A Cable-Driven Parallel Hip Exoskeleton for High-Performance Walking Assistance

Abstract

Misalignment between human and exoskeleton joints is a common issue in exoskeletons of rigid structure, as it can lead to discomfort or even injuries. Cable-driven exoskele-tons, by using human skeletal joints, remove misalignment as a potential issue. However, large parasitic forces due to cable pulling endure as a shortcoming of cable-driven exoskeletons. To address the problem, this paper proposes a novel cable-driven hip exoskeleton of parallel structures for assisted walking with eliminated parasitic force. The parasitic force potentially caused by either the misalignment or direct pulling can be removed mechanically. The new exoskeleton, conceptually different compared to existing anthropomorphic exoskeletons or soft exoskeletons, can conjugate flexibility and kinematic redundancy in flexion/extension and ab/adduction for self-alignment with anatomical joints. The unique design enables internal/external rotation for versatile walking gaits. In the work, the misalignment between the mechanical and biological hip joints is quantified both theoretically and experimentally. Moreover, an adaptive robust controller (ARC) is designed to provide desired force during assisted walking. Experimental results demonstrate the performance of the proposed cable-driven exoskeleton system and improved wearing comfort with parasitic forces eliminated.

Full text

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 1
A Cable-Driven Parallel Hip Exoskeleton for
High-Performance Walking Assistance
Xiangyang Wang, Sheng Guo, and Shaoping Bai, Senior Member, IEEE
Abstract—Misalignment between human and exoskeleton
joints is a common issue in exoskeletons of rigid structure, as it
can lead to discomfort or even injuries. Cable-driven exoskele-
tons, by using human skeletal joints, remove misalignment as
a potential issue. However, large parasitic forces due to cable
pulling endure as a shortcoming of cable-driven exoskeletons.
To address the problem, this paper proposes a novel cable-
driven hip exoskeleton of parallel structures for assisted walking
with eliminated parasitic force. The parasitic force potentially
caused by either the misalignment or direct pulling can be
removed mechanically. The new exoskeleton, conceptually differ-
ent compared to existing anthropomorphic exoskeletons or soft
exoskeletons, can conjugate flexibility and kinematic redundancy
in flexion/extension and ab/adduction for self-alignment with
anatomical joints. The unique design enables internal/external
rotation for versatile walking gaits. In the work, the misalignment
between the mechanical and biological hip joints is quantified
both theoretically and experimentally. Moreover, an adaptive
robust controller (ARC) is designed to provide desired force
during assisted walking. Experimental results demonstrate the
performance of the proposed cable-driven exoskeleton system
and improved wearing comfort with parasitic forces eliminated.
Index Terms—hip exoskeleton, anthropomorphic exoskeleton,
cable-driven parallel mechanism, self-alignment, parasitic force
elimination
I. INTRODUCTION
WEARABLE exoskeletons have strong potential for en-
hancing human motor functions and improving human
walking efficiency by delivering external power to the user
as needed [1]. For healthy users, portable and lightweight
hip exoskeletons are effective and desirable for augmenting
locomotion performance by enabling wearers to spend less
energy or carry an increased load [2].
A major goal for hip exoskeletons is to assist the wearer
walking through the surrounding environment without feeling
hindered [3]. However, this is very challenging because of the
rigid nature of existing exoskeletons [4]. A first requirement
on an exoskeleton that avoids hindering the wearer is to ensure
that its kinematic structure is compatible with the human
joints it assists [5]. If they are not properly aligned, kinematic
This work was supported in part by the National Natural Science Foundation
of China under grant 52275004, and in part by the Fundamental Research
Funds for the Central Universities under grant 2021YJS135. X. Wang ac-
knowledges financial support from Chinese Scholarship for his visiting study
at Aalborg University, Aalborg, Denmark. (Corresponding author: Sheng Guo)
X. Wang was with the Robotics Institute, School of Mechanical, Electronic
and Control Engineering, Beijing Jiaotong University, Beijing 100044, China,
and is currently with the Guangdong Provincial Key Lab of Robotics and Intel-
ligent System, Shenzhen Institute of Advanced Technology, Chinese Academy
of Sciences, Shenzhen 518005, China (e-mail:xy.wang2@siat.ac.cn).
S. Guo is with the Robotics Institute, School of Mechanical, Electronic
and Control Engineering, Beijing Jiaotong University, Beijing 100044, China
(e-mail:shguo@bjtu.edu.cn).
S. Bai is with the Department of Materials and Production, Aalborg
University, Aalborg 9220, Denmark (e-mail: shb@mp.aau.dk).
mismatch will lead to high parasitic forces, causing discomfort
or pain, and may even lead to long-term injury [6], [7].
Existing hip exoskeletons can be divided into two categories
according to their structure: anthropomorphic and soft [8]. In
anthropomorphic exoskeletons, it is desirable that the number
of degrees of freedom (DoF) of an exoskeleton matches the
human’s biological DoF [9], such that the exoskeletons possess
a motion pattern similar to that of human legs which helps
transfer of the torque from the exoskeleton to the wearer
properly [10]. In the last decade, many breakthroughs in
anthropomorphic exoskeletons have been reported, including
novel structural design and new control theories [11], [12].
However, it is still very challenging to eliminate the misalign-
ment between the mechanical and anatomical joints, and even
to identify and quantify the misalignment [13]. Currently, most
existing anthropomorphic exoskeletons provide one or two
DoFs for each leg and require manual alignment before use:
this is done by estimating the location of the hip center, which
requires fundamental knowledge of human anatomy. Because
the biological hip joint is not a perfect hinge, the exoskeleton
cannot adapt to the joint’s migration during movement [14].
Kinematic redundancy is usually utilized to remove mis-
matches. By introducing additional kinematic pairs, additional
DoFs were added. Hip exoskeletons developed by Beil et
al. [15] and Walsh et al. [16] are such examples. However,
exoskeletons of this form have typically featured bulky self-
aligning mechanisms and large inertias, which may increase
the metabolic cost and even reduce wearability [9].
Soft exoskeletons are lightweight and do not impose me-
chanical constraints on human motion [17]. Most soft ex-
oskeletons use Bowden cables to directly provide forces to
the joints. However, this can lead to a high pulling force,
with a peak pulling force up to 300 N [18]. Similar findings
were reported in [19] and [20] as well. High pulling force
would generate a large parasitic force in the direction along
the thigh. The parasitic force does not contribute to assisted
walking, instead, results in extra pressure in the hip due to the
absence of a rigid frame, possibly putting extra stress on the
joint ligaments and tissues.
