Control Engineering Practice 58 (2017) 307318

Contents lists available at ScienceDirect

Control Engineering Practice

journal homepage: www.elsevier.com/locate/conengprac

Design and control of an active knee orthosis driven by a rotary Series

Elastic Actuator$
Wilian M. dos Santos a, Glauco A.P. Caurin a,b, Adriano A.G. Siqueira a,b,n
a
b

Department of Mechanical Engineering, Engineering School of So Carlos, University of So Paulo, 13566-590 So Carlos, Brazil
Center for Robotics of So Carlos and Center for Advanced Studies in Rehabilitation, University of So Paulo, Brazil

art ic l e i nf o

a b s t r a c t

Article history:
Received 23 March 2015
Received in revised form
16 September 2015
Accepted 17 September 2015
Available online 1 October 2015

Active orthosis is one of the main research topics in the eld of motor recovery. This paper deals with the
design and control of an active knee orthosis driven by a customized rotary Series Elastic Actuator (SEA).
The proposed actuator includes a DC motor, a worm gear and a customized torsion spring. Since the
elastic element is the most important component in SEA design, a nite element analysis of the spring is
performed to meet the specic requirements for knee assistance. Torque and impedance control are
implemented to ensure secure interaction with the patient and to enable new strategies for rehabilitation. The torque controller, cascaded with an inner motor velocity control loop, is based on / criterion
to achieve good system performance with relation to parametric uncertainties and external disturbances.
The impedance control is implemented using a PD position controller in cascade with the torque controller, where the outer position controller determines the desired torque according to position and
velocity errors and impedance parameters. A variable impedance control strategy is then implemented to
show the possibility to regulate the impedance of the knee joint during walking. Experiments considering the interaction between the subject and the active orthosis are performed to evaluate the
proposed controllers.
& 2015 Elsevier Ltd. All rights reserved.

Keywords:
Rehabilitation robotics
Active orthosis
Series Elastic Actuator
Robust torque control
Impedance control

1. Introduction
The use of robotic devices for rehabilitation of neurological
patients is increasing rapidly due to the importance of functional
exercises that stimulate motor cortex and promote motor recovery
(Ferris, Sawicki, & Domingo, 2006). Studies suggest that rehabilitation of post-stroke patients was intensied with robotassisted therapy (Kwakkel, Kollen, & Krebs, 2008; Prange, Jannink,
Groothuis-Oudshoorn, Hermens, & Ijzerman, 2006). The advantages of robotic therapy compared with traditional ones also
include the ability to evaluate patient progress constantly through
objective measures, and the possibility of customizing the treatment according to the patient's level of commitment.
In general, robotic therapy includes a combination of exercises

This work was supported by So Paulo Research Foundation (FAPESP) under

Grants 2011/04074-3 and 2013/14756-0. An earlier version of this paper was presented at the 19th World Congress of the International Federation of Automatic
Control (IFAC 2014).
n
Corresponding author at: Department of Mechanical Engineering, Engineering
School of So Carlos, University of So Paulo, 13566-590 So Carlos, Brazil.
Fax: 55 1633739402.
E-mail addresses: [email protected] (W.M. dos Santos),
[email protected] (G.A.P. Caurin), [email protected] (A.A.G. Siqueira).

http://dx.doi.org/10.1016/j.conengprac.2015.09.008
0967-0661/& 2015 Elsevier Ltd. All rights reserved.

involving passive, active-assisted or active-resisted movements.

This combination of exercises can be obtained through the use of
impedance control, proposed by Hogan (1985), which allows to
congure the dynamic interaction between the device and the
patient. For example, an active orthosis can provide support torque
during gait training in a assist-as-needed basis. During remaining
part of the gait, the orthosis need to present a low impedance
behavior, so as to be fully compliant with the patient's actions. The
impedance control can also be used to impose a controlled resistance to the patient movement aiming to strengthen muscle
groups.
To effectively assist human motion and, at the same time,
guarantee patient safety, rehabilitation robots must satisfy certain
requirements such as precise and large torque generation, with a
bandwidth that approximates muscle movement. It is also essential to ensure a backdrivable behavior, characterized by a low
mechanical impedance. Traditional stiff and high-precision actuators do not meet these critical requirements (Ham, Sugar,
Vanderborght, Hollander, & Lefeber, 2009; Santis, Siciliano, Luca, &
Bicchi, 2008). A simple and effective solution, initially proposed in
Pratt and Williamson (1995), is the Series Elastic Actuator (SEA)
concept, where elasticity is intentionally introduced in series between a gear-motor and the load. This conguration allows decoupling the gear-motor inertia and other nonlinearities from the

308

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

output and isolates the drivetrain from shocks introduced by the

load. Another important feature is that the elastic element can be
used as a torque sensor considering the linear relationship between spring deection and torque.
A rotary SEA to assist the movement of lower limbs was presented in Kong, Bae, and Tomizuka (2009). It consists of a geared
DC motor, a helical torsion spring and two rotary potentiometers
used to detect the position of the output shaft and the deformation of the spring. In the proposed conguration, the spring is
directly placed between the gear-motor and the human joint,
therefore subjected to large torques. Helical torsion springs able to
support large torques are usually stiff for the considered application. Stiff springs are less sensitive to small torques, resulting in
lower torque control accuracy. Moreover, their nonlinearities are
not negligible. The spring constant values for gait assistance
usually lie in the range from 100 to 300 N m/rad (Carpino, Accoto,
Sergi, Tagliamonte, & Guglielmelli, 2012), with the maximum torque in the range from 10 to 100 N m (Kong, Bae, & Tomizuka, 2012;
Lagoda, Schouten, Stienen, Hekman, & van der Kooij, 2010; Sergi,
Accoto, Carpino, Tagliamonte, & Guglielmelli, 2012; Stienen et al.,
2010).
From these considerations, a new rotary SEA model was proposed by the same authors in Kong et al. (2012). In this new
conguration, the spring is inserted between the worm gear and
the output gears, thereby enabling the use of a spring with low
stiffness. The disadvantage of this conguration is that the nonlinearities associated with the output gears compromise the delity of measured torque, increasing the uncertainties in the system.
A solution adopted by some researchers is to propose an arrangement of linear springs so as to obtain a torsion elastic element (Tsagarakis, Laffranchi, Vanderborght, & Caldwell, 2009;
Yoon et al., 2005). This approach allows the insertion of elastic
elements with low stiffness directly between the gear-motor and
the load. However, a linear behavior in the torque versus angle
relationship is difcult to be obtained. Another solution is the
development of customized elastic element (Carpino et al., 2012;
Lagoda et al., 2010). In addition to allowing the elastic element to
be connected to the load in a direct-drive conguration, this approach may reduce some problems such as residual deection,
hysteresis and a non-linear behavior in the torque versus angle
relationship that negatively affect accurate torque estimation and
consequently control performance.
Considering these concepts, a new customized torsion spring
conguration for rotary SEA is proposed in this paper. The torsion
spring design was optimized using Finite Element Method (FEM)
simulations in order to satisfy the specic requirements to partially support the knee joint exion/extension during physical
therapy in individuals with low motor impairment. The value
15 N m, which corresponds to 60% of the peak torque from the gait
pattern, is dened as design requirement for the maximum torque
of the actuator. Robust / torque and impedance controllers are
designed to ensure secure interaction with the patient and good
system performance. The performance of the controllers is evaluated by frequency response analysis, where it can be veried that
a suitable bandwidth was achieved. Experimental results obtained
from the implementation of a variable impedance control strategy
during walking are presented. The design of / controllers for
robot-aided rehabilitation and their performance; as well as the
design of the customized torsion spring are the main contributions
of this paper.
This paper is organized as follows: Section 2 describes the
design requirements; Section 3 presents in detail the mechanical
design of the rotary SEA; Section 4 presents the torque and impedance controllers design; Section 5 presents the experimental
results obtained from the active knee orthosis; Section 6 presents
the conclusions.

