2014 14th IEEE-RAS International Conference on

Humanoid Robots (Humanoids)

November 18-20, 2014. Madrid, Spain

Feedback Parameter Selection for Impedance

Control of Series Elastic Actuators
Ye Zhao, Nicholas Paine and Luis Sentis
AbstractThe interest of series elastic actuators (SEAs) for
legged robots has recently increased to achieve compliant
interactions and efficient gaits. However, control of legged robots
with SEAs is difficult due to the need to design controllers that
take into account both torque and impedance feedback loops.
The work presented here addresses this issue by proposing a
critically-damped fourth order system gain selection criterion
for a cascaded SEA control structure with inner torque and
outer impedance feedback loops. Velocity filtering and feedback
delays are taken into consideration for stability and impedance
performance analysis. We observe and analyze the interdependence between torque and impedance feedback gains to achieve
the desired closed loop performance. Based on this analysis we
derive a simple gain design criterion to maximize the tracking
performance of SEAs. Our final goal is to maximize the output
impedance capabilities of SEAs in order to fulfill a wide range of
application needs. In contrast to low impedance design studies,
we focus here specifically on achieving the highest possible
impedance gains of SEAs. Finally, experiments using our UTSEA are conducted to verify our proposed approach. This
study serves as a stepping stone towards utilizing and designing
humanoid robots with SEA actuators for mobile behaviors and
interaction with cluttered and unknown environments.

I. I NTRODUCTION
Cascaded control structures for series elastic actuators
(SEAs) [1] have gained attention during the past decade [2],
[3], [4]. This type of structure utilizes multiple nested control
loops to achieve robustness and disturbance rejection. Existing studies proposed one type of cascaded structure with an
inner torque loop for controlling SEA dynamics and an outer
impedance loop for modulating high-level tasks [5], [3]. More
recently, investigations have been made to add an inner-most
motor velocity feedback loop nested inside the torque loop
[2], [3], [4]. This velocity loop enables the motor to behave
as a velocity source, eliminating the need to model drivetrain
friction. The results in [5] analyzed various cascaded loops
including torque, velocity and position feedback loops.
Impedance control is suitable for dynamic interaction control
between a robot and its environment [6]. In general, control
designers often choose low impedance as a design target of
SEAs for compliant performance [3], [4]. However, different
tasks require different impedances and high impedance for
SEAs will benefit stiff tasks. For instance, without feedforward compensation, legged robots require high stiffness to
counteract the effects of gravity. Our target is to maximize
the impedance range of series elastic actuator, i.e., the Zwidth [7], [8]. Since low or near-zero impedance has already
been explored successfully [9], [1], we focus more on the
achievable high impedance of SEAs, which is rarely explored
978-1-4799-7174-9/14/$31.00 2014 IEEE

Fig. 1. UT-SEA and Hume-SEA. This figure shows two types of series
elastic actuators. The upper one is the high-performance UT-SEA test bed.
The lower one is our Hume bipedal robot with series elastic actuators.

to the authors best knowledge. It should be kept in mind

that the impedance of SEAs still can not reach as high as
that of rigid actuators. In fact, SEA impedance is limited by
the user-defined virtual stiffness at low frequencies and by
spring stiffness at high frequencies [3].
Series elastic actuators have high-order system dynamics.
Designing controller gains for their several cascaded loops
is challenging and there is still a lack of convincing methodologies for this type of control design. Most existing results rely on empirical tuning [1], [5]. In [3], the authors
use a passivity criterion to derive controller gain ranges.
However, these ranges are inequalities and gain parameters
are coupled together, which makes controller gain selection
still complex. Recently, we proposed a critically-damped
gain selection criterion to achieve high impedance for rigid
actuators while incorporating effects from signal filtering and
feedback delays [10]. However, the field of gain design for
SEAs has not yet been explored. In this study, we propose a
fourth order gain selection criterion, which uniquely defines
both impedance and torque gains. This criterion only allows

999

Combining Equations (4) and (5) and defining q = qm qj ,

we have the relationship between spring deflection and motor
angle
q(s)
Ij s2 + bj s
r(s) =
=
.
(6)
qm (s)
Ij s2 + bj s + k
Combining with Equation (1), spring force can be calculated
k (s) = kq(s) = kr(s)qm (s).
Fig. 2. SEA Model Scheme. qm , qj represents motor and joint position,
respectively. k is the spring stiffness. Ij , Im is joint and motor inertia,
respectively. bj , bm are joint and motor damping coefficient, respectively.
Im and bm are mapped to the joint coordinates by multiplying by the square
of the gear reduction.

one variable to be determined by the designer, which is

convenient for selecting controller gains.
In this study, we design a fourth order gain selection criterion
to achieve critically-damped performance of series elastic
actuators. Making multiple gains depend on a natural frequency reduces the degree of freedom for gain selection to
one and this natural frequency represents the SEA impedance.
To the best of our knowledge, this is the first time a
critically-damped gain selection criterion has been proposed
for SEAs. This criterion is supported by a detailed analysis of
filtering and feedback delays and their influence in cascaded
impedance control of SEAs. Finally, our theoretical results
are validated through experimental implementations.