To overcome the limitations mentioned above, in this paper
we propose a cable-driven hip exoskeleton with a parallel
structure. As demonstrated in Fig. 1, the exoskeleton has two
DoFs per leg, of which the hip flexion-extension (HFE) is pow-
ered remotely by electrical actuators, and the hip adduction-
abduction (HAA) DoF is passive. Moreover, no constraint is
applied to internal/external rotation. The design differs from
the existing anthropomorphic or soft exoskeletons, taking ad-
vantage of the parallel structure and cable-driven mechanisms
to achieve high-performance assisted walking. In this work,
high performance refers to versatile gaits, reduced joints’s
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 2
misalignment, and eliminated parasitic force. By providing
these advances, the proposed exoskeleton is able to improve
the wearing comfort while maintaining a minimum system
weight that most soft exoskeletons have.
This paper extends our previous work [21], which proposed
the concept design. In this paper, we quantify the human-
exoskeleton joint’s misalignment and develop a wearable ex-
oskeleton system. The main contributions of this work are as
follows.
1) A wearable exoskeleton incorporating parallel structure and
cable-driven mechanism is developed. The design enables
hip motion in all DoFs can be achieved for versatile gaits.
An adaptive robust controller is designed for this novel
exoskeleton to provide accurate assistance during walking.
2) The human-exoskeleton joints’ misalignment is quantified,
and the elimination mechanism of parasitic force is studied.
The experimental results verify that the proposed design
can eliminate the parasitic force in most existing anthro-
pomorphic or soft exoskeletons.
The rest of this paper is organized as follows. In Section II,
the exoskeleton design is introduced and the joints’ misalign-
ment quantification is carried out. Section III introduces the
working principle and controller design. In Section IV, human
testing is conducted. Discussion is given in Section V. Section
VI concludes this paper.
II. SY ST EM OV ERVIEW AND MODELING
A. Mechanical Structure
Fig. 1 shows the overall system of the proposed parallel
exoskeleton, which consists of the following core elements:
1) Flexible parallel frame. The parallel frame consists of
a padded belt, two brace cuffs, and six resin-printed flexible
branch rods. The padded belt is composed of five segments
connected with each other with a revolute joint to fit different
users’ body sizes. For each leg, an independent parallel
structure is formed between the padded belt and brace cuff
by three branches. The brace cuff is held from dropping by
these branch rods, and each branch rod is connected to the
belt and cuff through some kinematic pairs.
2) Cable-driven actuation. The actuation unit is fastened to
the back of the padded belt. It consists of a Li-ion battery,
circuit boards, and two motor-reducer components installed in
an up-down arrangement. Each motor’s output shaft is coupled
to a pulley. Steel rope was used for the power transmission.
To avoid interference with the human leg, the shape of
the connecting rods is designed in a particular manner (see
Fig. 1(c)), so they can wrap around the human thigh without
hindering normal leg motion [22]. The shape of the rods in
a 3-dimensional local space is described by: x=d
2cosθ,
y=−d
2sinθ,z=m(π−θ)
π, where (θ∈[0, π]).
It is noted that the cuff is not tightened, so there is a
small clearance between the cuff and the human thigh, which
allows a sliding DoF between them. Moreover, the DoF of
internal/external rotation is allowed.
Table I lists the range of motion (ROM) of the parallel
exoskeleton and that of humans for different motions. As can
be seen from the table, the ROM of the exoskeleton is larger
Fig. 1. Design of the parallel hip exoskeleton. (a) CAD models. 1. Pulley; 2.
padded belt; 3. steel rope; 4. inertial measurement unit (IMU); 5. single-axis
force sensors; 6. waistband; 7. fastener; 8. strap; 9. portable control unit; 10.
motor& reducer; 11. brace cuffs; 12. resin printed rods; 13. fixed seat. (b)
Zoomed view of the human-machine system. (c) Projections of the spatial
curve of a single branch rod.
than the human ROM of walking, which enables walking
assistance to be achieved.
B. Kinematic Model and Misalignment Quantification
The parallel structure was analyzed to reveal the kinematic
characteristics. A kinematic model of the parallel structure
attached to right leg was established as an example. Three
coordinate systems are first defined, as shown in Fig. 2.
1) Global coordinate system {A}:{xa, ya, za}is attached to
the padded belt, whose xz-plane is parallel to the sagittal
plane. The za-axis is perpendicular to the ground, and the
xa-axis points in the walking direction.
2) Thigh coordinate system {B}:{xb, yb, zb}is attached to
the human thigh. Its origin is located at the intersection
point O2between the human thigh and the brace cuff. The
TABLE I
JOINT ROM OF THE PARALLEL HIP EXOSKELETON AND HUMANS
Joints Human walking
Max. [23] Human Max. Exoskeleton Max.
Flexion/
Extension 32.2◦/22.5◦≥50◦/30◦52◦/−52◦
Abduction/
Adduction 7.9◦/6.4◦55◦/31◦58◦/−58◦
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 3
zb-axis is in line with the thigh direction and points to the
biological hip center P. Frame {B}can be obtained from
{A}by rotation about xa-axis with angle α, and rotation
about ya-axis with angle θ.
3) Cuff coordinate system {C}:{xc, yc, zc}is also attached
to the brace cuff at O2. Its zc-axis is perpendicular to the
plane formed by the kinematic pairs S1,S2, and S3. Frame
{C}can be obtained from {A}by rotation about xa-axis
with angle α2, and rotation about ya-axis with angle θ2.
All parameters are set based on actual measurements of a
physical exoskeleton prototype, as shown in Table II. For each
chain, a loop-closure equation can be written as
−−→
UiSi=−−→
OP +−−→
P O2+−−−→
O2Si−−−→
OUifor i= 1,2,3.(1)
The length of a link, li, shown in Fig. 2(a), is given by
l2
i=Asi−AuiTAsi−Aui(2)
where Asi,Auiare position vectors of points Siand Uithat
are embedded in frame {A}. We have
Asi=Ap+A
BR[0,0,−L]T+A
CRCsi(3)
where A
BRand A
CRare rotation matrices, Lis the distance
between points Pand O2. Corresponding to each given hip
angle defined by (α,θ), Eq. (2) allows us to find solution
for L. Let L0be the initial value when both αand θare
Fig. 2. (a) Geometric model of human-machine system. (b) VRC loci
specified in terms of three velocity vectors on the cuff. vsi represents the
velocity at points Si(i=1, 2, 3), Pis the anatomical hip center and P′is the
VRC, lirepresents link length, sis the misalignment between joints.
