. Maufroy, "Generation and stabilization of quadrupedal dynamic walk using

phase modulations based on leg loading information," Ph.D. dissertation, The
University of Electro-Communications, 2009.

111