2. Design requirements
The design requirements are based on body mass normalized
data described in Kirtley (2006) for gait cycles. Considering that
the maximum power exerted by the knee joint is 0.739 W/kg, with
a maximum torque of 0.365 N m/kg, and that the active knee orthosis should be able to supply 60% of the peak torque from the
gait pattern of a healthy person with approximately 70 kg, the new
robotic device must provide a torque assistance up to 15 N m. The
minimum torque bandwidth is determined by the Power Spectral
Density (PSD) of knee joint torque. Regarding that more than 95%
of the PSD of knee joint torque is in the frequency range between
0 and 5 Hz a minimum bandwidth of 5 Hz is dened as a requirement to torque control.
The elastic element must be carefully designed, since it is the
most important component in the SEA. The spring constant is
dened through selection procedures described in Robinson
(2000). Basically, they dene lower and upper bounds for spring
constant based on desired large torque bandwidth and low output
impedance, respectively. First, consider the simplied model of
Fig. 1 which includes the equivalent inertia, Jeq, and damping, Beq.
At this point, these values consider only the inertia and damping
of the motor seen through the transmission (the complete values
of Jeq and Beq are dened in (10) and (11), respectively). The torque
applied to the load at the frequency domain, l (s ), is found as

l (s ) =

ks (Jeq s2 + Beq s )
ks

(
s
)

l (s ),
m
Jeq s2 + Beq s + ks
Jeq s2 + Beq s + ks

(1)

where m is the torque generated by the motor, l is the load position, and ks is the spring constant.
The ability of the actuator to produce large torques is limited in
frequency by the maximum torque which the motor can generate.
To dene the large torque bandwidth, the load position is considered xed and the torque from the motor is set to its maximum
value max (maximum continuous torque from motor datasheet).
Hence, the transfer function from max to the maximum output
torque, lmax , is given by

lmax (s )
ks
=
.
max (s )
Jeq s2 + Beq s + ks

(2)

The saturation frequency, i.e., the frequency at which the above

system begin to fall off, is dened from (2) as

0 =

ks
.
Jeq

(3)

Therefore, the higher the spring constant, the greater the large
torque bandwidth. Eq. (3) is used to dene the lower bound for the
spring constant. Fig. 2(a) shows the frequency responses for (2)
considering three values of spring constant (100 N m/rad,
200 N m/rad, and 300 N m/rad) and the parameters of the selected
motor.
To analyze the output impedance, a simple proportional controller is dened ( m = Kp (d l )). Assuming a constant desired
output torque ( d = 0 ), the output impedance is given by

Fig. 1. Schematic of rotary Series Elastic Actuator.

10

60

50
Magnitude (dB)

Magnitude (dB)

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

10
20
30

ks = 100 Nm/rad

40

ks = 200 Nm/rad

50

ks = 300 Nm/rad

60 1
10

40
30
20

ks = 100 Nm/rad

10

ks = 200 Nm/rad

0
1

10

309

10

10

Frequency (Hz)

10 1
10

ks = 300 Nm/rad
0

10

10

10

Frequency (Hz)

Fig. 2. The closed-loop large torque bandwidth and actuator output impedance. (a) Bode plot for large torque transfer function. (b) Bode plot for impedance transfer
function.

ks (Jeq s2 + Beq s )
l (s )
=
.
2
l (s )
Jeq s + Beq s + ks (1 + Kp )

(4)

Fig. 2 (b) shows the frequency responses for (4). The controller
gain Kp is dened so that the controlled natural frequency,
c = ks (1 + KP ) /Jeq , results in 5 Hz, the minimal desired closed
loop bandwidth. Note that decreasing ks linearly lowers the impedance prole. Also, at high frequencies, the output impedance
converges to the spring constant. Considering this result, a spring
constant of 200 N m/rad is dened as target value for the design. It
results in a large torque bandwidth of 0 = 3.3 Hz and a near 0 dB
output impedance for low frequencies.
Finally, the actuator must be as compact and lightweight as
possible. Therefore, it is specied a maximum weight of 2.5 kg.

3. Mechanical design
The mechanical design was conceived in order to obtain a
compact and lightweight architecture. All housing parts are made
of aluminum for the purpose of reducing weight. The nal assembly of the rotary SEA consists of: (a) Maxon Motor RE 40,
graphite brushes, 150 W DC motor; (b) worm gear set (M1-150 of
HPC Gears International Ltd.) with reduction ratio of 150:1;
(c) customized torsion spring; (d) angular contact bearings;
(e) magneto-resistant incremental encoder; and (f) opto-electronic
incremental encoder. The overall dimensions are shown in Fig. 3
and the resulting mass is 2.53 kg, allowing direct mounting of the
actuator on the frame of a knee orthosis.
The choice of gear-motor was made based on the characteristics of the knee joint considering gait pattern of a healthy person.
The angular velocities of the knee joint are in the range of
750 rpm and the maximum torque required for the project is
15 N m (see Section 2), while the maximum continuous torque and
the maximum velocity of the selected motor are respectively

Fig. 3. Cross section of the rotary SEA showing drivetrain components.