Since the motor current im and motor torque m is related

by
m (s)
= = N k ,
(8)
im (s)
with drivetrain efficiency (constant for simplicity, ignore
dynamic model of drivetrain losses), gear speed reduction N
and motor torque constant k , the control plant PF (s) is

PF (s) =

r(s)k
k (s)
=
.
2
im (s)
Im s + bm s + r(s)k

II. M ODELING OF S ERIES E LASTIC ACTUATORS

k = k(qm qj ).

k = Ij qj + bj qj .

(2)

Note that this model only considers the effects of viscous

joint friction; analysis of the effects of other types of friction
is left for future work. The load plant PL (s) is then
qj (s)
1
PL (s) =
=
.
2
k (s)
I j s + bj s

(3)

By Equations (1) and (2), we have the transfer function from

motor position qm to joint position qj .
qj (s)
k
=
.
2
qm (s)
Ij s + bj s + k

(4)

Motor torque m is represented as

m = Im qm + bm qm + k(qm qj ).

(5)

PF ( 1 + C)
k (s)
=
.
des (s)
1 + PF CeT s

(10)

The torque feedback loop has a feedback delay eT s and

a PD compensator C = K + B Q d s, where Q d is the
filtering for torque derivative.
Q d =

(1)

For the joint side, we assume external force d = 0 (i.e.,

spring torque is equal to load torque) and have

(9)

Then, from Figure 3, the closed-loop torque control plant PC

is

PC (s) =
In this section, we model a series elastic actuator that is
subject to two nested control loops: an inner torque loop
and an outer impedance loop. First, let us consider the SEA
dynamics. As shown in Figure 2, the spring force k is

(7)

2f d
,
s + 2f d

(11)

where f d is the filter cut-off frequency. A feedforward term

is designed to convert from desired torque des to motor
current im as shown in Figure 3. By Equations (3) and (10),
we have the following transfer function

PF ( 1 + C)
qj
= PL PC =
.
des
(1 + PF CeT s )(Ij s2 + bj s)

(12)

The impedance feedback has the following form

des (s) = Kq (qdes eTqs s qj ) Bq eTqd s Qqd sqj , (13)
where eTqs s and eTqd s represent the feedback delays in
stiffness and damping loops, respectively. The first order
low-pass filter Qqd for joint velocity has the same form as
Equation (11) with a cut-off frequency fqd . Using PL and PC
in Equations (3) and (10), we can derive the system closedloop transfer function PCL from qdes to qj ,

1000

Embedded Torque Control Loop

Embedded Damping Servo Loop

High-Level Stiffness Servo Loop

Fig. 3. Joint-Level SEA Control Block Diagram. The inner torque loop has proportional and derivative feedback and a feedforward loop with a mapping
scaler 1 . The outer impedance loop has stiffness and damping feedback. Our distributed controller proposes to embed the damping feedback at the low
level while maintaining the stiffness loop for high-level tasks. Delays in each loop are labeled as eT s . A first order low-pass filter is applied to both
velocity and torque derivatives. The motor has a current input im . k is the spring torque. PC represents the embedded torque control loop.

A. Critically-damped fourth order system gain criterion

qj (s)
PCL (s) =
qdes (s)
Kq PC PL
1 + PC PL (eTqd s Bq Qqd s + eTqs s Kq )
Kq (1 + K + B Q d s)
,
(14)
=
P4
i
i=0 Di s
=

with coefficients

For second order rigid actuators, impedance control gains

[10] can be designed based on the well-known criticallydamped criterion. For high order systems, there is no welldefined critically-damped criterion. However, high order systems can be often be represented as the product of first
and second order systems. To achieve a critically-damped
response for Equation (14), we represent this fourth order
system by two multiplied second order systems [11]
(s2 + 21 1 s + 12 )(s2 + 22 2 s + 22 ).

D4 =Im Ij /k,
D3 =(Ij bm + Im bj )/k + Ij B Q d eT s ,

(15)

Note that we ignore feedback delay and filtering effect when

designing controller gains. This allows the system to be
a fourth order system instead of a higher one. Now we
have four design parameters 1 , 2 , 1 , 2 , which are used to
design four gains Kq , Bq , K , B . First, we set 1 = 2 = 1
in Equation (15) for a critically-damped purpose. Second,
let us assume 2 = 1 for simplicity. We plan to study an
optimal pole placement in future work. Now we can define
the natural frequency fn of the fourth order system.

