Robotics and Autonomous Systems 73 (2015) 111122

Contents lists available at ScienceDirect

Robotics and Autonomous Systems

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

A proprioceptive neuromuscular facilitation integrated robotic

anklefoot system for post stroke rehabilitation
Zhihao Zhou a , Yuan Zhou b , Ninghua Wang b , Fan Gao c , Kunlin Wei d , Qining Wang a,
a
The Robotics Research Group, College of Engineering, Peking University, Beijing 100871, China
b
Department of Rehabilitation Medicine, First Hospital, Peking University, Beijing 100034, China
c
Department of Health Care Sciences, University of Texas Southwestern Medical Center, Dallas, USA
d
Department of Psychology, Peking University, Beijing 100871, China

highlights
We developed a proprioceptive neuromuscular facilitation (PNF) integrated robotic anklefoot system for post stroke rehabilitation.
It is the first time that PNF method has been used in ankle spasticity/contracture rehabilitation.
Five able-bodied subjects participated in the experiments and five stroke patients were recruited with a six-week PNF treatment.
The proposed system can offer more effective treatment than passive stretching in improvement of both passive and active joint properties.

article info abstract

Article history: Ankle joint with spasticity and/or contracture can severely disable the mobility and the independence of
Available online 2 October 2014 stroke survivors. In this paper, we developed a proprioceptive neuromuscular facilitation (PNF) integrated
robotic anklefoot system for post stroke rehabilitation. The system consists of a robotic platform and a
Keywords:
control system with graphic user interface. We employ five normal subjects to test the reliability and
Proprioceptive neuromuscular facilitation
feasibility of the proposed system. To validate the effectiveness of the PNF integrated robotic system, we
Robotic anklefoot system
Ankle rehabilitation recruit five stroke patients and carry out a six-week PNF treatment. Treatment outcome was evaluated
Spasticity/contracture quantitatively in passive and active joint properties. The passive hysteresis loop shows that the maximum
Stroke dorsiflexion angle increases from 32.9 1.5 to 42.0 3.2 (p = 0.014) while the resistance torque
decreases from 45.6 Nm 5.8 N m to 29.8 N m 4.4 N m (p = 0.019). The active joint properties are
improved significantly with the training score increasing from 5.7 0.9 to 8.1 0.6, and getting close
to that of normal subjects (9.5 0.3). In addition, muscle strength has a rising trend as time goes on. The
results demonstrate that the proposed PNF integrated robotic anklefoot rehabilitation system is effective
in improving ankle spasticity and/or contracture and is a promising solution in clinical rehabilitation.
2014 Elsevier B.V. All rights reserved.