0.181 N m and 8200 rpm. Therefore a worm gear set with reduction ratio of 150:1 is used to adjust the operating range of motor in
order to fulll the requirements for velocity and torque. Thereby,
the worm gear output can operate in a velocity range of 755 rpm
and, if the efciency of the gears is not considered, it can provide a
maximum continuous torque of 27.15 N m. However, the friction
between the gear reduces signicantly the efciency and the torque amplication ratio is not necessarily the same as the ratio
velocity reduction (Kong et al., 2012). For this reason, a safety
factor of 1.8 to the torque requirement is considered.
All relevant information to the control system, i.e. motor rotation, actuator output, and spring deection estimate, are obtained
by two encoders. A magneto-resistant incremental encoder Maxon
with a resolution of 4096 pulses per revolution in quadrature
decoding mode is used to measure the motor rotation and allows
estimation of the position of the worm wheel. An opto-electronic
incremental encoder Maxon HEDS 5540 with a resolution of 2000
pulses per revolution in quadrature decoding mode is used to
measure the actuator output. The spring deection estimate is
obtained by the difference between the position of the worm
wheel and the actuator output. The theoretical output torque resolution, bounded by the lower output encoder resolution, is given
by ks (2 /2000), where ks is the spring constant.
3.1. Customized torsion spring
To meet the requirements of the proposed application, the
elastic element must be compact, lightweight, and able to withstand high torque with low intrinsic stiffness. However, these
characteristics are not found in torsion springs usually available.
For this reason, a new topology torsion spring is proposed. Fig. 4
(a) shows a schematic perspective view of the customized torsion
spring. It is composed of two rings interconnected by exible
elements dened by nite element analysis. The material selected
for analysis and manufacture was chromiumvanadium steel (AISI
6150), with Young's modulus of 205 GPa and a yield strength of
approximately 1320 MPa after a heat treatment process. Although
there are other steels with better characteristics for such application, the AISI 6150 is considered in this work because it is an
inexpensive steel and can be easily found in the market.
The stress distribution and deformation of the customized
spring was analyzed by the FEM, using ANSYS software to ensure
that the maximum stress is less than the yield strength of the
material when subjected to the maximum torque. The analysis
consists in xing the inner ring of the spring while tangential
forces equivalent to the input torque are applied on the outer ring.
The theoretical spring stiffness is computed by the ratio between
the torque applied and the corresponding angular deformation
obtained in the simulation. In the rst analysis, it was observed
that the stress concentration is located in the inner corner radius,
R4, Fig. 4(b).
In order to nd the lowest stress value for a given stiffness, the

310

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

Fig. 4. Customized torsion spring design. (a) Schematic perspective view of the torsion spring. (b) Static simulation for stress distribution.

following methodology was adopted. The value of radius R4 was

varied from 2.5 to 3.5 mm with a step of 0.05 mm for each thickness (E) from 5 to 8 mm with a step of 0.5 mm. The values of D1, D2
and D3 were varied proportionally to the R4 whereas the values of
R1, R2, R3, R5 and L were kept constant in order to not change the
proposed topology. Fig. 5 shows the results of the adopted methodology. From Fig. 5(a) a linear approximation can be used to
characterize the directly proportional relationship between spring
geometry parameters (R4 and E) and spring constant. When Fig. 5
(b) is analyzed, a similar behavior is observed between geometric
parameter R4 and von Mises stress, however, in an inversely proportional way. It can also be observed that the inversely proportional relationship between geometry parameter E and Von Mises
stress is characterized by a non-linear behavior. It is observed in
Fig. 5(b) that the distances between the curves are decreasing as
the thickness increases.
The values obtained in the simulation for the geometry parameters R4 and E that had lower stress for the desired stiffness of
200 N m/rad are respectively 2.8 and 8 mm. In order to evaluate
the proposed topology, two other springs with thickness (E) of
6 and 7 mm, corresponding to the spring constant (theoretical) of
150 and 175 N m/rad, were manufactured. The other parameters
are shown in Table 1.
The maximum values of von Mises stress obtained in the simulation were 732, 622 and 541 MPa for each spring with thickness (E) of 6, 7 and 8 mm, respectively. Thus a safety factor of
1.8 from the highest value of stress is considered. The customized
torsion spring, shown in Fig. 6(a), has been manufactured using
the Wire Electrical Discharge Machining (WEDM) process. Fig. 6
(b) shows the nal assembly of the rotary SEA.
Experimental characterization of the spring stiffness was performed by coupling on the output shaft of the rotary SEA in a
torque sensor (Gamma SI-65-5 from ATI Industrial Automation,
Inc.). The rotary SEA was programmed to follow a position prole
consisting of a sequence of steps (amplitude 0.14, duration 10 s) in
both loading and unloading conditions, while the torque was

Table 1
Customized torsion spring geometry parameters (mm).
D1

D2

D3

R1

R2

R3

R4

R5

6, 7 & 8

17.6

17

17.1

62.5

52.5

15

2.8

5.3

1050
Von Mises Stress [MPa]

Spring Constant [Nm/rad]

measured by the sensor. Fig. 7 shows the obtained results.

A linear regression was performed in both directions (compression and extension). From the obtained results it is possible to
observe a backlash with amplitude of 70.0044 rad occasioned by
intrinsic feature of the worm gear set. In order to compensate the
backlash on the sensor measurement, two lines are dened connecting the origin to the regression lines at the points corresponding to 7 0.01 rad of deection, as illustrated in Fig. 7(d).
With this compensation, it is possible to estimate low torques
applied to the load. An initial positioning procedure is used to
align the motor shaft to the center of the backlash. The output
shaft is positioned with its longitudinal axis perpendicular to the
horizontal plane. The motor is set to move slowly in one direction
until the output encoder measures one pulse. From this point, the
same procedure is done in the opposite direction. The difference
between the measures of the motor position at the extreme points
is an estimate of the backlash amplitude.
The corresponding values of the spring constant (ks) are shown
in Table 2. Note that when the spring is compressed (positive
deection), the stiffness is larger than when it is extended (negative deection). The value of the spring constant determined
experimentally is approximately 50% lower than the obtained by
nite element analysis. This discrepancy is probably due to the
properties of the material being different from the nominal used in
the simulation and the imperfections in the model and mesh used
in the analysis. The assembly of the orthosis with the rotary SEA
attached to it is shown in Fig. 8.

