IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 18, NO.

6, DECEMBER 2013

1799

Control Strategies for Patient-Assisted Training

Using the Ankle Rehabilitation Robot (ARBOT)
Jody A. Saglia, Member, IEEE, Nikos G. Tsagarakis, Member, IEEE, Jian S. Dai, Member, IEEE,
and Darwin G. Caldwell, Member, IEEE

AbstractThis paper presents the control architecture and experimental results of the high-performance Ankle Rehabilitation
roBOT, ARBOT. The goal of this study is to design suitable control algorithms for assisted training and rehabilitation of the ankle
joint in presence of musculoskeletal injuries. A position control
scheme is used for patient-passive exercises, while an admittance
control technique is employed to perform patient-active exercises
with and without motion assistance. The selection and design of
the control algorithms are based on the analysis of the rehabilitation protocol taking into account the dynamics of the system and
the dynamics of the interaction between the human and the robot.
The performance of the proposed control algorithms is analyzed
through experiments on a group of healthy subjects.
Index TermsBiomechatronics, parallel robots, rehabilitation
robotics, robot control.

I. INTRODUCTION
VER the past decades, several studies demonstrated that
rehabilitation robots have a great potential in improving
diagnostics and physiotherapy outcome [1][4]. The main advantage of the automated rehabilitation systems is the capability
to perform a large number of repetitions, which was proved to
be extremely beneficial in the treatment of neuromuscular injuries [5]. Furthermore, such systems turn out to be extremely
precise diagnostic tools and can provide quantitative measures
of the patients recovery state after an injury [6]. As a result,
many systems are being currently developed and tested [7] and
require the implementation of advanced control strategies for
assisted training.
In the aspect of ankle rehabilitation systems, the Rutgers Ankle introduced by Girone et al. [8], [9] was the first system to
be based on a parallel mechanism, with a position controller for
passive training to drive the patients foot along certain trajectories and a force controller for active (only resistive) exercises.
A single degree of freedom (DOF) device proposed by Zhang
et al. [10] employed velocity control to mobilize the impaired

Manuscript received July 8, 2011; revised October 27, 2011, February 5,

2012, and May 20, 2012; accepted July 25, 2012. Date of publication September 6, 2012; date of current version December 11, 2013. Recommended by
Technical Editor A. Menciassi. This work was supported by the Istituto Italiano
di Tecnologia in collaboration with Kings College London.
J. A. Saglia, N. G. Tsagarakis, and D. G. Caldwell are with the Istituto Italiano di Tecnologia, 16163 Genoa, Italy (e-mail: jody.saglia@iit.it;
nikos.tsagarakis@iit.it; Darwin.Caldwell@iit.it).
J. S. Dai is with Kings College London, London, WC2R 2LS, U.K.,
and also with the Istituto Italiano di Tecnologia, 16163 Genoa, Italy (e-mail:
jian.dai@kcl.ac.uk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TMECH.2012.2214228

foot in a certain range of motion (ROM) and a torque sensor to

monitor the resistive torque provided by the patients foot. This
1-DOF system could perform only passive exercises.
A wide set of rehabilitation exercises were proposed by Yoon
et al. [11], ranging from passive mobilization (ROM recovery)
up to proprioceptive training such as balance exercises. Position and impedance control theories were used to implement
the rehabilitation regimes. Furthermore, a pseudo-assistive controller was developed to assist the patient to complete the given
task. The performance was limited by the actuation technology
used (i.e., pneumatic cylinders) and the indirect estimation of
interaction force/torque (FT).
Torque sensing for assistiveresistive exercises was used by
Lin et al. [12]. A fuzzy logic controller, which regulated the joint
angle and torque of a single DOF device, was implemented to
assist or resist the patient when trying to follow a target on a
screen. Constant assistive and resistive torques were used for
active training.
The passive mechatronic device MecDEAR by Bucca et al.
[13] used only position control for passive ROM recovery exercises and did not provide any control strategy for active exercises.
Although all the works mentioned previously are well justified and show promising results, none of the systems met the
requirements considered in this study such as providing control
algorithms for most of the exercises foreseen by standard rehabilitation protocols, including passive and active training with
effective assistive and resistive capabilities.
The Rutgers Ankle can be considered the most successful
system so far. However, the system did not provide assistance in
case of patients limited motion capabilities. The rehabilitation
systems introduced in [11] and [12] employed some sort of
assistive control; however, the amount of assistive torque/motion
needs to be set a priori by the physiotherapist and does not adapt
to patients motion during training.
The goal of this study was to design a control framework
for the ankle rehabilitation robot ARBOT [14], which allows us
to perform most of the rehabilitation exercises foreseen by the
standard rehabilitation protocols and provide assistance when
the patients ankle mobility is limited.
Based on the rehabilitation robot introduced in [14], see Fig. 1,
this paper presents the development of the control architecture
for this ankle rehabilitation system. The control system needs
to facilitate a wide range of exercises defined by the rehabilitation protocol (introduced in [14]) and serve as a tool for the
physiotherapist to treat patients in a faster and more effective
manner. To achieve this, a set of control algorithms, such as position, admittance, and admittance-based assistive control have

1083-4435 2012 IEEE

1800

Fig. 1.