D2 =Ij (1 + eT s K ) + Im + bj B Q d eT s
+ B Bq eTqd s Qqd Q d + bj bm /k,
D1 =bj (1 + eT s K ) + bm + B Q d Kq eTqs s
+ eTqd s (1 + K )Bq Qqd ,
D0 =eTqs s (1 + K )Kq .
This transfer function is a sixth order system since the low
pass filters Qqd and Q d increase the order by two in total.
An important issue to notice is that there is a zero in the
numerator of Equation (14). This zero is caused by the torque
derivative term. This induced zero will shorten the rise time
but also cause an overshoot in step response. However, it
does not influence system stability, which is determined by
the characteristic polynomial in the denominator.

1 = 2 , n = 2fn .

(16)

By comparing denominators of Equations (14) and (15), we

have the following gain selection criterion equations.

III. G AIN SELECTION OF SERIES ELASTIC ACTUATOR

Nested impedance and torque loops make controller gain
design more challenging. This section proposes a method to
design controller gains based on a critically-damped criterion.
1001

Ij bm + Im bj + Ij B k
= 4n ,
Im Ij
k(Ij (1 + K ) + Im + B (bj + Bq )) + bj bm
= 6n2 ,
Im Ij
k(bj + Bq )(1 + K ) + k(bm + B Kq )
= 4n3 ,
Im Ij
(1 + K )kKq
= n4 .
(17)
Im Ij

=Tqs =0.5 ms, Tqd =8 ms, fqd =50 Hz, f d =100 Hz

T =Tqd =0.5 ms, Tqs =8 ms, fqd =50 Hz, f d =100 Hz T(c)
(a)T =Tqs =Tqd =0.5 ms, fqd =50 Hz, f d =100 Hz (b)

Slighly
under-damped

Amplitude
Amplitude

1.2
1

f =18 Hz, with zero

0.8

fn =16 Hz, with zero

0.6

fn =12 Hz, with zero

fn =14 Hz, with zero

fn =18 Hz, w.o. zero

0.4
0.2
0
0

Critically-damped
0.1

0.15

Time (seconds)

1.2

1
0.8
0.6
fn =18 Hz, with zero

0.4

0.3

0
0

fn =14 Hz, with zero

Time (second)

0.1

0.15

0.2

Time
(seconds)
Time
(second)

0.25

0.8
0.6
fn =18 Hz, with zero
fn =16 Hz, with zero

Unstable

fn =14 Hz, with zero

0.2

fn =12 Hz, with zero

0.05

0.4

fn =16 Hz, with zero

0.2

fn =12 Hz, w.o. zero

0.25

1.4

1.2

fn =14 Hz, w.o. zero

0.2

Largely
under-damped

1.4

fn =16 Hz, w.o. zero

0.05

1.6

1.6

Overshoot due to
extra zero

Amplitude
Amplitude

1.4

Amplitude
Amplitude

1.6

0.3

0
0

fn =12 Hz, with zero

0.05

0.1

0.15

0.2

Time
Time(seconds)
(second)

0.25

0.3

Fig. 4. SEA Step Response with Effects from Feedback Delays and An Extra Zero. These subfigures demonstrate that large impedance feedback
delays deteriorate step response performance. Comparing subfigures (b) and (c), we observe that system stability is more sensitive to damping feedback
delays than its stiffness counterpart. Note that, in subfigure (a), the larger fn is, the larger the overshoot, which seems counterintuitive. However, if observed
closely, the solid magenta line with the largest fn already shows distortion and its phase margin value is 36.4 , smaller than other three cases. Also, to
analyze the effect of zero in Equation (14), we simulate step responses without this zero, shown in dashed lines of subfigure (a). By comparison, we can
observe this extra zero induces an overshoot.
TABLE I
C RITICALLY- DAMPED G AIN S ELECTION RULE
Frequency
(Hz)
fn = 12
fn = 14
fn = 16
fn = 18
fn = 20

Impedance Gains
(Nm/rad, Nms/rad)
Kq = 65
Bq = 0.46
Kq = 83
Bq = 0.76
Kq = 103
Bq = 1.02
Kq = 124
Bq = 1.26
Kq = 148
Bq = 1.49

Torque Gains
(A/Nm, As/Nm)
K = 1.18
B = 0.057
K = 1.80
B = 0.067
K = 2.56
B = 0.077
K = 3.45
B = 0.087
K = 4.48
B = 0.097

made simulations where a gain scale, GS, was used between

adjusted gains (Kia , Bia ) and nominal gains (Kin , Bin ),
i {, q}. Adjusted gains are the real gains used in the
simulations while nominal gains are the gains determined
from the fourth order system gain selection criterion. For
instance, we choose the gain set of fn = 14 Hz in Table I
as nominal gains (Kin , Bin ), i {, q}. Adjusted gains
(Kia , Bia ), i {, q} are obtained by a gain scale below

