Design and Analysis of Series Elasticity in Closed-loop

Actuator Force Control

by
David William Robinson
B.S., Mechanical Engineering
Brigham Young University
April 1994
S.M., Mechanical Engineering
Massachusetts Institute of Technology
June 1996

Submitted to the Department of Mechanical Engineering

in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Mechanical Engineering
at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

June 2000
c Massachusetts Institute of Technology 2000. All rights reserved.

Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Mechanical Engineering
May 11, 2000

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Gill A. Pratt
Associate Professor of Electrical Engineering and Computer Science
Thesis Supervisor

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
David Trumper
Associate Professor of Mechanical Engineering
Thesis Committee Chair

Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ain A. Sonin
Chairman, Department Committee on Graduate Students
 Design and Analysis of Series Elasticity in Closed-loop
Actuator Force Control
by
David William Robinson

Submitted to the Department of Mechanical Engineering

on May 11, 2000, in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy in Mechanical Engineering

Abstract
Series elastic actuators have a spring intentionally placed at the actuator output. Measuring the
spring strain gives an accurate measurement for closed-loop actuator force control. The low spring
stiffness allows for high control gain while maintaining actuator stability. This gives series elastic
actuators many desirable properties including high bandwidth at moderate force amplitudes, low
output impedance, large dynamic range, internal error rejection and tolerance to shock loading.
However, as a consequence of the elasticity, the large force bandwidth capabilities of the actuator
are reduced when operating at power saturation limits.
Series elasticity is examined with three models. First, it is generalized by using a minimal
actuator model. This mathematical model consists of an ideal velocity source actuator, linear spring
and proportional controller. Series elasticity is then demonstrated in two case studies of physical
actuator systems. The first is a linear hydraulic piston with a servo valve and the second is an electric
motor with a geared linear transmission. Both case studies have a linear spring and low complexity
control systems. The case studies are analyzed mathematically and verified with physical hardware.
A series elastic actuator under simple closed-loop control is physically equivalent to a second
order system. This means that an equivalent mass defined by the control system and physical
parameters, is effectively in series with the physical spring connected to the actuator load. Non-
dimensional analysis of the dynamics clarifies important parametric relationships into a few key
dimensionless groups and aids understanding when trying to scale the actuators. The physical
equivalent abstractions and non-dimensional dynamic equations help in the development of guidelines
for choosing a proper spring stiffness given required force, speed and power requirements for the
actuator.
Thesis Supervisor: Gill A. Pratt
Title: Associate Professor of Electrical Engineering and Computer Science

2
For Sarah

3
Acknowledgments
MIT has given me an incredible education the last four years. It has stretched me academically and
personally. I am very grateful to the people that have been with me to share this whole experience.
I want to thank my advisor Professor Gill Pratt for his advice, guidance, help, time, and enthu-
siasm. I appreciate all that I have learned from him as an advisor, teacher, mentor and especially
as a friend.
I also want to thank the other members of my doctoral thesis committee: Professor David
Trumper, Professor Haruhiko Asada, and Dr. J. Kenneth Salisbury. They have all made significant
contributions to this thesis. Their direction and encouragement as a group and as individuals has
been tremendous.
The Leglab is a great place to work because of the people. Dan Paluska, Ben Krupp, Jerry Pratt,
Chris Morse, Andreas Hofmann, Greg Huang, Robert Ringrose, Mike Wessler, Allen Parseghian,
Hugh Herr, Bruce Deffenbaugh, Olaf Bleck, Ari Wilkenfeld, Terri Iuzzolino, Joanna Bryson, Jianjuen
Hu, Chee-Meng Chew, and Peter Dilworth have been great friends and colleagues. They have filled
this experience with wonderful camaraderie in both work and fun.
I especially appreciate my good friends Brandon Rohrer, Sean Warnick, and Rick Nelson with
whom I have shared the MIT engineering Ph.D. road. The time I spent working and in counsel with
these men has been choice. They have each been marvelous examples to me in the ways that they
balance family, service, work, school, and fun.
I thank my parents, brothers, sisters, and their families for their constant support, encourage-
ment, and words of confidence. Mom and Dad are the ones who made it possible to start this journey
and the whole family have always been there to see me through.
I particularly appreciate the support of my two daughters. Hannah’s smiles, hugs and clever wit
have kept me going. MaryAnn’s recent arrival gave me the desire and motivation finish up quickly.
They have both helped me to keep life in its proper perspective.
Finally, I give deep appreciation to my eternal companion Sarah. Our experience together has
been full to overflowing. I am thankful for her faith, consistency, gentleness, kindness, and love.
When you are with Sarah, how can your experience be anything but great! 143, always.

This research was supported in part by the Defense Advanced Research Projects Agency under
contract number N39998-00-C-0656 and the National Science Foundation under contract numbers
IBN-9873478 and IIS-9733740.

4
Contents

1 Introduction 17
1.1 Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2.1 Actuation and Force Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2.2 Series Elastic Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Highlights of thesis results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.1 Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.2 Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.3 Output Impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.3.4 Load Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5 Thesis Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6 Note on thesis data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2 Background and Related Work 29

2.1 Force Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.1 Passive Compliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.1.2 Active Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Applications of Force Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3 Robot Actuators and Active Force Control . . . . . . . . . . . . . . . . . . . . . . . 33
2.3.1 Electro-Magnetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.2 Hydraulic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.3 Pneumatic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.4 Others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Intentionally Compliant Robot Actuators . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.1 Series Elastic Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.2 Other Electro-mechanical Compliance . . . . . . . . . . . . . . . . . . . . . . 38
2.4.3 Hydraulic Compliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3 Linear Series Elastic Actuators 43

3.1 General Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.1 Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1.2 Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1.3 Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.4 System Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 Minimal Linear Model Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.1 General Power Domain Open-Loop Model . . . . . . . . . . . . . . . . . . . . 47
3.2.2 General Closed-Loop Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3 Minimal Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.1 Case 1: Fixed Load – Closed-loop Bandwidth . . . . . . . . . . . . . . . . . . 49
3.3.2 Case 1: Fixed Load – Large force bandwidth . . . . . . . . . . . . . . . . . . 52

5
 3.3.3 Case 2: Forced Load Motion – Output Impedance . . . . . . . . . . . . . . . 54
3.3.4 Impact Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4 Mass Load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4.1 Forces on the load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Dimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.6 General Model Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Hydro-Elastic Case Study 65

4.1 Model Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1.1 Model Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1.2 Power Domain Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1.3 Closed-loop Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1.4 Two input cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.1 Saturation and Large Force Bandwidth . . . . . . . . . . . . . . . . . . . . . 70
4.2.2 Case 1: Closed-loop Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.3 Impedance: Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2.4 Proportional Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 Effect of Load Mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.1 Load Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.2 Load Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 Physical Actuator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4.1 Component Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4.2 Choosing the Spring Constant . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.3 Physical Actuator Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.5 Hydro-Elastic Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5 Electro-Magnetic Series Elastic Case Study 91

5.1 Model Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.1.1 Model Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.1.2 Power Domain Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.1.3 Closed-loop Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.1.4 Two input cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2 Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2.1 Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2.2 Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2.3 Impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2.4 Force Error Rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.3 Effect of Load Mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3.1 Load Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.3.2 Load Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.4 Physical Actuator Prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.4.1 Component Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.4.2 Choosing Sensor Spring Constant . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.4.3 Actuator Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.5 Case Study Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6 Conclusions 115
6.1 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.1.1 Actuator Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.1.2 Springs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.1.3 EM velocity mode control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.1.4 Control system design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

6
List of Figures

1-1 Series elastic actuator. The closed-loop actuator is topologically identical to any
motion actuator with a load sensor and closed-loop feedback controller. The major
difference is that the sensor is very compliant. Low spring stiffness removes gain from
the power domain and and thereby allows for increased controller gain in the signal
domain while still maintaining desired stability margins. . . . . . . . . . . . . . . . 20
1-2 CAD models of the hydro-elastic actuator (left) and EM series elastic actuator (right)
developed and analyized in this thesis. Both actuators have a force output range on
the order of 400–600 lbs. The power output for the hydro-elastic actuator is 1.5 kW
and 430 W for the EM actuator. The minimum resolvable force of both actuators is
approximately 1 lb due to the noise floor in the sensor. This gives the actuators a
dynamic range of the order 500:1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1-3 real Experimental bandwidth of the two prototype actuators. The hydro-elastic
actuator is on the left and the EM actuator is on the right. The magnitude of
oscillation for the actuators is 40 lbs. The bandwidth of both actuators is in the
range of 30-35 Hz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1-4 real Experimental large force bandwidth. The actuators ability to sinusoidally
oscillate at the maximum force, Fsat , at steady state is limited in frequency due to
the spring compliance and motor saturation in force and velocity. The frequency at
which that maximum force capabilities of the actuator begin to fall off is defined as
ωo . ωo = 25 and 8 Hz for the hydraulic and EM prototype actuators respectively. . 23
1-5 sim The EM actuator’s output impedance as a function of load motion input fre-
quency under PD control. The plot comes from the dynamic equation derived in
Chapter 5. The figure has been normalized in frequency by ωo and in magnitude by
the stiffness ks of the physical spring. At low frequency the impedance is small. The
impedance increases with increased frequency in the limit is equal to ks , the physical
spring stiffness. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1-6 Physical equivalent of output impedance. As xl drives at different frequencies, the
impedance of the actuator changes. At low frequency, the impedance looks like an
equivalent mass, meq . At high frequencies, the impedance looks like the physical
elasticity of the spring in the actuator. Even though damping is not shown it is
assumed present to limit uncontrolled oscillations at the natural frequency. . . . . . 24
1-7 Series elastic actuator connected to an inertial load. Ideally the actuator is a perfect
force source (left). However, when moving and inertial load, the actuator has its
physical equivalent impedance. Typically, meq is very small in comparison to the
load mass, ml . Nevertheless, understanding the relative magnitude of meq is important. 25
1-8 real Hydro-elastic actuator forces on an inertial load. The inertial load mass and
the actuator equivalent mass are 18 kg and 20 kg respectively. Since the two are so
close, at low frequency the actuator displays a significant non unity magnitude under
PI control. The response magnitude rises as frequency increases and then drops at
the closed-loop bandwidth of the actuator. . . . . . . . . . . . . . . . . . . . . . . . 25

7
2-1 P3. A commercial walking robot built by Honda Research and Development. It uses
a combination of passive compliance in both its actuators and feet as well as active
force control at the ankles. Reprinted with permission of Honda Motor Co., LTD. . . 31
2-2 Series Elastic Actuator. A spring is intentionally placed in series with an eletro-
magnetic motor, transmission and actuator output. The spring deflection is controlled
thus inferring force control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2-3 COG is a humanoid robot with upper torso, arms and head. There are series elastic
actuators for each of the six degrees of freedom in each arm. Here COG is turning a
crank using a dynamic oscillating controller [76]. . . . . . . . . . . . . . . . . . . . . 39
2-4 Spring Flamingo is a planar bipedal walking robot. It uses series elastic actuators to
actuate its six joints. Its top walking speed is 1.25 m/s [50]. . . . . . . . . . . . . . . 39
2-5 Corndog is a planar running robot representing half of large dog. It uses the electro-
mechanical series elastic actuators developed as part of this thesis to actuate its four
joints [34]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2-6 M2 is a 3D bipedal walking robot. It has 12 degrees of freedom: 3 at each hip, 1 at
each knee and 2 at each ankle. The goal of M2 is to extend work done on Spring
Flamingo to 3 dimensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2-7 Sarcos Dextrous Arm. It is a hydraulic force controlled teleoperated manipulator
that has been used for many applications including production assembly, undersea
manipulation and hazardous material handling. The manipulator actuators have ac-
cumulators on either side of the hydraulic fluid chambers. This gives the actuators
intrinsic compliance. Reprinted by permission of Sarcos. . . . . . . . . . . . . . . . . 41

3-1 Minimal model for a series elastic actuators. There are four parts to a series elastic
actuator: elasticity, control system, motor, and system inputs. The compliance is
a simple linear spring with a stiffness of ks . The force in the spring, fl , is equal
to the spring deflection times the stiffness. The control system consists of a simple
proportional controller with gain K. The motor is modeled as a high impedance
velocity source. The position output of the motor, xm , is the time integral of the
motor velocity. The system inputs or boundary conditions are the desired force, fd
and the load position, xl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3-2 Graphical representation of the minimal motor model. The minimal motor model is
a high impedance velocity source. It neglects inertia and has an instantaneous limit
of force and velocity. It represents idealized models of both a hydraulic piston as well
as an EM motor with a transmission. . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3-3 General motor saturation model. All motors have limits to the instantaneous force
and velocity output capabilities. The line connecting the maximum load force, Fsat ,
and the maximum actuator velocity, vsat , defines the envelope in which the motor can
operate. The slope of the saturation line is Ksat = Fvsatsat
. . . . . . . . . . . . . . . . . 46
3-4 Minimal model and block diagram for a series elastic actuator. The figure on top is a
time domain graphical representation of the the minimal series elastic actuator. The
bottom figure shows a block diagram representation of the actuator. . . . . . . . . . 47
3-5 Fixed load model and block diagram – case 1. (Top) A fixed load constraint defines
the closed-loop bandwidth of the system by isolating the relationship between desired
force and load force. (Bottom) A block diagram of case 1 shows that the model is a
simple first order system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3-6 Zero order model for moving the gain from the power domain to the signal domain. A
series elastic actuator removes gain from the stiffness of the spring. The control system
can then increase control gain while maintaining actuator stability. The stiffness of
the spring is not a limiting factor in the closed-loop stability of the actuator. . . . . 51
3-7 Zero order model for reducing sensitivity to internal position noise. With sensor
stiffness gain and increased control system gain, internal position noise due to the
transmission are greatly reduced. The provides for a very clean force output from the
actuator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

8
3-8 Large force equals large elastic deformation. In order to create a large force, there
must be a large elastic deformation. This requires the motor must move. In order to
oscillate at high force amplitude, the motor must move very quickly. . . . . . . . . . 53
3-9 sim Small force and large force bandwidth. The small force closed-loop bandwith
is unaffected by the large force saturation constraints. However, as the magnitude
of oscillation increases, the motor saturation dominates the large force output capa-
bilities of the actuator. The more compliant the spring, the smaller the large force
bandwidth. This is the key engineering tradeoff for series elastic actuators. . . . . . 55
3-10 Forced load motion model and equivalent block diagram. (Top) The desired force is
fixed constant, Fd = Fo , and the load motion is defined externally. The relationship
between load motion and output force is defined as the output impedance. (Bottom)
The block diagram shows that the feedback system and actuator dynamics are in the
feedback loop. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3-11 Equivalent impedance for the general series elastic actuator. The active impedance
of the actuator can be thought of as an equivalent damper in series with the physical
spring. At low driving frequencies, the actuator appears to be a damper. At high
frequencies, the impedance is the stiffness of the spring. . . . . . . . . . . . . . . . . 57
3-12 General series elastic actuator with load mass moving in free space. In this particular
case, the load motion, xl , is explicitly defined as a function of the loads mass, ml , and
the force in the spring, Fl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3-13 Equivalent model of the general series elastic actuator with load mass. The desired
force is seen through a damper before the spring. A simple proportional controller
does not make the actuator a force device. . . . . . . . . . . . . . . . . . . . . . . . 60
3-14 sim General closed-loop forward transfer function (left) and output impedance (right).
These figures are normalized to the saturation frequency ωo and have κ = 3. . . . . 62
3-15 sim General large force saturation bandwidth. The ability of an actuator to output
its full steady state force level is compromised by the introduction of a very compliant
spring. ωo is the break frequency and is defined by the open loop dynamics of the
actuator. The large force bandwidth is typically less than the controlled bandwidth
of the actuator. Even though the actuator cannot achieve full force output at high
frequency, it can and does operate above ωo . . . . . . . . . . . . . . . . . . . . . . . 63
3-16 sim General force bandwidth profile with load inertia moving in free space. The
variables in the figure are κ = 2 and L = 0.2. Any finite load inertia causes the system
to have damper like qualities at low frequency as well as the bandwidth reduction at
higher frequencies. As L → 0, the system behaves as if it has a fixed load end. . . . 64

4-1 Prototype Actuator CAD model. A 20MPa pressure source is connected to a MOOG
series 30 flow control servo valve (not shown) that directs flow to the two chambers
of the hydraulic cylinder. The piston is coupled to the output through four die
compression springs. The spring compression is measured with a linear potentiometer
which implies force. A closed-loop controller actively moves the piston to maintain a
desired spring deflection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4-2 Hydro-elastic actuator model. The servo valve directs fluid flow into the hydraulic
cylinder which moves the piston and thus compresses the elastic element. The strain
in the spring is measured and used in a proportional-integral feedback control system. 66
4-3 Power domain model for the hydro-elastic actuator. The force output of the actuator
is determined by the compression of the spring. There are two inputs to the power
domain. Q is the fluid flow from the servo valve and is a function of the input current
i. xl is a motion input from the environment. . . . . . . . . . . . . . . . . . . . . . . 67
4-4 sim The power saturation profile for a hydraulic servo valve is a square root relation-
ship between flow and pressure. In order to understand the effects of saturation on
the actuator through linear analysis, a linear saturation relationship is assumed. The
linear profile is a worse case than the square root saturation. . . . . . . . . . . . . . 70

9
4-5 sim This is a bode plot of the dimensionless closed loop system scaled to ωo . Values
for the plot were calculated from real parameters on the device: ωo = 152 rad/sec (25
Hz), κ = 2, I = 0.3 and V = 4.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4-6 sim The impedance of the actuator at low frequency is zero and is ks at high fre-
quency. The plot represents equation 5.14. It uses values ωo = 152 rad/sec (25 Hz),
κ = 2, I = 0.3 and V = 4.4 which were calculated from the prototype actuator. . . . 74
4-7 Rough characterization of hydro-elastic impedance. As shown in the simulation, the
impedance at low frequencies is equivalent to a mass and is equal to the spring constant
of the sensor at high frequencies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4-8 Hydro-elastic actuator model with load inertia moving in free space. Unlike the
previous model where xl is a system input defined by the environment, in this model
the load inertia defines the motion of xl as a function of the force in the spring. The
load mass is part of the power domain and stores and releases kinetic energy as it
moves. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4-9 Control abstraction for the hydro-elastic actuator pushing on an inertial load. Ideally,
the actuator produces the desired force directly on the load. However, the dynamics
of the actuator turn out to be an equivalent mass and spring. The equivalent mass
is defined by the control system and power domain characteristics. The spring is the
actual physical spring in the actuator. . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4-10 Prototype hydro-elastic actuator. The actuator has a piston with 0.2in2 area in one
direction and 0.15in2 in the other. The piston pushes on precompressed die compres-
sion springs. Spring deflection is measured with a linear potentiometer. Although
not shown, fluid is directed into the piston via a servo valve with a 3000psi pressure
source. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4-11 real Open-loop response for the servo, piston, spring and sensor. The solid line
represents the mathematical model and the dashed line is experimental data. The
simulation model uses a full third order servo valve. The model captures all important
features except the light damping on 800 Hz resonance of the real servo valve. . . . 81
4-12 real Closed loop response for the hydro-elastic actuator under PI control. The solid
line represents the mathematical model and the dashed line is experimental data. As
in figure 4-11, the simulation model uses a full third order servo valve. The two models
match very well in both magnitude and phase for the useful frequency range. . . . . 82
4-13 real PI control step response without and with servo valve dither. The integrator
in the controller keeps working on the errors in the system. Given that there is a
small amount of stiction on the piston-cylinder interface, the integrator keeps hunting
around the desired force. I add a small amount of dither on the to the servo valve and
the hunting behavior disappears. Regardless of the dither, notice the fast response
time of the actuator. The rise time is ≈10 msec and the system is settled in 50 msec. 83
4-14 real Large force saturation for the hydro-elastic actuator. This is the response of
the actuator when it is oscillating at its maximum force for different frequencies.
The large force performance begins to drop off near the predicted saturation point
of 25 Hz. After the response starts to die off, the first order bandwidth limitations
of the servo valve begin to come into effect. This effect is not accounted for in
the simulation. Therefore, there is a difference in the experimental and simulated
responses at frequencies above the saturation point. . . . . . . . . . . . . . . . . . . 84
4-15 Risky testing the hydro-elastic actuator. The author is suspended under the hydro-
elastic actuator. The actuator is supporting the load by a commanded force 1 lb above
the downward force of gravity (∼ 250lbs). The author is holding a 2lb weight which
drops the actuator. Releasing the weight causes the actuator to rise. The minimum
resolvable force is 1 lb and is due to the noise floor of the sensor. . . . . . . . . . . . 85
4-16 The hydro-elastic actuator rigidly connected to a hanging mass. The mass is 18kg.
By commanding desired force on the mass oscillating and different frequencies, the
effect of the load mass can be verified. . . . . . . . . . . . . . . . . . . . . . . . . . 85

10
4-17 real Actuator force output with an inertial load. The inertial load mass and and
the actuator equivalent mass are 18 kg and 20 kg respectively. Since the two are so
close, at low frequency the actuator displays a significant non unity magnitude under
PI control as expected. The response magnitude rises as frequency increases and
then drops at the closed-loop bandwidth of the actuator. Proportional control shows
that the damper characteristics are significant for this actuator and inertial load
combination. Using a proportional controller dramatically decreases low frequency
force magnitude far beyond that of the PI controller. . . . . . . . . . . . . . . . . . 86
4-18 real Results of drop testing an 18kg (40lb) mass attached to the actuator. At time
t = −0.2, the actuator is pulling the mass with an upwards force (force up is defined
negative). At about t ≈ −0.15, the mass is released to drop. Even though the desired
force of the actuator is set to zero, the impedance is not perfect and keeps pulling
upwards with a 20lb force. At t = 0, the mass hits the end stop and the force spikes.
The mass actually bounces a few times but the peak impact power is just after t = 0. 87
4-19 Model of the load mass during the drop test. Ideally, the load mass only feels the
force of gravity, F = ml g, since the actuator is set to output zero force (left). However
the actuator still has impedance. The gravity pull only affects the load mass. The
equivalent mass of the actuator does not feel the gravity. Therefore, in order for
the load to drop, the gravity force on the load must be distributed between the two
masses, one real, one apparent (right). In the drop test case, the meq ≈ ml = 20kg
and the gravity force is distributed equally between the two inertias. . . . . . . . . . 88
4-20 real Close-up of drop test impact forces. This shows the first impact of the load
mass. The impulse or change in momentum of the mass, J, is the area under the
force curve. A numerical integration shows that J = 12kg · m s . . . . . . . . . . . . . 88
4-21 real Impact power. The top figure shows the force in the spring on the first impact.
By taking the area under the force curve, we can calculate the impulse, J, of the
impact. J ≈ 12kg ṁ/s. The middle figure shows the energy stored in the spring. It is
F2
E = 12 kls . Finally, the impact power is estimated in the bottom figure by looking at
the rate of energy change P = ∆E ∆t . The series elasticity spreads the impact time out
so that the required power is well within the actuator’s output power capabilities. . 90

5-1 The prototype actuator has a brushless DC motor rigidly connected to a ballscrew
which drives the linear motion. The ballscrew nut is coupled to the output through
four die compression springs. The spring compression is measured with a linear po-
tentiometer. The actuator can output over 1350 N and move at 25 cm/second. . . . 92
5-2 Electro-magnetic series elastic actuator model. The motor mass has a driving force
and viscous friction. The controller drives the motor mass to compress the spring
which gives the desired force output. . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5-3 EM series elastic actuator power domain model. The motor mass has a driving force
and viscous friction. The Fl is the force output through the spring. . . . . . . . . . . 94
5-4 Electro-magnetic series elastic actuator model with a fixed load. This configuration
helps to define the saturation limits of the actuator. . . . . . . . . . . . . . . . . . . 96
5-5 Velocity saturation model for the EM series elastic actuator. Simple velocity satura-
tion limits the ability of the amplifier to produce high forces. This model is slightly
different than the force saturation model described in the previous chapters but has
very similar results. Here Ksat = Fvsat
sat
. . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5-6 sim EM large force bandwidth due to force and velocity saturation. The motor mass
spring resonance ωo defines the large force bandwidth of the actuator. The B group
(generalized damping) defines the shape of the profile around the saturation frequency.
The three lines represent underdamped, critically damped, and overdamped frequency
response. This plot only represents the large force capabilities of the actuator with
the load fixed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

11
5-7 sim Closed-loop bode plot of the EM series elastic actuator with PD control. This
is a bode plot of the closed loop system with κ = 3 and ζc = 1. The zero from the
controller adds the apparent resonance near the controlled natural frequency. . . . . 101
5-8 sim EM output impedance. The impedance of the actuator at small frequencies is
low and is equal to ks at high frequency. In this example κ = 4 and B = 0.1. For
low frequencies, the actuator has an equivalent mass characteristic. The equivalent
mass is, meq = mκm . The higher the controller gain the lower the impedance. As Γ
increases there is an effective reduction in the impedance at the controlled natural
frequency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5-9 Rough characterization if EM impedance. As shown in the simulation, the impedance
at low frequencies is equivalent to a mass and is equal to the spring constant of the
sensor at high frequencies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5-10 sim Inverse parabolic relationship of the force error due to stiction force. As control
gain, Kp , increases non-linear friction forces are dramatically decreased. The two
curves represent the esimate of κ2 . The solid line includes the +1 and the dashed line
does not. In the physical actuator, Kp = 25. As can be seen, at that level of gain
there is very little difference between the two curves. . . . . . . . . . . . . . . . . . 104
5-11 EM series elastic actuator model with load mass moving in free space. The load mass
defines the load motion xl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5-12 sim Force on a load that is free to move. This plot shows that the actuator can
output forces on a load mass even when it is free to move. . . . . . . . . . . . . . . 106
5-13 Control abstraction for the EM series elastic actuator. At frequencies below the
controlled bandwidth, the actuator can be considered a pure force source. Above the
controlled bandwidth, the dynamics of the equivalent mass meq (equation 5.25) and
the spring come into play. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5-14 EM series elastic actuator used in testing. The actuator uses a brushless DC motor
with a direct connection to a 2mm lead ball screw. The ball nut drives the springs
in series with the load. The springs are precompressed die compression springs. A
linear potentiometer measures the deflection of the springs and uses that signal for
feedback to the controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5-15 real Open loop bode plot for electro-magnetic series elastic actuator with fixed end
condition. This is a simple second order lumped mass spring system with a natural
frequency of 7.5 Hz. The dynamic mass is the motor mass as seen through the
transmission and the spring is the physical elasticity. . . . . . . . . . . . . . . . . . . 110
5-16 real Experimental maximum force saturation for the electro-magnetic series elastic
actuator with fixed end condition. The actuator can only generate the maximum
output force up to the ωo , the mass-spring resonance frequency. . . . . . . . . . . . . 111
5-17 real Experimental closed-loop bandwidth for the electro-magnetic series elastic ac-
tuator with fixed end condition. The -3dB point is at 35Hz. The zero in the closed-
loop equation causes the apparent resonance. . . . . . . . . . . . . . . . . . . . . . . 111
5-18 Risky testing. The EM prototype actuator is suspended from the ceiling and loaded
with 300 lbs (me, seat and weights). The displacement of the spring is commanded
so that it compensates for the gravity pull on the load mass. With small variations
in the set point, the actuator moves the load up and down. . . . . . . . . . . . . . . 112
5-19 Fully loaded actuator demonstrating small force sensitivity. The actuator is fully
loaded with 300 lbs and has a force equal to gravity pull on the load mass. Finger
force of approximately 1 lb can get the actuator to move up and down. This shows
that the dynamic range of the actuator is 300:1. . . . . . . . . . . . . . . . . . . . . . 113

12
List of Tables

2.1 Flow and effort for different power domains. Every power domain has variables that
define the flow and effort of the medium. The product of these two variables is the
power transferred in the system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.1 Comparison of sensor stiffness. The top load cell is a standard off the shelf component.
The bottom entry is the stiffness of the springs actually used in the prototype actua-
tors of this thesis. The value shown is the one listed in the catalog. Experimentally,
the springs were 12% more compliant. . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.1 Physical properties of Hydro-elastic prototype actuator. These properties are taken
from component literature and from experimental tests. . . . . . . . . . . . . . . . . 80
4.2 Important physical properties for force saturation. The point at which the actua-
tor begins to show signs of large force saturation is dependent upon these physical
parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5.1 Physical properties of EM prototype actuator. Some of the values are calculated from
motor literature and others are measured. . . . . . . . . . . . . . . . . . . . . . . . 109
5.2 Itemized mass distribution for EM series elastic actuator. This actuator is used in
two dynamically stable robots described in Chapter 2. The table shows that over
50% of the mass is contained in the motor and magnet. In the design of series elastic
actuators, it is important to keep the percentage power producing element mass high
in comparison to the overall actuator mass. . . . . . . . . . . . . . . . . . . . . . . . 113

13
14
Nomenclature
Force
Fd — Desired output force.
Fl — Measured output force in the spring.
Fe — Error between the desired and measured forces.
Fo — Constant desired output force.
Fsat — Maximum actuator output force at steady state.
Flmax — Maximum force amplitude of oscillation any a desired frequency.

Motion
xm — Position of the motor before the spring.
vm — Velocity of the motor before the spring.
xl — Position of the load.
vl — Velocity of the load.
xp — Position of the piston.
vp — Velocity of the piston.
xn — Position noise in the motor.
Vsat — Maximum actuator velocity.

Physical
Parameters
ks — Actuator spring constant.
mm — Lumped motor mass.
ml — Load mass.
bm — Motor friction.
A — Piston Area.
τv — First order time constant of the servo valve.
meq — Physical equivalent actuator inertia.
beq — Physical equivalent actuator damping.

Frequency
ωo — Saturation frequency.
ωc — Controlled natural frequency.
ωl — Load mass (ml ) spring (ks ) resonant frequency.

Gain
K — Controller gain.
Km — General model motor gain.
Kv — Servo valve gain.
Ka — Motor amplifier gain.
Ki — Integral gain.
Kp — Proportional gain.
Kd — Derivative gain.

