You are on page 1of 37

Elsevier Editorial System(tm) for ISA

Transactions
Manuscript Draft

Manuscript Number: ISATRANS-D-15-00379R3

Title: Robust disturbance rejection control of a biped robotic system


using high-order extended state observer

Article Type: Research Article

Section/Category: Analysis

Keywords: Active Disturbance Rejection; Extended State Observers; Biped


Robot; Nonlinear Mechanical Systems; Disturbance Observers

Corresponding Author: Dr. Alberto Luviano-Juarez, Ph.D.

Corresponding Author's Institution: UPIITA-IPN

First Author: Nadhynee Martinez-Fonseca, M. in Sc.

Order of Authors: Nadhynee Martinez-Fonseca, M. in Sc.; Luis A Castañeda,


M. in A.T.; Agustin Uranga, M. in A.T.; Alberto Luviano-Juarez, Ph.D.;
Isaac Chairez, Ph.D.

Abstract: This study addressed the problem of robust control of a biped


robot based on disturbance estimation. Active disturbance rejection
control was the paradigm used for controlling the biped robot by direct
active estimation. A robust controller was developed to implement
disturbance cancelation based on a linear extended state observer of high
gain class. A robust high-gain scheme was proposed for developing a state
estimator of the biped robot despite poor knowledge of the plant and the
presence of uncertainties. The estimated states provided by the state
estimator were used to implement a feedback controller that was effective
in actively rejecting the perturbations as well as forcing the trajectory
tracking error to within a small vicinity of the origin. The theoretical
convergence of the tracking error was proven using the Lyapunov theory.
The controller was implemented by numerical simulations that showed the
convergence of the tracking error. A comparison with a high-order
sliding-mode-observer-based controller confirmed the superior performance
of the controller using the robust observer introduced in this study.
Finally, the proposed controller was implemented on an actual biped robot
using an embedded hardware-in-the-loop strategy.

Suggested Reviewers:
Cover Letter

Wednesday, November, 23th, 2015

Dear Professor A.B. Rad


Editor in chief
ISA Transactions
School of Mechatronic Systems,
Simon Fraser University,
250-13450, 102 Avenue, Surrey,
V3T 0A3, British Columbia, Canada

Please find enclosed our manuscript, "Robust disturbance rejection control of a biped robotic
system using high-order extended state observer" by N. Martínez-Fonseca, L.A. Castañeda, A.
Uranga, A. Luviano-Juarez, and I. Chairez, which we would like to submit for publication as a
research paper in ISA Transactions.

This article describes the design of a multivariable observer-linear controller-based robust output
feedback scheme for output reference trajectory tracking tasks in a large class of nonlinear
mechanical systems, with an explicit application on a biped robotic system A robust controller was
developed to implement a disturbance cancelation based on a linear extended state observer of
high gain class. A robust high gain scheme was proposed for developing a state estimator of the
biped robot despite the poor knowledge of the plant and the presence of uncertainties. The
estimated states provided by the state estimator were used to implement a feedback controller
that was effective to reject actively the perturbations and forced the trajectory tracking to a small
vicinity of the origin. The theoretical convergence of the tracking error was proven using the
Lyapunov theory. The controller was implemented by some numerical simulations for showing the
convergence of the tracking error. A comparison with a high order sliding mode observer based
controller served to confirm the superior performance of the controller using the robust observer
introduced in this study. Finally, the proposed controller was implemented on an actual biped
robot using an embedded hardware in-the-loop strategy. We believe our findings would appeal to
the readership of the ISA Transactions Journal.

Please address all correspondence to:

Dr. Alberto Luviano Juárez


Instituto Politécnico Nacional - UPIITA
Av. IPN No. 2580, Col. Barrio la Laguna Ticomán 07340, D.F, Mexico
email:aluvianoj@ipn.mx

We confirm that this manuscript has not been published elsewhere and is not under consideration
by another journal. All authors have approved the manuscript and agree with its submission to ISA
Transactions.

We look forward to hearing from you at your earliest convenience.

Sincerely yours
The authors
*Title page showing Author Details

Robust disturbance rejection control of a biped robotic system using high-order


extended state observer

N. Martinez-Fonsecaa,, L.A. Castañedab,, A. Urangac,, A. Luviano-Juárezb,∗, I. Chaireza,


a Department of Bioprocessess UPIBI- IPN, Av. Acueducto S/N Col. Barrio La Laguna Ticomán C.P. 07340 México, D.F. México
b Instituto
Politécnico Nacional - UPIITA Av. IPN 2580 Col. Barrio La Laguna Ticomán D.F, México. Tel. +52 55 57296000 ext 56918
c Universidad Politécnica de Cuautitlán Izcalli. Department of Biomedical Engineering, Av Constitución 1000, Cumbria, 54740 Cuautitlán Izcalli,

México

Abstract

This study addressed the problem of robust control of a biped robot based on disturbance estimation. Active distur-
bance rejection control was the paradigm used for controlling the biped robot by direct active estimation. A robust
controller was developed to implement disturbance cancelation based on a linear extended state observer of high gain
class. A robust high-gain scheme was proposed for developing a state estimator of the biped robot despite poor knowl-
edge of the plant and the presence of uncertainties. The estimated states provided by the state estimator were used to
implement a feedback controller that was effective in actively rejecting the perturbations as well as forcing the trajec-
tory tracking error to within a small vicinity of the origin. The theoretical convergence of the tracking error was proven
using the Lyapunov theory. The controller was implemented by numerical simulations that showed the convergence
of the tracking error. A comparison with a high-order sliding-mode-observer-based controller confirmed the superior
performance of the controller using the robust observer introduced in this study. Finally, the proposed controller was
implemented on an actual biped robot using an embedded hardware-in-the-loop strategy.
Keywords: Active Disturbance Rejection, Extended State Observers, Biped Robot, Nonlinear Mechanical Systems,
Disturbance Observers


This work was supported by the Secretarı́a de Investigación y Posgrado (SIP-IPN) and CONACyT, México, under research grants SIP-
20150279, SIP-20140274
∗ Corresponding author

Email addresses: nyah.140408@gmail.com (N. Martinez-Fonseca ), castaneda.silux@gmail.com (L.A. Castañeda ),


auranga011@gmail.com (A. Uranga ), aluvianoj@ipn.mx (A. Luviano-Juárez), ichairezo@gmail.com (I. Chairez )

Preprint submitted to ISA Transactions January 21, 2016


*Detailed Response to Reviewers

January 21st, 2016

Lili Dong
Associate Professor
Associate Editor, ISA Transactions
Dept. of Electrical Engineering and Computer Science
Cleveland State University
2121 Euclid Ave.
Cleveland, OH 44115
l.dong34@csuohio.edu

Dear Professor Dong,

This file contains all our answers, replies and comments regarding the revision of the article
“Robust disturbance rejection control of a biped robotic system using high-order extended state
observer”, with reference ISATRANS-D-15-00379R2 as a regular paper.

According to the report of the reviews, we have proposed a new version of our manuscript. We
hope that after a new consideration of the paper, the article can be accepted for publication. We
have considered all the suggestions for changes proposed in the Associate Editor Report.

We understand that resubmitting our manuscript does not guarantee eventual acceptance.
However, we have made our best to prepare an acceptable manuscript with the expected quality
for the ISA Transactions Journal.