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 18, NO. 6, DECEMBER 2013

High-performance ankle rehabilitation robot.

been developed, implemented, and tested on a group of healthy

subjects.
In this paper, Section II briefly describes the ankle rehabilitation robot ARBOT, while Section III reports on the analysis of the dynamics of the interaction. Consequently, Section
IV presents the control algorithms and shows how those algorithms are suitable for the needs of the various rehabilitation
exercises. The stability of the proposed control schemes is studied and presented in Section V. Finally, the experimental results
are reported in Section VI and discussed in Section VII.
II. ANKLE REHABILITATION ROBOT
Fig. 1 demonstrates the ankle rehabilitation robot used in
this study. The robot is basically a 3UPS/U1 parallel mechanism with two rotational DOFs. The mechanical structure is
composed of a fixed base, a central strut [15], [16], a moving
platform, and three actuated limbs with a UPS kinematic chain.
The platform is attached to the central strut through a universal
joint. As depicted in Fig. 1, the patients foot is constrained to
the foot plate with Velcro stripes.
The limb prismatic joint is actuated by a custom-designed
linear actuator. This actuator makes use of a brushed DC motor
Maxon RE40 and a planetary gearbox with a reduction ratio
of 12:1. A capstan system of pulleys together with a steel cable
transmission transforms rotary motion of the motor into linear
motion of the piston. Optical encoders of 4095 ppr mounted on
the DC motor shafts provide a position resolution of 1.278 m
at the prismatic joint.
The custom-designed actuator can provide a peak force of
over 1100 N and a maximum speed of 60 cm/s. The resultant
maximum device output torque is 120 Nm and the maximum
speed is 500 /s. More details on the system can be found in [14].
An ATI six-axis FT sensor mounted between the moving
platform and the footplate senses the humanrobot interaction
force and torque. The device is interfaced to a standard PC
through a CAN BUS interface.

1 U, P, and S stand for universal, prismatic, and spherical joint, respectively.

An underlined letter represents an actuated joint.

Fig. 2. Equivalent kinematic tree. Note, S, U, P, and G stand for spherical,

universal, prismatic joints, and ground.

III. DYNAMICS OF INTERACTION

A. Mechanism Dynamics
The dynamics of the redundantly actuated parallel mechanisms can be analyzed by considering the dynamics of an equivalent tree system [17][22], generated by cut-opening some of
the passive joints in order to break all the kinematic loops (see
Fig. 2). In this case, kinematic loops have been cut at the spherical joints in order to create three identical limb tree systems and
a strut/platform tree system. The dynamics of the equivalent tree
system is expressed by
+ Ct (q,
q) q + Nt (q) =
Mt (q) q

(1)

where Mt , Ct , and Nt are the inertia, coriolis and centrifugal

and eventually gravity matrices of the tree system, respectively,
q is the vector of generalized coordinates which contains all active (actuated) and passive joint variables. The vector contains
forces and torques of the tree system.
Considering a set of independent generalized variables q in ,
in this case the two angular DOFs of the platform, the mapping
between the independent platform torques and those of the tree
system can be expressed as
in = JT

(2)

while the actuator forces can be mapped to the independent

platform torque with the relation
in = JT
r fa.

(3)

The matrix J is the nonsquare Jacobian matrix that relates

the time derivative of the generalized cordinates of the tree system q with the time derivatives of the platform independent
coordinates q in , while Jr is the nonsquare Jacobian matrix of
the redundantly actuated parallel mechanism and f a the actuator forces. Inserting (1) into (2) and (3) and expressing the

SAGLIA et al.: CONTROL STRATEGIES FOR PATIENT-ASSISTED TRAINING USING THE ANKLE REHABILITATION ROBOT (ARBOT)

generalized coordinates q in terms of independent coordinates

q in , yield the following equation of motion (EoM) of the parallel
mechanism:
M
q in + Cq in + N = JT
r fa

1801

TABLE I
CONTROL ALGORITHMS FOR REHABILITATION EXERCISES

(4)

where
M = JT Mt J
C = JT Mt J + JT Ct J
N = JT Nt .
Therefore, the inverse dynamics is given by
fa =

J+T
r

(M
q in + Cq in + N)

(5)

being the pseudo-inverse Jacobian matrix of the rewith J+T

r
dundantly actuated parallel mechanism, computed in the form
that minimizes the actuator forces [23][26].
B. Dynamics of HumanRobot Interaction
The dynamics of the robotic rehabilitation system is described
by (5). In a scenario where the robot is applied to rehabilitation,
the interaction between the robot and the human as defined
by the exercise must be included in the dynamic model. The
interaction FT can be included in (5) as
T
(M qin + C q in + N + patient )
fa = J+
r

(6)

where patient is the torque vector applied by the user to the

footplate. Further to this, when the patient needs to perform active exercises, it may be necessary to replicate certain dynamics
at the end-effector of the system via imposing specific inertia,
stiffness, and damping parameters. In this case, the torque felt by
the patient will be equal to those obtained from the replication
of the desired massspringdamper system as
patient = virtual .
Hence, the expression in (6) becomes
T
in + C q in + N + virtual )
(M q
fa = J+
r

