Joonbum Bae1

Department of Mechanical Engineering,

University of California,
Berkeley, CA 94720
e-mail: jbbae@me.berkeley.edu

Kyoungchul Kong
Department of Mechanical Engineering,
Sogang University,
Seoul, Korea 121-742
e-mail: kckong@sogang.ac.kr

Masayoshi Tomizuka
Department of Mechanical Engineering,
University of California,
Berkeley, CA 94720
e-mail: tomizuka@me.berkeley.edu

Gait Phase-Based Control for a

Rotary Series Elastic Actuator
Assisting the Knee Joint
Actuators for physical human-robot interaction (pHRI) such as rehabilitation or assistive
systems should generate the desired torque precisely. However, the resistive and inertia
loads inherent in the actuators (e.g., friction, damping, and inertia) set challenges in the
control of actuators in a force/torque mode. The resistive factors include nonlinear
effects and should be considered in the controller design to generate the desired force
accurately. Moreover, the uncertainties in the plant dynamics make the precise torque
control difficult. In this paper, nonlinear control algorithms are exploited for a rotary series elastic actuator to generate the desired torque precisely in the presence of nonlinear
resistive factors and modeling uncertainty. The sliding mode control smoothed by a
boundary layer is applied to enhance the robustness for the modeling uncertainty without
chattering phenomenon. In this paper, the rotary series elastic actuator (RSEA) is installed on the knee joint of an orthosis, and the thickness of the boundary layer is
changed by gait phases in order to minimize the torque error without the chattering phenomenon. The performance of the proposed controller is verified by experiments with
actual walking motions. [DOI: 10.1115/1.4004793]

Introduction

As applications that involve physical human-robot interaction

(pHRI), such as power augmentation or rehabilitation systems, are
gaining great attention, actuators for pHRI have been investigated.
Actuators for ideal pHRI are controlled in a force/torque mode,
which requires compensation of the inherent mechanical impedance including nonlinear friction. The nonlinear relation between
the control input and the generated torque output also obstructs
the natural pHRI.
Series elastic actuators have been proposed to fulfill the requirements of actuators for pHRI [14]. The springs utilized in the series elastic actuators enable compensation of the mechanical
impedance inherent in electric motor systems. In our previous
work [5], a rotary series elastic actuator (RSEA) was proposed
based on the similar principle but with a torsional spring. The performance of series elastic actuators, including the RSEA, depends
on the algorithm that controls the deflection of the spring. Such
control algorithms play the role of the lowest level controller in
pHRI systems to generate the desired torque accurately.
The control of series elastic actuators is challenged by the inherent nonlinear friction in geared motor systems and modeling
uncertainty. To compensate for such undesired factors, a disturbance observer (DOB) was applied in our previous work [5]. The
DOB treats the friction and modeling uncertainty as an external
disturbance and rejects them by feeding back the estimated disturbance. The proposed control method for the RSEA has been
deployed in actual pHRI applications and showed good performance. However, the difficulty in the design of DOB (e.g., unexpected instability shown in [6]) still remains as a challenge.
In this paper, nonlinear control algorithms are exploited for the
RSEA to generate the desired torque accurately in the presence of
nonlinear resistive factors and modeling uncertainty. In general,
the design of nonlinear control algorithms is more straightforward
than linear robust controllers for systems with known nonlinear1
Corresponding author.
Contributed by the (Division or Committee) of ASME for publication in the
JOURNAL OF MEDICAL DEVICES. Manuscript received December 31, 2010; final
manuscript received July 26, 2011; published online August 23, 2011. Assoc. Editor:
Ming-Yih Lee.