15
16
Chapter 1

Introduction

1.1 Thesis
Closed-loop control of series elasticity allows high-power density, high-impedance actuators to achieve
low output impedance and good force sensitivity while maintaining good bandwidth. Intentional
compliance in robot actuators is a deviation from traditional force control robot architecture. Nev-
ertheless, series elastic actuators offer many advantages in force control tasks.
This thesis develops quantitative understanding of the effects, requirements and limitations for
series elasticity in closed-loop actuator force control. The goals of the thesis are threefold:

1. Build a general framework for understanding series elastic actuators that is not dependent on
a specific power domain.
2. Demonstrate and verify properties and limitations of series elastic actuators both mathemati-
cally and with physical prototypes.
3. Give guidelines for choosing spring stiffness given force, speed and power requirements for the
actuator.

General Framework
The ideas of series elasticity can be applied generally to high impedance actuation systems. A
hydraulic piston and an electro-magnetic motor with a large reduction transmission are two examples.
This thesis develops quantitative measures for understanding and describing series elastic actuators.
Regardless of an actuator’s power domain, series elastic actuators exhibit good small force closed-
loop bandwidth, reduced large force bandwidth and low output impedance.

Demonstration
This thesis demonstrates series elasticity both mathematically and physically with three models.
First, a minimal actuator model generalizes the main ideas of series elasticity. This mathematical
model consists of an ideal high-impedance velocity source motor, linear spring and proportional
controller. Two case studies of physical actuator systems also demonstrate the principles of series
elasticity. The first is a linear hydraulic piston with a servo valve and the second is an electric motor
with a geared linear transmission. Both case studies have a linear spring and a low complexity
control system, PI and PD respectively. The case studies include both mathematical analysis and
physical hardware verification.

Choosing the Spring

The compliance of the spring in series elasticity is orders of magnitude less than traditional load
sensors. Choosing a proper spring stiffness and designing an actuator that houses the elasticity

17
can prove challenging since the spring travel is so large. In this thesis, I explain the fundamental
tradeoffs of spring stiffness in balancing the requirements for large force bandwidth and low output
impedance. Through the use of non-dimensional models, a designer can also understand the effects
of different spring stiffnesses in order to scale the actuator.

1.2 Motivation
The vast majority of commercial robot applications consist of tasks that control the position1 or tra-
jectory of each robot degree of freedom. Traditional robots can do this with great speed, endurance,
precision and accuracy. Robots have been used in this commercial niche because repetitive tasks
which require precision and accuracy are difficult and tedious for humans.
On the other hand, there are many tasks in which robot competence is inferior to that of biological
counterparts. Despite extensive research, actions such as walking, running, swimming, catching,
grasping and manipulation, considered easy to most able-bodied humans, are difficult for robots.
These tasks all require interacting with the real world.
Since our world is kinematically and inertially constrainted, it is important for a robot to be
able to sense and control the interaction forces between itself and the environment. This is typically
referred to as Force Control.2 There are two aspects to successful force control. The first is to use
an algorithm and sensory information to determine the desired robot joint torques in order to place
the proper forces on the robot’s environment. The second aspect of successful force control is to
generate accurate forces at the robot joints. This thesis deals with the second aspect of force control
and specifically on the actuators that create the forces.
Present-day actuation technology is typically poor at generating and maintaining accurate forces.
Using local closed-loop control of actuator forces can improve the quality of force output by reject-
ing disturbances. Furthermore, it has also been shown that using a compliant load sensor in a
configuration called Series Elasticity, increases an actuator’s ability to generate accurate forces.
The focus of this thesis is to analyze and quantify the effects of load sensor compliance for
closed-loop actuator force control. There are some unifying principles of series elasticity that can
be captured in a general mathematical model. These general principles are further developed with
specific models and physical prototypes for the hydraulic and electro-magnetic (EM) power domains.
This introductory chapter defines and outlines the qualities of good force control actuators. It
also describes the first implementations of Series Elastic Actuators as a deviation from traditional
actuator design. Finally, it gives a summary of the thesis results and contributions.

1.2.1 Actuation and Force Control

Actuation is the process of converting some form of energy into mechanical force and motion. An
actuator is the device or mechanism that accomplishes this energy conversion process. An electro-
magnetic motor with a gear transmission and a hydraulic piston are two examples of such actuators.
With few exceptions, standard robot actuation systems are poor at creating accurate forces in
robot joints. The reasons for this inaccuracy include friction, stick-slip, breakaway forces on seals,
backlash in transmissions, cogging in motors and reflected inertia through a transmission. All of
these real phenomena induce force noise in the actuator. The effects of force noise are minimized
when controlling the position or trajectory of the robot. The mass in both the robot and the
actuators low-pass filter force noise on positional output. This is one reason why robots are so
successful at trajectory control even in the face of force noise. However, for tasks requiring good
force control, the force noise in the actuators can be problematic.
Good force-controlled actuators have several important measures: force bandwidth, mechani-
cal output impedance, dynamic range, power density and force density. The first three measures
are related to one another, and typically an improvement in one yields an improvement overall.

1 The term position is the generalized position/orientation vector.

2 The term force denotes a generalized force/torque vector.

18
An increase in power or force density usually means that the other three characteristics suffer in
performance.
The force bandwidth is a measure of how quickly the actuator can generate commanded forces.
Bandwidth is the highest frequency at which the actuator creates a near one-to-one output force
to desired force.3 The bandwidth must be sufficient to transmit accurate forces through the robot
structure. Most force control tasks only require a few hertz bandwidth as demonstrated by the low
bandwidth of human muscle. However, improving the bandwidth of the actuator above the minimal
operational level can only improve the overall capabilities of the robot.
Mechanical output impedance4 is defined as the minimum amount of force an actuator outputs
for a given load motion. In the simplest terms it can be thought of as the stiffness of the actuator
output. Ideally, the impedance of a force controlled device is zero. Low impedance means that the
actuator appears to the robot as a pure force source with negligible internal dynamics. An actuator
with low impedance is sometimes referred to as being backdriveable.
Dynamic range is the maximum output force divided by the minimum resolvable output force
increment of the actuator. It gives a measure for how sensitive the actuator is to small forces with
respect to its full force output capability. A large dynamic range is desirable because it permits the
actuator to be used in a versatile way in both very sensitive activities and in large force situations.
Humans have very good dynamic range.
Force density and power density are not necessarily operational characteristics. Nevertheless,
these qualities refer to an actuator’s ability to generate force or deliver power to the robot per unit
mass and unit volume of the actuator. Having high force and power density means that a small
package can output a lot of energy. Lightweight actuators prevent over-burdening the robot structure
with excess mass and allow for quick, responsive robot performance.
Electric motors with gear reductions and hydraulics are two present-day actuator systems that
have high force and power density and can also achieve moderate to high bandwidth. Unfortunately,
neither system possesses low impedance or large dynamic range qualities required for good force
control.

1.2.2 Series Elastic Actuators

The force output capabilities of a standard actuator such as an electric motor with gear reduction or
a hydraulic piston can be dramatically improved by using a load sensor as feedback for a closed-loop
control system. This technique has been used for many years on both actuator closed-loop force
control and robot endpoint force control.
Standard load sensors are designed to be as stiff as possible for a given load. This allows the
sensor to transmit power interactions with no internal storage of energy. Unfortunately, a stiff load
cell in series with an already stiff actuator means that the open-loop gain of the actuator is very
high. This in turn implies that to maintain actuator control stability, control gains must be kept
low. Low control gains typically indicate poor overall closed loop performance. Low control gains,
high impedance actuators and stiff sensors make for a system that cannot handle shock loads from
the environment or filter out friction and other transmission non-linearities inherent in real systems.
To overcome these deficiencies, several investigators including Howard [29] and Pratt and Williamson
[46, 75] experimented with a closed-loop control of an electro-magnetic motor and gear reduction in
series with a compliant spring. Information about force in the spring is obtained by measuring the
spring strain or deflection. Pratt and Williamson named this actuator configuration a Series Elastic
Actuator.
Figure 1-1 shows that series elastic actuators are topologically similar to any motion actuator with
a load sensor and closed-loop control system. However, there are two significant differences resulting
from the compliant element. First, there is increased energy storage in the spring that cannot be
neglected. Second, integration of a compliant load-bearing sensor within a compact actuator can
prove challenging. Nevertheless, compact designs have been successfully developed.

3A -3dB magnitude of output force to desired force is typically considered the bandwidth. -3dB is roughly 70%.
4 Throughout this thesis mechanical output impedance is referred to as impedance or output impedance.

19
 Signal Power
Spring x load
+
f desired Σ Control
System
Motor
-

Sensor
f load

Figure 1-1: Series elastic actuator. The closed-loop actuator is topologically identical to any motion
actuator with a load sensor and closed-loop feedback controller. The major difference is that the
sensor is very compliant. Low spring stiffness removes gain from the power domain and and thereby
allows for increased controller gain in the signal domain while still maintaining desired stability
margins.

The primary advantage of series elasticity is that the compliant load bearing sensor lowers the
loop gain of the closed-loop system. The control gain can be proportionally increased to maintain the
overall loop gain of the actuator at desired stability margins. This allows series elastic actuators to
have low output impedance, be tolerant to shock loading and robust to changing loads. In addition,
the increased controller gain greatly reduce the effects of internal stiction and other transmission
non-linearities to give the actuators clean force output.
Many other research and commercial groups have recognized that purposeful compliance in ac-
tuators can be beneficial for force control. Several of these ideas and implementations in both stand-
alone actuators and complete robots are described in Chapter 2. For the most part, in the past, the
advantages of compliance in actuators, including series elastic actuators, have been explained with
intuitive reasoning and not with mathematical demonstration.
In this thesis, I build quantitative understanding for closed-loop control of series elasticity. I also
give performance measures for the actuator. These measures include closed-loop bandwidth, large
force bandwidth, output impedance and inertial loading effects. These performance measures can
be used to help guide actuator design, especially in choosing the stiffness of the elasticity. A brief
summary of thesis highlights are given next.

1.3 Highlights of thesis results

The series elastic actuator shown in figure 1-1 is representative of the general model and the two
physical actuators built for this thesis. Using minimal control the actuators achieve good bandwidth,
low output impedance and large dynamic range.
As a brief summary of the thesis, I focus on results from the physical actuators. I first describe
some of their general capabilities such as maximum force, power output and dynamic range. Then,
by looking at specific cases of the actuators in operation, we can discuss other actuator measures:
bandwidth, output impedance and the effects of an inertial load.

1.3.1 Actuators
Figure 1-2 shows CAD pictures of the two actuators designed, built and tested in this thesis. The
left picture is of the hydro-elastic actuator and the right picture is the EM series elastic actuator.
Both actuators have similar maximum force levels (400–600 lbs). The hydro-elastic actuator has

20
Figure 1-2: CAD models of the hydro-elastic actuator (left) and EM series elastic actuator (right)
developed and analyized in this thesis. Both actuators have a force output range on the order of
400–600 lbs. The power output for the hydro-elastic actuator is 1.5 kW and 430 W for the EM
actuator. The minimum resolvable force of both actuators is approximately 1 lb due to the noise
floor in the sensor. This gives the actuators a dynamic range of the order 500:1.

significantly higher power capabilities (1.5kW) compared with the EM actuator (430 W). The power
density is also very high.
Without series elasticity, both actuators have high output impedance and are poor at open-
loop force control. However, with series elasticity, the actuators are very force sensitive and can
achieve high bandwidth at moderate force amplitudes. The closed-loop actuators have low output
impedance. The lowest resolvable force, which is limited by the noise floor of the sensor, is approx-
imately one pound, giving a dynamic range on the order of 500:1. The low output impedance also
helps to significanly decouple the actuator dynamics from that of an inertial load.

1.3.2 Bandwidth
Bandwidth is a measure of how well the actuator can output a load force or force in the spring, Fl ,
given a desired output force, Fd . The closed-loop bandwidth, or the relationship between Fl and
Fd , changes with frequency and can be written as a transfer function of the form

Fl (s)
Gcl (s) = .
Fd (s)

The proposed bandwidth is measured with the load position of the actuator output held constant.
This is analogous to when the actuator is in rigid contact with an unyielding environment or if the
actuator is connected to an infinite load inertia.
I examine two cases of bandwidth. The first is the small force closed-loop bandwidth and the
second is the large force bandwidth when the actuator is operating at full force and velocity saturation
limits.

Small force bandwidth

The fundamental limits in small force bandwidth for series elastic actuators are independent of the
spring stiffness that senses the force. The control system gain and the spring stiffness gain each
contribute to the loop gain of the closed-loop system. With a lower spring stiffness the controller
gain can be increased a proportional amount to bring the loop gain to the desired phase and gain
stability margins.

21
 10 10

0 0
Fl/Fd (dB)

Fl/Fd (dB)
-10 -10

-20 -20
Experimental Experimental
Simulation Simulation
-30 -30
0 1 2 0 1 2
10 10 10 10 10 10

0 0
Phase (degrees)

Phase (degree)
-50 -50

-100 -100

-150 -150
Experimental Experimental
Simulation Simulation
-200 -200
0 1 2 0 1 2
10 10 10 10 10 10
f (Hz) f (Hz)

Figure 1-3: real Experimental bandwidth of the two prototype actuators. The hydro-elastic
actuator is on the left and the EM actuator is on the right. The magnitude of oscillation for the
actuators is 40 lbs. The bandwidth of both actuators is in the range of 30-35 Hz.

Assuming that the actuator is operating in a non-saturated state, we can measure the bandwidth.
Figure 1-3 shows the closed-loop bode plots for both the hydro-elastic actuator (left) and the EM
series elastic actuator (right). The force amplitude of oscillation for the test is 40 lbs. Each actuator
has a bandwidth in the range of 30-35 Hz.

Large force bandwidth

Elasticity requires that there is a large elastic deformation in order to achieve a large force. High
force amplitude oscillations require that the motor must move very fast. However, all motors have
force and velocity saturation limits. Closed-loop analysis of the series elastic actuator does not apply
when the actuator is operating at saturation levels. Motor saturation places a limit on the large force
bandwidth. Large force bandwidth is defined as the frequency range over which the actuator can
sinusoidally oscillate at a force amplitude equivalent to the maximum output force Fsat at steady
state.
The large force bandwidth or saturation bandwidth, referred to as ωo , is typically less than the
closed-loop bandwidth when using a series elastic force sensor. The actuator can still produce force
above ωo but the maximum force capability decreases with increased frequency.
Figure 1-4 shows the large force bandwidth for the hydro-elastic actuator on the left and the EM
series elastic actuator on the right. As with the small force closed-loop bandwidth, the large force
bandwidth is tested with the load end of the actuator fixed. ωo = 25 and 8 Hz for the hydraulic
and EM actuators respectively. It is important to note there that that the actuators are oscillating
sinusoidally at full saturation force (over 400 lbs)!
The large force bandwidth, ωo , for the hydro-elastic actuator is a function of the spring stiffness,
piston area, maximum flow rate of the valve and the supply pressure. For the EM actuator, the
large force bandwidth is defined at the natural frequency of spring and reflected motor inertia.
Understanding and carefully defining an appropriate ωo for a series elastic actuator is important in
the actuator design process.

22
 5 5

0 0

-5 -5

-10 -10

Fl/Fsat (dB)
F /P A (dB)

-15 -15
s
l

-20 -20

-25 -25

-30 -30
Experimental Experimental
Simulation Simulation
-35 -35
0 1 2 0 1 2
10 10 10 10 10 10
f (Hz) f (Hz)

Figure 1-4: real Experimental large force bandwidth. The actuators ability to sinusoidally oscil-
late at the maximum force, Fsat , at steady state is limited in frequency due to the spring compliance
and motor saturation in force and velocity. The frequency at which that maximum force capabilities
of the actuator begin to fall off is defined as ωo . ωo = 25 and 8 Hz for the hydraulic and EM
prototype actuators respectively.

1.3.3 Output Impedance

As defined previously, output impedance is the amount of force at an actuator output given a moving
load position. This can be written as a transfer function in the form

Fl (s)
Zcl (s) = .
xl (s)

Ideally the actuator is a pure force source and has zero output impedance. If this were the case, the
dynamics of the actuator would be completely decoupled from that of the load motion. However,
since the actuator is a real system, it has some output impedance.
Figure 1-5 shows the calculated the EM actuator’s output impedance as a function of load motion
input frequency under PD control. The plot comes from the dynamic equations derived in the thesis.
The impedance of the hydro-elastic actuator under PI control is almost identical to that of figure 1-5
and is therefore not shown. The figure has been normalized in frequency by ωo and in magnitude by
the stiffness ks of the physical spring. The load motion is assumed to be a high-impedance position
source. At low frequency the impedance is small. The impedance increases with increased frequency
in the limit is equal to ks , the physical spring stiffness.
The passive compliance in the spring and the active control system both contribute to the low
impedance of the series elastic actuator.
The spring’s stiffness can be more than two orders of magnitude lower than a standard load
sensor. Even without closed-loop control, the spring lowers the impedance. This effect is most
noticeable at high frequencies.
The second effect is due to the active control system. With a decreased sensor stiffness, the
control gain is increased. This effect is noticeable at low frequencies. Notice that the impedance is
rising at 40 dB per decade which means that the output impedance has two zeros at the orgin in
the dynamic equation. This implies that at low frequency

Fl = meq s2 Xl .

An equivalent impedance mass, meq , can be defined for both the hydro-elastic actuator and the EM

23
 10

Increased
0
damping

Fload/(ksx load) (dB)

-10

-20

-30
40 dB/dec
-40

-50

-60

-70
-1 0 1
10 10 10
ω/ωo

Figure 1-5: sim The EM actuator’s output impedance as a function of load motion input frequency
under PD control. The plot comes from the dynamic equation derived in Chapter 5. The figure
has been normalized in frequency by ωo and in magnitude by the stiffness ks of the physical spring.
At low frequency the impedance is small. The impedance increases with increased frequency in the
limit is equal to ks , the physical spring stiffness.

actuator. To give a sense of the equivalent mass, I define it for the EM series elastic actuator
mm
meq =
K
where mm is the reflected motor inertia and K is the overall gain in the control system. In a
series elastic actuator, the control system gain, K, is very large. Therefore, the equivalent inertial
impedance, meq , is small relative to the reflected motor mass, mm .
The output impedance is shown graphically in figure 1-6 as a physically equivalent mass in series
with the spring driven by the load motion. Eventhough the impedance of the actuator is not zero,
it is small.

m eq x load
ks

Figure 1-6: Physical equivalent of output impedance. As xl drives at different frequencies, the
impedance of the actuator changes. At low frequency, the impedance looks like an equivalent mass,
meq . At high frequencies, the impedance looks like the physical elasticity of the spring in the
actuator. Even though damping is not shown it is assumed present to limit uncontrolled oscillations
at the natural frequency.

24
 xl m eq xl
ks
Fd ml Fd ml

Figure 1-7: Series elastic actuator connected to an inertial load. Ideally the actuator is a perfect
force source (left). However, when moving and inertial load, the actuator has its physical equiv-
alent impedance. Typically, meq is very small in comparison to the load mass, ml . Nevertheless,
understanding the relative magnitude of meq is important.

0
Fl/Fd (dB)

-5

-10

-15
Experimental
-20 Simulation
-25
0 1 2
10 10 10
Phase(degrees)

50

-50

-100
Experimental
-150
Simulation
0 1 2
10 10 10
f (Hz)

Figure 1-8: real Hydro-elastic actuator forces on an inertial load. The inertial load mass and
the actuator equivalent mass are 18 kg and 20 kg respectively. Since the two are so close, at low
frequency the actuator displays a significant non unity magnitude under PI control. The response
magnitude rises as frequency increases and then drops at the closed-loop bandwidth of the actuator.

1.3.4 Load Motion

Series elastic actuators are meant to be used in real-world robots that contact the environment.
However, there are many times when such a robot is simply moving an inertial load in free space. In
this case, the load motion is defined by the mass of the load, ml , and the force through the actuator.
Ideally, a series elastic actuator can be considered a pure force source within the controlled
bandwidth independent of the load mass. However, as just explained, the impedance of the actuator
is not perfectly zero. The actuator does have minimal dynamics that must be considered when
looking at the force the actuator applies to a freely moving inertial load. Figure 1-7 graphically
depicts the difference between the ideal force source (left) and with the impedance model (right).
When ml  meq then the dynamics of the actuator are negligible to within the actuator band-
width and the actuator can be considered a force source. However, the closed-loop force transfer
function changes at low frequency if the load inertia is close to or less than that of the impedance
equivalent mass, meq . In the limit, if there were no load mass, the actuator has nothing to push on
and produces no force.
Figure 1-8 shows the results of applying hydro-elastic actuator forces to a freely moving inertial

25
load. The inertial load and equivalent mass of the actuator are 18 kg and 20 kg respectively. At
low frequency, the force on the load is a little less than half of the desired force. As force frequency
increases, the response magnitude rises and then drops at the closed-loop bandwidth of the actuator.
In summary, series elastic actuators have many desireable properties as force controllable ac-
tuators. The high bandwidth, low impedance, shock tolerance, internal non-linear error rejection
and large dynamic range are balanced with reduced large force bandwidth and reduced load force
response. Understanding these tradeoffs is the key to successfully applying series elastic actuators
in robot systems.

1.4 Thesis Contributions

The contributions of this thesis are:
1. A presentation of a minimal model for series elastic actuators that aids in understanding the
fundamental principles of the actuators.
2. The design and construction of a linear hydro-elastic actuator that uses a linear piston and
servo valve.
3. The design and construction of a modular electro-mechanical series elastic actuator that uses
a brushless DC motor and linear transmission.
4. The development of linear models for both the hydro-elastic actuator and EM series elastic
actuator.
5. The clarification between small and large force bandwidth and a discussion of bandwidth limits.
6. The development of intuitive physical equivalent models that represent actuator impedance.
7. The development of a linear model for the actuator connected to an inertial load moving in
free space.
8. The demonstration and explanation of impact tolerance in series elastic actuators.
9. The reformulation of the linear models into dimensionless terms that aids in understanding
important relationships in actuator scaling.
10. The explanation of choosing spring stiffness for actuator design and development.

1.5 Thesis Contents

The thesis is organized as follows:
Chapter 1 gives a brief introduction to the background and motivation of the thesis. It also
includes a summary of the thesis objectives and covers the key results from the research.
Chapter 2 presents detailed background material and related work covering general ideas of force
control, present-day actuator technology and actuators which use compliance.
Chapter 3 describes a minimal model of series elastic actuators. This model contains all of the
key characteristics of a compliant load sensor used in closed loop force control.
Chapter 4 applies results of chapter 3 to a case study for a hydraulic piston, a hydro-elastic
actuator. The theoretical model is compared with experimental results.
Chapter 5 applies results of chapter 3 to another case study of an electro-magnetic motor with a
gear reduction. The theoretical model is compared with experimental results.
Chapter 6 gives concluding remarks and outlines future avenues of investigation.

26
1.6 Note on thesis data
This thesis includes data from both mathematical models and hardware. The figure captions indicate
the source of the data. Simulated figures are marked sim and figures from hardware are marked
real . real is also used when both simulated and real data are compared in the same figure.
This notation method is adapted from Williamson [76].

27
28
Chapter 2

Background and Related Work

Robots are very successful at performing tasks that require movement in free space or known en-
vironments under position control. Tasks such as spray painting vehicle exteriors, pick and place
of IC chips, arc welding and spot welding fit very well into this paradigm. While many robotic
applications that use position control have found their way into the commercial sector, robots that
contact the surrounding environment and work within kinematic constraints have for the most part
been limited to laboratory research. In fact, simple tasks such as driving a screw, turning a crank,
assembling toys, and writing on a chalk board are extremely difficult for most robotic manipulators
[5]. Instead of position control, these robots require force control. The robots themselves must be
capable of accurately modulating and controlling actuator torques and forces.
This chapter explores the background of force control and force controllable actuators and pro-
vides explanations of applicable previous work. Some explanations of different techniques and algo-
rithms to achieve robot compliance are outlined. The bulk of the chapter discusses capabilities of
different actuator systems and how each performs with respect to force control. Actuators with in-
tentional compliance, specifically series elastic actuators, are also discussed with examples of related
work in both the electro-magnetic and hydraulic domains.

2.1 Force Control

Force Control is the term used to describe the necessary interaction between a robot and an external
unknown environment. The robot must have the capability of sensing and controlling forces in
addition to knowing where it is in its work space.
The necessity for force control deserves some discussion. Paynter gives a general explanation for
maintaining proper causality when any two physical systems interact energetically [43]. The two
systems’ interaction is defined at the interface by a set of orthogonal variables: flow and effort. In
a mechanical system, flow is the velocity or motion and effort is the force. In any power domain
(such as mechanical, hydraulic, electric, thermal, etc.) there are equivalent flow and effort variables
(Table 2.1). The product of these two variables is the power transferred across the system interface.
Interestingly, one physical system cannot independently control both the flow and the effort at the
point of interface. A system may define and control one of these variables or a relation between the
two variables, but not both.
This explanation of general physical system interaction has direct application for why force
control is necessary when a robot interacts with an unknown environment. Most workpieces and
environments with which a robot will interact are inertially and kinematically constrained. A pallet
has mass. A door swings on a hinge. A wall and floor are vertical and horizontal barriers respectively.
Given these constraints the environment often defines the flow interaction. This leaves the robot
with only the option to modulate the effort. In other words, while the robot can push on and give
energy to the environment, it cannot specify how the environment will respond. If the robot tries to
control the position at a constrained interface, there exists a basic incompatibility between the two
physical systems. Therefore, whenever a robot interacts with an environment or workpiece and is

29
 Power Domain Effort Flow
Mechanical Force (N) Velocity (m/s)
Hydraulic Pressure (Pa) Flow rate (m3 /s)
Electric Voltage (V) Current (A)
Thermal Temperature (K) Entropy rate (J/Ks)

Table 2.1: Flow and effort for different power domains. Every power domain has variables that define
the flow and effort of the medium. The product of these two variables is the power transferred in
the system.

not simply moving around in free space, it must use force control. While it is still important for the
robot to understand its own sense of position, it also needs some capability of being compliant in
order to effectively match the given constraints. As defined in the introduction, the robot must have
low impedance. This allows the causality of the robot-environment interface to maintain a proper
relationship.
It is important to point out that all environments do have some compliance. While most con-
straints have high stiffness, there are other environments which are very compliant such as sod or
carpet. In fact some systems, like water, can exhibit variable compliance depending on the con-
ditions. In these cases it may be necessary to control some linear combination of flow and effort.
Nevertheless the robot must be capable of being compliant. It cannot be limited to high stiffness.
Therefore, responsive force control is essential.
The following three subsections cover different techniques, ideas and algorithms that robot op-
erators use in order to match compliance requirements with the environment. This includes passive
methods as well as active control. The discussion also outlines some robotic applications where force
control can be very useful.

2.1.1 Passive Compliance

A simple method to achieve pseudo force control and low robot output impedance is to have a robot
use conventional position control but incorporate passive compliance. Passive compliance can be
found in the robot structure, joints, and at the end effector. It differs from active force control in that
passive compliance does not use force information for feedback to modulate the control algorithm.
There are two major ideas behind passive compliance: low servo stiffness and passive end effectors.
By introducing passive compliance into the joints, the loop gain for the position control loop
is effectively softened. A similar effect can be achieved by increasing the compliance in the robot
structure (links). With a combination of low servo stiffness and low structural rigidity, the robot
may be able to successfully interact with a work piece or environment simply by using position
control schemes. However, the risk of adding too much compliance to a robot is that the robot may
be sloppy in its task execution and may fail in its overall goal [16].
Successful examples of passive compliance in robot joints are the bipedal walking robots built
by Honda Research and Development called P2 and P3 (figure 2-1 shows two views of P3). P3 is
slightly smaller than P2. Both robots use a compliant damping material as a coupling between the
robot actuators and joints [17, 24]. The coupling protects the robot’s actuators from the shock loads
in each step and decouples the reflected inertia through the large Harmonic drive gear reduction.
To compensate for softened servos, P3 has active force sensors in the ankles which give feedback to
the control system and are critical to its walking stability.
Passive end effectors are another successful strategy for helping robots interact with the envi-
ronment. This is because many robots only touch the environment at the manipulator endpoint. A
robot can use position control to approximately locate the endpoint properly in its workspace. The
endpoint then performs its workpiece interaction task. For example, many end effectors have a soft
rubber covering. This effectively decouples the robot from the workpiece through the compliance.
Biological examples of passive compliance at manipulator end points are human hands and feet.

30
Figure 2-1: P3. A commercial walking robot built by Honda Research and Development. It uses a
combination of passive compliance in both its actuators and feet as well as active force control at
the ankles. Reprinted with permission of Honda Motor Co., LTD.

The soft compliant skin compensates for poor positioning accuracy of gross motions, allows for fine
manipulation and also aids shock reduction in locomotion [3].
Interestingly, soft endpoint covering has also been shown to be an effective tool for active force
control [16]. P3 (figure 2-1) has a compliant foot in addition to compliance in the actuators. It helps
to keep the supporting foot securely in contact with the ground during walking.
Besides a soft rubber covering, pneumatics are also common end effectors that effectively convert
the end of the robot into a force control device. This is because pneumatics are inherently compliant.
By controlling the pressure in a pneumatic piston, the piston converts gas pressure into a mechanical
force. Pneumatics also maintain relatively low output impedance regardless of the pressure or force
at the output.
Another clever passive end-effector is the remote center compliance (RCC) that is set up to au-
tomatically perform a peg-in-hole assembly [14]. The configuration of the RCC creates a compliance
center at a certain point within its structure. Forces acting on the compliance center of the RCC
result in pure translation and torques at that point cause pure rotation. By placing a peg tip right
at the compliance center, it can be passively dropped into a hole without jamming.

