2018 IEEE International Conference on Robotics and Automation (ICRA)

May 21-25, 2018, Brisbane, Australia

Adaptive Oscillator-based Control for Active Lower-Limb Exoskeleton

and Its Metabolic Impact
Keehong Seo Kyungrock Kim Young Jin Park Joon-Kee Cho Jongwon Lee
Byungjune Choi Bokman Lim Younbaek Lee Youngbo Shim

Abstract— We developed a robotic lower-limb exoskeleton for mild chronic stroke patients. Given the challenge of human-
those who have weakened muscle due to aging and experience robot interaction problem, we have developed several control
difficulty in walking or getting up without help. The exoskeleton approaches for tasks such as walking on level ground, stairs
covering both limbs from the feet to the waist has 6 electric
actuators in the hip abduction/adduction, hip extension/flexion and slopes. The controllers based on adaptive oscillator (AO)
and knee extension/flexion joints. For users with volitional [11], [13], [15], [16] and finite state machine (FSM) [17],
motion, delivering assistance power according to their intention [18], [14], [19] as well as environment recognition method
is a challenging task. We propose an adaptive oscillator-based [20] have been proposed and validated with healthy subjects
controller to assist users walk in the lower-limb exoskeleton. [13], [15], [16] and the elderly [21]. Lim et. al. in [18]
To adapt to changes in walking speed and environment, motion
command from the controller is modulated by estimate walking have worked on simulation-based optimization of control
speed and walking environment recognized as one of the parameters for patients with neuropathy.
following categories: level ground, stairs up/down and slope
up/down. Experimental results demonstrate the feasibility of While we are witnessing the efficacy of the single-joint
the proposed environment recognition method and the impact exoskeletons with the target users, there are people who
of assistance on the metabolic cost of walking on level and
inclined treadmills. need more than single-joint assistance; for example, one can
manage to walk alone and can be short of strength to stand
I. I NTRODUCTION up without help. He or she may also need help to climb up a
single step. In this case, a lightweight multi-joint exoskeleton
Increasing number of people suffer from loss of gait that is capable of partial body weight support should be
function due to various reasons including aging and neu- helpful.
ropathy. In industry, technologies to prevent injuries of
workers handling heavy loads are demanded. As a solution In this paper, we present a lower-limb type exoskele-
to these issues, robotic technologies are actively transferred ton, Gait Enhancing Mechatronic System for Lower Limb
to the domain of medical or industrial exoskeletons. In the (GEMS-L), with actuated hip abduction/adduction (a/a), hip
course, various types of exoskeletons have been developed by extension/flexion (e/f) and knee e/f joints and passive ankle
universities and industrial companies. Their purposes, control joints to partially support the body weight and its control
methods, and mechanisms are all different, which could framework. The GEMS-L in Fig. 1 features lightweight and
however fall into several categories as seen in the extensive fitting silhouette to minimize conflict with the environments
reviews [1], [2], [3]. Some exoskeletons like ATLAS [4], of daily lives.
Ekso [5], ReWalk [6] were developed for paraplegic or
quadriplegic people and cover full lower limb for assistance. To control the device, we propose to use AOs. An AO
In this application, wearers would have no volitional motion learns a periodic signal in terms of its frequency compo-
in the legs and the exoskeletons can only replay pre-recorded nents, amplitudes and phases online. The way we apply
leg motion when triggered. Other exoskeletons like HAL [7] AOs to GEMS-L is less complicated than other AO-based
and SMA have been developed for those with muscle weak- approaches for lower limb exoskeletons. Yan and colleagues
ness or partially impaired gait. Increased mobility of stroke [22] gave different control objectives at different FSM phases
patients using SMA was reported in [8]. A soft wearable for knee joints while hip joints were controlled by AOs;
robot Exosuit [9] has shown that it can improve mobility and [23] implemented force field toward predicted positions
of post-stroke patients [10]. In this type of exoskeleton or from periodic motion. In our approach, controller obtains
exosuit, the users have volitional motion and thus human- continuous gait phase from a single AO and then feeds it to
robot interaction becomes a main challenge. assistance torque generators to control heterogeneous joints
At Samsung Electronics Co. Ltd. (Suwon, Korea) we have such as hip a/a, hip f/e and knee f/e. For adaptability in
developed a series of single-joint exoskeletons — several hip- changing condition, the torque generation is modulated by
types [11], [12], [13] and an ankle-type [14] to support the walking speed and environment. The walking environment
elderly with degraded gait performance and the moderate-to- is recognized by using a support vector machine as one
of 5 classes: level, slope up/down, and stairs up/down.
Authors are with Samsung Advanced Institute of Technology, Samsung
Electronics. Co. Ltd. Suwon, South Korea; Correspondance should be sent Experimental test results on the recognition accuracy and
to Keehong Seo keehong.seo at samsung.com assistance performance of GEMS-L is presented.