Sincerely,

Dr. Alberto Luviano-Juárez


Corresponding Author
National Polytechnic Institute - UPIITA
Associate Editor’s comments:

AE’s comments and revisions are included in the sticky notes in the manuscript. Please address
the comments and revise the paper accordingly

We appreciate the revision as well as the suggestions. The new version of the article took into
consideration the suggestions concerning, both, the main text as well as the figure 2.

Page 2
*Highlights (for review)

Robust disturbance rejection control of a biped robotic system using high-


order extended state observer
Highlights of the presented article:

• A novel Lyapunov function is given to prove the stability of Active Disturbance


Rejection Control
• Generalized Proportional Integral Observers are used to solve the robust con-
troller
• The problem of trajectory tracking of biped robots is solved using the proposed
controller.
• Numerical and experimental tests show the effectiveness of the suggested con-
troller.
• Comparison against a robust technique (sliding modes) are provided.

1
*Blinded Manuscript - without Author Details
Click here to download Blinded Manuscript - without Author Details: ADRC_biped_blinded_r3cbib.pdf
Click here to view linked References

Robust disturbance rejection control of a biped robotic


system using high-order extended state observer

Abstract

This study addressed the problem of robust control of a biped robot based on distur-
bance estimation. Active disturbance rejection control was the paradigm used for con-
trolling the biped robot by direct active estimation. A robust controller was developed
to implement disturbance cancelation based on a linear extended state observer of high
gain class. A robust high-gain scheme was proposed for developing a state estimator
of the biped robot despite poor knowledge of the plant and the presence of uncertain-
ties. The estimated states provided by the state estimator were used to implement a
feedback controller that was effective in actively rejecting the perturbations as well as
forcing the trajectory tracking error to within a small vicinity of the origin. The theo-
retical convergence of the tracking error was proven using the Lyapunov theory. The
controller was implemented by numerical simulations that showed the convergence of
the tracking error. A comparison with a high-order sliding-mode-observer-based con-
troller confirmed the superior performance of the controller using the robust observer
introduced in this study. Finally, the proposed controller was implemented on an actual
biped robot using an embedded hardware-in-the-loop strategy.
Keywords: Active Disturbance Rejection, Extended State Observers, Biped Robot,
Nonlinear Mechanical Systems, Disturbance Observers

1. Introduction

Biped locomotion is considered an active field in robotics and mechatronics [1].


The number of possible applications for biped robots is continuously increasing. Among
others, human interactive robots, military exoskeletons, gait cycle rehabilitation sys-
5 tems, and augmented reality systems have benefited from recent advances in designing,

Preprint submitted to ISA Transactions January 20, 2016


constructing, and controlling this type of robotic system.
Automatic control design for biped robots has received special attention because of
the wide variety of challenges that this task entails. Most of the existing literature in
this field has considered the problem of generating a stable dynamical walking cycle.
10 A large set of of different control strategies has been successfully proposed and tested
on different types of biped robots [2].
A different type of solution emerged recently in which the climbing, running, and
one-legged hopping problems have been tackled. These controllers consider different
paradigms in which the robot is modeled as a switched system affected by intense
15 perturbations. Moreover, the new class of controllers used to solve these problems is
adjusted to control biped robots in rough and/or difficult terrains.
Despite the nature of these new problems in control design for biped robots, more
general and more robust control laws are demanded. One popular approach to solve
this task divides the robot into smaller and simpler subsystems; the result is usually
20 called a decentralized control system [3]. However, many of these studies consider
the possibility of using a state feedback control form, which is not a realistic solu-
tion. Measuring the position and velocity of biped robots is not impossible, but it is
expensive and compromises their mechanical design.
Considering the complications in implementing state feedback controls as a feasi-
25 ble solution for biped robots, it is more natural to consider output-based control strate-
gies. Some remarkable solutions have been developed considering this possibility.
However, many of them require an exact and unperturbed mathematical representa-
tion of the biped robot. Others have used high-gain observers that neglect the partial
knowledge over the biped robot model.
30 One remarkable option for controlling biped robots is using sliding mode control
schemes [4]. In particular, finite-time estimation of the velocity of biped robot joints
has been solved recently [5]. These estimations are injected into a discontinuous con-
troller that also provides finite-time tracking of a reference trajectory that corresponds
to a regular dynamical gait cycle.
35 Another popular reported technique uses adaptive schemes (see [6], [7], and refer-
ences therein). These methods have shown excellent results, and, in some cases, ro-

2
bustness is included as part of the solution. Even when the response of these schemes
is quite competitive, the computational cost involved may be higher than that of other
classic approaches such as proportional integral derivative (PID) controllers.
40 One additional possibility is using control schemes based on active disturbance
rejection control (ADRC) ([8], [9]). This control philosophy has been proposed for
solving many different problems (see [10], [11], [12], [13], [14], [15]), as the complex-
ity in the control design can be reduced by lumping all the nonlinearities as well as
the external disturbances into a generalized term. Indeed, this is a strong motivation to
45 design such a class of robust decentralized controllers.
The ARDC method proposes to estimate the lumped term mentioned above. If
an accurate estimation is obtained, then it can be used in the control action that im-
plements a disturbance cancellation method. The lumped disturbance input can be
estimated by nonlinear and linear extended state observer schemes. Concerning linear
50 methods, in [16], a high-order extended state observer was proposed to solve the prob-
lem of estimating high-frequency sinusoidal disturbances in the context of increasing
the bandwidth of the estimator. This work offers a comprehensive frequency-domain
analysis of the use of high-order extended state observers by means of linear analysis.
On the other hand, using the same principle, [17], [18], [19], and [20] proposed a high-
55 order extended state Luenberger-like observer (defined as a generalized proportional
integral observer, or GPI) for the robust trajectory tracking problem for nonlinear dif-
ferentially flat systems [21]. The observer includes a time-velocity dating, linear model
approximation of the perturbation input.
Thus, ADRC is a new robust approach that takes advantage of the classic PID con-
60 trol strategy and modern control technology. A perturbation and nonlinearity approx-
imation is generally placed in the feedforward path, and an extended state observer
appears in the feedback path. These two parts are designed to obtain a high-quality
and accurate approximation of unknown disturbance signals without establishing an
accurate plant model.
65 Some studies have considered the approximation of the uncertain section of a biped
robot [22], [23]. These control designs use adaptive algorithms to obtain accurate
approximations, but they require exhaustive preliminary adjustments before they can

3
be used. Moreover, they require additional efforts to include new scenarios that were
been considered in the pre-adjustment procedure. Therefore, a self-adjusting scheme
70 seems to be a more promising option for approximating the uncertain section of a biped
robot. Nevertheless, it demands a more complicated approximation algorithm structure
with more robust control strategies.
The main contribution of this study is the design of an ADRC output-based mul-
tivariable controller that uses the estimation of the lumped uncertainties provided by
75 a linear extended state observer. This new approach to biped robotic systems allows
tracking of a reference trajectory that represents a standard gait cycle obtained by a
biomechanical study. A new formal analysis based on the Lyapunov stability theory
provides the conditions for obtaining the ultimate boundedness of the tracking devia-
tion as well as the estimation error; this analysis allows the estimation of a larger class
80 of disturbances because it is free of differentiability conditions. The estimation of the
lumped uncertainties is solved by a least mean square algorithm, which is also included
in the Lyapunov analysis. A set of numerical simulations is presented to demonstrate
the effectiveness of the proposed output-based controller, and real experiments using
a biped robot prototype are presented in which the proposed controller provides rapid
85 convergence with small steady-state errors. In addition, comparisons with a variable
structure observer-based control scheme are provided.
The remainder of the paper is organized as follows: the model structure of the
biped robot is given in section 2. Section 3 describes the problem formulation, and
section 4 presents some preliminary properties necessary to solve the ADRC design.
90 Then, in sections 5 and 6, the main theoretical contributions are presented, including
a novel Lyapunov-based analysis of the GPI observer-based control and an alternative
gain optimization procedure based on the invariant ellipsoid algorithm. The reference
trajectory planning criterion is given in section 7. The numerical results are shown
in section 8. Section 9 presents the experimental results. In both the numerical and
95 experimental results, a comparison test is provided. Finally, some concluding remarks
are presented.