2.1.2 Active Control

Passive methods have limits when trying to increase the capability of force controlled robots. Often
passive techniques are task specific, such as RCC being applied to peg-in-hole assembly. In order for
the robot to have the capability to perform more general tasks, there is a need for active control and
task planning. This leads to the need for active feedback of measurements to modulate the control
of forces on either the manipulator endpoint, joint actuators or both.
The vast majority of active force control techniques and algorithms have been developed to
control and monitor the end point forces and torques via a force sensor at the robot tip. Mason
[36] and Whitney [74] give good summaries of formalized compliance control and other active force
control methods.
Even though this thesis deals with single degree of freedom force controlled actuators, I assume
that a high level force control algorithm is used in the coordination and control of the entire robot
system. Therefore, I present an overview of several force control methods and algorithms for multi-
degree of freedom robots. Also included are a few general applications which require force control.
More will be discussed about the physical capabilities of robots and actuators in the next section.
Various active force control methods include explicit force control, stiffness control and damping
control, impedance control, hybrid position/force control and virtual model control.

31
 • Explicit Force Control creates a virtual endpoint force servo. The controller input is the
error between the desired force and actual measured endpoint force. The difficulty in this type
of controller is that there is no positional feedback on the robot which means that there is no
control on the absolute endpoint position. Therefore, this approach is somewhat difficult to
apply to a multi-degree of freedom robot without a higher level control system modulating the
forces.

• Stiffness Control [60] and Damping Control [74] create a virtual six degree of freedom
spring or damper in robot end point coordinates respectively. Both techniques are related
because they measure the endpoint forces as well as the joint positions and velocities in order
to generate a desired force output relating to the virtual spring or damper. Good torque
control on the joints is helpful as the endpoint forces are created with the Jacobian transpose.
However, no inverse kinematic calculations are needed.

• Impedance Control [26] generalizes the ideas of stiffness control and damping control. The
position and velocity of the robot endpoint is commanded to follow some trajectory. In ad-
dition, the control system modulates a manipulator disturbance response which handles the
control of dynamic interactions. The robot disturbance response is modeled after a second
order mass-spring-damper behavior. In other words, the robot endpoint will behave as if it is
a second order system. Therefore the endpoint forces, joint positions and velocities are used
to generate actuator torques. The gain matrices which set the effective stiffness, damping and
inertia of the manipulator endpoint correlate directly with stability and bandwidth criteria for
the robot.

• Hybrid Position/Force Control [51] is a method that combines conventional position con-
trol and force control. The environment dictates natural constraints (such as being in contact
with a surface) where only force control can be used. Similarly, position control is used in
the directions where there are no constraints and the robot can move freely. The trajectory
and path that the robot follow are called artificial constraints. A systematic method for un-
derstanding and defining natural and artificial constraints has been outlined in several places
[5, 37, 67]. Once a task has been defined, actuator torques are then commanded to make the
robot operate within the given constraints. This technique requires a new control system for
each new task.

• Virtual Model Control [47, 49] is a motion control language that allows the control designer
to use simulations of virtual mechanical and/or other components, linear or non-linear, to
create reaction forces which are transmitted through the robot via joint torques. The robot
behaves as if the virtual components are actually attached to the robot. This algorithm can be
formulated into a state machine so that different virtual elements control the robot with each
change of state. Virtual model control is different from impedance control in that it does not
assume a fixed base and allows component connections between any two points on the robot,
not just between base and endpoint. While this control language has been applied to legged
robots that successfully accomplish their task of locomotion [48], there are currently no formal
methods for quantifying stability.

2.2 Applications of Force Control

There are many applications that require real time force control. These applications include: teach-
ing and skill acquisition both for people teaching robots and robots training people, telerobotics,
haptics and biomimetic robots. While each application does not have a specific technique associated
with it, the following are some of the applications to give emphasis to the importance of force control.

• Teaching and Skill Acquisition [7] – Off-line programming has had only partial success
in applications in manufacturing such as grinding and deburring. In these cases, interaction
forces for the task must be acquired or learned. These skills are often difficult to teach to

32
 people, let alone robots. Often it is easiest to train force interaction skill with a teach-by-
showing method. It is then possible to embed that information into one of the active control
methods listed above such as hybrid force/position control or impedance control. In order for
the robot to learn these skills, it must be compliant in the instructor’s hands. This requires
good force control.
Conversely, robots can teach humans. Experiments are being done with robots helping in
the rehabilitation process of upper limb control for people recovering from strokes [2]. These
robots must be force-sensitive and inherently safe.

• Telerobotics [62, 63] places a human in the control loop of a robot. One standard form
places a human on a master manipulator and the robot slave mimics the motion of the master.
To add to the capabilities of the system, a controller will reflect the forces felt by the slave
to the master. While it is not recommended that the slave manipulator have endpoint force
control from the standpoint of stability, local force control on each individual joint is highly
recommended (if not required) for proper teleoperation [41]. Advances in manipulator design
and control algorithms have allowed for the development of a teleoperated micro-surgery robot
[57].

• Haptics [58, 56] is closely related to teleoperation where the only difference is that there is
no slave manipulator. A haptic interface will generate forces at its end point as it follows the
users motion. The desired forces are created in a virtual environment rather than reflecting a
remote real environment. However, the haptic interface gives the illusion that an environment
actually exists. Good force control is critical in haptically representing a virtual environment.

• Biomimetic Robots are those that try to imitate motions and actions of humans and animals.
Robots that walk[48], run[52], catch[28], grasp[58], swim[71], etc. are all currently inferior to
their biological counterparts in both structure and technique. These robots interact with the
surrounding environment and therefore good actuator force control is very beneficial.

2.3 Robot Actuators and Active Force Control

The above algorithms and applications, while using different methods of calculation, generate desired
output forces for each of the joints in response to joint positions, endpoint forces, desired behavior,
and/or the environment. Regardless of the application, algorithm or technique, there is often an
assumption made in the control system that high level desired force equals actual force at the robot
actuator. For the most part, this assumption is poor, particularly for very low forces when desired
force is near zero. There are many reasons why this is the case. Friction, stick-slip, intermittent
environment contact (impacts), transmission dynamics, actuator dynamics, actuator saturation,
reflected inertia and noncolocated control are all limiting factors.
Although actuators are often assumed to be black box force devices, the limiting factors above
make that assumption invalid.1 It is important to understand how well various actuation meth-
ods generate clean and accurate forces or torques. Studies have been conducted that compare and
contrast different actuation domains including internal combustion engines, electro-magnetics, hy-
draulics and pneumatics as well as other developing technologies such as shape memory alloys (NiTi),
electro-active polymers and braided pneumatics (McKibbon muscles). In the comparisons, biological
muscle is used as a benchmark. These studies focused on characterizing and generalizing different
actuators by the shape of their flow-effort relationship [22] and on looking at the power density and
force density of each actuation method [27].
This thesis considers only the traditional actuation methods of hydraulic and electric actua-
tion. The following discussion focuses on how each method performs with respect to force control.
Pneumatics and other actuation methods are mentioned for completeness.

1 Itis interesting to note that many linear and non-linear position control methods and algorithms also assume
accurate force generation from the actuators in order to prove proper operation [13, 65].

33
2.3.1 Electro-Magnetic
The vast majority of robotic actuators in use today consist of some form of electro-magnetic (EM)
motor with a transmission. EM motors by themselves are well understood. They are fairly easy to
model and control because they are linear for the most part. In the simplest model, motor torque
is proportional to current. Transmissions, on the other hand, are not linear. They introduce several
problems discussed below.
The purpose of having a transmission is to increase the force and power density of the actuator.
This allows the EM motor to run at peak efficiency operating conditions (high speed and low torque)
while output power is at the low speed and high torques good for robot operation.
While mathematically a transmission works wonderfully, physically it presents problems for both
position as well as force control. Non-linearity in the form of backlash, increased dynamic mass and
increased output impedance are all a function of the transmission.
Backlash and other non-linearities in the transmission are especially problematic at low forces.
The actuator may hunt back and forth in the dead zone between gear teeth in an attempt to modulate
zero force. It is possible to create and use non-backlash gears to create a stiffer transmission, however
they correspondingly increase drive friction [19].
A gear reduction also increases the equivalent dynamic mass of the motor as seen after the
transmission by N 2 , where N is the reduction. For geared motors, the large reflected inertia leads
to broken gear teeth, damaged ball screws and ruined actuators when large unexpected loads appear
as the robot interacts with an environment. This is because the large reflected inertia needs time to
accelerate in order to store energy. Therefore, the impact energy goes directly into the compliance of
the transmission elements until they yield or break. In addition, the motor inertia, as seen through
the transmission, will typically be on the same order of magnitude as the robot link inertia and be
coupled to the link, significantly altering its dynamic behavior.
EM motors by themselves are typically easy to backdrive and have low impedance. This means
that a small external torque on the motor shaft will cause it to accelerate. However, with a large
transmission, due to the increased real and apparent friction and increased reflected inertia, an EM
motor becomes significantly more difficult to backdrive. With this increased impedance, the actuator
becomes much more of a position causal system rather than a force causal one.
To overcome the difficulties associated with linking an EM motor to a transmission, many novel
approaches have been tried. Three of these approaches are described below which include: direct-
drive, cable transmissions and evoloid gear sets.
In an effort to make a clean, low-friction, low-impedance actuator, Asada et. al. created direct-
drive actuators and robots [6]. Direct-drive eliminates the transmission and connects a DC brushless
motor directly to a robot link. This construction eliminates friction and backlash and creates a very
force sensitive actuator. Advanced torque sensors in the actuators have added to the capabilities of
direct-drive robots [1].
From a modeling and control standpoint, direct-drive robots are simple systems. The motor
(magnet) mass is included with the link inertia. Unfortunately, to compensate for the loss of trans-
mission, direct-drive motors must be large in order to achieve high torques. High torque requires
high current which in turn requires heavy wiring. In weight sensitive and power sensitive applications
such as mobile robots, direct-drive actuators are often unacceptable. Nevertheless, for ground based
robots with few degrees of freedom, direct-drive actuators have been very successful in both force
control applications and position control, particularly in semiconductor manufacturing automation.
In order to improve the force and power density of EM motors without sacrificing force sensitivity,
work has also been done to create stiff, low-friction, clean, light-weight cable transmissions [70].
These transmissions have zero backlash and high power efficency due to high tensioning in the
cables. Cable transmissions have been used on novel robots such as the Whole Arm Manipulator
[59] used in robot catching [28] and digging, the PHANToM [38] used as a three degree of freedom
haptic point interface, Robotuna used to study undersea oscillating foil propulsion [8, 71] and robots
for minimally invasive telerobotic surgery [35, 57]. However, because of the size constraints of pulleys,
cable transmission can only achieve moderate transmission ratios.
Force control with slightly higher ratio gearing has been achieved with very careful actuator

34
design and development in the construction of Artisan, a ten degree-of-freedom manipulator [72].
Instead of using single stage helical gears, Vischer and Khatib used evoloid [9] gear sets for each
joint. Evoloid gears have a much coarser pitch than traditional spur gears. These gears have
higher load capabilities than helical gears but are still backdrivable and have smooth torque transfer
characteristics. This transmission used in conjunction with a novel torque sensor, helped Artisan to
demonstrate compliant behavior.

2.3.2 Hydraulic
In contrast to electro-magnetic motors, hydraulic systems operate best at high force (high pressure)
and low speed (flow rate), which is ideal for robotics. Hydraulic systems also have the highest
power density of modern controllable actuation methods [27]. Systems often operate at 3000 psi and
higher. Hydraulics can hold large loads indefinitely while consuming minimal power. This is some-
thing that would cause electro-magnetic systems to quickly overheat. Hydraulics are particularly
useful in high force and high power density situations such as construction machinery, airplanes
and automobile steering systems. Nevertheless, there are several downsides to hydraulics which
have caused electro-magnetic servos to be favored in the robotics industry. These include potential
messiness (catastrophic failure), system complexity, high impedance, sensitivity to contamination
and non-linearity from a control perspective.
There is an inherent messiness with hydraulics especially when hydraulic oil is used. This is
especially true for catastrophic system failure such as a tree falling onto hydraulic forestry equipment
[40]. Modern system design has alleviated some concerns regarding leakage of hydraulic fluid but
the threat is always there.
Since pressure and return lines must be run to hydraulic actuators, there is increased system
complexity from a plumbing perspective. The hoses and pipes going to and from the actuator are
non-trivial in size. They must be accounted for in robot design and can add significant mass to the
robot structure.
Hydraulic systems also have high impedance. This is a result of the low compressibility of
hydraulic fluid and the fact that servo valves control fluid flow. All fluids are compressible to
some extent, but traditional hydraulic fluids, such as oil or water, can typically be modelled as
incompressible. The only way to backdrive (backflow) a hydraulic system is to create a load pressure
higher than the source pressure. In this case, fluid reverses flow direction because flow in an open
servo valve is a function of pressure diffenence. Hydraulics are usually designed for motion control,
not force control.
Hydraulics are very sensitive to contamination as foreign particles in the working fluid can cause
system failure. In order to limit contamination in the working fluid, hydraulics require tight seals
at the piston cylinder interface. Tight seals increase friction and stiction in the system.
The final downside to hydraulic systems is that they are highly non-linear from a control perspec-
tive. Null bias, null shift, hysteresis, threshold, internal leakage, square root flow-load characteristics
and unequal piston areas are all non-linear effects in hydraulics. Depending upon the level of preci-
sion required by the robot, some or all of these effects need to be accounted for.
The advantages of hydraulics often outweigh the disadvantages. Therefore, hydraulics are used
quite frequently in position controlling robots and heavy equipment. Hydraulic actuators typically
consist of a pressure source and a flow control valve (i.e., spool valve). A small signal current deflects
the spool valve which allows the high pressure fluid to flow. In the simplest linear model, the fluid
flow rate through the valve is proportional to the small input current. The valve directs high pressure
fluid into one of two chambers which drives a piston thus converting fluid flow into a mechanical
motion. Thus, in this configuration hydraulics are very good at position control.
It may seem that it would be possible to modulate the piston chamber pressure in order to get
good force control. Unfortunately, the pressure in the chamber is not a good representation of the
force at the actuator output. A few reasons for this include friction and stiction in the piston and
seals, supply pressure variations and non-linear flow characteristics, and high output impedance.
The friction and non-linearity create force noise on the actuator output. High impedance comes
from difficulty in back-flowing through a servo valve as well as the fact that hydraulic fluid has great

35
inertia. In general, force control of hydraulic actuators is a difficult problem [12].
There have been attempts to overcome some of the difficulties associated with hydraulics for use
in robotic contact tasks. These include mechanically decreasing the seal tolerance of the piston and
implementing advanced control algorithms.
In order to reduce the sliding friction and stiction, the seal tolerance can be lifted. However,
that has a direct effect on system efficiency, not to mention the problems of keeping the hydraulic
fluid and potentially the work environment clean. In Raibert’s hopping robots, the legs had two sets
of loose seals [52]. The first set allowed leakage flow from the pressure chambers. The second seals
simply scavenged the leakage. This configuration helped the piston to slide very smoothly. Even so,
using pressure as an estimate of actuator output force proved inaccurate.
Given the natural non-linearity of hydraulic systems, Alleyne created a non-linear Lyapunov
based hydraulic piston force controller [4]. This controller also includes adaptive behavior for un-
known or varying parameters since many parameters in the hydraulic system are either difficult to
measure off-line or are time varying (with temperature).
Hydraulic impedance control has been implemented on an industrial hydraulic robot. The al-
gorithm for control is a specialized one which recognizes that many robot actuators are position
causal and is called position-based end-point impedance control [44]. While the implementation was
somewhat successful, limitations in the positioning accuracy and the bandwidth of actuators caused
the robot to exhibit a hunting or chatter behavior when in contact with a semi-stiff environment
[23].
It is interesting to note that the hopping robots and the Lyapunov controller also included springs
in their actuators. This effect of the compliance will be discussed in the section 2.4.

2.3.3 Pneumatic
Pneumatics typically consist of a piston and a valve connected to a gaseous pressure source. Pressure
in a piston chamber converts the piston into a force compliant actuator. Pneumatic actuators have
many features similar to hydraulics but with significant differences resulting from the operating fluid.
Inherent compliance from the compressibility of gas make pneumatics useful as end effectors. Clever
design configurations and control schemes have demonstrated that it is possible to achieve moderate
positional bandwidths of 21-35 Hz [27].
There are several problems with pneumatics that limit applicability to robotics including safety,
thermodynamic effects and potential resonance. First, unlike hydraulics, pneumatic systems run
at relatively low pressures (100 psi) for safety reasons. They can operate at higher levels but
compressed air stores large amounts of energy. If there is a rupture in the line, pneumatics become
very dangerous. Thermodynamic effects of compressing and expanding air can heat and cool the
system dramatically. Valves and seals can be designed to operate within temperature variations;
however, it is usually not recommended. Finally, the natural compliance of pneumatics can also
resonate with robot link inertias. Clever damping schemes are often required to maintain stability.
A variation on the standard configuration above is an inflatable elastic tube covered by a flexible
braided mesh typically called McKibben muscles [27]. When pressurized, the elastic tube inside
expands but is constrained by the mesh. The flexible mesh shortens or contracts like a muscle due
to the expanding tube. It has been found that McKibben muscles can exhibit passive behavior very
similar to biological muscle [33] since it has both series and parallel elasticity.

2.3.4 Others
There are many emerging actuator technologies with potential for use in macro scale robots such
as shape memory alloy (NiTi), electro-active polymers, polymer gels, piezoelectric and dielectric
elastomers [10, 27, 45]. Some of these systems demonstrate moderate force density and look very
promising. However, there are some technical hurdles that must still be overcome. These include
long actuation time, small scale, low power density and low efficiency.
Several of the systems have very long actuation time constants. This is the case for polymer
gels as well as shape memory alloy. Specifically for shape memory alloy, actuation time is thermally

36
limited. The elongation time is fairly short because high current can be pulsed through the actuator.
However, in order for shape memory alloy to return to its initial shape, the latent heat in the actuator
must be removed which can take a long time.
For the most part, electro-active polymers, piezoelectric actuators and dielectric elastomers have
current working prototype systems on a very small scale. This is ideal for actuation of micro robots.
In order to work on macro robots, however, work must be done to scale the systems in size.
The final difficulty with many of these actuators is that they have low output power density
and are very inefficient. The power density is orders of magnitude lower than electro-magnetic
motors and hydraulics. Also, the input power to these systems is still quite high, which gives a poor
efficiency rating. One would hope that future development of these systems will make them more
competitive with current macro scale robot actuators.

2.4 Intentionally Compliant Robot Actuators

Each actuation method described above has strengths and weaknesses with respect to force control,
force density and power density. The predominant theme for the actuators is that those with high
force and high power density typically have high impedance. Therefore, as is, these actuators are
difficult to use in force control situations regardless of sensory feedback information.
In an attempt to overcome this impasse, it is possible to decouple the dynamics of the actuator
from that of the robot by placing a compliant element between the two. This gives passive compliance
to the actuators. In addition, by measuring the deflection of the spring, an estimate of force in the
joint is obtained and can be used for active feedback control.
Introducing compliance into the drive system is contrary to conventional machine design wisdom.
Traditionally, machines and drive systems are built to be as stiff as possible to increase bandwidth.
However, it is interesting to note that the actuator benchmark for force control applications, biolog-
ical muscle, is connected in series to a link output through an elastic tendon [3]. There are many
reasons why this is thought to be the case. Ideas include energy storage and increased efficiency,
stability when contacting environments and filtering shock loads to the body. Even though muscles
do not have many of the problems inherent to EM motors and hydraulics, elastic elements can be
useful to actuation regardless of whether the system is biological or aritficial.
Active measurement of a compliant element in an actuator was first applied to electro-magnetic
motors with a transmission and called Series Elastic Actuators. This section presents a history of
these actuators as well as other systems that use active compliance.

2.4.1 Series Elastic Actuators

Howard [29] built an actuator with a spring in series with an EM motor and transmission. He
then measured the relative spring deflection by taking the difference between the motor rotation
and output shaft rotation. By controlling this deflection he essentially was controlling the actuator
output torque by Hooke’s Law:
F = kx. (2.1)
Unfortunately, use of a differential measurement between motor and output shaft misses the actual
strain in the spring due to transmission compliance and is also noise prone. Nevertheless, he was
able to show that by using a smaller motor and larger transmission ratio, the output force and power
density increased while maintaining good control over force.
Pratt and Williamson [46, 75] also used elasticity and created an actuator similar to that
of Howard’s. They showed that to get a better force measurement for closed-loop control, the
strain/deflection of the spring should be measured directly and not as the difference of the motor
and output shaft. This actuator configuration is called a Series Elastic Actuator (figure 2-2). As in
Howard’s actuator, the series elastic actuator mechanically filters friction and backlash and other
non idealities in high ratio transmissions drives. It also filters shock loading from the environment.
In addition to initial prototypes, versions of series elastic actuators have been demonstrated in
a few robots.

37
 Signal Power
Spring x load
+
Control EM Trans-
f desired Σ System Motor mission
-

f load
Sensor

Figure 2-2: Series Elastic Actuator. A spring is intentionally placed in series with an eletro-magnetic
motor, transmission and actuator output. The spring deflection is controlled thus inferring force
control.

• COG (figure 2-3) [11] is a humanoid robot with upper torso and head. Series elastic actuators
are used in its arms, which have 6 degrees of freedom. Using the natural dynamics of the arms,
low noise force controllable actuators and dynamic oscillators for control, COG was able to
do many tasks previously thought to be difficult for robots such as hammering and turning a
crank [76].
• Spring Flamingo (figure 2-5) is a planar bipedal walking robot with series elastic actuators
driving 6 degrees of freedom (three in each leg for hip, knee and ankle) [48]. It can walk at
1.25 m/s on flat terrain. It has also walked “blindly” over terrain of uphill and downhill 15
degree slopes.
• Corndog (figure 2-4) is a planar running robot with one fore and one aft leg [34]. It represents
half of a large dog. The actuators used in Corndog are the electro-mechanical series elastic
actuators developed as part of this thesis. These actuators are described in chapter 5.
• M2 (figure 2-6) is a 3 dimensional bipedal walking robot. It has 12 degrees of freedom; 3 at
each hip, 1 at each knee and 2 at each ankle. M2 is an extension of the work started with
Spring Flamingo. M2 also uses the actuators characterized in this thesis described in chapter
5 [54]. At the time of writing this thesis, the robot could stand on its own.
Series elastic actuators work successfully over a moderate range of spring stiffnesses. Unfortu-
nately no previous work gives design guidelines for choosing an appropriate spring constant in the
actuator. The actuators in the robots were mostly designed through trial and error iterations. This
thesis provides a more quantified understanding of the effects of the springs in the system and also
provides direction to spring selection in actuator design.

2.4.2 Other Electro-mechanical Compliance

There are other examples in which springs have been used in series with transmissions. Elastic
transmissions are an integral part of Morrell’s parallel coupled micro-macro actuators [39]. In this
scheme, a small direct drive actuator is coupled in parallel with a larger geared motor through an
elastic transmission. The large motor creates a force with gross sensitivity and the small direct drive
motor accounts for the error. While increasing the complexity of the actuator, a substantial increase
in dynamic range is achieved.
In the control and handling of arbitrary two dimensional objects, Hanafusa created a three finger
gripper with springs at the end of each finger [21]. The fingers were driven by a single motor and
by measuring the deflection of each spring, stable grasping forces were achieved.

38
Figure 2-3: COG is a humanoid robot with upper torso, arms and head. There are series elastic
actuators for each of the six degrees of freedom in each arm. Here COG is turning a crank using a
dynamic oscillating controller [76].

Figure 2-4: Spring Flamingo is a planar bipedal walking robot. It uses series elastic actuators to
actuate its six joints. Its top walking speed is 1.25 m/s [50].

39
Figure 2-5: Corndog is a planar running robot representing half of large dog. It uses the electro-
mechanical series elastic actuators developed as part of this thesis to actuate its four joints [34].

Figure 2-6: M2 is a 3D bipedal walking robot. It has 12 degrees of freedom: 3 at each hip, 1 at
each knee and 2 at each ankle. The goal of M2 is to extend work done on Spring Flamingo to 3
dimensions.

40
Figure 2-7: Sarcos Dextrous Arm. It is a hydraulic force controlled teleoperated manipulator that
has been used for many applications including production assembly, undersea manipulation and
hazardous material handling. The manipulator actuators have accumulators on either side of the
hydraulic fluid chambers. This gives the actuators intrinsic compliance. Reprinted by permission of
Sarcos.

In the manipulator design for mobile robots to perform cooperative object handling, Sugar built
a three degree-of-freedom parallel manipulator which could generate forces in the x and y directions
and a moment in the z direction [69]. The actuators for the manipulator have springs between the
transmission and load, and the force is controlled by measuring the deflection of the spring. This
design is very similar to the series elastic actuators described above.
Sugano designed and build a compliant one degree-of-freedom finger [68]. The finger’s compliance
could be set by changing the effective length of a leaf spring which was in series with the motor and
transmission. The actuator required three motors. One would drive the position and one would set
the spring stiffness. An additional motor was required as a brake and used to damp out the residual
vibration from the spring.

2.4.3 Hydraulic Compliance

Springs and other compliant mechanisms have also been used in conjunction with hydraulic actu-
ators. Following examples from biological studies of running [3], Raibert’s hydraulic legged robots
use springs to recover impact energy from each step during locomotion [52]. Using the stored energy
in the springs during stance, these robots were able to perform simple gymnastics such as a flip [25].
These robots were not under force control; however, the springs in the legs are necessary to maintain
stability under high shock loading with each step.
As discussed previously, Alleyne designed and demonstrated a Lyapunov based controller for
hydraulic piston force control [4]. The piston had a standard load cell attached to its output.
However, rather than using an inertial load, the force sensor pressed on a spring fixed to ground.
The spring compliance dominated over the sensor compliance and essentially lowered the loop gain
and allowed the actuator to perform as well as it did.
The first force reflective teleoperated hydraulic manipulator was built by Sarcos, Inc. (Salt Lake
City, UT) [30] (figure 2-7). Using one of the actuators for this robot, Wells characterized a hydraulic
piston with accumulators which gives the actuator intrinsic compliance [73]. Instead of a mechanical
spring at the actuator output, the piston has a fluid spring on the inside. Wells demonstrated that
due to the compliance in the piston chambers, the actuator was much better able to tolerate external
disturbances.

41
 As with electro-mechanical series elastic actuators, little published work has been done to quantify
the effects of compliance for hydraulic actuators. This thesis applies the ideas of series elastic
actuators to the hydraulic domain. Similar ideas about the spring design tradeoffs and considerations
are discussed as in the EM series elastic actuator. Initial work on hydro-elastic actuators done as
part of this thesis has also been reported previously [53].

2.5 Summary
There are many applications for robotic force control, especially when the robot needs to interact with
an external unknown environment. While there exist several types of algorithms and methods for
implementing force control, there is a lack of actual robots than can achieve expected performance.
One primary reason for this limitation is the robot actuators.
Robot actuators are especially poor at force control when high force and power density are
required. In order to achieve the high power density, electro-magnetic motors require power trans-
mission which introduces many problems including increased friction, large reflected inertia and high
impedance. Hydraulic actuators achieve high force and power density requirements. However, force
at the end of the hydraulic piston is difficult to control due to non-linearities, pressure variation and
seal friction. Hydraulic and electro-mechanical actuators are much better suited for position control.
Closed-loop actuator force control has been shown to be improved with the introduction of
compliance and controlling the deflection of the compliant element. The following chapters of the
thesis describe and quantify this finding. I describe a general theory for series elastic actuators and
give two case studies for implementing a series elastic design in the hydro-mechanical and electro-
mechanical domains.

42
Chapter 3

Linear Series Elastic Actuators

This thesis focuses on understanding the effects of placing a significantly compliant spring in series
between a motor output and load. By measuring the compression of the spring, the force in the
spring and thus the force on the load can be inferred. Closed-loop feedback control then modulates
the actuator system to obtain the desired output force. This technique has been termed series elastic
actuation.
This chapter discusses the fundamental principles of series elasticity using a minimal complexity
model which is not power domain specific. The general model demonstrates closed-loop bandwidth,
internal error rejection, large force bandwidth, closed-loop output impedance and impact tolerance.
It also discusses the effects on the actuator of an inertial load moving in free space. This chapter
provides a unifying framework and descriptive language for understanding series elastic actuators.

3.1 General Model

As discussed in the previous chapter there are many methods of actuation, each with its own set
of desirable and detrimental characteristics. Typically, actuators with high force and high power
density also have high output impedance. They are better at control of position or velocity output
rather than force control.
The main idea behind series elasticity is to take a standard actuation method with high output
impedance, add compliance at the end point and use feedback control to modulate the force acting
through the spring. This is done by directly measuring spring deflection or strain in the spring.
Figure 3-1 shows the main components of a series elastic actuator with the shaded line indicating
the division between power domain and signal domain. The main elements are:

• Elasticity

• Control System

• Motor

• System Inputs

It is important to point out that topologically, this minimal model is no different from any
actuator with closed-loop force control. Any actuator of this sort will have a motor, a sensor with
compliance and a control system. The major difference between series elasticity and other actuators
with closed-loop force control lies in stiffness of the sensor. This subtle yet very important difference
is the key to series elastic actuators.
In section 3.1.4 through section 3.1.2, I define and describe the four main elements of the general
model for series elastic actuators. The definitions lead to the derivation of the open-loop model and
subsequently the closed-loop general model for series elastic actuators in section 3.2.

43
 Signal Power ks
Spring xl
+
Control
fd Σ System
Motor
-

fl

Figure 3-1: Minimal model for a series elastic actuators. There are four parts to a series elastic
actuator: elasticity, control system, motor, and system inputs. The compliance is a simple linear
spring with a stiffness of ks . The force in the spring, fl , is equal to the spring deflection times the
stiffness. The control system consists of a simple proportional controller with gain K. The motor
is modeled as a high impedance velocity source. The position output of the motor, xm , is the time
integral of the motor velocity. The system inputs or boundary conditions are the desired force, fd
and the load position, xl .

