Control Engineering Practice 80 (2018) 49–60

Contents lists available at ScienceDirect

Control Engineering Practice

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

Robust Active Disturbance Rejection Control via Control Lyapunov

Functions: Application to Actuated-Ankle–Foot-Orthosis
J.F. Guerrero-Castellanos a, *, H. Rifaï b , V. Arnez-Paniagua b , J. Linares-Flores c,1 ,
L. Saynes-Torres a , S. Mohammed b
a
Facultad de Ciencias de la Electrónica, Benemérita Universidad Autónoma de Puebla (BUAP), Ciudad Universitaria, C. P. 72570, Puebla, Mexico
b Laboratoire Images, Signaux et Systèmes Intelligents (LISSI), Université Paris-Est Créteil (UPEC), 94400 Vitry-sur-Seine, France
c Universidad Tecnológica de la Mixteca, Oaxaca 69000, Mexico

ARTICLE INFO ABSTRACT

Keywords:
Control Lyapunov Function (CLF)
Actuated-Ankle–Foot-Orthosis (AAFO)
Active Disturbance Rejection Control (ADRC)
Extended State Observer (ESO)
Input-to-State Stability (ISS)
In this paper, the problem of trajectory tracking, for an Actuated-Ankle–Foot-Orthosis (AAFO) to assist the gait
of paretic patients, is addressed. The control strategy is based on the system’s flatness property, which allows the
development of an Active Disturbance Rejection Control (ADRC). For this purpose, an Extended State Observer
(ESO) is designed to estimate on-line the unknown disturbances and canceled by injecting the output of ESO
into the feedback loop. A stability analysis of the estimation error dynamics is carried out in the Input-to State
Stability (ISS) framework, stating the observer’s robustness. On the other hand, the feedback design is based on
the existence of a Control Lyapunov Function (CLF) and the Sontag’s formula. The stability analysis discloses
that the tracking trajectory controller is ISS, i.e. robustly stable. Finally, the effectiveness of the ADRC strategy
is validated by performing real-time tests with a healthy subject walking on a treadmill at self-selected speed.
The experimental results validate the theoretical analysis.

1. Introduction
1.1. Motivations and background
Neurological injuries such as Spinal Cord Injuries (SCI) and stroke,
results frequently in Foot-Drop symptom (Stewart, 2008). Indeed, every
year an augmenting number of people are diagnosed with disabilities
that prevent them from performing daily living activities such as
walking, stairs ascent/descent, standing up, etc. According to the World
Health Organization, around 15% of the world’s population lives with
some form of disability; among them 2%–4% are facing important
difficulties while performing daily activities. Patients suffering from gait
pathologies may either have total loss of ankle muscle forces and are
unable to initiate a movement with the affected limbs or they may have
partial loss of muscle forces and are able to move their limbs within lim-
ited ranges. Furthermore, spasticity can occur during a gait cycle due to
the involuntary contraction reflex of antagonist muscles spanning at the
ankle joint. The most common gait pathologies at the ankle joint is foot-
drop, foot slap, and insufficient push-off power. Foot-drop patients are
unable to lift their feet and toes properly during walking, affecting thus
their movement coordination, their balance and increasing the risk of
falling. Treatments for this pathology range from conventional therapy,
use of passive orthoses, functional electrical stimulation (FES) (Coste,
Jovic, Pissard-Gibollet, & Froger, 2014; Peckham & Knutson, 2005), to
the use of wearable robotics based solutions, known also as actuated
ankle–foot orthosis (AAFO).
Conventional therapy include lower limb muscle strengthening ex-
ercises, joint stretching to enhance ankle dorsiflexion and plantar-
flexion, and ground walking with the assistance of therapists. It is worth
noting, however, that this therapy is difficult and effort demanding
to be performed continuously for more than few minutes by both
therapists and patients (Jamwal, 2011; Krishnamoorthy, Hsu, Kesar,
Benoit, Banala, Perumal, Sangwan, Binder-Macleod, Agrawal, & Scholz,
2008; Nicholas Romansky, Kelly Scollon-Grieve, & McGinness, 2012).
Therefore, the inclusion of robotic devices such as AAFOs could poten-
tially increase the dosage and intensity of the therapy while reducing
the effort required from the therapists. In robotic-assisted therapy,
some examples of AAFOs used to prevent foot-drop provide assistance
torque in the dorsiflexion direction (Roy, Krebs, Iqbal, Macko, Macko,

* Corresponding author.
E-mail address: fermi.guerrero@correo.buap.mx (J.F. Guerrero-Castellanos).
URL: http://www.fguerrero.ece.buap.mx (J.F. Guerrero-Castellanos).
1
On sabbatical leave at Benemérita Universidad Autónoma de Puebla (BUAP).