978-1-5386-3081-5/18/$31.00 ©2018 IEEE 6752

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.
 pose and spatiotemporal information of the exoskeleton on-
line as in Fig. 3. Walking speed is also available at this stage
and is fed to torque generating module.
The environment classification module, by observing se-
lected features from the pose and the foot motion data,
predicts the current walking environments into 5 classes:
level, slope up, slope down, stair up, and stair down. The
prediction is then provided to the torque generation module.
The foot-to-foot distance projected onto the sagittal plane,
marked in Fig. 3, is computed from the pose and then used
as input for an AO module which estimates the gait phase.
The gait phase is used by the torque generation module.
All the actuated DOFs are torque-controlled throughout
the whole gait cycle as long as the user is walking. The initia-
Fig. 1: A lower limb exoskeleton, Samsung GEMS-L tion and termination of the walking assistance is determined
by the states of the left and right foot FSMs. Assistance
torque command for each actuated joint is generated from
II. T HE E XOSKELETON D ESIGN a predefined torque function of gait phase. The function
GEMS-L in Fig. 1 is a lower-limb exoskeleton to support parameters are modulated by walking speed and walking
users with their daily activities including walking, sit-to- environment. Recognized walking environment is rendered
stand and so on. Joints at hip a/a, hip e/f, and knee e/f continuous via a low-pass filter to guarantee the continuity
are actuated by electrical motors with maximum torque of of torque command. Torque patterns for hip a/a and knee e/f
30 Nm. Without output torque sensors, motor torque is esti- were determined by referring to the biological data in [25]
mated and controlled by sensing motor electrical current. No and hip e/f pattern was adopted from the previous study [13]
foot pressure sensor is equipped; the foot contact information on the GEMS-H.
and overall pose are estimated from 5 IMUs placed on the Torque from each motor is controlled by an off-the-
pelvis, shank, and foot segments. shelf motor controller sensing the electric current and then
Featuring lightweight and fitting design, the overall weight delivered to a wearer through gears with reduction ratio of
including the batteries is 9.98 kg and the extruding height of about 36:1 for the knee actuators, and 47:1 for the hip e/f and
the exoskeleton out of the body silhouette is maximal as 4.2 a/a actuators. It is open-loop torque control without sensors
cm at the hip joint and thigh frame. To minimize unwanted to measure output torque delivered to a wearer. Preliminary
tension due to misalignment of the knee joint, a novel joint experiments on the actuators had been performed to identify
mechanism is devised, which adapts to changes in relative gear efficiency and the high-level torque command is scaled
joint positions of the device and the human by applying up accordingly to compensate for loss.
under-actuated multiple rolling contact joints as presented The overall control architecture is illustrated in Fig. 2.
in [24]. The exoskeleton is also capable of supporting body The AO module and environment classification are further
weight of 10 kg as experimentally validated in [24] by lifting explained below.
10.3 kg from a squatting posture.
A. Gait Phase Estimation with AO module
III. C ONTROL A RCHITECTURE
Particularly-shaped adaptive oscillator (PSAO) estimates
Using IMU sensors attached on the back of the waist,
gait phase as a continuous and periodic value from 0 to 2π.
the shanks, and the feet, the estimation module obtains
Unlike other AOs, its basis function for the lowest frequency
information on the exoskeleton pose, foot state, and foot
is a mapping from gait phase to the nominal pattern of input,
motion. The pose describes relative position and orientation
making estimated gait phase interpretable. The difference
of each segment and joint angles. The foot state block
between the nominal pattern and the actual input is filled
consists of an FSM that determines the foot state as one
by harmonic oscillators that learns it. For example, input to
of the following: foot-impact, foot-rest, heel-rise, and foot-
PSAO in [11] was hip joint angles and the basis function was
swing by using IMU signals. Transition rules in the FSM
the nominal hip joint pattern with respect to conventional
was hand tuned until the foot FSM was accurate enough to
gait phase which starts with heel strike, resulting in the
use with dead reckoning. The foot motion block describes
synchronization of actual heel strike and zero gait phase.
the velocity of a foot in the global coordinates obtained
The dynamics of PSAO is as follows.
by integrating the acceleration signals from IMU on the
foot. The computed foot velocity is used for environment φ̇1 = ω + kφ eg(φ1 ) (1)
classification.
α̇1 = kα ef (φ1 ) (2)
The dead reckoning module then processes the information
of current pose and foot state to obtain spatiotemporal φ̇i = iω + kφ e cos(φi ) (3)
information of user motion. As a result, we can monitor the α̇i = kα e sin(φi ) (4)