3.1.1 Elasticity
The elasticity in the actuator is a transducer and thus has two parts. One part, the spring, is in the
power domain and the other part, the strain or deflection sensor, is in the signal domain.
The elasticity is a simple linear spring with a spring constant of ks . It is load bearing but passive
and is placed at the output of the motor. Depending on the spring stiffness, the deflection can be
significant in order to cover the full force output range of the actuator.
The sensor measures the deflection or strain in the spring which is a representation of the force,
fl , acting through the spring. I assume that there is no hysteresis in the spring and therefore the
relationship between force and spring deflection is defined by Hooke’s Law [18]:

fl = ks xs (3.1)

where xs is the spring deflection.

In an ideal world xs is defined by the difference between the actuator position xa and the load
position xl .
xs = xa − xl . (3.2)
While it is possible to measure both xa and xl and take the difference for the measurement,
it is not recommended. There may be uncorrelated noise between these two measurements due to
transmission dynamics and other non-linearities. In order to reduce noise from the compliant sensor,
it is much better to make a single direct measurement of the spring strain xs . Regardless, I assume
that the measurement of the spring strain is a very good representation of the for force output of
the actautor.

3.1.2 Control System

The spring strain measurement xs implies the load force fl (equation 3.1) or force in the spring and
is used for feedback in a control system. Given that there is some desired force fd for the actuator,
the error between these two signals will define the input to the controller fe :

fe = fd − fl . (3.3)

44
 vc x m= vm

Figure 3-2: Graphical representation of the minimal motor model. The minimal motor model is
a high impedance velocity source. It neglects inertia and has an instantaneous limit of force and
velocity. It represents idealized models of both a hydraulic piston as well as an EM motor with a
transmission.

Throughout this thesis, emphasis is placed on controller simplicity in order to understand the
basic characteristics of the complete actuator. For the purpose of this discussion, the control system
is represented by the controller gain K. Therefore, the control law for the general series elastic
model is:
vc = Kfe (3.4)
K has the proper units to convert force error to desired actuator velocity.
With further controller development, improvements in overall system performance can be re-
alized. However, a proportional control law K clearly demonstrates the closed-loop effects of the
compliant load sensor. Premature introduction of complex control systems would cloud understand-
ing of the underlying issues. There are ample methods for controller design once the basic concepts
of the plant characteristics are clarified [42, 65].

3.1.3 Motor
The motor converts energy into mechanical force and motion. For the electro-mechanical domain,
the actuator system consists of a motor, transmission and motor amplifier. Force on the motor mass
creates a high impedance velocity output through the transmission. Similarly, for the hydraulic
domain, the motor consists of a piston and servo valve connected to supply and return pressure.
The servo valve modulates fluid flow and thus velocity of the piston output. The piston too can
be considered a high impedance velocity source. Therefore, the motor in the minimal model is also
assumed to be a high impedance velocity source and has a simple relationship between controller
output, vc , and motor output, vm .
vm = Km vc (3.5)
where Km is the gain of the actuator. In this representation, motor inertia is neglected and it is
assumed that the actuator can perfectly generate the output velocity given the commanded velocity
vc . This is analogous to flow rate control in the hydraulic domain and velocity control in the
electro-magnetic. The minimal motor model is graphically depicted in figure 3-2.
Since the force in the spring is a function of motor position, I relate the motor position as the
time integral of the velocity Z
xm = vm (t).

In the Laplace domain it can be written as

Vm
Xm = . (3.6)
s
Substituting equation 3.6 into equation 3.5 and solving for the actuator position output Xm ,
creates a transfer function relating controller velocity to motor position
Km
Xm = Vc . (3.7)
s
Since the actuator has such high impedance, it is assumed that the force in the spring, fl , is

45
 vm
Vsat
Ksat

fl
Fsat

Figure 3-3: General motor saturation model. All motors have limits to the instantaneous force and
velocity output capabilities. The line connecting the maximum load force, Fsat , and the maximum
actuator velocity, vsat , defines the envelope in which the motor can operate. The slope of the
saturation line is Ksat = Fvsat
sat
.

equal to the force generated in the motor, fm . Real motors typically have instantaneous limits on
force and velocity output capabilities. Therefore, I include a simple saturation limit for this motor
model. This is graphically shown in figure 3-3.
Fsat represents the maximum steady state force output of the actuator. Vsat is the maximum
motor velocity under zero load. Therefore, the motor velocity, vm , is limited by
 
fl
vm ≤ Vsat 1 − . (3.8)
Fsat

It is important to point out that typically motor systems should be designed to be low impedance
devices when used in force control situations. However, in this case it is actually advantageous to have
a high impedance motor modulating the position of the spring for two reasons. First, the causality
of the spring dictates that it should have a position input to give a force output. The better that
the motor can modulate the spring position, the cleaner the force output of the spring. The second
reason is that it is easier to control the velocity and thus the position output of real motors than it
is the force. This is due to the fact that real motors and transmissions have non-linearities such as
backlash and stiction that create force noise in the system.
Such a simple motor model is hardly representative of real motor and assumptions to justify
this model are extreme. Nevertheless, a first order actuator model captures the essence of series
elasticity. More complex actuator models will be applied when discussing the case studies of two
real actuators in the following chapters.

3.1.4 System Inputs

The final part of the minimal model to discuss is the system input or the boundary conditions.
There are two boundary conditions:
• fd – desired output force.
• xl – motion from the load or robot link.
The desired force is in the signal domain. The value of fd is generated by a higher level controller.
Ideally, the output force of the actuator equals fd with one-to-one correspondence in magnitude and
phase over the frequency spectrum.

46
 xm ks xl
+ vc
fd Σ K
-

fl

Xl
+ Vc K Xm+ - Xs Fl
Fd Σ K m Σ ks
- s

Figure 3-4: Minimal model and block diagram for a series elastic actuator. The figure on top is a
time domain graphical representation of the the minimal series elastic actuator. The bottom figure
shows a block diagram representation of the actuator.

The motion of the load, xl , is in the power domain. Since the actuator is used in force control
situations, it may be either directly in contact with the environment or it may be connected to
a freely moving inertial load. When in contact with the environment, the load position can be
considered a position source either fixed or moving. However, an inertial load moving in free space
defines the load position as a function of the force in the spring and the load mass. There is energy
interaction between the spring and the load that must be taken into account.
It is instructive to first look at the case when the load is in contact with the environment. In a
subsequent section (section 3.4), I consider the case of the inertial load mass.

3.2 Minimal Linear Model Derivation

With each of the main elements of a series elastic actuator defined, the minimal model can now be
derived. Figure 3-4 shows the model in both time and frequency domain representations.
The goal of the actuator is to control the force in the spring, Fl . I use the model in figure 3-4
to derive the linear dynamic equation for Fl as a function of the two inputs Fd (desired input force)
and Xl (position of the load).

3.2.1 General Power Domain Open-Loop Model

The force in the spring Fl is directly proportional to the spring’s displacement Xs (equation 3.1).
Xs is a function of the motor output Xm and the boundary condition for position of the load Xl
(equation 3.2). Substituting equation 3.2 into equation 3.1 for Xs

Fl = ks Xs

47
 = ks (Xa − Xl )

and subsequently putting equation 3.7 into the result for Xa yields the open-loop dynamics for the
force in the spring of a general series elastic actuator:
 
Ka
Fl = ks Vc − Xl . (3.9)
s

Equation 3.9 shows that the open-loop power domain model for the force in the spring, Fl , is a
function of the input from the controller and the position of the load.

3.2.2 General Closed-Loop Model

The closed-loop dynamic equation can be derived by including feedback and a control law. From
equation 3.4, the input force to the actuator Vc is proportional to the error difference between the
desired and load force.
Vc = K (Fd − Fl ) (3.10)
Combining the open-loop power domain model (equation 3.9) with the control law (equation
3.10) gives
 
Ka
Fl = ks Vc − Xl
s
 
Ka
= ks K(Fd − Fl ) − Xl .
s

Solving the equation for Fl , we get the closed-loop dynamic equation of a general series elastic
actuator
ks KKa Fd (s) − ks sXl (s)
Fl (s) = . (3.11)
s + ks KKa
The inputs for equation 3.11 are as described in the previous section 3.1.4, the desired force Fd
and the motion of the load Xl .

3.3 Minimal Model Analysis

The open-loop model (equation 3.9) and the closed-loop model (equation 3.11) are symbolic and do
not represent a physical actuator system. This generic model does, however, help us look at the
key characteristics of series elastic actuators. In particular, the open-loop power domain model of
equation 3.9 and the closed-loop model of equation 3.11 help us discuss:

• Bandwidth

– Small force closed-loop bandwidth

– Large force bandwidth

• Output Impedance

• Impact Tolerance

There are two inputs to the series elastic actuator, desired force, Fd , and load motion, xl . By
looking at the effects of each of these two inputs independently, the closed-loop bandwidth and
output impedance can be understood as two separate cases. These two cases completely define the
linear characteristics of the actuator.
The first case has fixed load motion where xl is constant. Rewriting equation 3.11 with this
constraint defines the closed-loop bandwidth. The relative values of the control gain, K, and the

48
spring stiffness of the sensor, ks , play an important role in defining the actuator bandwidth. In the
control of a series elastic actuator, a decrease in spring stiffness is compensated for an increase in
control gain in order to maintain a high closed-loop bandwidth.
Under the same conditions of fixed load motion, it is also helpful to investigate the large force
bandwidth capabilities of the actuator. Large force bandwidth is defined as the ability of the actuator
to oscillate at a full steady-state output force level. Force and velocity saturation limits of the
actuator and the low stiffness of the spring, reduce the ability of the actuator to output large forces
at increased frequency. Low forces can be generated to the full closed-loop bandwidth. However, the
large force bandwidth is typically smaller and is limited by the open loop dynamics of the system.
In the second input case, the load end is forced to follow a specific trajectory and the desired
force is held constant. Rewriting equation 3.11 under this constraint defines the output impedance
which is a measure of how well the system responds to external disturbances. In other words, the
output impedance defines how well the internal dynamics of the actuator are decoupled from load
motion. I explain the important role that the the spring stiffness, ks , plays in minimizing output
impedance.
Impact tolerance has to do with the ability of the actuator to handle impact loading. This is a
function of the external motion constraint in the general model and the output impedance described
in section 3.3.3. A soft spring stiffness in the sensor and an increased controller gain help the actuator
lengthen the time of impact that in turn minimizes peak impact forces.

3.3.1 Case 1: Fixed Load – Closed-loop Bandwidth

By assuming a fixed load, Xl = Xo where Xo is a constant (figure 3-5), we can write the closed-loop
forward transfer function relating the load force Fl to desired force Fd .
Fl ks KKm
Gcl (s) = = . (3.12)
Fd s + ks KKm
For low frequencies, the transfer function is unity. There is a one-to-one correspondence between
the desired force and measured output force. As frequency increases, the actuator response begins
to drop off (bandwidth). Since this is a first order model, the controlled bandwidth, wc , can be
defined as:
ωc = ks KKm (3.13)
It is a function of the spring constant, ks , controller gain, K, and motor gain, Km . In this simple
case with an ideal actuator, there seems to be no limit to controlled actuator bandwidth. Increasing
the gain K will give any desired bandwidth. However, practical intuition says that actuators cannot
have infinite bandwidth.

Bandwidth limits
Research into the fundamental limits to force control bandwidth has been done previously [15, 16].
This previous work considers the problem of robot endpoint force control. In order to maintain
clarity, the analysis in that research is simplified to focus on a one degree-of-freedom robot. The
results of that research directly apply to this thesis which focuses on a one degree-of-freedom actu-
ator. Topologically, the model used in the previous research and the general series elastic actuator
are equivalent.
Since this work has been done previously, I do not perform an analysis of force controlled actuator
bandwidth limitations. For the sake of completion, I list the findings of the previous work which
gives several potential bandwidth limits. The fundamental factors that limit the force bandwidth
applicable to series elastic actuators are five fold:
• Non-colocated dynamics – Dynamics between the the power modulator (i.e. motor amplifier
or servo valve and the sensor.
• Plant nonlinearities (i.e. transmission backlash and friction, actuator saturation, intermittent
contacts such as impacts)

49
 xm ks
+ vc
fd Σ K
-

fl

+ Vc Xm Fl
Km
Fd Σ K ks
- s

Figure 3-5: Fixed load model and block diagram – case 1. (Top) A fixed load constraint defines the
closed-loop bandwidth of the system by isolating the relationship between desired force and load
force. (Bottom) A block diagram of case 1 shows that the model is a simple first order system.

• Bandwidth limitations of the power modulator.

• Finite sensor resolution

• Controller Dynamics

Depending on the design of the actuator system, any of these five can be the limiting factor. I use
the analysis in the previous work and discuss how each limitation applies to series elastic actuators.
Physical examples of non-colocated dynamics include compliance in the transmission of electro-
mechanical actuators and compressibility of the fluid in hydraulic actuators. Non-colocated dynamics
represent higher-order dynamics that may be excited as the closed-loop bandwidth of the actuator
is increased. These, typically unmodelled dynamics, can drive the system to instability with very
high controller gain.
For most actuator systems, the plant non-linearities, such as stiction or backlash in the trans-
mission or intermittent contacts, hinder the ability to effectively modulate closed-loop force control.
Control systems often cause these actuators to hunt around the desired force and cause chatter.
To eliminate hunting and chatter, control gains are reduced which in turn reduces the controlled
bandwidth. It is interesting to note that when using a stiff load sensor, the magnitude of backlash in
a transmission can represent a significant portion of the sensor deflection. Using a compliant sensor
reduces the effect of backlash and other non-linearities. This is discussed later in further detail.
Even though the power modulator in an actuator system is often assumed to have infinite band-
width, it too has bandwidth limitations. A motor amplifier for the electro-mechanical domain and
a servo valve for the hydraulic are physical examples of the power modulator. The motor amplifier
operates in the kHz range. This is far above the frequency range of interest in series elastic actua-
tors and probably does not significantly affect the overall bandwidth performance. Hydraulic servo
valves, however, have low frequency first order dynamics that strongly affect actuator bandwidth
performance.

50
 fd +
Σ
-
K ks fl

fd +
Σ
-
K ks fl

Figure 3-6: Zero order model for moving the gain from the power domain to the signal domain. A
series elastic actuator removes gain from the stiffness of the spring. The control system can then
increase control gain while maintaining actuator stability. The stiffness of the spring is not a limiting
factor in the closed-loop stability of the actuator.

Sensor resolution can take on many forms: encoder ticks, noise floor of a potentiometer, hysteresis
in strain sensors and discrete motor poles. Each of these reduces information passed in either the
forward or feedback loops. The sensitivity and bandwidth of an actuator cannot be raised above the
limitation set by the sensors. Noise in the sensors is amplified with increased controller gain, which
can lead to instability.
Finally, the controller dynamics can affect the closed-loop bandwidth. Low-pass filtering a noisy
sensor adds poles to the controller dynamics. A similar effect comes from using an integral term in
a controller which adds phase lag. However, it may be necessary to filter signals or have PI control
to get zero steady state error. Therefore, it is important to remember that the controller itself can
be the limit in bandwidth.

Moving system gain

There are various factors that limit the actuator bandwidth. Depending on the design and configu-
ration one of the limitations will be dominant. Increasing closed-loop bandwidth is an optimization
problem between all of the potential limitations. It is interesting to point out, however, that the
spring stiffness, ks , in the sensor is not one of the limits. ks may represent high or low stiffness, but
the fundamental bandwidth limit is independent of ks .
As an aid to understanding how the bandwidth is independent of ks , notice that the spring
constant ks and the control gain K are always together in the loop gain. The key point of series
elastic actuators is that the spring is very compliant in comparison to the stiffness of a load cell.
Therefore, by decreasing ks it is possible to increase the control gain a proportional amount and
maintain the same bandwidth of the actuator. Series elastic actuators take the gain out of the sensor
and put it into the control system. (figure 3-6)

51
 xn
+
fd
+
Σ
-
K +
Σ ks fl

Σ
+
xn ks fl
-

K
Figure 3-7: Zero order model for reducing sensitivity to internal position noise. With sensor stiffness
gain and increased control system gain, internal position noise due to the transmission are greatly
reduced. The provides for a very clean force output from the actuator.

Internal motion error rejection

Mathematically it is unimportant how gain is distributed in a perfectly linear system. However, in
real systems it is important. A real motor does not have perfect velocity output and always has
some position or motion error. A motion error in the motor means that for a given desired position
or velocity output from the motor, there is some variation from that output as expected in a linear
model. An example of this is backlash in a gear train or ballscrew.
I represent a motion error as a noise function xn adding to the system just before the elasticity.
Figure 3-7 shows a zero order model of the internal position noise in the motor. If the desired force
is held constant, then we can see the reduction in sensitivity of the internal position noise on the
output force of the actuator and written as:
fl ks
= . (3.14)
xn 1 + ks K

Essentially, the sensitivity to position noise is reduced by 1/K. As discussed above, decreasing
ks means that K can be increased. Using a compliant sensor and increasing the control gain
dramatically reduces the effect of motion errors such as backlash in a series elastic actuator. The
high control gain helps the actuator to create a very clean force output.

3.3.2 Case 1: Fixed Load – Large force bandwidth

Decreasing spring stiffness and increasing control gain does come with one tradeoff. Keeping the
same boundary conditions with the load position constant, let us look at the tradeoff. In order to
generate large forces, there must be a large elastic deformation (Figure 3-8). At steady state, the
actuator can produce the maximum force output as defined previously, Fsat .

52
 xm xm

Figure 3-8: Large force equals large elastic deformation. In order to create a large force, there must
be a large elastic deformation. This requires the motor must move. In order to oscillate at high
force amplitude, the motor must move very quickly.

Remember that the motor has limits in its instantaneous force and velocity capabilities. As long
as the motor operates within those limits, the overall linear bandwidth of the actuator is not affected
by the compliance in the spring. However, the large force bandwidth is significantly altered. Large
force bandwidth is defined as the frequency range over which the actuator can oscillate at a force
amplitude equivalent to the maximum force, Fsat .
Because of the compliance, the actuator can only modulate the maximum force over a certain
frequency range. This large force bandwidth, hereafter referred to as ωo , is less than the bandwidth
of the complete closed-loop system when using a very compliant sensor. The actuator can still
produce force above ωo but maximum force capability decreases with increased frequency.
This can be understood intuitively by remembering that the power producing element in the
actuator must move a significant amount in order to compress the compliant springs and generate
the maximum force. If the actuator has velocity and force saturation then there is a limit to the
frequency at which the actuator can oscillate at maximum force. Decreasing the stiffness of the
spring means that the distance the actuator must move to achieve maximum force is increased,
thereby decreasing the large force bandwidth. The opposite is true for increasing the stiffness.
I introduce a function Flmax that is the maximum amplitude the actuator can oscillate at any
given frequency.
fl (t) = Flmax (ω)sin(ωt) = ks xm (t) (3.15)
I also relate Flmax to the velocity of the motor.

f˙l (t) = Flmax (ω)ωcos(ωt) = ks vm (t) (3.16)

Take the force relationship in equation 3.15 and the motor velocity relationship in equation 3.16 and
substitute them into the motor saturuation constraint of equation 3.8 for vm and fl .

 
fl
vm ≤ Vsat 1 −
Fsat
 
Flmax ωcos(ωt) Fl sin(ωt)
≤ Vsat 1 − max (3.17)
ks Fsat

Solving the inequality for Flsat gives

V
Flmax (ω) ks Fsat
≤ sat
(3.18)
Fsat ωcos(ωt) + ks FVsat
sat
sin(ωt)

In order to better understand the function Flsat , let us look only at the magnitude of the function
in equation 3.18.

53
 Flmax (ω) ks FVsat
≤ r 
sat
2 (3.19)
Fsat
ω 2 + ks FVsat
sat

Equation 3.19 has been normalized to the maximum force Fsat . In this minimal model case, the
large force bandwidth or saturation bandwidth follows a first order profile. It can be thought of as

Flmax (ω) ks FVsat

≤ sat
(3.20)
Fsat s + ks FVsat
sat

At low frequency Flmax equals Fsat . As frequency increases, the magnitude of Flmax begins to
decrease. The break point is at the saturation frequency. Here the saturation frequency or large
force bandwidth is defined explicitly
ωo = ks Ksat (3.21)
where Ksat is equal to FVsat
sat
or the slope of the instantaneous force and velocity limit of the motor.
ωo is a function of two elements in the open-loop dynamics of the power domain, Ksat and ks .
Ksat represents the saturation characteristics of the actuator. It defines how much power can be
put into the spring of the force sensor. ks is the spring constant of the sensor. Since, ks is very
compliant in a series elastic actuator, the large force bandwidth is significantly lower than it would
be if a stiff load cell were used in the feedback loop.
Figure 3-9 graphically demonstrates an example of large force bandwidth. The closed-loop band-
width of the actuator is four times higher than the saturation bandwidth, ωc = 4ωo . The maximum
output force for the actuator is 500 pounds. The plot has been normalized in frequency to ωo . The
actuator can oscillate up to the full closed-loop controlled bandwidth as long as it does not exceed
the saturation limits of the motor. However, as the magnitude of oscillation increases, the motor
saturation dominates the large force output capabilities of the actuator.
For many robotic applications, especially force control situations, having a large saturation
bandwidth is unnecessary. The maximum output force of the actuators are only required at low
frequencies. Having a reduced saturation bandwidth is the key engineering trade off of series elastic
actuators. It is important to match the stiffness of the sensor to the requirements of a task. There-
fore, choosing a proper ωo for the actuator is critical for making closed-loop actuator force control
work properly.
Equivalent saturation frequencies can be defined for the large force bandwidth as a function of
the open loop dynamics of the power domain for both electro-mechanical and hydraulic systems.
This will be discussed in the case studies of the following chapters.
A reduced large force bandwidth is the main reason for performance gains in the improved output
impedance and impact tolerance discussed next.

3.3.3 Case 2: Forced Load Motion – Output Impedance

The second case of interest is when the load is free to move and the desired force is fixed Fd = Fo
(figure 3-10). With Fd fixed, the dynamics due to the desired force are eliminated. Therefore, the
transfer function relating the change in load force Fl due to the load position xl can be written as
Fl −ks s
Z(s) = = . (3.22)
xl s + ks KKm
This equation is referred to as the output impedance. Fl has been adjusted to be centered on Fo .
Therefore, this is the impedance around the constant desired force Fo . However, often it is important
to know the impedance of the actuator when Fo = 0.
As described in the thesis introduction, mechanical impedance is adapted from the idea of
impedance in electrical circuits [26]. Output impedance is a measure of the stiffness of a system for a
given load motion. For example, the impedance of a spring is its spring constant because the spring
constant relates an input position to output force. For robot actuators, low impedance is desirable.

54
 0 500

Saturation limit
-5
250
/Fsat(dB)

-10

Fl (lbs)
max

100
Fl

-15

Small force bandwidth

-20 50

-25
-1 0 1
10 10 10
ω/ωo

Figure 3-9: sim Small force and large force bandwidth. The small force closed-loop bandwith
is unaffected by the large force saturation constraints. However, as the magnitude of oscillation
increases, the motor saturation dominates the large force output capabilities of the actuator. The
more compliant the spring, the smaller the large force bandwidth. This is the key engineering
tradeoff for series elastic actuators.

55
 xm ks xl
+ vc
Fo Σ K
-

fl

Xl - Xs Fl
Σ ks
-
Xm
K m Vc
K
s

Figure 3-10: Forced load motion model and equivalent block diagram. (Top) The desired force is
fixed constant, Fd = Fo , and the load motion is defined externally. The relationship between load
motion and output force is defined as the output impedance. (Bottom) The block diagram shows
that the feedback system and actuator dynamics are in the feedback loop.

Sensor Stiffness (lbs/in)

Transducer Techniques LBO-500 500, 000
Century Die Spring D-1222-A 4 × 450 = 1, 800

Table 3.1: Comparison of sensor stiffness. The top load cell is a standard off the shelf component.
The bottom entry is the stiffness of the springs actually used in the prototype actuators of this
thesis. The value shown is the one listed in the catalog. Experimentally, the springs were 12% more
compliant.

The output impedance in equation 3.22 is minimal at low frequency. At high frequencies, the
impedance is equal to the spring stiffness of the physical spring in the sensor. Very low impedance
at low frequencies is equivalent to having a zero rate spring connected to the load. Any load mass
will move around and react to the actuator force commands almost entirely decoupled from the
dynamics of the actuator itself.
There are two factors that reduce the impedance of the series elastic actuator, one passive and
one active. The first is the fact that the spring stiffness ks is greatly reduced from that of a stiff
load cell. This is a passive effect and the lower spring stiffness may reduce the impedance by several
orders of magnitude by itself. Table 3.3.3 shows the difference in stiffness from an off-the-shelf load
cell verses the springs used in the series elastic actuator prototypes. The difference is greater than
two orders of magnitude.
The second effect in reducing output impedance is due to the active control system. Note that the
characteristic equation for both closed-loop forward transfer function (equation 3.12) and closed-loop

56
 beq ks xl

Figure 3-11: Equivalent impedance for the general series elastic actuator. The active impedance of
the actuator can be thought of as an equivalent damper in series with the physical spring. At low
driving frequencies, the actuator appears to be a damper. At high frequencies, the impedance is the
stiffness of the spring.

output impedance (equation 3.22) is the same.

s + ks KKm = 0 (3.23)

The argument previously given in case 1 for increasing the control gain K proportional to the
reduction of spring stiffness ks holds for the impedance as well. Therefore, not only is there a
passive impedance reduction with the compliant spring, but the control system has an increased
gain which further extends the low impedance range.
The output impedance has an active physical equivalent at low frequencies. Given that the load
motion is slow enough or has low frequency content then the impedance looks like a physical damper.

Fl = beq sxl (3.24)

Where the equivalent damping coefficient is:

1
beq = . (3.25)
KKm
The impedance of the actuator is graphically depicted in figure 3-11. It shows the equivalent
damper in series with the physical spring. The damping effect is strongest at low frequencies and
the passive impedance of the spring dominates at higher frequencies.
Depending on the actuator dynamics and the control system, the impedance can take on other
physical equivalents at low frequencies such as an equivalent mass, meq . Although, the impedance
is always equal to the spring stiffness at high frequencies.
The closed-loop bandwidth of the system, ωc , is the frequency at which the damping effect
and the passive impedance of the spring are the same. Regardless of the impedance bode profile,
the increased compliance in the sensor dramatically reduces the overall output impedance of the
actuator.

3.3.4 Impact Tolerance

Whenever a robot works in unknown surroundings it makes and breaks contact with that environ-
ment. The interaction power at the system interface is subsequently transferred through the robot
structure and can be significant. To minimize interaction power, most force controlled robots move
very slowly in anticipation of contact. If the impact power is too high, there is the potential for
damage to the robot structure, actuators and to the environment. Series elastic actuators maximize
impact tolerance by the use of the load bearing compliant sensor and can operate at high speeds
even in the presence of potential impacts.
While the energy transferred to the actuator is independent of spring stiffness, the spring and
control system reduce the peak impact power to a range within the actuators output power capa-
bilities. This power matching insures that the impact energy does not exceed the load capability

57
of the actuator. The compliance allows the actuator to move out of the way of the impact while
maintaining the proper desired force.
As an example of the impact power in the actuator from an unexpected load, I assume that the
environment defines a sudden load motion vl . The impact power, Pl , from the load is defined by the
load force and velocity at the spring output.

Pl = Fl vl (3.26)

The spring compression defines Fl .

Assuming that the desired force in the actuator is Fd = 0, the general model for impedance
(equation 3.22) can be redefined in terms of velocity rather than position.

Fl ks
= (3.27)
vl s + ks KKm
Combining equations 3.26 and 3.27 gives the controlled impact power.
ks
Pl = vl2 (3.28)
s + ks KKm
There are two parts to equation 3.28. The first is vl which is defined by the environment. There
is no control over this constraint. Therefore, in order to minimize interaction power, I focus on the
second part of the equation which is due to the actuator.
The second part contains information about the actuator characteristics namely spring stiffness,
control gain and bandwidth. Remember that this is the load interaction power regardless of the
spring stiffness. The power contribution due to the actuator is at worst 1/KKm. As mentioned
above, because of the sensor’s compliance decreasing loop gain, the control gain K can be increased.
Clearly from equation 3.28, as control gain increases, the impact power decreases. The impact
power spectrum rolls off at higher frequency with the break point at the controlled bandwidth of
the actuator but this effect is secondary.

3.4 Mass Load

The two inputs to the system model are desired force, Fd , and load motion, xl . The use of load
motion as an input irrespective of potential load dynamics has to do with the fact that the actuator
load condition is typically unknown. The robot can be in contact with the environment but it could
also be moving in free space. When the robot is in contact with the environment, the situation
is analogous to case one with the load motion fixed. When the robot is not in contact with the
environment, there is an inertial load on the actuator that can move in free space.
Let us look at the specific case for a mass load moving in free space under force control from
a series elastic actuator in one dimension. Since the robot may be operating like this often, it is
helpful in understanding how the load responds to the series elasticity in the actuator as a function
of the input force Fd (figure 3-12).
I show that there are important differences when an actuator is connected to a finite inertial
load from when it is rigidly connected to ground. Understanding these differences helps quantify
actuator limitations.

3.4.1 Forces on the load

Since the load mass is moving in free space, its motion, xl , is defined by the force through the spring,
Fl , and by its own mass, ml .
Fl
xl = (3.29)
ml s2
This is equivalent to Newton’s classic equation F = ma written with Laplace variables.

58
 xl
xm ks
+ vc
fd Σ K ml
-

fl

Figure 3-12: General series elastic actuator with load mass moving in free space. In this particular
case, the load motion, xl , is explicitly defined as a function of the loads mass, ml , and the force in
the spring, Fl .