(7)

where virtual is the vector of torques required to replicate the

desired dynamics of the regime (see Section V).
IV. CONTROL STRATEGIES FOR REHABILITATION EXERCISES
The rehabilitation protocol for ankle injuries can be seen in
Table I [14]. To permit effective execution of these regimes, the
control architecture employs a specific control scheme appropriate to satisfy the requirements of particular exercises.
In the early stage of the therapy, the patient can hardly move
his/her foot; therefore, a passive exercise which delicately moves
the patients foot is needed.
Such a task can be accomplished by a position control scheme
which drives the injured foot/ankle along a certain trajectory at
a moderate speed. Trajectory parameters, such as wave type,
speed, amplitudes, and number of repetitions, can be set by the
physiotherapist.
In order to allow the patient to fully regain his/her ROM and to
evaluate the patients progress of the first stage of rehabilitation,

active exercises can be executed by the device using an assistive

control scheme based on admittance techniques.
Suppose the patient is capable of providing moderate torque
levels to initiate the motion; however, he/she cannot provide
enough torque to complete the exercise trajectory. In this case,
the application of the patient torque to the footplate can be monitored by the installed FT sensor, therefore providing information
to the assistive controller about the patient intentioned motion.
The assistive control (see Section IV-C) supplies the additional
torque effort required in order to assist the patient to complete
the motion.
The position control algorithm is also used for isometric
strengthening exercises. In this case, the ankle rehabilitation
robot is controlled to maintain a fixed position while the patient tries to apply a certain level of torque to the footplate. The
applied torque is measured by the FT sensor. Strength training includes also isotonic exercises, as in Table I. For this kind
of regime, an admittance controller is implemented in order to
provide a certain resistance to the patients motion.
In the last stage of the rehabilitation process, the patient has to
undergo proprioceptive training. Balance exercises are typical
for this kind of training and in such a case, the patient has to stand
on top of the rehabilitation robot and try to keep the balance,
as if he/she was using a wobble board. Since position, velocity,
and the dynamic behavior can be controlled, more sophisticated
exercises can be performed with this system, than those allowed
by traditional tools (foam rollers, wobble boards, etc.). Hybrid
control (combination of position and force control) is used to
design this type of exercises.
The study presented in this paper focuses on the first two steps
of the rehabilitation protocol while proprioceptive training will
be treated elsewhere.
A. Patient-Passive Exercises
As mentioned previously, when the patient is passive, the
robot is controlled to follow a reference trajectory imposed by
the therapist or to hold a certain position. In order to obtain
high position tracking accuracy [27], a computed-torque controller was implemented. The inverse dynamics introduced in
(5) is used to compute the actuation forces required by the rehabilitation robot to follow a certain trajectory. The EoM is
therefore used in the control loop to linearize the system and
a proportional-integral-derivative controller is used to compensate for modeling errors and its contribution is added to the
reference acceleration.

1802

Fig. 3.

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 18, NO. 6, DECEMBER 2013

Assistive control model.

B. Patient-Active Exercises: Strengthening

When considering strengthening exercises, two different
cases need to be distinguished. As mentioned previously, isometric training requires a fixed position; therefore, a computedtorque control algorithm as the one used for patient-passive
exercises can be implemented.
To perform isotonic exercises or other types of resistive training, an admittance control is chosen. As in (7), a certain dynamic
behavior of the rehabilitation device can be simulated in relation
to the patientrobot interaction.
By measuring the interaction torque applied by the user to
the footplate, it is possible to compute the reference position required to render certain mass, stiffness, and damping parameters.
Hence, the reference position q in r is obtained with an admittance
filter as
m
q in r =
(8)
(ms2 + bs + k)
where m, b, and k are the desired mass, damping, and spring
parameters, and m is the interaction torque measured with the
six-axis FT sensor.
Using the computed-torque control scheme, the dynamics
expressed by the denominator of (8) can be realized, since all
the other dynamic components are compensated by the system
linearization.
C. Patient-Active Exercises: Assistance
Assistive control is required in the early stage of rehabilitation
when the patient cannot complete the movement alone and needs
to reacquire his/her ROM. Assistance can be provided with
the robot, by monitoring the interaction torque between the
user/patient and the robotic device.
To achieve motion assistance, the robot is controlled in admittance with the reference position of the assistive (i.e., admittance) filter being modified in the direction toward which the
patient is attempting to move. The patients intended direction
of motion is considered to be the direction of the measured interaction FT. As in Fig. 3, when the patient applies a torque along a
certain direction, the equilibrium point of the assistive network
will be moved toward the same direction, therefore generating
a pulling effect.
This behavior is achieved by measuring the interaction FT
which is then integrated over time and used to update the robot
reference position, in terms of equilibrium point of an assistive
spring-damper network.

Fig. 4.

Patientrobot dynamic model.

In particular, the reference position is obtained from

q in r =

m
+ q as
(bs + k)

with


q as =

(9)

ka m dt

(10)

where q as is the assistive component of the position reference

and ka is a weight constant that determines the level of assistance. The greater the value of ka , the higher the assistance
provided. The torque is measured with the six-axis FT sensor.
V. STABILITY OF THE CONTROLLED SYSTEM
The stability of the position control, based on the computation
of the actuator torques via computedtorque control, is well
known and documented in the literature. This is also true for the
admittance controller, where the reference position is computed
through the measured torque and an admittance filter as in (8).
On the other hand, the stability of the assistive controller
needs to be proved and details on the stability analysis are given
in this section.
A. Stability Analysis of the Assistive Control Algorithm
Fig. 4 shows the dynamic model of the rehabilitation robot
controlled in assistive mode, linked with the patients leg.
The model represented in Fig. 4 includes the mechanics of
the human ankle complex and the mechanics of the assistance
filter. The interaction between the patient and the rehabilitation
robot device is measured by the FT sensor mounted between
the platform and the footplate. The measured torque m results
from the difference of the torque applied by the human patient
and the assistive torque provided by the device as
m = patient as .