6753

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.
 Fig. 2: Functional Block Diagram for Control of GEMS-L

ω̇ = kω eg(φ1 ) (5) in general, which comes with trade-off in noise rejection,

α̇0 = ko e (6) making the system susceptible to noise.
n
 Using PSAO gives an advantage over AOs in the conver-
θ̂ = α0 + α1 f (φ1 ) + αi sin(φi ), (7) gence because the basis function f (φ) already contains prior
i=2 information on the input to learn making the system less
where i = 2, . . . , n is the index of oscillators; θ is the input; dependent to initial values. In addition to such advantage,
φ1 and φi are the phases of the oscillators in [0, 2π); α1 ≥ 0 in order to enhance the convergence we propose to couple
and αi ≥ 0 are the amplitudes of the oscillators; α0 is an the oscillator with an FSM that detects the frequency of
offset; ω ≥ 0 is the frequency of the base oscillator; k’s with transitions from the input. In fact, FSMs have been widely
subscripts are gains; θ̂ is the estimation for θ; e = θ − θ̂ is used in the literature to estimate gait phase as a discrete
an estimation error. The function f (φ) is the basis for the value such as stance and swing. For example, Lim et. al.
base oscillator, which is the pattern of interest with respect [17] have demonstrated discrete gait phase detection, where
to the phase; and g(φ) is its derivative. FSM transited at zero-crossings in hip joint angles and hip
In this paper, foot-to-foot distance is chosen for input θ. joint angular velocities. In our controller, the gait frequency
Although the basis function f (φ) should be the nominal ωF SM measured by an FSM, updated at every transition, is
pattern of foot-to-foot distance over a gait cycle, it is simply then coupled as follows.
set to a cosine due to its similarity as shown in the top plot ω̇ = kω eg(φ1 ) + k(ωF SM − ω) (8)
of Fig. 4.
The effect of a coupling from FSM frequency was tested in
B. Enhanced Convergence of AO
simulation as in Fig. 4, where the gait phase is estimated
As is often the case with error-driven dynamic system, correctly in shorter time when the coupling is stronger.
PSAO or AOs converges to an equilibrium asymptotically.
When the input changes its periodic behavior the phase C. Environment Classification
estimation can be delayed until it converges to a new equilib- We categorized walking environment into 5 classes: level
rium. The convergence can be shortened by increasing gains ground, stair-up, stair-down, slope-up, slope-down because
the joint torque patterns should be generated differently for
each environment. The following features have been selected
as descriptors to distinguish gaits in the 5 environments.
1) φ1 : maximum foot swing speed on vertical axis before
crossing with other foot
2) φ2 : maximum foot swing speed on vertical axis after
crossing with other foot
3) φ3 = φ1 − φ2
4) φ4 : vertical foot speed when crossing with other foot
Fig. 3: 3-d pose and foot contact state are computed in real 5) φ5 : difference in the hip angles at foot-impact
time from IMU sensor data. Using dead-reckoning we can 6) φ6 : difference in the knee angles at foot-impact
obtain the spatio-temporoal locomotion data and visualize it. 7) φ7 : ankle angle at foot-impact
The foot-to-foot distance for AO module is marked on the 8) φ8 : ankle height difference from the other ankle at foot-
side view panel. impact

6754

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.
 ,QSXWIRRWWRIRRWGLVWDQFH


*DLW3KDVHN 

OHIW ULJKW



Fig. 5: A subject is walking over stairs and ramps installed
*DLW3KDVHN  indoors to test an environment recognition model.