We know from equations 3.1 and 3.2 that the force in the spring is

Fl = ks (xm − xl ) (3.30)

where xm is the actuator motion defined previously. I rewrite the expression for xm using equations
3.7 and 3.10
Km
xm = Fc
s
Km
= K(Fd − Fl )
s
By putting the actuator motion, xm (equation 3.31), and the load motion, xl (equation 3.29), into
the equation for the force in the spring (equation 3.30) gives

Fl = ks (xm − xl )
 
Km Fl
= ks K(Fd − Fl ) − .
s ml s2

Solving this equation for Fl , we get the force in the spring as a function of the desired force.
Remember that in this case the load mass is free to move.
Fl ks KKm s
= 2 ks
(3.31)
Fd s + ks KKm s + ml

Notice that the characteristic equation of the transfer function in equation 3.31 is now second
order due to the load mass and spring resonance. Our previously derived transfer function (equation
3.12) relating desired force to load force was only first order.
Let us focus on the last term in the denominator of equation 3.31. I define this term as the load
frequency, which is the resonant frequency of oscillation for the load mass and spring.
ks
ωl2 = (3.32)
ml
Ideally, ωl should be as small as possible meaning that ml should be large. As ωl → 0, equation
3.31 reduces to
Fl ks KKm
= .
Fd s + ks KKm
This is identical to closed-loop bandwidth in equation 3.12 discussed previously with the load position

59
 beq ks
Fd
ml

Figure 3-13: Equivalent model of the general series elastic actuator with load mass. The desired
force is seen through a damper before the spring. A simple proportional controller does not make
the actuator a force device.

held constant.
There are two terms in the definition of ωl : ml and ks . Both of these terms can reduce the effect
of the resonance on closed-loop force performance.
As ml increases, ωl gets smaller. In the limit, as ml → ∞ =⇒ ωl → 0, and shows that infinite
load mass is equivalent to a fixed end condition. However, load mass of a robot is always finite
when moving in free space. If the the link inertia is large enough, it may be negligible. In the other
extreem, when load inertia is zero, the actuator has nothing to push on and therefore generates no
output force. Regardless, it is important to understand the robot link inertia to determine whether
or not its effects are important.
In series elastic actuators, the ks of the elasticity is small in comparison to standard load cell
stiffness. This is good since a small ks lowers ωl just like an large ml . This shows one more reason
for the benefit of intentional elasticity in series with the load.

Load mass motion

Now that we know the force in the spring, we can look at the load motion as a function of the input
force. Substituting equation 3.31 into equation for load motion as a function of force (equation 3.29),
yields
xl 1 ks KKm s
= ks
. (3.33)
Fd ml s2 s2 + ks KKm s + m
l

1
Remember that beq = KKm and the equation further simplifies to

ks
xl 1 beq s
= ks ks
. (3.34)
Fd ml s2 s2 + beq s + ml

A careful examination of equation 3.34 shows that it is equivalent to a force pushing on a damper,
spring and the load mass all in series. This is graphically represented in figure 3-13. The transfer
function for the load motion consists of two parts. The first part is
1
. (3.35)
ml s2
This is the equivalent of a force pushing on a mass to produce motion. Ideally, the second part of
the transfer function would be the equation for the closed-loop bandwidth of the actuator.
Fl ks KKm
=
Fd s + ks KKm
While equation 3.34 shows this, because of the free zero in the numerator, the force pushing on the
mass changes a little bit at low frequencies. In order help clarify the meaning of all the term in the

60
load motion transfer function (equation 3.34), the parameters can be rearranged to give
xl 1 1
= ml 2 ml . (3.36)
Fd beq s ks s + beq s +1

The motion of the load is a function of the desired force and appears to be a damper rather than
a mass at low frequency. It has significantly decreased magnitude and increase phase lead. This
effect is true up to the load mass spring resonant frequency ωl . After that point in the frequency
spectrum, the actuator behaves as expected with performance beginning to drop off above the
controlled bandwidth.
The impedance of the actuator at low frequencies is a damper. This is perhaps a mismatched
impedance with the load mass. A more appropriate impedance for the actuator could be an equiv-
alent mass. Since the actuator model is first order, in order to have equivalent mass impedance, it
requires an additional pole. Using an integrator in the control system, changes the dynamics of the
actuator to second order. A second order actuator can match the impedance of the mass load and
produce much better forces within the actuator closed-loop bandwidth. 1
Regardless of the impedance of the actuator (equivalent damper or inertia), the load inertia does
have effect on the closed-loop force generation properties of the actuator. Creating a controlled
system that at low frequencies looks like an equivalent mass is an important lesson used in the
construction and control of the physical actuators. The case studies show that by understanding
the actuator plant and developing an appropriate controller, it is possible to create an actuator that
can be more reliably considered a virtual force source throughout the frequency spectrum of its
controlled bandwidth.

3.5 Dimensional Analysis

Dimensional analysis often helps in understanding the basic relationships of the different variables in
an equation. The equations in this chapter are simple enough to understand without going through
the formality of dimensionless group reformulation. However, rewriting the equations for closed-loop
bandwidth, large force bandwidth, output impedance and inertial load force from this chapter is a
good exercise and will be helpful when applying dimensional analysis to the models in the case
studies.
For the general model there are three dimensionless groups. The large force bandwidth, ωo ,
defined in section 3.3.2 is used to normalize frequency. The other groups are a generalized system
gain, κ, and a normalized mass-spring resonant frequency of the inertial load, L.
s s
S = =
ks Ksat ωo
ks KKm ks 1
κ = = (3.37)
ωo beq ωo
r
ks 1 ωl
L = =
ml ωo ωo
• S is a scaled Laplace variable which normalizes s to ωo .

• κ is the normalized gain in the system. In this case it represents the increase in frequency of
the controlled system above the saturation bandwidth.

• L is the normalized load resonance of the actuator attached to a load mass. It represents the
inertial load resonance in comparison to the saturation bandwidth.

1 Derivation of the closed-loop properties of general series elastic actuator with PI control is not done here. This

derivation is similar enough to the hydro-elastic actuator contained in chapter 4 that the reader is encouraged to refer
to that case study. Closed-loop dynamic equations that both include and do not include the load mass are explained
there.

61
 0 0

-5 -5

-10 -10

Fl/ksXl (dB)
F /F (dB)

-15 -15
d

-20 -20
l

-25 -25

-30 -30

-35 -35
-1 0 1 2 -1 0 1 2
10 10 10 10 10 10 10 10
ω/ωo ω/ωo

Figure 3-14: sim General closed-loop forward transfer function (left) and output impedance (right).
These figures are normalized to the saturation frequency ωo and have κ = 3.

Closed-loop Bandwidth

The closed-loop forward transfer function relating the output force as a function of the desired force
is written in equation 3.12. It is rewritten here for convenience.
Fl ks KKm
G(s) = = .
Fd s + ks KKm
Applying the dimensionless groups from equation 3.38 gives

Fl (S) κ
G(S) = = . (3.38)
Fd (S) S+κ

This equation shows that the DC gain of the closed-loop system is still unity. κ indicates the
breakoff frequency above the saturation frequency where the controlled system rolls off.

Closed-loop Output Impedance

The closed-loop output impedance is taken from equation 3.22. It is rewritten here.
Fl ks s
G(s) = = .
xl s + ks KKm

Applying the dimensionless groups from equation 3.38 gives

Fl (S) S
G(S) = = . (3.39)
ks Xl (S) S+κ

This equation is normalized to ks and shows that the impedance is very low and increases at
20dB/dec up to the controlled bandwidth of the actuator. Remember that this is an equivalent
damper with a low damping coefficient. At the point of transition, the impedance is equal the
spring constant of the physical spring.
Both the normalized closed-loop forward dynamics and the output impedance are shown graph-
ically in figure 3-14.

62
 0

-2

-4

-6

sat
-8

/F
lmax
-10

F
-12

-14

-16

-18

-20
-1 0 1
10 10 10
ω/ωo

Figure 3-15: sim General large force saturation bandwidth. The ability of an actuator to output
its full steady state force level is compromised by the introduction of a very compliant spring. ωo
is the break frequency and is defined by the open loop dynamics of the actuator. The large force
bandwidth is typically less than the controlled bandwidth of the actuator. Even though the actuator
cannot achieve full force output at high frequency, it can and does operate above ωo .

Large force bandwidth

Similarly, I normalize the large force bandwidth from equation 3.20.

Flmax (S)
= 1S + 1 (3.40)
Fsat
This equation is pictured in figure 3-15. The saturation frequency is independent of the control
system and is defined completely by the open-loop actuator dynamics and the spring constant ks .
Reduced saturation bandwidth is the tradeoff that must be made in the design of series elastic
actuators.

Load Forces
Finally, I normalize the force in the spring as a function of the desired force when a load mass is
attached to the actuator and is free to move. I rewrite equation 3.31
Fl ks KKm s
= 2 ks
.
Fd s + ks KKm s + ml

Using the dimensionless groups gives

Fl (S) κS
= 2 . (3.41)
Fd (S) S + κS + L2

This equation is graphically depicted in figure 3-16. In this case, κ = 2 and L = 0.2. The figure
shows that up to the load natural frequency, the forces are equivalent to a damper. However, after
that frequency, the actuator follows the expected bandwidth profile similarly depicted in figure 3-14.

3.6 General Model Summary

The minimal model helps to demonstrate some of the key ideas of using a significantly compliant
force sensor for closed-loop actuator force control. In summary, the simple linear actuator model

63
 0

-5

-10

F /F (dB)
-15

l d
-20

-25

-30

-35
-2 -1 0 1 2
10 10 10 10 10
ω/ωo

Figure 3-16: sim General force bandwidth profile with load inertia moving in free space. The
variables in the figure are κ = 2 and L = 0.2. Any finite load inertia causes the system to have
damper like qualities at low frequency as well as the bandwidth reduction at higher frequencies. As
L → 0, the system behaves as if it has a fixed load end.

demonstrates and quantifies closed-loop bandwidth, large force bandwidth, output impedance, im-
pact tolerance and the effects of an inertial load.

• Closed-loop bandwidth is independent of the spring constant. The reduction in sensor gain
can be compensated for an increase in controller gain. The bandwidth is limited by other
factors such as non-colocated dynamics and valve and amplifier limitations. The increased
control gain gives a very clean force output.
• Large force bandwidth (saturation bandwidth) is reduced with the compliant sensor. The
actuator can still operate above the saturation bandwidth even though the maximum force
output capability is limited. Large force bandwidth is independent of control system. Under-
standing the difference between large force bandwidth and closed-loop bandwidth is key to the
success of series elastic actuators.
• Low output impedance in the actuator is achieved by decreased spring stiffness and in-
creased controller gain. Even stiff flow source actuator systems can be made to be actively
backdriveable and have an impedance no greater than the spring stiffness ks of the sensor.
• Impact tolerance comes from the spring absorbing energy from the environment which
allows the actuator control system to minimize peak impact power. This protects actuator
components from overload damage.
• The effects of an inertial load moving in free space are also demonstrated. In order to have
the actuator be able to apply load forces to a mass, it must have the proper dynamics. It will
be shown in the case studies that making the actuator appear as an equivalent mass at low
frequencies is important for proper actuator operation.

The above ideas are discussed in further detail in specific case studies of applying series elasticity
to the hydraulic domain and electro-mechanical domain.

64
Chapter 4

Hydro-Elastic Case Study

The minimal model for a series elastic actuator in the previous chapter is helpful for quantifying the
important characteristics and tradeoffs when using series elasticity. The deficiency of the minimal
model is the oversimplification of the motor. There are key differences between the minimal model
and specific actuator implementations that deserve discussion.
This first case study looks at series elasticity applied to a hydraulic piston. The case is presented
in a manner similar to the general model. The open and closed-loop models are derived and refor-
mulated in terms of dimensional analysis. These models are used to discuss closed-loop bandwidth,
saturation and impedance. The effects of load mass attached to the actuator moving in free space
is also explained. A physical prototype has also been developed (figure 4-1). Where appropriate,
experimental results are compared to the mathematical models.

4.1 Model Derivation

The hydro-elastic actuator is similar to the minimal model, with a few key differences in the control
system and motor model. After clarifying these differences in the model definition, I derive two
relationships from the closed-loop dynamic equations which quantify closed-loop bandwidth and
output impedance. These equations are at the heart of understanding the hydro-elastic actuator.

4.1.1 Model Definition

There are four main components to the hydro-elastic actuator as seen in figure 4-2.

• System Inputs

• Motor

• Compliant Sensor

• Control System

The system inputs and the compliant sensor are the same as the general model. Specifically
there are two system inputs. Fd is the desired output force and xl is the motion of the load. The
compliance of the sensor is defined by its spring constant, ks . The spring is assumed to be linear in
both tension and compression.
While the system inputs and elasticity are the same as the previous model, figure 4-2 shows that
motor and the control system are different. Therefore, I focus on these differences.

Motor model
The power domain model for the hydro-elastic actuator consists of a servo valve connected to a
hydraulic cylinder and piston. The piston is directly connected to the elastic element (figure 4-3).

65
Figure 4-1: Prototype Actuator CAD model. A 20MPa pressure source is connected to a MOOG
series 30 flow control servo valve (not shown) that directs flow to the two chambers of the hydraulic
cylinder. The piston is coupled to the output through four die compression springs. The spring
compression is measured with a linear potentiometer which implies force. A closed-loop controller
actively moves the piston to maintain a desired spring deflection.

Signal Power
Servo Valve
i Q
ks
Spring
xl
+
Fd Σ K (1 + K i ) A
- s
Hydraulic Piston
Fl

Figure 4-2: Hydro-elastic actuator model. The servo valve directs fluid flow into the hydraulic
cylinder which moves the piston and thus compresses the elastic element. The strain in the spring
is measured and used in a proportional-integral feedback control system.

66
 Servo Valve
i
Q
Fl xl
A

Hydraulic Piston Spring ks

Figure 4-3: Power domain model for the hydro-elastic actuator. The force output of the actuator is
determined by the compression of the spring. There are two inputs to the power domain. Q is the
fluid flow from the servo valve and is a function of the input current i. xl is a motion input from
the environment.

The servo valve directs high pressure fluid flow, Q, into one of the two chambers in the hydraulic
cylinder. The literature for the servo valve defines it as a third order system consisting of a low
first order pole and a higher second order pole pair. Higher order dynamics in the valve are more
than an order of magnitude above the frequency range of interest for the actuator. Therefore, in
the mathematical analysis, the servo valve is assumed to be first order. Also, assuming no power
saturation in the actuator for now, flow is a direct function of input current as seen in equation 4.1.

Kv
Q(s) = i. (4.1)
τv s + 1
Kv is valve gain relating electrical current to fluid flow rate. τv is the first order time constant for
the valve. Flow and pressure limits of the actuator are discussed later in the chapter.
The hydraulic cylinder houses the piston. Potentially, the piston could be double acting and the
piston areas would be the same on both sides. Having different piston area essentially changes the
transmission ratio from the fluid to the mechanical domain depending on which side of the piston
is pressurized. This in turn alters the loop gain of the closed-loop system as well. Even though the
physical actuator prototype is not double acting, in the model I assume that the piston areas are
equivalent and defined by the area A. I do this for three reasons. First, in the real actuator the two
areas are not significantly different from each other. Second, the gain margin of the actuator is very
high. Therefore, there is a built in stability margin for change in loop gain. Finally, it simplifies
analysis and aids in understanding the basic principles of the actuator.
A servo valve and hydraulic piston are very much like the ideal velocity controlled actuator system
described in the previous chapter. The velocity of the piston output, vp , is directly proportional to
the flow rate, Q.
Q
vp = (4.2)
A
The position of the piston can be written in laplace terms directly from equation 4.2
Q
Xp =
As
Substituting equation 4.1 in for Q yields the piston position as a function of valve current.
Kv
Xp = i (4.3)
As(τv s + 1)

Comparing the piston position (equation 4.3) to the actuator position output of the general

67
model in the previous chapter (equation 3.7 rewritten here for convenience)

Km
Xm = Vc
s
shows that they are essentially the same with two exceptions. The first is differences from simple
parameter name variation. KAv compared to Ka and i to Fc . The second difference comes from the
bandwidth limitation of the servo valve, τv s + 1. Up until the frequency 1/τv the two systems are
essentially the same. Therefore, much of the analysis applied to the general model also applies to
the hydro-elastic actuator.

Control System
Feedback control of the actuator is closed by measuring the deflection of the spring which implies
the force output, Fl , of the actuator and comparing that to a desired output force, Fd .
I use a PI controller on the error with controller gain K and integral gain Ki . As can be seen
in equation 4.3, there is already one free integrator in the system. This is due to the fact that the
valve controls fluid flow and the force in the spring is defined by position. The additional integrator
is important for two reasons. The first is to help automatically compensate for non-linearities in the
servo valve not accounted for in linear analysis. The second reason is that it is beneficial to make
the actuator second order. This is especially important for the impedance at low frequencies as the
actuator can be made to appear as an equivalent mass meq . The calculation for meq is defined later.
Returning to the control definition, I assume that the control gain K and Ki have the proper
units to take voltage signals Fd (desired force) and Fl (load force) in the spring to control current,
i, sent to the servo valve. This relationship is determined by:
Ki
i = K(1 + )(Fd − Fl ). (4.4)
s
This controller puts a pole at the origin and a zero at −ωi on the negative real axis where ωi = Ki .

4.1.2 Power Domain Model

The spring deflection determines the load force, Fl . This is a function of the piston position, Xp ,
and load motion, Xl . The hydraulic piston deflects the spring. Assuming that there is some counter
force from the load, the force in the spring, Fl , is defined as a function of two variables: the flow
from the servo valve, Q, and the relative position of the load, Xl . This relationship is derived to be:

Fl (s) = ks (Xp − Xl )
 
Q
= ks − Xl
As
 
Kv
= ks i − Xl (4.5)
As(τv s + 1)

where ks is the spring stiffness of the series elasticity.

4.1.3 Closed-loop Model

I derive the closed-loop model by substituting the controller equation (equation 4.4) into the power
domain model (equation 4.5) for the valve current i. Rearranging and solving for Fl yields the
closed-loop equation for the force through the spring as a function of the desired force, Fd , and the
load position, Xl

( Ksi + 1)Fd (s) − A 2

KKi Kv s (τv s + 1)Xl (s)
Fl (s) = A 2 s
. (4.6)
ks KKi Kv s (τv s + 1) + Ki + 1

68
 With the addition of the integral term in the controller, the closed loop system characteristic
equation is third order.

4.1.4 Two input cases

As in the general model, it is helpful to write equation 4.6 as a function of the isolated input variables.
For the first case, I assume that the load position is constant (Xl = 0). This eliminates the load
motion. The second case is with the load end free to move and the desired force fixed. This is
the output impedance and eliminates the dynamics due to desired force modulation. The transfer
functions for these two cases completely specify the linear characteristics of the actuator.

Case 1: Fixed Load

By imposing a fixed end condition on the output and using equation 4.6, I write the closed-loop
forward transfer function for case one which relates the desired force to the output force:

Fl (s) ( Ksi + 1)
= A 2 s
. (4.7)
Fd (s) ks KKi Kv s (τv s + 1) + Ki +1

At low frequency, this transfer function is equal to unity regardless of the controller values. In the
limit as frequency goes to infinity the transfer function goes to zero. The servo valve may or may not
have a significant effect on the closed-loop dynamics depending on the bandwidth of the actuator.
As the controller gain increases, increasing bandwidth, the servo valve plays a more dominant role.
Therefore, the true bandwidth limit of the system will depend on the servo valve characteristics.

Case 2: Free end with zero force

The other case is to assume that the desired force is zero and the output is free to move. This will
be the zero load impedance. For case two, we write:

Fl (s) − KKAi Kv s2 (τv s + 1)

= A 2 s
. (4.8)
xl (s) ks KKi Kv s (τv s + 1) + Ki +1

Opposite of the closed loop forward transfer function, the impedance at low frequency is small.
This is the property we desire in order to make the actuator very insensitive to position inputs.
A low output impedance implys that the actuator dynamics are decoupled from that of the load
motion. With active control, the actuator is backdrivable. At high frequency, the transfer function
is equal to the value of ks , the spring constant of the physical spring. Essentially at higher frequency
the piston cannot move and therefore the load only sees the spring.

4.2 Model Analysis

I discuss the hydro-elastic actuator with respect to the following characteristics:

• Saturation

• Closed-loop bandwidth

• Output impedance

• Proportional Control

Regardless of the control system, when using series elasticity, saturation plays an important roll
in defining the large force output capabilities of the actuator. Taking some simplifying assumptions
about the load/flow characteristics of the servo valve, I define a frequency at which the performance

69
 1

0.9

0.8
Hydraulic Saturation
0.7

0.6

Q/Qo
0.5
Linear Simplification
0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(Fload/A)/Ps

Figure 4-4: sim The power saturation profile for a hydraulic servo valve is a square root relationship
between flow and pressure. In order to understand the effects of saturation on the actuator through
linear analysis, a linear saturation relationship is assumed. The linear profile is a worse case than
the square root saturation.

of the actuator begins to diminish due to saturation. This frequency is then used to normalize the
dynamic equations.
I next investigate equation 4.7 to understand the closed-loop bandwidth limitations of the actu-
ator. Since the actuator is third order, I look at limitations of the control gains.
The output impedance of the hydro-elastic actuator is small at low frequencies. I look at the
output impedance of equation 4.8 and quantify how well the actuator dynamics are decoupled from
that of the load.
In an exercise of comparing the hydro-elastic model with the general model in the previous
chapter, I eliminate the integral term of the controller and look at the effects on the closed-loop
dynamic equations.

4.2.1 Saturation and Large Force Bandwidth

Regardless of the control system, the open loop characteristics of the actuator dominate the response
when operating at the force and velocity limits of the servo valve. The maximum flow rate assuming
a fully open valve is r
Fl
Q = C Ps − (4.9)
A
where Ps is the supply pressure, C is a servo valve sizing factor and Fl and A given previously.
Equation 4.9 assumes a negligible return pressure in comparison to the supply.
Square root saturation is difficult to use in linear analysis. By assuming a linear saturation profile
that is close to but a worse case than the square root saturation (equation 4.9), it is possible to
understand the effect that the spring has on the large force bandwidth of the actuator (figure 4-4).
The linear saturation profile is defined as
 
Fl
Q = Ksat Ps − . (4.10)
A

In this case Ksat = Q

Ps , where Qo is the maximum flow rate for the rated current of the valve given
o

no load. C in equation 4.9 and Ksat in equation 4.10 perform the same function relating pressure

70
to flow but they are not equivalent.
As a baseline for understanding the large force or saturation bandwidth, I assume that the load
is fixed as in case 1. The load motion does effect the large force bandwidth but I take this one data
point as a fixed reference. I also assume that the controller knows just what to do in order for the
actuator to generate maximum force over the entire frequency spectrum. Since the actuator is good
at low frequencies the large force bandwidth is independent of controller and is solely a function
saturation characteristics of the hydraulic power domain.
As in the minimal model analysis, I define a function Flmax that is the maximum force amplitude
of oscilation for the actuator at any given frequency. I insert the linear flow rate saturation into the
open-loop power domain dynamic equation for the maximum force in the spring (equation 4.5).

Qmax
Flmax = ks
As  
Flmax
Ksat Ps − A
= ks (4.11)
As
Simplifying this equation to solve for Flmax and normalizing by the maximum steady state output
force, Ps A, shows that the actuator oscillating at its saturation force level has a first order bandwidth
profile.
Flmax 1
= A2 (4.12)
Ps A ks Ksat s + 1

Oscillating at low frequency, the system can achieve the maximum force output of the supply pressure
times the piston area. However, the capability of the actuator begins to decline after the break
frequency
ks Ksat
ωo = . (4.13)
A2
The careful selection of this frequency is critical to the performance of the actuator for its specified
tasks. Even though, the actuator can still operate above ωo , its maximum force capabilities are
reduced. The large force bandwidth requirement of the task should be well defined. Note that this
relationship was made independent of the control system.

Dimensionless formulation
In order to generalize and see the effects of the various parameters, both the closed-loop transfer
function, equation 4.7, and the output impedance, equation 4.8, can be written in dimensionless
form.
I normalize frequency about ωo (equation 4.13) and generate the following dimensionless groups.

s
S =
ωo
Ki
I =
ωo
1
V =
τv ωo
ks KKv
κ = (4.14)
ωo A
• S is a scaled Laplace variable which normalizes s to ωo .

• I can also be written as ωi /ωo . It is the scaled placement of the zero on the negative real axis
due to the integral term in the controller.

• V is the scaled time constant from the servo valve first order model. It is a measure of how
much faster the servo valve is in comparison to the saturation bandwidth.

71
 10

Fload/Fdesired
0

−10

−20

−30
−2 −1 0 1
10 10 10 10
ω/ω
o
0

Phase (degrees) −50

−100

−150

−200
−2 −1 0 1
10 10 10 10
ω/ω
o

Figure 4-5: sim This is a bode plot of the dimensionless closed loop system scaled to ωo . Values
for the plot were calculated from real parameters on the device: ωo = 152 rad/sec (25 Hz), κ = 2,
I = 0.3 and V = 4.4.

• κ represents the scaled loop gain determined by the spring, size of the cylinder, valve gain and
controller gain.

Using the dimensionless groups and the scaled Laplace variable, the dimensionless forward closed-
loop and output impedance transfer functions can be written as:

Fl (S) κV (S + I)
= (4.15)
Fd (S) (S + V )S 2 + κV S + κIV
Fl (S) −(S + V )S 2
= (4.16)
ks xl (S) (S + V )S 2 + κV S + κIV

In order to make the impedance, equation 4.16, truly dimensionless, it has been normalized to
the spring constant ks .

4.2.2 Case 1: Closed-loop Bandwidth

Figure 4-5 shows a bode plot for equation 4.15. The plot is normalized to ωo . For reference, in the
real actuator ωo = 150 rad/sec (≈ 25 Hz). The other values used in the plot are κ = 2, I = 0.3 and
V = 4.4. These values are taken from the physical actuator discussed later. At low frequency, the
system tracks very well. The system response dies quickly at a frequency above ωo . The apparent
resonance in figure 4-5 is a strong function of the zero introduced by the control system.
There are two major closed-loop bandwidth and stability limitations for the hydro-elastic actu-
ator. The first is due to the dynamics of the servo valve. The second is due to the introduction of
an integral term in the controller. The second limitation is not completely independent of the first
but it is helpful to look at each separately.

Valve dynamics
The servo valve used in this thesis is specified as third order. However, given that the frequency
range of interest for the actuator is low enough, a first order model is deemed sufficient for the basic
analysis of understanding the fundamental characteristics of the actuator.

72
 For the moment assume that the servo valve is linear and does not require integral control.
Eliminating I from equation 4.15 and simplifying yields:

Fl (S) 1
= S2
. (4.17)
Fd (S) κV + κ1 S + 1

The natural frequency and the damping of the system are coupled. Nevertheless, increasing gain
does not limit system bandwidth. The actuator may have very low damping at high gain but it still
stable. This is true until the point at which the servo valve higher order dynamics come into play.
The addition of just a small amount of phase from the unmodelled dynamics push the dominant
poles of the controlled system into the right half plane. I realize that a simple lead controller could
effectively extend the usable bandwidth of the actuator further. The important point though is that
the actuator has fundamental limits in its bandwidth and stability due to the limitations of the
valve.

Integral limitations

Unfortunately, the servo valve is not linear. At steady state there is a wandering null offset that
must be accounted for using an integral term in the controller.
When using integral control, there is a controller pole placed at the origin that eliminates steady
state error and a controller zero placed at a frequency equal to the integral gain Ki . The dimen-
sionless group I represents the controller zero frequency with respect to the saturation frequency.
In order to understand the effect and limit of the integral term, it is helpful to apply Routh’s
stability criterion to equation 4.15. The expanded characteristic equation is:

S 3 + V S 2 + κV S + κIV = 0.

All of the coefficients are positive numbers. The array of coefficients becomes

S3 1 κV
S2 V κIV
κV 2 −κIV (4.18)
S1 V
0
S0 κIV 0

The condition that all roots have negative real parts is given by the term in the S 1 row.

κV 2 − κIV
≥0
V
Reducing this equation to its simplest form, shows an important relationship between the servo valve
time constant and the placement of the controller zero due to the integral term.

V ≥ I
1
≥ Ki (4.19)
τv
The controller zero must be placed at a frequency less than the first order bandwidth of the servo
valve. It is interesting that this relationship is independent of the loop gain of the system.
Since our frequency range of interest is so low relative to the higher order dynamics in the servo
valve they are typically neglected. However, similar to the proportional control case, the higher
order dynamics also limit the overall loop gain of the actuator.
In summary, the stability and thus the bandwidth of the actuator is primarily limited by the
relative slow dynamics of the servo valve. Regardless, the control system is capable of increasing
gain high enough that the bandwidth of the actuator is sufficient for operation.

73
 20

Fload/(ksxload)
0

−20

−40

−60

−80
−2 −1 0 1
10 10 10 10
ω/ω
o
0

Phase (degrees) −50

−100

−150

−200
−2 −1 0 1
10 10 10 10
ω/ω
o

Figure 4-6: sim The impedance of the actuator at low frequency is zero and is ks at high frequency.
The plot represents equation 5.14. It uses values ωo = 152 rad/sec (25 Hz), κ = 2, I = 0.3 and
V = 4.4 which were calculated from the prototype actuator.

4.2.3 Impedance: Case 2

Equation 4.16 shows the dimensionless form of output impedance. I rewrite it here for convenience.

Fl (S) −(S + V )S 2
Zcl (S) = =
ks xl (S) (S + V )S 2 + κV S + κIV

At low frequencies the impedance is small but in the limit as frequency increases, the impedance
is equal to the stiffness of the physical spring, ks . Equation 4.16 normalizes to ks and figure 4-6
shows an example of the impedance. The figure uses the same values as in the closed loop forward
transfer function in figure 4-5 and equation 4.15. In this example, the impedance up to and including
the saturation frequency, ωo , is significantly reduced.
Remember that the reduction in impedance comes from two characteristics of the series elastic
actuator.