(11)

By introducing the variable s, it is possible to analyze the

system in the frequency domain. Therefore, the expression of
the assistive motion component becomes
q as = Ga m

(12)

SAGLIA et al.: CONTROL STRATEGIES FOR PATIENT-ASSISTED TRAINING USING THE ANKLE REHABILITATION ROBOT (ARBOT)

with


Ga =

Ga P D
Ga I E


1 ka P D
=
s

1803


ka I E

where Ga P D and Ga I E represent the transfer functions between

the measured torque m and the equilibrium position q as of the
assistive network. The assistive network is expressed as
Gv = bv s + kv

(13)

where bv and kv are the damping and stiffness parameters.

The springdamper network in (13) relates the assistive
torque and the difference between the equilibrium position of
the network and the actual position of the patientdevice system
as
as = Gv (q as q patient ).

(14)

The dynamics of the human can be expressed as the difference

of the assistive torque and the dynamics of the foot, resulting in
patient = Gh q patient as
with Gh = [

(15)

Gh P D

], where Gh P D and Gh I E are the equivaGh I E

lent impedance of the human ankle for foot plantardorsiflexion
and inversioneversion, respectively (where for simplicity the
coupling effects have been neglected).
Considering a rehabilitation exercise where only plantar
dorsiflexion are involved, with Gh P D = Ih P D s2 + bh P D s +
kh P D , and by substituting (11) into (12), then combining with
(14) and (15) yield
as P D
Gv Ga P D Gh P D G v
=
patient P D
Gv Ga P D Gh P D + Gh P D + Gv

(16)

which represents the assistance ratio.

By substituting (14) and (15) into (11) and rearranging yields
qpatient P D
1 + 2Ga P D Gv
=
qas P D
Ga P D Gh P D + 2Ga P D Gv

(17)

which represents the transfer function between the equilibrium

position of the assistive springdamper network as input and the
humanrobot measured position. Expanding the transfer function results in
qpatient P D
((1/ka P D ) + 2bv ) s + 2kv
.
=
2
qas P D
Ih P D s + (bh P D + 2bv ) s + (2kv + kh P D )
(18)
It is possible to notice that the transfer function of the human
device system, controlled in assistive mode, results in a secondorder system with certain inertia, damping, and stiffness parameters.
The characteristic equation of the transfer functions in (18) is

Fig. 5. Bode diagram of the transfer function that relates the displacement of
the foot with the equilibrium position of the virtual admittance filter.

rewritten as
2kv + kh P D
= n2
Ih P D
=

bh P D + 2bv
2Ih P D n

bh P D + 2bv
= 
2 Ih P D (2kv + kh P D )

(19)

where n is the natural frequency for a stable, second-order,

closed-loop system and is the damping ratio. Through the control parameters of the assistive springdamper network (kv , bv ),
independent regulation of the natural frequency and damping ratio can be obtained.
Considering the scenario in which the dynamics of the human
ankle is precisely known, the stiffness and damping values of the
assistive network could be selected in order to obtain a desired
system response. Since this is not true in the real case, certain
assumptions need to be made in order to guarantee the stability
of the system.
According to [28], the impedance of a youngsters ankle is characterized by a mean damping value of bh P D =
2 Nmsrad1 , an inertia value of the foot of Ih P D = 0.22 kgm2 .
Given the dynamic characteristics of the human ankle, it is
possible to calculate the parameters of the springdamper system such as the complete humandevice system features a certain bandwidth and is stable. Therefore, to obtain device system
bandwidth of 30 Hz, the stiffness parameter kv and the damping
parameter bv were chosen as
kv = 10 N m rad1
bv = 20 N m s rad1

Ih P D s + (bh P D + 2bv ) s + (2kv + kh P D ).

This characteristic equation is subject to change as a result of
the variation in the mass and damping coefficients of the human
ankle/foot. Note that if (bh P D + 2bv ) > 0 and 2kv + kh P D > 0,
then the characteristic equation is Hurwitz. This may also be

and the value of ka P D , which determines the level of assistance,

was set to ka P D = 1.
Using these parameters, together with those relative to the
human foot dynamics aforementioned, it is possible to draw a
Bode diagram of the transfer function in (18). Fig. 5 shows the

1804

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 18, NO. 6, DECEMBER 2013

TABLE II
ADMITTANCE CONTROL PARAMETERS

Fig. 7.

