Mechatronics xxx (2013) xxxxxx

Contents lists available at ScienceDirect

Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics

Control design of a novel compliant actuator for rehabilitation robots

Haoyong Yu a,b, Sunan Huang a,, Gong Chen a,b, Nitish Thakor a
a
SiNAPSE, Singapore Institute of Neurotechnology, Singapore 117575, Singapore
b
Department of Bioengineering, National University of Singapore, Singapore 117575, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: Rehabilitation robots have direct physical interaction with human body. Ideally, actuators for rehabilita-
Received 1 October 2012 tion robots should be compliant, force controllable, and back drivable due to safety and control consid-
Accepted 23 August 2013 erations. Series Elastic Actuators (SEA) offers many advantages for these applications and various designs
Available online xxxx
have been developed. However, current SEA designs face a common performance limitation due to the
compromise on the spring stiffness selection. This paper presents a novel compact compliant force con-
Keywords: trol actuator design for portable rehabilitation robots to overcome the performance limitations of current
Compliant actuator
SEAs. Our design consists of a servomotor, a ball screw, a torsional spring between the motor and the ball
Optimal control
Sliding mode
screw, and a set of translational springs between the ball screw nut and the external load. The soft trans-
Smooth switching control lational springs are used to handle the low force operation, while the torsional spring with high effective
stiffness is used to deal with the large force operation. It is a challenging task to design the controller for
such a novel design as the control system needs to handle both the force ranges. In this paper, we develop
the force control strategy for this actuator. First, two dynamical models of the actuator are established
based on different force ranges. Second, we propose an optimal control with friction compensation
and disturbance rejection which is enhanced by a feedforward control for the low force range. The pro-
posed optimal control with feedforward term is also extended to the high force range. Third, a switching
control strategy is proposed to handle a transition between low force and high force control. The math-
ematical proof is given to ensure the stability of the closed-loop system under the proposed switching
control. Finally, the proposed method is validated with experimental results on a prototype of the actu-
ator system and is also veried with an ankle robot in walking experiments.
2013 Elsevier Ltd. All rights reserved.

1. Introduction research on compliant actuators (CAs) because of their ability to

In classic industrial applications such as the autonomous weld-
ing systems for the automobile industry, robots are always de-
signed with stiff actuators for precise and rapid position control
with good repeatability. Stiff actuators are good for handling exter-
nal disturbance forces and internal frictions, but cannot handle
external impacts and shocks. Therefore, they are most suitable
for working environments that are well dened with no direct
physical interactions with humans. In recent years, due to the rap-
idly aging populations in most developed nations, there is a strong
need for service robots, assistive and rehabilitation robots in both
domestic [1,2] and hospital settings [3,4]. In these applications,
the robots need to adjust with versatile unstructured environ-
ments as well as human demands [5]. Stiff actuators may be
unsuitable for unstructured environments and the environments
where humans safety is a vital issue. This motivates the need of
Corresponding author. Address: BLK 611, #07-280, Clementi, West Street 1,
Singapore 120611. Tel.: +65 67795219; fax: +65 6872 3069.
E-mail address: elehsn@gmail.com (S. Huang).
safely interact with the user, and their ability to store and release
energy in passive elastic elements. The most well-known CA is the
series elastic actuator (SEA) [6,7] where a spring is placed between
the motor and the load. Compared to a stiff actuator, series elastic
actuators have the following benets [8]:
 The actuators exhibit lower output impedance and back-
driveability.
 Shock tolerance is greatly improved by the springs.
 The force transmission is smooth.
 Energy can be stored and released in the elastic element,
thereby improving efciency in applications.
Many different SEAs have been developed for rehabilitation ro-
bots to capitalize the advantages of the SEA concept. In [9,10], the
SEA is designed based on a linear spring coupled to a ballscrew
which is connected to a dc motor. In [11,12], a Bowden cable is
connected to linear spring to achieve a rotary SEA. In [13,14], ro-
tary SEAs are designed in which torsional spring is used to transmit
the output force. Although current SEAs have achieved reasonable

0957-4158/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mechatronics.2013.08.004