Journal of Medical Devices

ities. For example, the nonlinear Coulomb friction in the RSEA

can be directly rejected by the nonlinear control input. In this paper, the sliding mode control is applied since it shows better
robustness for modeling uncertainty than the feedback linearization control. However, the chattering phenomenon caused by the
sliding mode controller should be avoided for natural pHRI. By
applying a boundary layer around the sliding surface, the chattering effect can be decreased, but it deteriorates tracking performance. Since the desired motor angle is dependent on the human
motions, the thickness of the boundary layer is changed according
to human motions to minimize the tracking error without chattering phenomena. In this paper, the RSEA is installed on the knee
joint and the thickness of the boundary is changed by two major
gait phases, i.e., swing and stance phases. The performance of the
proposed control algorithm is verified by experiments with actual
human walking motions.
This paper is organized as follows. In Sec. 2, preliminary materials such as design of the RSEA and motor friction identification
are introduced. System modeling and nonlinear controller design
for the RSEA are presented in Sec. 3. The experimental results of
the proposed control algorithm are shown in Sec. 4. Summary and
conclusions are given in Sec. 5.

Preliminary Works

2.1 A Rotary Series Elastic Actuator (RSEA). A rotary series elastic actuator (RSEA) has been proposed for assisting
human motions [5]. The RSEA consists of a dc motor, a torsional
spring, and two encoders each on the human side and the motor
side. The torsional spring is used to generate the desired torque,
and acts as an energy buffer between the actuator and the human
joint. The similar approaches are shown in [14], where they
applied a linear spring for the same purpose. Since a linear spring
requires torque arms for generating torque, a torsional spring is
directly installed between the human joint and the motor in the
design of the RSEA. The position of the dc motor is controlled to
have the proper spring deflection such that the RSEA generates
the desired torque precisely. The overall design of the RSEA is
shown in Fig. 1, and the details are given [5].

C 2011 by ASME
Copyright V

SEPTEMBER 2011, Vol. 5 / 031010-1

Downloaded From: http://medicaldevices.asmedigitalcollection.asme.org/ on 02/07/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Table 1

Coefficient of resistive torque in the RSEA [5]

a1 Nm

a2 Nm

a3 Nm=rads1

1:509  100

9:572  101

1:414  101

Coefficient (unit)
Value

a1 , a2 , and a3 were found using a curve fitting method with given

data, and the values are shown in Table 1.

3 Nonlinear Controller Design for a Rotary Series

Elastic Actuator

Fig. 1

A rotary series elastic actuator (RSEA) [5]

2.2 Identification of Motor Resistive Torque. The actuating torque is generated by a geared dc motor in the RSEA. Geared
motors have been widely used in pHRI applications due to their
great controllability and flexibility, but nonlinear resistive torque
introduced by a gear reducer makes the accurate generation of the
desired torque difficult. The resistive torque is undesirable characteristics for pHRI actuators since a human has to make an additional effort to overcome the resistive torque. Thus the resistive
torque should be compensated for improved pHRI.
The measured resistive torque in the RSEA is shown in Fig. 2.
In an ideal case the resistive torque should be zero regardless of
the angular velocity when the control input is zero (labeled Target
in Fig. 2). However, due to the friction force, there is the resistive
torque in actual device (labeled Actual in Fig. 2). Note that the
magnitude of the bias force is so large that the motor rotates even
under zero control input. This phenomenon often occurs in applications of dc motors. The discontinuity at zero angular velocity
represents the Coulomb friction force.
The resistive torque shown Fig. 2 is modeled as follows:
sresistive a1 a2 sgnh_M a3 h_M

(1)

where the coefficients a1 , a2 , and a3 represent terms due to bias,

nonlinear friction, and linear damping, respectively. The values of

3.1 System Modeling. A schematic diagram of the RSEA installed on a human joint is depicted in Fig. 3. IM is the inertia of
the motor, and hM and hH are the angles of the motor and the
human joint, respectively. sM represents the motor torque, and
sresistive is the resistive torque. The motor and the human joint are
connected via a torsional spring with spring constant k, and the
controlled output is the spring torque which is proportional to the
spring deflection, i.e., the difference between the motor and the
human joint angle.
The following relations are obtained from Fig. 3 by applying
Newtons law and Hookes law,
IM hM sM t khH  hM  sresistive

where sresistive is defined in (1).