• Decreased the spring constant.

• Increased the control gain.

Increasing the control gain is desirable for increasing the bandwidth of the actuator. Here we
see that a larger control gain will also decrease impedance. At high frequency the impedance will be
equal to the stiffness of the spring in the load sensor, ks . The unique part of series elastic actuators
is that ks is relatively low. Increasing the spring constant will actually hurt the impedance. For
example, if ks → ∞ we will have the equivalent of a stiff load cell in series with the output. On
the contrary, ks → 0 will be an actuator with almost zero impedance. Including a compliant spring
dramatically reduces the impedance. Nevertheless, there must be a engineering decision which
balances the desire for large force bandwidth and low impedance. Later, we discuss this tradeoff in
more detail.
Notice that the impedance is rising at low frequencies at 40dB/dec. This means that there are
two free zeros at low frequency. This can also represent an equivalent mass meq where

Fl = meq s2 xl . (4.20)

74
 F
l
x
l

k
s

40 dB/dec

Figure 4-7: Rough characterization of hydro-elastic impedance. As shown in the simulation, the
impedance at low frequencies is equivalent to a mass and is equal to the spring constant of the
sensor at high frequencies.

meq does not represent a physical mass but rather and apparent one as defined by the impedance
equation when frequency is very small.
A
meq = (4.21)
KKiKv
The higher the gain, the lower the equivalent mass.
Using a simple figure, let me summarize impedance. Figure 4-7 shows the impedance. At
low frequencies, up to the controlled bandwidth of the actuator, the impedance can be considered
and equivalent mass. The load is essentially decoupled from the actuator dynamics all except this
equivalent mass. At high frequencies, the load only sees the physical spring. Shock loads and other
high frequency disturbances are filtered through this spring. Overall the impedance of the actuator
is very low.

4.2.4 Proportional Control

It is interesting to look at the effects of the closed-loop system when the integral term of the controller
is turned off. This leaves the controller as a simple proportional controller.
I rewrite equations 4.7 and 4.8 with Ki = 0.

Fl (s) 1
= A
Fd (s) ks KKv s(τv s + 1) + 1
Fl (s) − KK
A
v
s(τv s + 1)
= A
(4.22)
xl (s) ks KKv s(τv s + 1) + 1

Eliminating the valve dynamics makes the equations identical to those of 3.12 and 3.22 except
that the parameter names are different.

Fl (s) 1
= A
Fd (s) ks KKv s +1

75
 Signal Power
Servo Valve
i Q
ks Load
Spring Mass
+
Fdesired Σ K (1 + K i ) A ml
- s
Hydraulic Piston
Fload

Figure 4-8: Hydro-elastic actuator model with load inertia moving in free space. Unlike the previous
model where xl is a system input defined by the environment, in this model the load inertia defines
the motion of xl as a function of the force in the spring. The load mass is part of the power domain
and stores and releases kinetic energy as it moves.

Fl (s) − KK
A
v
s
= A
(4.23)
xl (s) ks KKv s +1

Notice that both equations are now first order. This is particularly important for the impedance.
Since now it has only one free zero. The low frequency impedance now looks like a damper. In fact
it is possible to even define and equivalent damping term for the hydro-elastic actuator just like was
done for the general model.
A
beq =
KKv
From an impedance perspective, this exercise shows how important it is for the hydro-elastic
actuator to have integral control. The extra zero in the impedance helps to further decouple the
actuator dynamics from that of the load. This is further explained in the next section.

4.3 Effect of Load Mass

To this point in the case study, the load motion has been considered a system input. The reason
for this is that the load has intermittent dynamics as it contacts the environment and then releases
from it. Those types of interactions are difficult to model comprehensively. When in contact, the
dynamics equations derived and discussed so far can be considered accurate. However, the actuator
is most likely attached to an inertial load. When the load inertia is moving in free space, it becomes
part of the power domain (figure 4-8). This situation is very important to understand. Ideally the
low impedance of the actuator decouples the actuator dynamics from that of the load inertia. I show
that if the load is above a minimum mass then the actuator can still be though of as a pure force
source within its closed-loop bandwidth capabilities.

4.3.1 Load Forces

Let us first derive the transfer function for the force in the spring, Fl , as a function of the desired
force, Fd , with the inertia load free to move. The force in the spring is

Fl = ks (Xp − Xl )

76
where Xp is the piston position and Xl is the load position. Xp is defined previously (equation 4.3)

Kv
Xp = i.
As(τv s + 1)

The motion of the load moving in free space comes from the force in the spring. I can write the
force/position relationship as
Fl
Xl = . (4.24)
ml s2
Finally, recall that the current to the servo valve is defined by the controller as (equation 4.4)

Ki
i = K(1 + )(Fd − Fl ).
s
Substitute the controller into the equation for Xp . Use that result and the definition of the load
motion to solve for the force in the spring, Fl as a function of the desired force Fd .
s
Fl (s) Ki + 1
= A
(4.25)
Fd (s) 2
ks KKi Kv s (τv s + 1) + Ksi + 1 + A
ml KKi Kv (τv s + 1)

Assume that the servo valve dynamics are negligible and the equation simplifies to
s
Fl (s) +1
= Ki
  (4.26)
Fd (s) A 2 s A
ks KKi Kv s + Ki + 1+ ml KKi Kv

With one exception, equation 4.26 is identical to that of equation 4.7 derived earlier assuming
that the load was fixed. The one exception is the s0 term of the characteristic equation. This is the
only term that contains the load inertia. In the case the ml → ∞, equation 4.26 becomes equal to
that of the fixed load case.
Assuming that there is a finite load mass, let us analyze equation 4.26. Remember that the
equivalent impedance mass of the actuator defined in equation 4.21 is
A
meq =
KKiKv
The transfer function in equation 4.26 can be rewritten as
s
Fl (s) +1
= Ki
  (4.27)
Fd (s) meq 2 s meq
ks s + Ki + 1+ ml

m
This equation clarifies that it is important to understand the relative magnitude of the ratio meql .
The small the magnitude or the greater the load mass the less power interaction the load mass has
with the spring.

Proportional control and load motion

Referring back to the transfer function in equation 4.26, I rewrite this equation eliminating the
integral control gain, Ki = 0.

Fl (s) s
= A 2 A
(4.28)
Fd (s) ks KKv s +s+ ml KKv

Again, as in the PI control case, if ml is very big, equation 4.28 reduces to closed-loop dynamic
equation for a fixed load. However, given a finite load, equation 4.28 ends up exactly as described
in the general model case. Instead of being an equivalent mass and spring in series with the load, it

77
 xl m eq xl
ks
Fd ml Fd ml

Figure 4-9: Control abstraction for the hydro-elastic actuator pushing on an inertial load. Ideally,
the actuator produces the desired force directly on the load. However, the dynamics of the actuator
turn out to be an equivalent mass and spring. The equivalent mass is defined by the control system
and power domain characteristics. The spring is the actual physical spring in the actuator.

is an equivalent damper and spring. The equivalent damper for the hydro-elastic actuator is defined
as
A
beq = . (4.29)
KKv
An equivalent damper forcing the spring and load inertia has two problems having to do with the
free zero in the numerator. The free zero causes the system to have phase lead at low frequencies.
It also high pass filters all low frequency forces. If these limitations are not important or if these
effects are unnoticeable, then the using a simple proportional controller is fine. However, caution
should be used when looking at this problem. The effects of using a proportional controller with the
hydro-elastic actuator is demonstrated experimentally as well.

4.3.2 Load Motion

Let us now look at the effect of the load motion with this dynamic equation. This relationship can
be written as s
Xl 1 Ki + 1
=   (4.30)
Fd ml s2 meq s2 + s + 1 + meq
ks Ki ml

There are two parts to equation 4.30. The first part is the dynamics of the load mass as a function
of a simple force Fd = ml s2 Xl . The second part of the transfer function comes from equation 4.27
and represents the closed-loop dynamics of the actuator assuming a load inertia.
As long as the load mass is significant relative to the equivalent inertia of the actuator at low
frequencies, the actuator performs quite well as a force source.
An abstraction of the actuator operation can be seen in figure 4-9. Ideally, the actuator pushes
on a load with the exact desired force. The actuator, however, is limited. In its simplest model,
the desired force pushes on and equivalent mass and spring in series with the load. The equivalent
mass has been defined before and the spring is the physical spring in the actuator. Therefore, at low
frequencies, the actuator works very well assuming that ml  meq . Then the performance begins
to taper off as the actuator reaches its bandwidth limit.

4.4 Physical Actuator

The following discussion elaborates on decisions made during the design and construction of a
prototype hydro-elastic actuator (figure 4-10). This section also includes experimental results taken
from the actuator.

4.4.1 Component Selection

The design space for hydro-elastic actuators is very large. Along with geometry and topology there
are five major components in addition to the supply pressure source: servo valve, piston, spring,

78
Figure 4-10: Prototype hydro-elastic actuator. The actuator has a piston with 0.2in2 area in one
direction and 0.15in2 in the other. The piston pushes on precompressed die compression springs.
Spring deflection is measured with a linear potentiometer. Although not shown, fluid is directed
into the piston via a servo valve with a 3000psi pressure source.

sensor, and controller. Of all five components, choosing the spring is the only part of the actuator
which requires unique perspective and is discussed in the next section.
The specifications for the servo valve, piston, and supply pressure need to be done based on the
force, speed, and power requirements for a given task. These design requirements are not unique
to hydro-elastic actuators and would be done for any hydraulic actuator. It is just important to
remember that the bandwidth of the actuator will typically be some fraction of the servo-valve
bandwidth. Therefore, when choosing the servo valve, it must have sufficient bandwidth to allow
for that reduction.
Previous work done on force controllable hydraulics sacrificed seal tolerances on the piston and
cylinder to help reduce stiction and coulomb friction during sliding [52, 73]. In the hydro-elastic
actuator, since there is such a large motion to signal ratio, the effects of stiction and friction are
virtually eliminated. Therefore, seals can be very tight with little to no degradation in performance.
The other limiting component to achieve high feedback gain is the sensor. The sensor needs to
directly measure the spring deflection. This insures that the feedback measurement is a representa-
tion of true force regardless of any hysteresis or other losses in the spring. In the prototype, we use
a linear potentiometer to measure spring deflection.

4.4.2 Choosing the Spring Constant

Selecting the spring constant is a balance between large bandwidth needing a high ks and impedance
needing a low one. We present some guidelines for choosing a spring constant.

1. Select a servo valve, piston, and supply pressure based on the force, speed, and power require-
ments for the given task. The characteristics of the servo valve will then define the maximum
bandwidth of the actuator.
2. Define a minimum acceptable break point ωo for large force bandwidth. Since the characteris-
tics of servo valve, piston area and spring constant define this value, ωo defines a lower bound
on the stiffness of ks .
3. Define a minimum tolerable impedance level which must also come from the task description.
This places an upper bound on ks .
4. Choose a spring within the two set bounds. It may be necessary to iterate. The non-
dimensional equations will help in guiding to know whether to increase or decrease ks and

79
 Parameter A1 A2 Units
Area 1.29 0.97 cm2
Max. Force 2500 2000 Newton
Cont. Force 2500 2000 Newton
Min. Force 0.5 0.5 Newton
Max Speed 1.25 1.5 m/s
Spring Constant 286 286 kN/m
Max. Power 1500 Watts
ωo 25 Hz

Table 4.1: Physical properties of Hydro-elastic prototype actuator. These properties are taken from
component literature and from experimental tests.

how much effect it will have.

4.4.3 Physical Actuator Characteristics

A prototype actuator is used to demonstrate and test series elasticity in the hydraulic domain.
Figure 4-10 shows the prototype actuator used in the experiments and table 4.4.3 gives the phys-
ical properties of the actuator. The experimental actuator is used to verify and demonstrate the
following.
• Open-loop dynamics
• Controlled bandwidth
• Saturation (Large force bandwidth)
• Dynamic range
• Inertial load forces
• Shock tolerance
Where appropriate the experimental data is compared with theoretical models.

Open-loop dynamics
The prototype was built as an experimental test bed and did not necessarily have specific task
requirements. Nevertheless, as a guideline, I attempted to design both the hydro-elastic actuator
and the EM series elastic actuator to have similar characteristics. Table 4.4.3 shows the actual
physical characteristics of the actuator.
I use a MOOG series 30 servo valve with a supply pressure of 20MPa (3000 psi). The 12.5 mm
diameter cylinder is from Custom Actuator Products. The springs are Century die compression
springs. The actuator output is connected in series with these springs. Regardless of whether the
piston is pushing or pulling, the actuator maintains a linear measurable stiffness and deflection. The
spring deflection is measured by a Novotechnik linear potentiometer. This measurement is then used
in a feedback PI controller which has a desired force input. The output of the controller then defines
a control input current to the servo valve.
The actuator is tested to verify its open-loop dynamic model. Frequency response for the open
loop system of the servo valve, piston, spring and sensor system is shown in figure 4-11 for the both
mathematical model and experimental setup. The figure is in units of Hertz and is not scaled to
the saturation frequency as is done with the dimensionless model. For comparison sake only, the
simulation includes a full third order servo valve. Model and experimental data match well in both
magnitude and phase. The model even predicts the frequency of the higher order resonance but

80
 0

i/Fl (dB)
−20

−40

−60

−80
1 2 3
10 10 10
Hz
−50

Phase (degrees)
−100
−150
−200
−250
−300

−350
1 2 3
10 10 10
Hz

Figure 4-11: real Open-loop response for the servo, piston, spring and sensor. The solid line
represents the mathematical model and the dashed line is experimental data. The simulation model
uses a full third order servo valve. The model captures all important features except the light
damping on 800 Hz resonance of the real servo valve.

does a poor job of matching the damping of that pole pair. Nevertheless, since the frequency range
of interest is more than an order of magnitude below this frequency and the magnitude of this pole
pair is virtually nil. Thus, the first order servo valve assumption is justified.

Closed-loop bandwidth
The open-loop model is closed with a PI controller. Remember that closed-loop bandwidth is
measured with the load end fixed (case 1). For figure 4-12, K = 3 and Ki = 50. Ki = 50 indicates
that the zero in the controller is placed at approximately 8Hz.
The closed loop actuator experimental frequency response is close to the mathematical model.
Figure 4-12 shows there is good low frequency response. Somewhere in the 35-40 Hz range the
response begins to drop off. Again, the 800 Hz pole pair is visible but has no effect on the overall
system performance since its magnitude is so low in the closed loop system.
With the load end fixed, it is interesting to show a step response for the actuator (figure 4-13).
The rise time is ≈10 msec and the system is settled in 50 msec. The actuator step response not only
shows the speed of the actuator but it also demonstrates controller interaction non-linear stick-slip
present in the open-loop system and how dither on the servo-valve helps to over come it.
Figure 4-13 shows that the integrator in the controller keeps working on the errors in the system.
Given that there is a small amount of stiction on the piston-cylinder interface or perhaps in the servo
spool valve itself, the integrator slowly winds up the current to the servo valve. At some point, the
stiction gives and then stops again. The integrator causes the actuator to hunt around the desired
force in steady state.
To overcome this problem, I add a small amount of dither on the to the servo valve and the
hunting behavior disappears. For this particular servo valve, I put the dither at 1% of its rated
current oscillating at 100 Hz. Dither increases fluid leakage through the valve decreasing efficiency.
However, the tradeoffs for closed-loop steady state performs are probably worth it.

Large force saturation

In section 4.2.1 we explain that with the introduction of significant elasticity, the actuator may
often be operating at saturation levels because it must physically move a large distance to compress

81
 0

Fl/Fd (dB)
−20

−40

−60

0 1 2 3
10 10 10 10
Hz
0

Phase (degrees) −100

−200

−300

−400
0 1 2 3
10 10 10 10
Hz

Figure 4-12: real Closed loop response for the hydro-elastic actuator under PI control. The solid
line represents the mathematical model and the dashed line is experimental data. As in figure 4-11,
the simulation model uses a full third order servo valve. The two models match very well in both
magnitude and phase for the useful frequency range.

Parameter A2 Units
lbs
Spring Constant (ks ) 1600 in
in3
Max. Flow Rate (Qo ) 12 s
lbs
Supply Pressure (Ps ) 3000 in2 (psi)
2
Area (A) 0.2 in

Table 4.2: Important physical properties for force saturation. The point at which the actuator
begins to show signs of large force saturation is dependent upon these physical parameters.

the spring to full force levels. The point at which the actuator performance begins to drop off at
maximum force levels is defined as the saturation frequency and is solely a function of the open-loop
characteristics of the actuator. For the hydro-elastic case, this saturation frequency is written in
equation 4.13 and is rewritten here as
ks Ksat rad
ωo = = 160 = 25Hz
A2 s
and
Qo
Ksat =
Ps
where Qo is the maximum valve flow rate given no load and Ps is the supply pressure. The values
for the equation are taken from the literatur and are shown in table 4.2. It is important to point out
that the value of Ksat is a constant for a given servo valve. If the actual operating supply pressure
is less than the rated value, the maximum flow rate compensates such that the Ksat stays constant.
As a demonstration of the effects of force saturation on the hydro-elastic actuator, I command
the actuator to oscillate at its maximum force level at different frequencies from 2 to 100 Hz. Figure
4-14 shows the results of the experiment.
The figure shows both the experimental data as well as a simulated large force profile. Both real
and simulated data break around 25 Hz. However, the real data falls off at -40 dB/dec rather and

82
 300

250

200
Force (lbs)

150

100

50

0
Actual
Desired
-50
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (seconds)
300

250

200
Force (lbs)

150

100

50

0
Actual
Desired
-50
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (seconds)

Figure 4-13: real PI control step response without and with servo valve dither. The integrator
in the controller keeps working on the errors in the system. Given that there is a small amount of
stiction on the piston-cylinder interface, the integrator keeps hunting around the desired force. I add
a small amount of dither on the to the servo valve and the hunting behavior disappears. Regardless
of the dither, notice the fast response time of the actuator. The rise time is ≈10 msec and the
system is settled in 50 msec.

83
 5

-5

-10

Fl/PsA (dB)
-15

-20

-25

-30
Experimental
Simulation
-35
0 1 2
10 10 10
f (Hz)

Figure 4-14: real Large force saturation for the hydro-elastic actuator. This is the response of
the actuator when it is oscillating at its maximum force for different frequencies. The large force
performance begins to drop off near the predicted saturation point of 25 Hz. After the response
starts to die off, the first order bandwidth limitations of the servo valve begin to come into effect.
This effect is not accounted for in the simulation. Therefore, there is a difference in the experimental
and simulated responses at frequencies above the saturation point.

-20 as in a first order system. This is because at higher frequencies, the effect of the first order valve
dynamics begin to play a role in the large force bandwidth.

Dynamic range
Formal impedance spectrum measurements have not been performed on the actuator. However,
informally one can observe (experience) that it is easy to backdrive the actuator with finger force.
The actuator feels massive with zero little friction it is commanded to output zero force. The
minimum resolvable DC force was measured to be 1lb. This corresponds to a spring deflection
equivalent to the noise floor of the sensor.
In an attempt to verify high force impedance and force sensitivity, we put the actuator is a
configuration referred to as the risky tester. (figure 4-15). We suspend the actuator from the ceiling
and then suspend me under the actuator in a swing. When the actuator is given a commanded force
just above my weight, the actuator rises to the top of its stroke. If I pick up a 2 lb weight, then the
actuator drops to the bottom. Releasing the weight causes it to rise again.
Depending on which side of the piston is pressurized, the maximum force level is 450lbs to 600lbs.
With the minimum resolvable force equal to 1 lb then the dynamic range is ∼500:1.

Inertial load forces

Earlier in the chapter, I discussed the effects of inertial loads moving in free space. Those results
showed that an inertial load, ml , has greatest effect on the actuator when the load mass is small in
comparison the equivalent mass, meq , of the actuator. The closed-loop transfer function of equation
4.27 is rewritten here. s
Fl (s) Ki + 1
=  
Fd (s) meq 2
s + s + 1 + eq
m
ks Ki ml

The most important thing to notice about the above equation is that if meq ∼ ml then at low
frequency, the steady state force on the load is reduced.

84
Figure 4-15: Risky testing the hydro-elastic actuator. The author is suspended under the hydro-
elastic actuator. The actuator is supporting the load by a commanded force 1 lb above the downward
force of gravity (∼ 250lbs). The author is holding a 2lb weight which drops the actuator. Releasing
the weight causes the actuator to rise. The minimum resolvable force is 1 lb and is due to the noise
floor of the sensor.

Figure 4-16: The hydro-elastic actuator rigidly connected to a hanging mass. The mass is 18kg. By
commanding desired force on the mass oscillating and different frequencies, the effect of the load
mass can be verified.

85
 5

Fl/Fd (dB)
-5

-10

-15 P control
PI control
-20 PI Simulation
-25
0 1 2
Phase(degrees) 10 10 10

50

-50
P control
-100
PI control
-150
PI Simulation
0 1 2
10 10 10
f (Hz)

Figure 4-17: real Actuator force output with an inertial load. The inertial load mass and and
the actuator equivalent mass are 18 kg and 20 kg respectively. Since the two are so close, at low
frequency the actuator displays a significant non unity magnitude under PI control as expected.
The response magnitude rises as frequency increases and then drops at the closed-loop bandwidth
of the actuator. Proportional control shows that the damper characteristics are significant for this
actuator and inertial load combination. Using a proportional controller dramatically decreases low
frequency force magnitude far beyond that of the PI controller.

In order to experimentally verify the effect of an inertial load, I rigidly attach an 18 kg mass to
the output of the actuator suspended from the ceiling as seen in figure 4-16. I then command an
oscillating force signal from 2 to 100 Hz that causes the mass to go up and down. The results of the
experiment are shown in figure 4-17.

The equivalent mass of the actuator in the experiment is calculated to be

A
meq = = 20kg.
KKi Kv
Since the inertia load, ml = 18, and meq are so close, the mass load cannot be neglected. The load
mass was explicitly chosen to demonstrate this fact.

The low frequency response of the actuator matches with theory in that the magnitude function
is reduced (figure 4-17). This is because the load mass begins to accelerate when it feels the force
through the spring. The spring needs something to push on in order to compress. Therefore, the
bigger the load mass in comparison to the equivalent actuator mass, the less effect the load mass
has on actuator performance.

Remember also, that the hydro-elastic actuator under P control exhibits damper like character-
istics at low frequency. As a damper at low frequency there is a free zero in the forward closed-loop
transfer function. The experimental data in figure 4-17 verifies this by showing the 20dB/dec rise
in the magnitude and phase lead at low frequencies under P control. For this particular actuator,
simple proportional control in not recommended.

86
 150
Actual
Desired

100

Force (lbs)
50
Impact

Dropping

-50
-0.2 -0.1 0 0.1 0.2 0.3 0.4
Time (seconds)

Figure 4-18: real Results of drop testing an 18kg (40lb) mass attached to the actuator. At time
t = −0.2, the actuator is pulling the mass with an upwards force (force up is defined negative). At
about t ≈ −0.15, the mass is released to drop. Even though the desired force of the actuator is set
to zero, the impedance is not perfect and keeps pulling upwards with a 20lb force. At t = 0, the
mass hits the end stop and the force spikes. The mass actually bounces a few times but the peak
impact power is just after t = 0.

Shock tolerance

The physical elasticity in the actuator allow it to have tolerance to shock loads. The spring in the
actuator maintains stability with an impact and spreads the impulse out over time. In so doing it
minimized the peak impact power.
In order to test the shock tolerance of the actuator, a load mass is again connected to the
actuator. This is the same as in figure 4-16. The load inertia is suspended at the top of the actuator
stroke with a force equal to the gravity pull on the mass, F = ml g. The desired force is then set to
zero and the actuator drops to the bottom of the actuator stroke exerting a sharp impulse load to
the actuator. Figure 4-18 shows the forces in the spring during the drop test.
At time t = −0.2, the actuator is pulling the mass with an upwards force (force up is defined
negative) of 40 lbs. At about t ≈ −0.15, the desired force is set to zero and mass begins to drop. As
the mass is dropping, there is still an upwards force of roughly 20 lbs. This is due to the impedance
of the actuator. Ideally the actuator would have zero impedance and the load would only feel the
pull of gravity. However, recall that the impedance is an equivalent mass and spring in series with
the load as shown in figure 4-19. Only the inertial load feels the pull of gravity. Nevertheless, in
order to move, the gravity force is distributed between the two masses and the load and actuator
drop.
At t = 0, the mass hits the end stop and the force spikes. The mass actually bounces a few
times but the peak impact power is at about 8-10 msec after impact and spreads out over 40 msec
(figure 4-20). The impulse, J, is equal to the area under the curve of the force profile. Doing a
rough numerical integration shows that the impulse is
m
J = 12kg · .
s

87
 ml g ks
m eq ml ml g

mlg

Figure 4-19: Model of the load mass during the drop test. Ideally, the load mass only feels the force
of gravity, F = ml g, since the actuator is set to output zero force (left). However the actuator still
has impedance. The gravity pull only affects the load mass. The equivalent mass of the actuator
does not feel the gravity. Therefore, in order for the load to drop, the gravity force on the load must
be distributed between the two masses, one real, one apparent (right). In the drop test case, the
meq ≈ ml = 20kg and the gravity force is distributed equally between the two inertias.

150
Actual
Desired

100
Force (lbs)

50

-50
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
Time (seconds)

Figure 4-20: real Close-up of drop test impact forces. This shows the first impact of the load
mass. The impulse or change in momentum of the mass, J, is the area under the force curve. A
numerical integration shows that J = 12kg · m
s .

88
 The impulse can also be verified by looking at the expected impact velocity of the mass load, vl .
vl can be calculated by applying work and energy balances. The potential energy of the load mass,
ml , is equal to its kinetic energy just before impact plus work done on the mass over the travel of
the actuator, h.
1
ml gh = ml vl2 + F h (4.31)
2
Rearranging yields the velocity just before impact.
s  
F m
vl = 2h g − = 0.7
ml s

Therefore, the momentum just before impact is

m
pl = ml vl = 13kg · .
s
The calculations are close enough for verification.
The springs and the control system passively and actively spread out the impulse transfer so
that impact power is minimized. Because of the springs, the time of impact is longer. With a slower
impact, the control system can respond to it. By doing so the components are protected and the
system maintains stability. Figure 4-21 shows three plots regarding the drop test. The top is the
force in the spring again. The middle plot shows the energy stored in the spring. The actual energy
is not that high. The bottom plot shows the time derivative of the energy which is instantaneous
mechanical power. Notice that the maximum power levels are well within the power capabilities of
the actuator.
It is interesting to imagine what sort of forces are present in a comparable actuator with a
stiff spring. The required impulse is the same as the analysis just presented. However, the major
difference is the time of impact. Conservative estimates for time of impact are in the range of a few
milliseconds [20]. Assume that the impact takes 2 msec. The average force of the impact is

J
Fave = = 4000N !
∆t
This force is an order of magnitude greater than without elasticity! This is also just the average
force not to mention the peak force.
In summary, series elasticity helps with shock tolerance.

4.5 Hydro-Elastic Summary

We use a third order model of a Hydro-Elastic Actuator to investigate the closed loop forward
transfer function and the impedance of the system. These two cases completely describe the linear
characteristics of the actuator. The model is generalized by using dimensional analysis. We examine
force bandwidth and minimizing impedance as a function of controller gain and spring constant.
The model has helped to create a basic design guideline for hydro-elastic actuators with particular
emphasis placed on choosing the spring constant for the elastic element. Large force bandwidth
requires a higher spring constant. Minimizing impedance requires a low spring constant. The choice
of spring constant has been shown to be a compromise between these competing requirements.
In both the analysis and physical actuator, we use a linear spring. Future work will investigate
nonlinear stiffening springs which may ease the design tradeoffs currently required. Further testing
will also be done to experimentally verify impedance and saturation.

89
 800

600
Force (N)

400

200

0
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06

0.8

0.6
Energy (J)

0.4

0.2

0
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06

150
Power (W)

100

50

0
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06
Time (seconds)

Figure 4-21: real Impact power. The top figure shows the force in the spring on the first impact.
By taking the area under the force curve, we can calculate the impulse, J, of the impact. J ≈
F2
12kg ṁ/s. The middle figure shows the energy stored in the spring. It is E = 12 kls . Finally, the
impact power is estimated in the bottom figure by looking at the rate of energy change P = ∆E∆t . The
series elasticity spreads the impact time out so that the required power is well within the actuator’s
output power capabilities.

90
Chapter 5

Electro-Magnetic Series Elastic

Case Study

This chapter presents the second case study for series elasticity applied to an electro-magnetic motor
in series with a linear transmission. The case is presented in a similar manner as the general model.
Unlike the hydraulic case, the power domain of the EM actuator is significantly different from that
of the general model. Despite these differences in the actuator model, the actuator displays all of
the same characteristics of series elasticity.
The open and closed-loop models for the actuator are derived and reformulated in terms of
dimensional analysis. These models are used to discuss bandwidth, impedance, error rejection,
impact tolerance and saturation. The effects of load mass moving in free space are also discussed.
A physical prototype has been developed (figure 5-1). Where appropriate experimental results are
compared to the mathematical models.

5.1 Model Derivation

There are some key differences between the EM series elastic actuator and the general model. After
defining and explaining the parameters of the actuator, I derive three relationships. The first is the
open-loop power domain model. The second two come from the closed-loop dynamic equation and
describe the closed-loop bandwidth and the actuator output impedance. These equations are at the
heart of understanding EM series elastic actuators.

5.1.1 Model Definition

As in the general model, there are four main components to the EM series elastic actuator as seen
in figure 5-2 .
• System Inputs
• Motor and Transmission
• Compliant Sensor
• Control System
The system inputs and the compliant sensor are the same as the general model. Specifically
there are two system inputs. Fd is the desired output force and Xl is the motion of the load. Also,
identical to the general model, the compliance of the sensor is defined by the linear spring constant
ks .
While the system inputs and sensor are the same as in the general model, figure 5-2 shows that
motor model and the control system are different. Therefore, I focus on the discussion on these
differences.