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
2 H. Yu et al. / Mechatronics xxx (2013) xxxxxx
performance, they still face a common fundamental limitation,
which is the xed spring constant of the elastic element as dis-
cussed by Pratt et al. in [9,15]. The performance of the SEAs largely
depends on spring constant [15]. Soft spring produces high delity
of force control, low output impedance, and reduces stiction, but
also limits the force range and the force control bandwidth at high
force range. On the other hand, stiff spring increases large force
bandwidth, but reduces force delity. In order to achieve the de-
sired output force/torque, most current SEAs are designed with
very stiff springs, leading to compromised force control perfor-
mance, low intrinsic compliance and back-drivability, and bulky
and heavy systems. The novel design presented in this paper aims
to overcome the above limitations of the conventional SEAs while
improve the performance. This design concept was rst proposed
in [16,17] and supported with simulation results and simple
experiments.
Apart from the mechanical development of SEAs, the control de-
sign of SEAs is also gaining attention in recent years. Many control-
lers have been presented for SEA and their performance have also
been analyzed. The following literature will give a brief insight in
this area. In [10,11,15,18], pure PID control is used to produce a de-
sired output force. In [9], PD plus feedforward control is used to
improve the dynamical performance for a class of SEAs. In
[13,1921], a type of cascaded control is presented to ensure sta-
bility in human interacting devices where a PI torque control is
used in the outer loop, while a PI velocity control is used in the in-
ner loop. In [6], a modied PID with feedforward term and human
joint compensator is designed to generate desired force and low
impedance. In [14], based on the model of a SEA, the authors fur-
ther improve the human joint compensator used in [6] by adding
a low pass lter. A disturbance observer is also designed to com-
pensate the modeling error. In [22], the authors use a feedback plus
feedforward force control which is enhanced by a disturbance ob-
server to compensate for plant variations, where feedback and
feedforward controls are optimally designed. However, all these
current controllers are designed based on the actuators with xed
stiffness spring, either linear spring or torsional spring. For an actu-
ator working with both types of springs in different force ranges,
existing results are not available.
In this paper, we present a novel compliant actuator system and
its force control design. Unlike existing SEAs, our system uses two
types of springs (torsional and translational) at different force
ranges. In order for this SEA to work well, there are challenges to
be addressed in the control system. An adequate control strategy
should be designed to deal with system dynamics and make a tran-
sition between high force and low force. The goal of the paper is to
address these issues. First, two dynamical models are established
with respect to low force range and high force range. Second, an
optimal control is designed for low force range. Since there is a fric-
tional behavior at the low force, we design a compensator in our
controller to deal with the friction. As sliding mode control has
been used as general tools for handling unknown nonlinear uncer-
tainties [23], we incorporate it into our controller to deal with the
unknown disturbance. The proposed optimal control plus feedfor-
ward term is also extended to high force range. Third, a switching
control is proposed to make a smooth transition between low force
and high force. The mathematical analysis is given to prove the sta-
bility of the closed-loop system. Finally, experimental results are
provided to verify the effectiveness of the proposed method.
2. Compliant actuator design for gait rehabilitation robot
Gait disorders are common for patients post stroke and in most
cases cannot be treated medically or surgically. Therefore, treat-
ment often relies on rehabilitation service. Rehabilitation robotics
Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
has shown promise in providing patients with intensive therapy
leading to functional gains. This involves the use of a robot exoskel-
eton device or end-effector device to help the patient retrain motor
coordination by performing gait movement. Here, we are develop-
ing a portable wearable knee ankle robot for gait rehabilitation [17]
based on the novel actuator design presented in this paper.
Fig. 1 shows the concept design of the robot. The modular sys-
tem consists of an ankle foot module and a knee module. Each
module is driven with the same compact compliant force control-
lable actuator as shown in Fig. 1. Based on the human lower limb
biomechanics, it is known that the range of movement of the lower
limb joints is within a known range during normal walking. We
can make use of this property and use the simple rockerslider
mechanism to achieve a compact design that is most suitable for
the exoskeleton. The core component in this exoskeleton is the
compliant actuator as shown in Fig. 2.
Our novel actuator design consists of a servomotor with a rotary
encoder, one torsional spring assembly with another rotary enco-
der, a pair of spur gear with appropriate gear ratio to transmit
the motion to the ball screw which converts the rotary motion of
the shaft to linear motion of the nut, a set of linear springs attached
to the ball screw nut to transmit the force to a carriage which has a
force output pin to transmit the force to the load (prospective ro-
bot link), and a linear position sensor installed in the carriage to
measure the displacement of the linear spring. The two rotary
encoders measure the angular deection of the torsional spring.
In this design, the stiffness of the linear spring is chosen to pro-
vide the average targeted assistive torque, which is usually about
one third of the peak torque. Therefore the linear spring is soft,
small and light-weight. A very small torsional spring is used to
achieve a very big effective spring constant at the output end be-
cause it is located in the high speed range. As we can see in the
next section, the effective spring constant at the output is more
than one hundred times that of the linear spring. Due to the differ-
ence in spring constant, when the actuator is working in the low
force range, the force control is based on the linear spring and
the torsional spring behaves like a rigid link. However, when the
actuator is working in the high force range, the soft linear is fully
compressed and the force control is based on the torsional spring.
Therefore, we can achieve a much smaller physical size of the over-
all actuator compared to existing SEA designs and make it ideal for
wearable exoskeleton application.
Fig. 3 shows actuator prototype built based on this design for an
exoskeleton system for lower limb rehabilitation for stroke
patients. The actuator is designed to be able to provide up to
Fig. 1. Prototype of the knee ankle robot.
 H. Yu et al. / Mechatronics xxx (2013) xxxxxx 3

where the output load is not xed, one linear potentiometer which
is placed in linear spring as shown in Fig. 2, is used to check the
linear spring deection.
The actuator is driven with a brushless dc motor. Three hall
effect sensors built in the dc motor are connected to the driver
and provide commutation information which are used to identify
the position of the rotor. The commutators and current control
maintain the optimum torque angle.
The controller is the main part of the control system. In our
system, a NI CompactRIO 9074 embedded control and data acqui-
sition system and a personal computer (PC) are used to implement
two force range control laws.
Fig. 2. A novel series elastic actuator.
3.2. Modeling of compliant actuator system

In order to analyze the actuator performance at the output end,

Fig. 3. A prototype of the novel compliant actuator.
60 Nm assistive torque at the lower limb joints. A Maxon DC brush-
less motor (EC 4-pole 120 Watt 36 V) is used for the design due to
its lightweight (0.175 kg) and low moment of inertia. The ball
screw selected from Eichenberger Gewinde AG has a pitch of
2 mm/rev and can output over 1000 N force. The linear springs
have spring constant of 24 N/mm and a working stroke of
10 mm. They can provide an output force of 240 N before fully
compressed. They are used to operate in the range of about 25%
of the full force. The torsional spring has a spring constant of
0.29 Nm/rad and a deection range of 72. The incremental rotary
encoder has a resolution of 1024 lines/rev. The total mass of the
actuator is less than 0.85 kg.
which produces linear output force, the actuator is modeled as a
system consists of only translational elements by converting the
rotary elements to equivalent translational elements. The actuator
model for the equivalent translational motion is shown in Fig. 4(a).
In this model, F is motor input force, m1 is equivalent mass of the
motor plus the torsional spring coupler and the encoder as derived
in (2) where J1 refers to moment of inertia of the motor and the tor-
sional spring coupler and the encoder, p is the pitch of the ball
screw, m2 is equivalent mass of the ball screw and the gears as de-
rived in (3) where J2 refers to moment of inertia for the ball screw
and the gears, k1 is considered as the equivalent translational
spring constant of the torsional spring kt as derived in (4), k2 is
spring constant of the translational spring, b1 and b2 are the viscous
damping for motor and ball screw respectively, and F2 is output
force.

3. Control system and controller design

Controlling a compliant force actuator can be dened as achiev-

ing output forces according to certain requirements. In this section,
we describe a detailed control design for our novel actuator,
including hardware conguration and model-based control
algorithms.

3.1. Hardware of actuation system

To achieve the control objective, a closed-loop force control sys-

tem must be designed. This involves sensor, driver and controller.
In our system, the force is measured according to Hookes law:

F k Dx 1

where k is spring constant and Dx is the displacement of a spring.