For a state space system model, let x1 hM , x2 h_M , u sM ,
and y x1 . Then, the resulting state space system equation is
x_ 1 x2
x_ 2 

k
1
sresistive
k
x1 u 
hH
IM
IM
IM
IM

Experimental results for the motor friction test [5]

031010-2 / Vol. 5, SEPTEMBER 2011

(3)

y x1
3.2 Feedback Linearization. The feedback linearization
control technique algebraically transforms a nonlinear system into
a linear one such that linear control methods can be applied [7,8].
The nonlinearities in the system are canceled by the feedback linearization control input; thus the closed-loop dynamics is in a linear form. To define the control law for the feedback linearization
controller, the output y is differentiated until the control input u
appears. In case of the system in (3), the control input u appears in
the second derivative of y, which means the relative degree of the
system is two.
y_ x_1 x2

Fig. 2

(2)

Fig. 3

(4)

Schematic plot of the human joint and the RSEA

Transactions of the ASME

Downloaded From: http://medicaldevices.asmedigitalcollection.asme.org/ on 02/07/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

y x_ 2 

k
1
sresistive
k
x1 u 
hH  v
IM
IM
IM
IM

(5)

where v is the synthetic input. To determine the synthetic input, a

pole placement method is applied, i.e.,
v c0 y  yd  c1 y_  y_d yd

Note that the human joint angle hH appears in the feedback linearization control law. This contributes to compensation of the
human factors imposed on the motor side. However, there are
drawbacks in feedback linearization such as (1) it requires accurate model parameters, and (2) there exists internal dynamics if
the relative degree is smaller than the dimension of the system.
The internal dynamics is problematic if it is unstable. The internal
dynamics does not exist in the present problem, since the relative
degree is equal to the dimension of the system. The requirement
of an exact model; however, may cause problems in actual implementation since there is always modeling inaccuracy in the nominal model.
3.3 Robustness Enhancement by Sliding Mode Control. The
drawback of the feedback linearization control, i.e., the requirement of accurate model parameters, can be overcome by sliding
mode control. Sliding mode control deals with modeling uncertainty by applying additional terms to a nominal model [7,8]. In
this section, the design of a typical sliding mode controller for the
system in (3) is reviewed briefly.
Given the spring deflection E hM  hH , the error between the
actual and the desired deflection and its derivative are defined as
follows:
e E  Ed hM  hH  hMd  hH hM  hMd

(8)

e_ h_M  h_Md

(9)

where Ed is the desired spring deflection to generate the desired

torque. The sliding surface S is defined by


d
S
k e
dt
e_ ke
(10)
(11)

where k is a positive constant.

To assure that the tracking error e converges to zero, the sliding
variable must converge to zero, which takes place if
SS_ < gjSj

(12)

where g is a positive constant.

S_ can be calculated as follows:
S_ hM  hM kh_M  h_Md
k
a3
a2
a1
k
 x1  x2  sgnx2  hH  hMd
IM
IM
IM
IM IM
1

u kh_M  h_Md
IM
f x CEx bxu
Journal of Medical Devices

f x 

k
a3
a2
a1
k
x1  x2  sgnx2  hH  hMd
IM
IM
IM
IM IM
CEx kh_M  h_Md

(6)

The values of c0 and c1 are chosen such that the equation is

asymptotically stable, i.e., c0 > 0 and c1 > 0 in (6). Thus the
overall control input for feedback linearization is


k
sresistive
k
(7)
 hH
u I M v x1
IM
IM
IM

S h_M  h_Md khM  hMd

where

(13)

1
IM

(14)
(15)
(16)

The parameters in f x and bx are to be obtained by system

identification. However, the assumed model shown in (3) may not
fully reflect the actual dynamics due to modeling uncertainties.
Therefore, f x and bx may depend on unmodeled dynamics
and time varying dynamics as well as the nominal dynamics. Thus
suppose f x and bx are composed of the nominal model f^x
^ and uncertain parts Df x and Dbx, respectively, i.e.,
and bx
f x f^x Df x