TABLE II
PARAMETERS USED IN THE KINEMATIC ANALYSIS
Description Unit (cm)
The positions of the
kinematic pairs’ center
Au1=−−→
OU1= [17.1,3.6,0]T
Au2=−−→
OU2= [−12.8,3.6,0]T
Au3=−−→
OU3= [2.4,−12.1,0]T
Ap=−−→
OP = [0,0,−21.2]T
Cs1=−−−→
O2S1= [1.8,8.8,0]T
Cs2=−−−→
O2S2= [10.9,−6.5,0]T
Cs3=−−−→
O2S3= [−6.6,−6.5,0]T
Length of links l1=l2= 54; l3= 52
Fig. 3. Models of misalignment. (a) Calculated sliding distance ∆Lduring
the leg HFE motion; (b) Overview of the VRC distribution and workspace
of the cuff in space for the given human ROM. 1. VRC distribution; 2.
Workspace; 3. Human thigh; 4. Brace cuff; (c) Misalignment quantification
results for different human ROM; (d) Diagram showing the relationship
between misalignment and sliding for the anthropomorphic design (left) and
the proposed design (right). Nand N′are contacting points on human leg
and brace cuff, respectively.
equal to zero. For any given ROM, the sliding distance of the
contacting point can be obtained as
∆L=L−L0.(4)
Assuming that the ROM of the human hip joint is within
±55◦in both the HFE and HAA directions (the same as the
human maximum angle listed in Table I), the sliding distance
can be theoretically determined based on (4) and is shown
in Fig. 3(a). The result shows relative sliding generates with
the leg movements. When the hip angle reaches 55◦,∆Lalso
reaches a maximum of 1.23 cm.
Moreover, the instantaneous virtual rotation center (VRC)
of the brace cuff and its workspace can also be determined.
Referring to Fig. 2 (b), the location of the VRC in 3D space,
Ap′= [x, y, z]T, can be determined by solving
vT
si ·−−→
P′Si= 0 (i= 1,2,3) (5)
where −−→
P′Si=Asi−Ap′is normal to the instantaneous
velocity vsi.vsi can be found by taking the time derivative
of both sides of (1)
vsi =wi×−−→
UiSi=wt×−−→
P O2+˙
Ln+wb×−−−→
O2Si(6)
where wiand wtare the angular velocities of the ith branch
and human thigh, respectively. Moreover, wb=h˙α2,˙
θ2,0iT
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 4
is the angular velocity of the brace cuff, ndenotes the unit
vector pointing from Pto O2, and ˙
Lis the relative sliding
speed between the cuff and thigh. Dot-multiplying both sides
of (6) by −−→
UiSiyields
wt·−−→
P O2×−−→
UiSi+˙
L−−→
UiSi·n+wb·−−−→
O2Si×−−→
UiSi= 0.
(7)
Equation (7) stands for three scalar equations for i= 1,
2, 3, thus all unknowns - ˙
L,˙α2, and ˙
θ2- can be uniquely
determined. Hence, vsi can be obtained from (6) to calculate
the location of the VRC Ap′, which is done by substituting vsi
into (5). Fig. 3(b) shows the distribution of the instantaneous
VRC for the given ROM. Based on this distribution, the
misalignment s, which is the distance between the VRC and
the hip center, can be quantified as
s=

Ap−Ap′

2(8)
where ∥·∥2denotes the 2-norm of a vector.
The anatomical hip center position Apis estimated by a
motion capture system (Cortex, Motion Analysis Co., USA)
based on placement of marker points on human body. So the
misalignment scan be calculated for any leg postures. Fig.
3(c) shows the quantification results of s, among them the
maximum value is found as 4.47 cm for walking and 10.9 cm
for maximum ROM of the human.
Compared with the anthropomorphic exoskeletons, the pro-
posed exoskeleton has a VRC which is very close to the
biological hip without adjustment after donning. Fig. 3(d)
shows how the misalignment caused the kinematic mismatch
(the green line) in both cases. Most existing anthropomorphic
exoskeletons have their brace cuffs tightened, as a result,
the kinematic mismatch at the interface could result in large
parasitic forces. In contrast, the proposed exoskeleton allows
relative sliding, which is a passive DoF and has no restriction
on relative motion. As a consequence, no parasitic force
would generate during leg movements and wearing comfort
is therefore improved.
C. Experimental quantification
To verify the theoretical identification result of the mis-
alignment and sliding, we quantified these two parameters
experimentally. In the experiment, two IMUs were placed on
the front and back sides of the brace cuff, recording the angular
position and velocity. The relative sliding between the cuff and
the thigh was measured using a magnetic grating transducer
(MR50, Milont Technology Co. Ltd, Shenzhen, China) by
attaching a magnetic stripe to the skin. The subject was asked
to wear the exoskeleton and move his leg back and forth
in three different ranges (marked as small: ±15◦, medium:
±22◦, and large: ±30◦) in the sagittal plane. For each range,
motion data for the last 30 gait cycles was recorded for post-
processing.
By calculating the intersection points of two planes that
are normal to the velocity vectors of the IMU placement
points, the actual VRC locations that are mapped into the
sagittal plane can be obtained and visualized in space. The
instantaneous VRC locations corresponding to three different
ROM levels are plotted, as shown in Fig. 4(a). They are
Fig. 4. (a) Experimentally obtained VRC locations in three ROM levels; (b)
measured sliding distance between the cuff and the thigh.
all around the hip center with the shape having the same
‘U’ arrangement as the theoretical result shown in Fig. 3(b).
The misalignment varies between 1.24 and 5.68 cm, which
is close to the obtained result from the kinematic analysis.
The maximum sliding is 10.43 mm, which is also close to the
results shown in Fig. 3(a). A paired t-test with the significant
level α= 0.05 on the mean misalignment and sliding was
conducted to test the statistical significance. The p-value for
the misalignment results was 0.057, and for the sliding was
0.741. Both show no significance were found.
With the parasitic force elimination mechanism, the ex-
oskeleton can transfer the pulling force from the rope to the
cuff where the assistive force is completely perpendicular to
the human thigh, generating pure assistive torque in the hip.
III. CON TRO LL ER DESCRIPTION
A. Human-exoskeleton Interaction Model
The exoskeleton was driven by a single-cable mechanism.
A simplified unilateral exoskeleton was used to model the
human-machine interaction. As shown in Fig. 5. A virtual axis
is used to indicate the cuff’s position, which passes through the
instantaneous VRC and is collinear with the cuff centerline.