This means that a spring with a displacement Dx generates a force
F. With this forcedisplacement relationship, we can use encoder
sensor to determine the spring deection and spring force. In the
control system, two encoders are used: one encoder is placed in
spur gear and the other is placed in dc motor. The former is used
to measure the position of linear spring when the output load is
xed, while the difference between both encoders is used to mea- Fig. 4. Modeling of the series elastic actuator in translational motion. (a) The
sure the angular movement of the torsional spring. For the case general case. (b) Model for low force range. (c) Model for high force range.

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
4 H. Yu et al. / Mechatronics xxx (2013) xxxxxx

Table 1
Parameters of the actuator prototype.

Values of the hardware parameters Values of the hardware parameters for equivalent
for rotational motion translational motion
J1 = 5.517  106 kg m2 m1 = J1(2p/p)2 = 54.4 kg
J2 = 7.406  106 kg m2 m2 = J2(2p/p)2 = 73 kg
Torsional spring kt = 0.29 Nm/rad k1 = kt(2p/p)2 = 2.86  106 N/m
Linear spring k2 = 24  103 N/m k2 = 24  103 N/m
Pitch of the ball screw (p) = 2  103 m p = 2  103 m
Rotary encoder resolution = 1024/rev Force resolution 5.65 N
Linear potentiometer 25 mm Force resolution 0.0092 N
Total weight of actuator 0.84 Kg

m1 J 1 2p=p2 2 where c is the same as in (7).

2 The control objective in this paper is stated as follows: to nd a
m2 J 2 2p=p 3
control mechanism for a given trajectory Fd so that the output force
2
k1 kt 2p=p 4 F1 approaches to Fd as closely as possible.
The physical parameters of the actuator prototype are listed in
3.3. Controller design
Table 1, where both the original number and the equivalent trans-
lational values are given for the dynamic modeling and analysis
3.3.1. Low force control
purpose. We can see that the spring constant k1 derived from the
We rst design the model-based feedback control. Dene the
torsion spring is about 120 times that of k2, meaning the torsional
force error e1(t) = Fd(t)  F1(t) and the derivative error is given by
spring can be considered to be a rigid link when the output force is
e2 e_ 1 t F_ d t  F_ 1 t. Thus, we dene a state being e = [e1, e2]
low and the actuator behaves like a normal SEA with the linear
and the state equation is given by
spring only. The model of the actuator can be simplied as shown  
in Fig. 4(b). When the output force reaches a level when the linear m1 m2 b2 bm 1
e_ A1 e  B1 u B1 Fd F d F d Df x_ 1 d1
spring is fully compressed, the actuator will behave like a SEA with ck2 ck2 c
only the torsional spring as shown the model in Fig. 4(c). 11
Based on the working force range and the assumptions made
above we derive those mechanical models. with
" # " #
0 1 0
3.2.1. Modeling for low force case A1 ; B1 12
 m1km
2
 mb21bm ck2
m1 m2
Referring to Fig. 4(b), the equation of motion according to New- 2 m2

tons second law is given by and Df x1 l1 sgnx_ 1 .

m1 m2 x1 F  k2 x1  x3  b2 x_ 1  x_ 3  bm x_ 1 5 Consider the state [e1, e2]T. It is natural to think that a propor-
tional-derivative (PD) controller should be employed, that is
We can get the system force output by applying the Hooks law,
i.e., F1 = k2(x1  x3). Assuming a xed load end, x3 = 0, the Eq. (5) ut K p F d  F 1 K d F_ d  F_ 1 13
can be written as
where Kp and Kd are the PD parameters which should be chosen
k2 k2 b2 bm _ appropriately. Since the dominant linear model (without nonlinear
F1 F F1  F1 6
m1 m2 m1 m2 m1 m2 and uncertain terms) is given by
However, this model does not consider the effects of the uncer- e_ A1 e  B1 u 14
tainties, such as frictional force and disturbance. Thus, a complete
we will design the PD control based on this model. Substituting the
model for low force case should be given by
control into the above equation yields
ck2 k2 b2 bm _ " #
F1 u F1  F 1 Df x_ 1 d1  7 0 1
m1 m2 m1 m2 m1 m2 e_ e 15
 m1km
2
2
 m1ckm
2
2
Kp  mb21b
m2
m
 m1ckm
2
2
Kd
_ is the frictional force, d1 is the disturbance, c is the con-
where Df x
version coefcient and u is applied current to the dc motor. Since the characteristic equation of the closed-loop system is
 
b2 bm ck2 k2 ck2
3.2.2. Modeling for high force case s2 Kd s Kp 16
Referring to Fig. 4(c), we have the following equation for high
m1 m2 m1 m2 m1 m2 m1 m2
force case. the pole assignment method can be used to design the PD
m1 x1 F  k1 x1  x3  b1 x_ 1  x_ 3  bm x_ 1 8 parameters.
A more quantitative method to select PD parameters is to apply
Similar to the low force case, assuming a xed load end, x3 = 0, the Linear Quadratic Regulator (LQR) design theory.
we have the following model for the output force The LQR performance index is
Z 1
k1 k1 b1 bm _
F1 F F1  F1 9 J eT Q 1 e q1 u2 ; 17
m1 m1 m1 0

Thus, a complete model for high force case is given by where Q1 is a symmetric positive-denite matrix, and q1is a posi-
tive constant. The state feedback controller for the criteria (17) is
ck1 k1 b1 bm _
F1 u F1  F1 10 T
m1 m1 m1 u q1
1 B1 P 1 et; 18

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
 H. Yu et al. / Mechatronics xxx (2013) xxxxxx 5
" #
1
p11
1
p12 " # " #
where P 1 1 1 > 0 is the unique positive denite solution 0 1 0
p21 p22 A2 ; B2 27
to the following equation
 mk11  b1mb1 m ck1
m1