IV. I MPACT ON THE M ETABOLIC C OST OF WALKING


*DLW3KDVHN 
To evaluate the impact of the exoskeleton and its control
framework on wearers, we conducted an experiment with 2
 healthy male subjects. Full understanding on the impact of
assistance at different joints and the synergy or interference
 between them would require testing all the possible combi-
       
7LPHVHF nations of actuated joints. In this study, however, we had to
limit ourselves to the following cases: no actuation; actuate
Fig. 4: Adaptive oscillator coupled to discrete phase estima- hip e/f only; actuate hip e/f and hip a/a; and actuate all 6
tor with different values for coupling gain k shows different joints. For all these cases, joints were bilaterally actuated for
convergence behaviors. the left and right sides.
To compare the impact of exoskeleton for walking on level
and inclined treadmill, the subjects put on K5 (Cosmed, Italy)
9) φ9 : ankle advance from the other ankle at foot-impact wearable metabolic system and went through the protocol in
10) φ10 : knee height difference from the other knee at foot- Fig. 6. The first subject stood still for 5 minutes and then
impact walked on the level treadmill of 3.6 km/h speed for 6 minutes
without the exoskeleton. After wearing the exoskeleton, he
We used support vector machine (SVM) to classify the walked again under the following assistance conditions for
environments. To train an SVM model we collected data 6 minutes each: (i) Exo-Off: zero-torque command (ii) Exo-
from 4 subjects walking on level ground, a slope and a stair- HP: hip e/f on, (iii) Exo-HP/HR: hip e/f and a/a on (iv)
case installed in the lab (Fig. 5). An instance of observation Exo-HP/HR/K: hip e/f, a/a and knee on. He then took
was obtained from each step and the total numbers of the off the exoskeleton to walk for 6 minutes again and then
instances collected were 362, 71, 82, 59, and 58 respectively stood still for 5 minutes. The treadmill was then inclined to
for level, slope-up, slope-down, stair-up, and stair-down. 12% gradient and set to 2.4 km/h speed. The subject then
We then used a software library for SVM model, LIBSVM proceeded to the following set of conditions. He walked for
[26] with the following setting: C-SVC mode, RBF kernel 6 minutes, and then put on the exoskeleton to go through the
and γ = 0.16. The trained model was tested for two new following 4 conditions for 6 minutes each: (i) Exo-Off, (ii)
subjects wearing the full exoskeleton without actuation. The Exo-HP, (iii) Exo-HP/HR, (iv) Exo-HP/HR/K. The subject
number of test instances for each environment ranged from then took off the exoskeleton and walked on the inclined
20 to 43 and all the 142 instances were classified correctly treadmill for 6 minutes and finally stood still for 5 minutes
implying 100% accuracy as in Table I. Both for the training on level ground. For the second subject we reversed the order
and the test of the SVM model, we excluded steps in to rule out time effect. For the brevity of presentation, the
transition such as start or stop of walking, which should standing condition and walking without exoskeleton will be
addressed in our future work. referred to as Stand and No-Exo, respectively. Stand(1)/No-
Exo(1) and Stand(2)/No-Exo(2) for level/incline walking
TABLE I: Environment Classification Test refer to the standing/walking-without-exoskeleton condition
before and after walking on the level/inclined treadmill in
Subject 1 2
Level 11/11 9/9 the exoskeleton, respectively. Two successive Stand trials
Slope-Up 11/11 16/16 between the level walking and the incline walking were
Slope-Down 23/23 20/20 supposed redundant and therefore replaced by a single Stand
Stair-Up 14/14 14/14
Stair-Down 12/12 12/12 trial as illustrated in Fig. 6.
Total 71/71 71/71 The time series data of GEMS-L in Fig. 7 shows how
AO estimated gait phase in walking as well as how joint
kinematics changed as the hip e/f joint starts actuation during