91
Figure 5-1: The prototype actuator has a brushless DC motor rigidly connected to a ballscrew which
drives the linear motion. The ballscrew nut is coupled to the output through four die compression
springs. The spring compression is measured with a linear potentiometer. The actuator can output
over 1350 N and move at 25 cm/second.

Signal Power
Motor
ks
Mass Spring xl
+ Fm
Motor
Fd Σ Kp + Kd s K a Amp mm
-
bm
Fl

Figure 5-2: Electro-magnetic series elastic actuator model. The motor mass has a driving force and
viscous friction. The controller drives the motor mass to compress the spring which gives the desired
force output.

92
Motor model
The motor model in the power domain consists of a motor amplifier, motor mass and a viscous
friction element.
The motor amplifier creates a force on the motor mass given a signal from the controller. For the
initial linear analysis of the actuator, I make two assumptions about the motor amplifier. I assume
that the actuator can produce any desired output force up to the force saturation limit, Fsat , and
that the amplifier dynamics are fast enough that they are assumed to be negligible. In the physical
prototype actuator, the amplifier gain Ka , that relates the desired force (input voltage) to torque
on the motor mass (output current), is unity. Ka is therefore lumped into the controller gain and
eliminated throughout the chapter to reduce equation complexity.
The lumped mass mm includes the dynamic motor mass and the mass of the transmission ele-
ments as seen through the transmission. Remember that this reflected inertia is N 2 the actual mass
value where N is the transmission reduction.
The motor friction bm is also seen through the transmission. It too is N 2 of its actual value.
Even so, depending on motor and transmission construction, the viscous friction can be very low.
There are two modes of operations for an EM motor, torque control and velocity control. In this
model, I choose to use torque control. This is a deviation from the previous two models. Velocity
control, to some degree, masks the effect of motor mass which happens to be very important in
defining the large force bandwidth for the EM series elastic actuator. Therefore, while torque
control is perhaps the more difficult method of controlling a series elastic actuator, I do so in order
to highlight effects of the motor mass that makes the actuator inherently second order.

Control System
Feedback control of the actuator is closed by measuring the deflection of the spring which implies
the force output, Fl , of the actuator. I use a PD controller on the error of the signal with gains Kp
and Kd for the proportional and derivative terms respectively. Since the system is already second
order, I do not include an integral term. By remembering that I assume the motor amplifier to be
ideal, the force on the motor mass, Fm is determined by

Fm = (Fd − Fl )(Kp + Kd s). (5.1)

The actuator can be controlled with proportional control alone. However, it is helpful to have a
damping term as well to help reduce the natural resonance on the mass and spring in the sensor.
It would be very easy to also use a feed-forward term in the controller and attempt to keep the
set point of the system equal always centered on the desired force. This would allow the control
system to focus on controlling the error in the spring rather than requiring a large error signal just
to compress the spring at high forces. However, in order to keep the control system complexity to
a minimum, feed-forward is not included.

5.1.2 Power Domain Model

Figure 5-3 shows the model for the power domain for a series elastic actuator. The two constitutive
equations for the power domain are defined by the force in the spring and the motion of the motor
mass.

Fl = ks (xm − xl )
mm ẍm + bm ẋm = Fm − Fl (5.2)

These equations can also be written with laplace variables.

Fl = ks (Xm − Xl )
Fm − Fl
Xm = (5.3)
mm s2 + bm s

93
 xm xl
Fm mm

bm ks

Figure 5-3: EM series elastic actuator power domain model. The motor mass has a driving force
and viscous friction. The Fl is the force output through the spring.

By combining the two equations in 5.3, the force in the spring, Fl , can be written as a function
of two variables: the force input from the motor, Fm , and the relative position of the load, Xl . This
relationship is derived to be:

Fm (s) − (mm s2 + bm s)Xl (s)

Fl (s) = mm 2 bm
. (5.4)
ks s + ks s + 1

This is the open-loop dynamic equation for the force Fl in the spring as a function of the motor
force and the load motion.

5.1.3 Closed-loop Model

As shown in figure 5-2 the open-loop power domain model is modulated by a force feedback loop.
Combining power domain model (equation 5.4) and the feedback signal commands from the controller
(equation 5.1) yields the closed-loop dynamic equation for the force through the spring:

(Kd s + Kp )Fd (s) − (mm s2 + bm s)Xl (s)

Fl (s) = mm 2 bm +ks Kd
. (5.5)
ks s + ks s + (Kp + 1)
The closed-loop dynamic equation for the force Fl in the spring is a function of the desired force,
Fd , and load motion, Xl .

5.1.4 Two input cases

As in the previous chapters, it is helpful to write the force in the spring, Fl , in the power domain
model and the closed-loop model as a function of isolated input variables. For the first case, the
load end is fixed eliminating the load motion. The second case is with load end free to move and
the desired force fixed. The second case eliminates the dynamics due to desired force modulation.

Case 1: Fixed Load

By eliminating the load motion, the transfer function between motor force input and actuator output
can be taken from equation 5.4 and written explicitly as:

Fl (s) 1
= mm 2
. (5.6)
Fm (s) ks s + bkms s + 1

Equation 5.6 will be referred to as is the open-loop transfer function.

94
 By imposing the fixed end condition again and using the closed-loop dynamic equation (equation
5.5), we can write the closed-loop forward transfer function for case one which relates the desired
force to the output force:

Fl (s) Kd s + Kp
= mm 2 bm +ks Kd
. (5.7)
Fd (s) ks s + ks s+ (Kp + 1)

Depending on the magnitude of Kp , at low frequency, this transfer function is close to unity. In
the limit at high frequency it goes to zero. The zero in the numerator is an artifact of the taking
the derivative of the desired force via the error signal.
For series elastic actuators, KP is typically very high. It is at least an order of magnitude greater
than 1 if not more. Therefore,
Kp + 1 ' Kp .

Case 2: Free end with zero force

With the desired force fixed constant, the force in the spring due to the load motion under closed-loop
force control can be written from equation 5.5 as

Fl (s) −(mm s2 + bm s)
= mm 2
. (5.8)
Xl (s) ks s + bm +k
ks
s Kd
s + (Kp + 1)

This is the output impedance. The impedance at low frequency is ideally equal to zero. At high
frequency, it is equal to ks , the spring constant of the physical spring.
Whenever, either of the closed-loop equations are used in the rest of the chapter, it is assumed
that Kp + 1 ' Kp . Even though this assumption is not technically and mathematically precise, it is
accurate enough.

5.2 Model Analysis

As done in the previous chapters, I discuss the EM series elastic actuator with respect to the following
characteristics:

• Saturation

• Closed-loop bandwidth

• Output impedance

• Error rejection

I deviate slightly from the pattern in the previous chapter and begin by discussing actuator
saturation. I use the power domain model of equation 5.6 to help explain saturation. I define a
saturation frequency for the EM series elastic actuator and rewrite the dynamic equations in terms
of dimensionless groups. Power domain saturation is a function of natural viscous damping and back
emf as well as the motor mass–sensor spring resonance. It is important to remember that saturation
is independent of control system.
In order to understand the closed-loop bandwidth, I investigate equation 5.7. I discuss the
selection of control gains to get proper actuator performance as well as the tradeoffs of reduced
sensor gain and increased control gain.
The output impedance of the actuator is defined by equation 5.8. This equations demonstrates
how well the actuator responds to external motion disturbances. The lower the impedance, the
better the actuator dynamics are decoupled from the load.
As mentioned in Chapter 3, internal errors such as stiction are reduced when using a compliant
sensor. This is explained specifically for the controller case of the EM series elastic actuator.

95
 xm
Fm mm

bm ks

Figure 5-4: Electro-magnetic series elastic actuator model with a fixed load. This configuration
helps to define the saturation limits of the actuator.

5.2.1 Saturation
I begin by discussing actuator saturation. This allows for a definition of the saturation bandwidth.
With this definition in place, I can reformulate the dynamic equations into dimensionless terms using
the saturation bandwidth as a normalization point.
In order to clarify saturation at the most basic level in the EM series elastic actuator, I fixed
the load as in case 1 (figure 5-4). Obviously, saturation is affected by load motion. Depending on
the frequency and phase of the load motion, it can either help or hinder the actuator. However, by
fixing the load, I have a fixed point of reference for understanding saturation.
Assuming no velocity saturation for the moment, the amplifier can produce large motor forces
over a wide frequency range. However, in order to create large forces in the spring, the force from
the amplifier must move the motor mass. The force from the amplifier is independent of the force in
the spring. For the EM motor with a transmission in series with an elastic element, the mass-spring
resonance of the motor mass and series elasticity define the plant dynamics. Therefore, the ability
of the actuator to produce large forces is limited in frequency by these dynamics independent of
control system. The frequency at which the open-loop characteristics of the plant begin to fall off is
defined as the saturation frequency ωo .
r
ks
ωo = (5.9)
mm
Since the sensor spring is very compliant, ωo can be low. Typically, it may be a few hertz to
tens of hertz depending on the reflected inertia as seen through the transmission.1 The saturation
bandwidth is well within the controlled bandwidth discussed later. The actuator still operates and
can generate forces at frequencies above ωo . The actuator is simply limited in the magnitude of
large forces at high frequency.
Other saturation effects, such as back emf and viscous friction, also play a role in defining the
large force bandwidth profile near the saturation frequency ωo . These effects are explained after
reformulating the open and closed-loop transfer functions in dimensionless terms.

Dimensionless formulation

Both the open and closed loop transfer functions, equation 5.6 and equation 5.7, and the output
impedance, equation 5.8, can be written in dimensionless form.
First, notice that there are two natural frequencies in the dynamic equations. The first is the
saturation frequency of the motor mass and spring just described ωo . The second is the controlled

1 These number are taken from macro scale actuators built previously.

96
natural frequency ωc .
r
ks
ωn =
mm
r
ks Kp
ωc = (5.10)
mm

Using these two frequencies and the following dimensionless groups we can non-dimensionalize
both the open and closed loop equations equations of case one and the impedance of case two.

s
S =
ωo
ωc p
κ = = Kp
ωo
bm
B = ωo
ks
Γ = K d ωo (5.11)

• S normalizes the equation to the saturation frequency or the natural resonance of the motor
mass-spring system of the actuator.

• κ represents the gain in the system and is a measure of the increase in controlled bandwidth
above the saturation frequency. Remember that the definition of κ assumes that Kp  1.

• B is a scaled natural damping term due to viscous friction physically present in the construction
of the actuator.

• Γ represents the scaled controller damping gain.

The dimensionless open-loop, closed-loop, and output impedance transfer functions are:

Fl (S) 1
Gol (S) = = 2 (5.12)
Fm (S) S + BS + 1
Fl (S) ΓS + κ2
Gcl (S) = = 2 (5.13)
Fd (S) S + (Γ + B)S + κ2
Fl (S) −S(S + B)
Zcl (S) = = 2 (5.14)
ks xl (S) S + (Γ + B)S + κ2

Back-emf, Friction and Large Force Bandwidth

Back-emf and friction play an important role in defining the frequency profile of large force bandwidth
or as stated previously saturation bandwidth. In order to better understand these effects, I introduce
a simple velocity motor saturation model:

Fsat (1 − |vm|
Vsat ) : |vm | ≤ Vsat
|fm | ≤ (5.15)
0 : |vm | > Vsat

where Fsat and Vsat are the maximum force from the amplifier and maximum velocity of the motor
respectively. vm is the actual motor mass velocity. This is graphically depicted in figure 5-5.
When the motor is not pegged at saturation conditions, it still follows a linear back-emf model
where:
Fsat
Femf = vm . (5.16)
Vsat

97
 Fm
Fsat
Ksat

vm
v sat

Figure 5-5: Velocity saturation model for the EM series elastic actuator. Simple velocity saturation
limits the ability of the amplifier to produce high forces. This model is slightly different than the force
saturation model described in the previous chapters but has very similar results. Here Ksat = Fvsat sat
.

The back-emf can be thought of as an equivalent damping because it relates a loss in motor force
due to motor velocity just like the term bm . The equivalent damping coefficient for the back EMF
is Ksat = Fsat /Vsat . I then modify the dimensionless B group to be
Fsat
bm + Vsat
B= ωo . (5.17)
ks
The B group plays an important role in the large force bandwidth. Also, by incorporating the
velocity saturation as an equivalent force we can investigate large force bandwidth in terms of the
saturation force only.
Large force bandwidth is independent of the control system and is only a function of the open-
loop dynamics of the power domain, equation 5.12. While the magnitude of equation 5.12 relates
the load force Fl as a function of the motor force Fm in the linear range, it can also relate maximum
load force output, Flmax , as a function of the maximum motor output, Fsat .

Flmax 1
= 2 . (5.18)
Fsat S + BS + 1
There are three frequencies of interest in the large force bandwidth transfer function: above,
below and near the saturation frequency ωo . At low frequencies, those below ωo , the transfer
function is unity and the actuator can output the maximum force capable by the motor. At high
frequencies above ωo , the maximum force output capability of the actuator is limited. This is because
the motor cannot accelerate the motor mass fast enough to deflect the spring to full compression.
The dimensionless group B defines what the large force bandwidth profile looks like near ωo .
Figure 5-6 shows different values for B with respect to the maximum available force output
due to force saturation. B is an equivalent damping term for the saturation profile. The damping
affects how well the system handles large forces near the natural frequency. With an overdamped
system, the available force amplitude at resonance dramatically reduces the available force amplitude.
Regardless of the damping in the system, high amplitude forces at high frequency are limited.
This analysis confirms previous work stating that elasticity in an actuator reduces large force
bandwidth and increases control effort at high frequencies [29, 46]. The primary variable in defining
ωo is the stiffness of the elasticity, ks . Increasing ks will increase large force bandwidth. When
choosing an appropriate ks , an actuator designer must understand the large force requirements for
the application. The frequency, ωo , is used as a reference point and defines the lower bound for the
spring constant, ks .
The profile of the large force bandwidth near ωo is solely a function of the dimensionless group

98
 5

-5

-10
sat(dB)

-15
/F

-20
lmax
F

-25

-30

-35

Underdamped
-40 Critally Damped
Overdamped
-45
-1 0 1
10 10 10
ω/ωo

Figure 5-6: sim EM large force bandwidth due to force and velocity saturation. The motor mass
spring resonance ωo defines the large force bandwidth of the actuator. The B group (generalized
damping) defines the shape of the profile around the saturation frequency. The three lines represent
underdamped, critically damped, and overdamped frequency response. This plot only represents
the large force capabilities of the actuator with the load fixed.

99
B. In standard control textbooks [42], the resonance for second order systems is defined by the
magnitude of ζ, the damping ratio. Since the natural frequency has been normalized to one for the
open loop transfer function, ζ is related to B by:
B
ζ= . (5.19)
2
Therefore, we can talk about B in terms of under, critical, and over damping as the three curves in
figure 5-6 show.
Let us now take a closer look at B.
Fsat
bm + Vsat
B = ωo
ks
r
bm + FVsat
sat
ks
=
ks mm
 r
Fsat 1
= bm + (5.20)
Vsat ks mm

Equation 5.20 helps us see the relationship of B to the physical model parameters. In general,
we want B to be as small as possible to increase bandwidth. Equation 5.20 can help define design
guidelines regarding friction.

• bm — Make the actuator with as little drive friction as possible.

• Fsat
Vsat — Minimize back EMF.

• mm — Using a larger motor mass will decrease resonant frequency thus making the apparent
damping smaller.

• ks — A larger spring constant decreases spring deflection. Spring and motor velocity subse-
quently decrease and therefore damping decreases.

It is also important to remember that both bm and mm as represented in this model are seen
through a transmission reduction and are scaled by N 2 , where N is the reduction. Therefore a
decrease in reduction decreases the damping.

5.2.2 Bandwidth
Up to this point in the analysis I have only worked with the open-loop forward transfer function.
From here, I will be dealing with the closed-loop behavior.
The controlled bandwidth is significantly higher than the saturation bandwidth for the EM
series elastic actuator. In the previous chapter, I outlined several things that can potentially limit
the bandwidth of the actuator. None of those limitations were a function of the compliance in the
sensor. Therefore, the loop gain reduction from sensor compliance can be proportionally increased
in the controller. This essentially takes the gain out of the spring and into the control system.

Controller Gains
Using the closed-loop characteristic equation (denominator of equation 5.13 and 5.14)

S 2 + (Γ + B)S + κ2 = 0 (5.21)

in conjunction with the dimensionless groups (equation 5.11), the controller gain values for a PD
controller can be explicitly defined to get a desired closed loop natural frequency wc and damping
ratio ζc .

100
 10

Fl /Fd (dB)
-10

-20

-30
-1 0 1 2
10 10 10 10
ω/ω
o
0
Phase (degrees)

-20

-40

-60

-80

-100
-1 0 1 2
10 10 10 10
ω/ωo

Figure 5-7: sim Closed-loop bode plot of the EM series elastic actuator with PD control. This is a
bode plot of the closed loop system with κ = 3 and ζc = 1. The zero from the controller adds the
apparent resonance near the controlled natural frequency.

The proportional gain is found from the definition of κ.

κ2 = Kp
wc2
=⇒ Kp = (5.22)
wo2

The derivative gain is found from the damping term of the characteristic equation.

2ζc κ = B + Γ
2ζc ωc mm − bm
=⇒ Kd = (5.23)
ks

As an example figure 5-7, shows a bode plot of the closed loop forward transfer function: κ = 3
and ζc = 1.

5.2.3 Impedance
Equation 5.14 shows the dimensionless form of output impedance. I rewrite it here for convenience

Fl (S) −S(S + B)
Zcl = = 2
ks xl (S) S + (Γ + B)S + κ2

At low frequencies the impedance is small. As input frequency increases, so too does the
impedance. In the limit, the output impedance levels off and is equal to the spring constant, ks , of
the force sensor. Equation 5.14 normalizes to ks and figure 5-8 shows an example of the impedance

101
 10

Increasing Γ
0

-10

-20
Fl /(ksXl )
-30

-40

-50

-60

-70
-1 0 1
10 10 10
ω/ωo

Figure 5-8: sim EM output impedance. The impedance of the actuator at small frequencies is low
and is equal to ks at high frequency. In this example κ = 4 and B = 0.1. For low frequencies,
the actuator has an equivalent mass characteristic. The equivalent mass is, meq = mκm . The higher
the controller gain the lower the impedance. As Γ increases there is an effective reduction in the
impedance at the controlled natural frequency.

for different values of Γ which represents the derivative gain.

There are four regions of interest in figure 5-8. The first is at very low frequency: very low,
mid-range near ωo , near ωc and high frequencies.
The natural damping of the system B is quite low but its effects are still just noticeable. For the
most part however, the effects of B can be neglected.
For the rest of the low frequencies up to the controlled natural bandwidth the impedance has
an equivalent mass characteristic because it is rising at 40dB/dec. This means that it has two free
zeros in the numerator2
Fl = meq s2 Xl (5.24)
meq is some fraction of the reflected motor inertia mm defined by the controller gain.
mm mm
meq = = 2 (5.25)
Kp κ

Therefore, the higher the gain, the lower the equivalent mass.
The break point or resonance is at κ. In figure 5-8, κ = 4.
As mentioned previously, at frequencies above the controlled natural frequency ωc , the actuator’s
output impedance is the spring constant of the sensor. All high frequency disturbances and shock
loads are filtered through the spring.
There are three ways to decrease the output impedance for the series elastic actuator.

• Increase the control gain. This drives the impedance resonance further away from the opera-
tional bandwidth.

2 Forall practical purposes, a zero with a very slow time constant can also be considered a free zero in the impedance
equation when trying to calculate the equivalent mass.

102
 F
l
x
l

k
s

40 dB/dec

ω
ω c

Figure 5-9: Rough characterization if EM impedance. As shown in the simulation, the impedance
at low frequencies is equivalent to a mass and is equal to the spring constant of the sensor at high
frequencies.

• Decrease the spring constant. This linearly lowers the impedance profile.

• Increase the derivate gain. As seen in figure 5-8, Γ’s major contribution is to reduce impedance
resonance at the controlled natural frequency.

It is important to remember that the key benefit of series elastic actuators is low impedance.
Controlled bandwidth can remain constant by increasing control gain proportionally to the reduction
in spring stiffness. Impedance however, is dramatically reduced when the spring stiffness is lowered.

5.2.4 Force Error Rejection

As described in chapter 3, coulomb friction and stiction are dramatically affected by the introduction
of the spring into the actuator. The compliant spring allows for an increase in controller gain. Stiction
is one of the main reasons why the actuator cannot respond to low forces and thus reduces dynamic
range.
Because of stiction, the lumped mass will not move until

Fm − Fl = Fm − Fd + Fe = Fs (5.26)

where Fe is the feedback force error and Ff is the force due to stiction. Since, stiction is a low
frequency phenomenon, we assume that the derivative term of the controller is negligible. The force
to the motor due to the error then becomes

Fm ≈ Kp Fe + Fd . (5.27)

Using equations 5.26 and 5.27 we can solve for the force error Fe .
1
Fe ≈ Fs (5.28)
Kp + 1

This shows that the force error due to stiction is related by the dimensionless group κ.
1
Fe ≈ Fs (5.29)
κ2

103
 1
κ 2 = Kp + 1
0.9 κ2 = Kp

0.8

stiction 0.7

0.6
/F

0.5
error

0.4
F

0.3

0.2

0.1

0
5 10 15 20 25
Kp

Figure 5-10: sim Inverse parabolic relationship of the force error due to stiction force. As control
gain, Kp , increases non-linear friction forces are dramatically decreased. The two curves represent
the esimate of κ2 . The solid line includes the +1 and the dashed line does not. In the physical
actuator, Kp = 25. As can be seen, at that level of gain there is very little difference between the
two curves.

Remember that κ is an approximate ratio between the resonant frequency of the mass-spring
system, omegao , and the controlled natural frequency, ωc . Increasing κ reduces the effect of stiction.
There are two ways to effectively reduce stiction: 1–increase ωc by increasing the gain or 2–decrease
ωo by reducing the spring constant. The relationship is inverse parabolic and can be visualized in
figure 5-10. Theoretically, to eliminate stiction we should have as soft a spring as possible. However,
part of the design is to evaluate an acceptable level of stiction reduction.
As the gain is increased, there is a rapid decrease in the noticeable stiction force in the system.
If we set our controlled natural frequency 5 times greater than the mass-spring resonance (Kp = 25)
as is done in the physical actuator, the apparent stiction is reduced to 5% of its original value.

5.3 Effect of Load Mass

Even though series elastic actuators have variable load conditions due to intermittent contact with
the environment, it is helpful to analyze the dynamics of a mass load moving in free space (figure
5-11). Ideally, the mass load is decoupled from the actuator dynamics such that the actuator looks
like a pure force source. I show that given the load has some minimum mass, then the actuator can
be considered a pure force source within its closed-loop bandwidth capabilities.

104
 Signal Power
Motor
ks Load
Mass Spring Mass
+ Fm
Motor
Fdesired Σ Kp + Kd s K a Amp mm ml
-
Fload

Figure 5-11: EM series elastic actuator model with load mass moving in free space. The load mass
defines the load motion xl .

5.3.1 Load Forces

Consider the force balance relationships for force in the spring, Fl , and motor mass motion, xm
derived previously for the laplace domain (equation 5.3).

Fl = ks (Xm − Xl )
Fm − Fl
Xm = (5.30)
mm s2

With the load mass moving in free space, I can also write the laplace domain equation for the
load motion, Xl . It is solely a function of the force in the spring, Fl .
Fl
Xl =
ml s2

For enhanced clarity and understanding, I eliminate the viscous damping on the motor mass.
Remember that the motor force Fm is a function of the control system.

Fm = (Kp + Kd s)(Fd − Fl )

Substitute the controller into the equation for Xm . Then use this result and the equation for Xl
to solve for the force in the spring, Fl as a function of desired force Fd .
Fl Kd s + Kp
=   (5.31)
Fd mm 2 mm
ks s + Kd s + Kp + 1 + ml

With two exceptions, equation 5.31 is the same as the closed-loop dynamic transfer function for
the fixed load case (equation 5.7). The first exception is the elimination of the viscous damping
term. Only controller damping remains. The second difference is in the s0 term of the characteristic
equation in the denominator. Instead of just being Kp + 1, there is and additional term, m ml , that
m

is the ratio of the motor mass to the load mass.

There are two important things to remember about the s0 term. The first is that Kp  1. The
proportional gain is high due to the fact that ks is low. As mentioned before, Kp + 1 ' Kp . This
leads to the second important point. The additional term, mml , is only significant for the case where
m

there is very little load inertia in comparison to the motor mass. Fortunately, this is hardly ever
true.
Typically these actuators are used in robots with articulated joints. Therefore, ml is a linear

105
 10

Fl /Fd (dB)
0

-10

-20

-30
-1 0 1 2
10 10 10 10
ω/ω
o
20
Phase (degrees)

-20

-40

-60

-80

-100
-1 0 1 2
10 10 10 10
ω/ωo

Figure 5-12: sim Force on a load that is free to move. This plot shows that the actuator can output
forces on a load mass even when it is free to move.

equivalent of a rotary inertia

Irot
ml =
r2
where Irot is the robot link inertia about the joint and r is the radius through which the actuator
force is converted into a torque. Practical experience with using the actuators on robots have shown
that ml ≈ mm . Therefore, again Kp dominates and m ml can be assumed small in comparison. At
m

low frequency, the actuator has good force response.

q
ks
Remember that the load mass resonates with the spring at a frequency ωl = ml and that ωl
is part of the non dimensionless group L where
ωl
L= .
ωo
I write the load force equation in non dimensional terms as

Fl (S) ΓS + κ2
= 2 (5.32)
Fd (S) S + (Γ + B)S + κ2 + L2

This equation is depicted in figure 5-12 with κ = 3, Γ = 6 and L = 1. At low frequency, there is
a small reduction in the output magnitude due to the load force. However, the force is in phase and
is close to the actual desired value.

106
 xl m eq xl
ks
Fd ml Fd ml

Figure 5-13: Control abstraction for the EM series elastic actuator. At frequencies below the con-
trolled bandwidth, the actuator can be considered a pure force source. Above the controlled band-
width, the dynamics of the equivalent mass meq (equation 5.25) and the spring come into play.

5.3.2 Load Motion

Assuming that the load mass is significant, we can refer back to the load motion in equation 5.30,
we can write the load motion as a function of the desired input force.
xl 1 Kd s + Kp
=   (5.33)
Fd ml s2 mm 2
s + K s + Kp + 1 + mm
ks d ml

There are two parts to equation 5.33. The first part is just the dynamics of the load mass as
a function of a simple force Fd = ml s2 Xl . The second part of the transfer function come from
equation 5.31 and represents the closed-loop characteristics of the actuator. Equation 5.33 verifies
that the actuator performs well as pure force source over over its closed-loop bandwidth as long as
load mass is significant.
In summary, a load mass feels a pure force at low frequency for the EM series elastic actuator.
As drive frequency increases, the dynamics of the actuator begin contribute to the force on the load.
In the very simplest model, it follows the diagram in figure 5-13. The actuator dynamics are a small
equivalent mass,meq (equation 5.25) and the physical elasticity followed by the load mass.

5.4 Physical Actuator Prototype

The theoretical work just presented has a practical side too. The following discussion elaborates on
design decisions made during the design and construction of the actuator prototype (figure 5-1). The
discussion focuses on component selection, with particular emphasis on picking the correct series
elasticity. Experimental results are also given which help to quantify actuator capabilities.

5.4.1 Component Selection

The design space for Series Elastic Actuators is very large. Besides geometry and topology there
are six major components: motor, amplifier, transmission, elasticity, sensor, and controller. Of all
five components, choosing the spring in the elasticity is the only part of the actuator which requires
unique perspective and is discussed in the next section.
The specifications for the motor, amplifier and transmission need to be done based on the force,
speed, and power requirements for a given task. These design requirements are not unique to Series
Elastic Actuators and would be done for any actuator.
As a helpful aid to the final operation, the motor and transmission should be selected with the
idea of keeping friction and motor saturation low. Motor friction bm which is seen through the
transmission with a multiiplication of N 2 , is kept at a minimum in the prototype actuator by using
a brushless DC motor. The actuator design also reduces drive friction by designing the center of
mass, center of friction, and center of stiffness to be colinear along the ball screw axis [64].
Transmission dynamics should also be kept to a minimum as these can be a primary limiting
factor in using high feedback gain [16]. We used a frameless motor configuration where the motor

107
Figure 5-14: EM series elastic actuator used in testing. The actuator uses a brushless DC motor
with a direct connection to a 2mm lead ball screw. The ball nut drives the springs in series with the
load. The springs are precompressed die compression springs. A linear potentiometer measures the
deflection of the springs and uses that signal for feedback to the controller.

magnets are mounted directly onto an extended ballscrew journal instead of using a coupling, a belt
drive, or gears.
The other limiting component to achieve high feedback gain is the sensor. The sensor needs to di-
rectly measure the spring deflection. This insures that the feedback measurement is a representation
of true force. Noise in the sensor is also detrimental to operation.

5.4.2 Choosing Sensor Spring Constant

Selecting the spring constant for the sensor needs to be balanced when trading off the effects of large
force bandwidth with low impedance and error rejection. I present some guidelines for choosing a
spring constant.

1. Select a motor and transmission based on the force, speed, and power requirements for the
given task. This will define the lumped motor mass, damping, and saturation characteristics
seen through the reduction.
2. Define an operational bandwidth, ωo , for which the actuator will need large forces. In other
words, define a required large force bandwidth profile (section 5.2.1). This profile is indepen-
dent of controller and depends solely on ωn and B. This places a lower bound on ks . Most
likely, the operational bandwidth may be a little greater than but close to the selected natural
frequency.
3. Minimizing impedance places an upper bound on ks . This is a function of the controller values
κ and Γ. Insure that the controller gains can be raised to acceptable levels of stiction and
impedance reduction.
4. It may be necessary to iterate. The non-dimensional equations will help to guide to know
whether to increase or decrease ks and how much effect it will have.