The misalignment sis bounded, i.e., |s| ≤ δ1.δ1is the
maximum misalignment specified in Fig. 3 (c).
As the relative sliding doesn’t cause interaction force, the
interaction force Fhm is mainly due to the compressing of the
soft tissue of the human leg. The amount of deformation of
the soft tissue is dependent on the stiffness kat the interface.
For simplicity, only the interface’s main compliant property
was considered, and the other unmodeled uncertainties were
Fig. 5. Schematic of the human-machine interface model. (a) The config-
uration at initial state. (b) The configuration in arbitrary posture. qeand qh
are cuff’s angle and thigh angle, respectively. 
xhis a vector pointing from P
to N, which is used to indicate the thigh’s position. 
xeis a vector pointing
from P′to N′, which is used to indicate the cuff’s position.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 5
Fig. 6. Working principle of the cable-driven exoskeleton and walking
assistance. The blue curve indicates the force control reference to be delivered,
and the red curve is the integral of the force reference.
included in the lumped disturbance ˜
D1. Hence, the human-
machine interaction model [24] can be obtained as
Fhm =k∥xh−xe∥2+e
D1.(9)
The position of 
xeis determined by (qe, L +s), whereas

xhis determined by (qh, L). Hence, the term ∥xh−xe∥2can
be written as
∥xh−xe∥2=L(qh−qe)−∆(10)
where ∆is the nonlinear uncertainties caused by s.
Eq. (4) suggests that Lvaries during leg movements. How-
ever, the variation is experimentally tested to be about 1 cm,
which is much smaller than the thigh length and therefore has
little influence on Fhm. In our implementation, Lis regarded
as a constant. So (9) can be written as
Fhm =kL(qh−qe) + e
D1−k∆.(11)
B. Working Principle
The exoskeleton was designed to assist hip extension during
walking. The gait cycle can be divided into three phases:
mid-stance, end-stance, and swing. Accordingly, assistance is
delivered as the leg moves from the maximum hip flexion
(MHF) to mid-stance, during which the controller works in a
cascade force control mode (CFCM) and enables the desired
assistive force to be delivered. In the rest of the cycle, the
controller works in a rope release control mode (RRCM) and
feeds out rope, so it becomes slack, imposing no mechanical
constraints on human motion. In this work, we use a simplified
trapezoidal force profile as the control reference, as shown in
Fig. 6. The profile is adapted from the one used in [18], [25],
aiming to provide roughly the normative equivalent biological
hip joint moment during hip extension. The control mode
switches based on the detection of the key gait events, which
is achieved by a detection algorithm based on the IMU data.
According to (12), the force tracking problem is transferred
to the force integral tracking. By controlling the integral of the
measured human-machine interaction force to track the control
reference (the red line in Fig. 6), the desired assistive force is
achieved.
C. Cascade Force Control Mode (CFCM)
A first-order dynamical system can be formed from 11, with
which a controller is developed, as described presently.
d
dt Zt
0
Fhmdt =kL(qh−qe) + e
D1−k∆.(12)
In the first-order system, the lumped disturbance and the
unknown parameters can be defined as
e
∆2=kqh+e
D1/L −k∆/L = ∆2n+ ∆2m
β1=k, β2= ∆2n
(13)
where ∆2nis the constant part of e
∆2, and ∆2mis the time-
varying part. We define the following state variables:
x1=Zt
0
Fhmdt, x2=qe.(14)
The state space equation can thus be linearly parameterized
in terms of βas
˙x1=−β1x2L+β2L+ ∆2mL. (15)
For a desired force integral x1d, the state space equation of
the tracking error e=x1d−x1can be written as
˙e=β1x2L−β2L−∆2mL+ ˙x1d.(16)
The following assumption is made:
Assumption 1: The extents of the parametric uncertainties
and uncertain nonlinearities are known, that is,
β= [β1, β2]T∈Ωβ
∆
={β:βmin ≤β≤βmax}
|∆2mL| ≤ δ(t, x)(17)
where βmin = [β1 min, β2 min ]T,βmax = [β1 max, β2 max]T,
which represent the lower and upper bounds of unknown
parameters. δ(t, x)is a known number.
Based on the state space equation, the CFCM is grouped
into two control levels.
1) High-level adaptive robust controller: As shown in Fig.
7, the goal of the high-level controller is to synthesize the
proper position of the brace cuff x2din (16), so that the
tracking error econverges or is bounded. The established
model is nonlinear and contains the unknown parameters qh
and k. An adaptive robust controller is utilized to estimate the
unknown parameters, while achieving accurate force tracking.
The synthesized control law x2dis given by
x2d=qhm +qfb
qhm =−1
Lˆ
β1
( ˙x1d−ˆ
β2L)
qfb =1
β1max
(−k1e) + qd
(18)
Fig. 7. Diagram of the cascade force control mode.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 6
where ˆ
βi(i= 1,2) is an estimation of βi,˙x1dis the desired
force, and k1is a positive feedback gain. qhm is the model
compensation term. qfb is the robust feedback term. qdre-
moves the unmodeled part (e.g., the nonlinear uncertainties
caused by s) and is chosen to satisfy the following constraints:
e(−ϕTe
β−∆2mL+β1Lqd)≤ε
eβ1Lqd≤0(19)
where εis a positive error bound that can be arbitrarily small,
ϕ= [−Lqhm, L]T. The proof of the existence of εcan be
found in the Appendix.
In our case, qdis given by
qd=−1
4εh2( ˙e+k1e)
h≥ ∥βmax −βmin∥2· ∥ϕ∥2+δ(t, x).
(20)
Parameter estimates will be adjusted by learning mecha-
nisms of discontinuous projection [26]. The adaptation law
for ˆ
βis
˙
ˆ
β=P roj ˆ
β(Γ1ϕe)˜
β=h˜
β1,˜
β2iT
P roj ˆ
β(•) = 




0 if ˆ
βi=βimax and •>0
0 if ˆ
βi=βimin and •<0
•otherwise
(21)
where Γ1>0is a positive definite gain matrix.
For the stability analysis of the controller, the Lyapunov
function is considered
V=1
2e2
˙
V=e(β1Lx2−β2L−∆2mL+ ˙x1d)
=−k1β1L
β1max
e2+e(β1Lqd−∆2mL) + eβ1(−˙x1d−ˆ
β2L
ˆ
β1
−˙x1d−β2L
ˆ
β1
+˙x1d−β2L
ˆ
β1
+˙x1d−β2L
β1
)
=−k1β1L
β1max
e2+e(−ϕT˜
β−∆2mL+β1Lqd)
≤ −k1β1L
β1max
e2+ε≤ −2k1LV +ε.