280
250
220
190
160
130
Thickness = (5, 5.5, ..., 7.5, 8) [mm]
100
2.5

2.6

2.7

2.8 2.9
3
3.1 3.2
Value of radius R4 [mm]

3.3

3.4

3.5

950

Thickness = (5, 5.5, ..., 7.5, 8) [mm]

850
750
650
550
450
2.5

2.6

2.7

2.8 2.9
3
3.1 3.2
Value of radius R4 [mm]

3.3

3.4

3.5

Fig. 5. Results of the methodology adopted. The arrows indicate the direction of increasing spring thickness. (a) Geometry parameters vs. spring constant. (b) Geometry
parameters vs. von Mises stress.

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

311

5
4
3
2
1
0
1
2
Compression spring
3
Extension spring
4
Linear regression
5
0.05 0.04 0.03 0.02 0.01
0
0.01 0.02 0.03 0.04 0.05
Deflection [rad]

Torque [Nm]

5
4
3
2
1
0
1
2
Compression spring
3
Extension spring
4
Linear regression
5
0.05 0.04 0.03 0.02 0.01
0
0.01 0.02 0.03 0.04 0.05
Deflection [rad]

5
4
3
2
1
0
1
2
Compression spring
3
Extension spring
4
Linear regression
5
0.05 0.04 0.03 0.02 0.01
0
0.01 0.02 0.03 0.04 0.05
Deflection [rad]

2
1.5
Torque [Nm]

Torque [Nm]

Torque [Nm]

Fig. 6. Customized spring and rotary Series Elastic Actuator. (a) Customized torsion spring. (b) Rotary SEA.

backlash compensation

0.5
0
0.5
Compression spring
Extension spring
Linear regression

1
1.5
2
0.02 0.015 0.01 0.005
0
0.005
Deflection [rad]

0.01

0.015

0.02

Fig. 7. Characterization of the customized torsion spring. (a) Spring 1 thickness of 6 mm and stiffness equal to 84 N m/rad. (b) Spring 2 thickness of 7 mm and stiffness
equal to 94 N m/rad. (c) Spring 3 thickness of 8 mm and stiffness equal to 105 N m/rad. (d) Spring 1 backlash compensation on the sensor measurement.
Table 2
Customized torsion spring physical parameters.
Spring

1
2
3

Mass (kg)

0.292
0.338
0.384

ks simulated (N m/rad)

150
175
200

Fig. 8. Rotary SEA attached to a knee orthosis.

ks real (N m/rad)
Extension

Compression

82
92
103

85
96
106

312

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

Fig. 9. Block diagram of the torque control with inner velocity loop.

4. Controller design

in (s ) = m (s ) (Jm + Jwg ) m (s ) s2 Bm m (s ) s,

4.1. Control hardware

The control hardware consists of a EPOS 24/5 Positioning
Controller, Maxon Motor, and an ordinary computer hardware
equipped with a CAN communication card sold by National Instruments. The EPOS is a full digital smart motion controller capable of operating in position, velocity and current modes. The
device is also responsible to decode the signals from quadrature
encoders. The communication interface between the computer
hardware and the EPOS is performed by CANopen communication
protocol. The frequency of the control loop is set to 200 Hz.
In adopted approaches traditionally for SEA torque control, the
motor is treated as a torque source. This assumption is justied
since the current applied to the motor is directly proportional to
the output torque. However, this approach becomes difcult to
implement due to the intrinsic nonlinearities of the drivetrain
such as backlash, stiction, and efciency losses. In Robinson
(2000), it is suggested to treat the motor as a velocity source rather
than as a torque source. According to Wyeth (2006), this approach
helps to overcome these undesirable effects of the gear-motor.
The block diagram of the torque control loop is illustrated in
Fig. 9. The inner velocity loop is performed by the built-in EPOS
velocity control. The PI controller parameters were dened by the
automatic tunning procedure available in the device conguration
set-up.
The torque applied to the load is calculated as l = ks (ww l ),
where l is the load position and ww is the worm wheel position
given by ww = m N1, where m is the motor position and N  1 is
the reduction ratio. It is ltered through a second-order low-pass
Butterworth lter with a cut-off frequency of 50 Hz.
4.2. Dynamic model
The transfer function of the rotary SEA can be obtained as described in the sequence. Regarding Fig. 10, the input torque (in)
can be expressed by

Jwg

in

Jm, Bm

(5)

where m represents the torque generated by the motor, Jm and Jwg

are the rotational inertias of the motor and the worm gear, respectively, and Bm is the damping coefcient of the motor.
Similarly, the output torque (out) can be written as

out (s ) = (Js + Jww ) ww (s ) s2 + Bww ww (s ) s + l (s ),

(6)

where Js and Jww are, respectively, the rotational inertias of the

spring and worm wheel, Bww is the damping coefcient of the
drivetrain, and l is the torque applied to the load.
Since the torque generated by the motor is amplied by the
velocity reduction ratio, the following kinematic conditions can be
considered:

out (s ) = Nin (s )

(7)

and

ww (s ) = N (1) m (s ),

(8)

where N (1) is the reduction ratio.

Substituting (5) and (6) into (7), and (8) into the resulting
equation, one obtains

Jeq ww (s ) s2 + Beq ww (s ) s = Nm (s ) l (s ),

(9)

where Jeq is the equivalent rotational inertia and Beq is the

equivalent damping coefcient, dened as

Jeq = Jww + Js + (Jm + Jwg ) N 2

(10)

and

Beq = Bww + Bm N 2.

(11)

As in the adopted approach the motor is treated as a velocity

source and considering that the desired current is equal to the
actual current, the motor torque can be computed as

k
k
m (s ) = kt kp + i md (s ) kp + i m (s ) ,

s
s

(12)

where kp and ki are the proportional and integral gains of the

velocity controller, respectively, kt is the motor torque constant, md
is the desired motor velocity, and m is the motor velocity, which
can be expressed by m (s ) = Nww (s ) s .
Substituting (12) into (9) and rewriting ww (s ) as l (s ) /ks + l (s ),
one gets the open-loop transfer function of the system as a function of two variables md and l as

Nk k k

s t p s + Nks kt ki
J
J

eq
eq
l (s ) =
d (s )
B + k k N2
k + k k N2 m
eq
t p
s
t i
3
2

s +
s +
s
Jeq
Jeq

Js
Jww, Bww

Fig. 10. Dynamic model of the rotary SEA.