1. Introduction from the hypertonus and reflex hyperactivity of skeleton muscles

Human ankle joint as a very flexible and complex skeletal struc-
ture plays an important role in providing forward propulsion force
during terminal stance phase and maintaining body balance and
smooth gait during the whole gait cycle [1]. Cerebrovascular ac-
cident (CVA), or stroke, is one of the leading causes of ankle dis-
ability [2,3]. For those patients, the ankle joint with spasticity
and/or contracture can severely disable the mobility and the in-
dependence of stroke survivors [48]. The spasticity is resulted
Corresponding author. Tel.: +86 10 6276 9138; fax: +86 10 6276 9138.
E-mail address: qiningwang@pku.edu.cn (Q. Wang).
http://dx.doi.org/10.1016/j.robot.2014.09.023
0921-8890/ 2014 Elsevier B.V. All rights reserved.
[9,10]. Such spasticity in flexor muscles of stroke patients is a more
common syndrome than extensor muscles. It reduces the range of
motion (ROM) of ankle joint and may cause severe physical pain.
Moreover, lack of mobilization and prolonged spasticity may fur-
ther change the structure of muscle fibers and connective tissues
and finally lead to permanent contracture as a result. About 34%
of stroke survivors have developed ankle contracture [8,11,12].
Therefore, one of the greatest challenges in stroke survivors re-
habilitation is to improve their ankle spasticity and/or contracture,
which can seriously influence its normal functional activities.
In clinic, the ankle joint with spasticity and/or contracture
is generally rehabilitated via physiotherapy [1315]. During the
treatment, patients ankle is manually moved within its ROM by
112 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122
a physical therapist. Physical rehabilitation is in need of a long-
term continuous operation as short-term treatment is less effective
and usually insufficient to make patients fully recuperation [16].
Even if the patients have temporarily recovered from short-term
treatment, they tend to relapse and have future problems [13,17].
In addition, for some severe patients, their ankle joints have a very
high stiffness and can hardly be stretched by therapists, even with
strong arms. Above all, manual stretching is very time-consuming,
strenuous and laborious to physical therapists. Therefore, manual
rehabilitation may not last long, partly due to the limitation of
stretching frequency and duration time.
In view of shortcomings of physiotherapy, a robotic an-
klefoot rehabilitation system has been proposed to support
physicians in providing a high-intensity therapy for the stroke
patients [18]. Robotic technology can transform rehabilitation
from labor-intensive operations to robot-assisted operations,
which can implement different kinds of rehabilitation meth-
ods [19]. The robotic system can offer an adequate stretching force
and sustaining long-term training, which can cover the limita-
tion of manual stretching [20]. It can also record rich information,
such as velocity, ROM, joint torque, electromyography (EMG) sig-
nals. Those useful signals can facilitate patient diagnosis, functional
assessment, therapy customization and rehabilitation history
recording. Thus, robotic rehabilitation is gradually being thought
to be as good as or even better than manual therapy [21]. There are
mainly two kinds of robotic anklefoot rehabilitation systems [20]:
one kind is mobile systems, e.g. [2224], mainly focusing on im-
proving walking gait and the other one is platform-based systems
aiming at improvement of ankle performance. For ankle joint with
spasticity/contracture induced by the hypertonus and reflex hy-
peractivity of flexor muscles, the primary work is to alleviate the
spasticity of crus muscle [5,6]. Before walking rehabilitation using
wearable systems, patients with ankle joint spasticity/contracture,
especially those severe patients, have to use a platform-based sys-
tem to ensure reliability and improve current performance.
Recently, several research groups have developed different
robotic platform-based anklefoot rehabilitation devices [2533].
Continuous passive motion (CPM) is mainly applied in those de-
vices. It has been confirmed in their studies that passive stretch-
ing is effective in treating the ankle joint with spasticity and/or
contracture. CPM devices can provide regular and consistent pas-
sive stretching. The ankle joint is moved between two predefined
positions which usually not cover the whole ankle ROM. There-
fore, calf muscle may not be fully stretched into the extreme po-
sition of dorsiflexion where the spasticity and/or contracture is
significant. In addition, most of the ankle CPM devices run at a set
velocity and do not provide motions with velocity change during
one reciprocation. Different from those devices, Zhang et al. has
developed an intelligent stretching device for the patients with
contracture/spasticity and the stretching velocity is inversely pro-
portional to the joint resistance torque [2527]. However, during
passive stretching since lower limb is totally relaxed, the improve-
ment of muscle strength and coordination are limited and patients
can hardly get functional recovery. In addition, passive stretching is
only a kind of mechanically reciprocating motion without involv-
ing patients active participation, which makes their acceptance
and initiative not high.
To address these problems, we choose an active rehabilitation
method, namely proprioceptive neuromuscular facilitation (PNF)
technique. Different from the passive stretching, here the active
method is specified from the subject point of view, which means
the active participation of the patient. Common PNF stretching
involves a shortening contraction of the opposing muscle to place
the target muscle on stretch. It was firstly proposed by Kabat and
Knott for the rehabilitation of polio patients with paralysis [34].
Klein et al. reported that PNF treatment in elderly will significantly
improve flexibility, ROM, muscle strength and ADL function [35].
The PNF is even found effective to increase muscle volume and
alter muscle fiber types [36]. Thus, PNF is widely used for physical
therapist and athletic trainers [37]. Above all, PNF technique can
cover the problems in the previous treatment and is more effective
than passive stretching [38]. Moreover, the active participation in
PNF treatment can improve their compliance and initiative.
In this paper, we develop a PNF integrated robotic anklefoot
system for ankle joint with spasticity/contracture of post stroke
rehabilitation. The robotic system can provide the required mo-
tion of plantar flexion and dorsiflexion, and has nine degrees of
freedom (DOFs) to conveniently adjust the position of footplate for
the sake of avoiding misalignment between biological ankle axis
and robotic system axis [39]. A graphic user interface (GUI) de-
veloped in the Labview environment is customized friendly and
concisely for both patients and operators. Moreover, in consid-
eration of the safety in humanmachine interaction (HMI) [40],
protection on control system and GUI, mechanical limits and emer-
gency switches are all designed in the proposed system. Five post
stroke patients participate in our pilot experiment and accept a
course of six-week PNF rehabilitation treatment using the robotic
anklefoot system. Experimental results including changes of both
passive and active properties of ankle joint after training show the
improvement of spasticity and/or contracture.
This paper is organized as follows. Section 2 presents the details
of the robotic anklefoot rehabilitation system. PNF technique,
experiment protocol and evaluation methods are illustrated
in Section 3. Performance of the proposed robotic anklefoot
rehabilitation system and results of PNF rehabilitation for five post
stroke patients are shown in Section 4. Finally, we make discussion
in Section 5 and conclude in Section 6.
2. Robotic anklefoot system
This section presents the main technical solutions of the robotic
anklefoot system (see Fig. 1) which is functionally divided into
two parts: GUI (top layer) and hardware (bottom layer). Three
subsystems are implemented on the platform and described
hereafter. They contains the mechanical design, the sensory and
control system, and the graphical user interface. Safety of patients
for the rehabilitation devices is placed at the first place. Once the
danger happens, it would cause a destructive injury to patients.
Therefore, protection on control system and GUI, mechanical stops
and emergency switches are redundantly designed to ensure the
absolute safety of patient in our system.
2.1. Mechanical design
The proposed robotic anklefoot rehabilitation system consists
of an immobile base that contains a comfortable seat, a motor
suite (dunkermotoren Inc.), an adjustable sliding platform in three
degrees of freedom which is used to move the motor bracket to
an appropriate position, an adjustable leg support and a control
cabinet (see Fig. 2). The motor and the footplate are fixed on the
sliding platform. The leg support can be adjusted in four degrees
of freedom and the leg was strapped to the leg support by the
leg belt. The adjustable sliding platform and leg support together
ensure that ankle axis is aligned with the motor shaft while knee
is flexed at a fixed degree for each training. They are all locked
by the lock plungers after being adjusted to the desired position.
Since the rehabilitation technique that we are adopted is a kind of
active method which need the patient to perform his maximum
force during the treatment, the hardware structure must have
high mechanical strength and stiffness to make the patient and
guarantee the device to be at a relative standstill. Therefore the
 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122 113

Fig. 1. The overall structure of the proposed robotic anklefoot rehabilitation system. The top layer is the GUI and the bottom layer is the hardware.

Fig. 2. The proposed robotic anklefoot rehabilitation system. It consists of the immobile base, the sliding platform, the leg support, the control and power supply cabinet
and the sensory system including the inclinometer and the EMG acquisition system. The sliding platform has three DOFs of translation along x, y and z. The leg support has
four DOFs of translation along x, y and z and rotation along y.