4
2. Mathematical description of the biped robot

The robotic system considered in this study is presented in Figure 1. This robot
can move freely only in the x − y plane; it contains five links and is defined by seven
100 degrees of freedom. The corresponding seven coordinates are shown on the left side
h iT
of Figure 1. These are given by the set q = x0 y0 α βL βR γL γR , where
the coordinates (x0 , y0 ) correspond to the position of the torso center of mass; α is the
angle formed between the torso and the vertical axis; βL and βR are the angles formed
between the upper sections of the left and right legs, respectively, and the torso; and γL
105 and γR are the angles formed between the upper and lower sections of the left and right
legs, respectively.
The links’ lengths are denoted as l0 , l1 , l2 , and their masses are denoted as m0 , m1 , m2 .
The links’ centers of mass are located at the distances r0 , r1 , r2 from the corresponding
joints (see right side of Figure 1). The model is actuated with four moments given by
h iT
110 M = ML1 ML2 MR1 MR2 . Two of these moments act between the torso and
either thigh, and the other two act at the knee joints. The walking surface is modeled
h iT
using a set of external forces, denoted by F = FLx FLy FRx FRy , that affects the
leg tips. When the leg is touching the ground, the corresponding forces are switched to
support the leg. As the leg rises, the forces are zeroed.
115 Using the Euler–Lagrange modeling process, the dynamic equations for the biped
system can be derived. These equations can be represented as
d2q dq
= b(q, , M(q), F(t), τ (t)) (1)
dt 2 dt
Here M (q) ∈ Rn×n , n = 7 is the inertia matrix, and b(q, dq
dt , M(q), F(t), τ (t)) ∈ R
n

is a vector containing the right-hand sides of the seven differential equations. The
vector τ corresponds to the external torques applied over each articulation (hip, ankles,
120 and knees). The matrix M (q) and the vector b(q, dq
dt , M(q), F(t), τ (t)) are defined in
[24].
The right-hand side of the differential equations describing the biped robot can be
represented as follows:

dq dq
b(q, , M(q), F, τ ) = G(q) + R(q, )λ + τ (2)
dt dt

5
Figure 1: Biped robot diagram showing state variables as well as the forces involved in its movement.

where the matrix G(q) ∈ Rn is the gravity vector, the matrix R(q, dq 2n
dt ) : R → R
n×p

125 represents the restriction matrix, λ ∈ R p is the constraints vector (associated with the
Lagrange multipliers), and τ ∈ Rn is the input vector.

3. Control challenge in biped robots

The objective of the control design for the biped robot is defined as follows.
Problem Statement: Consider an alternative way of describing the mathematical
model of the biped robot (1) affected by any possible external perturbations and con-
sidering the presence of unmodeled dynamics (both lumped in the term ξ (q,t)):

d2q dq
= M −1 (q)τ (t) + ξ (q,t) − M −1(q)[G(q) + R(q, )λ ] (3)
dt 2 dt

Now, consider ψ (q,t), a vector that


 includes some of  the dynamical model terms
dq
130 taken as uncertain, such as −M −1 (q) G(q) + R(q, )λ + ξ (q,t), and some additive
dt
disturbance terms due to mechanical coupling in the gear train of the motors, or various

6
unmodeled perturbation terms. Then, the dynamic system (3) can be presented as

d2q
= M −1 (q)τ (t) + ψ (q,t) (4)
dt 2

Consider the alternative mathematical representation of the biped robot, and define
a set of desired articular positions and torso coordinates q∗ . Then the control chal-
135 lenge is to ensure convergence of the error between the articulation positions and the
reference (desired) trajectories in an ultimate boundedness sense (see [25] and [26] for
further information), despite the presence of the set of disturbances and uncertainties
lumped in ψ (according to Assumption 1). Formally, the control design must solve the
problem lim supt→∞ maxψ ∈Ψ kq(t) − q∗(t)k ≤ β , where β is a positive scalar that


140 must be proportional to the power of the uncertainties and perturbations affecting the
nominal model of the biped robot. A feasible control design must consider that only
position measurements can be used, and no velocity data are available.

4. Biped robot as interconnected uncertain second-order systems

The model (4) belongs to a class of coupled second-order nonlinear systems with
uncertain structure. On the basis of the state variable theory, the biped robot dynamic
system can be represented as follows:

dxa (t)
= xb (t)
dt
dxb (t)
= f (x) + g(xa (t))u(t) + ξ (x,t) (5)
dt

where x = [x⊤ ⊤ ⊤
a , xb ] , x ∈ R
2n is the state of the robotic system. Indeed, x = q and
a
dq
145 xb = dt . The initial condition of the system (5) is given by x(0) = x0 , kx0 k < ∞.
Let X be the set such that x ∈ X ⊂ R2n . The bounded function u ∈ Rn is referred to
as the control function. All the controls belong to the following admissible set:

Uadm = u : kuk2 ≤ u0 + u1kxk2 , u0 , u1 ∈ R+



(6)

The nonlinear Lipschitz function f : R2n → Rn is composed of n-uncertain nonlin-


ear functions that describe the drift term of (5).

7
The matrix function g : Rn → Rn×n is invertible and meets the following constraint:

0 < g− ≤ kg(·)k ≤ g+ g − , g + ∈ R+ (7)

150 As mentioned before, the term ξ (x,t) was included to consider the effects of per-
turbations/uncertainties that can also include, for example, parameter variations and
modeling errors. In this study, an important assumption regarding ξ (x,t) is consid-
ered:
Assumption 1: The class of uncertainties ξ (x,t) considered in this study satisfies
the following inequality:

kξ (x,t)k2Λ ≤ γ0 + γ1 kxk2 , ∀t ≥ 0
ξ
(8)
γ0 , γ1 ∈ R+ ; 0 < Λξ = Λ⊤
ξ ∈R
n×n

The system (5) can be rewritten as follows:


dx(t)
= Ax(t) + G(xa(t))u(t) + F(x,t) (9)
dt    
0n×n In×n 0
A=  , G(x) =  n 
0n×n 0n×n g(xa )
h i⊤
F(x,t) := 0⊤
n f ⊤ (x) + ξ ⊤(x,t)

The function F(x) satisfies the following inequality for all t ≥ 0: kF(x,t)k2 ≤
155 f0 + f1 kxk2 , f0 , f1 ∈ R+ . The following extra assumption regarding the function F(x)
is needed to propose the controller design.
Assumption 2: There is a constant vector a such that the function f (x) + ξ (x(t),t)
evaluated over the system trajectories, that is, x = x(t), can be represented as f (x) +
ξ (x,t) = a⊤ κ (x) + f˜(x,t), where the vector a ∈ R(p+1)×n is formed with a set of con-
160 stant parameters ak that must be adjusted to improve the approximation of f (x) +
ξ (x,t). According to the structure of the so-called GPI extended state observers, the
vector κ ∈ R p+1 is composed as follows: κ = [1,t, · · · ,t p ]. This is a regular decom-
position of f (x) + ξ (x,t) in terms of a finite number of elements that form a basis. In
particular, the last set of polynomials is considered in this study (see [27, 20] for further
165 details). The term f˜(x,t) is called the modeling error produced by the approximation
of f (x) + ξ (x,t) by a finite number p of elements in the basis.