https://doi.org/10.1016/j.conengprac.2018.08.008
Received 2 March 2018; Received in revised form 9 June 2018; Accepted 14 August 2018
0967-0661/© 2018 Published by Elsevier Ltd.
J.F. Guerrero-Castellanos et al.
& Forrester, 2014). However, another advantage of actuated orthosis
compared to passive ones is the capacity of the former group to provide
assistance in the plantar-flexion direction to promote a better foot
push-off power prior to the swing phase (Blaya & Herr, 2004; Park,
Chen, Young, Stirling, Wood, Goldfield, & Nagpal, 2011). There are
several strategies to determine the assistance level provided by the
AAFOs (Boehler, Hollander, Sugar, & Shin, 2008; Jimenez-Fabian &
Verlinden, 2011; Shorter, Xia, Hsiao-Wecksler, Durfee, & Kogler, 2013):
(1) by pre-selecting the assistance torque and applying it in a feedfor-
ward scheme with respect to the gait phase detected (Ab Patar, Said,
Mahmud, Majeed, & Razman, 2014; Arnez-Paniagua, Huo, Colorado-
Cervantes, Mohammed, & Amirat, 2016; Boehler et al., 2008; Jamwal,
2011; Shorter, Kogler, Loth, Durfee, & Hsiao-Wecksler, 2011), (2) as a
function of electromyography (EMG) signals (Ferris, Gordon, Sawicki,
& Peethambaran, 2006; Pérez-Ibarra & Siqueira, 2017), (3) by adapting
the stiffness, inertia or impedance of the coupled human-exoskeleton
system based on the gait phase detected, (Blaya & Herr, 2004; Lawn,
Takashima, Ninomiya, Yu, Soma, & Ishimatsu, 2015; Roy et al., 2014;
el Zahraa Wehbi, Huo, Amirat, El Rafei, Khalil, & Mohammed, 2017),
or (4) as a function of the tracking error between the current ankle
joint angle and a reference trajectory pattern commonly generated from
healthy subject walking profiles (Arnez-Paniagua, Rifaï, Amirat, &
Mohammed, 2017a; Bharadwaj, Sugar, Koeneman, & Koeneman, 2005;
Brahmi, Saad, Ochoa-Luna, & Rahman, 2017; Hitt, Oymagil, Sugar,
Hollander, Boehler, & Fleeger, 2007; Holgate, Bohler, & Suga, 2008;
Jamwal, Xie, Hussain, & Parsons, 2014; Madani, Daachi, & Djouani,
2014; Rifaï, Mohammed, Hassani, & Amirat, 2013; Veneva & Ferreira,
2014; Ward, Sugar, Standeven, & Engsberg, 2010; Wolbrecht, Chan, Le,
Cramer, Reinkensmeyer, & Bobrow, 2007; Zhang, Cao, Xie, Zhu, Zeng,
Huang, & Xu, 2017). The control method used for each exoskeleton
has a direct impact on the level and rate of human adaptation to
the active device; i.e., the nervous system can adapt more easily to
a continuous, smooth and proportionate stimulus (Cain, Gordon, &
Ferris, 2007). In this sense, it is clear that there exist an intimal relation
between AAFO and the control strategies design based on rehabilitation
objectives (Jiménez-Fabián & Verlinden, 2012). Feedforward strategies
are simple to implement but the lack of feedback could produce an
insufficient or excessive assistance. EMG based strategies provide a high
rate of adaptation for the nervous system, but requires a minimum
residual muscular activities in order to be effective, which might not
be the case for some patients with acute stroke symptoms. The systems
that adapt their impedance have the advantage of not requiring a
predefined trajectory but they require a minimum residual voluntary
effort to initiate movements. Therefore, a trajectory tracking strategy
could potentially provide more appropriate assistance when the wearer
of the active orthosis is not fully able to initiate movement.
The present paper is more in the spirit of the fourth group. Since
AAFO is modeled as a nonlinear system subject to disturbances from the
external environment and uncertainties such as unmodeled dynamics,
parameter perturbations, and nonlinear couplings, the control problem
for gait tracking to achieve ankle rehabilitation is in itself a chal-
lenge (Gregg, Bretl, & Spong, 2010). Proportional control (also called
High-Gain Control) and PID control results in an attractive solution
due to its simplicity to real-time implementation (Aström & Murray,
2008) and it has been applied to AAFO, for instance in Boiadjiev and
Veneva (2009) and Bai, Gao, Zhao, Jin, Dai, and Lv (2015). However,
it is well known that proportional control cannot completely remove
the effects caused by disturbance (even a constant one) and a higher
control gain has to be designed to diminish the disturbance effects. On
the one hand, when PID is used, the integral term could effectively
reject the constant disturbance but achieves poor performance in the
presence of time-varying disturbances. On the other hand, PID is often
implemented without the D part because of the noise sensitivity (Han,
2009). Adaptive Control (AC) is quite effective in dealing with model
parameter uncertainties and has gained wide applications in practical
engineering. The successful applications of AC methods usually depend
50
Control Engineering Practice 80 (2018) 49–60
highly on the design of identification or estimation laws on time-varying
model parameters (Krstic, Kanellakopoulos, & Kokotovic, 1995) and
has been recently applied to AAFO in Ibarra, dos Santos, Krebs, and
Siqueira (2014) and Arnez-Paniagua, Rifai, Mohammed, and Amirat
(2017b). However, when external disturbances are present, e.g. mus-
cular torque due to gait pathologies or the impact with the ground, AC
is limited since the error produced by the disturbance could negatively
affect the adaptive performance and key parameters are difficult to
identify or estimate online (Arnez-Paniagua et al., 2017b). Sliding
Model Control (SMC) has fine abilities in suppressing the effects of
parameter perturbations as well as external disturbances (Spurgeon,
2014). However, it is well known that the discontinuous switching of the
controller is prone to induce high-frequency chattering of mechanical
systems which would be unacceptable for the AAFO. The employment
of some modification methods could effectively reduce the chattering
problem (Bartolini, Pisano, Punta, & Usai, 2003), as has been shown
recently in Mohammed, Huo, Huang, Rifa, and Amirat (2016) for a
knee joint orthosis. In robotics field, impedance control (IC) has been
largely used for manipulator robots (Hogan, 1985). IC establishes a
relationship between the force, the velocity and the environment. In the
last years, remarkable works have been proposed to deal with simple,
safe and robust motion/force tracking controllers (Mehdi & Boubaker,
2012a, b) and applied to rehabilitation robotic systems with parameters
uncertainties (Mehdi & Boubaker, 2016). Furthermore, when the
contact with the ground is taken into account, Jerk-Impedance (Aloulou
& Boubaker, 2015) and Jerk-Stiffness (Aloulou & Boubaker, 2016) con-
trollers have been designed for gait pattern generation and safe walking
with applicability to a breadth of rehabilitation robotics applications.
In the context of nonlinear control systems, since the seminal works
of Artstein and Sontag (Sontag, 1998), Control Lyapunov Functions, so
called CLFs, have become central to feedback design. A main reason is
that the existence of a CLF is necessary and sufficient for the stabiliz-
ability of a system with a control input. Domains of application include
robust nonlinear feedback design (Sepulchre, Jankovic, & Kokotović,
1997), receding horizon control of nonlinear systems (Primbs, Nevistic,
& Doyle, 1999), stabilization of hybrid systems (Sanfelice, 2013) and
stabilization of nonlinear system with event-based control (Marchand,
Durand, & Guerrero-Castellanos, 2013), to name only a few. In the
bipedal walking robot framework, CLF approach has been successful
used to exponentially stabilize periodic orbits of the hybrid zero dynam-
ics by shaping the energy (Ames, Galloway, Sreenath, & Grizzle, 2014;
Galloway, Sreenath, Ames, & Grizzle, 2015), where the control laws are
based on the Sontag’s formula which is well known to possess robustness
to static and dynamic input uncertainties (Jankovic, Sepulchre, &
Kokotović, 1999). Former properties represent a main motivation to use
CLFs in the present work.
The objective of the above-mentioned control approaches is to reject
disturbances via feedback, which is based on the tracking error between
the measured outputs and their setpoints or desired trajectories. As a
consequence, these controllers cannot react fast enough in the presence
of strong disturbances. In order to overcome this limitation, Active
Disturbance Rejection Control (ADRC) was introduced by Han (2009).
ADRC is fundamentally based on the possibility of on-line estimating
adverse effects so called ‘‘total disturbance’’ caused by the coupling
between unknown system dynamics (endogenous) and external (exoge-
nous) disturbances. This estimation is then canceled via an appropri-
ate feedback-feedforward control law (Sira-Ramírez, Luviano-Juárez,
Ramírez-Neria, & Zurita-Bustamante, 2017). The most remarkable fea-
ture of ADRC lies in its estimation/cancellation nature, where the total
disturbance is considered as an extended state and is estimated, in real-
time, through an Extended State Observer (ESO) (Chan, Naghdy, &
Stirling, 2013; Mehdi & Boubaker, 2015; Xue, Huang, & Gao, 2016) so
called disturbance-observer (Yu, Yang, Han, & Liu, 2018). ADRC has
been exploited in almost all domains of control engineering for example:
motion control of humanoid robots (Orozco-Soto & Ibarra-Zannatha,
2017), power filter design (Fuentes, Cortés-Romero, Zou, Costa-
Castelló, & Zhou, 2015), energy storage (Chang, Li, Zhang, Wang, &
J.F. Guerrero-Castellanos et al.
Xue, 2015), power plants (Huang, Wu, Han, Feng, & Lin, 2004), DC Con-
verters (Sira-Ramírez, Hernández-Méndez, Linares-Flores, & Luviano-
Juarez, 2016), electric machines and servomechanisms (Linares-Flores,
García-Rodríguez, Sira-Ramírez, & Ramírez-Cárdenas, 2015; Linares-
Flores, Sira-Ramírez, Barahona-Avalos, & Contreras-Ordaz, 2012; Sira-
Ramírez, Linares-Flores, García-Rodríguez, & Contreras-Ordaz, 2014),
renewable energy together with cooperative control (Hernandez-
Méndez, Linares-Flores, Sira-Ramírez, Guerrero-Castellanos, & Mino-
Aguilar, 2017). This list, is of course far from being exhaustive. Recently,
ADRC was applied to the area of biped locomotion (Martnez-Fonseca,
Castan̄eda, Uranga, Luviano-Jurez, & Chairez, 2016) and rehabilitation
systems (Long, Du, Cong, Wang, Zhang, & Dong, 2017; Yang, Yu, &
Zhang, 2017). The work presented in Long et al. (2017) is within
the framework of the presented paper. However, there are important
differences, for instance, the assistance is provided to the knee and hip
joints, not to the ankle joint. Furthermore, the controller uses a PD
control strategy instead of a CLF-based control as the one introduced
in the present paper. Finally, the experiments were conducted with a
healthy subject following a healthy trajectory pattern which is timed
fixed, meaning the subject has to synchronize to the trajectory before
starting the assistance.
In the above-mentioned works, the ADRC have been constructed
by combining ESO with other control methods, such as PID control,
optimal control, backstepping, sliding modes, adaptive control, predic-
tive control. Actually, the heart of this ADRC methodology lies the
need for flatness property (Sira-Ramírez & Agrawal, 2004) or of a
partially linearizable structure with stable zero dynamics (minimum
phase systems) (Sira-Ramírez et al., 2017). Flatness is natural in the
ADRC design and offers no obstacle to either observer-based design or
direct disturbance cancellation.
On the other hand, the notion of Input-to-State Stability (ISS) is
now recognized as a central concept in linear and nonlinear systems
analysis (Sontag, 1998, 2008). It provides a nonlinear generalization of
finite gains with respect to supremum norms and also of finite 𝐿2 gains.
ISS allows to quantify sensitivity to disturbances, and, more generally,
of the dependence of state trajectories on actuator and measurement
errors, magnitudes of tracking signals, and the like. It plays a central
role in recursive design, controllers for non-minimum phase systems,
observers, and many other areas.
1.2. Contributions
The aim of the present paper is to merge the philosophy of ADRC
design with CLFs and to analyze the stability and robustness of the
rejector, i.e Extended State Observer, and the closed-loop system, i.e
rejector-controller-plant, using the notion of ISS. Furthermore, the paper
shows the application of these theoretical concepts to an engineering
control problem of great societal importance and high public impact,
namely, active ankle foot orthoses (AAFO) which are used for correcting
gait following ankle dorsiflexion or plantar flexion deficits. This feature
is important to improve the benefits of the physical assisted therapy. The
ADRC strategy is based on the flatness property together with the design
of a robust ESO and the use of CLFs and a stabilizing control obtained
via the Sontag’s formula.
The preliminary theoretical aspects are introduced in Section 2. The
control law is based on the AAFO model, which is proven to be flat
and can be written as an affine in the control dynamical system, such as
detailed in Section 3. Furthermore, it is shown that the torques acting on
the AAFO system at the ankle level can be classified in endogenous dis-
turbances (which are dependent upon internal variables) and exogenous
disturbance (which are generated by the environment). The sum of these
unknown disturbances will be considered the ‘‘total disturbance’’. The
Proposition 4.1 represents the first contribution of this work. It presents
the design of an Extended State Observer (ESO) to estimate the ‘‘total
disturbance’’ as well as a redundant estimate of angular position and
velocity. Although, the ESO design is no novelty, a stability analysis
51
Control Engineering Practice 80 (2018) 49–60
in the ISS sense has never been reported in the ADRC context. This
analysis discloses that the estimation error behavior satisfies the ISO
property (input-to-output stability) (Sontag & Wang, 1995), i.e., which
is equivalent to say that it is robustly stable. Section 4.2 is devoted
to tracking control design for the AAFO and it contains the second
contribution of this work. Actually, it presents the existence of Control
Lyapunov Function (CLF) relative to the origin of the space–state of
the tracking error. This CLF is then used in Proposition 4.5 for the
design of a nonlinear control law based on the Sontag’s formula. This
control law incorporates a feedforward term, a feedback term and the
estimation of the ‘‘Total disturbance’’ provides by the ESO in order to
cancel it. The stability analysis disclaims that the tracking trajectory
controller is ISS, i.e. robustly stable, when the difference between the
unknown total disturbance and its estimation obtained by the ESO is
viewed as the input and the tracking error as the state. The Fig. 3 states
that the proposed ADRC, composed of the Extended State Observer
(ESO) and the trajectory-tracking controller is ISS, when the unknown
total disturbance is viewed as the input and the tracking error as the
state. This point, represent the main theoretical contribution, because
a system, which is ISS, exhibits low overshoot and low total energy re-
sponses when excited by uniformly bounded or energy-bounded signal,
respectively. These are highly desirable qualitative characteristics for
the AAFO. Finally, the effectiveness of the ADRC strategy is validated by
performing real-time tests with a healthy subject walking on a treadmill
at self-selected speed. The AAFO is controlled such that the ankle joint
angle tracks a predefined desired trajectory during the whole gait cycle.
This fact, represent the third and main experimental contribution of this
work.
2. Preliminaries
2.1. Flatness and feedback linearization revisited
A single input nonlinear system is said to be differentially flat if
there exists a differential function of the state, called flat output such
that all variables in the system (i.e. states, outputs, inputs, etc.) are,
in turn, expressible as differential functions of the flat output. It is a
property that trivializes the exact linearization problem in a nonlinear
system. Furthermore, flatness immediately yields the required open-
loop (nominal) behavior of the system for a particular desired trajectory
tracking.
The relation of single input flat systems with systems linearizable
by means of static state feedback is quite interesting since they are
equivalent.
Theorem 2.1 (Sira-Ramírez & Agrawal, 2004). A single input nonlinear
system of the form 𝑥̇ = 𝑓 (𝑥, 𝑢) with 𝑥 ∈ R𝑛 and 𝑢 ∈ R is differentially flat if
and only if it is feedback linearizable.
2.2. Input-to-state stability (ISS)
In the following, ‖⋅‖ denotes Euclidean norm for vectors and the
induced 2-norm for matrices, respectively. A scalar continuous function
𝛾(𝑟) defined for 𝑟 ∈ [0, 𝑎] is said to belong to class  if it is strictly
increasing and 𝛾(0) = 0, and it is said to belong to class ∞ if it is defined
for all 𝑟 ≥ 0 and 𝛾(𝑟) ⟶ ∞ as 𝑟 ⟶ ∞. A scalar continuous function
𝛽(𝑟, 𝑠) defined for 𝑟 ∈ [0, 𝑎], 𝑠 ∈ [0, ∞] is said to belong to class  if for
each fixed 𝑠 it belongs to class  and for each fixed 𝑟 it is decreasing in 𝑠
and 𝛽(𝑟, 𝑠) ⟶ 0 as 𝑠 ⟶ ∞. A dynamical system with state 𝑥 and input
𝑤 is called input-to-stable state (ISS) if there exist a class  function
‖ ‖
𝛽 and a class  function 𝛾 such that ‖𝑥(𝑡)‖ ≤ 𝛽(‖𝑥(0)‖ , 𝑡) + 𝛾(‖𝑤|0.𝑡| ‖ )
‖ ‖∞
‖ ‖
for all 𝑡 ≥ 0. For a signal 𝑤, ‖⋅‖∞ denotes the ∞ -norm: ‖𝑤|0.𝑠| ‖ =
‖ ‖∞
sup0≤𝑡≤𝑠 ‖𝑤(𝑡)‖. For linear systems, ISS is equivalent to global asymptotic
stability of the unforced system (Sontag & Wang, 1995).
J.F. Guerrero-Castellanos et al. Control Engineering Practice 80 (2018) 49–60

Definition 2.2 (Sontag & Wang, 1995). A smooth function 𝑉 ∶ R𝑛 ⟶

R≥0 is called an ISS-Lyapunov function for system 𝑥̇ = 𝑓 (𝑥, 𝑢) if there
exist ∞ -functions (𝛼1 , 𝛼2 ), and -functions 𝛼3 and 𝜒, such that

𝛼1 (|𝜉|) ≤ 𝑉 (𝜉) ≤ 𝛼2 (|𝜉|) (1)

for any 𝑥(0) = 𝜉 ∈ R𝑛 and

∇𝑉 (𝜉) ⋅ 𝑓 (𝜉, 𝜇) ≤ −𝛼3 (|𝜉|) (2)

for any 𝜉 ∈ R𝑛 and any 𝜇 ∈ R𝑚 so that |𝜉| ≥ 𝜒(|𝜇|).