6755

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.
 Hip Extension Torque (Nm)
5
0
-5

Hip Flexion Angle (rad)

0.8
0.6
0.4
0.2
0
-0.2

Fig. 6: Protocol for the experiment: subject 1 and subject 2

Knee Flexion Angle (rad)
went through in opposite orders.
1
0.5
0
the experiment. For example, one can see the knee joint angle
trajectory change as the hip joint actuation starts. Gait Phase Estimated by AO
Ensemble average of the GEMS-L time series data over 1

gait cycle for a subject at different conditions is shown in Fig. 0.5

8 for level and incline walking. Other than the data collected 0
by the exoskeleton, human motion itself was not measured 450 455 460 465 470 475 480
in this experiment. Time (sec)

For level and incline walking, we applied different assis- Fig. 7: Time series GEMS-L data from walking on the level
tance torque patterns except for hip a/a. For level walking, treadmill
hip extension and knee extension torque in average ranged
from -5.6 to 7.4 Nm and from -2.3 to 6.4 Nm, respectively.
For incline walking, the averages ranged respectively from We can also compute net metabolic rate (NMR), EEm for
-5.3 to 10.3 Nm and from -2.3 to 10.6 Nm especially walking minus EEm for standing, to estimate the energetic
emphasizing extension torque. We applied the same torque cost of walking only. We then computed the mean NMR of
pattern in the hip abduction for level and incline walking the subjects for each condition. The lowest NMR for level
but peak hip abduction assistance torque was 9.8 Nm for walking was found at Exo-HP/HR: it was 19.2% lower than
level walking and 8.1 Nm for incline walking because the Exo-Off although 27.7% greater than No-Exo. For incline
controller was designed to generate less torque at lower walking the lowest NMR was at Exo-HP/HR/K, which was
speed. 28.6% lower than Exo-Off and 8.8% lower than No-Exo.
Exoskeleton joint power, the product of joint angular No other condition but Exo-HP/HR/K at incline walking
velocity and assistance torque, is also plotted in the figure. decreased metabolic cost below that of No-Exo. Fig. 10
Positive values indicate that power was generated by mo- shows the NMR for each subject and the mean NMR in
tors and then assisted the wearer unless dissipated by the comparison to the NMR of No-Exo.
exoskeleton. Negative values indicate that the exoskeleton From the metabolic data in Fig. 10, it is clear that
absorbed power generated by the wearer or other joint assistance in hip-e/f reduces the metabolic cost of walking
actuators. The power pattern for hip e/f indicates that the both in level and incline walking. We observed assistance in
assistance power flowed into the user motion with positive hip a/a saved metabolic cost at some cases and at least did
peaks in the stance and the swing phase. not act adversely. It is however difficult to understand why
The metabolic energy expenditure rate for each subject is assistance in the knee joint reduced NMR in incline walking
shown in Fig. 9. The energy expenditure per minute (EEm) and increased NMR in level walking. It could be related
was recorded per breath by the measurement system and then to the knee assistance power pattern in Fig. 8. Knee joint
we computed median for the second half of the duration of power in level walking has smaller area under the positive
each condition — 3 minutes for walking and 2.5 minutes for values and larger area under the negative values than incline
standing. By observing EEm data in the figure, we notice walking. The positive peak values is also higher in incline
some similarity in the way how the subjects react to each walking by the factor of 2.
assistance condition. First of all, incline walking is more
demanding than level walking as we expected. Wearing the V. C ONCLUSION
exoskeleton without any assistance torque increased EEm We presented a lower-limb exoskeleton GEMS-L and its
substantially. Actuating the hip e/f joint lowered EEm in level control framework along with the impact on the metabolic
and incline walking. Actuating hip a/a in addition further cost of walking.
lowered EEm but the change is relatively small. Actuating Walking environment recognition based on IMU sensors
the knee joint brought conflicting results. It lowered EEm and joint angle sensors proved its feasibility by marking
even further for incline walking; but raised EEm for level 100% accuracy in the test with 2 subjects. The lack of recog-
walking in both subjects. nition for transient actions such as start and stop walking

6756

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.
 (a) Level Walking (3.6 km/h) (b) Incline Walking (2.4 km/h, 12% gradient )

Fig. 8: Ensemble average of GEMS-L data of a subject walking on the level and inclined treadmill. Solid lines are the
ensemble average over gait cycles; shades represent 5 to 95 percentiles.