AT1 P1 P1 A1  q1 T The LQR theory is used to design the feedback control. The lin-
1 P 1 B1 B1 P 1 Q 1 0 19
ear model in the above system is
1 1
with Q 1 diagfq1 ; q2 g.
e_ A2 e  B2 u 28
Note that the uncertainties such as nonlinear friction and dis-
turbance are not considered in the above controller design. Next, The performance index for the LQR theory is given by
the design of the sliding mode for the uncertain part of the system Z 1
(11) will be discussed. Although the term Df x_ 1 is nonlinear, it can J eT Q 2 e q2 u2 ; 29
be estimated by using an off-line experiment. Thus, a relay mode is 0

used to compensate the friction where Q2 is a symmetric positive-denite matrix, and q2is a posi-
ur l1 sgnx_ 1 20 tive constant. The feedback control is given below.
T
where l1 should be estimated before the control design. For the u q1
2 B2 P 2 et; 30
term d1, it is uncertain but it is assumed to be bounded by a value, " #
2 2
p11 p12
jd1 j 6 r1 21 where P 2 2 2 > 0 is the positive denite solution to the
p21 p22
Thus, a sliding mode control is used to reject the disturbance following equation
T
us c1 sgne P1 B1 22 AT2 P2 P2 A2  q1 T
2 P 2 B2 B2 P 2 Q 2 0 31
where c1 should be chosen to meet the conditions c1 > r1. with Q 2
2 2
diagfq1 ; q2 g.
Note that the LQR design can guarantee
Combining (18) with (22), the proposed low force control is gi- the stability of the closed-loop system.
ven by For the other terms in the system (26), the feedforward control
u q1 T T is used to compensate them.
1 B1 P 1 et l1 sgnx1 c 1 sgne P 1 B1 uff 1
_ 23
with uff 1 m1 m2
Fd b2 bm
Fd 1
F . m1 b1 bm 1
ck2 ck2 c d uff 2 Fd Fd Fd 32
ck1 ck1 c
Theorem 1. Consider the system (11). If the proposed low force Finally, the proposed composite high force control law is given
controller (23) is applied to the system (11) and the condition c1 > r1 by
is satised, then the state e is asymptotically stable.
T
u q1
2 B2 P 2 et uff 2 ; 33

Proof 1. Consider a Lyapunov function Taking a similar proof as in Theorem 1, we have the following
stability result for the high force.
V eT P1 e 24
Its time derivative is given by Theorem 2. Consider the system (26). If the proposed high force
controller (33) is applied to the system (26), then the state e is
V_ eT AT1 P1 P1 A1 e  2eT P1 B1 u 2eT P1 B1 Df x_ 1 d1  asymptotically stable.
6 eT AT1 P1 P 1 A1 e  2eT P1 B1 u 2eT P1 B1 Df x_ 1 2jeT P 1 B1 jr1
3.3.3. Switching control
where we have used the condition (21). Whenever the state of the system goes from low force range to
Substituting the control (23) into the above inequality yields high force range or from high force range to low force range, the
control law should also change accordingly. The switching law is
V_ 6 eT AT1 P1 P 1 A1  q1 PB1 BT1 P1 e  q1 eT P1 B1 BT1 P1 e designed according to the mechanism model as shown in Fig. 4.
 2jeT P1 B1 jc1 2jeT PBjr1 It can be seen that the force control is naturally switched to the
high force when the force being greater than the maximum value
Since the Eq. (19) holds, it follows that of the low force range, while it is switched to the low force when
V_ 6 eT Q 1 q1 P1 B1 BT1 P1 e  2jeT P1 B1 jc1  r1 25 the force being less than the maximum value of the low force
range. Based on these principles of the switching, we develop the
This implies that V_ < 0 if c1 > r1. The proof of Theorem 1 is com- following control law which results in a switching controller for
pleted. h the compliant actuator
Next, let us see the high force control. 
uH ; F d P F s
u 34
3.3.2. High force control uL ; Fd < Fs
The controller is proposed according to the design method as where uH and uL are the individual control for the high force or the
presented in previous subsection. low force control as described in previous sections, Fd is the desired
Dene the error e1(t) = Fd(t)  F1(t) and the derivative error force, and Fs is the switching point which is the maximum value of
e2(t) = e1(t). Let e(t) = [e1(t), e2(t)]T. The following error equation is the low force range.
obtained Theoretically speaking, the force transition control from the low
  force FL to the high force FH (or from the high force FH to the low
m1 b1 bm 1
e_ A2 e  B2 u B2 Fd Fd Fd 26 force FL) should be seamlessly switched at the switching point Fs
ck1 ck1 c
and executed on the corresponding mechanism system. Unfortu-
with nately, due to control error, such transition is sometime not

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
6 H. Yu et al. / Mechatronics xxx (2013) xxxxxx

 
perfect. For example, when executing the low force control uL, it b1 bm b2 bm  
F_ d  
may act at the mechanism system which belongs to the high force ck1 ck2
   
range; or the high force control uH executed may work at the 
kekkmin Q 2 kek  2kP2 B2 k l1 c1 Fd m1  m1 m2
mechanism system which belongs to the low force range. If such  ck1 ck2
situations occur, a theoretical issue then arises as to what about  
b1 bm b2 bm  
the stability of the closed-loop system under the proposed switch- _
Fd  41
ck1 ck2 
ing control? It is quite important to respond to this question for the
safety reason. The following theorem is given to establish a stabil- where kmin Q 2 minfeigenv aluesQ 2 g.
ity result for the switching control. It follows that V_ < 0 as long as
h 

i
Theorem 3. Consider both the systems 11, 26. If the proposed  
2kP 2 B2 k l1 c1 Fd ck
m1
 m1ckm 2
F_ d b1ckb1 m  b2ckb2 m 
switching controller (34) is applied to the system (11), or the system kek >
1 2

(26) and the following inequalities hold kmin Q 2

T T x2
Q 1 Q 1 2q1 1
2 P 1 B1 B2 P 2  q1 P 1 B1 B1 P 1 > 0 35
T T 42
Q 2 Q 2 2q1 1
1 P 2 B2 B1 P 1  q2 P 2 B2 B2 P 2 > 0 36
then all the states e are uniformly bounded. Then, V_ is negative outside the compact set. This demonstrates that
the state e is uniformly bounded at the high force range.
Proof 2. When applying the switching law (34) to the system, this Case 4. Consider the case where the switching control (34) is
results in four cases: Case 1 when Fd < Fs, the control uL acts at the applied to the actuator which is at the low force range. Recalling
mechanism system which belongs to the low force range; Case 2 the model (11) of the low force range, we have
when Fd P Fs, the control uH acts at the mechanism system which
e_ A1 e  B1 u
belongs to the high force range; Case 3 when Fd < Fs, the control  
uL acts at the mechanism system which belongs to the high force m1 m2 b2 bm 1
B1 Fd F d F d Df x_ 1 d1 43
range; Case 4 when Fd P Fs, the control uH acts at the mechanism ck2 ck2 c
system which belongs to the low force range. The switching control takes the form given by
For the cases 1 and 2, the stability of the closed-loop system can
be guaranteed by Theorems 1 and 2, respectively. Here, we uH q1 T