mechanical structure of the device will be more difficult and
complex than CPM devices.
As shown in Fig. 3, the motor is fixed on the motor bracket
and the footplate is fixed on the motor shaft through the side
plate. The motor suite consists of a DC motor and an inline gearbox
with a 250:1 gear ratio which increases the loading capacity of the
motor up to 100 N m. It also has an inline rotation encoder for
speed-closed control. Considering interaction between human and
machine, the safety of the robotic platform is quite important. In
this study, rotation limits of the footplate are set both in the motor
drive and control module. The system will stop running if the
obliquity of the footplate is out of the prescribed range. In addition,
a mechanical limit stop is set to constrain the range of motion.
The motor bracket with location holes on the perimeter is used
to place the mechanical limit stops in our system. The separation
angle of two adjacent location holes is 5. Two stop pins on the
motor bracket are used for limiting the rotation range of the swing
pin, which is fixed on the motor shaft. Besides, the operator and the
patient all have their own handhold emergency switch and either
of them could shut down the motor by pressing their own switch.
The footplate can move along the side plate to ensure that
the rotating axis of footplate is aligned with biological ankle axis
(see Fig. 3). The two DOFs can satisfy the foot size for different
patient. In experimental protocol, we firstly measure the height
of foot (distance from the bottom of foot to the lateral malleolus)
which is represented by D0 and the length of foot (distance from
the heel of foot to the lateral malleolus) which is represented by D1 .
Before the ankle is placed on the footplate, we adjust the footplate
position depending on D0 and D1 to avoid the misalignment in the
process of rotation. Then the foot is secured on the footplate by
velcro at the dorsal foot and the heel.
2.2. Control system
To obtain as much useful information as possible, placement
positions of sensors used in the system are carefully determined
(see Figs. 1 and 2). One uni-axial torque sensor (Transducer
Techniques, Inc.) is mounted on the shaft to measure the resistance
torque. An inclinometer is attached to the underneath of footplate
which can record the joint angle with the reference to the ground.
It is self-designed and its output range is 90. Two-channels
EMG system (Delsys Inc.) is used to measure EMG signals of
gastrocnemius muscle and soleus muscle. In addition, an inline
114 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122
Fig. 3. Schematic diagram of the motor suite and the footplate. The footplate has
two DOFs of translation along x and z which is adjusted according to D0 and D1 .
rotation encoder can calculate rotational speed of footplate. All
of those signals are fed into the Labview program through a USB
data acquisition (DAQ) card (NI Inc.) with the sampling rate at
1 kHz. Thereinto, the resistance torque and the joint angle are
simply processed through a second order Butterworth low-pass
filter with a cut-off frequency of 20 Hz. The raw EMG signals are
firstly bandpass filtered through a fourth order Butterworth filter
with a cutoff frequency of 20 and 450 Hz, then full-wave rectified,
and low-pass filtered through another fourth order Butterworth
filter with a cutoff frequency of 5 Hz. Then it is filtered by moving
average with a window length of two hundred sample points. The
processed EMG signals also need to subtracts the measured offset
when the muscles are relaxed. The relationship between torque
and EMG signals of gastrocnemius and soleus is shown in Fig. 4. The
correlation coefficient between torque and EMG of gastrocnemius
is 0.4392 while the correlation coefficient between torque and
EMG of soleus is 0.9576. The gastrocnemius muscle runs from its
two heads above the knee to the heel, so we choose soleus muscle
as feedback in our treatment also due to its high correlation. It can
demonstrate that the filter method can make soleus EMG signal
being accurately in correspondence with joint torque.
The proposed robotic anklefoot rehabilitation system is
modeled as a double pendulum using NewtonEuler formulation.
The combined dynamics of robotic system and human subject is
given by
Js (t ) + Bs (t ) + Ks (t ) = Tm (t ) + Tgrav (t ) + Tank (t )
(1)
where Js , Bs and Ks are the inertia, damping and stiffness of the
system which are estimated according to the motor specification
and the mechanical model in the Solidworks Software. (t ) was the
joint flexion angle. Tm (t ), Tgrav (t ) and Tank (t ) are the motor output
torque, gravitational torque and ankle joint torque, respectively.
Fig. 4. Relationship between torque and EMG signals of gastrocnemius (Gas) and soleus (Sol). All of them were filtered.
The joint torque when the muscles are relaxed can be obtained
by
Tank (t ) = Tres (t ) = Ja (t ) + Ba (t ) + Ka (t ) (2)
where Ja , Ba and Ka are joint stiffness, viscous damping and foot
inertia of ankle. And the joint torque when the muscles are
contracted can be obtained by
Tank (t ) = Tmuscle (t ). (3)
The system is driven by a DC motor controlled by a customized
driver. As shown in Fig. 5, the control strategy implements a double
closed-loop system with the current control as the inner-loop and
the speed control or position control as the outer-loop which is
switched based on motor status.
The velocity is controlled in such a way that it was inversely
proportional to the resistance torque [27]. Thus, when the position
of the ankle joint gets closed to the extreme position of ROM, the
increasing resistance slows down the motor gradually and at the
same time the muscletendons involved are stretched slowly and
safely. In the middle of ROM where the resistance is usually low,
the motor stretches the slack muscles at higher speed. Certainly, if
the resistance is high in the middle, the movement would also be
slowed down accordingly. Specifically, d and p are the extreme
position of the joint ROM in ankle dorsiflexion and plantar flexion,
respectively (both are positive numbers). Tp is a specified peak
resistance torque when the motor reaches the mechanical stops.
If Tres (t ) is beyond the Tp , the system will shut down immediately
as a kind of self-protection. It can serve as a safe value to protect the
patient and the system. vmin and vmax (two positive numbers) are
the minimum and maximum speeds (for stretching in the middle
of ROM) which are predetermined, respectively. C1 and C2 are
two constants and are empirically selected according to the range
of velocity and resistance torque. C2 scales the 1/Tres (t ) to the
appropriate stretching velocity.
When the position (t ) is in the middle of ROM, the motor
velocity v(t ) is calculated by
vmin vt 6 vmin
C2
v(t ) = C1 + vmin < vt < vmax (4)
|T (t )|
vmax vt > vmax
where vmin and vmax (two positive numbers) are the minimum
and maximum speeds (for stretching in the middle of ROM) which
are predetermined, respectively. C1 and C2 are two constants
and are empirically selected according to the range of velocity
and resistance torque. C2 scales the 1/Tres (t ) to the appropriate
 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122
Fig. 5. Control flow diagram of the ankle rehabilitation system.
Fig. 6. The main interface (or homepage) of the customized graphical user interface developed in the Labview environment.
stretching velocity. In our system, C1 and C2 are 7.5 and 0.05
respectively.
When the position (t ) reaches the end of ROM, namely d or
p , the motor velocity v(t ) is zero. The motor driver implements
a position close-loop by PID controller and current closed-loop
serves as the inner-loop.
2.3. Graphical user interface
A customized GUI is developed in Labview environment for
the robotic anklefoot rehabilitation system. All interfaces are
concise and friendly so that the patients can easily understand.
They are also convenient to operate by the therapist. As shown
in Fig. 6, it is the main interface (or called homepage) of our
GUI, where we can operate patients database and enter into each
training interface according to the rehabilitation protocol. Some
individual information like name, age and height needs to input
when adding a new patient. Also we can choose a previous patient
from patients database and set wanted paths of saving data.
According to the rehabilitation protocol in the homepage, we can
enter into corresponding sub interfaces in sequence.
PNF stretching as the key interface during treatment is shown
in Fig. 7. Two windows display to the operator and the patient,
like Fig. 7(a) and (b). ROM of dorsiflexion and maximum voluntary
contraction (MVC) of EMG have been recorded in the step of ROM
Measurement and MVC Measurement, which are the default
value in PNF stretching. Total trials number, hold time, relax time
and target need to set by the operator. Those parameters all can
be updated when the software is running. The patient is asked to
115
perform isometric contraction with the soleus muscle activated
and maintains the soleus EMG in the target percentage range of
MVC for seconds (Hold time). A countdown clock is used to display
the remaining time of relax time and hold time. And each trial
will have a training score to the patient which will be explained
in Section 3.
3. Method
3.1. Proprioceptive neuromuscular facilitation (PNF)
PNF stretching is commonly used in clinical environments to
enhance both active and passive ROM with the ultimate goal
being to optimize motor performance and rehabilitation. Propri-
oception means sense of self. In human limbs, the propriocep-
tor provides information about joint angles, muscle length, and
muscle tension, which give information about the position of the
limb in space. The Golgi tendon organs (GTO) serves as one kind of
proprioceptive sensory receptor organs in our body. It can provide
information about changes in muscle tension. One end of GTO is
connected to the muscle fibers and the other end merges into the
tendon bundles. When the central nervous system sends a message
to the agonist muscle to contract, (here the agonist muscles are
gastrocnemius and soleus muscle), these target muscles develop
active force. Due to the applied force, GTO gets compressed, and
triggers Golgi tendon reflex (GT reflex), which can relax and
lengthen the target muscle (here the target muscles are also gas-
trocnemius and soleus muscle). So the patient actively contracts
his gastrocnemius and soleus muscle, meanwhile makes these
116 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122

(a) Interface for operators.