Fig. 6. (a) Equilibrium position of the admittance filter and the real position
of the patient. (b) Human and the assistive torques versus time.

one inputone output Bode plot relative to a plantardorsiflexion

motion.
From the Bode diagram shown in Fig. 5, it is possible to
notice that the frequency response of the humanrobot system,
under assistive control, is that of a single-order system with a
cutoff frequency of about 30 Hz.
B. Simulation of Assistive Control
A simulation was performed by giving as input a pulse of
1 Nm for a period of 0.5 s to the human torque and evaluating
the humanrobot displacement together with the equilibrium
point of the admittance filter. Fig. 6(a) reports the trajectory of
the equilibrium point of the admittance filter compared with real
position of the patients ankle.
It is clear that the equilibrium point moves faster than the
humanrobot system and, as a result, the patient is pulled by
the rehabilitation robot toward the direction of attempted motion. Fig. 6(b) shows the torque applied by the patient and that
provided by the assistance. The simulation results had been
obtained with an assistance level ka P D = 1. The amount of assistive torque can be increased by setting ka P D to a higher value.

Experimental evaluation of stiffness.

Increasing the assistance level implies an increase of the

amount of assistive torque provided by the device.
On the other hand, reducing the level of assistance (e.g.,
ka P D < 1), it reduces the amount of help provided to the patients intentioned motion. In such a case, the patient initially
perceives a certain resistance given by the admittance filter. At
the same time, the equilibrium point starts to be upgraded according to the torque measured. Once the difference between
the equilibrium point and the patient position becomes positive,
the rehabilitation robot will start assisting the motion.
Furthermore, increasing damping and stiffness values of
the assistive network improves the position tracking performance when considering the equilibrium point as the reference
position.
VI. EXPERIMENTAL RESULTS
A. Torque-Sensing-Based Admittance Control
Using the computed-torque control scheme and the expression in (8), experiments were performed in order to evaluate the
ability of the system to reproduce certain stiffness and damping
levels.
The parameters of the admittance filters are given in Table II.
Note that the mass m has been set to zero and, in the case of pure
stiffness simulation, a small damping of bv = 1 Nmsrad1 was
necessary to guaranty system stability. The experiments have
been performed at low speeds (when reproducing pure stiffness)
in order to avoid the effect of inertial torques and forces. The
trials have been performed with a young volunteer, sitting on a
chair with his right foot constrained to the robotic device. The
subject was instructed to perform voluntary plantardorsiflexion
movements at self-selected low speed (maximum angular speed
reached was 0.06 rad/s) for the stiffness and high speed (as in

SAGLIA et al.: CONTROL STRATEGIES FOR PATIENT-ASSISTED TRAINING USING THE ANKLE REHABILITATION ROBOT (ARBOT)

Fig. 8.

Experimental evaluation of damping.

TABLE III
ADMITTANCE CONTROL RESULTSSTIFFNESS

TABLE IV
ADMITTANCE CONTROL RESULTSDAMPING

Fig. 8 the speed ranged between 0.5 and 2 rad/s) for damping
experiments. The results are shown in Figs. 7 and 8 and reported
in Tables III and IV. The graphs present the measured torque
applied by the user to the footplate versus the platform position
and velocity computed through the forward kinematics from the
measured limb lengths. It is possible to see that the stiffness
felt by the user in Fig. 7 and Table III matches the reference
stiffness value. Table III reports the slopes of the curves shown
in the graph.
Looking at the last column of Table III, it is evident that the
robot performance decreases with the increase of desired stiffness. The reason is that for high desired output stiffness, the
effect of the actuator passive compliance (due to the elasticity
of the transmission cable) on the overall stiffness increases. This
can be improved by either adding pretension in the redundant
parallel mechanism using the actuation force null space or controlling the actual position of the piston rather than the position
of the motor (see [25]).
Fig. 8 shows the results for the rendering of damping. The
noise that can be observed in the curves is due to the numerical
derivation of the angular position. It can be seen in Table IV that
the efficacy of the robot in rendering damping is high for all the
reference damping values.

Fig. 9.

1805

Subjects leg with two pairs of electrodes for EMG signals collection.

B. Torque-Sensing-Based Assistive Control

A first experiment was conducted in order to evaluate the
proposed assistive control algorithm. A young male subject was
asked to perform the experiment. The subject right foot was
constrained to the footplate and two pairs of electrodes were
applied to his leg to measure the muscle activity during motion
according to [29], see Fig. 9.
During the experiment, the electromiographic (EMG) signals
were recorded in order to evaluate the amount of subjects muscular effort. The signals were collected via surface electrodes
placed on leg flexor (tibialis anterior) and extensor (gastrocnemius) muscles. The EMG signals collected by a portable device
at the frequency of 128 Hz had to be filtered to remove artifacts
and interference [30]. In particular, a preliminary high-pass filtering at 20 Hz was employed to remove artifacts and trend
components. Subsequently, the signal was rectified and finally
low-pass filtered at 5 Hz to obtain an envelope.
Two trials were conducted and consisted in extending and
contracting the foot to reach about 15 of maximum extension
and return back to the start position in presence and absence of
motion assistance. During the first extension/flexion trial, the
subject experienced a zero counteracting torque with the rehabilitation device being controlled for zero torque. In the second
trial, the admittance parameters are set to kv = 80 Nm/rad and
bv = 8 Nms/rad, resulting in an overall system bandwidth of
about 10 Hz. These springdamper network parameters kv and
bv were used with the addition of the assistive control component described in Sections IV-C and V. The results of this
experiment can be seen in Fig. 10.
Fig. 10(a) reports the amount of EMG signals required to
perform the motion, while Fig. 10(b) depicts the angular displacement of the foot/platform. Dashed lines represent motion
without assistance (first trial), while continuous lines show motion with assistance (second trial).
It is possible to notice that, during the first trial, the subject
has to provide a certain effort in terms of muscle activity [see
Fig. 10(a)]. Differently, in the second trial, the application of
the assistive control keeps the peak force applied by the subject
at lower level during the initiation of the movement.