2 B2 P 2 e uff 2 44
consider the stability issue under the cases 3 and 4.
Case 3. Consider the case where the switching control (34) is Consider the Lyapunov function given by (24). Its time deriva-
applied to the actuator which is at the high force range. Recalling tive along the dynamics (11) with the switching control is given by
the model (26) of the high force range, we have h
  V_ 6 eT AT1 P1 P1 A1 e  2q1 T T
2 e P 1 B1 B2 P 2 e
m1 b1 bm 1
e_ A2 e  B2 u B2 Fd Fd Fd 37 2eT P1 B1 l1 sgnx_ 1 d1
ck1 ck1 c     
m1 m2 m1 b2 bm b1 bm
The switching control takes the form given by 2eT P1 B1 Fd  F_ d 
ck2 ck1 ck2 ck1
T T
uL q1 T T
1 B1 P 1 e l1 sgnx1 c 1 sgne P 1 B1 uff 1
_ 38 eT Q 1 2q1 1
2 P 1 B1 B2 P 2  q1 P 1 B1 B1 P 1 e

Consider the Lyapunov function given by 2eT P1 B1 l1 sgnx_ 1 d1

    
m1 m2 m1 b2 bm b1 bm
V eT P 2 e 39 2eT P1 B1 Fd  F_ d 
ck2 ck1 ck2 ck1
Its time derivative along the dynamics (26) with the switching 45
control is given by
where we have used the result of Theorem 1. Since the inequality
V_ eT AT2 P 2 P2 A2 e  2q1 T T
1 e P 2 B2 B1 P 1 e
(35) holds, it follows that
 2eT P2 B2 l1 sgnx_ 1 c1 sgneT P1 B1 V_ kmin Q 1 2q1 T 1 T 2
2 P 1 B1 B2 P 2  q1 P 1 B1 B1 P 1 kek
          
m1 m1 m2 b1 bm b2 bm 
m1 m2  m1 F_ d b2 bm  b1 bm 

2eT P2 B2 Fd  F_ d  2kekP1 B1 l1 sgnx_ 1 d1 F d 
ck1 ck2 ck1 ck2 ck
2 ck
1 ck ck2 1
T T 2
T T 6 kmin Q 1 2q1 1
2 P 1 B1 B2 P 2  q1 P 1 B1 B1 P 1 kek
eT Q 2 2q1 1
1 P 2 B2 B1 P 1  q2 P 2 B2 B2 P 2 e      
 
2kekkP1 B1 k l1 r1  m1 m2  m1 F_ d b2 bm  b1 bm 
 2eT P2 B2 l1 sgnx_ 1 c1 sgneT P1 B1 F d ck2 ck1 ck2 ck1 
    
m1 m1 m2 b1 bm b2 bm kekkmin Q 1 kek
2eT P2 B2 Fd  F_ d       
ck1 ck2 ck1 ck2  
m1 m2  m1 F_ d b2 bm  b1 bm 
 2kP1 B1 k l1 r1 
F d 
40 ck2 ck1 ck2 ck1

where we have used the Eq. (31). Since the inequality (36) holds, it
where kmin Q 1 minfeigenv aluesQ 1 g.
follows that
It follows that V_ < 0 as long as
 h 

i
V_ 6 kmin Q 2 kek2 2kekkP2 B2 l1 sgnx_ 1 c1 sgneT P 1 B1
    2kP1 B1 k l1 r1 Fd m1 m2
ck2
m1
 ck

F_ d b2ckb2 m  b1ckb1 m 
m1 m1 m2 b1 bm b2 bm kek >
1
x1
Fd  F_ d  k kmin Q 1
ck1 ck2 ck1 ck2
   

6 kmin Q 2 kek2 2kekkP 2 B2 k l1 c1  d m1  m1 m2
F Then, V_ is negative outside the compact set. This demonstrates
 ck1 ck2
that the state e is uniformly bounded at the low force range.

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
 H. Yu et al. / Mechatronics xxx (2013) xxxxxx 7

3277 1
2723

Input Signal
0.5

Voltage (V)
723
0

1277
0.5

3277 1
0 0.1 0.2 0.3 0.4 0.5
Time (s)
Fig. 5. Experimental set-up. (a)
4267 200
4. Experimental results
2400 100

Output Signal

Force (N)
In this section, the proposed control is applied to a real compli- 200 0
ant actuator. The experimental setup is shown in Fig. 5. The system
consists of the linear compliant actuator, two encoders, a motor 2000 100
driver and a PC with the controller. The compliant actuator is actu-
ated by one brushless dc motor (MAXON 311537), capable of gen- 4200 200
erating 1130mNm stall torque. The rst encoder (it is Encoder 1 in
6400 300
Fig. 5) is used to check the low force, while this encoder with other 0 0.1 0.2 0.3 0.4 0.5
encoder (it is Encoder 2 in Fig. 5) is used to measure the high force. Time (s)
The motor driver is Elmo Harmonica 5/60 which can provide a
maximum power output 200 W. The controller is the NI Compac- (b)
tRIO 9074 programmable automation controller that is an ad-
Fig. 7. Signals used in identication: low force range. (a) Input signal. (b) Output
vanced embedded control and data acquisition system designed response.
for applications that require high performance and reliability. In
the controller, we use NI 9215 (Analog Input module), NI 9263
(Analog Output module), and NI 9516 (Encoder module) for the signal and the output force response of the open-loop actuator sys-
data acquisition. The entire control algorithm is written into two tem are shown in Fig. 7. It should be noted that the input/output
modules: monitoring (signal generator and program monitoring) signals obtained are based on 32 bit FPGA and have no units. To
and FPGA (control algorithm). The functional block-diagram of illustrate their corresponding actual meanings, the second labels
the entire control system is shown in Fig. 6. The sampling period of the Y-axis are given in the gures. Based on the input/output sig-
for our test is chosen as 0.5 ms. nals, we use MATLAB to identify the model, where the least square
identication algorithm is used. With the help of the System
Identication Toolbox of MATLAB, the linear model of the actuator
4.1. Model identication is given by