(22)
So we have
|e|2≤exp(−2k1Lt)|e(0)|2+ε
k1L[1 −exp(−2k1Lt)] .(23)
Thus, the tracking error decays exponentially and is
bounded within a fixed region. The size of the error |e(∞)| ≤
qε
k1Lcan be adjusted by selecting an arbitrarily small value
for ε.
2) Low-level motion tracking controller: The synthesized
control reference x2dis exactly the desired brace cuff’s
position qed, as defined in Eq. (14). The task of the low-
level controller is to track the desired position by controlling
the pulley position and maintain the position tracking error
e2=qed −qeconverging to zero. To achieve better dynamic
performance, field-oriented control (FOC) [27] was adopted in
our low-level controller. Fig. 8 illustrates the implementation.
The controller includes three control loops: position, speed,
and current loops. A PI controller is adopted in each loop.
A command motor position can be computed based on the
synthesized position qed, with the transmission ratio of both
the reducer and cable transmission taken into account. The
control law for each PI controller is given as
uj=KP j esj +KIj Zesj dt (j= 1,2,3).(24)
where KP j and KI j are P and I gains and esj is the error.
D. Rope Release Control Mode (RRCM)
In this mode, the pulley rotation is controlled to feed out
ropes, so clearance is given for the human to swing his leg.
When the controller is switched to the RRCM, it records the
position of the motor rotor and generates a new position related
to the angular velocity of the human leg:
qed new =qed +|ω|(25)
where ωis a scaled factor related to the angular velocity
measured from the IMU. qed new is the newly updated motor
position and qed is the old position.
This approach prevents the excessive release of the rope at
different walking speeds, and reduces the possible slack during
control mode switching.
IV. EXP ER IM EN TS A ND RE SU LTS
A prototype of the proposed parallel hip exoskeleton was
built, along with electronics and controller. The total weight
of the exoskeleton is 3.62 kg, including the battery and
actuation units. Fig. 9 shows the architecture of the human-
machine system control. Each leg was equipped with a PMSM
motor driver (BJTU PMSM driver) to drive a 100-W servo
motor (ASM100-36-1250, XinYuan Electronic Technology
Ltd., Haikou, China) that was connected to a planetary re-
ducer (NPF60, NanRui Automation Technology Ltd., Nanjing,
China) and to provide feedback from the encoder (embedded
in the motor, 1250 ppr) and motor current data to the digital
signal processor (DSP, TMS320F28335, TI Inc., USA). The
reduction ratio of the reducer is 10:1, and the ratio between the
pulley’s radium and the force arm L is approximately 9.1:1.
Therefore, the total transmission ratio is 91:1. As the torque
capacity of the PMSM motor is 0.38 Nm, the output torque
can reach 34.62 Nm.
To measure the human-machine interaction force, a single-
axis load cell (ZNLBM-10 kg CHINO SENSOR Ltd., Anhui,
China) was installed on the inner side of each brace cuff.
The signal of the load cell was filtered and amplified by an
amplifier and processed by a DSP chip. Therefore, the integral
of the acquired forces can be computed in real time as the
feedback of the control system.
Two IMUs (HWT905-CAN, Wit-motion Ltd, Shenzhen,
China) were used to measure the key gait event of both legs.
Each IMU incorporates a triaxial accelerometer, gyroscope,
and magnetometer sensor. Kinematic information (Euler an-
gles, angular velocities, accelerations) for all three axes can
be read at a sampling rate of 200 Hz and filtered with a built-
in sensor fusion algorithm based on an extended Kalman filter
with a cutoff frequency of 5 Hz. The controller area network
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 7
Fig. 8. Schematic of the low-level control.
(CAN) communication protocol is used for communication
between the IMUs, DSP chip, and PC. The DSP reads the
angle information of both legs and determines the control
mode for each leg independently. In our case, the high-level
controller generates the desired exoskeleton position qed at 1
kHz, while performing low-level motion tracking at 10 kHz.
A. Experimental Protocol
Walking trials were tested on four healthy subjects (all
male, Means & SD: age: 24 ±5years, mass: 76.9±8.2kg,
height: 178.5±5.6cm). All procedures in this study were
carried out in accordance with the protocol approved by the
Local Ethics Committee of Beijing Jiaotong University, and
all human subjects provided informed written consent prior
to participation. Two experimental trials were conducted. Ex-
periment I was performed to evaluate the human-exoskeleton
kinematic compatibility and the generated parasitic force when
the relative sliding between the brace cuff and human thigh is
restricted. Experiment II assesses the performance of assisted
walking.
During Experiment I, the subjects were first asked to per-
form required motions to test the available DoFs provided by
the exoskeleton. Then the subjects were required to stand up
straight and then swing his right leg back and forth, when the
brace cuff was restrained with a fixing device. The device has
a plate holding a force sensor, and is fixated on a thigh point
with cables that limit the downward sliding. The force sensor
was used for the recording of the generated parasitic force as
a result of restricting the relative sliding DoF.
During Experiment II, the subjects were asked to walk
on a treadmill at three speed levels (3, 4, and 5 km/h), as
shown in Fig. 10. For each speed level, the walking duration
was 1 min, and the subjects were required to have a five-
minute rest between adjacent sets of gait tests. To evaluate
the exoskeleton’s biological contribution, electromyography
(EMG) was recorded using NeXus-10 MKII hardware with
electrodes (Covidien H124SG, Ag/AgCl, 24 mm diameter)
placed over the belly of the biceps femoris, in line with the
muscle fiber direction. A ground electrode was placed on
Fig. 9. Architecture of the hip exoskeleton control.
Fig. 10. Experimental setup of the treadmill walking tests. 1. Motion capture
system; 2. treadmill; 3. host computer; 4. exoskeleton.
the anterior tibialis. Before the electrodes were applied, the
skin was shaved and cleaned with single-use 70% isopropyl
alcohol wipes in accordance with the SENIAM protocol [28].
Assistance effect can be observed, as indicated from the
EMG measurement in different conditions, i.e., no exoskeleton
(‘NE’), assisted walking (‘AW’), and walking without any
assistance (but with the exoskeleton on, ‘WA’).
B. Experimental Results
Experiment I: Fig. 11 shows the available motion patterns
that can be achieved by the wearer with the exoskeleton.
It is noted that some larger-scale actions, including long
stride and squat, can be achieved in addition to the basic
HFE and HAA. Internal/external rotation was also performed
without any mechanical restriction imposed, and the pilots can
freely move their legs to 45◦for both directions. No manual
alignment procedures are required before the experiment.