Theorem 2.3 (Sontag & Wang, 1995). The following properties are
equivalent for any system:

1. It is ISS.
2. It admits an ISS-Lyapunov function.
3. It is robustly stable.
Fig. 1. On the left, a healthy subject wearing the system. On the right, the angles
determining the foot and the shank orientations. Foot frame  (⃗ 𝑥𝑓 , 𝑦⃗𝑓 , 𝑧⃗𝑓 ), where
2.3. Stability via control Lyapunov functions 𝑥⃗𝑓 is in the same plane as the heel-half of the insole. Ground frame (⃗ 𝑥𝑔 , 𝑦⃗𝑔 , 𝑧⃗𝑔 ).

Consider the affine system in the control dynamical systems defined

by:
In order to model the AAFO system, the frame  and  are used (see
Fig. 1(b)). Denote by 𝜃 the angle between the foot and the shank, by 𝜃𝑠
𝑥̇ = 𝑓 (𝑥) + 𝑔(𝑥)𝑢 (3)
the angle between the shank and the vertical axis and by 𝛼 the angle
where 𝑥 ∈  ⊂ R𝑛 , 𝑢 ∈  ⊂ R𝑝 , and 𝑓 and 𝑔 are Lipschitz functions between the foot and the horizontal axis. 𝜃 and 𝜃𝑠 can be measured by
vanishing at the origin. If the system (3) admits an asymptotic stabilizing the encoder and an inertial measurement unit (IMU), respectively and
feedback 𝑘 ∶  →  then there exists a Control Lyapunov Function 𝛼 = 𝜃 + 𝜃𝑠 − 𝜋2 . In this work, 𝜃 is the variable to be controlled and it
(CLF) 𝑉 ∶  → R, that is a smooth function, positive definite such that: is measured with a differential encoder in the AAFO. Other variables
𝜕𝑉 𝜕𝑉 measured in this study are used to evaluate the performance of the
𝑉̇ = 𝑓 (𝑥) + 𝑔(𝑥) 𝑘(𝑥) < 0 (4)
𝜕𝑥 𝜕𝑥 proposed controller.
⏟⏞⏟⏞⏟ ⏟⏞⏟⏞⏟
∶=𝑎(𝑥) ∶=𝑏(𝑥)
Several torques acting on the AAFO system at the ankle level are
being considered. They can be classified in endogenous (which are
It is worth noting that if 𝑘 is assumed to be smooth, then 𝑉 is known to
generated by the environment) and exogenous disturbances (which are
exist and to be as smooth as 𝑘. In the present work, only the smoothness
dependent upon internal variables). The former group are the gravity
of 𝑉 is required which is less restrictive than the one of 𝑘. An equivalent
torque 𝜏𝑔𝑟𝑎𝑣𝑖𝑡𝑦 , the solid and viscous friction torques 𝜏𝑓 , and system’s
requirement on the time derivative of 𝑉 is that:
joint stiffness torque 𝜏𝑠 . The later group are the torque induced by the
𝑏(𝑥) = 0 ⇒ 𝑎(𝑥) < 0 translational acceleration of the foot 𝜏𝑎 , the torque created by the ground
reaction forces 𝜏𝑟 , and the torque created by the plantar flexion and
Theorem 2.4 (Sontag’s Universal Formula (Sontag, 1998)). Assume that dorsiflexion muscle groups 𝜏ℎ . Also, the torque developed by the AAFO’s
system (3) admits 𝑉 as CLF. For any real analytic function 𝑞̄ ∶ R → R such actuator, namely, the control torque 𝑢, is considered. Since the AAFO
that 𝑞(0)
̄ = 0 and 𝑏𝑞(𝑏)̄ > 0 for 𝑏 ≠ 0, let 𝛾𝐶 ∶ R2 → R be defined by used for this study has a degree of freedom only in the plantar and
⎧ 𝑎(𝑥) + √𝑎(𝑥)2 + 𝑏(𝑥)𝑞(𝑏(𝑥))
̄
dorsiflexion direction, the modeling of the system is restricted to the
⎪ if 𝑏 ≠ 0 sagital plane. Thus, the AAFO-human’s dynamics can be expressed as
𝛾𝐶 (𝑎, 𝑏) ∶= ⎨ 𝑏(𝑥) (5)
⎪0 follows (Arnez-Paniagua et al., 2017a):
⎩ if 𝑏 = 0

Let the feedback 𝑢 ∶  →  , smooth on  ⧵ {0} be defined by 𝐽 𝜃̈ = 𝜏𝑓 + 𝜏𝑎 + 𝜏𝑠 + 𝜏𝑟 + 𝜏𝑔𝑟𝑎𝑣𝑖𝑡𝑦 + 𝜏ℎ + 𝑢 (7)

𝑢𝑖 (𝑥) ∶= −𝑏𝑖 (𝑥)𝛾𝐶 (𝑎(𝑥), 𝛽(𝑥)), 𝑖 ∈ {1, … , 𝑝} (6) All the torques are considered positive if they induce a counter
clockwise rotation. Each of the aforementioned torques is defined as
with 𝛽(𝑥) ∶= ‖𝑏(𝑥)‖2 . Then 𝑢 is such that for all non zero 𝑥, 𝜕𝑉 𝜕𝑥
𝑓 (𝑥) + follows:
𝜕𝑉
𝑔(𝑥)𝑢(𝑥) < 0. Moreover, if the CLF satisfies the so called small control
𝜕𝑥 𝜏𝑓 = − 𝑘𝑓𝑆 sign𝜃̇ − 𝑘𝑓𝑉 𝜃̇
property (Sontag, 1998), then taking 𝑞(𝑏(𝑥))
̄ ∶= 𝑏(𝑥), the control is
continuous at the origin. 𝜏𝑎 = − 𝑘𝑎 (𝑎𝑦 cos 𝛼 − 𝑎𝑥 sin 𝛼)
𝜏𝑠 = − 𝐾𝑠 (𝜃 − 𝜃𝑟 ) (8)
3. Actuated-Ankle–Foot-Orthosis (AAFO) dynamics model and 𝜏𝑟 = − 𝑘𝐺𝑅𝐹 (𝑥𝑔 𝐹𝑟 ) cos 𝛼
properties 𝜏𝑔𝑟𝑎𝑣𝑖𝑡𝑦 = − 𝜏𝑔 cos 𝛼

3.1. System description and modeling
The orthosis used in this study is an AAFO attached to the subject’s
left leg, which was provided by the Laboratory of Images, Signals and
Intelligent Systems (LISSI), as shown in Fig. 1(a). The orthosis has one
active and one passive degrees of freedom (DoF), at the ankle and the
knee joints, respectively. The active DoF is driven by a DC motor and a
gearbox.
where, 𝐽 is the moment of inertia of the foot, 𝑘𝑓𝑆 and 𝑘𝑓𝑉 are the solid
and viscous friction coefficients, 𝐾𝑠 is the system’s stiffness coefficient,
𝑘𝑎 is the system’s acceleration torque coefficient, 𝑎𝑥 and 𝑎𝑦 are the
longitudinal and vertical linear accelerations, expressed in the  frame,
𝐹𝑟 is the equivalent ground reaction force (GRF) applied to the center of
mass of the foot, 𝑥𝑔 is the position of the center of mass expressed in the
 frame, and 𝜏𝑔 = 𝑚𝑔𝑥𝑔 is the system’s gravity torque coefficient, where
𝑚 and 𝑔, represent the mass of the foot and the gravity acceleration

52
J.F. Guerrero-Castellanos et al. Control Engineering Practice 80 (2018) 49–60

coefficient. Replacing (8) in (7) we obtain: 4.1. ESO design for the AAFO
𝐽 𝜃̈ = −𝜏𝑔 cos 𝛼 − 𝑘𝑓𝑆 𝑠𝑖𝑔𝑛𝜃̇ − 𝑘𝑓𝑉 𝜃̇ − 𝐾𝑠 (𝜃 − 𝜃𝑟 )
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ Let 𝑒 = 𝐹 − 𝐹1 be the estimation error, through (15), we propose the
̇
∶=𝜉𝑒𝑛 (𝜃,𝜃,𝛼,𝑡) following extended state observer
(9)
−𝑘𝑎 (𝑎𝑦 cos 𝛼 − 𝑎𝑥 sin 𝛼) − 𝑘𝐺𝑅𝐹 (𝑥𝑔 𝐹𝑟 ) cos 𝛼 + 𝜏ℎ +𝑢 ⎧𝐹̇ = 𝐹 + 𝑙 𝑒
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⎪ 1 1
2 3

̇
⎪𝐹̇ 2 = 𝑢 + 𝜂1 + 𝑙2 𝑒
∶=𝜉𝑒𝑥 (𝜃,𝜃,𝛼,𝑡) 𝛴𝐸𝑆𝑂 ∶= ⎨ 𝐽 (16)
̇ 𝛼, 𝑡) and 𝜉𝑒𝑥 (𝜃, 𝜃,
̇ 𝛼, 𝑡) represent the sum of all endogenous ⎪𝜂̇ 1 = 𝜂2 + 𝑙1 𝑒
where, 𝜉𝑒𝑛 (𝜃, 𝜃, ⎪𝜂̇ = 𝑙 𝑒
and exogenous torques, respectively. ⎩ 2 0