For the control application, it is necessary to identify the model F1 208:3304F 1  11:838F_ 1 197:22u 46
of the compliant actuator. From the analysis of Section 3, it is In order to evaluate the identication performance, the
known that the model structure is a linear second order model variance-accounted-for factor is chosen as Bestt model, which is
which is described by a relationship from the input signal u to dened as
the output signal F1. The system identication is to identify the !
parameters of the model from available inputoutput data. The v arF 1  bF 1
Bestfit 100  1 47
choice of an input signal should contain many frequencies. As v arF 1
the actuator works at low frequency, square wave signal is enough
for our application. In the experiment, the model of the low force where b
F 1 is the model output. The comparison of the simulated out-
range is identied rst. The input signal for the test consists of a put and the actual measured output is shown in Fig. 8 and Bestt in
square and 1 Hz which is around the working frequency. The input MATLAB Toolbox is 88.18. As observed, the model output is closed

Fig. 6. Functional block of system.

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
8 H. Yu et al. / Mechatronics xxx (2013) xxxxxx

Force: Measured and Model Output to the measured actual output force, although they are slightly
4267
Measured output different at some points. The nonlinear model is given by
Model output
2400 F_ A1 F  B1 u B1 Df x_ 1 d1  48
Output Signal

with
200
 
T 0 1
F F 1 ; F_ 1  ; A1 ;
2000 208:3304 11:8381
 
0
4200
B1 49
197:22
where Df x_ 1 is friction force and d1 is the disturbance.
6400
0 0.1 0.2 0.3 0.4 0.5 Next, the high force model is identied. Since the high force is
Time (s) large, square wave signal may damage the hardware due to fre-
quent jumping. To prevent this occurrence, the identication for
Fig. 8. Model validation at low force range: dashdot line represents measured the high force range is based on the step response. The input and
signal; solid line represents model output. output signals used in the experiment are shown in Fig. 9. As indi-
cated in the identication of the low force range, since the signals
obtained have no units, the second label is also given in Fig. 9 to
show the corresponding physical meanings. With the help of the
13107 4
12000 System Identication Toolbox of MATLAB, the following linear
3.5
model is obtained.
10000 3
Input Signal

Voltage (V)

8000 2.5 F1 1:3587  106 F 1  6:6911  103 F_ 1 1:1198  106 u 50

2
6000 The comparison of the simulated output and the actual mea-
1.5
4000 sured output is shown in Fig. 10 and Bestts in MATLAB Toolbox
1
are 94.63. The model is given by
2000 0.5
0 0 F_ A2 F  B2 u 51
0 0.5 1 1.5 2
Time (s) with
(a)  
0 1
A2 ;
8533 400 1:3587  106 6:6911  103
 
0
6733 300 B2 52
1:1198  106
Output Signal

4533 200
Force (N)

2333 100 4.2. Controller design

133 0
Based on the models obtained above, we design the low force
2067 100
and high force controllers according to the schemes proposed in
4267 200 Section 3.
0 0.5 1 1.5 2
Low force controller. According to our design, the low force is
Time (s) generated by the linear spring and its range is from 0 to 240 N.
(b) The PD controller is designed based on the LQR approach. The
selection of the weighting matrix is Q1 = diagf700; 0:5g and
T
Fig. 9. Signals used in identication: high force range. (a) Input signal. (b) Output q1 = 1. Therefore, the PD gain becomes q1 1 B1 P 1 which is obtained
response. from (19), giving Kp = 25.42, Kd = 0.81. The eigenvalues of the linear
closed-loop system are {5.92 13.16i} which are all negative.
This implies that the designed PD feedback control system is stable
in the sense of the linear model. On the other hand, we choose the
Force:Measured and Model Output
8533
Model output
following mode to compensate the friction and disturbance.
Measured output
6733
us l1 sgnx_ 1 c1 sgneT P1 B1 53
Output Singal

4533 The coefcient of the friction can be estimated by checking

whether the force commences for a given a small amount of input
2333
signal; if it does not have, i.e., it is still zero, we have to increase the
133 input signal until the force occurs. Fig. 11 shows the output force
commences when the input signal is 850 which is the value of
2067 l1. For the coefcient c1 of the sliding mode, it can be estimated
by checking whether the tracking performance is improved when
4267 a small value of c1 is given; if no, increase its value until the perfor-
0 0.5 1 1.5 2
mance is improved. Finally, it is determined by experimental test,
Time (s)
that is 300. Then the proposed control law (23) is given by
Fig. 10. Model validation at high force range: dashdot line represents measured
signal; solid line represents model output. u 25:42e 0:81e_ 850sgnx_ 1 300sgneT P1 B1 uff 1 54

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
 H. Yu et al. / Mechatronics xxx (2013) xxxxxx 9

1.5 100
Actual
1.25 Desired

80
Force (N)

Force (N)
0.75 60
0.5

0.25 40

0
0 0.1 0.2 0.3 0.4 0.5 20
Time (s)
0
Fig. 11. Friction estimation. 0 0.5 1 1.5 2
Time (s)
(a)
100
Actual
Desired
3

80
2

Force error (N)

Force (N)

60
1

40 0

20 1

0 2
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (s) Time (s)
(a) (b)
1.4 Fig. 13. Tracking control performance of low force with a frequency of 2 Hz (the PD
control).
1.2
Force error (N)

0.8 100
Actual
Desired
0.6
80
0.4
Force (N)

0.2 60

0
0 0.5 1 1.5 2 40
Time (s)

(b) 20

Fig. 12. Tracking control performance of low force range 1 Hz. 0

0 0.5 1 1.5 2
Time (s)
where uff 1 1:056F d 0:06F_ d 0:0051Fd . (a)
High force controller. Taking a similar design as in the low force
control, the PD control of the high force is also obtained based on 1.5
the LQR approach. For given Q2 = diagf500; 0:000001g and q2 = 1,
the PD gains are Kp = 21.18, Kd = 0.002. The proposed control law
Force error (N)

1
is given by

u 21:18e 0:02e_ uff 55 0.5

where uff 1:2133F d 0:006F_ d 8:93  106 Fd .