8
The so-called nominal model a⊤ κ (t) can be represented as a⊤ κ (t) = a0 + a1t +
a2t 2 + · · · + a pt p . Each term a j ∈ Rn , j = 0, ..., p. This representation has been used
in various articles regarding the application of high-order extended observers [17].
However, this approximation method usually requires the differentiability of the term
f (x) + ξ (x,t), which is a strong restriction. In this study, an alternative to this problem
is proposed in which the function a⊤ κ (x(t)) can be represented as a chain of integra-
tors. Thus, the approximation presented above states that f (x) + ξ (x,t) must be the
solution of integration of an uncertain function plus the approximation error. This con-
dition relaxes the constraint that is usually a drawback in this class of extended state
observers. Thus, the previous equation is written as
Z t Z t Z τ1

a κ (t) = a0 + a1 d τ1 + 2a2 d τ2 d τ1 + · · ·+ (10)
τ1 =0 τ1 =0 τ2 =0
Z t Z τ p−1
+ ··· p!a pd τ p · · · d τ1
τ1 =0 τ p =0

The last equation can be expressed in differential form:

a⊤ κ (t) = ρ0 (t) (11)


d ρi (t)
= ρi+1 (t), i = 0, . . . , p − 1
dt
d ρ p (t)
=0 ρ j (0) = a j , j = 0, 1, ..., p
dt

Now the problem formulation given above can be rephrased as follows. Given an
output reference trajectory x∗ for the system (5), design an output feedback controller
that, regardless of the unknown unmodeled dynamics or external disturbances (both
170 lumped in an additive signal F(x)), forces the states x to track asymptotically the de-
sired reference trajectories, with the tracking error restricted to a small neighborhood
near the origin and proportional to a power of the uncertainties and perturbations. The
first stage in solving this problem is designing an extended state observer to reconstruct
the unmeasurable part of the state.

9
175 5. The high-order extended state observer

The extended state observer for the system (5) [with the corresponding feedback
controller using the observer states u(t) := u(x̂(t))] is proposed as
d x̂(t)
= Ax̂(t) + G(xa (t))u(t) + Bρ̂0(t) + He(t) (12)
dt
d ρ̂ (t)
= Φρ̂ (t) + Le(t)
dt
 
0n×n In×n 0n×n · · · 0n×n  
 


0n×n 0n×n In×n · · · 0n×n 
 


 .. .. .. .. 
 
Φ=  . . . · · · .  p − times
 
 

0n×n 0n×n 0n×n · · · In×n  
 



0n×n 0n×n 0n×n · · · 0n×n 

ρ̂0 (t) = D⊤ ρ̂ (t); e(t) = xa (t) − C⊤ x̂(t)

Here the vectors x̂ and ρ̂ are the estimated states for x and ρ , respectively. The ma-
trix H ∈ R2n×n is the state observer gain. The matrix L ∈ Rn(p+1)×n is the extended ob-
server gain. The matrix D ∈ Rn(p+1)×p is represented as D⊤ = [In×n 0n×n · · · 0n×n ].
The matrix Φ ∈ Rn(p+1)×n(p+1) is called the extended state self matrix. The output
180 matrix C ∈ R2n×n is given by C⊤ = [In×n 0n×n ]. Finally, the matrix B ∈ R2n×n is rep-
resented as B⊤ = [0n×n In×n ].
On the basis of the extended state observer proposed in (12), the output-based con-
troller can then be designed using the following structure:
h i
u(t) = −g(xa (t))−1 K ⊤ σ (x̂(t), x∗ (t)) + ρ̂0 (t) − s(x∗ (t)) (13)

The variable σ is the tracking error, σ := x̂−x∗ . The vector x∗ ∈ R2n is the reference
trajectory and satisfies
dx∗ (t)
= Ax∗ (t) + Bs(x∗ (t)), x∗ (0) = x∗0 ∈ R2n (14)
dt
where s : Rn → Rn is a Lipschitz function. The reference trajectory is designed to be
bounded ∀t ≥ 0; that is, kx∗ (t)k2 ≤ [x∗max ]2 , x∗max ∈ R+ .
The controller gain matrix K is designed such that the matrix A − BK is Hurwitz.
To solve this part of the problem, by using Ackermann’s approach [28], the matrix K

10
can be obtained as

K = [0n×n 0n×n 0n×n · · · In×n] S−1ν (A) (15)

S = B BA BA2 · · · BAn
 

ν (A) = An + a1An−1 + · · · + anIn×n

The main result of this study is developed as follows.

Theorem 1. Consider the state observer given in (12) and the output feedback con-
troller proposed in (13) with the gain adjusted according to (15) for the class of nonlin-
ear uncertain systems (5) fulfilling the condition (7) with incomplete information and
affected by perturbations that obey the constraints given in (8). Define the matrices
Π(H, L, K), R, and Q described by
 
A − HC⊤ BD⊤ 02n×2n
 
Π(H, L, K) =  −LC
 ⊤ Φ 0n(p+1)×n(p+1)

 (16)
 
HC⊤ 02n×2n A − BK
 
2f I 0 0
 1 
R = Λ+ 0 (17)
 
0 0 
 
0 0 4 f1 I

185 where Λ = Λ⊤ ∈ RNp ×Np , Λ > 0, Q = Q0 , Q0 = Q⊤


0 ∈R
N p ×N p , Q > 0, and N =
0 p

4n + n(p + 1). If there is a positive definite matrix Q0 such that the algebraic Riccati
equation Ric(P) = 0 with

Ric(P) = PΠ(H, L, K) + Π(H, L, K)⊤ P + PRP + Q (18)

has a positive definite symmetric solution P ∈ RNp ×Np , then the tracking trajectory
f 0 +4 f 1 [x∗max ]+
error ε = x − x∗ is ultimate bounded with an upper bound given by β = 2 αQ ,
αQ := λmin P−1/2Q0 P−1/2 .