Phase
Margin
49.1
47.0
43.6
39.9
36.4

GS =

These nonlinear equations can be solved using Matlabs

fsolve function. Table I shows five natural frequency cases.
As fn increases, all four gains increase simultaneously, which
is consistent with our intuition that increasing torque (or
impedance) gains will increase system torque (or impedance)
bandwidth. Note that, when calculating phase margins, we
incorporate filters fvd = 50 Hz, f d = 100 Hz and feedback
delays T = Tqs = Tqd = 0.5 ms.
Now let us consider how impedance feedback loop delays
affect the SEA step response. Since torque feedback is
regarded as the inner loop, it always has the smallest delay,
T = 0.5 ms in our case, than the other two impedance
loops. As observed in Figure 4, system stability is more
sensitive to damping delays than stiffness delays. The case
with T d = Tqd = 0.5 ms, Tqs = 8 ms is analogous to our
proposed distributed control architecture in [10].
B. Trade-off between impedance control and torque control
The motivation first came from gain tuning work on our
bipedal robot, which uses a similar cascaded control architecture in Figure 3. We observed that, when torque gains
increased or when impedance gains decreased, the robots
motions became less stable. To validate this phenomenon, we

Kqn
Ka
=
.
Kn
Kqa

Similarly, the damping gains are related to GS by

GS =

B a
Bq n
=
,
B n
Bq a

which guarantees multiplication of cascaded proportional (or

damping) torque and impedance gains are the same for both
normal and adjusted cases, i.e.,
Ka Kqa = Kn Kqn ,
Ba Bqa = Bn Bqn .

(18)

A similar observation is obtained in [12] where increasing

the inner loop controller bandwidth reduces the stable range
of impedance parameters. A trade-off exists between a high
torque bandwidth to provide accurate torque tracking and
a low torque bandwidth to increase the stable range of
impedance gains. However, the experimental validation in
[12] is not completely persuasive because impedance gains
are not reduced when torque gains are increased. This means
that the multiplication of two cascaded gains increases. In this
case, system stability obviously decreases. Instead, our study
assumes a constant gain multiplication and then focuses on
the trade-off between torque and impedance gains. Subfigure
(a) in Figure 5 represents the nominal case with GS = 1. As
observed, when torque gains increase and impedance gains

1002

(a)

60

Ideal Response without Filter, without Delay

(b)

nominal GS = 1

(c)
70

1.2

40
30
20

GS = 0.4
When GS > 1, phase margin
monotonously decreases
as GS increases

10
0

10
0

6
8
GainScale
Scale GS
Gain
(GS)

10

12

nominal GS = 1

fn =12 Hz
fn =16 Hz
fn =20 Hz
fn =24 Hz
fn =28 Hz

60

GS = 0.4

0.8
GS =4.0
GS =3.6
GS =3.2
GS =2.8
GS =2.4
GS =2.0
GS =1.6
GS =1.2
GS =0.8
GS =0.4

0.6

0.4

0.2

0
0

0.05

0.1

0.15

Time
(seconds)
Time
(seconds)

0.2

0.25

Phase
Margin (degree)
Phase Margin (degree)

Amplitude
Amplitude

Phase
Margin [degree]
Phase
Margin
(degree)

50

50
40
30

At GS = 1, four cases achieve

largest phase margin and
one achieves second largest

20
10
0
10
0

Gain Scale GS
Gain
Scale (GS)

10

12

Fig. 5. Different Gain Scales between Impedance and Torque Gains. Subfigure (a) shows that our critically-damped gain selection is a sub-optimal
solution since its phase margin is the second best. As subfigure (b) shows, increasing GS will slow down the rise time and cause larger overshoot. This
means larger GS deteriorates system stability. However, decreasing GS may decrease phase margin as well. GS = 0.4 has a phase margin around 34
in subfigure (a) and its step response has a distortion in subfigure (b), although behaving more over-damped. Subfigure (c) samples a range of natural
frequencies for the nominal gains and all of them demonstrate similar results. Filtering and delays are ignored to focus on effects of different gain scales.

decrease (i.e., GS > 1), system shows larger oscillatory

step response in subfigure (b) and its phase margin becomes
smaller in subfigure (a). This result validates our observation
that using larger torque gains deteriorates SEA phase margin
based stability. Note that, when GS 6= 1, there is no welldefined fn , which is only proposed in the case of criticallydamped gains, i.e., GS = 1. Indeed, in our implementations
we only use the critically-damped gains due to their resulting
optimal behaviors. The purpose of considering GS 6= 1 is to
study the trade-off phenomenon.