0

4.3. Performance assessment

0.5
0 0.5 1 1.5 2
In order to verify the effectiveness of the proposed control law, Time (s)
the actuator system is controlled to follow the force reference tra- (b)
jectory which is a sinusoidal signal. We will evaluate the control
performance under the proposed low force control, high force con- Fig. 14. Tracking control performance of low force with a frequency of 2 Hz (the
trol and switching control modes. proposed control).

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
10 H. Yu et al. / Mechatronics xxx (2013) xxxxxx

310
Actual
Desired
300

290
Force (N)

280

270

260

250

240
0 1 2 3 4 5
Time (s)

(a)
15
Fig. 17. Ankle rehabilitation robot.
10
4.3.1. Low force tracking control
Force error (N)

5 Using the proposed control law (54), Fig. 12 shows the control
results, where the top gure is the output force response and the
0 bottom one is the force tracking error. Thanks to the low-compli-
ance spring, it is observed that the maximum tracking error
5
achieved is very small, that is 1.25 N. Alternative way to evaluate
10 the control performance is based on the variance-accounted-for
factor, checking the delity of the generated output force (readers
15 can refer to [11] for the detailed content). The value for the force
0 1 2 3 4 5
delity according to the given denition of [11] is 99.9%. As ex-
Time (s)
pected, the force prole generated by the low-compliance spring
(b) has very high accuracy.
To further evaluate the performance of the proposed controller,
Fig. 15. Tracking control performance of high force with a frequency of 1 Hz. it is also compared with the PD control without the friction com-
pensation and sliding mode scheme. The reference with a fre-
quency of 2 Hz is used as a desired force prole. For a fair
comparison, the PD control parameters are the same as the pro-
posed controller. Fig. 13 shows the results using the PD control.
320
Actual
With the same reference trajectory, the control performance using
300 Desired the proposed controller is also measured and the results are shown
SP
in Fig. 14. It is observed that the force control reaches to a maxi-
280 mum tracking error of 3 N by the PD control, compared to a max-
Force (N)

260 imum tracking error of 1.25 N when the proposed control scheme
is used. This shows that the proposed controller can offer better
240
control performance than that of the conventional PD control.
220

200 4.4. High force tracking control

180
0 0.5 1 1.5 2 The high force is generated by the torsional spring and it is
Time (s) greater than 240 N in our design. When entering a high force con-
trol, the proposed high force control (55) has to be used. Fig. 15
(a) shows a high force control result with a frequency of 1 Hz. It can
be seen that the force generated can track the desired force trajec-
20
tory well. However, due to coarse force resolution which is 5.65 N,
15 the output force has an error of about 15 N, about two times of the
resolution. The force delity obtained for this tracking is 91.5%.
Force error (N)

10 Considering the high force range, the result is acceptable.

5
4.4.1. Switching control
0 In a practical situation, a more complicated case for the force
5
prole is that it is involved in both low force and high force ranges.
In this case, to track the force prole, we have to use the switching
10 control, which is given by (34). Since the switching control is
0 0.5 1 1.5 2
served for both force ranges. It is important to check whether or
Time (s) not the switching conditions (35), (36) hold. Here, it is found that
(b) the proposed low and high force controllers (54), (55) can achieve
Q 1 > 0; Q 2 > 0. This implies that the designed switching control
Fig. 16. Switching control results where SP is denoted as switching point. law can guarantee the stability of the closed-loop system. Fig. 16

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
 H. Yu et al. / Mechatronics xxx (2013) xxxxxx 11

Fig. 18. Walking experiment.

5.65 N with the current encoder and torsional spring stiffness used
80
in the design. From the gure, it can be seen that although the con-
Actual troller switches between low and high forces frequently, the
60 Desired
closed-loop system is still stable. This further veries our theoret-
40 ical analysis in Theorem 3.
In summary, it can be concluded that: (1) the LQR-based PD
Force (N)

20 parameter tuning is a good way to guarantee the closed-loop

0 system performance and (2) switching control law is necessary
to implement a stable transition between low and high force
20
ranges.
40

60 4.5. Experiments with a human subject

0 1 2 3 4 5
Time (s) Although the proposed actuator can generate accurate forces
(a) following a desired trajectory as shown in the experiments above,
it is necessary to further test the rehabilitation robot with a human
10 subject. The actuator presented in this paper was used to construct
an ankle rehabilitation robot, as shown in Fig. 17.
In actual rehabilitation therapy, the desired force prole has to
5
be determined based on the user needs at different gait phases. As
Force error (N)

the focus of this paper is on generating accurate forces from the

0 actuator controlled, we use sinusoidal force signals as desired force
5
10
0 1 2 3
Time (s)
(b)
Fig. 19. Ankle walking with a frequency of 1 Hz.
shows a switching force control result with a frequency of 1 Hz,
where the switching point is Fs = 240 N. Initially, the low force con-
trol (54) is employed until the desired output force reaches to
240 N. Then, the switching control is activated and the force con-
trol enters the high force range. It is observed from the gure that
a satisfactory result (<18 N) is achieved using the proposed switch-
ing law. This is also veried by measuring the force delity which
is 97.7%. It can also be noted that at the low force range the signal
is very smooth, but at high force range, the measured signal
zigzags. This is because the force resolution at high force range is
trajectories for testing purpose.
The subject was asked to follow the force generated by the
rehabilitation robot to do motions, as shown in Fig. 18. Since the
load (the subject) is not xed, i.e., x3 0 (see Fig. 4), the generated
output force is measured by using the linear potentiometer whose
4 5
physical specications are shown in Table 1. We tested the system
performance at a frequency of 1Hz. Fig. 19 shows the control re-
sults, where the top gure is the output force response, the bottom
gure is the force tracking error. It is observed that the desired
force trajectory is tracked with a maximum tracking error of
6.5N which is increased when compared to the actuator control
without human factor. This is because the human motions intro-
duce some uncertainties and disturbances into the robot, which af-
fect the control performance. But the control performance is
acceptable as well as the closed-loop system is still stable.
5. Conclusion
We have presented a novel compliant actuator design and the
control system for this actuator. By introducing one extra torsional