190

Proof. The estimation error ∆ satisfies

d∆(t) h i
= A − HC⊤ ∆(t) + F(x(t)) − Bρ̂0(t) (19)
dt

11
The approximation proposed in (11) makes it possible to transform the previous
differential equation into
d∆(t)
= [A − HC⊤]∆(t) + B f˜(x(t),t) + BD⊤ [ρ (t) − ρ̂ (t)] (20)
dt
On the other hand, the extended state error, defined as δ = ρ − ρ̂ , and the tracking
error σ satisfy
d δ (t)
= Φδ (t) − LC⊤ ∆(t) (21)
dt
d σ (t)
= Ax̂(t) + G(xa (t))u(t) + Bρ̂0(t) + He(t) − s(x∗(t)) (22)
dt
If the control action described by (13) is substituted in the previous differential
equation, one gets
d σ (t)
= (A − BK)σ (t) + HC⊤∆(t) (23)
dt
To prove that the equilibrium point of (20) and (21) is stable, the following Lya-
punov function candidate is proposed:

V (z) = z⊤ Pz (24)

where the vector z ∈ RNp is constructed as z⊤ = [∆⊤ δ ⊤ σ ⊤ ]. Taking the Lie derivative
dV (t)
of V (z) with respect to time, one has dt = 2z⊤ (t)P dz(t)
dt .
From (20) and (21), the time derivative of z is
   
A − HC ⊤ BD ⊤ 02n×2n ˜
B f (x(t),t)
dz(t)  ⊤
  
=  −LC Φ 0n(p+1)×n(p+1) z(t) +  0n(p+1)  (25)
   
dt    
HC⊤ 02n×2n A − BK 0n

Using this result in the previous differential equation, one gets


dV (t)
= 2z⊤ (t)PΠ(H, L, K)z(t) + 2z⊤ (t)Pη (t) (26)
dt
where η represents all the internal uncertainties and external perturbations.
The direct application of the matrix inequality X ⊺Y + Y ⊺ X ≤ X ⊺ NX + Y ⊺ N −1Y ,
which is valid for any X,Y ∈ Rr×s and any 0 < N = N ⊺ ∈ Rs×s [29], yields
dV (t)  
≤ z⊤ (t) PΠ(H, L, K) + Π(H, L, K)⊤ P + PΛP z(t) + η ⊤ (t)Λ−1 η (t) (27)
dt

12
An algebraic manipulation involving the term z⊤ (t)Q0 z(t) yields
dV (t)
≤z⊤ (t)(PΠ(H, L, K) + Π(H, L, K)⊤ P + PΛP + Q0 )z(t)
dt
+ η ⊤ (t)Λ−1 η (t) − z⊤ (t)Q0 z(t) (28)

By taking the upper bound presented in (8), the term η ⊤ Λ−1 η for all t ≥ 0 is
bounded from above by

η ⊤ Λ−1 η ≤ f0 + 2 f1 k∆k2 + 4 f1 kσ k2 + 4 f1 kx∗ k2 (29)

On the basis of this inequality, the differential inclusion presented in (28) is trans-
formed into
dV (t)
≤ z⊤ (t)Ric(P)z(t) − z⊤ (t)Q0 z(t) + β0 (30)
dt
where β0 = f0 + 4 f1 [x∗ ]+ . Using the assumption regarding the existence of a positive
definite solution to the Riccati equation Ric(P) = 0, the previous differential inclusion
is modified to
dV (t)
≤ −z⊤ (t)Q0 z(t) + β0 (31)
dt
The right-hand side of the previous differential inclusion can be bounded from
above as
dV (t) n o
= −λmin P−1/2 Q0 P−1/2 z⊤ (t)Pz(t) + β0 (32)
dt
The last inequality was obtained using the so-called Rayleigh inequality and the
Cholesky decomposition [29]. According to the definition of β , one has the following
differential inequality:
dV (t)
≤ −αQV (t) + β0 (33)
dt
Taking the equality in the previous differential inclusion, one has
dV eq
= −αQV eq + β0 (34)
dt
The solution of the previous differential equation is
β0
V eq (t) = e−αQt V eq (0) + (1 − e−αQt ) (35)
αQ

13
By applying the comparison principle [25] to the previous equation and using the
fact that V (t) ≥ 0, one has

β0
V (t) ≤ e−αQt V (0) + (1 − e−αQt ) (36)
αQ
Taking the upper limit of the previous inequality, the main result presented in the
195 theorem is finally obtained.

6. Optimization of convergence region

The ultimate boundedness (practical stability) condition obtained in the previous


theorem is a consequence of the unmatched uncertainties and perturbation effect. The
so-called invariant ellipsoid technique [30, 31] was applied recently to minimize the
200 size of the region characterized by the ultimate boundedness condition. This method
uses the concept of an ellipsoid:

Definition 2. [31] The ellipsoid


n o
ε (P) = z ∈ Rn : z⊤ Pz ≤ 1, P > 0 (37)

with the center at the origin and the configuration matrix P is said to be state-invariant
for the system (25) with respect to the uncertainties η if

1) The initial condition z(0) ∈ ε (P) implies z(t) ∈ ε (P) for all t ≥ 0.

205 2) The initial condition z(0) 6= ε (P) implies z(t) → ε (P) as t → ∞.

The possibility of finding a possible invariant ellipsoid with the largest matrix P
yields the minimal deviation of all the possible trajectories from the origin. This fact
is easily understood because of the inequality presented in (37), but it also requires a
measure or criterion to define the size of the matrix P. In this case, the trace operator
is used as the criterion, that is, tr {P} → max or tr P−1 → min.

210

In the theorem presented above, the existence of the invariant set in which the tra-
jectories of z converge has already been proven. According to the definition presented
in (37), the configuration matrix Pα = αQ P/β satisfies the invariant ellipsoid concept
if one considers the inequality (36) and the application of the upper limit when t → ∞.

14
215 To obtain the solution of the minimal invariant ellipsoid, the following lemma gives the
sufficient conditions for designing the gain controller.

Lemma 3. If the tuple (ν1 , ν2 , αQ ,P) is a solution of the optimization problem tr {P} →
max subject to (8) and

 αQ 
PA + A⊤P + ν2 P P
β0
1 −1  ≤ 0 (38)
 

P −ν1 Λ
β0
then the corresponding controller proposed in (13) guarantees that any trajectory of
220 the system (5) converges to a quasi-minimal ellipsoid ε (P).

Proof. Consider the Lyapunov candidate function used in Theorem 1 and given in (24).
dV (t)
The ellipsoid ε (P) is invariant for the trajectories of z if and only if dt ≤ 0 holds for

any z such that z Pα z ≥ 1 and for any uncertainties η satisfying the condition (8).
The expression in (26) can be represented as
 ⊤    
dV (t)  z(t)  z(t) PA + A⊤P P
≤ Ω0   , Ω0 =   (39)
dt η (t) η (t) P 0N×N

An algebraic manipulation makes it possible to express the inequality (29) as


 ⊤    −4 f 
1
z z IN×N 0N×N
  Ω1   ≤ 1, Ω1 =  0
 β 
(40)
1 −1 
η η 0N×N Λ
β0

The condition z⊤ (t)Pα z(t) ≥ 1 can be represented as


 ⊤    
z(t) z(t) −P 0N×N
  Ω2   ≤ −1 Ω2 =  α  (41)
η (t) η (t) 0N×N 0N×N

According to the S-procedure [29], the previous matrix inequalities imply that (39)
225 is negative if and only if there exist two scalars ν1 and ν2 such that Ω0 < ν1 Ω1 +
ν2 Ω2 , ν1 − ν2 ≤ 0.
Finally, the Schur complement provides the inequality in (38).