B k + N 2k k k
2

eq s
s t p
s + N ks kt ki
ks s2 +
Jeq
Jeq

(s ).
B + k k N2
k + k k N2 l
eq
t p
s
t
i
2
s +

s +
Jeq
Jeq

(13)

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

313

Table 3
Rotary SEA model parameters.
ks (N m/
rad)

Jeq (kg m2) Beq (N m s/

rad)

kt (A/
N m)

84, 94 &
105

0.47

0.03

150 11.9

60

kp (N m s/
rad)

ki (N m/
rad)
1.19

However, for the controller design it was considered the case

where the load is xed (l 0). Therefore, open-loop transfer
function of the system is given by

Nk k k

s t p s + Nks kt ki
J
J

eq
eq
(s )
=
G (s ) = ld
,
B + k k N2
k + k k N2
m (s )
eq
t
p
s
t i
3
2

s +
s +
s
Jeq
Jeq

This section presents the linear / control design and the

procedures to apply it for torque control of rotary SEA. It adopts
the mixed-sensitivity / control procedure over the sensitivity (S
(s)) and the control sensitivity ( K (s ) S (s )) transfer functions, as
presented in Skogestad and Postlethwaite (2005), where
S (s ) = (I + G (s ) K (s ))1 and K(s) is the resulting controller. It ensures robust tracking performance, disturbance rejection, and
denes the size and bandwidth of the controller (to limit the
control energy), with relation to the worst-case bounded ( 3 2)
external inputs acting on the system (robustness with relation to
uncertainties is not guaranteed). The procedure consists in dening functions We(s) and Wu(s) to weight S(s) and KS(s),
respectively.
The transfer function We(s) is dened as

(15)

where s is the closed-loop bandwidth, Ms is the maximum peak

for the sensitivity function, and s > 0 is the maximum steadystate error. Ms denes the maximum amplication of high frequency noises and the limits for the overshoot. The transfer
function Wu(s) is dened as

Wu (s ) =

1 s + u
,
Mu

s + u

u
K(s)

(14)

4.3. / torque control

1 s + s
,
Ms (s + s s )

z
P(s)

where the parameters are dened in Table 3. Although the

equivalent rotational inertia, Jeq, depends on the spring rotational
inertia, Js, its value has not changed since the difference between
the rotational inertia of the springs is small and the predominant
term in (10) is dened by the high reduction ratio N.

We (s ) =

Fig. 11. Block diagram of the / torque control.

(16)

where u is the controller bandwidth, Mu is the maximum controller gain for high frequencies, and u > 0 is a small value set to
ensure that the cost function is dominated by We(s) at low frequency, and to guarantee that the controller design is well-posed.
The weighting functions parameters are shown in Table 4.
Fig. 11 shows the block diagram of the SEA torque control represented in / control framework. It considers the weighting
Table 4
Weighting functions parameters.
Ms (dB)

b (rad/s)

Mu (dB)

bc (rad/s)

29

600

0.0012

44

600

0.01

Fig. 12. Block diagram for / control.

functions We(s) and Wu(s) which scale the error (e) and the input
(u), respectively. The parametric uncertainties and external disturbances acting on the plant output are grouped in the combined
disturbance d(t).
/ control framework considers the system block diagram as
shown in Fig. 12, where P(s) represents the augmented plant
(nominal plant G(s) plus weighting functions) and K(s) the controller. The plant has two sets of input signals, the internal input u
and the external input w, and two sets of output signals, the
measured signal y and the regulated output z. Regarding Fig. 11:

u = md,

y = e,

w1 d
w = = l ,
w2 d
z1 We e
z= =
.
z2 Wu u

The objective of / controllers is to guarantee that / norm of

the transfer function from input w to output z, Tzw (s ), be bounded
by an attenuation level . Regarding the mixed-sensitivity /
control, this condition is given by

Tzw (s ) =

We (s )
Wu (s )

where We (s) = [We (s) S (s) We (s) S (s)] and Wu (s) = [Wu (s) K (s) S (s) Wu (s)
K (s ) S (s )] . The parameter indicates the level of robustness of the
control system with relation to the worst-case bounded disturbance w. The resulting values of S(s) and K (s ) S (s ) satisfy the
following inequalities:

(S (j)) (We1 (j)),

and

(K (j) S (j)) (Wu1 (j)),

where (H (j)) and (H (j)) are the greatest and least singular
values of the transfer function H(s) over the frequency , respectively. Fig. 13 shows the graphics of the sensitivity function
S (j) versus the weighting function We1 (j), and of K (j) S (j)
versus Wu1 (j), for the resulting controller K1 (j). Note that the
controlled plant curves are lower than the weighting functions for
all frequency. The level of attenuation for this case is 0.137.
4.3.1. Torque control bandwidth
Torque control bandwidth was determined through Frequency
Response Function (FRF) analysis, obtained from H1 (j) and H2 (j)

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

100

150

50

100

Magnitude (dB)

Magnitude (dB)

314

0
50
1

We (j)
S(j)

100
150 2
10

10

10

10
10
Frequency (rad/s)

10

10

10

50
0
50

100 0
10

Wu1(j)
K(j)S(j)
10

10

10
10
Frequency (rad/s)

10

10

10

Fig. 13. Sensitivity function S (j) vs. We1 (j) and K (j) S (j) vs. Wu1 (j) .

estimators as proposed in Maia and Silva (1997). The estimator

H1 (j) considers the presence of random noise in the output signal,
and in this case, the FRF is calculated as H1 (j) = SXY (j) /SXX (j),
where SXY (j) is the cross-spectrum of the input and output signal,
and SXX (j) is the auto-spectrum of the input signal; the estimator
H2 (j) considers the presence of random noise in the input signal,
and in this case, the FRF is calculated as H2 (j) = SYY (j) /SXY (j),
where SYY (j) is the auto-spectrum of the output signal. The results obtained by the estimators can be evaluated through the
ordinary coherence function C (j), dened as the ratio of the
H1 (j) estimator over the H2 (j) estimator. The ordinary coherence
function is always a real value between 0 and 1, where a value near
to 1 represents a good estimate of both estimators.
To implement the estimation method, the desired torque ( ld ),
seen here as an input signal, was dened as a chirp signal with
amplitude of 5 N m and frequency varying from 0 to 20 Hz and
back to 0 Hz in 20 s. The signal is repeated four times, totaling a
sampling period of 80 s. Fig. 14 shows the Bode plot of the closed
loop transfer function computed from H1 (j) and H2 (j) estimators
and the coherence function C (j) for the rotary SEA assembled
with springs 1 and 3. The values of the torque control bandwidth
for the proposed SEA assembled with springs 1, 2 and 3 are respectively 6.8, 8.5 and 9.6 Hz. Note that there is a directly proportional relationship between spring stiffness and torque control
bandwidth. This is consistent with the literature, e.g. Robinson,
Pratt, Paluska, and Pratt (1999).