When the relative sliding was restricted, the kinematic
redundancy was eliminated and hence parasitic force was
observed with leg movements. The parasitic force from the
last 20 gait cycles was recorded for post-processing, as shown
in Fig. 12. It shows the parasitic force increases with the thigh
angle, and reaches approximately 50N when the leg is at MHF.
Due to the deformation of the rods the parasitic force is not
very large and therefore endurable for the wearers.
Experiment II: During assisted walking, the controller
switches between the two control modes. As the CFCM
occupies almost 2/3of a unit gait cycle, and only during which
period the learning mechanism exist, the unknown parameters
were therefore not continuously learned due to the control
mode switching. When the controller is switched to RRCM
the unknown parameters are always set to 0 or the final value
before switching. When we talk about the learning process,
only the period of CFCM is discussed.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 8
Fig. 11. Kinematic compatibility test. (a) Long stride; (b) squat; (c) HFE;
(d) HAA.
Fig. 12. Measured parasitic force when the brace cuff is fixated on the thigh
during HFE motion. 1. Force sensor; 2. fixing device; 3. IMU.
The experimental results of parameter adaptive learning
from a representative trial are shown in Fig. 13. The results
showed good performance in parameter adaptive learning dur-
ing CFCM. Both the unknown parameters β1and β2updated
their values fast with gaits. At the beginning of each gait cycle,
a new learning session was launched.
It can be seen from (13) that the parameter β1is actually
the human-exoskeleton stiffness k. The result shows that k
is between 2.44 and 2.52N/mm. We used a force sensor and
motion capture system to measure the stiffness at the human-
machine interface based on Hook’s law. The test reveals that
the real stiffness is approximately 2.7N/mm, which is pretty
close to the learned stiffness. In addition, as the term kqhis
the main slow-varying part included in β2, the thigh angle qh
can also be estimated with the learned kas input. It can be
seen that the magnitude of the estimated thigh angle follows
the measured angle data well, with a slight time lag.
As those unknown parameters were adaptively learned in
Fig. 13. Estimation of the unknown parameters during assisted walking.
The measured thigh angle was from the IMU, ˆ
β1and ˆ
β2are estimated
parameters based on the learning mechanism of the controller. The red dashed
line indicates the beginning of each gait cycle.
Fig. 14. Assistance delivered in three consecutive gait cycles. (a) 3 km/h;
(b) 4 km/h; (c) 5 km/h.
Fig. 15. (a) Representative results of EMG measurement in three consecutive
gait cycles of a subject at 3 km/h. (b) Peak muscle activities across the subjects
at 3 km/h (N = 4).
real time, the force tracking shows good performance. Fig.
14 shows the force tracking results from a representative trial
during assisted walking. When the controller was switched to
the CFCM from the RRCM, the rope was dragged immediately
by the pulley, resulting in a rapid increase in the human-
machine interaction force Fhm. Compared with the force
tracking results in our previous work, the results in this work
show that the delivered force can follow the designed force
control reference, and a quick response as well. The result of
tracking error is given in Table III.
The assistance effect can be observed, as indicated by
EMG measurements. The results are presented in Fig. 15
and Table III. The measured muscle activities of all subjects
were normalized to the maximal voluntary contraction (MVC).
It can be seen that the biceps femoris muscle activity was
significantly lower during AW when the controller worked
in CFCM mode, but not for the period of RRCM. Among
the peak muscle activities (Max %MVC), it is noted that the
reduction was 7.8 %MVC for AW compared with NE. For WA,
the peak value was slightly higher than that of NE. This can be
partly attributed to the introduced weight of the exoskeleton
system, which can cause penalty.
V. DISCUSSION
Misalignment is a problem that exists in most exoskeletons
with anthropomorphic structures, including both lower- and
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 9
TABLE III
COL LEC TE D DATA ACROS S THE S UB JEC TS I N THR EE T RIA LS (N=4)
3km/h 4km/h 5km/h
Variation of the Hip angle
(Mean±SD) 49.1◦±4.7◦58.0◦±7.3◦71.3◦±7.9◦
Max. tracking error e1.81Ns 2.72Ns 5.15Ns
Reduction of muscle
activity (%MVC,
Mean±SD)
3.42±1.51 2.83±1.25 2.57±1.44
upper-limb exoskeletons [14], [29]. In contrast, overstress due
to parasitic forces is a problem in current soft exoskeletons.
The cable-driven parallel hip exoskeleton provides an alter-
native to existing anthropomorphic and soft exoskeletons to
address these problems.
In Experiment I, all the hip movements can be achieved
by the subjects with the exoskeleton, which shows good
human-exoskeleton kinematic compatibility. Most existing hip
exoskeletons don’t provide the internal/external rotational DoF
as their brace cuffs are usually fixed to the thigh. The unique
unbound design in this work allows this DoF, therefore the
versatile gaits, to be achieved. Given the deformation of the
rods, restriction of the sliding DoF cause endurable parasitic
forces. However, this could lead to injury for long-term use.
In this regard, the proposed design can obviously improve the
wearing comfort with eliminated parasitic force.
In Experiment II, it is noted that when the walking speed
increases, the tracking error increases correspondingly. At 5
km/h, the maximum tracking error is 5.15 N.s, which appears
approximately between 24% and 54% gait cycle. During this
period, the cable needs to be dragged immediately to provide
the assistance. However, due to the delay of the mechanical
transmission, it takes time for the pulley to drag into excessive
cable before cable tension generates. This is the main cause
that leads to relatively larger errors with increased walking
speeds. It is worth noting that there is a spike for each trial,
as shown in the square area of Fig. 14. At that point, the
controller is in RRCM, and no external force is applied. The
unexpected increase in Fhm is caused by the flexion motion
of human leg that pushes against the brace cuff, generating
contact pressure on the sensor.
Moreover, unexpected disturbances such as gait dynamics,
would also cause measurement errors during human walking.
Noise and oscillatory waves were observed in the measured
force, and the controller could still work stably and suppress
transient oscillations without generating a rapidly varying
control reference of the motor position. This can be attributed
to the benefit of using the force integral rather than force as
the control goals, which is more robust to the measurement.
By delivering the desired assistance, a reduction trend in the
EMG activity was observed.
Some limitations of this work are noted. The mechanical
delay needs to be further reduced to lower the force tracking
error and achieve better dynamic performance. Moreover, the
diameter of the brace cuff should be adjustable to avoid
it is neither too big to become unfettered from adhesion,
nor too small so friction would cause uncomfort. Further
improvements and iterations of the design is needed in our
future work.