Nz2 =Im k + kbm (Tf + Tf v ) + kk

(Tf + (B + K Tf ))(Bq eTqd s + Kq Tf v eTqs s ),
Nz1 =bm k + Bq kk (1 + K )eTqd s + Kq kk
(Tf v + Tf + (B + K (Tf + Tf v )))eTqs s ,
Nz0 =Kq kk eTqs s (1 + K ),
and the denominator coefficients

Impedance control is suitable for dynamic interaction between robots and their environment [6]. Since legged robots
inevitably have contact impact, it is necessary to study how
SEA impedance behaves within different frequency ranges
based on the gain selection criterion proposed above.

Dz5 =Im Tf Tf v ,
Dz4 =Im (Tf v + Tf ) + Tf v Tf bm ,
Dz3 =Im + bm (Tf + Tf v ) + Tf v k(Tf
+ k (B + K Tf )eT s ),
Dz2 =(bm + Tf k + kk (B + K Tf )eT s )

IV. SEA IMPEDANCE ANALYSIS WITH FILTERING AND

+ Tf v k(1 + K k eT s ),

FEEDBACK DELAY EFFECTS

Dz1 =k(1 + K k eT s ), Dz0 = 0.

In this section, we derive the SEA impedance transfer

function based on the control block diagram in Figure 3
and analyze different frequency range characteristics with
considerations of filtering and delays. Recall that the higher
the natural frequency fn , the higher the SEA impedance. Let
us first derive the SEA impedance transfer function.
A. SEA impedance transfer function
To study the impedance of SEA dynamics, we choose joint
velocity qj as input and joint torque j as output. Given zero
desired joint position qdes , the impedance transfer function
Z(s) = j (s)/(sqj (s)) is formulated as follows.
Z(s) =

P4
Nzi si
j (s)
,
= P5i=0
i
sqj (s)
i=0 Dzi s

with numerator coefficients

Nz4 =Im Tf Tf v k,
Nz3 =k(Im (Tf + Tf v ) + Tf Tf v bm ),

(19)

Note that, joint inertia Ij and damping bj at the load side

are not included in Z(s) since they are regarded as the interacted environment. When analyzing this SEA impedance, we
incorporate effects of filtering and feedback delays. As far as
we know, results in literature are still missing that analyze the
filtering and delay effects on SEA impedance of a cascaded
control structure with nested inner PD torque control and
outer impedance control. Results in [3] analyzes another
typical cascaded control architecture with a proportionalintegral torque loop and an inner-most motor velocity loop.
However, they only consider sampling rate delay instead of
large delays such as communication bus delays. Also, they
neglect integral gains during SEA impedance analysis for
simplicity. Instead, our study uses the exact transfer function
in Equation (19) without any approximations.
B. Effects of filtering and feedback delays on SEA impedance
SEA impedance results are shown in Figure 6. Four cases are
studied with or without velocity filtering and feedback delays:

1003

25

Physical Stiffness k/j

Virtual Stiffness Kq/j

5
0

No Filter, No Delay
With Filter, No Delay
No Filter, With Delay
With Filter, With Delay

45

Phase twisted

90
0

10

15
10
5
0
5
10
0

Passivity stability
violation region

10

Frequency
(Hz)
Frequency
(Hz)

10

Small disparity
due to filtering

20

Physical Stiffness k/j

Virtual Stiffness K /j
No Filter
fqd =80 Hz, f d =150 Hz
=50 Hz, f

=100 Hz

=20 Hz, f

=50 Hz

qd
qd

d
d

25

Medium disparity
due to delays

20
15

10
Physical Stiffness k/j
Virtual
5 Stiffness K /j

vir

SEA Impedance
different
delays
(c) Impedance
effected by with
feedback
delays
(no filtering)

vir

0
No Delay
T =T =T =1 ms
qs 5 qd

T =5 ms, T =2 ms, T =1 ms
qs10
qd

T =10
ms,
T =3 ms, T =1 ms
qs 0
qd

45

No phase twisted

Phase (deg)

Converge to
virtual stiffness
10
15

10
0

Phase
Phase(degree)
(deg)

Converge to
spring stiffness

Magnitude
(dB)
Magnitude
(dB)

20

Phase
Phase(degree)
(deg)

Magnitude
(dB)
Magnitude
(dB)

25

Magnitude (dB)

SEA Impedance
witheffected
different
frequency
SEA Impedance
with Virtual
Spring
Damper
and filtering
(b) Impedance
byfilter
filtercut
(nooff
delay)
(a) SEA impedance
with/w.o
filter
and delay