Note that the system’s parameters described in (8) are difficult or where 𝐹1 , 𝐹2 are the estimated variable of the flat output 𝐹 and its
impossible to measure directly, or require a mathematical model to derivative, respectively. 𝜂1 and 𝜂2 represent the disturbance estimation
estimate the values through prior experiments (Son, Hwang, & Kim, and its time derivative, respectively, and 𝑢 is the control input. The set
2010). Furthermore, the values of these parameters could be changing of coefficients {𝑙3 , 𝑙2 , 𝑙1 , 𝑙0 } are constant values, which are selected with
throughout a single session. For example, the torque generated by the the assistance of a desired closed-loop Hurwitz polynomial of fourth-
interaction with the ground could be estimated using force sensitive order for the linear dominating dynamics of the injected estimation error
resistors (FSR) embedded in the insoles of the shoes (Arnez-Paniagua dynamics. Then, the following result is obtained.
et al., 2017a). However, this estimation would be inaccurate if a small
number of sensors is used and only the GRF perpendicular to the ground Proposition 4.1. Consider the ESO (16) and the AAFO system (15). Then,
is considered. It is for these reasons that a linear extended state observer the estimation error behavior satisfies the ISO property (Input to Output
that is able to estimate the total endogenous and exogenous disturbances Stability) (Sontag & Wang, 1995), i.e. the solutions for the error dynamics
was developed and it is presented in Section 4.1. converges to a sphere centered at the origin of the estimation error phase
space with radius
3.2. Flatness differential of the AAFO
𝛼𝐾0
𝜌= (17)
Let 𝑦 = 𝜃 be an associated output for the AAFO system (9). This 𝐽 𝜆1
results in: where 𝛼 and 𝜆1 are parameters that depend of the selection of the set
1( ̇ 𝛼, 𝑡) + 𝑢
) of coefficients {𝑙3 , 𝑙2 , 𝑙1 , 𝑙0 }. Furthermore, the error dynamics exhibits
𝑦̈ = 𝜉(𝜃, 𝜃, (10)
𝐽 asymptotic stability to 𝑒 = 0 for 𝜉 = 0.
where, 𝜉(𝜃, 𝜃,̇ 𝛼, 𝑡) = 𝜉𝑒𝑛 (𝜃, 𝜃,
̇ 𝛼, 𝑡) + 𝜉𝑒𝑥 (𝜃, 𝜃,
̇ 𝛼, 𝑡). Inspection of (10) shows
that choosing 𝑢 = −𝜉(𝜃, 𝜃, ̇ 𝛼, 𝑡) + 𝐽 𝑣 results in the linear system 𝑦̈ = 𝑣, Proof. Consider the estimation error 𝑒 = 𝐹 − 𝐹1 and its successive
where 𝑣 is a feedback component. From Theorem 2.1, the AAFO system derivatives together with the observer dynamics (16) and the AAFO
(9) is differentially flat, such that its Flat Output is the angular position dynamics (15). Then, estimation error 𝑒 satisfies the following linear
𝜃. The flatness of the system implies that all variables of the system and differential equation
the control vector 𝑢 can be parameterizable in terms of 𝐹 = 𝜃 and a 1̈
finite number of its derivatives with respect to time, that is: 𝑒(4) + 𝑙3 𝑒(3) + 𝑙2 𝑒̈ + 𝑙1 𝑒̇ + 𝑙0 𝑒 =
𝜉 (18)
𝐽
The space–state realization of (18) is the following
𝜃 = 𝐹 (11) {
𝑥̇ = 𝐴𝑥 + 𝐵𝜉
𝜃̇ = 𝐹̇ (12) 𝛴𝐸𝑂 ∶= (19)
𝑒 = 𝐶𝑥
𝜃̈ = 𝐹̈ (13)
where
𝑢 = 𝐽 𝐹̈ − 𝜉(𝐹 , 𝐹̇ , 𝛼, 𝑡) (14)
⎛ 0 1 0 ⎛0⎞
0 ⎞
The function, 𝜉(𝐹 , 𝐹̇ , 𝛼, 𝑡) represents the included endogenous and ex- ⎜ ⎜0⎟ ⎟ ( )
0 0 1 0 ⎟
𝐴=⎜ 𝐵 =⎜0⎟ 𝐶 = 0 0 1 0 (20)
ogenous disturbances which is unknown, but is assumed to be uniformly ⎜ 0 0 0 ⎜ ⎟
1 ⎟
absolutely bounded. ⎜−𝑙 −𝑙1 −𝑙2 ⎜ ⎟
−𝑙3 ⎟⎠
1
⎝ 0 ⎝𝐽 ⎠
4. ADRC design for the Actuated-Ankle–Foot-Orthosis (AAFO) The variation of parameters formula gives the following solution
𝑡
In this section a position trajectory tracking for a lower limb orthosis 𝑥(𝑡) = exp (𝐴𝑡) 𝑥(0) + exp(𝐴(𝑡 − 𝜏))𝐵𝜉𝑑𝜏
∫ 0
performed at the level of the ankle joint is addressed. In order to tackle (21)
𝑡
this problem, a control strategy based on the Active Disturbance Rejec-
𝑒(𝑡) = 𝐶 exp (𝐴𝑡) 𝑥(0) + 𝐶 exp(𝐴(𝑡 − 𝜏))𝐵𝜉𝑑𝜏
tion Control (ADRC) technique is proposed. The function, 𝜉(𝐹 , 𝐹̇ , 𝛼, 𝑡), ∫ 0
will be estimated through of the Extended State Observer (ESO), which If the set of coefficients {𝑙3 , 𝑙2 , 𝑙1 , 𝑙0 } are selected such that the matrix
is based on the system’s dynamic: 𝐴 is Hurwitz, the following inequality is obtained:
1[ ] 𝑡
𝐹̈ = 𝑢 + 𝜉(𝐹 , 𝐹̇ , 𝛼, 𝑡) (15) ‖𝑒(𝑡)‖ ≤ ‖ exp (𝐴𝑡) 𝑒(0)‖ + ‖ exp(𝐴(𝑡 − 𝜏))𝐵𝜉𝑑𝜏‖
𝐽 ∫0
For this purpose, the following assumptions stand: 𝑡
(22)
≤ ‖ exp (𝐴𝑡) ‖‖𝑒(0)‖ + |𝜉|‖𝐵‖ ‖ exp(𝐴(𝑡 − 𝜏))‖𝑑𝜏
∙ Only the flat output is measured, i.e. the angular position 𝐹 = 𝜃; ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ∫0
∙ The nominal value of inertia moment 𝐽 is known; 𝛽(‖𝑥(0)‖,𝑡)∈ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
𝛾(‖𝜉‖∞ )∈∞
∙ The perturbation function 𝜉(𝐹 , 𝐹̇ , 𝛼, 𝑡) is a uniformly absolutely
bounded disturbance, i.e. sup𝑡 |𝜉(⋅)| = ‖𝜉(𝐹 , 𝐹̇ , 𝛼, 𝑡)‖∞ ≤ 𝐾0 . Since there exist 𝛽 ∈  and 𝜚 ∈ ∞ the error dynamics is ISO (Sontag
∙ The disturbance estimation and its time derivative will be de- & Wang, 1995), i.e. the solutions for the error dynamics are ultimately
noted by 𝜂1 and 𝜂2 , respectively; bounded by 𝛾(‖𝜉‖∞ ) and the system exhibits asymptotic stability to
∙ The estimated variable of the flat output 𝐹̂ , and its successive 𝑒(𝑡) = 0 for 𝜉 = 0. Furthermore, if the set of coefficients {𝑙3 , 𝑙2 , 𝑙1 , 𝑙0 }
derivative 𝐹̂̇ , will be denoted by 𝐹1 = 𝐹̂ , and 𝐹2 = 𝑑 𝐹̂ ∕𝑑𝑡. are selected such that the matrix 𝐴 is Hurwitz, with real eigenvalues