VI. CONCLUSION
This paper proposes a novel cable-driven parallel hip ex-
oskeleton and its implementation on human body. In this
design, a cable-driven mechanism is combined with a parallel
exoskeleton to provide walking assistance for able-bodied
individuals. The proposed exoskeleton can be an alternative
to existing hip exoskeletons owning to its benefits in three
aspects. Firstly, the exoskeleton has virtual rotation center
that is closely distributed to the biological hip. The quantifi-
cation experiment shows that the misalignment between the
mechanical hip joint and biological hip can be reduced to
1.24 cm for normal walking. Secondly, hip internal/external
passive rotation is allowed for walking with versatile gaits.
Experiment I shows that the subjects can move their legs to 45◦
in both directions, which is the normal ROM required for daily
locomotion. Thirdly, the exoskeleton can provide pure assistive
torque with eliminated parasitic force during assisted walking.
In Experiment I, a maximum of 50 N parasitic force was found
if the relative sliding between the thigh and cuff is restricted,
which on the flip side, this showing that the proposed design
can eliminate the parasitic force completely for improved
comfort. All these advances make high-performance assisted
walking can be achieved with good human-machine kinematic
compatibility, which is a major contribution of this work.
Another contribution of this study pertains to the control
design. Experiment II shows that the controller enables stable
and promising assistance to be delivered at different speed
levels in the presence of gait dynamics. A maximum of 7.8
%MVC is reduced for the biceps femoris muscle activity,
which shows the effectiveness of the proposed exoskeleton
system for assistive walking.
APPENDIX
Proof of the existence of error bound ε:
For the left-hand side of (19), it satisfies:
e(−ϕTe
β−∆2mL+β1Lqd)≤eLqhm e
β1−Le
β2+
eδ(t, x) + eβ1Lqd
As qdis chosen to satisfy eβ1Lqd≤0in (19), then
e(−ϕTe
β−∆2mL+β1Lqd)≤eL(qhm e
β1−e
β2) + eδ(t, x)
In (18), qhm is bounded because the ranges of β1and β2
are fixed. Assuming λis the maximum value of qhm, we can
obtain
eL(qhm e
β1−e
β2) + eδ(t, x)
≤e(Lλ |β1max −β1min|+L|β2max −β2min |+δ(t, x) )
To simplify the notation, let η=Lλ |β1max −β1min|+
L|β2max −β2min|+δ(t, x). When t= 0, there is always a
design parameter εthat is large enough to satisfy e(0)η≤ε
because e(0) is a real number. Consequently, the tracking error
edecays exponentially as (23) is satisfied. Then, a smaller
value of εcan be chosen to satisfy the inequality e(t)η≤ε.
Ultimately, the final tracking error satisfies |e(∞)| ≤ qε
k1L.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 10
REFERENCES
[1] S. Bai, G. S. Virk, and T. G. Sugar, Wearable exoskeleton systems:
design, control and applications. UK: Institution of Engineering and
Technology, Jan. 2018.
[2] X. Wu, Y. Yuan, X. Zhang, C. Wang, T. Xu, and D. Tao, “Gait phase
classification for a lower limb exoskeleton system based on a graph
convolutional network model,” IEEE Trans. Ind. Electron., vol. 69, no. 5,
pp. 4999–5008, May. 2021.
[3] M. B. N¨
af, K. Junius, M. Rossini, C. Rodriguez-Guerrero, B. Vander-
borght, and D. Lefeber, “Misalignment compensation for full human-
exoskeleton kinematic compatibility: state of the art and evaluation,”
Appl. Mech. Rev., vol. 70, no. 5, Feb. 2019, Art. no. 050802.
[4] Y. Ding, I. Galiana, A. T. Asbeck, S. M. M. De Rossi, J. Bae,
T. R. T. Santos, V. L. De Araujo, S. Lee, K. G. Holt, and C. Walsh,
“Biomechanical and physiological evaluation of multi-joint assistance
with soft exosuits,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 25,
no. 2, pp. 119–130, Jan. 2016.
[5] K.-W. Park, J. Choi, and K. Kong, “Hybrid filtered disturbance observer
for precise motion generation of a powered exoskeleton,” IEEE Trans.
Ind. Electron., vol. 70, no. 1, pp. 646–656, Jan. 2023.
[6] A. H. Stienen, E. E. Hekman, G. B. Prange, M. J. Jannink, A. M.
Aalsma, F. C. van der Helm, and H. van der Kooij, “Dampace: Design
of an exoskeleton for force-coordination training in upper-extremity
rehabilitation,” J. Med. Devices, vol. 3, no. 3, Sep. 2009, Art. no. 031003.
[7] M. Sposito, C. Di Natali, S. Toxiri, D. G. Caldwell, E. De Momi,
and J. Ortiz, “Exoskeleton kinematic design robustness: An assessment
method to account for human variability,” Wearable Technol., vol. 1,
Nov. 2020, Art. no. e7.
[8] X. Wang, S. Guo, B. Qu, and S. Bai, “Design and experimental
verification of a hip exoskeleton based on human-machine dynamics
for walking assistance,” IEEE Trans. Hum.-Mach. Syst., vol. 53, no. 1,
pp. 85–97, Feb. 2023.
[9] N. Naghavi, A. Akbarzadeh, S. M. Tahamipour-Z, and I. Kardan,
“Assist-As-Needed control of a hip exoskeleton based on a novel
strength index,” Rob. Auton. Syst., vol. 134, Dec. 2020, Art. no. 103667.
[10] T. Xue, Z. Wang, T. Zhang, and M. Zhang, “Adaptive oscillator-based
robust control for flexible hip assistive exoskeleton,” IEEE Rob. Autom.
Lett., vol. 4, no. 4, pp. 3318–3323, Jul. 2019.
[11] G. Aguirre-Ollinger and H. Yu, “Lower-limb exoskeleton with variable-
structure series elastic actuators: phase-synchronized force control for
gait asymmetry correction,” IEEE Trans. Rob., vol. 37, no. 3, pp. 763–
779, Jun. 2020.
[12] B. Lim, J. Lee, J. Jang, K. Kim, Y. J. Park, K. Seo, and Y. Shim,
“Delayed output feedback control for gait assistance with a robotic hip
exoskeleton,” IEEE Trans. Rob., vol. 35, no. 4, pp. 1055–1062, Aug.
2019.
[13] L. Zhou, W. Chen, S. Bai, J. Wang, Z. Zhao, X. Zhao, and X. Yu,
“Lower limb exoskeleton parasitic force modeling and minimizing with
an adaptive trajectory controller,” Mech. Mach. Theory, vol. 170, Apr.