(17)

b b^  Db

(18)

It is assumed that Df x and Dbx are state dependent, and that

their upper bound can be found as follows:
jDf xj  ax

(19)

bmin
bmax
 jDbj 
^
b
b^

(20)

A control input u that satisfies the condition in (12) is

u

1
f^x  CEx  K  sgnS
^
bx

(21)

where K is chosen to guarantee the condition (12) such that the

system is properly controlled even in the presence of modeling
uncertainties. To select a proper K, the following worst cases for
Df x and Dbx are assumed, i.e.,
Df x ax

(22)

bmin
 bmin
b^

(23)

1  bmin f^x CEx a g

bmin

(24)

Db
Then K can be given by
K

The control law in (21) with K in (24) always satisfies the condition in (12). Note that once S 0 is achieved, the error et converges to zero for any k > 0 and 1=k is the time constant of the
error convergence. However, since the control law is discontinuous across the sliding surface S 0, it introduces chattering,
which is not desirable in general. In particular, the noise and
vibration caused by the chattering phenomenon are not desirable
in the pHRI application.
3.4 Gait Phase-Based Smoothed Sliding Mode Control. To
reduce the chattering phenomenon in the sliding mode control, a
saturation function shown in Fig. 4 is often applied instead of a
signum function [7,8]. The boundary layer of the saturation function can be considered as the extended sliding surface of the
signum function S 0. Outside of the boundary layer, the control
law is chosen to satisfy the condition in (9), which guarantees that
the boundary layer is attractive [7]. In the boundary layer, the discontinuous control activity is replaced by a continuous control.
The signum function used in (21) can be thought that the thickness
of the boundary layer is zero.
SEPTEMBER 2011, Vol. 5 / 031010-3

Downloaded From: http://medicaldevices.asmedigitalcollection.asme.org/ on 02/07/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Fig. 4

S
Saturation function, sat(U
)

The control input of the smoothed sliding mode control is given

by
u


 
1
S
f^x  CEx  K  sat
^
U
bx

(25)

By applying the saturation function instead of the signum function, the chattering phenomenon can be decreased, but the tracking performance is deteriorated. By adjusting the thickness of the
boundary layer, the chattering phenomenon and the tracking error
can be traded off. That is, if the thickness of the boundary layer is
close to zero, then the controller acts like the sliding mode controller with a signum function, which shows more chattering and
less tracking error. On the contrary, if the thickness of the boundary layer is large, then the chattering phenomenon disappears but
the tracking performance is much deteriorated.
The torque output of the RSEA s is generated by the spring
deflection between the motor and the human joint, i.e.,
s khM  hH

(26)

where k is the spring constant. Given the desired torque sd the

desired spring deflection is skd . Then the desired motor angle hMd is
determined by
hMd

sd
hH
k

The swing and stance phases are detected by Smart Shoes

[912]. Smart Shoes measure the ground contact forces (GCFs)
by four force sensing units embedded under the insole. In the
swing phase, all GCFs are zero since the foot is in the air. In
the stance phase, the sum of the GCFs indicates the body
weight since the whole body weight is transferred to the foot.
At the heel strike phase (the first phase of the stance phase), the
sum of the GCFs is over the body weight due to the impact
force at the heel strike. In the experiment in the next section, if
the sum of the GCFs is greater than 5% of the body weight
(body weight is about 650 N in the experiment), then the
human is considered in a stance phase. Otherwise he/she is
regarded in a swing phase.