45

Phase twisted

90

90
0

10

10

Frequency (Hz)
(Hz)
Frequency

10

10

10

Frequency (Hz)

10

Frequency (Hz)

Fig. 6. SEA Impedance Effected by Filtering and Feedback Delays. In subfigure (a), the yellow and blue dashed lines represent impedance of physical
spring stiffness k and a virtual stiffness gain, respectively. The red dashed line is the ideal SEA impedance without filtering and delay. At low frequencies,
SEA impedance approaches the virtual stiffness, similar to that in [3]. At medium and high frequencies, it converges to another impedance asymptote.
Subfigure (b) analyzes the effect of the filter while subfigure (c) analyzes the effect of feedback delays. Similarly, the sensitivity to different feedback
delays can be analyzed but is not discussed here due to space limitations. A natural frequency of fn = 30 Hz is used for all simulations in this figure
with Kq = 293.6N m/rad, Bq = 2.49N ms/rad, K = 11.71A/N m, B = 0.146As/N m.

(I) Zi (j), ideal impedance without filtering and delays;

(II) Zf (j), impedance with filtering only; (III) Zd (j),
impedance with delays only; (IV) Zf d (j), impedance with
both filtering and delays. At low frequencies, it is observed
that SEA impedance approaches a virtual stiffness in any of
the four cases above (if feedback delays exist, eTqs j
1, eT j 1 as 0)
Nz0
Kq k ( 1 + K )
=
, (20)
0 j Dz1
j (1 + K k )

lim Zc (j) = lim

where c {i, f, d, f d}. We keep j in final expression

for indicating the existence of a 20 dB/dec asymptote.
The low frequency impedance Zc (j) behaves like a pure
virtual stiffness impedance Kq /j with a scaling k ( 1 +
K )/(1 + K k ). This scaling exists for any PD-type cascaded impedance controller. However, k 1 is normally
very small. When k K is larger, Zc (j) approaches to
Kq /j, more behaving like a pure spring Kq /j.
In the case of high frequencies, the impedance magnitude
approaches or twists around an asymptote, depending on
specific cases. First, let us consider the ideal case (I), i.e.,
without filtering and delays. We have Dz5 = Dz4 = 0.
Nz2
k(Im + k B Bq )
=
lim Zi (j) = lim
+
+ j Dz3
j Im
(21)
which is a constant stiffness impedance but differs from
the pure spring stiffness. Figure 6 (b)-(c) shows that the
red dashed line represents an ideal SEA impedance and it
approaches another asymptote at high frequencies.

Since the complex number eTqd s periodically rotates along

the unit circle, SEA impedance twists around the impedance
of pure spring stiffness at high frequencies. This can be
checked at the high frequency range of both magnitude and
phase plots in Figure 6 (c).
Third, in Case (II) with filtering only, i.e. Tqs = Tqd = T =
0.
Nz4
k
lim Zf (j) =
(23)
=
+
jDz5
j
which has a 20 dB/dec asymptote as shown in Figure 6 (b).
No curve twist shows up since the limit value is a constant.
By comparing Figure 6 (b) and (c), it is obvious to notice
that delay has significantly large effect on the deviation from
ideal SEA impedance.
Practical factors from filtering and delay indeed influence
practical impedance. Based on the passivity stability criterion
[7], [3], [13], the phase value should remain within 90
and 90 (the blue region). As we see, the SEA impedance
has phase value smaller than 90 , which violates the
passivity criterion. However, a non-passive system does not
mean the system is unstable. For instance, all four cases in
Figure 6 (a) violate the passivity condition, but two cases
without filtering have phase margins larger than 54 while
two cases with filtering have phase margins larger than 16 .
Moreover, filtering and delays make the magnitude and phase
curve twisted at high frequency range. These effects are
commonly ignored in existing literatures but they certainly
play important roles in real implementations.
C. Effects of load inertia on SEA impedance

Second, let us consider case (III) with delay only, i.e. Tf v =

Tf = 0. We also have Dz5 = Dz4 = 0.

lim Zd (j) = lim

Nz2
j Dz3

In this subsection, we study how load inertia affects SEA

impedance. In practice, the port of interaction with the
environment contains a load inertia. Thus, a load inertia Ij s
k(Im + k B Bq eTqd s ) is added into Equation (19), i.e., Zl (j) = Z(j) + Ij s.
=
Then, SEA impedance behaves like a spring-mass instead of
j Im
(22) a pure spring at high frequencies. Specifically, it is dominated
1004

SEA Impedance with different load inertias

No Filter, No Delay
Magnitude (dB)

30
20
10
0
10

Load Inertia 0.045 kgm2

Phase (deg)

90

Load Inertia 0.09 kgm2

Load Inertia 0.135 kgm2

45
0
45
90

10

10

Frequency (Hz)

10

SEA Impedance with different load inertias

Magnitude (dB)

30
20

0
10
90

Phase (deg)

Fig. 8.
Bode Diagrams with Different Natural Frequencies fn .
This figure shows that increasing fn increases closed-loop bandwidth. At
low frequencies, experiments match simulations quite well. At frequencies
around resonant frequency, experiments show a larger resonant peak and a
slightly larger bandwidth than simulations.