53
J.F. Guerrero-Castellanos et al. Control Engineering Practice 80 (2018) 49–60
𝑇 𝑇
listed in increasing order 𝜆1 (𝐴) < 𝜆2 (𝐴) < 𝜆3 (𝐴) < 𝜆4 (𝐴), one knows 𝑉̇ (𝑧) = 𝑧̇ 𝑃 𝑧 + 𝑧 𝑃 𝑧̇
that, ‖ exp (𝐴𝑡) ‖ ≤ 𝛼 exp(−𝜆1 𝑡) with 𝛼 = ‖𝑇 ‖‖𝑇 −1 ‖ where 𝑇 is a matrix ( )
𝜉
such that 𝑇 −1 𝐴𝑇 is diagonal. Using this fact in (22), one obtains = 𝑧𝑇 (𝐴𝑇 𝑃 + 𝑃 𝐴)𝑧 + 2𝑧𝑇 𝑃 𝐵 𝑣 +
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏟⏟ 𝐽
∶=𝑏(𝑧) (29)
( ) 𝛼𝐾0 𝑡 ( ) ∶=𝑎(𝑧)
‖𝑒(𝑡)‖ ≤ exp −𝜆1 𝑡 ‖𝑥(0)‖ + exp −𝜆1 (𝑡 − 𝜏) 𝑑𝜏
𝐽 ∫ 0 = 𝑧𝑇 (𝐴𝑇 𝑃 + 𝑃 𝐴 − 2𝜀𝑃 𝐵𝐵 𝑇 𝑃 ) 𝑧 < 0
(23) ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
( ) 𝛼𝐾0 ( ) −𝑄
≤ exp −𝜆1 𝑡 ‖𝑥(0)‖ + (1 − exp −𝜆1 𝑡 )
𝐽 𝜆1
Furthermore, for 𝑄 = diag(𝑞1 , 𝑞2 ) and 𝜀 > 0, the solution of the Riccati
Consequently ‖𝑒(𝑡)‖ converges exponentially to a sphere with radius equation is satisfied for 𝑃 given in (30)
𝛼𝐾
𝜌 = 𝐽 𝜆 0 when 𝑡 → ∞. ■ ( )
1 𝑝11 𝑝12
𝑃 = (30)
𝑝21 𝑝22
Remark 4.2. The first term in (23) may dominate for small 𝑡, and this
√ √ √ √ √ √
serves to quantify the magnitude of the observer’s transient behavior as 𝑞 𝑞 𝑞
where 𝑝11 = 2𝜀 2𝜀1 1𝜀 2𝜀1 + 2𝜀2 , 𝑝12 = 𝑝21 =
𝑞1
and 𝑝 = 1 𝑞1
2𝜀 22 𝜀 2𝜀
a function of the size of the initial state 𝑥(0) and the value of 𝜆1 . This 𝑝
Besides, from (29), note that 𝑏(𝑥) = 0 implies 𝑧1 = − 𝑝22 𝑧2 , as a
property allow us to chose a judicious set of coefficients {𝑙3 , 𝑙2 , 𝑙1 , 𝑙0 } √ 12
𝑞
in order to provide an acceptable transient term. consequence 𝑎(𝑥) = −2( 2𝜀1 + 𝑞2 )𝑧22 < 0. Thus, the time derivative of
𝑉 meets the following requirement:
Remark 4.3. The first term in (22) vanishes for a 𝑡 sufficiently large.
𝑏(𝑥) = 0 ⇒ 𝑎(𝑥) < 0 for all 𝑧 ≠ 𝑧𝑒
Thus the dynamics of estimation error satisfies the asymptotic gain (AG)
property. That is, for all large enough 𝑡, the trajectory exists, and it Then 𝑉 given by (26) is a Control Lyapunov Function (CLF) for the
gets arbitrarily close to a sphere whose radius is proportional to the system (25) relative to the equilibrium state 𝑧𝑒 = (0 0)𝑇 . ■
bound of 𝜉 and inversely proportional to 𝜆1 . In the language of robust The Proposition 4.4 was used only to construct a CLF with which the
control, the estimate (AG) would be called an ‘‘ultimate boundedness’’ active disturbance rejection control (ADRC) and the observer (16), used
condition. as a rejector, are designed. Then, the feedback component 𝑣 is defined
as:
( )
4.2. Tracking control design for the AAFO 𝑣(𝑧) = −𝑏(𝑧) 𝛾𝐶 (𝑎(𝑧), 𝑏(𝑧)) + 𝜅 − 𝜂1 (𝑡) (31)
where 𝜅 ∈ R>0 . Using (5):
It is desired to drive the flat output 𝐹 = 𝜃 of the system (15), to track
⎧ 𝑎(𝑧) + √𝑎2 (𝑧) + 𝑏4 (𝑧)
a given smooth reference trajectory 𝐹 ∗ , defined by a healthy profile ⎪
𝛾𝐶 (𝑎(𝑧), 𝑏(𝑧)) ∶= ⎨ if 𝑏(𝑧) ≠ 0 (32)
of the ankle joint during the gait cycle, regardless of the unknown 𝑏2 (𝑧)
⎪0 if 𝑏(𝑧) = 0
but uniformly bounded nature of 𝜉. Then, the objective is to design a ⎩
control law using the estimated disturbance in order to cancel it. Since Furthermore, the system exhibits asymptotic stability to the origin of
the state is available, they will be used in the feedback. The proposed the state space tracking error for 𝛥 = 0.
trajectory tracking controller with disturbance rejector has the following
topology Proposition 4.5. The AAFO system (15) with control (24) with 𝑣 given
( ) by (31) is ISS with respect to 𝛥, where 𝛥 = (𝜉∕𝐽 − 𝜂1 ), i.e. the difference
𝑢 = 𝐽 𝐹̈ ∗ (𝑡) − 𝑣 (24) between the total unknown disturbance and its estimation obtained by (16).
𝐹̈ ∗ (𝑡) the feed-forward component and 𝑣 the feedback component to be
Proof. Let 𝑉 ∶ R2 ⟶ R be a Control Lyapunov Function (CLF) for the
determined. Let 𝑧1 = 𝐹 − 𝐹 ∗ and 𝑧2 = 𝐹̇ − 𝐹̇ ∗ be the flat output tracking
system (25) defined by (26). The derivative of (26) along the trajectories
error and its derivative, the AAFO system (15) can be expressed as of the closed-loop becomes
⎧ 𝑧̇ = 𝑧 𝑉̇ (𝑧) = 𝑧̇ 𝑇 𝑃 𝑧 + 𝑧𝑇 𝑃 𝑧̇
⎪ 1 2
𝛴𝐸𝐶 ∶= ⎨ 𝜉 (25) ( )
⎪ 𝑧̇ 2 = 𝑣 + 𝐽 𝜉
= 𝑧𝑇 (𝐴𝑇 𝑃 + 𝑃 𝐴)𝑧 + 2𝑧𝑇 𝑃 𝐵 𝑣 +
⎩ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏟⏟ 𝐽
( ) ∶=𝑎(𝑧) ∶=𝑏(𝑧)
which has the form 𝑧̇ = 𝐴𝑧 + 𝐵 𝑣 + 𝐽𝜉 . Since the proposed control law ( )
( ) 𝜉
in this work is based on an CLF approach, the first step is to find a CLF = 𝑎(𝑧) + 𝑏(𝑧) −𝑏(𝑧) 𝛾𝐶 (𝑎(𝑧), 𝑏(𝑧)) + 𝜅 − 𝜂1 (𝑡) +
for the system (25). This is summarized in the following proposition. 𝐽
√ 𝜉
= − 𝑎2 (𝑧) + 𝑏4 (𝑧) − 𝜅𝑏2 (𝑧) + 𝑏(𝑧) ( − 𝜂1 )
Proposition 4.4. The function 𝑉 ∶ R2 ⟶ R defined by 𝐽
⏟⏞⏟⏞⏟ (33)
𝛥
𝑉 (𝑧) = 𝑧𝑇 𝑃 𝑧 (26) √ [ ]
2𝑏(𝑧)𝛥 𝛥2 𝛥2
= − 𝑎2 (𝑧) + 𝑏4 (𝑧) − 𝜅 𝑏2 (𝑧) − + +
is a Control Lyapunov Function (CLF) for the system (25) relative to the 2𝜅 4𝑘2 4𝑘
equilibrium state 𝑧𝑒 = (0 0)𝑇 with the stabilizing control: √ [ ]2 2
𝛥 𝛥
= − 𝑎2 (𝑧) + 𝑏4 (𝑧) − 𝜅 𝑏(𝑧) − +
𝜉 2𝑘 4𝑘
𝑣 = −𝜀𝐵 𝑇 𝑃 𝑧 − (27)
𝐽 √ 𝛥2
≤ − 𝑎2 (𝑧) + 𝑏4 (𝑧) +
with 𝑃 the solution of the Riccati equation: ⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟ 4𝑘
𝑊 (𝑧)
𝐴𝑇 𝑃 + 𝑃 𝐴 − 2𝜀𝑃 𝐵𝐵 𝑇 𝑃 = −𝑄 (28) since 𝑊 (𝑧) is positive definite and radially unbounded, there exist a
class ∞ function 𝜚̄ such that 𝑊 (𝑧) ≥ 𝜚(|𝑧|)
̄ and therefore
Proof. Clearly 𝑉 is smooth and positive definite. Now, consider the
𝛥2
derivative of (26) along the trajectories of the closed-loop 𝑉̇ (𝑧) ≤ −𝜚(|𝑧|)
̄ + (34)
4𝑘