7.67
Level 3.6 km/h Level 3.6 km/h
7.34

8 12% incline 2.4 km/h 8 12% incline 2.4 km/h

6.82

6.72

6.67

6.64
6.51
6.47

6.25

6.11
5.88

5.84

5.83
5.82
EEm (kcam/min)

EEm (kcam/min)

5.58
5.55
5.53

5.51
5.18

5.06

6 6

4.79
4.35

4.33

4.5

4 4
1.85

1.85
1.71

1.68

1.67
1.58
1.4

1.4

2 2

0 0
/K

/K
1)

2)

1)

2)
P

P
R

R
1)

2)

1)

2)
ff

ff
O

-O
H

-H
H

H
d(

d(

d(

d(
o(

o(

o(

o(
o-

o-

P/

P/
H

H
o

o
an

an

an

an
Ex

Ex

Ex

Ex
Ex

Ex

Ex

Ex
P/

P/
H

H
St

St

St

St
o-

o-

o-

o-
o-

o-
H

-H
N

N
Ex

Ex
o-

o
Ex

Ex

(a) Subject 1 (b) Subject 2

Fig. 9: For the 2 subjects, EEm is compared for the conditions in the protocol.

should be resolved for more sophisticated interaction. energy expenditure on 2 subjects. Among the various con-
Estimation of gait phase using AO was demonstrated. As ditions we tested, the action of the exoskeleton was found
an input for AO, we used foot-to-foot distance without a effective under some of the conditions. For example, in
reason to choose it over other candidates such as joint angles incline walking, metabolic cost of Exo-HP/HR/K was lower
or IMU signals. We need a comparative study on using than that of No-Exo by 8.8%. In level walking, however,
signals from different sensors in the exoskeleton as inputs actuated knee e/f had an adverse effect. To address this issue,
to AO. we need online or offline optimization of assistance torque
Using biological and hand-tuned torque patterns, hip e/f, patterns, investigation for mechanical interference between
hip a/a and knee e/f joints were actuated. We investigated actuated joints, or trying a new control approach such as
its energetic effect through the measurement of metabolic

6757

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.
 Subject 1 Level 3.6 km/h
70
[10] L. N. Awad, J. Bae, K. O’Donnell, S. M. M. De Rossi, K. Hendron,
Subject 2 Level 3.6 km/h
Subject 1 12% Incline 2.4 km/h
L. H. Sloot, P. Kudzia, S. Allen, K. G. Holt, T. D. Ellis, and C. J.
60 58.2 Subject 2 12% Incline 2.4 km/h Walsh, “A soft robotic exosuit improves walking in patients after
Mean Level 3.6 km/h stroke.” Science translational medicine, vol. 9, no. 400, p. eaai9084,
50 Mean Incline 2.4 km/h Jul. 2017.
[11] K. Seo, S. Hyung, B. K. Choi, Y. Lee, and Y. Shim, “A new adaptive
Impact on NMR (%)

40 37.0 frequency oscillator for gait assistance,” in 2015 IEEE International

30.5 Conference on Robotics and Automation (ICRA). IEEE, May 2015,
30 27.7 27.7
pp. 5565–5571.
20 [12] Y. Lee, B. Choi, J. Lee, M. Lee, S.-g. Roh, J. Kim, H. Choi, and Y.-J.
11.1 Kim, “Flexible sliding frame for gait enhancing mechatronic system
10 7.0 (GEMS),” in 2016 38th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC). IEEE, Aug.
0
2016, pp. 598–602.
-8.8
-10 [13] K. Seo, J. Lee, Y. Lee, T. Ha, and Y. Shim, “Fully autonomous hip
exoskeleton saves metabolic cost of walking,” in Proceedings - IEEE
-20 International Conference on Robotics and Automation, vol. 2016-June.
IEEE, May 2016, pp. 4628–4635.