(b) Interface for patients1. (c) Interface for patients2.

Fig. 7. Sub interface: PNF stretching. (a) is the interface for operator. (b) is the interface for patients when patients relax muscles in neural position and (c) is the one when
patients contract muscles to step the footplate.

(a) Direct hold-relax PNF.

(b) Indirect hold-relax PNF.

Fig. 8. Principles of two kind of PNF methods. Direct hold-relax PNF and indirect hold-relax PNF are shown in subfigure (a) and (b), respectively.

muscles get further relaxed. The repetition of this process facili-
tates the patient to further contract and relax his ankle joint. Since
the target muscle is also the agonist, this technique is usually called
direct hold-relax PNF [41,42] (see Fig. 8(a)).
On the other hand, there is an indirect hold-relax PNF tech-
nique (see Fig. 8(b)). In human body, there exists a neural phe-
nomenon called reciprocal inhibition. When the agonist muscle
(muscle causing movement) starts to contract, the tension in the
antagonist muscle (muscle opposing movement) is inhibited by
impulses from motor neurons, and thus must simultaneously relax.
With reciprocal inhibition as one muscle contracts, the opposing
(antagonist) muscle will relax and allow more movement around
the joint. In this case the target muscle is the opposing muscle (an-
tagonist) and this technique is called indirect hold-reflex PNF.
In this study, we currently only adopt the direct hold-relax PNF
technique in our system. We can intuitively discover that direct
hold-relax PNF is more easy to take in than indirect hold-relax PNF
because it can directly improve the lost or deficient dorsiflexion
motion. But considering hypertonus of back muscle due to spastic-
ity and/or contracture, the patient is very hard to actively dorsiflex-
ion according to indirect hold-reflex PNF. Therefore indirect PNF is
more suitable for their primary rehabilitation. When spasticity has
been largely alleviated, indirect hold-relax PNF maybe be needed.
3.2. Subjects and experiment protocol
Five chronic stroke patients with ankle spasticity and/or
contracture (Ashworth scale >1 around ankle joint) participated
in the study. Detail information was shown in Table 1. The stroke
subjects age was 65.6 9.0 years, height was 170.4 11.0 cm
and weight was 71.0 12.6 kg. Two of them were left impaired
side and others were right. Their first stroke occurred 42.8 13.3
months ago. All the stroke patients were able to walk in an
abnormal gait without any mechanical aid, and able to generate
plantar flexion using the calf muscles. They were all from the
Department of Rehabilitation Medicine, First Hospital, Peking
 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122
Table 1
Subject data.
Patient no. Age (year) Gender (M/F) Height (cm)
1 77 M 156
2 65 M 170
3 72 M 165
4 58 M 185
5 56 M 176
Mean Std 65.6 9.0 / 170.4 11.0
University and they gave written and informed consent before
the experiment. All procedures were approved by the institutional
review board of the First Hospital, Peking University. The training
was carried out 3 times a week and lasted for 6 weeks. In addition,
we also recruited five normal subjects to check the system before
patients experiments. Their age was 27.2 1.8 years, height was
165.6 8.7 cm and weight was 65.8 9.9 kg, and two of them
were female. Test data of both sides were measured. The purpose
of those normal subjects was to evaluation the feasibility of the
overall system.
The experiment protocol mainly involved three steps, namely
preparation, initialization and stretching:
3.2.1. Step I: preparation
D0 and D1 of the subjects foot were measured in order to prop-
erly adjust the position of the device. It could ensure that the an-
kle joint was able to align with the axis of the motor shaft. The
subject seated comfortably with knee flexed at 30 which was de-
termined after multiple comparison. The lower leg was strapped
to the leg support and the foot was attached to the footplate. The
skin was cleaned and conditioned with warm water before attach-
ing the electrode pads. Surface EMG electrodes were placed ac-
cording to the recommendation of the SENIAM Protocol (Surface
ElectroMyoGraphy for the Non-Invasive Assessment of Muscles) to
detect soleus muscle EMG [43].
3.2.2. Step II: initialization
At the beginning of each stretching, the extreme position in an-
kle dorsiflexion was measured which was the maximum angle the
patient was able to reach. At first, subjects foot was moved pas-
sively to its dorsiflexion. When the extreme position was reached
according to real-time feedback of patient, the operator would shut
down the motor and the system would record the maximum angle
value. With the ankle at the extreme dorsiflexion position, subject
was asked to perform maximum voluntary contraction (MVC) in
plantar flexion direction by activating the soleus muscle and the
EMG signal was collected at the same time. MVC value of EMG gen-
erally needed to be measured up to three times. The peak value of
MVC was recorded for normalization in the PNF stretching.
3.2.3. Step III: stretching
During the PNF stretching, the ankle joint was passively rotated
from its neutral position to the extreme dorsiflexion position. Then,
the subject was asked to perform isometric contraction with the
soleus muscle activated and maintained the soleus EMG in the
range of 50% 10% MVC for 15 s. The target range and the hold
duration time were adjusted based on actual condition of different
subjects and the rehabilitation phase. The processed EMG feedback
and the target range were provided through a customized GUI in
the patient monitor. After the 15 s muscle activation, ankle joint
was moved back to its neutral position to relax the muscle. The
break between each PNF stretching was 10 s. Each training session
was 30 min including about 30 trails. Before and after each training,
ankle joint was passively rotated between its neutral position and
extreme dorsiflexion position for about 1 min to warm up and relax
the soleus and gastrocnemius muscle, respectively.
117
Weight (kg) Impaired side (L/R) Injury duration (Month)
51 L 50
85 L 59
69 R 26
76 R 46
74 R 33
71.0 12.6 / 42.8 13.3
3.3. Evaluation
The robotic anklefoot system can be available not only to treat
the ankle joint with spasticity and/or contracture but also to quan-
titatively evaluate the effectiveness based on the rehabilitation-
induced changes in joint biomechanical properties. Improvement
of passive and active properties of ankle joint will be described
hereafter.
3.3.1. Passive properties of ankle joint
Since spasticity and/or contracture can reduce the joint ROM
(mainly in dorsiflexion direction) and largely increase dorsiflex-
ion resistance torque (plantar flexor muscle resistance torque),
changes of them are effectively physical size improvements.
Throughout the stretching treatment, the robotic system measures
the passive ROM and the resistance torque. The ROM measures
were divided into dorsiflexion ROM, plantar flexion ROM and to-
tal ROM. Here we only pay more attention to the dorsiflexion ROM
since spasticity/contracture can cause drop-foot and we do not
want them to do some plantar flexion motion within its ROM.
Passive joint dynamic properties in terms of joint stiffness and
viscosity before/after PNF rehabilitation are evaluated quantita-
tively. They are measured as the slope of the torqueangle rela-
tionship (hysteresis loop) and the slope of torquevelocity rela-
tionship [4,25]. In hysteresis loop, the ascending curve represents
the passive dorsiflexion phase and the descend curve represents
the plantar flexion phase, as show in Fig. 9. and repre-
sent the torqueangle relationship before and after rehabilitation,
respectively. The changes of them can reflect improvement of pas-
sive biomechanical properties.
Before the whole rehabilitation treatment, we measure and
record the maximum dorsiflexion angle m which the patient can
be ability to reach passively. Meanwhile, corresponding resistance
torque m at the position m is also measured. We did not directly
compare (m , m ) and (m , m ) between before and after the PNF
rehabilitation treatment and it may be meaningless, because
obviously the larger joint angle, and the larger resistance torque.
Therefore, maximum dorsiflexion angle and m under the same
resistance torque m before and after rehabilitation are compared
and so are maximum resistance torque m and at the same angle
position m , as shown in Fig. 9.
3.3.2. Active properties of ankle joint
For stroke patients, because of brain dysfunctions related to dis-
ease of the blood vessels supplying the brain, it leads to loss of
muscle inhibition and cause spasticity. PNF can improve patients
control for their back muscles through reconstruction of nerve
function because the treatment needs patients to perform isomet-
ric contraction with the soleus muscle activated and to track tar-
get curve. In addition, PNF as a active rehabilitation therapy can
strengthen the muscle force indicated by the active output torque
of ankle joint. Those advantages are what the passive stretching
does not have.
We propose a training Score to quantitatively evaluate the
outcome of PNF stretching. It can reflect active properties of ankle
joint comprehensively. We assume that P is the feedback value
118 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122