54
J.F. Guerrero-Castellanos et al.
Fig. 2. ADRC: ESO and control law based on CLF approach.
Fig. 3. ISS interpretation for the ADRC.
( 2)
It follows that if |𝑧| ≥ 𝜚̄−1 𝛥2𝑘 , the 𝑉̇ (𝑧) ≤ − 21 𝜚(|𝑧|).
̄ By Definition 2.2
and Theorem 2.3, the AAFO system (15) with control (24), (31) and (32)
is ISS with respect to 𝛥. ■
Consider the Active Disturbance Rejection Controller (ADRC) com-
posed of the Extended State Observer (ESO) (16) and the trajectory
tracking controller (24) working together to drive the AAFO system (15)
to follow a desired angular trajectory (see Fig. 2). From (19) and (25) the
observer dynamics error and the tracking dynamics error can be viewed
as two ISS systems in cascade (see Fig. 3). Then, the following result for
the ADRC stands.
Proposition 4.6. The Active Disturbance Rejection Controller (ADRC)
composed of the Extended State Observer (ESO) (16) and the trajectory
tracking controller (24) is ISS, when the unknown total disturbance 𝜉 is
viewed as the input and the tracking error 𝑧 as the output.
Proof. The claim follows from the properties for the interconnection
of nonlinear systems that are Input-to-State Stable (ISS) (Sontag,
1998, 2008). Consider the estimation error dynamics (19) and the
tracking error dynamics (25). These systems can be viewed as a cascade
system as shown in Fig. 3. Since each system is ISS as was proven in
Propositions 4.1 and 4.5 , then the overall system is ISS. That is, the
ADRC is ISS when the unknown total disturbance 𝜉 is the input and the
tracking error 𝑧 the output. ■
5. Experimental results
In this section, the effectiveness of the AAFO’s controller to assist the
wearer during walking is assessed through real-time experiments. The
55
Control Engineering Practice 80 (2018) 49–60
objective in this early stage of the study is to prove the feasibility and
safety of the system, i.e., the ankle joint needs to track the reference
even in the presence of the endogenous and exogenous disturbances
(which are considered bounded and unpredictable), with a high level of
repeatability and stability. To evaluate the feasibility of the system, the
ankle position and velocity errors are measured and compared between
unassisted and assisted sessions. Furthermore, the profiles of both errors
are normalized with respect to the gait cycle and analyzed in order to
evaluate the repeatability of the assistance even in the presence of the
wearer’s voluntary action.
To guarantee the safety of the subject while walking, the mechanical
design of the AAFO ensures an ankle joint movement within a limited
range. This range limit is set to −32◦ for the dorsiflexion and 22◦ for
the plantar flexion. Furthermore, the maximum torque produced by the
AAFO in both directions is 15 Nm, which is relatively small compared
to the torque that can be produced by the ankle joint muscles of a
healthy person. Finally, if at any moment the torque calculated by the
controller exceeds a 15 Nm threshold, the system automatically disables
the AAFO’s motor for the remaining of the session.
Since it is understood that the healthy subjects produce healthy ankle
joint profiles, the reference profile is deliberately different from the
wearer’s normal ankle joint trajectory. In the scenario of applying the
system to a paretic patient, the reference trajectory would be adjusted to
better represent a healthy ankle joint profile. The algorithm to generate
the ankle joint reference trajectory is presented in the following sub-
section.
5.1. Adaptive reference trajectory
The ankle joint angle profile used as reference was obtained from
analyzing the gait of 20 healthy subjects using a motion capture
system (Motion Analysis Corporation, Santa Rosa, CA, USA, six cameras,
Sampling Frequency 100 Hz) and two force plates (AMTI, Watertown,
MA, USA, Sampling Frequency 1000 Hz), for more details please refer
to (Hutin, Pradon, Barbier, Bussel, Gracies, & Roche, 2012). This
ankle joint angle reference is then divided in two main phases: the
stance and the swing phases, as it is shown in Fig. 4. The stance
phase of the gait cycle can be divided into several sub-phases that
are: loading response (LR), mid-stance (MS), terminal stance (TS), and
pre-swing (PS). Similarly, the swing phase can be divided into several
J.F. Guerrero-Castellanos et al.
Fig. 4. On the top, the description of the gait by means of the gait sub-phases. On
the bottom, the events in the gait cycle that trigger a transition of gait sub-phases
or that describe a relative minimum or maximum in the ankle joint angle profile.
The ankle angle is measured from the relaxed ankle angle, such that the positive
values are considered dorsiflexion and the negative values are plantar-flexion.
sub-phases based on the analysis of the opposite foot that is in contact
with the ground. These sub-phases are: initial swing (ISw), mid-swing
(MSw), and terminal swing (TSw). The transition between gait phases is
characterized by specific events such as initial contact (IC), toe landing
(TL), heel off (HO), toe off (TO), and maximum dorsiflexion (MD) during
the swing phase. By detecting these specific events during the stance
phase, it is possible to determine the key moments at which the ankle
joint angle profile switches from dorsiflexion to plantar-flexion and vice-
versa as it is shown in Fig. 4. Particularly, during the swing phase, the
maximum dorsiflexion (MD) value happens around the 62% of the swing
phase duration, that is about 85% of the whole gait cycle.
In order to detect the gait events, three force sensitive resistors (FSR)
are placed in each of the insoles of both feet. A fuzzy-based gait detection
algorithm is used to identify in real time the gait phases and measure
the walking speed of the subject. For more details regarding this fuzzy-
logic based algorithm, the reader is invited to refer to the following
study (Arnez-Paniagua et al., 2016).
The adaptive reference trajectory is calculated by connecting the
ankle joint angle values at the gait events defined above using a cubic
spline function. For example, if the TL event is detected, a cubic
spline trajectory that connects the current value of the ankle joint
angle reference to the HO key point value is calculated. The duration
of the cubic spline is defined by the measured walking speed. The
adaptive reference of the ankle joint angular velocity and accelerations
are calculated afterwards by numerically deriving the calculated cubic
spline trajectory. For the swing phase, if the TO event is detected, the
cubic spline is calculated from the current ankle joint angle reference,
then follows the MD key point in the middle of the path, and ends with
the IC event.
5.2. Experimental setup
One healthy subject walking on a treadmill has participated in the
experiments. The subject is a 31 years old male, 180 cm tall, and
56
Control Engineering Practice 80 (2018) 49–60
weighting 63 Kg. The subject was informed of the experimental protocol
and gave his consent before participating in the experiment.
The orthosis used in this study is an AAFO attached with straps to
the subject’s left calf and thigh, as shown in Fig. 1(b). The orthosis has
an active rotational degree of freedom (DoF) at the ankle joint, which
is driven by a DC motor with a gearbox, and one passive rotational DoF
at the knee joint. The AAFO is equipped with an incremental encoder
to measure the ankle joint angle 𝐹 = 𝜃 at sampling rate of 1 KHz. The
angular velocity of the ankle joint 𝐹̇ is obtained by numerically deriving
the angular position 𝐹 , and then filtering with a lowpass, fourth order
Butterworth filter with a cutoff frequency of 20 Hz. For detecting the
gait cycle, three force sensitive resistors (FSR) are embedded in both
left and right insoles (a total of six FSR) and are connected to a wireless
system. A gait cycle is considered from the initial heel contact of the left
foot to the next initial heel contact of the same foot. All data were time
normalized to 100% of the gait cycle. At the end of the experiments, the
ankle joint angle, the reference profile, and the assistance torque have
been re-sampled at 2000 samples per gait cycle, such that each point
represents 0.05% of the gait cycle. Afterwards, these three signals have
been averaged at each sample for all strides to obtain the mean profiles
for each signal normalized to the gait cycle.
5.3. Experimental protocol
The scenario of the experiment with the healthy subject is as follows:
first, the subject walks on a treadmill at a speed of 2 Km/h at a self-
selected step duration. The ankle joint reference profile is deliberately
adjusted to be different from the proper ankle joint angle profile of the
subject in order to test the capability of the system to modify the current
ankle joint profile during the gait cycle. Then, three unassisted sessions
and ten assisted sessions, each lasting 60 s, are performed to evaluate
the repeatability and the consistency of the results. A resting time of 30
s between the sessions is provided.
To tune the controller gains, several calibration sessions were done
prior to the experiments on the treadmill. The subject was seated on
a chair with the leg above the ground (no GRF) and was asked not to
produce any muscular torque (𝜏ℎ = 0). The system then tracks a healthy
ankle joint walking profile, shown in Fig. 4. After each session, the
tracking performance is evaluated and the subject is asked about the
comfortability of the assistance provided. With this information, the so-
lution of the Riccati equation (30) was adjusted until the tracking error
was considered sufficiently small while remaining comfortable for the
subject. The values of the observer gains were selected from a simulation
in Matlab, Simulink, and further refined in the calibration sessions.
The exact value of 𝐽 cannot be measured but it can be estimated
using bio-mechanic tables from Winter (2009) to estimate the mass and
the center of mass of the foot. Afterwards, the moment of inertia 𝐽 is
calculated.
𝐽 = (𝑚𝑓 𝑜𝑜𝑡 + 𝑚𝑠ℎ𝑜𝑒 )𝑑𝐺2 (35)
where 𝑚𝑓 𝑜𝑜𝑡 and 𝑚𝑠ℎ𝑜𝑒 are the mass of the foot and the shoe, respectively,
and 𝑑𝐺 is the distance between the ankle joint to the center of mass of
the foot.
5.4. Results
First of all, the performance of the ESO (16) is evaluated with the
design parameters: 𝑙0 = 17160, 𝑙1 = 6026, 𝑙2 = 791, 𝑙3 = 46, and the
nominal value 𝐽 = 0.03. The controller (31) uses the CLF (26) with 𝑃
set to:
( )
58.8087 1.7321
𝑃 =
1.7321 0.1698
which corresponds to 𝜀 = 106.3, 𝑞1 = 637 and 𝑞2 = 2. Besides, 𝜅 = 20.
Note that these parameters allow tuning the controller and they were
selected, first in simulation, and finely adjusted in practice, in order to
J.F. Guerrero-Castellanos et al. Control Engineering Practice 80 (2018) 49–60

Fig. 5. The performance of the estimation of the angular position of the ankle joint for the first 30 s (𝐹1 vs 𝜃). The first second of the session is rescaled in order to
observe the transient of the estimation convergence.

̇ The first second of the session is rescaled in order to

Fig. 6. The performance of the estimation of the angular velocity of the ankle joint for the first 30 s (𝐹2 vs 𝜃).
observe the transient of the estimation convergence.

provide the control influence in the position and velocity errors. All the
observer’s states are initialized to zero.
The measured angular position and angular velocity, as well as the
redundant estimate of the angular position (Fig. 5) and the estimate
of the angular velocity (Fig. 6) are presented. Note that there is a
small transient of approximately 0.2 s (showed in a zoom for the first
second of the session); when the estimations of the angular position
and velocity, 𝐹1 and 𝐹2 , overshoot the current value of the ankle joint
angle and velocity, 𝜃 and 𝜃. ̇ After this transient, the estimated angular
position and angular velocity converge to the measured ones and the
variables’ trajectories become indistinguishable. The average error of
the estimation of the angular position of the ankle joint across all ten
assisted sessions is 1.41◦ with a standard deviation of ±0.07◦ , and for the
estimation of the angular velocity is 86.77◦ /s with a standard deviation
of ±3.66◦ /s. The average error in the estimation of the angular position
and velocity for each session is presented in Table 1.
The proposed controller was able to track the generated adaptive
desired ankle trajectory profile with an average root-mean-square (RMS)

57
J.F. Guerrero-Castellanos et al. Control Engineering Practice 80 (2018) 49–60

Fig. 7. Mean ankle joint angle position errors across the gait cycle for the
Fig. 9. The mean ankle joint velocity profile, normalized with respect to the gait
assisted and unassisted sessions with patient.
cycle, for the assisted and unassisted sessions and the reference. The standard
deviation is presented for each ankle joint velocity profile.

Table 1
Estimation error of the angular position and velocity in RMS for each session
(error±std)
Assisted session 𝐹 − 𝐹1 (◦ ) 𝐹̇ − 𝐹2 (◦ ∕s)
1 1.3382 ± 1.3382 82.4550 ± 82.4455
2 1.4302 ± 1.4302 87.9357 ± 87.9236
3 1.4276 ± 1.4276 87.5018 ± 87.4862
4 1.4016 ± 1.4016 86.2706 ± 86.2625
5 1.4104 ± 1.4104 86.7709 ± 86.7609
6 1.2993 ± 1.2993 80.3256 ± 80.3139
7 1.5318 ± 1.5318 92.4811 ± 92.4718
8 1.5002 ± 1.5002 91.7413 ± 91.7260
9 1.3910 ± 1.3910 86.4614 ± 86.4458
10 1.3940 ± 1.3940 85.7532 ± 85.7419

Table 2
RMS error of the angular position for each gait group for each session
(error◦ ±std◦ ∕𝑠). A = assistance session, NA = no assistance session.
Fig. 8. The mean ankle joint angle profile, normalized with respect to the gait Session Loading response Roll-over Push off Swing
cycle, for the assisted and unassisted sessions and the reference. The standard
NA1 4.2793 ± 0.5803 4.8386 ± 0.7933 9.4753 ± 7.0040 5.7494 ± 4.5516
deviation is presented for each ankle joint angle profile. NA2 5.5842 ± 0.9772 4.4168 ± 0.9698 10.2550 ± 7.1190 5.1530 ± 5.1546
NA3 6.5880 ± 0.6745 4.8336 ± 1.2353 10.3196 ± 7.1570 5.3097 ± 5.2984
A1 2.6521 ± 0.3778 2.1085 ± 1.2706 2.6645 ± 2.5814 2.4389 ± 2.1589
A2 2.3647 ± 0.2528 2.2412 ± 1.0942 2.6866 ± 2.6885 2.5164 ± 2.1465
angular position error of 5.42◦ and a standard deviation of 2.52◦ across A3 2.4122 ± 0.2325 2.2089 ± 1.0579 2.6906 ± 2.6902 2.7672 ± 2.2588
all ten assisted sessions. This represents a reduction in the angular A4 2.6950 ± 0.2642 1.9628 ± 1.2724 2.9073 ± 2.7199 2.8354 ± 2.0806
position error of 53.37% compared to the average RMS angular position A5 2.6496 ± 0.2596 2.0826 ± 1.1097 3.1649 ± 2.9709 2.5263 ± 1.9478
A6 3.0005 ± 0.2949 2.4264 ± 1.2929 2.6185 ± 2.5833 3.0150 ± 1.9721
error of the three unassisted sessions. The RMS angular position error is
A7 2.9084 ± 0.3658 2.3892 ± 1.1309 3.2182 ± 3.1869 3.1200 ± 2.7394
classified in four gait groups relative to the sub-phases: loading response A8 2.7258 ± 0.2328 2.1099 ± 1.2132 2.5300 ± 2.3874 3.0981 ± 2.1004
(LR), roll over (MS plus TS), push-off (PS), and swing (ISw plus MSw plus A9 3.1901 ± 0.2471 2.3458 ± 1.2967 2.6972 ± 2.5914 2.9332 ± 2.0527
TSw). Fig. 7 shows the position error values for each session, both with A10 2.6322 ± 0.3016 2.1597 ± 1.1076 2.9059 ± 2.7573 2.7109 ± 1.9190
and without assistance, classified by the aforementioned gait groups. It
can be observed that the normalized tracking error is reduced in average
by 50.34%, 53.08%, 71.96%, and 48.26% for the loading response, roll provide dorsiflexion assistance to begin the swing phase. It can be seen
over, push-off, and swing gait groups, respectively, when the assistance from the standard deviation that the control law has a high repeatability,
torque is provided. The values of the RMS angular position error for which makes it suitable for gait assistance applications.
each gait group of every session is presented in Table 2.
The angular position and velocity ankle joint profiles have been 6. Conclusions
normalized with respect to the gait cycle to compare the reference
trajectory with the assisted and unassisted mean ankle joint profiles. Traditional proportional, integral derivative (PID) controllers have
The results are presented in Figs. 8 and 9, where it can be seen that the been widely used with robotic orthoses in the rehabilitation environ-
controller is able to track the ankle joint reference profiles, especially ment (Roy et al., 2014; Veneva & Ferreira, 2014; Ward et al., 2010).
during push-off (this can also be seen in Fig. 7) While these controllers are simple to implement and can be used to track
In Fig. 10, the assistance torque presents a plantar flexion assistance a predefined trajectory for the ankle joint, they usually lack the perfor-
during the stance phase, with an peak value during terminal stance (TS), mance achieved by controllers that consider a mathematical model of
which represents an increasing assistance for push-off at the end of the the system. Control approaches that require accurate knowledge of the
stance phase. Immediately after this, the torque changes direction to system’s parameters, i.e. wearer-active orthosis’ parameters, to calculate