15
7. Reference trajectories

The reference trajectories proposed in (14) were obtained by a biomechanical anal-


230 ysis of regular gait cycles. The angles observed for the ankles, knees, and hips were
reported in several studies [32], [33]. This information was collected from several arti-
cles and specialized books. These angles are usually reported at very specific moments
of the gait cycle. Therefore, these values were interpolated using a nonlinear algo-
rithm based on a least mean square approximation that uses cubic polynomials. This
235 algorithm is already implemented in MATLAB (the interp1 function with a cubic in-
terpolation method). The corresponding function s(·) was obtained after differentiating
the third-order polynomials obtained by this procedure.

8. Numerical results

A set of numerical simulations was developed in MATLAB for testing the robust
240 observer and the controller. The numerical simulations were executed using the fixed
step fourth-order Runge–Kutta method with a simulation step of 0.005 s. The biped
robot simulation used the MATLAB model Bipedsim proposed by [24]. This system
was used to simulate the dynamics of a biped robot. The simulation was executed using
the licensed file provided by the original authors of the study. However, the original
245 controller based on a proportional derivative scheme was modified to include ADRC.
The biped robot parameters required by [24] and used for the numerical simulations
are the torso length (l0 = 0.8 m), thigh length (l1 = 0.5 m), and shank length (l2 = 0.5
m). The distances with respect to the center of mass of the torso (r0 ) and the thigh (r1 )
from the hip are r0 = 10 /2 and r1 = l1 /2, respectively. The distance from the knee
250 to the center of mass of the shank is r2 = l2 /2. The masses of the torso, thigh, and
shank are m0 = 5 kg, m1 = 2 kg, and m2 = 1 kg, respectively. In addition, the link
inertias of the torso, thigh, and shank are given by J0 = (1/12)m0l02 , J1 = (1/12)m1l12 ,
and J2 = (1/12)m2 l22 , respectively, in units of kgm2. The gravitational acceleration is
g = 9.81 m/s2 . The initial conditions for the positions (velocities) of the system are
255 x0 (0) = 0 m (0.1 m/s), y0 (0) = 1.3749 m (−0.05 m/s), a(0) = 0.04 rad (−0.06 rad/s),

16
bl (0) = −0.11 rad (−0.6 rad/s), cl (0) = 0.1696 rad (−0.06 rad/s), and cr (0) = 0.1616
rad (0.0 rad/s).
The estimated information obtained by the observer was injected into the controller
structure introduced in (13). The control and observer parameters are given by H =
h iT h iT
46 80 60 25 30 25 40 , L = 7.5× 25 25 25 35 35 35 35 , and
h iT
K = 6 × 15 15 15 25 25 25 25 . The optimization parameter obtained
from the Riccati equation is
 
1.715 −0.291 0.001 −0.040 −0.043 −0.080 0.464
 
−0.291 1.2526 −0.1697 0.0049 0.0008 0.0001 −0.025 
 
 
 0.001 −0.1697 1.3312 −0.1097 0.0001 −0.0063 0.0816 
 
 
Pα = −0.040 −0.1097 −0.1064 −0.0214 0.3442  .
 
0.0049 1.1720
 
−0.043 −0.1064 −0.1785
 
0.0008 0.0001 0.8959 0.3078 
 
 
−0.080 0.0001 −0.0063 −0.0214 −0.1785 1.1559 −0.098 
 
0.464 −0.025 0.0816 0.3442 0.3078 −0.098 12.5111
The simulation was evaluated in two steps. The first one considers only the observer
performance. The observer trajectories converged to a bounded zone around the actual
260 trajectories of (5). This performance was obtained without any knowledge of the drift
section obtained from the biped robot model.
Figure 2 compares the actual velocity measured at the left knee of the biped robot
versus two estimated trajectories. The first was produced by the robust observer de-
scribed in (12), whereas the second was produced by a multivariable super-twisting
265 algorithm (ST) used as a robust differentiator [34], [35].
The green solid line corresponds to the actual angular velocity of the biped robot
knee. The blue dashed line depicts the corresponding estimated velocity of the same
articulation according to the robust observer proposed in this study. The red dashed
line describes the velocity obtained by applying the ST.
270 Even when both trajectories seem to reproduce the actual angular velocity of the left
knee at the same time, the observer proposed in this study does not use discontinuous
functions. This feature yields a relative advantage because no high-frequency oscil-
lations appear in the velocity estimation. This behavior reportedly affects the output-
based controller energy [36].

17
275

Comparison of actual angular velocity of left knee


1.5
Angular vel

1
(rad/s)

0.5
Biped System
ST
0 GPI
0 1 2 3 4 5 6 7
Time (s)
Comparison of actual angular velocity of left knee: Zoom
1.5
Angular vel

1
(rad/s)

0.5
Biped System
0 ST
GPI
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)
Comparison of MSE of the whole estimation error
1
Estimation error

0.8

0.6

0.4

0.2 GPI
ST
0
0 0.5 1 1.5 2 2.5 3
Time (s)

Figure 2: Top and center: Comparison of actual angular velocity of left knee (green solid line) with that
obtained using the proposed observer (blue dashed line) and the ST used as a differentiator (red dashed line).
Bottom: Comparison of mean square error (MSE) of the entire estimation error obtained by the proposed
observer (blue solid line) and the ST used as a differentiator (red dashed line).

The observer proposed in this study exhibited better velocity estimation for all
the articulations of the simulated biped robot. These comparisons are typically made
using the mean square error (MSE) for the observation error, which corresponds to
MSE = k∆k. This value was obtained both when the observer (12) was applied and
280 when the ST was used. The results are shown in Figure 2. The superiority observed

18
in each comparison was confirmed by the MSE value. This value showed that the esti-
mation error obtained by the robust observer was closer to zero than that of the ST, as
shown by the ultimate bound obtained after 7 s (0.01 vs. 0.04). The same controller,
(13), was implemented using the information provided by either the robust observer
285 or the ST. This strategy was useful for comparing the effect of the estimation quality
on the tracking error. Although the controller structure used the estimated velocity, it
successfully tracked the reference trajectory that was based on the regular gait cycle.
However, the controller based on the robust observer produced better tracking perfor-
mance than the one that used the ST-estimated velocities. Figure 3 shows the tracking
290 comparison reference trajectory proposed for the left knee of the biped robot vs. the
two controlled trajectories. Again, the green solid line corresponds to the reference an-
gular trajectory of the biped robot knee. The blue dashed line depicts the corresponding
angle of the same articulation based on the robust observer proposed in this study. The
red dashed line describes the result of applying ST as the velocity estimator over the
295 controller structure.
The MSE of the tracking process (MSEtr ), which corresponds to MSE = k∆ +
σ k, was determined. Again, this value was obtained both when the observer (12)
was applied and when the ST was used as a velocity estimator to feed the controller
structure. The results are shown in Figure 3. The MSE values confirm that the tracking
300 error of the robust observer approached zero more rapidly than that of the ST (0.3 vs.
0.5 s). The corresponding ultimate bound of the tracking process was 4 times better at
a time of 7 s (0.09 vs. 0.41) for the controller that used the estimated velocities and the
uncertain section of the system.