1806

Fig. 10.
control.

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 18, NO. 6, DECEMBER 2013

Comparison of humanrobot interaction with and without assistive

Once the subject applies a torque to initiate the motion, the

assistive control becomes active, by updating the equilibrium
position q as [shown in (9)] of the assistive springdamper network. This results in the device starting to pull the subjects foot
toward the direction he previously begun the movement. Note
that the maximum speeds of the two trials [when the foot moves
from the starting position to the target position of about 15 ,
Fig. 10(a)] are comparable and reach about 40 /s. As for the
return phase, the foot motion is helped by muscle relaxation.
In order to further evaluate the effectiveness of the proposed
assistive controller, more experiments have been performed with
a group of five healthy subjects. The experimental group was
a sample of male subjects, aged between 30 and 35, with an
average height of 1.73 0.15 m and a mean weight of 75.4
3.8 kg. Each subject was in a sitting position with the right
foot constrained to the ankle rehabilitation robot, as in Fig. 9.
The robot was equipped with a screen displaying the angular
position of the subjects foot (which corresponded to the robot
platform as well). The subject was instructed to plantarflex his
foot from the initial position of 0 to the target position which
was represented by the area between 15 and 20 . Fig. 11 shows
the experimental results.
The curves reported in Fig. 11(a)(c) represent the average
and standard deviation of the results obtained from the five
subjects. Fig. 11(a) shows the equilibrium position of the assistive admittance filter qas P D generated by (10) and the effective
position of the robot and the subjects foot qpatient P D . It is possible to notice that the equilibrium position is moved toward
the target position and the subjects position follows due to the
pulling assistive torque as P D , as shown in Fig. 11(b). Moreover, Fig. 11(b) depicts the subjects torque patient P D which
has been computed using (11) (where as is the assistive torque
produced by the robotic device estimated through motor current
measurement). It is noticeable that the magnitude of the assistive
torque is comparable with and greater than the subjects torque.
Fig. 11(c) shows the instantaneous work done by the interaction
between the subject and the robotic device which was obtained
considering the interaction torque m P D and the instantaneous
foot-robot displacement qpatient P D . In other words, it is the

Fig. 11

Experimental results of assisted motion.

work that the patient does on to the device. As reported in the

graph, the work becomes negative during the motion meaning
that the device is pulling the subject in the direction of motion.
These results demonstrate the effectiveness of the proposed
assistive controller and confirm the results reported in Fig. 10. It
should be noted that by varying the assistance constant in (12),
it is possible to regulate the level of assistive torque.
VII. CONCLUSION
This paper presented the control schemes of a highperformance parallel robot used for robot-aided ankle exercises.

SAGLIA et al.: CONTROL STRATEGIES FOR PATIENT-ASSISTED TRAINING USING THE ANKLE REHABILITATION ROBOT (ARBOT)