Fig. 9. Torqueangle relationship (hysteresis loop) during ankle stretching in dorsi- and plantar flexion, obtained from subject 3. The x-axis is the dorsiflexion angle and the
Y -axis is the resistance torque. (, ) and ( , ) represent the torqueangle relationship before and after rehabilitation, respectively. The variable with subscript m refers
to the maximum value.

Table 2
Comparison of maximum dorsiflexion angle and m under the same resistance torque m and maximum resistance torque m and at the same position m between before
and after PNF rehabilitation.
Patient no. m (deg) m (N m) (N m) m (deg) m (N m) (deg)
1 43.8 45.4 20.7 51.5 35.6 37.4
2 40.1 61.3 36.5 44.0 42.9 30.4
3 37.0 55.6 43.2 39.0 46.7 32.5
4 31.5 31.7 28.1 32.0 27.5 29.3
5 35.9 33.8 21.4 43.7 34.4 35.1
Mean SE 37.6 2.1 45.6 5.8 29.8 4.4 42.0 3.2 37.4 3.4 32.9 1.5

for patient during PNF training, which refers to the ratio between
current measured EMG value and MVC or the ratio between current
torque and maximum active joint torque. Pt refers to the target
ratio and the target range is Pt Er in our system which are showed
in GUI (Fig. 7). Patients are asked to perform isometric contraction
to reach Pt and maintain P within the target range during the
predefined time T . Then we calculate the average value across T
and the training Score of each trial is defined as
Score = mean(sgn(P Pt )) (5)
where sgn(x) is defined as

| x| < E d

1
Er |x|

sgn(x) = Ed 6 |x| 6 Er (6)
Er Ed Fig. 10. Function sgn(x). Ed (>0) is the value of deadband around 0 and Er (>0) is
| x| > E r

0 the value of target range around 0.

where Ed (>0) is the value of deadband around 0 and Er (>0) is

the value of target range around 0 (Fig. 10). They are empirically
determined to be 0.02(2%) and 0.1(10%), respectively. Finally we
multiply Score by 10.0 to map it from 01 to 010. In our experi-
ment, default value of T and Pt are 15 s and 50%, respectively. For
instance, if real-time P is 56% or 44%, sgn (P Pt ) equals to 0.5.
To avoid some disturbance, data including two seconds of start-
ing and ending respectively during each stretching are discarded
and not counted for the score. Each treatment is 30 min containing
about 2030 trials PNF stretching. Final score for one treatment is
calculated through averaging all trials score in the day.
4. Experimental results
4.1. Changes in passive properties of ankle joint
The increase in ankle ROM and the decrease in resistance
torque were consistently observed, indicating that ankle joint was
more compliant than before and spasticity/contracture had been
alleviated effectively. Comparison of the maximum dorsiflexion
angle and m under the same resistance torque m before and
after the treatment were shown in 6th and 4th columns of Table 2.
And we also conducted a comparison of the maximum resistance
torque m and in the same position m before and after the
PNF rehabilitation treatment, as shown in 2nd and 3rd columns of
Table 2. Over multiple patients, the dorsiflexion angle increased
from 32.9 1.5 to 42.0 3.2 while the resistance torque
decreased from 45.6 N m 5.8 N m to 29.8 N m 4.4 N m.
Furthermore, the showed significantly lower than m (p = 0.019)
while the m showed significantly larger than (p = 0.014).
Those can demonstrate that PNF is effective for treatment for ankle
spasticity and contracture.
The rehabilitation treatment of our system can result in consid-
erable changes in joint properties characterized by joint stiffness
and viscosity. They are calculated from the relationship between
 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122 119

Fig. 11. Changes in torqueangle relationship (1st row), stiffnessangle relationship (2nd row), and viscosityangle relationship (3rd row) of spastic ankles caused by
the six-week PNF treatment in two stroke patients. The blue and red line correspond evaluations done immediately before and after treatment, respectively. Each column
corresponds to a stroke patient. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 12. Changes of active torque of ankle joint during six weeks. The x-axis is the trial number (about 3 times per week) and the y-axis is the increment of active joint
torque relative to the maximum torque before treatment.