/K
P

R
ff
O

R
o-

o-

P/

H
[14] H. Choi, Y. J. Park, K. Seo, J. Lee, S.-e. Lee, and Y. Shim, “A
Ex

Ex

P/
H
o-

H
Multifunctional Ankle Exoskeleton for Mobility Enhancement of Gait-
Ex

o-
Ex Impaired Individuals and Seniors,” IEEE Robotics and Automation
Fig. 10: Bars indicate relative changes in NMR for 4 assis- Letters, vol. 3, no. 1, pp. 411–418, Jan. 2018.
[15] K. Seo, J. Lee, and Y. Jin, “Autonomous Hip Exoskeleton Saves
tance conditions and for the 2 subjects in comparison with Metabolic Cost of Walking Uphill,” 2017 International Conference
No-Exo on the level/inclined treadmill. The mean NMR of on Rehabilitation Robotics (ICORR), pp. 246–251, Jul. 2017.
the two subjects are marked with circles. [16] J. Lee, K. Seo, B. Lim, J. Jang, K. Kim, and H. Choi, “Effects of
assistance timing on metabolic cost, assistance power, and gait param-
eters for a hip-type exoskeleton,” in 2017 International Conference on
Rehabilitation Robotics (ICORR). IEEE, Jul. 2017, pp. 498–504.
impedance control. For example, in [27], finding the optimal [17] B. Lim, Kyungrock Kim, Jusuk Lee, Junwon Jang, and Youngbo Shim,
“An event-driven control to achieve adaptive walking assist with gait
torque pattern for each joint online was possible if one can primitives,” in 2015 IEEE/RSJ International Conference on Intelligent
reduce the number of parameters to optimize. It was also Robots and Systems (IROS). IEEE, Sep. 2015, pp. 5870–5875.
possible in [18] to simulate with a neuromuscular model and [18] B. Lim, S. Hyoung, J. Lee, K. Seo, J. Jang, and Y. Shim, “Sim-
ulating gait assistance of a hip exoskeleton: Case studies for ankle
search for optimal assistance pattern with evolution strategy. pathologies,” in 2017 IEEE International Conference on Robotics and
Automation (ICRA). IEEE, May 2017, pp. 1022–1027.
R EFERENCES [19] J. Jang, K. Kim, J. Lee, B. Lim, and Y. Shim, “Assistance strategy
for stair ascent with a robotic hip exoskeleton,” in 2016 IEEE/RSJ
[1] T. Yan, M. Cempini, C. M. Oddo, and N. Vitiello, “Review of assistive International Conference on Intelligent Robots and Systems (IROS).
strategies in powered lower-limb orthoses and exoskeletons,” Robotics IEEE, Oct. 2016, pp. 5658–5663.
and Autonomous Systems, vol. 64, pp. 120–136, Feb. 2015. [20] ——, “Online gait task recognition algorithm for hip exoskeleton,” in
[2] H. Herr, “Exoskeletons and orthoses: classification, design challenges 2015 IEEE/RSJ International Conference on Intelligent Robots and
and future directions.” Journal of neuroengineering and rehabilitation, Systems (IROS). IEEE, Sep. 2015, pp. 5327–5332.
vol. 6, no. 1, p. 21, Jan. 2009. [21] H.-J. Lee, S. Lee, W. H. Chang, K. Seo, Y. Shim, B.-O. Choi, G.-H.
[3] B. Chen, H. Ma, L.-Y. Qin, F. Gao, K.-M. Chan, S.-W. Law, L. Qin, Ryu, and Y.-H. Kim, “A Wearable Hip Assist Robot Can Improve
and W.-H. Liao, “Recent developments and challenges of lower Gait Function and Cardiopulmonary Metabolic Efficiency in Elderly
extremity exoskeletons,” Journal of Orthopaedic Translation, vol. 5, Adults.” IEEE transactions on neural systems and rehabilitation
pp. 26–37, Apr. 2016. engineering, vol. 25, no. 9, pp. 1549–1557, Sep. 2017.