(27)

Note that the desired motor angle is dependent on the human joint
angle as in (27) since the appropriate spring deflection is required
to generate the desired torque. When the human joint moves fast
with a large angle change, the desired motor angle trajectory is
also large and changes fast, which makes the tracking error large.
To decrease the tracking error without chattering phenomenon,
the thickness of the boundary layer needs to be adjusted according
to human motion.
In this paper, the RSEA is installed on the knee joint and the
thickness of the boundary layer is changed according to two major
gait phases, i.e., swing and stance phases. In the swing phase, the
movement of the knee joint is large and fast, which makes the
tracking error large. On the contrary, the knee hardly moves in the
stance phase. Thus the boundary layer is set thinner in the swing
phase than in the stance phase as shown Fig. 5 to decrease the torque error without the chattering phenomenon. The actual value of
the thickness of the boundary layer in each phase, USW and UST ,
is adjusted manually since discomfort feeling caused by the chattering phenomenon and tracking error must both be considered.
Also the boundary layer is changed smoothly as shown in Fig. 5
to avoid discomfort feeling when the thickness of the boundary
layer is changed.
031010-4 / Vol. 5, SEPTEMBER 2011

Fig. 5 Varying boundary layer as gait phases

Performance Analysis by Experiments

4.1 Smoothed Sliding Mode Control. The performance of

the smoothed sliding mode control is verified by comparing with
that of typical sliding mode control. In this experiment, the performance of smoothed sliding mode control is verified comparing
with that of sliding mode control. The desired spring deflection
was set to sine wave with an amplitude of 1 rad at 1 Hz, and arbitrary motions were given to the human joint side. Figure 6 shows
the experimental results with sliding mode control with control
input in (21) and smoothed sliding mode control with the control
input in (25). The motor moves around the human joint to make
appropriate spring deflection for the generation of the desired joint
torque. The generated torque is calculated by the spring deflection
multiplied by the spring constant k [5]. As shown in Fig. 6(a),
sliding mode control with the signum function shows extremely
high performance, but the chattering phenomenon is observed as
shown in the tracking error graph in Fig. 6(a). The vibration and
noisy sound by the chattering effect can be easily felt from the
human side, which is highly undesirable for pHRI. Smoothed sliding mode control shows a slightly worse performance comparing
to pure sliding mode control as shown in Fig. 6(b), but the chattering phenomenon is not observed.
Transactions of the ASME

Downloaded From: http://medicaldevices.asmedigitalcollection.asme.org/ on 02/07/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Fig. 8 Desired knee joint torque profile in one stride

or the extension directions are applied in turn to help the knee

motions in a swing phase.
The experimental result with actual walking motion is shown
in Fig. 9. The experimental results with a constant boundary
layer are shown in Fig. 9(a). The spring is deflected appropriately by the motor angle change. The torque error is large when
the leg is in swing phases as shown in Fig. 9(a) due to the large
and fast knee motion. The large torque error can be decreased by
reducing the thickness of the boundary layer in the swing phase.
Figure 9(b) shows the experimental result with a varying boundary layer. The thickness of the boundary layer is reduced only in
the swing phases and remains the same in the stance phase. The
root-mean-square (rms) values of the torque error of each control

Fig. 6 Experimental results: the desired torque is set as a sinusoidal wave

4.2 Gait Phase-Based Smoothed Sliding Mode Control. The

experiments in Fig. 6 shows the performance of the smoothed
sliding mode controller without actual human motions. To verify
the performance of the proposed method, i.e., the smoothed sliding mode controller with a varying boundary layer on an actual
human, the experimental setup shown in Fig. 7 was utilized. One
RSEA was installed on the knee joint, and controlled by the proposed control algorithm. The subject wearing this experimental
setup walked on a treadmill at about 1.5 km/h. The desired joint
torque pattern in one stride is given in Fig. 8. Just after the heel
strike, the knee torque to the flexion direction is applied for the
subject to bend the knee easily, and the knee torques to the flexion

Fig. 7

The RSEA and Smart Shoes installed on the orthosis

Journal of Medical Devices

Fig. 9

Experimental results with actual walking motions

SEPTEMBER 2011, Vol. 5 / 031010-5

Downloaded From: http://medicaldevices.asmedigitalcollection.asme.org/ on 02/07/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

Table 2 rms values of torque error in swing and stance phases

Fixed boundary layer (N m)

Varying boundary layer (N m)

Swing

Stance

0.1283
0.0984

0.0558
0.0507

algorithm in each phase are calculated, and the values are shown
in Table 2. The rms values of the torque error in the stance phase
for the fixed boundary layer and the varying boundary layer are
almost the same since the thickness of the boundary layer is the
same, but the rms values of the torque error in the swing phase is
significantly decreased by reducing the thickness of the boundary layer.