5.4.3 Actuator Characteristics

A prototype actuator is used to demonstrate and test series elasticity in the electro-mechanical
domain. Figure 5-14 shows the prototype actuator used in the experiments and table 5.4.3 shows
the physical properties of the actuator. The experimental actuator is used to verify the following.

108
 Parameter Value Units
Max Force 1350 N
Cont. Force 750 N
Min Force 4.0 N
Max Speed 0.25 m/s
Max Power 300 W
Cont. Power 150 W
Actuator Mass 1.13 kg
Dynamic Mass mm 128 kg
Spring Constant ks 286 kN/m
Damping bm 2500 Ns/m
ωo 7.5 Hz
ωc 30 Hz
ζ 0.2 no units
κ 4 no units
B 0.1 no units

Table 5.1: Physical properties of EM prototype actuator. Some of the values are calculated from
motor literature and others are measured.

• Open-loop dynamics

• Saturation

• Controlled bandwidth

• Power density

• Dynamic range

Where appropriate the experimental data is compared with theoretical models.

Open-loop Dynamics
As done previously, the actuator is tested for its open-loop properties. By fixing the load end of the
actuator and forcing the mass with motor torque, the actuator exhibits its second order behavior.
Figure 5-15 shows the loop gain of the actuator compared with simulated data.
There are three things that this plot demonstrates: natural frequency, damping and open-loop
gain.
The natural frequency of the actuator is at 7.5 Hz as calculated from the dynamic mass and spring
constant in table 5.4.3. This frequency is ωo or the frequency at which large force performance of the
actuator begins to drop off. The system is second order as the gain plot drops off at 20 dB/decade
and the phase goes to −180o. There are signal filters on the feedback from the sensor. However,
they are set high enough that their effects are not noticeable.
The damping term, bm , in table 5.4.3 was figured experimentally from figure 5-15. The mass-
spring resonance is underdamped with ζ = 0.2 [42].
The final item to notice is that the open loop gain is not unity. This deviation from unity is
compensated for in the control system gain discussed later.

Saturation
Because of the mass-spring resonance, the actuator is limited in its ability to control large forces. The
resonance defines the frequency at which the actuator begins to decrease in large force performance.
This constraint is independent of control system.

109
 0

-20

Fm/Fd
-40
Experimental
Simulation
-60
0 1 2
10 10 10

0
Phase (degree)

-50

-100

-150 Experimental
Simulation
-200
0 1 2
10 10 10
f (Hz)

Figure 5-15: real Open loop bode plot for electro-magnetic series elastic actuator with fixed end
condition. This is a simple second order lumped mass spring system with a natural frequency of 7.5
Hz. The dynamic mass is the motor mass as seen through the transmission and the spring is the
physical elasticity.

The maximum output force capability of the experimental actuator is 1350 N (300 lbs). In order
to test the theory of large force bandwidth, the actuator is commanded to oscillate through the
frequency spectrum at an amplitude of 1350 N with the output end of the actuator fixed. Figure
5-16 shows the results of the experiment. As predicted, the large force capability of the actuator is
limited. At 8 Hz the performance begins to drop off and follows a second order profile.
It is important to remember that the actuator can still produce some fraction of the maximum
output force at high frequency. The controller can command high frequency forces but its peak to
peak magnitude will be limited due to saturation.

Bandwidth

It is possible to find the closed-loop bandwidth of the actuator assuming small forces. Experimentally
this is found by again fixing the load of the actuator and sweeping the frequency spectrum. Figure
5-17 shows both the experimental and mathematical closed-loop model.
Experimentally, the -3dB point is at 35 Hz. The calculated controlled natural frequency is
ωc = 30Hz and the damping was fixed at ζ = 0.707. The derivative term of the controller introduces
a zero into the system that represents the apparent resonance in the closed loop system.

Power Density

In order to test the mechanical output power of the actuator, and to prove that the actuator can
operate as designed, I suspend myself hanging from the actuator in a chair with 60 lbs in my lap
(figure 5-18). This is referred to as the risky tester. The total combined load on the actuator is 300
lbs (1350 N).
The desired force in the actuator is set to just balance the gravity pull on the mass load. By
oscillating the desired force around the equilibrium point, the actuator moves up when it has greater
force and drops when it has less. In this configuration, the actuator can move a load mass with a
continuous mechanical power output of 150 W. The peak inertial load power is more than 300 W.

110
 5

-5

-10

Fl/Fsat (dB) -15

-20

-25

-30
Experimental
Simulation
-35
0 1 2
10 10 10
f (Hz)

Figure 5-16: real Experimental maximum force saturation for the electro-magnetic series elastic
actuator with fixed end condition. The actuator can only generate the maximum output force up
to the ωo , the mass-spring resonance frequency.

10

0
Fl/Fd (dB)

-10

-20
Experimental
Simulation
-30
0 1 2
10 10 10

0
Phase (degree)

-50

-100

-150
Experimental
Simulation
-200
0 1 2
10 10 10
f (Hz)

Figure 5-17: real Experimental closed-loop bandwidth for the electro-magnetic series elastic ac-
tuator with fixed end condition. The -3dB point is at 35Hz. The zero in the closed-loop equation
causes the apparent resonance.

111
Figure 5-18: Risky testing. The EM prototype actuator is suspended from the ceiling and loaded
with 300 lbs (me, seat and weights). The displacement of the spring is commanded so that it
compensates for the gravity pull on the load mass. With small variations in the set point, the
actuator moves the load up and down.

An itemized breakdown of the experimental actuator mass is given in table 5.4.3. The power
density of this specific actuator is on the order of 300W/kg. This power density is about 3 times
that of an direct drive actuator with equivalent force output [27] assuming no forced cooling of the
motor coils.

Dynamic Range
The actuator has a large dynamic range despite having a high power density. The minimum resolv-
able force is 4.0 N and it can output over 1350 N which gives a dynamic range over 300. Considering
the actuator has 2mm/rev reduction in the ball screw, the dynamic range is impressive [27]. The
limitations on the bottom end of force sensitivity is the noise floor of the sensor and static friction.
The dynamic range of the actuator is demonstrated by fully loading the actuator and setting it
in equilibrium with its load. The load is moved up and down with finger force. Figure 5-19 shows
the test conditions.

5.5 Case Study Summary

This chapter presented the first case study for series elasticity applied to an electro-magnetic motor
in series with a transmission. The chapter looks at both the mathematical model of the actuator
as well as a physical prototype. These models are used to discuss bandwidth, impedance, error
rejection and saturation bandwidth. Where appropriate experimental results are compared to the
mathematical models.

112
 Part mass (g) % of total
Motor and Magnet 580 51.3%
Ball Screw and Nut 130 11.5%
Thrust Bearing Block 115 10.2%
Springs (4) 60 5.3%
CFRP Pushrods 2) 26 2.3%
CFRP Rails (2) 26 2.3%
Spring Carriage Clamps (2) 44 3.9%
Motor Mount 44 3.9%
End Clamp 33 2.9%
Pushrod End 25 2.2%
Nut Rider 20 1.8%
Sensor 10 0.9%
Bushings (8) 8 0.7%
Screws and Socket Caps 4 0.4%
Simple Support Bearing 5 0.4%
Total 1130 100%

Table 5.2: Itemized mass distribution for EM series elastic actuator. This actuator is used in two
dynamically stable robots described in Chapter 2. The table shows that over 50% of the mass is
contained in the motor and magnet. In the design of series elastic actuators, it is important to keep
the percentage power producing element mass high in comparison to the overall actuator mass.

Figure 5-19: Fully loaded actuator demonstrating small force sensitivity. The actuator is fully loaded
with 300 lbs and has a force equal to gravity pull on the load mass. Finger force of approximately 1
lb can get the actuator to move up and down. This shows that the dynamic range of the actuator
is 300:1.

113
114
Chapter 6

Conclusions

Many of the key characteristics of series elastic actuators are described and demonstrated in this
thesis. Three models are used to explain the actuators. A minimal-complexity mathematical model
captures the essence of the actuator’s properties. This is followed by two case studies of physical
actuators with both mathematical models and physical prototypes. The first is a hydro-elastic
actuator consisting of a hydraulic piston and servo valve under PI control. The second is an electro-
magnetic motor with a linear transmission under PD control. Each of these models help to explain
closed-loop bandwidth, large force bandwidth, output impedance, internal error rejection, impact
tolerance and inertial load forces.

Closed-loop bandwidth
The fundamental limits in closed-loop force bandwidth for series elastic actuators are independent of
the spring stiffness that senses the force. The control system gain and the spring stiffness gain each
contribute to the loop gain of the closed-loop system. With a lower spring stiffness, the controller
gain can be increased a proportional amount to bring the loop gain to the desired phase and gain
stability margins. Series elastic actuators take the gain out of the spring and put it into the control
system.

Large force bandwidth

The ability of a series elastic actuator to oscillate at a force equal to the maximum steady state
force is limited due to force and velocity saturation of the actuator and low stiffness of the spring.
The motor in the actuator must move a significant amount in order to compress the spring and
generate the maximum force. While, low forces can still be generated to the full controlled closed-
loop bandwidth, the large force bandwidth is limited. Large force bandwith can be increased by
using a less compliant spring. Nevertheless, understanding the tradeoff of large force bandwidth is
important in designing and using series elastic actuators.

Output impedance
At low frequencies, the output impedance of a series elastic actuator is very small and increases
with increased frequency. In the limit at high frequencies, the output impedance is equal to the
stiffness of the physical spring in the actuator. The output impedance is significantly lower than if
the actuator used a stiff load cell. The low frequency impedance can be considered as a physically
equivalent first or second order system depending on the control system. This means that an
equivalent mass or damper, defined by the control system parameters, is effectively in series with
the actual spring connected to the actuator load. Overall, the lower the spring stiffness the lower the
output impedance. Having a very soft spring, and thus low output impedance, needs to be balanced
with the actuators requirements for the large force bandwidth.

115
Internal error rejection
There are internal non-linearities in any physical actuator that hinder closed-loop force control
performance. Closed-loop force control with a compliant sensor reduces the magnitude of these
errors. The high gain controller provides for a clean force output.

Impact tolerance
Series elastic actuators can successfully handle shock loading. Because of the spring in the system,
the actuator can minimize impact power from the environment and effectively spread an impulse
load over time to minimize peak forces. This protects the internal actuator components, the robot
structure, and the environment from damage.

Inertial load forces

Series elastic actuator behavior is altered when connected to a finite inertial load moving in free
space rather than when in contact with an external environment. This deviation from expected
performance is a function of the magnitude of the load inertia and actuator output impedance.
As the load mass increases, better actuator performance is realized. In the limit as mass becomes
infinite, the actuator behaves exactly as predicted in the closed-loop bandwidth case. Regardless,
the dynamics of the actuator are significantly decoupled from that of the load.

6.1 Further Work

There are many topics that deserve further investigation that relate to the understanding and design
of series elastic actuators.

6.1.1 Actuator Scaling

This thesis has presented several models of actuators. Each of the models have been generalized by
reformulating them with non-dimensionless groups. This gives the models flexibility in understanding
how to scale a given actuator system to a different situation.
This thesis does not, however, explain when it is beneficial to use one actuator system over
another. For example, at low forces it may be that a simple cable drive mechanism, with no
elasticity, has the force and power density required to complete the task. Using simple feed-forward
current control on the motor may be sufficient for that case. When the force and power requirements
increase to a certain level, a series elastic actuator with an EM motor transmission is required. At
very high force and power density levels, a hydro-elastic actuator is appropriate.
The above scenario is only a guess at the appropriate use of different embodiments of series
elasticity. It would be helpful if there were force and power density models of these actuator system
to aid as rules of thumb when the actuators are used robot construction.

6.1.2 Springs
Springs are at the core of the series elastic actuators. This thesis focused attention only on linear
springs. However, in order to more fully understand the actuators, the springs themselves need to be
investigated. Spring hysteresis, materials, nonlinear stiffening springs, space constraints, geometry
and life cycle are all important. I briefly elaborate on the first two.

• Hysteresis and materials – Most physical springs have some hysteresis. The die compres-
sion springs used in this thesis have very little hysteresis and were assumed to be linear in
the control. It is suspected that with increased spring hysteresis, the actuator becomes more
difficult to control. For example rubber springs are often used for vibration isolation because

116
 of inherent hysteresis and damping properties. Air springs are often used for high force appli-
cations. However, they too have thermodynamic hysteresis properties. Quantifying the effects
of hysteresis would be an important step to using new materials and springs in the actuators.

• Non-linear stiffening springs – Biological springs (tendons) are stiffening. One theory
behind this is that at low forces or during contact tasks the low stiffness helps maintain
stability. At much higher forces, low stiffness is unnecessary and may be even undesirable.
Some initial theoretical and experimental work has been done previously in applying non-linear
stiffening springs to the actuators in fixed base robots [32, 55, 61]. In this thesis, I show that
there is a fundamental trade off between large force bandwidth and impedance when choosing
a spring stiffness. Low impedance requires a soft spring whereas a soft spring reduces the
large force bandwidth. Using a non-linear stiffening spring may be a bridge between these
requirements.
Stable control of series elastic actuators with non-linear springs may be as simple as using a
scheduling control gain that is a function of spring deflection. As the spring stiffness increases,
the control gain decreases proportionally. This keeps the loop gain of the system constant and
within stability margins.

6.1.3 EM velocity mode control

The motor of EM series elastic actuator used in this thesis is current controlled. Under current
control, it is easy to analytically demonstrate the large force bandwidth limitations of the actuator.
Due to space constraint requirements for the actuator, a tachometer and encoder were eliminated
from the design which also forced the move to current control.
It is strongly suspected that if the motor were run with velocity control, that the overall actuator
performance could potentially improve. Velocity control would make the closed-loop model more
similar to that of the hydro-elastic actuator model by partially masking some of the effects of the
motor inertia. The same fundamental limits would apply to the large force bandwidth. However,
the small force sensitivity and error rejection would get better.

6.1.4 Control system design

This thesis focused attention on simplicity of control system design. P, PI and PD control operating
on the force error signal adequately demonstrated the properties of the actuators.
There exist many other more complex control systems that could incrementally enhance series
elastic actuator performance. Proper use of state feedback could eliminate the zeros in the closed-
loop system. The use of feed-forward could help further eliminate internal errors such as stiction.
Adaptive control could be used to better understand actuator load conditions and accommodate for
load variations.
The use of these and many other control system possibilities, may help to reach closer to the
fundamental closed-loop bandwidth limits of the actuator and improve the decoupling of the actuator
dynamics from that of the load. However, if using a linear compliant spring, regardless of the
simplicity or complexity of control system, the large force bandwidth is constrained by the open-
loop characteristics of the actuator.

6.2 Summary
This thesis investigates the closed-loop force control characteristics of series elastic actuators. A
minimal mathematical model and two physical case studies help demonstrate and quantify actuator
properties.
The ideas of force control and using compliance in actuation are not new. Chapter 2 presented
detailed background material on the general ideas of force control and discusses the capabilities of
present-day actuator technology to accomplish good force control. This chapter also introduces the

117
related work on actuators which use compliance including series elastic actuators. It also describes
some of the robots that use previous versions of series elastic actuators and robots the use the
actuators designed in this thesis.
The principles of series elastic actuators are quite general. Chapter 3 describes a general model
of series elastic actuators with minimum complexity. Using this model, I explain all of the key
characteristics of using series elasticity as a compliant load sensor for feedback in closed-loop force
control. I discuss the ideas of closed-loop bandwidth in conjunction with large force bandwidth. I
explain actuator output impedance, internal error rejection and impact tolerance. Finally, I look at
the case of the actuator being connected to an inertial load.
The principles of series elastic actuators discussed in Chapter 3 are applied in Chapter 4 to a
case study of a hydraulic piston and servo valve with a series elastic force control system, which is
called a hydro-elastic actuator. A theoretical model of the actuator is compared with experimental
results.
In a second case study, Chapter 5 applies results of chapter 3 to an electro-magnetic motor with
a gear reduction. Again, the theoretical model is compared with experimental results. Both case
studies demonstrate good closed-loop bandwidth, low output impedance, shock tolerance and large
dynamic range.
This thesis has helped to lay a foundation for understanding, characterizing and designing series
elastic actuators in both the hydro-mechanical and electro-mechanical power domains. It shows that
actuators under closed-loop force control using a compliant sensor can achieve good force bandwidth
and have low output impedance. The actuators show great promise in advancing the state-of-the-are
of real world robotics.

118
Bibliography

[1] Farhad Aghili, Martin buehler, and John M. Hollerbach. Sensing the torque in a robot’s joints.
Mechanical Engineering, pages 66–69, Sept. 1998.

[2] M. L. Aisen, H. I. Krebs, N. Hogan, F. McDowell, and B. T. Volpe. The effect of robot-assisted
therapy and rehabilitative training on motor recovery following stroke. Archives of Neurology,
54:443–6, 1997.

[3] R. McNeill Alexander. Elastic Mechanisms in Animal Movement. Cambridge University Press,
1988.

[4] Andrew Alleyne. Nonlinear force control of an electro-hydraulic actuator. Japan/USA Sympo-
sium on Flexible Automation, 1:193–200, 1996.

[5] H. Asada and J.J. Slotine. Robot Analysis and Control. John Wiley and Sons, New York, 1985.

[6] H. Asada and K. Youcef-Toumi. Direct Drive Robots: Theory and Practice. MIT Press, Cam-
bridge, 1987.

[7] Haruhiko Asada. Manufacturing robotics: Basic issues and challenges. IFAC, pages 319–330,
1996. San Francisco, CA.

[8] David Scott Barrett. Propulsive efficiency of a flexible hull underwater vehicle. PhD thesis,
Massachusetts Institute of Technology, May 1996. Ocean Engineering.

[9] B. E. Berlinger. Evoloid – a new concept in high ratio gearing. AGMA Paper, 5(109.37), Oct.
1975.

[10] David Brock and Woojin Lee. Dynamic model of a linear actuator based on polymer hydrogel.
Journal of Intelligent Material Systems & Structures, 5(6):764–771, Nov. 1994.

[11] Rodney A. Brooks, Cynthia Breazeal, Matthew Marjanovic, Brian Scassellati, and Matthew
Williamson. The Cog Project: Building a Humanoid Robot. Springer-Verlag, 1999. To appear
in a Springer-Verlag Lecture Notes in Computer Science Volume.

[12] F. Conrad and C. Jensen. Design of hydraulic force control systems with state estimate feedback.
IFAC, pages 307–312, 1987. Munich, Germany.

[13] John J. Craig. Introduction to Robotics: Mechanics and Control. Addison-Wesley, 1989.

[14] Samuel Hunt Drake. Using compliance in lieu of sensory feedback for automatic assembly. PhD
thesis, Massachusetts Institute of Technology, 1978. Mechanical Engineering.

[15] S.D. Eppinger and W.P. Seering. Understanding bandwidth limitations in robot force control.
IEEE International Conference on Robotics and Automation, 2:904–909, 1987. Raleigh, NC.

[16] Steven D. Eppinger. Modeling Robot Dynamic Performance for Endpoint Force Control. PhD
thesis, Massachusetts Institute of Technology, 1998. Mechanical Engineering.

119
[17] Masato Hirose et. al. Legged mobile robot and a system for controlling the same. United States
Patent, (#5,455497). Issued October 3, 1995.

[18] James M. Gere and Stephen P. Timoshenko. Mechanics of Materials. PWS-KENT Publishing
Company, Boston, MA., 1990.

[19] Layton C. Hale and Alexander H. Slocum. Design of anti-backlash transmissions for precision
position control systems. Precision Engineering, 16(4):244–258, 1994.

[20] David Halliday and Robert Resnick. Fundamentals of Physics. John Wiley & Sons, New York,
1988.

[21] H. Hanafusa and H. Asada. A robotic hand with eleastic fingers and its application to assembly
process. IFAC Symposium on Information and Control Problems in manufacturing Technology,
pages 127–138, 1977. Tokyo.

[22] Blake Hannaford and Jack Winters. Actuator properties and movement control: Biological and
technological models. In Winters and Woo, editors, Multiple Muscle Systmes: Biomechanics
and Movement Organization, pages 101–120. Springer-Verlag, 1990.

[23] B. Heinrichs, N. Sepehri, and A.B. Thornton-Trump. Position-based impedance control of an

industrial hydraulic manipulator. IEEE Control Systems Magazine, 17(1):46–52, 1997.

[24] Kazuo Hirai, Masato Hirose, Yuji Haikawa, and Toru Takenaka. The development of honda
humanoid robot. International Conference on Robotics and Automation, 1998.

[25] Jessica K. Hodgins and Marc H. Raibert. Biped gymnastics. International Journal of Robotics
Research, September 13 1989.

[26] N. Hogan. Impedance control: An approach to manipulation: Part i - theory, part ii - imple-
mentation, part iii - applications. J. of Dynamic Systems, Measurement and Control, 107:1–24,
1985.

[27] John Hollerbach, Ian Hunter, and John Ballantyne. A comparative analysis of actuator tech-
nologies for robotics. In Robotics Review 2, pages 299–342. MIT Press, 1991.

[28] Won Hong. Robotic catching and manipulation using active vision. Master’s thesis, Mas-
sachusetts Institute of Technology, September 1995. Mechanical Engineering.

[29] Russel D. Howard. Joint and Actuator Design for Enhanced Stability in Robotic Force Control.
PhD thesis, Massachusetts Institute of Technology, September 1990.

[30] S.C. Jacobsen, C.C. Smith, D.K. Backman, and E.K. Iversen. High performance, high dexterity,
force reflective teleoperator i. ANS topical meeting on robotics and remote systems, 2:180–185,
1990. Washington, DC, November.

[31] S.C. Jacobsen, C.C. Smith, K.B. Biggers, and E.K. Iverson. Behavior based design of robot
effectors. Proc. of the Fourth Int’l Symp. on robotics research, pages 41–55, 1989.

[32] Arrin Katz. The design and applicaiton of a nonlinear series compliance actuator for use in
robotic arms. Master’s thesis, Massachusetts Institute of Technology, September 1999. Mechan-
ical Engineering.

[33] Glenn K. Klute, Joseph M. Czerniecki, and Blake Hannaford. Mckibben artificial muscles:
Pneumatic actuators with biomechanical intelligence. IEEE/ASME Conference on Advanced
Intelligent Mechatronics, Sept. 19-22, 1999. Atlanta, GA.

[34] Benjamin T. Krupp. Design and control of a planar robot to study quadrupedal locomotion.
Master’s thesis, Massachusetts Institute of Technology, 2000. Mechanical Engineering.

120
[35] Akhil J. Madhani. Design of teleoperated surgical instruments for minimally invasive surgery.
PhD thesis, Massachusetts Institute of Technology, 1998. Mechanical Engineering.
[36] Matthew Mason. Compliant motion. In Robot Motion: Planning and Control, pages 305–322.
MIT Press, 1982.

[37] Matthew Mason. Compliance and force control for computer controlled manipulators. IEEE
Transactions on Systems, Man, and Cybernetics, SMC-11(6):419–426, June 1981.

[38] Thomas H. Massie. Design of a three degree of freedom force-reflecting haptic interface. 1993.
Undergraduate Thesis, Massachusetts Institute of Technology.
[39] John Bryant Morrell. Parallel coupled micro-macro actuators. PhD thesis, Massachusetts
Institute of Technology, 1996. Mechanical Engineering.
[40] Charles J. Murray. All-in-one hydraulics. Design News, pages 84–89, Nov. 15, 1999.
[41] Gunter Niemeyer. Using Wave Variables in Time Delayed Force Reflecting Teleoperation. PhD
thesis, Massachusetts Institute of Technology, 1996. Aeronautics and Astronautics.
[42] Katsuhiko Ogata. Modern Control Engineering. Prentice Hall, 1990.

[43] Henry M. Paynter. Analysis and Design of Engineering Systems. MIT Press, Cambridge, MA,
1960.
[44] M. Pelletier and M. Doyon. On the implementation and performance on impedance control on
position controlled robots. IEEE Conference on Robotics and Automation, pages 1228–1233,
1994. Atlanta, GA.
[45] Ron Pelrine, Roy Kornbluh, Qibing Pei, and Jose Joseph. High-speed electrically actuated
elastomers with strain greater than 100%. Science, 287:836–839, 4 February 2000.
[46] Gill A. Pratt and Matthew M. Williamson. Series elastic actuators. IEEE International Con-
ference on Intelligent Robots and Systems, 1:399–406, 1995.

[47] Jerry Pratt. Virtual model control of a biped walking robot. Master’s thesis, Massachusetts
Institute of Technology, May 1995. Electrical Engineering and Computer Science.
[48] Jerry Pratt and Gill Pratt. Intuitive control of a planar bipedal walking robot. ICRA, 1998.

[49] Jerry Pratt, Ann Torres, Peter Dilworth, and Gill Pratt. Virtual actuator control. IROS, 1996.
Osaka, Japan, Nov. 3-9.

[50] Jerry E. Pratt. Exploiting Inherent Robustness and Natural Dynamics in the Control of Bipedal
Walking Robots. PhD thesis, Massachusetts Institute of Technology, June 2000. Electrical
Engineering and Computer Science.
[51] M. H. Raibert and J. J. Craig. Hybrid position/force control of manipulators. Journal of
Dynamic Systems, Measurement, and Control, 102, 1981.
[52] Marc. H. Raibert. Legged Robots That Balance. MIT Press, Cambridge, MA., 1986.
[53] David W. Robinson and Gill A. Pratt. Force controllable hydro-elastic actuator. IEEE Inter-
national Conference on Robotics and Automation, April 24-28, 2000. San Francisco, CA.
[54] David W. Robinson, Jerry E. Pratt, Daniel J. Paluska, and Gill A. Pratt. Series elastic actuator
development for a biomimetic walking robot. IEEE/ASME AIM99, Sept 20-23, 1999. Atlanta,
GA.
[55] J. Kenneth Salisbury. Kinematic and force Analysis of Ariticulated hands. PhD thesis, Stanford
University, May 1982. Department of Mechanical Engineering.

121
[56] J. Kenneth Salisbury. Making graphics physically tangible. Communications of the ACM,
42(8):75–81, August 1999.

[57] J. Kenneth Salisbury. The heart of microsurgery. Mechanical Engineering, pages 46–51, Dec.
1998.

[58] J. Kenneth Salisbury and Mandayam A. Srinivasan. Proceedings of The First PHANToM User’s
Group Workshop. MIT AI Laboratory, 1996.

[59] J.K. Salisbury, W.T. Townsend, B.S. Eberman, and D.M. DiPietro. Preliminary design of a
whole-arm manipulation system (wam). Proceedings 1988 IEEE International Conference on
Robotics an d Automation, 1988. Philadelphia, PA, April 1988.

[60] K. Salisbury. Active stiffness control of a manipulator in cartesian coordinates. 19th IEEE
Conference on Decision and Control, pages 83–88, Dec. 1980.

[61] Vinay K. Shah. Design and control of a nonlinearly compliant robotic finger. Master’s thesis,
Massachusetts Institute of Technology, September 1997. Mechanical Engineering.

[62] Thomas B. Sheridan. Telerobotics. Automatica, 9(5):487–507, 1989.

[63] Thomas B. Sheridan. Telerobotics, Automation, and Human Supervisory Control. MIT Press,
Cambridge, MA, 1992.

[64] Alexander H. Slocum. Precision machine design. Society of Manufacturing Engineers, 1992.

[65] Jean-Jacques E. Slotine and Weiping Li. Applied Nonlinear Control. Prentice-Hall, Cambridge,
MA., 1991.

[66] C.C. Smith, S.C. Jacobsen, L.A. Robins, D.W. Wilcox, and S.J. Bohn. Design and control
of electromechanical actuation systems. ASME DSC Modelling and Control of Compliant and
Rigid Motion Systems Winter Annual Meeting, pages 137–144, 1991. Dec 1-6.

[67] Mark W. Spong and M. Vidyasagar. Robot Analysis and Control. John Wiley & Sons, 1989.

[68] S. Sugano, S. Tsuto, and I. Kato. Force control of the robot finger joint equipped with mechan-
ical compliance adjuster. International Conference on Intelligent Robots and Systems, pages
2005–2012, 1992.

[69] Thomas Sugar and Vijay Kumar. Design and control of a compliant parallel manipulator for
a mobile platform. ASME Design Engineering Technical Conference, September 13-16, 1998.
Atlanta, GA.

[70] W.T. Townsend and J.K. Salisbury. Mechanical bandwidth as a guideline to high-performance
manipulator design. IEEE International Conference on Robotics and Automation, pages 1390–
1395, 1989. Scottsdale, AZ.

[71] M.S. Triantafyllou and G.S. Triantafyllou. An efficient swimming machine. Scientific American,
272(3):64–70, March 1995.

[72] Dieter Vischer and Oussama Khatib. Design and development of high-performance torque-
controlled joints. IEEE Transactions on Robotics and Automation, 11(4):537–544, 1995.

[73] D.L. Wells, E.K. Iversen, C.C. Davis, and S.C. Jacobsen. An investigation of hydraulic actuator
performance trad-offs using a generic model. IEEE International Conference on Robotics and
Automation, pages 2168–2173, 1990. May 13-18.

[74] Daniel E. Whitney. Historical perspective and state of the art in robot force control. Interna-
tional Journal of Robotics Research, 6(1), 1987.

122
[75] Matthew M. Williamson. Series elastic actuators. Master’s thesis, Massachusetts Institute of
Technology, June 1995.
[76] Matthew M. Williamson. Robot Arm Control Exploiting Natural Dynamics. PhD thesis, Mas-
sachusetts Institute of Technology, June 1999.

123