58
J.F. Guerrero-Castellanos et al.
Fig. 10. Mean assistance torque normalized with respect to the gait cycle for the
ten assisted sessions and the standard deviation. The negative values represent
a dorsiflexion assistance while the positive values represent a plantar-flexion
assistance.
the needed assistance torque might not be adequate in a rehabilitation
environment. This is due to the fact that the identification process has
to be done prior to each session and with each subject, which is time
and effort consuming. Furthermore, these system’s parameters might
change through a single rehabilitation session due to the fatigue for
example, which increases the risk of inappropriate assistance, reduces
the rehabilitation benefits, or even may cause injuries. Therefore, a
control approach with a robust active disturbance rejection presents a
clear advantage since it does not require to use extra sensors to measure
the system parameters.
Early intervention in rehabilitation of the gait in paretic patients
could require strategies that make the ankle joint follow a predefined
healthy trajectory profile. For example, for a paretic patient with a
drop foot pathology, the adaptive reference trajectory can be defined
with an increased dorsiflexion during the swing phase. This definition
of the desired trajectory based on the requirements of the wearer
to compensate for gait pathologies could promote appropriate and
meaningful assistance provided by the AAFO.
The purpose of the experiments is to evaluate the feasibility of the
system to assist the gait of a healthy subject. The linear extended state
observer (ESO) was able to estimate the ankle position and velocity
after the first 200 ms. The active disturbance rejection control with
the Lyapunov function selected was able to track the adaptive ankle
reference with high repeatability even in the presence of unpredictable
endogenous and exogenous disturbances. Note that in a pathological
gait, the gait pathologies could be considered as unpredictable distur-
bance patterns. These results are important for the use of an AAFO for
a trajectory tracking strategy in a rehabilitation environment.
Further studies will be conducted with paretic patients to validate
the proposed system’s efficacy in terms of correcting observed gait
pathologies and providing rehabilitation benefits.
Acknowledgment
The authors would like to thank the Mexican government through
Consejo Nacional de Ciencia y Tecnología (CONACyT), Mexico for the
financing of this study via scholarship. This work has been supported
by Vicerrectoría de Investigación y Estudios de Posgrados (VIEP-BUAP)
under grant 100305333-VIEP2018.
59
Control Engineering Practice 80 (2018) 49–60
References
Ab Patar, M. N. A., Said, A. F., Mahmud, J., Majeed, A. P. A., & Razman, M. A.
(2014). System integration and control of dynamic ankle foot orthosis for lower
limb rehabilitation. In International symposium on technology management and emerging
technologies (pp. 82–85). IEEE.
Aloulou, A., & Boubaker, O. (2015). A minimum jerk-impedance controller for planning
stable and safe walking patterns of biped robots. In G. Carbone, & F. Gomez-Bravo
(Eds.), Motion and operation planning of robotic systems: background and practical
approaches (pp. 385–415). Cham: Springer International Publishing.
Aloulou, A., & Boubaker, O. (2016). An optimal jerk-stiffness controller for gait pattern
generation in rough terrain. ROBOMECH Journal, 3(1), 13.
Ames, A. D., Galloway, K., Sreenath, K., & Grizzle, J. W. (2014). Rapidly exponentially
stabilizing control lyapunov functions and hybrid zero dynamics. Transactions on
Automatic Control, 59(4), 876–891.
Arnez-Paniagua, V., Huo, W., Colorado-Cervantes, I., Mohammed, S., & Amirat, Y. (2016).
A hybrid approach towards assisting ankle joint of paretic patients. In International
functional electrical stimulation society conference.
Arnez-Paniagua, V., Rifaï, H., Amirat, Y., & Mohammed, S. (2017a). Adaptive control
of an actuated-ankle-foot-orthosis. International Conference on Rehabilitation Robotics
(ICORR), 1584–1589.
Arnez-Paniagua, V., Rifa, H., Mohammed, S., & Amirat, Y. (2017b). Adaptive control of
an actuated ankle foot orthosis for foot-drop correction. International Federation of
Automatic Control (IFAC), 50(1), 1384–1389.
Aström, K., & Murray, R. (2008). Feedback systems: an introduction for scientists and
engineers. Princeton and Oxford: Princeton University Press.
Bai, Y., Gao, X., Zhao, J., Jin, F., Dai, F., & Lv, Y. (2015). A portable ankle-foot
rehabilitation orthosis powered by electric motor. The Open Mechanical Engineering
Journal, 9, 982–991.
Bartolini, G., Pisano, A., Punta, E., & Usai, E. (2003). A survey of applications of second-
order sliding mode control to mechanical systems. International Journal of Control,
76(9–10), 875–892.
Bharadwaj, K., Sugar, T. G., Koeneman, J. B., & Koeneman, E. J. (2005). Design of a robotic
gait trainer using spring over muscle actuators for ankle stroke rehabilitation. Journal
of Biomechanical Engineering , 127 (6), 1009–1013.
Blaya, J. A., & Herr, H. (2004). Adaptive control of a variable-impedance ankle-foot
orthosis to assist drop-foot gait. Transactions on Neural Systems and Rehabilitation
Engineering , 12(1), 24–31.
Boehler, A. W., Hollander, K. W., Sugar, T. G., & Shin, D. (2008). Design, implementation
and test results of a robust control method for a powered ankle foot orthosis (afo).
In International conference on robotics and automation (pp. 2025–2030). IEEE.
Boiadjiev, G. V., & Veneva, I. P. (2009). Control system for data acquisition and processing
of ankle-foot orthosis. IFAC Symposium on Robot Control, 42(16), 597–602.
Brahmi, B., Saad, M., Ochoa-Luna, C., & Rahman, M. H. (2017). Adaptive control of
an exoskeleton robot with uncertainties on kinematics and dynamics. International
Conference on Rehabilitation Robotics (ICORR), 1369–1374.
Cain, S. M., Gordon, K. E., & Ferris, D. P. (2007). Locomotor adaptation to a powered
ankle-foot orthosis depends on control method. Journal of Neuroengineering and
Rehabilitation, 4(1), 1.
Chan, L., Naghdy, F., & Stirling, D. (2013). Extended active observer for force estimation
and disturbance rejection of robotic manipulators. Robotics and Autonomous Systems,
61(12), 1277–1287.
Chang, X., Li, Y., Zhang, W., Wang, N., & Xue, W. (2015). Active disturbance rejection
control for a flywheel energy storage system. IEEE Transactions on Industrial Electronics,
62(2), 991–1001.
Coste, C. A., Jovic, J., Pissard-Gibollet, R., & Froger, J. (2014). Continuous gait cycle
index estimation for electrical stimulation assisted foot drop correction. Journal of
Neuroengineering and Rehabilitation, 11(1), 118.
Ferris, D. P., Gordon, K. E., Sawicki, G. S., & Peethambaran, A. (2006). An improved
powered ankle–foot orthosis using proportional myoelectric control. Gait & Posture,
23(4), 425–428.
Fuentes, G. A. R., Cortés-Romero, J. A., Zou, Z., Costa-Castelló, R., & Zhou, K. (2015).
Power active filter control based on a resonant disturbance observer. IET Power
Electronics, 8(4), 554–564.
Galloway, K., Sreenath, K., Ames, A. D., & Grizzle, J. W. (2015). Torque saturation in
bipedal robotic walking through control lyapunov function-based quadratic programs.
Access, 3, 323–332.
Gregg, R. D., Bretl, T. W., & Spong, M. W. (2010). A control theoretic approach to robot-
assisted locomotor therapy. In Conference on decision and control (pp. 1679–1686).
IEEE.
Han, J. (2009). From pid to active disturbance rejection control. Transactions on Industry
Electronics, 56(3), 900–906.
Hernandez-Méndez, A., Linares-Flores, J., Sira-Ramírez, H., Guerrero-Castellanos, J., &
Mino-Aguilar, G. (2017). A backstepping approach to decentralized active disturbance
rejection control of interacting boost converters. Transactions on Industry Applications,
53(4), 4063–4072.
Hitt, J., Oymagil, A. M., Sugar, T., Hollander, K., Boehler, A., & Fleeger, J. (2007).
Dynamically controlled ankle-foot orthosis (dco) with regenerative kinetics: incre-
mentally attaining user portability. International Conference on Robotics and Automation
(ICRA), 1541–1546.
J.F. Guerrero-Castellanos et al.
Hogan, N. (1985). Impedance control: an approach to manipulation: part itheory. J. Dyn.
Sys., Meas., Control., 1, 1–7.
Holgate, M. A., Bohler, A. W., & Suga, T. G. (2008). Control algorithms for ankle
robots: a reflection on the state-of-the-art and presentation of two novel algorithms.
International Conference on Biomedical Robotics and Biomechatronics, 97–102.
Huang, H., Wu, L., Han, J., Feng, G., & Lin, Y. (2004). A new synthesis method for unit
coordinated control system in thermal power plant-adrc control scheme. In Power
system technology. International conference on: Vol. 1 (pp. 133–138). IEEE.
Hutin, E., Pradon, D., Barbier, F., Bussel, B., Gracies, J. M., & Roche, N. (2012). Walking
velocity and lower limb coordination in hemiparesis. Gait and Posture, 36(2), 205–
211.
Ibarra, J. C. P., dos Santos, W. M., Krebs, H. I., & Siqueira, A. A. G. (2014). Adaptive
impedance control for robot-aided rehabilitation of ankle movements. In 5th IEEE
RAS/EMBS international conference on biomedical robotics and biomechatronics (pp. 664–
669).
Jamwal, P. (2011). Design analysis and control of wearable ankle rehabilitation robot. (Ph.D.
thesis), The university of Auckland, New Zealand.
Jamwal, P. K., Xie, S. Q., Hussain, S., & Parsons, J. G. (2014). An adaptive wearable