the joint angle and resistance torque (hysteresis loop). As shown
in the representative cases (Fig. 11), such as the subject 1 (the 1st
column), joint stiffness is reduced remarkably after rehabilitation
across the range of muscle contraction (2nd row of Fig. 11). At the
same time, the treatment also reduces the joint viscosity (3rd row
of Fig. 11). Other subjects except subject 4 are similar to the sub-
ject 1, having significant improvement. For subject 4 (the 2nd col-
umn), we do not see increase in maximum dorsiflexion angle and
just find decrease in resistance torque. Also, stiffness and viscos-
ity also do not decrease evidently. Since he suffered from one time
cerebrovascular disease in the third week, this maybe result that
the effect of rehabilitation was not obvious for him.
4.2. Change of active properties of ankle joint
As shown in Fig. 12, we can see that the maximum active
torque of every subject was showing a rising trend as time goes
on. For a representative case, such as subject 5, his joint torque has
increased 24.24 N m after treatment which was a large promotion.
Other subjects are also similar. On the whole, increment of joint
torque is 15.3 N m 5.9 N m. As shown in Fig. 13, the Score
of before/after six weeks rehabilitation and normal subjects were
5.7 0.9, 8.1 0.6 and 9.5 0.3 respectively. Over multiple
patients, the increase in training Score is observed, and it is close to
that of the healthy subjects. Thereinto, the normal value are from
five normal subjects about 20 trials per subject and we can find that
120 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122
Fig. 13. Comparison of training Score of patients before and after rehabilitation and
normal subjects. The Scores are 5.7 0.9, 8.1 0.6 and 9.5 0.3 respectively.
the task is relatively easy for all normal subjects. The result present
that the patients were becoming faster to track the target line after
the PNF rehabilitation.
5. Discussion
We proposed a PNF integrated robotic anklefoot system for
rehabilitation of ankle joint with spasticity/contracture in post
stroke patients. Different from existing treatments, PNF is a kind
of method based on proprioception with the ultimate goal being to
optimize motor performance and rehabilitation. The proposed sys-
tem intelligently stretches the ankle forcefully and safely to its ex-
treme positions where spasticity and contracture are significant. It
is reliable and convenient for patients and therapists. Furthermore,
rehabilitation effectiveness is verified in two aspects including
active and passive joint properties. Our system is considerable im-
provement over previous study and more importantly, it is promis-
ing for treatment of ankle joint with spasticity/contracture in
clinical rehabilitation.
The proposed system has following advantages over previous
robotic devices for post stroke rehabilitation. First, the PNF is very
effectiveness for post stroke rehabilitation over passive stretch-
ing. In PNF treatment, the ankle is stretched to its extreme po-
sitions where spasticity and contracture are significant and then
the patient performs isometric contraction with the soleus mus-
cle according to biological feedback. For spasticity/contracture re-
habilitation, the barrier to further progress lies not in developing
new hardware but rather in finding the most effective way to en-
hance neuro-recovery [44]. It is clear that PNF can be qualified to
carry the neurologic recovery. Active joint properties are improved
significantly with the training score increasing from 5.7 0.9 to
8.1 0.6, and getting close to that of normal subjects (9.5 0.3)
after treatment (Fig. 13). The results indicate that neural control is
improved by the PNF training. In addition, muscle strength has a
rising trend as time goes on (Fig. 12) and increment of joint torque
is 15.3 N m 5.9 N m. Those are what the passive stretching does
not have.
To verify the effectiveness of our system with passive stretch-
ing, we can compare some common indicators. In our results, the
passive hysteresis loop shows that the maximum dorsiflexion an-
gle increases from 32.9 1.5 to 42.0 3.2 (p = 0.014)
while the resistance torque decreases from 45.6 N m 5.8 N m
to 29.8 N m 4.4 N m (p = 0.019) (Table 2). All these changes
are quite evident. In addition, changes in joint stiffness and viscos-
ity are also significantly improved (Fig. 11). However, in [25], dor-
siflexion range increased from 11.9 to 16.5 at the same torque
(10 N m) after stretching. Changes in joint stiffness and viscosity
are not significant. In [27], the resistance torque was 11.9 N m and
10.5 N m and the stiffness was 0.60 N m/ and 0.49 N m/ at 20
dorsiflexion, pre- and post-intervention, respectively. In [29], the
dorsiflexion angle increases of 5. Above all, we can find that the
proposed PNF method is more effective than the passive stretching.
Now, there are several commercial products called isokinetic ma-
chines (e.g. BIODEX, CYBEX NORM) whose working modes include
continuous passive motion (CPM) and resistance training. Differ-
ent from them, the working mode of our system is originated from
PNF technique as a kind of neurophysiological technique most of-
ten used to inhibit spasticity based on two inhibitory concepts
Autogenic inhibition and Reciprocal inhibition which is respond to
direct hold-relax and indirect hold-relax respectively. The prelimi-
nary clinical results demonstrate its effectiveness on rehabilitation
of spasticity and/or contracture.
Second, the proposed system can raise compliance and ini-
tiative of patients which are key points to make them training
zealously and persistently [45,46]. Stroke rehabilitation is in need
of a long-term continuous operation as short-term treatment is
less effective and usually insufficient to make patients fully recu-
peration [13,16,17]. In our system, the friendly GUI can provide
biological feedback and the patient needs to perform isometric
contraction with the soleus muscle according to the biomedical
feedback, which have formed a reasonable humanmachine inter-
action. From behaviors and experiences of all subject, it is found
that they show more interested in training than passive stretching.
Thus, our system can make the patients persevere in a long-term
treatment with better initiative.
Moreover, our device stretches the ankle joint forcefully and
safely. The stretching velocity is controlled similar to Zhang
et al. [25] and the control method can enhance safety. The system
can cover the limitation of stretching frequency and force in phys-
ical therapist and do not have the shortcoming of labor, so it can
make long-term and sufficient treatment [20]. Safety of HMI [40]
is ensured by redundantly design of mechanical limit, emergency
switch and control strategy. In addition, multi-DOFs of platform
can be adjusted to avoid misalignment between biological ankle
axis and robotic device rotation axis [39].
The robotic anklefoot rehabilitation system, though being
promising, has some limitations in this paper and future improve-
ments are needed. Firstly, the small sample of patients takes part in
the pilot experiments and we will verify more subjects to achieve
more comprehensive and reliable evaluation of our system and
the rehabilitation therapy comparing with different training meth-
ods. Secondly, only a six-week rehabilitation was done for those
patients and we will continue the long-term treatment for those
patients and to further investigate the effectiveness of PNF treat-
ment. Thirdly, we will try more effective methods integrated with
robotics for stroke rehabilitation in future, such as other neuro-
physiologic techniques or motor learning techniques [47].
6. Conclusion
In this paper, we have proposed a PNF integrated robotic
anklefoot system for rehabilitation of ankle joint with spastic-
ity/contracture induced by stroke. The proposed system is in-
telligent and reliable, and can offer more effective treatment
than passive stretching in improvement of both passive and ac-
tive joint properties. We verify its effectiveness through clini-
cal experiments. The system can be a good solution for ankle
spasticity/contracture. Future endeavors should further improve
its performance and make it widely used in clinical rehabilitation.
Acknowledgments
This work was supported by the National Natural Science
Foundation of China (61005082, 31000456, 31371020, J1103602,
61020106005), the Beijing Nova Program (No. Z141101001814001)
and the PKU-Biomedical Engineering Joint Seed Grant 2014.
 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122 121