10

Load Inertia 0.045 kgm

Load Inertia 0.09 kgm2

45

Load Inertia 0.135 kgm

1.6

0
45

1.4

90
1

10

Frequency (Hz)

GS increases

1.2

10

Amplitude
Amplitude

10

Fig. 7.
Effects of Load Inertia on SEA Impedance. This figure
shows three different load scenarios. At high frequencies, SEA impedance
approaches the load inertia impedance Ij s. Filtering and delays provides
very limited effect since two figures perform very similar.

1
0.8
0.6

Simu GS =1
Simu GS =1.3
Simu GS =1.9
Exp GS =1
Exp GS =1.3
Exp GS =1.9

0.4

by the inertia in Figure 7. Four different loads are simulated.

For Equation (19). Z(j) 0 as +. Thus we have

0.2
0
0

lim Zl (j) = lim (Z(j) + Ij j) = Ij j

0.05

0.1

0.15

Time
(seconds)
Time
(second)

0.2

0.25

Thus, Ij j is the asymptote at high frequencies in Figure 7

when an inertia is attached to the load side. Thus, the SEA
impedance will exhibit the high impedance caused by the
load inertia at high frequencies.
V. E XPERIMENTAL VALIDATION
In this section, we use our UT-SEA testbed [14] to validate
theoretical and simulation results. All of the controller gains
are designed based on the gain selection criterion. In Figure 8,
we observe that the larger the natural frequency is, the higher
the bandwidth of the closed-loop system is. Simulations and
experiments match each other with the exception of a slight
discrepancy at high frequency range. Refer back to Table I
for the normalized values of virtual stiffness and damping
used in this test. With respect to the trade-off between torque
gains and impedance gains, Figure 9 demonstrates that when
GS > 1, system step response exhibits larger overshoot and
slower rise time. This validates our conjecture that increasing
torque gains while decreasing impedance gains deteriorates

Fig. 9. Comparison of Simulations and Experiments with Different

Gain Scales Experimental data matches simulations with a small discrepancy on the overshoot. This discrepancy probably comes from unmodelled
dynamics and Coulomb friction. It validates that increasing torque gains and
decreasing impedance gains deteriorate system stability.

system stability. As to the discrepancy in these two tests,

one main factor is that our SEA uses a timing belt, which
induces inevitable elasticity. However, it is difficult to model
its elasticity and incorporate it into controller design.
For dynamic legged locomotion [15], torque tracking under
impact dynamics is important for stable and agile walking.
By implementing an impulse test, we show how accurately
our controller reacts to the varied desired torques under
external impulse disturbances. As shown in Figure 10, when
a ball free falls from 20 cm height and hits the arm with
an impulse force, the SEA actuator settles down promptly
and recovers after approximately 0.3 seconds. Joint torque
tracking is quite accurate with only minor distortion around
peak values.

1005

R EFERENCES
[1] G. A. Pratt, P. Willisson, C. Bolton, and A. Hofman, Late motor
processing in low-impedance robots: Impedance control of serieselastic actuators, in American Control Conference, 2004. Proceedings
of the 2004, vol. 4. IEEE, 2004, pp. 32453251.

20 cm
height

Joint
JointPosition
Angle (degree)
ArmTorque
Torque(Nm)
(Nm) Arm
(degree)

[2] M. Hutter, C. D. Remy, M. A. Hoepflinger, and R. Siegwart, Efficient

and versatile locomotion with highly compliant legs, Mechatronics,
IEEE/ASME Transactions on, vol. 18, no. 2, pp. 449458, 2013.
[3] H. Vallery, J. Veneman, E. Van Asseldonk, R. Ekkelenkamp, M. Buss,
and H. Van Der Kooij, Compliant actuation of rehabilitation robots,
Robotics & Automation Magazine, IEEE, vol. 15, no. 3, pp. 6069,
2008.

Impulse Response
1

[4] N. L. Tagliamonte, D. Accoto, and E. Guglielmelli, Rendering viscoelasticity with series elastic actuators using cascade control, in
Robotics and Automation (ICRA), 2014 IEEE International Conference
on. IEEE, 2014.

0
1
2
3
0.1

0.1

0.2

Time (second)

0.3

[5] M. Mosadeghzad, G. A. Medrano-Cerda, J. A. Saglia, N. G.