19
Comparison of reference angle proposed for the left knee

Angle (rad)
8 Biped System
6 ST
GPI
4
0 1 2 3 4 5 6 7
Time (s)
Angle (rad) Comparison of reference angle proposed for the left knee: Zoom
6
Biped System
5.5 ST
GPI
5
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)
Comparison of MSE of the whole tracking error
Angle (rad)

3
2
GPI
1
ST
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)

Figure 3: Top and center: Comparison of reference angle proposed for left knee (green solid line) with the
trajectories obtained using the proposed observer as a velocity estimator (blue dashed line) and using the ST
as a differentiator (red dashed line). Bottom: Comparison of MSE of the entire tracking error caused by the
proposed observer (blue solid line) and the ST used as a differentiator (red dashed line).

Figure 4 compares the reference states and controlled trajectories produced by the
305 PID controller using the states estimated by a set of STs as well as the GPI controller
using the proposed observer. In all cases, the GPI controller exhibited better tracking
of the reference trajectories.

20
Ang. vel. (rad/s)
1.4

Angle (rad)
0.4
1.38 0.2
0
1.36
0 2 4 6 0 2 4 6
a) Time (s) d) Time (s)

Ang. vel. (rad/s)


Angle (rad)

0.1
0.05 0.5
0
-0.05 0
0 2 4 6 0 2 4 6
b) Time (s) e) Time (s)

Ang. vel. (rad/s)


Angle (rad)

0.4
0.2 0.5
0
-0.2
0
0 2 4 6 0 2 4 6
c) Time (s) f) Time (s)

Figure 4: Comparison of reference states (green solid lines) and controlled trajectories produced by the PID
controller using the states estimated by a set of STs (red dashed lines) as well as the GPI controller using the
high-gain observer (blue dashed lines). All the sections of the biped robot are shown: a) left hip, b) left knee,
c) right knee, d) right hip, e) left ankle, f) right ankle.

Figure 5 shows the numerical results in such a way that the anatomical posture
can be appreciated. This figure shows the behavior of both the proposed method and
310 the ST. The comparison confirms the enhanced tracking performance of the proposed
method introduced in this study. The complete reproduction of the reference gait cycle
is more accurate and faster when the robust observer is used.

21
T=0.17 s T=0.41 s
0.3 0.3 Reference

Y (m)

Y (m)
0.2 0.2 GPI
0.1 0.1 ST
0 0
0 0.04 0.08 0.12 0.16 0.2 0.24 0 0.04 0.08 0.12 0.16 0.2 0.24
X (m) X (m)
T=0.65 s T=1.05 s
0.3 0.3
Y (m)

Y (m)
0.2 0.2
0.1 0.1
0 0
0 0.04 0.08 0.12 0.16 0.2 0.24 0 0.04 0.08 0.12 0.16 0.2 0.24
X (m) X (m)
T=2.01 s T=2.97 s
0.3 0.3
Y (m)

Y (m)
0.2 0.2
0.1 0.1
0 0
0 0.04 0.08 0.12 0.16 0.2 0.24 0 0.04 0.08 0.12 0.16 0.2 0.24
X (m) X (m)
T=3.29 s T=3.53 s
0.3 0.3
Y (m)

Y (m)
0.2 0.2
0.1 0.1
0 0
0 0.04 0.08 0.12 0.16 0.2 0.24 0 0.04 0.08 0.12 0.16 0.2 0.24
X (m) X (m)

Figure 5: Comparison of the reference postural evolution tracking of the simulated biped robot (reference in
blue lines) obtained by the proposed observer (red lines) and the ST used as a differentiator (green lines).

9. Hardware-in-the-loop implementation

The experiment was executed using an adaptation of the robot “Brat” produced by
315 the Lynxmotion company (see Figure 6). The robot had the following specifications
originally:

• Number of degrees of freedom per leg: 3

• Servo motion control by local closed loop

• Total height: 8.5 in.

320 • Total width: 6.0 in.

• Ground clearance: 2.5 in.

• Weight (without batteries): 22.2 oz

• Range of motion per axis: 180◦

22
• Voltage input for each motor: 6 VDC

Figure 6: Experimental robotic system.

325 The original servo-motors (HITEC HS-422) in the robot were modified to serve
as simple DC motors. Each motor was driven by an external power DC motor driver
(L293D). The angle of each articulation was measured by a single-turn potentiome-
ter (Suntan, TSR-3590) of 10 KΩ. The output-based controller was implemented in a
hardware-in-the-loop scheme. Simulink/MATLAB was connected to an Arduino UNO
330 board through the Arduino IO Library. Each driver was connected to an optoisolation
system based on 4N35 circuits. A pair of Arduino UNO boards was attached to the
set of optoisolators. These boards implemented only a data acquisition board emula-
tor. The information acquired by the boards was fed into a Simulink/MATLAB-based
simulation system. This simulation system implemented both the high-order extended
335 state observer and the subsequent automatic controller. The gains of the observer and

23
the controller were not adjusted with respect to those reported in the numerical section.
No information from the actual robot was used to design the observer.
The controller action produced by the Simulink program was adjusted (normal-
ized to 5.0 V) in order to use the analog output block from the Arduino IO library.
340 This information was combined with an additional digital output used to characterize
the movement direction of each articulation. All the experimental conditions and the
reference trajectories proposed in this study yield a robot step size of 5 cm.
As in the numerical simulation, the controller structure tracked the reference tra-
jectory that was proposed on the basis of the regular gait cycle (Figure 7). In this case,
345 only the tracking performance of the controller using the estimated velocities from the
robust observer was evaluated. The same figure shows the tracking comparison ref-
erence trajectory proposed for the left knee of the biped robot versus the closed loop
controlled trajectories. The solid line corresponds to the reference biped robot knee
angular trajectory. The dashed line depicts the corresponding angle of the same articu-
350 lation based on the robust observer proposed in this study.
Angle, [rad]

ST Ref GPI
0.6
0.4
0.2
0 2 4 6
Time, (s)
0.2
ST GPI
0.1
MSE

0
-0.1
0 2 4 6
Time, (s)

Figure 7: Upper figure: Comparison of reference angle proposed for the left knee articulation in the exper-
imental biped robot (red solid line) with the trajectory obtained using the proposed observer as a velocity
estimator (blue lines) and the ST used as a differentiator (green lines). These trajectories were measured di-
rectly from the instrumented robot presented above. Bottom figure shows the mean square error of estimation
generated by the GPI observer (Red lines) as well as the generated by the ST algorithm (blue lines).

24
The MSE of the tracking process obtained in the actual robot (MSEtr,r ) was calcu-
lated. In this case, this value was obtained only when the observer (12) was applied as
velocity estimators to feed the controller structure. The result is shown in Figure 7.
Two additional reference trajectories were proposed to evaluate the controller per-
355 formance. These trajectories forced the biped robot to sit down and stand up. The
results of the controller action were similar to those obtained when the regular gait cy-
cle was used as a reference trajectory. These results are not shown to avoid repetition
of the already reported results.