References [34] H. Kabat, M. Knott, Proprioceptive facilitation technics for treatment of

paralysis, Phys. Ther. Rev. 33 (1953) 5364.
[1] N. Tejima, Rehabilitation robotics: a review, Adv. Robot. 14 (2000) 551564. [35] D.A. Klein, W.J. Stone, W.T. Phillips, J. Gangi, S. Hartman, PNF training and
[2] World Health Organization (2008). The World Health Report. [Online]. physical function in assisted-living older adults, J. Aging Phys. Act. 10 (2002)
Available: http://www.who.int/whr/2008/en. 476488.
[36] N. Kofotolis, I.S. Vrabas, E. Vamvakoudis, A. Papanikolaou, K. Mandroukas,
[3] T. Ingall, Stroke-incidence, mortality, morbidity and risk, J. Insur. Med. 36
Proprioceptive neuromuscular facilitation training induced alterations in
(2004) 143152.
muscle fibre type and cross sectional area, Br. J. Sports Med. 39 (3) (2005) e11.
[4] S.G. Chung, E.M.T. van Rey, Z. Bai, E.J. Roth, L.Q. Zhang, Biomechanic chans in
[37] S.M. Marek, J.T. Cramer, A.L. Fincher, L.L. Massey, S.M. Dangelmaier,
passive properties of hemiplegic ankles with spastic hypertonia, Arch. Phys.
S. Purkayastha, K.A. Fitz, J.Y. Culbertson, Acute effects of static and
Med. Rehabil. 85 (2004) 16381646.
proprioceptive neuromuscular facilitation stretching on muscle strength and
[5] J. Harlaar, J.G. Becher, C.J. Snijders, G.J. Lankhorst, Passive stiffness characteris-
power output, J. Athl. Train. 40 (2005) 94103.
tics of ankle plantar flexors in hemiplegia, Clin. Biomech. 15 (2000) 261270.
[38] D.C. Funk, A.M. Swank, B.M. Mikla, T.A. Fagan, B.K. Farr, Impact of prior exercise
[6] R.T. Katz, W.Z. Rymer, Spastic hypertonia: mechanisms and measurement,
on hamstring flexibility: a comparison of proprioceptive neuromuscular
Arch. Phys. Med. Rehabil. 70 (1989) 144155.
facilitation and static stretching, J. Strength Cond. Res. 17 (2003) 489492.
[7] F. Gao, L.Q. Zhang, In vivo biomechanical evaluations of the medial
[39] A. Schiele, F.C. van der Helm, Kinematic design to improve ergonomics in
gastrocnemius: changes in muscle properties in stroke survivors, in:
human machine interaction, IEEE Trans. Neural Syst. Rehabil. Eng. 14 (4)
Proceedings of the 28th IEEE EMBS Annual International Conference, 2006, pp.
(2006) 456469.
20832086.
[40] A.D. Santis, B. Siciliano, A.D. Luca, A. Bicchi, An atlas of physical humanrobot
[8] W. Vattanasilp, L. Ada, J. Crosbie, Contribution of thixotropy, spasticity, and
interaction, Mech. Mach. Theory 43 (2008) 253270.
contracture to ankle stiffness after stroke, J. Neurol. Neurosurg. Psychiatry 69
[41] W.E. Prentice, M.I. Voight, Techniques in Musculoskeletal Rehabilitation,
(2000) 3439.
McGraw Hill, New York, NY, 2001.
[9] J.W. Lance, Symposium synopsis, in: R.G. Feldman, R.R. Young, W.P. Koella
[42] P.R. Surburg, J.W. Schrader, Proprioceptive neuromuscular facilitation tech-
(Eds.), Spasticity: Disordered Motor Control, Year Book Medical Publishers,
niques in sports medicine: a reassessment, J. Athl. Train. 32 (1) (1997) 3439.
Chicago, 1980, pp. 485495. [43] Surface ElectroMyoGraphy for the Non-Invasive Assessment of Muscles
[10] B.D. Schmit, Mechanical measures of spasticity in stroke, Top. Stroke Rehabil. [http://www.seniam.org].
8 (2001) 1326. [44] H.I. Krebs, Rehabilitation robotics: an academic engineer perspectiv, in:
[11] M.M. Pfeffer, M.J. Reding, Stroke rehabilitation, in: R.B. Lazar (Ed.), Principles Proceedings of the 33rd Annual International Conference of the IEEE EMBS,
of Neurologic Rehabilitatio, McGraw-Hill, New York, 1998, pp. 105119. 2011, pp. 67096712.
[12] M.J. Reding, F. McDowell, Stroke rehabilitation, Neurol. Clin. 5 (1987) 601630. [45] J. Nithianantharajah, A.J. Hannan, Enriched environments, experience-
[13] J. Dettori, C. Basmania, Early ankle mobilization, part II: a one-year followup of dependent plasticity and disorders of the nervous system, Nat. Rev. Neurosci.
acute, lateral ankle sprains (a randomized clinical trial), Mil. Med. 159 (1994) 7 (2006) 697709.
1520. [46] S.A. Eraker, J.P. Kirscht, M.H. Becker, Understanding and improving patient
[14] J. Carr, R. Shepherd, Neurological Rehabilitation: Optimizing Motor Perfor-
compliance, Ann. Intern. Med. 100 (2) (1984) 258268.
mance, Butterworth-Heinemann, Oxford, 1998.
[15] M.L. Dombovy, B.A. Sandok, J.R. Basford, Rehabilitation for stroke: a review, [47] J.M. Belda-Lois, S.M. Horno, I. Bermejo-Bosch, J.C. Moreno, J.L. Pons, D. Farina,
Stroke 17 (1986) 363369. M. Iosa, M. Molinari, F. Tamburella, A. Ramos, A. Caria, T. Solis-Escalante, C.
Brunner, M. Rea, Rehabilitation of gait after stroke: a review towards a top-
[16] P.K. Jamwal, Design analysis and control of wearable ankle rehabilitation
down approach, J. Neuroeng. Rehabil. 8 (2011).
robot, (PhD thesis), The University of Auckland, Mechanical Engineering
Department, 2011.
[17] S. Trevino, P. Davis, P. Hecht, Management of acute and chronic lateral Zhihao Zhou received his Bachelor degree in Mechani-
ligament injuries of the ankle, Orthop. Clin. North Am. 25 (1994) 116. cal Engineering from China University of Geosciences (Bei-
[18] G. Kwakkel, R.C. Wagenaar, J.W. Twisk, G.J. Lankhorst, J.C. Koetsier, Intensity of jing). He is currently working towards the Ph.D. degree
leg and arm training after primary middle-cerebralartery stroke: a randomized under the supervision of Prof. Qining Wang in the Col-
trial, Lancet 354 (1999) 191196. lege of Engineering, Peking University, China. His research