Tsagarakis, and D. G. Caldwell, Comparison of various active
impedance control approaches, modeling, implementation, passivity,
stability and trade-offs, in Advanced Intelligent Mechatronics (AIM),
2012 IEEE/ASME International Conference on. IEEE, 2012, pp. 342
348.

0.4

6
Desired
Measured

4
2

[6] N. Hogan, Impedance control: An approach to manipulation: Part

iiimplementation, Journal of dynamic systems, measurement, and
control, vol. 107, no. 1, pp. 816, 1985.

0
2
0.1

0.1

0.2

Time
Time(second)
(second)

0.3

0.4

[7] J. E. Colgate and J. M. Brown, Factors affecting the z-width of a

haptic display, in Robotics and Automation, 1994. Proceedings., 1994
IEEE International Conference on. IEEE, 1994, pp. 32053210.

Fig. 10. Impulse Response of UT-SEA. A ball is dropped from a height

(20 cm) and exerts an impulse force on the arm end-effector. The maximum
angle deviation is around 2.5 degrees. The arm recovers its initial position
within 0.3 seconds. Joint torque exhibits accurate torque tracking.

[8] T. Boaventura, G. A. Medrano-Cerda, C. Semini, J. Buchli, and D. G.

Caldwell, Stability and performance of the compliance controller of
the quadruped robot hyq, in Intelligent Robots and Systems (IROS),
2013 IEEE/RSJ International Conference on. IEEE, 2013, pp. 1458
1464.

VI. C ONCLUSION AND D ISCUSSIONS

This study proposed a controller gain design method for the
fourth order SEA system and analyzed SEA impedance at
different frequency ranges. By considering the trade-off between torque gains and impedance gains, we demonstrate our
critically-damped gain selection is optimal, and increasing
torque gains while decreasing impedance gains will deteriorate system stability. Meanwhile, filtering and feedback
delays are incorporated into our framework. Additionally, our
method is not limited by the conservative passivity criterion.
Our on-going work focuses on the following aspects: (I)
Systematically quantify the effects of filtering and delays
on closed-loop stability and SEA impedance. Additionally,
we are exploring the fundamental causes of the phase margin sensitivity discrepancy analogous to that in [10]; (II)
Since critically-damped gain selection achieves near-optimal
stability in terms of system phase margin, we will explore
its fundamental reasons; (III) Study a novel stability criterion which removes conservative assumptions for passivity
condition. A good reference is the stability Routh-Hurwitz
criterion proposed in [16]; (IV) Implement the fourth order
gain selection criterion on our Hume bipedal robot. The gain
selection criterion proposed here will serve as a foundation
of our long term goal of implementing Cartesian impedance
control on dynamic legged locomotion on rough terrain [15].

[9] K. Kong, J. Bae, and M. Tomizuka, Control of rotary series elastic

actuator for ideal force-mode actuation in humanrobot interaction
applications, Mechatronics, IEEE/ASME Transactions on, vol. 14,
no. 1, pp. 105118, 2009.
[10] Y. Zhao, N. Paine, and L. Sentis, Sensitivity comparison to loop
latencies between damping versus stiffness feedback control action in
distributed controllers, in ASME 2014 Dynamic Systems and Control
Conference. American Society of Mechanical Engineers, 2014.
[11] F. Petit and A. Albu-Schaffer, State feedback damping control for a
multi dof variable stiffness robot arm, in Robotics and Automation
(ICRA), 2011 IEEE International Conference on. IEEE, 2011, pp.
55615567.
[12] M. Focchi, G. A. Medrano-Cerda, T. Boaventura, M. Frigerio, C. Semini, J. Buchli, and D. G. Caldwell, Robot impedance control and
passivity analysis with inner torque and velocity feedback loops, arXiv
preprint arXiv:1406.4047, 2014.
[13] C. Ott, A. Kugi, and G. Hirzinger, On the passivity-based impedance
control of flexible joint robots, Robotics, IEEE Transactions on,
vol. 24, no. 2, pp. 416429, 2008.
[14] N. Paine, S. Oh, and L. Sentis, Design and control considerations for
high-performance series elastic actuators, IEEE/ASME Transactions
on Mechatronics, vol. 19, no. 3, pp. 10801091, 2014.
[15] Y. Zhao and L. Sentis, A three dimensional foot placement planner for
locomotion in very rough terrains, in Humanoid Robots (Humanoids),
2012 12th IEEE-RAS International Conference on. IEEE, 2012.
[16] T. Hulin, R. G. Camarero, and A. Albu-Schaffer, Optimal control for
haptic rendering: Fast energy dissipation and minimum overshoot, in
Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International
Conference on. IEEE, 2013, pp. 45054511.

1006