Phase [deg]

Magnitude [dB]

20
10
0
10
20 1
10
0

10

10

100
200

Coherence

300 1
10
1

10

10

Fig. 15. Block diagram of the impedance control with inner loop / torque
control.

4.4. Impedance control

The impedance control is implemented through a PD position
controller in cascade with / torque controller. This approach is
similar to that proposed in Santos and Siqueira (2014), however,
this paper includes an inner robust controller. Fig. 15 shows the
block diagram of the proposed control scheme. The outer position
controller determines the desired torque ( ld ) according to position
and velocity errors and impedance parameters. Thus, the desired
torque can be expressed by
d

ld = k v (ld l ) + bv (l l ),

(17)

where kv and bv are the virtual stiffness and damping, respectively.

Two experiments were conducted in order to evaluate the
ability of rotary SEA to reproduce the behavior of the knee joint
impedance within a frequency range suitable for the given application. First, considering a pure damping control through transfer
function Tb (j) = l (j) /l (j). In the second experiment, considering a pure stiffness control through transfer function
Tk (j) = l (j) /l (j). The input signals used to estimate the FRF in
both experiments were generated by a human operator by imposing oscillating movements to the SEA output shaft.
4.4.1. Damping control
d
The pure damping control was obtained by setting l = 0 and
kv 0, and evaluated varying the value of bv in 5, 10 and 15 N m s/
rad, for the rotary SEA assembled with the different springs. Fig. 16
shows the Bode plot of the closed loop transfer function Tb (j),
computed from H1 (j) and H2 (j) estimators, and the coherence
function between them, considering the SEA assembled with
Spring 3. The damping control bandwidth is approximately 3 Hz
and no signicant variations in bandwidth were observed for the
different values of bv and different values of spring constant.
However, a difference of 3 N m s/rad was evidenced between the
desired value of bv and the actual value, demonstrating the existence of an intrinsic damping in the system.

0.5
Spring 1
Spring 3
0 1
10

10
Frequency [Hz]
Fig. 14. Frequency response of the torque control.

10

4.4.2. Stiffness control

Similar to the rst experiment, pure stiffness control was also
evaluated by setting ld = 0 and bv 0, and varying the value of kv
in 10, 30 and 60 N m/rad, for the rotary SEA assembled with the
different springs. Fig. 17(a) shows the Bode plot of the closed loop
transfer function Tk (j) and the coherence function, considering

30

315

Table 5
Stiffness control bandwidth.

25

Spring

20

Bandwidth (Hz)

15
10 1
10
90

10

10

1
2
3

kv 10 N m/rad

kv 30 N m/rad

kv 60 N m/rad

0.6
0.5
0.4

1.5
1.4
1.3

3.8
2.9
2.5

180
270
10

10

bv = 5 Nms/rad
b = 10 Nms/rad

0.5

bv = 15 Nms/rad
0 1
10

10
Frequency [Hz]

10

Fig. 16. Frequency response of the damping control (Spring 3).

Torque [Nm]

180

Coherence

360 1
10
1

10

10

10

kv = 10 Nm/rad
0.5

kv = 30 Nm/rad
kv = 60 Nm/rad

0 1
10

10

10
Frequency [Hz]

0
0
70

4
5
Time [s]

35
0
35
70
0
5
2.5
0
2.5

10

45
40
35

30 1
10
120

10

Phase [deg]

10

15

Fig. 18. Experimental data of active orthosis during walking when the desired
torque is set to zero (Spring 3).

Magnitude [dB]
0

10

30

5
0

10

10

10

180
240
300 1
10
1

Coherence

Phase [deg]

Magnitude [dB]

the SEA assembled with Spring 3. The stiffness control bandwidth

considering the values kv equal to 10, 30 and 60 N m/rad is respectively 0.4, 1.4 and 2.7 Hz. Note that a larger bandwidth is
achieved with higher values of kv. This is justiable, since the
largest bandwidth is obtained when kv is equal to ks (approximately 105 N m/rad), and at high frequencies, the virtual stiffness
value (kv) tends to the spring stiffness (ks).
Fig. 17(b) shows the Bode plot of the closed loop transfer
function Tk (j) considering a constant value of virtual stiffness (kv
60 N m/rad) for the SEA assembled with the different springs.
Note that, the lower the spring constant, the higher the stiffness
control bandwidth. The other results obtained are shown in
Table 5.
The results, if compared with knee joint dynamic properties
presented in the literature Zhang, Nuber, Butler, Bowen, and Rymer (1998), show that the proposed rotary SEA is able to render a
desired stiffness and damping within a range of appropriate amplitude and frequency to assist the knee joint.

50
40
30
20
10
0 1
10
0

45

Joint Angle [deg]

Coherence

360 1
10
1

Joint Velocity [deg/s]

Phase [deg]

Magnitude [dB]

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

0.5
0 1
10

10

10

10

Spring 1
Spring 2
Spring 3
0

10

10
Frequency [Hz]

Fig. 17. Frequency response of the stiffness control. (a) (Spring 3). (b) (kv 60 Nm/rad).