[4] D. Sanz-Merodio, M. Cestari, J. Arevalo, X. Carrillo, and E. Garcia, [22] T. Yan, A. Parri, M. Fantozzi, M. Cortese, M. Muscolo, M. Cempini,
“Generation and control of adaptive gaits in lower-limb exoskeletons F. Giovacchini, G. Pasquini, M. Munih, and N. Vitiello, “A novel
for motion assistance,” Advanced Robotics, vol. 28, no. 5, pp. 329– adaptive oscillators-based control for a powered multi-joint lower-limb
338, Mar. 2014. orthosis,” in 2015 IEEE International Conference on Rehabilitation
[5] K. A. Strausser and H. Kazerooni, “The development and testing of Robotics (ICORR). IEEE, Aug. 2015, pp. 386–391.
a human machine interface for a mobile medical exoskeleton,” 2011 [23] R. Ronsse, B. Koopman, N. Vitiello, T. Lenzi, S. M. M. De Rossi,
IEEE/RSJ International Conference on Intelligent Robots and Systems, J. Van Den Kieboom, E. Van Asseldonk, M. C. Carrozza, H. Van
pp. 4911–4916, Sep. 2011. Der Kooij, and A. J. Ijspeert, “Oscillator-based walking assistance: A
[6] A. Esquenazi, M. Talaty, A. Packel, and M. Saulino, “The ReWalk model-free approach,” in IEEE International Conference on Rehabil-
Powered Exoskeleton to Restore Ambulatory Function to Individuals itation Robotics. IEEE, Jun. 2011, pp. 1–6.
with Thoracic-Level Motor-Complete Spinal Cord Injury,” American [24] B. Choi, Y. Lee, Y.-J. Kim, J. Lee, M. Lee, S.-g. Roh, Y. J. Park,
Journal of Physical Medicine & Rehabilitation, vol. 91, no. 11, pp. K. Kim, and Y. Shim, “Development of adjustable knee joint for walk-
911–921, Nov. 2012. ing assistance devices,” in 2017 IEEE/RSJ International Conference
[7] Y. Sankai, “HAL: Hybrid Assistive Limb Based on Cybernics,” in on Intelligent Robots and Systems (IROS). IEEE, Sep. 2017, pp.
Springer Tracts in Advanced Robotics. Springer, Berlin, Heidelberg, 1790–1797.
2010, vol. 66, pp. 25–34. [25] R. Riener, M. Rabuffetti, and C. Frigo, “Stair ascent and descent at
[8] C. Buesing, G. Fisch, M. O’Donnell, I. Shahidi, L. Thomas, C. K. different inclinations,” Gait and Posture, vol. 15, no. 1, pp. 32–44,
Mummidisetty, K. J. Williams, H. Takahashi, W. Z. Rymer, and Feb. 2002.
A. Jayaraman, “Effects of a wearable exoskeleton stride manage- [26] C.-C. Chang and C.-J. Lin, “Libsvm,” ACM Transactions on Intelligent
ment assist system (SMA) R on spatiotemporal gait characteristics
Systems and Technology, vol. 2, no. 3, pp. 1–27, 2011.
in individuals after stroke: a randomized controlled trial.” Journal of [27] J. Zhang, P. Fiers, K. A. Witte, R. W. Jackson, K. L. Poggensee,
neuroengineering and rehabilitation, vol. 12, no. 1, p. 69, Aug. 2015. C. G. Atkeson, and S. H. Collins, “Human-in-the-loop optimization
[9] M. Wehner, B. Quinlivan, P. M. Aubin, E. Martinez-Villalpando, of exoskeleton assistance during walking,” Science, vol. 356, no. 6344,
M. Baumann, L. Stirling, K. Holt, R. Wood, and C. Walsh, “A pp. 1280–1284, 2017.
lightweight soft exosuit for gait assistance,” in 2013 IEEE Interna-
tional Conference on Robotics and Automation. IEEE, May 2013,
pp. 3362–3369.

6758

Authorized licensed use limited to: Nanyang Technological University. Downloaded on November 02,2020 at 01:40:07 UTC from IEEE Xplore. Restrictions apply.