The rehabilitation protocol has been considered as the basis for

design of the control strategies. Both patient-passive and active
exercise modes have been addressed using position and admittance control schemes and the stability of the assistive controller
was addressed. The experimental trials of the stiffness/damping
tracking demonstrated the high performance of the rehabilitation robot and the experimental results obtained on a group
of healthy subjects showed the effectiveness of the proposed
admittance-based assistive controller.
The control algorithms presented in this study will serve as
framework for the design and development of patient-oriented
rehabilitation exercises.
Future work will look at the development of rehabilitation
exercises in collaboration with a team of clinicians and the
integration of a virtual environment to stimulate the patient
during training. Further studies will also focus on the integration
of the EMG information in the control algorithms.
REFERENCES
[1] H. I. Krebs, N. Hogan, M. L. Aisen, and B. T. Volpe, Robot-aided neurorehabilitation, IEEE Trans. Rehabil. Eng., vol. 6, no. 1, pp. 7587,
Mar. 1998.
[2] B. T. Volpe, H. I. Krebs, and N. Hogan, Is robot-aided sensorimotor
training in stroke rehabilitation a realistic option? Curr. Opin. Neurol.,
vol. 14, no. 6, pp. 745752, Dec. 2001.
[3] P. S. Lum, C. G. Burgar, P. C. Shor, M. Majmundar, and M. Van der
Loos, Robot-assisted movement training compared with conventional
therapy techniques for the rehabilitation of upper-limb motor function
after stroke, Arch. Phys. Med. Rehabil., vol. 83, no. 7, pp. 952959, Jul.
2002.
[4] J. Hidler, D. Nichols, M. Pelliccio, and K. Brady, Advances in the understanding and treatment of stroke impairment using robotic devices,
Stroke Rehabil., vol. 12, no. 2, pp. 2235, 2005.
[5] H. I. Krebs, B. T. Volpe, M. L. Aisen, and N. Hogan, Increasing productivity and quality of care: Robot-aided neuro-rehabilitation, J. Rehabil.
Res. Dev., vol. 37, no. 6, pp. 639652, Nov./Dec. 2000.
[6] S. E. Fasoli, H. I. Krebs, J. Stein, W. R. Frontera, R. Hughes, and N. Hogan,
Robotic therapy for chronic motor impairments after stroke: Follow-up
results, Arch. Phys. Med. Rehabil., vol. 85, no. 7, p. 1106-1111, Jul.
2004.
[7] M. Hillman, Rehabilitation robotics from past to presentA historical
perspective, in Proc. 8th Int. Conf. Rehabil. Robot., 2003, pp. 2325.
[8] M. J. Girone, G. C. Burdea, and M. Bouzit, The Rutgers ankle orthopedic
rehabilitation interface, in Proc. ASME Int. Mech. Eng. Congr. Dyn. Syst.
Control Div., Nashville, TN, Nov. 1999, vol. 67, pp. 305312.
[9] M. Girone, G. Burdea, M. Bouzit, and J. Deutsch, Orthopedic rehabilitation using the Rutgers ankle interface, in Proc. Med. Meets Virtual
Reality, Jan. 2000, pp. 8995.
[10] L.-Q. Zhang, S. Chung, Z. Bai, D. Xu, E. van Rey, M. Rogers,
M. Johnson, and E. Roth, Intelligent stretching of ankle joints with
contracture/spasticity, IEEE Trans. Neural Syst. Rehabil. Eng., vol. 10,
no. 3, pp. 149157, Sep. 2002.
[11] J. Yoon, J. Ryu, and K.-B. Lim, Reconfigurable ankle rehabilitation robot
for various exercises, J. Rob. Syst., vol. 22, no. S1, pp. 1533, 2005.
[12] C.-C. K. Lin, M.-S. Ju, S.-M. Chen, and B.-W. Pan, A specialized robot
for ankle rehabilitation and evaluation, J. Med. Biol. Eng., vol. 28, no. 2,
pp. 7686, 2008.
[13] G. Bucca, A. Bezzolato, S. Bruni, and F. Molteni, A mechatronic device
for the rehabilitation of ankle motor function, J. Biomech. Eng., vol. 131,
no. 12, pp. 125001-1125001-7, 2009.
[14] J. A. Saglia, N. G. Tsagarakis, J. S. Dai, and D. G. Caldwell, A high
performance redundantly actuated parallel mechanism for ankle rehabilitation, Int. J. Rob. Res., vol. 28, no. 9, pp. 12161227, 2009.
[15] J. S. Dai and D. R. Kerr, Six-component contact force measurement
device based on the Stewart platform, J. Mech. Eng. Sci., vol. 214, no. 5,
pp. 687697, 2000.
[16] J. A. Saglia, J. S. Dai, and D. G. Caldwell, Geometry and kinematic
analysis of a redundantly actuated parallel mechanism that eliminates

[17]
[18]
[19]
[20]
[21]
[22]

[23]
[24]

[25]
[26]
[27]
[28]
[29]
[30]

1807

singularity and improves dexterity, Trans. ASME, J. Mech. Design,

vol. 130, no. 12, pp. 124501-1124501-5, 2008.
H. Cheng, G. F. Liu, Y. K. Yiu, Z. H. Xiong, and Z. X. Li, Advantages
and dynamics of parallel manipulators with redundant actuation, in Proc.
IEEE/RSJ Int. Conf. Intell. Robot. Syst., Maui, HI, 2001, pp. 171176.
F. Caccavale, B. Siciliano, and L. Villani, The tricept robot: Dynamics
and impedance control, IEEE/ASME Trans. Mechatronics, vol. 8, no. 2,
pp. 263268, Jun. 2003.
Z. M. Bi and S. Y. T. Lang, Kinematic and dynamic models of a tripod
system with a passive leg, IEEE/ASME Trans. Mechatronics, vol. 11,
no. 1, pp. 108111, Feb. 2006.
Y. Li and Q. Xu, Design and development of a medical parallel robot
for cardiopulmonary resuscitation, IEEE/ASME Trans. Mechatronics,
vol. 12, no. 3, pp. 265273, Jun. 2007.
K. S. Sollmann, M. K. Jouaneh, and D. Lavender, Dynamic modeling of
a two-axis, parallel, H-frame-type XY positioning system, IEEE/ASME
Trans. Mechatronics, vol. 15, no. 2, pp. 280290, Apr. 2010.
L. Wang, J. Wu, J. Wang, and Z. You, An experimental study of a
redundantly actuated parallel manipulator for a 5-DOF hybrid machine
tool, IEEE/ASME Trans. Mechatronics, vol. 14, no. 1, pp. 7281, Feb.
2009.
H. Cheng, Y.-K. Yiu, and Z. Li, Dynamics and control of redundantly
actuated parallel manipulators, IEEE/ASME Trans. Mechatronics, vol. 8,
no. 4, pp. 483491, Dec. 2003.
sek, V. Bauma, Z. sika, K. Belda, and P. PIsa, Design-byM. ValA
optimization and control of redundantly actuated parallel kinematics sliding star, Multibody Syst. Dyn., vol. 14, no. 34, pp. 251267, Nov.
2005.
J. A. Saglia, N. G. Tsagarakis, J. S. Dai, and D. G. Caldwell, Inversekinematics-based control of a redundantly actuated platform for rehabilitation, J. Syst. Control Eng., vol. 223, no. 1, pp. 5370, 2009.
X. Zhu, G. Tao, B. Yao, and J. Cao, Adaptive robust posture control
of parallel manipulator driven by pneumatic muscles with redundancy,
IEEE/ASME Trans. Mechatronics, vol. 13, no. 4, pp. 441450, Aug. 2008.
H. Abdellatif and B. Heimann, Advanced model-based control of a 6DOF hexapod robot: A case study, IEEE/ASME Trans. Mechatronics,
vol. 15, no. 2, pp. 269279, Apr. 2010.
C. Cornu, M.-I. Almeida Silveira, and F. Goubel, Influence of plyometric
training on the mechanical impedance of the human ankle joint, Eur. J.
Appl. Physiol. Occup. Physiol., vol. 76, no. 3, pp. 282288, 1997.
F. P. Kendall, E. K. McCreary, P. G. Provance, M. Rodgers, and W. Romani,
Muscles: Testing and Function, with Posture and Pain. Baltimore, MD:
Williams & Wilkins, 2005.
C. J. DeLuca, The use of surface electromyography in biomechanics,
J. Appl. Biomech., vol. 13, no. 2, pp. 135163, 1997.