[19] H.I. Krebs, J.J. Palazzolo, L. Dipietro, M. Ferraro, J. Krol, K. Rannekleiv, B.T. interests include biomedical engineering and robotics for
Volpe, N. Hogan, Rehabilitation robotics: performance-based progressive healthcare.
robotassisted therapy, Auton. Robots 15 (2003) 720.
[20] M. Zhang, T.C. Davies, S. Xie, Effectiveness of robot-assisted therapy on ankle
rehabilitationa systematic review, J. Neuroeng. Rehabil. 10 (2013).
[21] H.I. Krebs, N. Hogan, M.L. Aisen, B.T. Volpe, Robot-aided neurorehabilitation,
IEEE Trans. Rehabil. Eng. 6 (1998) 7587.
[22] S. Mohammed, Y. Amirat, H. Rifai, Lower-limb movement assistance through Yuan Zhou received her Bachelor degree from Peking Uni-
wearable robots: state of the art and challenges, Adv. Robot. 26 (2012) 122. versity, China and the Ph.D. degree from the Department of
[23] K.A. Shorter, J. Xia, E.T. Hsiao-Wecksler, W.K. Durfee, G.F. Kogler, Technologies Rehabilitation Medicine, First Hospital, Peking University,
for powered ankle-foot orthotic systems: possibilities and challenges, China. Her research interests include stroke rehabilitation
IEEE/ASME Trans. Mechatronics 18 (1) (2013) 337347. and robotics.
[24] A. Roy, H.I. Krebs, D.J. Williams, C.T. Bever, L.W. Forrester, R.M. Macko,
N. Hogan, Robot-aided neurorehabilitation: a novel robot for ankle rehabili-
tation, IEEE Trans. Robot. 25 (3) (2009) 569582.
[25] L.Q. Zhang, S.G. Chung, Z. Bai, D. Xu, E.M.T. van Rey, M.W. Rogers, M.E. Johnson,
E.J. Roth, Intelligent stretching of ankle joints with contracture spasticity, IEEE
Trans. Neural Syst. Rehabil. Eng. 10 (2002) 149157.
[26] F. Gao, L.Q. Zhang, Altered contractile properties of the gastrocnemius muscle
poststroke, J. Appl. Physiol. 105 (2008) 18021808. Ninghua Wang received her Bachelor degree from Bei-
[27] F. Gao, Y.P. Ren, E.J. Roth, R. Harvey, L.Q. Zhang, Effects of repeated ankle jing Medical University, China and Ph.D. degree from The
stretching on calf muscletendon and ankle biomechanical properties in Hong Kong Polytechnic University, Hong Kong. She is cur-
stroke survivors, Clin. Biomech. 26 (2011) 516522. rently a full professor and the Dean of the Department
[28] P. Cordo, H. Lutsep, L. Cordo, W.G. Wright, T. Cacciatore, R. Skoss, Assisted of Rehabilitation Medicine, First Hospital, Peking Univer-
movement with enhanced sensation (AMES): coupling motor and sensory sity, China. Her research interests include rehabilitation
to remediate motor deficits in chronic stroke patients, Neurorehabil. Neural medicine, stroke rehabilitation and rehabilitation devices.
Repair. 23 (2009) 6777.
[29] K. Homma, M. Usuba, Development of ankle dorsiflexion/plantarflexion
exercise device with passive mechanical joint, in: Proceedings of the IEEE 10th
International Conference on Rehabilitation Robotics, 2007.
[30] A. Mirelman, P. Bonato, J.E. Deutsch, Effects of training with a robot-virtual
reality system compared with a robot alone on the gait of individuals after
Fan Gao received his Bachelor degree in Mechanical
stroke, Stroke 40 (2008) 169174.
Engineering from Peking University, China and Ph.D.
[31] J.E. Deutsch, J.A. Lewis, G. Burdea, Technical and patient performance using a
degree in Mechanical Engineering from Pennsylvania State
virtual reality-integrated telerehabilitation system: preliminary finding, IEEE
University, USA. He is currently an Assistant Professor from
Trans. Neural Syst. Rehabil. Eng. 15 (2007) 3035.
Department of Health Care Sciences, University of Texas
[32] E. Bressel, P.J. McNair, The effect of prolonged static and cyclic stretching on
Southwestern Medical Center, Dallas, USA. His research
ankle joint stiffness, torque relaxation, and gait in people with stroke, Phys.
interests include rehabilitation device, muscle mechanics,
Ther. 82 (2002) 880887.
orthotics and prosthetics.
[33] J.A. Saglia, N.G. Tsagarakis, J.S. Dai, D.G. Caldwell, Control strategies for patient-
assisted training using the ankle rehabilitation robot (ARBOT), IEEE/ASME
Trans. Mechatronics 18 (6) (2013) 17991808.
122 Z. Zhou et al. / Robotics and Autonomous Systems 73 (2015) 111122

Kunlin Wei received his two M.A. degrees (in Kinesi-
ology and Electrical Engineering, respectively) and Ph.D.
degree in Kinesiology from the Pennsylvania State Uni-
versity, USA. Before that, he obtained his B.E. in Biome-
chanics from Beijing Sports University, China. He is an
Associate Professor from Department of Psychology,
Peking University, Beijing, China. Dr. Weis research inter-
ests include sensorimotor control, robotics, biomechanics,
humanmachine interaction, and motor rehabilitation.
search interests are in the fields of bio-inspired robots and rehabilitation systems.
Qining Wang received his Bachelor degree in Computer
Science and Technology from China University of Geo-
sciences (Beijing) in 2004, and the Ph.D. degree in Dynam-
ics and Control from Peking University in 2009. He was an
Assistant Professor in the Center for Systems and Control,
College of Engineering, Peking University, from July 2009
to July 2012. He is currently an Associate Professor in the
College of Engineering, Peking University, and the Director
of the Beijing Engineering Research Center of Intelligent
Rehabilitation Engineering. He is the Project Leader of the
Robotic Prosthesis R&D Group, Peking University. His re-