10

316

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

5. Experimental results with active knee orthosis

70

5.1. Torque control

60

According to Kong et al. (2009), to perform a good torque

control in robots interacting with humans, an accurate actuation
system with low impedance is desired. In other words, the active
orthosis must follow the knee joint motion, so that the subject
does not feel any resistance, when the desired torque is set to zero.
Therefore, the following experiment was performed: one subject
walked on a treadmill using the active knee orthosis with the
desired torque set to zero.
As expected, the SEA followed the human joint motion to keep
zero spring deection. Fig. 18 shows the angle and velocity of the
knee joint and the interaction torque between the user's limb and
the active orthosis, for a gait cycle duration of 3 s. Note that the
resistive torque generated by the actuator is proportional to the
joint angular velocity. This is due to the presence of an intrinsic
damping in the actuator. The active knee orthosis works as a
damper when the desired torque is set to zero. Note also that the
stability was not affected by the human motions.
The active knee orthosis was also commanded to track a torque
prole dened as a sine wave with amplitude of 6 N m and frequency of 0.1 Hz in order to evaluate accuracy and robustness of
the adopted control strategy while interacting with a human.
Fig. 19(a) shows knee joint angle, reference and actual torque as
well as magnitude of the torque error when the subject resists the
generated torque. The subject is able to resist the applied torque,
maintaining a near zero joint position. The root mean square (rms)
value of the torque error is 0.1 N m, demonstrating that the active
orthosis is able to generate the desired torque accurately. Fig. 19
(b) shows knee joint angle, reference and actual torque, and
magnitude of the torque error when the subject is complaint with
the generated torque. The rms value of the torque error is
1.06 N m. The torque error shown in Fig. 19(b) is greater than that
shown in Fig. 19(a) because of the human motion. However, it
must be noted that the stability was not affected.

50

5.2. Impedance control

0.45
0.3
0.15
5

10

15

20

25

6
0
6
12
0
0.5

10

15

20

25

0.5
0

10

15
Time [s]

20

25

30

0.6 1.2 1.8 2.4

3 3.6 4.2 4.8 5.4

Time [s]

6.6

5.2.1. Variable impedance control strategy

In this section, a variable impedance control strategy is evaluated, where two different values of the virtual stiffness are dened for the stance and swing phases of the gait cycle, K1 = 100
and K2 = 25 N m/rad, respectively. For simplicity, in both cases, the
virtual damping was set to zero. At the rst moment, a subject
wearing the knee orthosis and walking on a treadmill was asked to
follow the orthosis trajectory without imposing constraint to its
movement. In the second moment, the subject was asked to force
the device against the desired trajectory, trying to modicate it.
These tests were proposed to show the orthosis ability to perform variable impedance during the gait phases. This feature will
be useful in our on-going efforts to design adaptive impedance
control strategies for robot-aided rehabilitation. For example, in
Ibarra, Santos, Krebs, and Siqueira (2014) the robot's stiffness is

30

20

was adjusted, respectively, to 20 and 100 N m/rad. In both cases,

the virtual damping was set to zero. The trajectory prole was
computed by interpolation of several points of a normal gait obtained in Kirtley (2006), for a gait cycle duration of 2.2 s, Fig. 20.
Note that the higher the stiffness value, the greater the accuracy
will be.

30

Reference
Actual

30

Fig. 20. Experimental data of active orthosis during walking with low and high
impedance (Spring 3).

Torque [Nm]

0
0
12

High impedance

40

Joint Angle [deg]

0.6

Low impedance

10

Torque Error [Nm]

Torque Error [Nm]

Torque [Nm]

Joint Angle [deg]

As a preliminary validation of the impedance control, the active

orthosis was commanded to track a knee joint trajectory prole
with low and high impedance, the controller's stiffness parameter

Joint Angle [deg]

Reference

60
45
30
15
0
0
12

10

15

20

25

30

Reference
Actual

6
0
6
12
0
2.5

10

15

20

25

30

10

15
Time [s]

20

25

30

2.5
0

Fig. 19. Experimental data of active orthosis when the desired torque is set as a sinusoidal wave (Spring 3). (a) With movement constraint. (b) Without movement constraint.

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

70

70
Reference
Actual

50
Stance Phase

Stance Phase

40
30
20

Swing
Phase

Swing
Phase

10

Reference
Actual

60
Joint Angle [deg]

Joint Angle [deg]

60

0
0

317

50
Stance Phase

Stance Phase

40
30
20

Swing
Phase

Swing
Phase

10
0.4

0.8

1.2

1.6

2 2.4
Time [s]

2.8

3.2

3.6

4.4

0
0

0.4 0.8

1.2 1.6

2 2.4 2.8
Time [s]

3.2 3.6

4.4

Fig. 21. Analysis of the variable impedance control (Spring 3). (a) Variable impedance without external torque. (b) Variable impedance with external torque.

adjusted based on the patient's stiffness estimation and on her/his

performance on playing distinct video games. The same approach
is now under investigation for walking conditions.
Fig. 21(a) shows the active orthosis behavior in the absence of
external torque, when there is no movement constraint. It is
possible to note that the reference trajectory is followed by the
orthosis in the two different impedance states (K1 and K2). On the
other hand, when a torque is applied by the subject against the
desired trajectory, Fig. 21(b), the reference trajectory is followed
only in the highest impedance state, K1. In this case, the applied
torque by the orthosis compensates the external torque and
guarantees the trajectory tracking. However, in the low impedance
state, K2, the torque decreases and the reference trajectory is not
followed, allowing the user to move the orthosis easily out of the
desired trajectory.

the active orthosis, were performed in order to evaluate the performance of the developed prototype. The initial results showed in
this paper are part of our on-going efforts to design adaptive impedance control strategies for robot-aided rehabilitation. A clinical
study with patients will be conducted in future works to evaluate
if the proposed orthosis is more effective than other available
devices.

Appendix A. Supplementary data

Supplementary data associated with this paper can be found in
the online version at http://dx.doi.org/10.1016/j.conengprac.2015.
09.008.