2022, Art. no. 104731.
[14] J. Bessler-Etten, L. Schaake, G. B. Prange-Lasonder, and J. H. Buurke,
“Assessing effects of exoskeleton misalignment on knee joint load during
swing using an instrumented leg simulator,” J. NeuroEng. Rehabil.,
vol. 19, no. 1, pp. 1–18, 2022.
[15] J. Beil and T. Asfour, “New mechanism for a 3 DOF exoskeleton hip
joint with five revolute and two prismatic joints,” in 2016 6th IEEE
International Conference on Biomedical Robotics and Biomechatronics
(BioRob), pp. 787–792, Jun. 2016.
[16] C. J. Walsh, K. Endo, and H. Herr, “A quasi-passive leg exoskeleton for
load-carrying augmentation,” Int. J. Humanoid Rob., vol. 4, no. 03, pp.
487–506, 2007.
[17] R. W. Nuckols, S. Lee, K. Swaminathan, D. Orzel, R. D. Howe, and
C. J. Walsh, “Individualization of exosuit assistance based on measured
muscle dynamics during versatile walking,” Sci. Rob., vol. 6, no. 60,
Nov. 2021, Art. no. eabj1362.
[18] Y. Ding, I. Galiana, C. Siviy, F. A. Panizzolo, and C. Walsh, “IMU-
based iterative control for hip extension assistance with a soft exosuit,”
in 2016 IEEE International Conference on Robotics and Automation
(ICRA), pp. 3501–3508, May. 2016.
[19] W. Cao, C. Chen, H. Hu, K. Fang, and X. Wu, “Effect of hip assistance
modes on metabolic cost of walking with a soft exoskeleton,” IEEE
Trans. Autom. Sci. Eng., vol. 18, no. 2, pp. 426–436, Oct. 2020.
[20] Q. Chen, S. Guo, L. Sun, Q. Liu, and S. Jin, “Inertial measurement
unit-based optimization control of a soft exosuit for hip extension and
flexion assistance,” J. Mech. Rob., vol. 13, no. 2, Apr. 2021, Art. no.
021016.
[21] X. Wang, S. Guo, M. Song, and P. Wang, “Mechanical design and
experimental verification of a parallel hip exoskeleton with virtual ro-
tation center,” in 2021 6th IEEE International Conference on Advanced
Robotics and Mechatronics (ICARM), pp. 230–235, Jul. 2021.
[22] X. Wang, S. Guo, B. Qu, M. Song, P. Wang, and D.-X. Liu, “Design
and experimental verification of a parallel hip exoskeleton system for
full-gait-cycle rehabilitation,” J. Mech. Rob., vol. 14, no. 5, Oct. 2022,
Art. no. 054504.
[23] C. Kirtley, CGA Normative Gait Database. Hong Kong Polytechnic
University, 2016. [Online]. Available: http://guardian.curtin.edu.au/cga/
data/
[24] S. Chen, Z. Chen, B. Yao, X. Zhu, S. Zhu, Q. Wang, and Y. Song,
“Adaptive robust cascade force control of 1-DOF hydraulic exoskeleton
for human performance augmentation,” IEEE/ASME Trans. Mechatron.,
vol. 22, no. 2, pp. 589–600, Oct. 2016.
[25] G. Lee, J. Kim, F. Panizzolo, Y. Zhou, L. Baker, I. Galiana, P. Malcolm,
and C. Walsh, “Reducing the metabolic cost of running with a tethered
soft exosuit,” Science Robotics, vol. 2, no. 6, May. 2017, Art. no.
eaan6708.
[26] L. Xu and B. Yao, “Adaptive robust precision motion control of linear
motors with negligible electrical dynamics: theory and experiments,”
IEEE/ASME Trans. Mechatron., vol. 6, no. 4, pp. 444–452, Dec. 2001.
[27] T.-H. Liu, J.-R. Fu, and T. A. Lipo, “A strategy for improving reliability
of field-oriented controlled induction motor drives,” IEEE Trans. Ind.
Appl., vol. 29, no. 5, pp. 910–918, Oct. 1993.
[28] H. J. Hermens, B. Freriks, C. Disselhorst-Klug, and G. Rau, “Devel-
opment of recommendations for SEMG sensors and sensor placement
procedures,” J. Electromyogr. Kines., vol. 10, no. 5, pp. 361–374, Sep.
2000.
[29] S. Bai, M. Islam, V. Power, and L. O´
Sullivan, “User-centered develop-
ment and performance assessment of a modular full-body exoskeleton
(axo-suit),” Biomimetic Intell. and Robot., vol. 2, no. 2, Jun. 2022, Art.
no. 100032.
Xiangyang Wang (S’21) received his B.S. and
Ph.D. degrees in mechanical engineering from Bei-
jing Jiaotong University, Beijing, China, in 2017
and 2023, respectively. From 2021 to 2023, He was
a visiting researcher with the Aalborg University,
Aalborg, Denmark.
He is currently a Postdoctoral Fellow with the
Shenzhen Institute of Advanced Technology, Chi-
nese Academy of Sciences. His research interests in-
clude wearable robotics and rehabilitation robotics.
Dr. Wang was the recipient of the Best Student
Paper from the IEEE International Conference on Robotics and Biomimetics
(IEEE ROBIO), in 2022.
Sheng Guo received his Ph.D. degree from Beijing
Jiaotong University in 2005. He was then a post-
doctoral fellow at National Cheng Kung University
in 2005-2006. He was a Visiting Scholar at the
University of California, Irvine, US, in 2010-2011.
Currently, He is a Full Professor, Vice Director of
Robotics Institute, and Dean of School of Mechan-
ical, Electronic and Control Engineering, at Beijing
Jiaotong University. His research interests include
robotics mechanism and mechatronics.
Shaoping Bai (M’01-SM’18) is currently a full
professor with the Department of Materials and
Production, Aalborg University, Aalborg, Denmark.
His research interests include exoskeletons, wearable
sensors, parallel manipulators, and walking robots.
He is a member of ASME, a senior member of
IEEE ROBOTICS AND AUTOMATION and deputy
chair of IFToMM Technical Committee on Mecha-
tronics. He serves as an Associate Editor for both
the ASME Journal Mechanisms and Robotics and
IEEE Robotics and Automation Letters.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3270494
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Shenzhen Institute of Advanced Technology CAS. Downloaded on May 04,2023 at 00:48:44 UTC from IEEE Xplore. Restrictions apply.