Jody A. Saglia (M09) received the B.Eng. degree

in automation engineering and the M.Sc. degree in
mechatronics engineering from the Polytechnic of
Turin, Turin, Italy, in 2004 and 2007, respectively,
and the Ph.D. degree from Kings College London,
London, U.K., in 2010.
He was a researcher at Kings College London
in 2006. Since 2007, he has been a Research Fellow
at the Istituto Italiano di Tecnologia, Genoa, Italy,
where is currently a Postdoctoral Researcher in the
Department of Advanced Robotics. His research interests include mechanisms design, rehabilitation robotics, mechatronic design,
actuation systems, and humanrobot interaction.
Dr. Saglia received the Professional Engineering Publishing Award from the
Journal of Systems and Control Engineering in 2009.

1808

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 18, NO. 6, DECEMBER 2013

Nikos G. Tsagarakis (M00) received the B.Eng. degree in electrical and computer science engineering
from the Polytechnic School of Aristotle University
of Thessaloniki, Thessaloniki, Greece, in 1995, and
the M.Sc. degree in control engineering and the Ph.D.
degree in robotics from the University of Salford, Salford, U.K., in 1997 and 2000, respectively.
He is currently a Senior Researcher at the Istituto
Italiano di Tecnologia, Genoa, Italy, where he is responsible for humanoid design and human-centered
mechatronics. He is the author or coauthor of more
than 150 papers in research journals and international conference proceedings,
and is the holder of six patents.
Dr. Tsagarakis received the Professional Engineering Publishing Award from
the Journal of Systems and Control Engineering in 2009 and the Best Paper
Award at the International Conference on Advanced Robotics in 2003. He was
also a finalist for the Best Entertainment Robots and Systems20th Anniversary Award at the International Conference on Intelligent Robots and Systems
(IROS) in 2007 and for the Best Manipulation Paper at the International Conference on Robotics and Automation (ICRA) in 2012. He has been on the program
committees of more than 40 international conferences including ICRA, IROS,
Robotics Science and Systems, and Humanoids.

Jian S. Dai (M95) received the B.Sc. and M.Sc. degrees from Shanghai Jiao Tong University, Shanghai,
China, and the Ph.D. degree from the University of
Salford, Salford, U.K.
He became a Lecturer at Shanghai Jiao Tong University in 1985. In late 1980s, he joined the University of Salford as a Research Scholar. In 1997, he
was a Senior Lecturer in mechanisms and robotics
at the University of Sunderland. In 1999, he joined
Kings College London, University of London, London, U.K., as a Lecturer, where he later became a
Reader in mechanisms and robotics and the Chair of mechanisms and robotics.
He is also with the Istituto Italiano di Tecnologia, Genova, Italy. He has authored
or coauthored more than 400 papers. His research interests include screw theory, mechanisms development, reconfigurable mechanisms and robotics, multifingered robot hands, grasping and manipulation, rehabilitation robotics, and
mechanisms and robotics in assembly and packaging.
Dr. Dai is the recipient of a number of best journal papers and conference
papers awards, and many IEEE and American Society of Mechanical Engineers
(ASME) service awards. He is the Chair of Mechanisms and Robotics, and a
Fellow of ASME and the Institution of Mechanical Engineers.

Darwin G. Caldwell (M92) received the B.Sc. and

Ph.D. degrees in robotics from the University of
Hull, Kingston upon Hull, U.K., in 1986 and 1990,
respectively.
He is currently a Director at the Italian Institute of
Technology, Genoa, Italy. He is a Visiting/Honorary
Professor at The University of Sheffield, The University of Manchester, The University of Bangor,
Kings College London, all in the U.K., and Tianjin University, China. He is the author or coauthor of
more than 350 academic papers, and holds 15 patents.
His research interests include innovative actuators, humanoid and quadrupedal
robotics and locomotion (iCub, HyQ, and COMAN), haptics, exoskeletons,
dexterous manipulators, and rehabilitation and surgical robotics.
Dr. Caldwell is the recipient of several awards from international journals
and conferences. He is an Associate Editor for the IEEE/ASME TRANSACTIONS
ON MECHATRONICS and a member of the Editorial Board of the International
Journal of Social Robotics and Industrial Robot.