References
6. Conclusions
This paper presents the design of an active knee orthosis driven
by a customized rotary Series Elastic Actuator to assist in exion/
extension of the knee joint during physical therapy. A customized
torsion spring is designed through simulations based on nite
element method with the aim of fullling a set of requirements
dened in terms of admissible peak load, low stiffness and a
compact and lightweight design. The value of the spring constant
determined experimentally is signicantly lower than obtained by
simulations. This discrepancy is probably due to the actual properties of the material being different from the nominal used in the
simulation and the imperfections in the model and mesh used in
the analysis. The resulting mass of the rotary SEA is 2.5 kg, thus
allowing direct mounting of the actuator on the frame of a knee
orthosis.
Torque and impedance controllers are implemented to ensure
secure interaction with the patient, enabling new strategies for
rehabilitation, for example, a variable impedance control strategy
is performed and shows that it is possible to regulate the impedance during walking. The performance of the controllers was
evaluated through frequency response analysis. The adopted torque control strategy, based on / criterion, allows the rotary SEA
to provide a continuous torque of 5 N m with a bandwidth of
9.6 Hz. The results obtained for impedance control show that the
actuator is able to render a desired stiffness and damping within a
range of appropriate amplitude and frequency to assist the knee
joint. A comparison of the proposed / controller with similar
controllers designed for rehabilitation are under investigation.
Experiments, involving the interaction between a subject and

Carpino, G., Accoto, D., Sergi, F., Tagliamonte, N. L., & Guglielmelli, E. (2012). A novel
compact torsional spring for series elastic actuators for assistive wearable robots. Journal of Mechanical Design, 134(12), 110.
Ferris, D. P., Sawicki, G. S., & Domingo, A. (2006). Powered lower limb orthoses for
gait rehabilitation. Topics in Spinal Cord Injury Rehabilitation, 11(2), 3449.
Ham, R. V., Sugar, T. G., Vanderborght, B., Hollander, K. W., & Lefeber, D. (2009).
Compliant actuator designs. IEEE Robotics Automation Magazine, 16(3), 8194.
Hogan, N. (1985). Impedance control: An approach to manipulation. Journal of
Dynamic Systems Measure Control, 107(Parts 13) (1), 124.
Ibarra, J. C. P., Santos, W. M., Krebs, H. I., & Siqueira, A. A. G. (2014). Adaptive impedance control for robot-aided rehabilitation of ankle movements. In 5th IEEE
RAS & EMBS international conference on biomedical robotics and biomechatronics
(pp. 664669), So Paulo, Brazil.
Kirtley, C. (2006). Clinical gait analysis: Theory and practice (1st ed.). Philadelphia,
PA: Elsevier Churchill Livingstone.
Kong, K., Bae, J., & Tomizuka, M. (2009). Control of rotary series elastic actuator for
ideal force mode actuation in human robot interaction applications. IEEE/ASME
Transactions on Mechatronics, 14(1), 105118.
Kong, K., Bae, J., & Tomizuka, M. (2012). A compact rotary series elastic actuator for
human assistive systems. IEEE/ASME Transactions on Mechatronics, 17(2),
288297.
Kwakkel, G., Kollen, B. J., & Krebs, H. I. (2008). Effects of robot-assisted therapy on
upper limb recovery after stroke: A systematic review. Neurorehabilitation and
Neural Repair, 22(2), 111121.
Lagoda, C., Schouten, A., Stienen, A., Hekman, E., & van der Kooij, H., (2010). Design
of an electric series elastic actuated joint for robotic gait rehabilitation training.
In Proceedings of 3rd IEEE RAS and EMBS international conference on biomedical
robotics and biomechatronics (pp. 2126), Tokyo, Japan.
Maia, N. M. M., & Silva, J. M. M. (1997). Theoretical and experimental modal analysis. Engineering dynamics series. Research Studies Press Limited, Baldock,
Hertfordshire, UK.
Prange, G. B., Jannink, M. J. A., Groothuis-Oudshoorn, C. G. M., Hermens, H. J., &
Ijzerman, M. J. (2006). Systematic review of the effect of robot-aided therapy on
recovery of the hemiparetic arm after stroke. Journal of Rehabilitation Research
and Development, 43(2), 171183.
Pratt, G., & Williamson, M. (1995). Series elastic actuators. In Proceedings of the 1995
IEEE/RSJ international conference on intelligent robots and systems (Vol. 1, pp.
399-406), Pittsburgh, USA.

318

W.M. dos Santos et al. / Control Engineering Practice 58 (2017) 307318

Robinson, D. W. (2000). Design and analysis of series elasticity in closed-loop actuator force control (Ph.D. thesis). Cambridge: Massachusetts Institute of
Technology.
Robinson, D. W., Pratt, J. E., Paluska, D. J., & Pratt, G. A. (1999). Series elastic actuator
development for a biomimetic walking robot. In IEEE/ASME international conference on advanced intelligent mechatronics (pp. 561568), Atlanta, USA.
Santis, A. D., Siciliano, B., Luca, A. D., & Bicchi, A. (2008). An atlas of physical human
robot interaction. Mechanism and Machine Theory, 43(3), 253270.
Santos, W. M., & Siqueira, A. A. G. (2014). Impedance control of a rotary series elastic
actuator for knee rehabilitation. In: 19th World congress of the international
federation of automatic control (pp. 48014806), Cape Town, South Africa.
Sergi, F., Accoto, D., Carpino, G., Tagliamonte, N. L., & Guglielmelli, E., (2012). Design
and characterization of a compact rotary series elastic actuator for knee assistance during overground walking. In The Fourth IEEE RAS/EMBS international
conference on biomedical robotics and biomechatronics (pp. 19311936), Roma,
Italy.
Skogestad, S., & Postlethwaite, I. (2005). Multivariable feed back control: analysis and
design (2nd ed.). NewYork: John Wiley and Sons.

Stienen, A. H. A., Hekman, E. E. G., ter Braak, H., Aalsma, A. M. M., van der Helm, F. C.
T., & van der Kooij, H. (2010). Design of a rotational hydroelastic actuator for a
powered exoskeleton for upper limb rehabilitation. IEEE Transactions on Biomedical Engineering, 57(3), 728735.
Tsagarakis, N., Laffranchi, M., Vanderborght, B., & Caldwell, D., (2009). A compact
soft actuator unit for small scale human friendly robots. In Proceedings of IEEE
international conference on robotics and automation (pp. 43564362), Kobe,
Japan.
Wyeth, G. (2006). Control issues for velocity sourced series elastic actuators. In
Proceedings of the Australasian conference on robotics and automation (pp. 16),
Auckland, New Zealand.
Yoon, S. S., Kang, S., Yun, S. K., Kim, S. J., Kim, Y. H., & Kim, M. (2005). Safe arm
design with mr-based passive compliant joints and viscoelastic covering for
service robot applications. Journal of Mechanical Science and Technology, 19(10),
18351845.
Zhang, L. Q., Nuber, G., Butler, J., Bowen, M., & Rymer, W. Z. (1998). In vivo human
knee joint dynamic properties as functions of muscle contraction and joint
position. Journal of Biomechanics, 31(1), 7176.