Conclusions

In this paper the nonlinear control algorithms are proposed for

a rotary series elastic actuator (RSEA) to generate the desired
torque accurately while overcoming the nonlinear resistive torque and modeling uncertainty. To guarantee the robustness for
modeling uncertainties, sliding mode control is applied, but it
introduces chattering, which is not desirable for physical humanrobot interaction. Thus smoothed sliding mode control by utilizing a boundary layer is applied. The rotary series elastic actuator
is installed on the knee joint on an orthosis, and the thickness of
the boundary layer is varied for the swing and stance phases to
decrease the torque error without the chattering phenomenon.
The experimental results verify the advantage of the sliding
mode control with the varying boundary layer based on the gait
phases.

031010-6 / Vol. 5, SEPTEMBER 2011

Acknowledgment
This work was supported by National Science Foundation
(NSF) Grant CMMI-0800501.

References
[1] Paluska, D., and Herr, H., 2006, Series Elasticity and Actuator Power Output,
in IEEE International Conference on Robotics and Automation (ICRA), pp.
18301833.
[2] Pratt, J. Krupp, B. and Morse, C., 2002, Series Elastic Actuators for High Fidelity Force Control, Industrial Robot: An International Journal, 29, pp. 234
241.
[3] Pratt, G., and Williamson, M., 1995, Series Elastic Actuators, in IEEE International Conference on Intelligent Robots and Systems, pp. 399406.
[4] Robinson, D., Pratt, J., Paluska, D. J., and Pratt, G. A., 1999, Series Elastic Actuator Development for a Biomimetic Walking Robot, in IEEE/ASME International Conference on Advanced Intelligent Mechatronics, pp. 561568.
[5] Kong, K., Bae, J., and Tomizuka, M., (2009), Control of Rotary Series Elastic
Actuator for Ideal Force-Mode Actuation in Human-Robot Interaction
Applications, IEEE/ASME Trans. Mechatron., 14, pp. 105118.
[6] Kong, K., 2009, Mechatronic Considerations for Human Assistive and Rehabilitation Systems, PhD Thesis, University of California at Berkeley, Berkeley, CA.
[7] Slotin, J., and Li, W., 1991, Applied Nonlinear Control, Prentice-Hall, Englewood Cliffs, NJ.
[8] Young, K. D., Utkin, V. I., and Ozguner, U., 1999, A Control Engineers
Guide to Sliding Mode Control, IEEE Trans. Control Syst. Technol., 7, pp.
328342.
[9] Kong, K., and Tomizuka, M., 2009, Sensor-Embedded Shoes Based Gait Monitoring System for Rrehabilitation, IEEE/ASME Trans. Mechatron., 14, pp.
358370.
[10] Bae, J., and Tomizuka, M., 2010, Gait Phase Analysis Based on a Hidden Markov Model, in IFAC Symposium on Mechatronics Systems, pp. 746751.
[11] Bae, J., Kong, K., Byl, N., and Tomizuka, M., 2009, A Mobile Gait Monitoring System for Gait Analysis, in IEEE Proceeding on Rehabilitation Robotics
(ICORR), pp. 7379.
[12] Bae, J., Kong, K., Byl, N., and Tomizuka, M., 2011, A Mobile Gait Monitoring System for Abnormal Gait Diagnosis and Rehabilitation: A Pilot Study for
Parkinsons Disease Patients, ASME J. Biomech. Eng., 133, p. 041005.

Transactions of the ASME

Downloaded From: http://medicaldevices.asmedigitalcollection.asme.org/ on 02/07/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use