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004
12 H. Yu et al. / Mechatronics xxx (2013) xxxxxx

spring at the high force range, a truly compliant actuator with [11] Veneman JF, Ekkelenkamp R, Kruidhof R, van der Helm F, van der Kooij H. A
series elastic- and bowden-cable-based actuation system for use as torque
excellent force control delity. Theoretical analysis for the
actuator in exoskeleton-type robots. The International Journal of Robotics
controller design has been given to guarantee the stability of the Research 2006;25(3):26181.
closed-loop system, especially for the switching control law. [12] Veneman JF, Kruidhof R, Hekman E, Ekkelenkamp R, Van Asseldonk E, van der
Experimental results have conrmed that the proposed controller Kooij H. Design and evaluation of the lopes exoskeleton robot for interactive
gait rehabilitation. IEEE Transactions on Neural System, Rehabiliation and
can achieve good performance for the low, high force, and Engineering 2007;15(3):37986.
switching control. In the next stage of our research, we will [13] F. Sergi, D. Accoto, G. Carpino, N.L. Tagliamonte, E. Guglielmelli, Design and
collaborate with clinical partners to test our exoskeleton robot characterization of a compact rotary series elastic actuator for knee assistance
during overground walking, in: The fourth IEEE RAS/EMBS International
for gait rehabilitation. Conference on Biomedical Robotics and Biomechatronics, Roma, Italy, June
2427, 2012, pp. 19311936.
References [14] M. Grun, R. Muller, U. Konigorski, Model based control of series elastic
actuators, in: The fourth IEEE RAS/EMBS International Conference on
Biomedical Robotics and Biomechatronics, Roma, Italy, June 2427, 2012, pp.
[1] Yu H, Spenko M, Dubowsky S. An adaptive shared control system for an
538543.
intelligent mobility aid for the elderly. Autonomous Robots 2003;15:5366.
[15] D. Robinson, Design and Analysis of Series Elasticity in Closedloop Actuator
[2] C. Zhu, M. Oda, H. Yu, H. Watanabe, Y. Yan, Walking support and power
Force Control, Ph.D. thesis, Massachusetts Institute of Technology, Cambridge,
assistance of a wheelchair typed omnidirectional mobile robot with
June, 2000.
admittance control mobile robots current trends, in: Zoran Gacovski (Ed.),
[16] H.Yu, S.N. Huang, G. Chen, S.L. Toh, M. STA Cruz, Y. Ghorbel, C. Zhu, Y. Yin,
ISBN: 978-953-307-716-1, InTech., 2011.
Design and analysis of a novel compact compliant actuator with variable
[3] T. Nef, M. Mihelj, G. Colombo, R. Riener, ARMin-robot for rehabilitation of the
impedance, in: Proceedings of IEEE International Conference on Robotics and
upper extremities, in: Proceedings of the IEEE International Conference on
Biomimetics (ROBIO), 2012, pp. 11881193.
Robotics and Automation, 2006, pp. 31523157.
[17] H.Yu, M. STA Cruz, G. Chen, S.N. Huang, C. Zhu, E. Chew, Y.S. Ng, N.V. Thakor,
[4] Jezernik S, Colombo G, Keller T, Morari F. Robotic orthosis lokomat: a
Mechanical design of a portable kneeanklefoot Robot, in: Proceedings of
rehabilitation and research tool. Neuromodulation 2003;6(2):10815.
2013 IEEE International Conference on Robotics and Automation (ICRA), 2013,
[5] M. Wassink, R. Carloni, S. Stramigioli, PortHamiltonian analysis of a novel
pp. 21752180.
robotic nger concept for minimal actuation variable impedance grasping, in:
[18] D.W. Robinson, G.A. Pratt, Force controllable hydro-elastic actuator, in:
Proceedings of the 2010 IEEE International Conference on Robotics and
Proceedings of 2000 IEEE International Conference on Robotics and
Automation, 2010, pp. 771776.
Automation (ICRA), vol. 2, 2000, pp. 13211327.
[6] G.A. Pratt, M.M. Williamson, Series elastic actuators, in: IEEE International
[19] Vallery H, Veneman J, Asseldonk E, Ekkelenkamp R, Buss M, Kooij H. Compliant
Workshop on Intelligent Robots and Systems (IROS 1995), Pittsburg, USA,
actuation of rehabilitation robots. IEEE Robotics & Automation Magazine
1995, pp. 399406.
2008;15(3):609.
[7] Sugar T. A novel selective compliant actuator. Mechatronics
[20] G. Wyeth, Control issues for velocity sourced series elastic actuators, in:
2002;12(9):115771.
Proceedings of the Australasian Conference on Robotics and Automation,
[8] Pratt J, Krupp B, Morse C. Series elastic actuators for high delity force control.
Auckland, 2006.
Industrial Robot 2002;29(3):23441.
[21] H. Vallery, R. Ekkelenkamp, H. van der Kooij, M. Buss, Passive and accurate
[9] D.W. Robinson, J.E. Pratt, D.J. Paluska, G.A. Pratt, Series elastic actuator
torque control of series elastic actuators, in: Proceedings of the IEEE/RSJ
development for a biomimetic walking robot, in: Proceedings of the 1999 IEEE/
International Conference on Mechatronic and Embedded Systems and
ASME International Confernece on Advanced Intelligent Mechatronics, 1999,
Applications, 2006, pp. 16.
pp. 561568.
[22] Kong K, Bae J, Tomizuka M. Control of rotary series elastic actuator for ideal
[10] J.E. Pratt, B.T. Krupp, C.J. Morse, S.H. Collins, The roboknee: an exoskeleton for
force-mode actuation in humanrobot interaction applications. IEEE/ASME
enhancing strength and endurance during walking, in: Proceedings of the 2004
Transactions on Mechatronics 2009;14(1):10518.
IEEE International Conference on Robotics and Automation, 2004, pp. 2430
[23] Slotine JJE, Li W. Applied Nonlinear Control. Prentice-Hall; 1991.
2435.

Please cite this article in press as: Yu H et al. Control design of a novel compliant actuator for rehabilitation robots. Mechatronics (2013), http://dx.doi.org/
10.1016/j.mechatronics.2013.08.004