parallel robot for the treatment of ankle injuries. Transactions on Mechatronics, 19(1),
64–75.
Jankovic, M., Sepulchre, R., & Kokotović, P. V. (1999). Clf based designs with robustness
to dynamic input uncertainties. Systems & Control Letters, 37 (5), 4554.
Jimenez-Fabian, R., & Verlinden, O. (2011). Review of control algorithms for robotic ankle
systems in lower-limb orthoses, prostheses, and exoskeletons. Medical Engineering &
Physics, 34(4), 397–408.
Jiménez-Fabián, R., & Verlinden, O. (2012). Review of control algorithms for robotic ankle
systems in lower-limb orthoses, prostheses, and exoskeletons. Medical Engineering &
Physics, (ISSN: 1350-4533) 34(4), 397–408.
Krishnamoorthy, V., Hsu, W. L., Kesar, T. M., Benoit, D. L., Banala, S. K., Perumal, R.,
Sangwan, V., Binder-Macleod, S. A., Agrawal, S. K., & Scholz, J. P. (2008). Gait
training after stroke: a pilot study combining a gravity-balanced orthosis, functional
electrical stimulation, and visual feedback. Journal of Neurologic Physical Therapy,
32(4), 192–202.
Krstic, M., Kanellakopoulos, I., & Kokotovic, P. V. (1995). Nonlinear and adaptive control
design. New York: Jhon Wiley & Sons, Inc..
Lawn, M. J., Takashima, M., Ninomiya, M., Yu, J., Soma, K., & Ishimatsu, T. (2015).
Development of an actuation system for a rotary hydraulic brake on a low cost light
weight knee-ankle-foot orthosis. Sensors, 1–4.
Linares-Flores, J., García-Rodríguez, C., Sira-Ramírez, H., & Ramírez-Cárdenas, O. (2015).
Robust backstepping tracking controller for low-speed pmsm positioning system:
design, analysis, and implementation. Transactions on Industrial Informatics, 11(5),
1130–1141.
Linares-Flores, J., Sira-Ramírez, H., Barahona-Avalos, J. L., & Contreras-Ordaz, M. (2012).
Robust passivity-based control of a buckboost-converter/dc-motor system: an active
disturbance rejection approach. Transactions on Industry Applications, 48(6), 2362–
2371.
Long, Y., Du, Z., Cong, L., Wang, W., Zhang, Z., & Dong, W. (2017). Active disturbance
rejection control based human gait tracking for lower extremity rehabilitation ex-
oskeleton. ISA Transactions, 67 , 389–397.
Madani, T., Daachi, B., & Djouani, K. (2014). Finite-time control of an actuated orthosis
using fast terminal sliding mode. International Federation of Automatic Control (IFAC),
47 (3), 4607–4612.
Marchand, N., Durand, S., & Guerrero-Castellanos, J. F. (2013). A general formula for
event-based stabilization of nonlinear systems. Transactions on Automatic Control,
(2013), 58(5).
Martnez-Fonseca, N., Castaeda, L. A., Uranga, A., Luviano-Jurez, A., & Chairez, I. (2016).
Robust disturbance rejection control of a biped robotic system using high-order
extended state observer. ISA Transactions, 62, 276–286.
Mehdi, H., & Boubaker, O. (2012a). New robust tracking control for safe constrained
robots under unknown impedance environment. In Advances in autonomous robotics.
ISBN: 978-3-642-32527-4, (pp. 313–323). Springer.
Mehdi, H., & Boubaker, O. (2012b). Stiffness and impedance control using lyapunov theory
for robot-aided rehabilitation. International Journal of Social Robotics, 4(1), 107–119.
Mehdi, H., & Boubaker, O. (2015). Robust impedance control-based lyapunov-hamiltonian
approach for constrained robots. International Journal of Advanced Robotic Systems,
12(12), 190.
Mehdi, H., & Boubaker, O. (2016). PSO-Lyapunov motion/force control of robot arms with
model uncertainties. Robotica, 34(3), 634651.
Mohammed, S., Huo, W., Huang, J., Rifa, H., & Amirat, Y. (2016). Nonlinear disturbance
observer based sliding mode control of a human-driven knee joint orthosis. Robotics
and Autonomous Systems, 75, 41–49.
Nicholas Romansky, D., Kelly Scollon-Grieve, M., & McGinness, C. James G. (2012).
Current concepts in diagnosing and treating drop foot. Podiatry Today, 25(6), 68–75.
Orozco-Soto, S. M., & Ibarra-Zannatha, J. M. (2017). Motion control of humanoid robots
using sliding mode observer-based active disturbance rejection control. In Colombian
conference on automatic control (pp. 1–8). IEEE.
Park, Y. L., Chen, B.-r., Young, D., Stirling, L., Wood, R. J., Goldfield, E., & Nag-
pal, R. (2011). Bio-inspired active soft orthotic device for ankle foot pathologies.
In International conference on intelligent robots and systems (pp. 4488–4495). IEEE.
60
Control Engineering Practice 80 (2018) 49–60
Peckham, P. H., & Knutson, J. S. (2005). Functional electrical stimulation for neuromus-
cular applications. Annu. Rev. Biomed. Eng., 7 , 327–360.
Pérez-Ibarra, J. C., & Siqueira, A. A. (2017). Comparison of kinematic and emg parameters
between unassisted, fixed-and adaptive-stiffness robotic-assisted ankle movements in
post-stroke subjects. International Conference on Rehabilitation Robotics (ICORR), 461–
466.
Primbs, J., Nevistic, V., & Doyle, J. (1999). Nonlinear optimal control: a control lyapunov
function and receding horizon perspective. Asian Journal of Control, 1(1), 14–24.
Rifaï, H., Mohammed, S., Hassani, W., & Amirat, Y. (2013). Nested saturation based
control of an actuated knee joint orthosis. Mechatronics, 23(8), 1141–1149.
Roy, A., Krebs, H. I., Iqbal, K., Macko, N. R., Macko, R. F., & Forrester, L. W. (2014).
Facilitating push-off propulsion: A biomechanical model of ankle robotics assistance
for plantarflexion gait training in stroke. In International conference on biomedical
robotics and biomechatronics (pp. 656–663). IEEE.
Sanfelice, R. G. (2013). On the existence of control lyapunov functions and state-feedback
laws for hybrid systems. Transactions on Automatic Control, (2013), 58.
Sepulchre, R., Jankovic, M., & Kokotović, P. V. (1997). Constructive nonlinear control.
London: Springler.
Shorter, K. A., Kogler, G. F., Loth, E., Durfee, W. K., & Hsiao-Wecksler, E. T. (2011).
A portable powered ankle-foot orthosis for rehabilitation. Journal of Rehabilitation
Research & Development, (2011), 48(4).
Shorter, K. A., Xia, J., Hsiao-Wecksler, E. T., Durfee, W. K., & Kogler, G. F. (2013).
Technologies for powered ankle-foot orthotic systems: possibilities and challenges.
Transactions on Mechatronics, 18(1), 337–347.
Sira-Ramírez, H., & Agrawal, S. K. (2004). Differentially flat systems. Marcel Dekker, Inc..
Sira-Ramírez, H., Hernández-Méndez, A., Linares-Flores, J., & Luviano-Juarez, A. (2016).
Robust flat filtering dsp based control of the boost converter. Control Theory and
Technology, 14(3), 224–236.
Sira-Ramírez, H., Linares-Flores, J., García-Rodríguez, C., & Contreras-Ordaz, M. (2014).
On the control of the permanent magnet synchronous motor: an active disturbance
rejection control approach. Transactions on Control Systems Technology, 22(5), 2056–
2066.
Sira-Ramírez, H., Luviano-Juárez, A., Ramírez-Neria, M., & Zurita-Bustamante, E. W.
(2017). Active disturbance rejection control of dynamic systems: a flatness-based approach.
Butterworth-Heinemann.
Son, J., Hwang, S., & Kim, Y. (2010). An emg-based muscle force monitoring system.
Journal of Mechanical Science and Technology, 24(10), 2099–2105.
Sontag, E., & Wang, Y. (1995). On characterizations of the input-to-state stability property.
Systems & Control Letters, 24, 351–359.
Sontag, E. D. (1998). Mathematical control theory, deterministic finite dimensional systems
(second ed.). New York Berlin Heidelberg: Springler Verlag.
Sontag, E. D. (2008). Input to state stability: basic concepts and results. In P. Nistri, & G.
Stefani (Eds.), Nonlinear and optimal control theory: Lectures given at the C.I.M.E. summer
school held in Cetraro, Italy June 19–29, 2004 (pp. 163–220). Berlin, Heidelberg:
Springer Berlin Heidelberg.
Spurgeon, S. (2014). Sliding mode control: a tutorial. In European control conference
(pp. 2272–2277).
Stewart, J. D. (2008). Foot drop: where, why and what to do?. Practical Neurology, 8(3),
158–169.
Veneva, I., & Ferreira, N. (2014). Adaptive system for control of active ankle-foot orthosis
and gait analysis. In Mathematical Methods in Engineering (pp. 153–163). Springer.
Ward, J., Sugar, T., Standeven, J., & Engsberg, J. R. (2010). Stroke survivor gait adaptation
and performance after training on a powered ankle foot orthosis. In International
conference on robotics and automation (pp. 211–216). IEEE.
Winter, D. A. (2009). Biomechanics and motor control of human movement. John Wiley &
Sons.
Wolbrecht, E. T., Chan, V., Le, V., Cramer, S. C., Reinkensmeyer, D. J., & Bobrow, J. E.
(2007). Real-time computer modeling of weakness following stroke optimizes robotic
assistance for movement therapy. International Conference on Neural Engineering , 152–
158.
Xue, W., Huang, Y., & Gao, Z. (2016). On adrc for non-minimum phase systems: canonical
form selection and stability conditions. Control Theory and Technology, 14(3), 199–
208.
Yang, H., Yu, Y., & Zhang, J. (2017). Angle tracking of a pneumatic muscle actuator
mechanism under varying load conditions. Control Engineering Practice, 61, 1–10.
Yu, Y., Yang, Z., Han, C., & Liu, H. (2018). Disturbance-observer based control for
magnetically suspended wheel with synchronous noise. Control Engineering Practice,
72, 83–89.
el Zahraa Wehbi, F., Huo, W., Amirat, Y., El Rafei, M., Khalil, M., & Mohammed, S. (2017).
Active impedance control of a knee-joint orthosis during swing phase. International
Conference on Rehabilitation Robotics (ICORR), 435–440.
Zhang, M., Cao, J., Xie, S. Q., Zhu, G., Zeng, X., Huang, X., & Xu, Q. (2017). A preliminary
study on robot-assisted ankle rehabilitation for the treatment of drop foot. Journal of
Intelligent and Robotic Systems, 1–9.