10. Conclusions

360 In this paper, we propose an active disturbance rejection multivariable control mixed
structure based on a robust observation and output-based controller to adjust the move-
ment of biped robot articulations. This mixed structure was useful for controlling the
class of uncertain nonlinear systems represented by the biped robot. Convergence of
this controller was ensured by a simple quadratic Lyapunov function. The last proof
365 contributes to pioneering work in the area of high-order extended state observers for
ADRC, such as those by [16] and [20].The novelty in this work is the fact that the set
of disturbance inputs does not need to be differentiable, which allows a larger set of
admissible disturbance inputs to be estimated.
A regular linear feedback controller was designed to show how the observer states
370 can be used to replace the real variables of an uncertain system. A basic representation
of the biped robot was used to generate a set of numerical simulations that validated
the theoretical result achieved in this study. The controller provided the solution to the
trajectory tracking problem. As future work, the observer gain may be tuned in the
context of low or adaptive gain control, which is an alternative for a class of constraint
375 systems, especially in actuators or when the working conditions may vary with respect
to time. On the other hand, the observer can be modified in the context of low-gain
control, which may have a remarkable impact on energy saving, a demanding open
problem in the active prosthesis field.
Even when the proposed controller can handle a wide variety of systems, there are

25
380 still problems to consider when using linear extended state observers for active distur-
bance rejection, where Lyapunov analysis can be used and extended. For example, this
scheme can be improved for dealing with a set of nonlinearities in the control input
such as dead zones and hysteresis. These effects can be reduced if further actions are
introduced in the proposed controller, from adaptation schemes to classical procedures.
385 On the other hand, adaptive forms in the gain tuning can be applied for energy opti-
mization, which is relevant for the class of mobile robotic systems representing a biped
robot. For this purpose, alternative procedures such as implicit Lyapunov functions
can be introduced. These schemes are mentioned in particular because the proposed
biped robot control design can be improved by introducing such actions. Because the
390 class of systems to consider can be subject to additive noise effects, a further develop-
ment would consider these effects in the design process or in the system structure (see
[37] for an authoritative analysis of the effects of output noise on high-gain observers).
Additional improvement can be obtained if the linear controller action proposed in
this study can be combined with some type of variable structure forms such as sliding
395 modes. This mixed strategy can enlarge the type of perturbations that can be handled
by our proposed method.

References

[1] S. Kajita, T. Yamaura, A. Kobayashi, Dynamic walking control of a biped robot


along a potential energy conserving orbit, IEEE Transactions on Robotics and
400 Automation 8 (4) (1992) 431–438.

[2] Z. Liu, C. Li, Fuzzy neural networks quadratic stabilization output feedback con-
trol for biped robots via h∞ approach, IEEE Transactions on Systems Man and
Cybernetics, Part B: Cybernetics 33 (1) (2003) 67–84.

[3] M. Mahmoud, Decentralized systems with design constraints, Springer, 2011.

405 [4] V. Lebastard, Y. Aoustin, F. Plestan, Observer-based control of a walking biped


robot without orientation measurement, Robotica 24 (03) (2006) 385–400.

26
[5] H. Oza, Y. Orlov, S. Spurgeon, Y. Aoustin, C. Chevallereau, Finite time tracking
of a fully actuated biped robot with pre-specified settling time: a second order
sliding mode synthesis, in: Proceedings of the IEEE International Conference on
410 Robotics and Automation (ICRA 2014), Hong Kong, China, 2014.

[6] S. Ge, Z. Li, H. Yang, Data driven adaptive predictive control for holonomic
constrained under-actuated biped robots, IEEE Transactions on Control Systems
Technology 20 (3) (2012) 787–795.

[7] Z. Li, S. Ge, Adaptive robust controls of biped robots, IET Control Theory &
415 Applications 7 (2) (2013) 161–175.

[8] J. Han, From PID to active disturbance rejection control, IEEE Transactions on
Industrial Electronics 56 (3) (2009) 900–906.

[9] Y. Huang, W. Xue, Active disturbance rejection control: methodology and theo-
retical analysis, ISA transactions 53 (4) (2014) 963–976.

420 [10] X. Chang, Y. Li, W. Zhang, N. Wang, W. Xue, Active disturbance rejection con-
trol for a flywheel energy storage system, IEEE Transactions on Industrial Elec-
tronics 62 (2) (2015) 991–1001. doi:10.1109/TIE.2014.2336607.

[11] L. Dong, Y. Zhang, Z. Gao, A robust decentralized load frequency controller for
interconnected power systems, ISA transactions 51 (3) (2012) 410–419.

425 [12] Z. Gao, Active disturbance rejection control: A paradigm shift in feedback control
system design, in: American Control Conference, Minneapolis, Minnesota, USA,
2006, pp. 2399–2405.

[13] F. Goforth, Q. Zheng, Z. Gao, A novel practical control approach for rate inde-
pendent hysteretic systems, ISA transactions 51 (3) (2012) 477–484.

430 [14] C. D. Johnson, Accommodation of external disturbances in linear regulator and


servomechanism problems, IEEE Transactions on Automatic Control AC-16 (6)
(1971) 1259–1369.

27
[15] S. Zhao, Z. Gao, An active disturbance rejection based approach to vibration
suppression in two-inertia systems, Asian Journal of Control 15 (2) (2013) 350–
435 362.

[16] A. Godbole, J. Kolhe, J. Jaywant, S. Talole, Performance analysis of generalized


extended state observer in tackling sinusoidal disturbances, IEEE Transactions on
Control Systems Technology 21 (6) (2013) 2212–2223.

[17] H. Sira-Ramı́rez, M. Ramirez-Neria, A. Rodrı́guez-Angeles, On the linear control


440 of nonlinear mechanical systems, in: Decision and Control (CDC), 2010 49th
IEEE Conference on, IEEE, 2010, pp. 1999–2004.

[18] R. Morales, H. Sira-Ramı́rez, J. Somolinos, Robust control of underactuated


wheeled mobile manipulators using gpi disturbance observers, Multibody Sys-
tem Dynamics (2013) 1–23.

445 [19] J. Cortés-Romero, G. A. Ramos, H. Coral-Enriquez, Generalized proportional


integral control for periodic signals under active disturbance rejection approach,
ISA transactions 53 (6) (2014) 1901–1909.

[20] H. Sira-Ramı́rez, J. Linares-Flores, C. Garcı́a-Rodrı́guez, M. Contreras-Ordaz,


On the control of the permanent magnet synchronous motor: An active distur-
450 bance rejection control approach, IEEE Transactions on Control Systems Tech-
nology 22 (5) (2014) 2056–2063.

[21] M. Fliess, J. Lévine, P. Martin, P. Rouchon, Flatness and defect of non-linear


systems: introductory theory and examples, International journal of control 61 (6)
(1995) 1327–1361.

455 [22] C. Zhou, Q. Meng, Dynamic balance of a biped robot using fuzzy reinforcement
learning agents, Fuzzy sets and systems 134 (1) (2003) 169–187.

[23] T. Li, Y. Su, S. Liu, J. Hu, C. Chen, Dynamic balance control for biped robot
walking using sensor fusion, kalman filter, and fuzzy logic, IEEE Transactions on
Industrial Electronics 59 (11) (2012) 4394–4408.

28
460 [24] H. Havisto, Hytyniemi, Simulation tool of a biped walking robot model, Tech.
rep., Helsinki University of Technology, Control Engineering Laboratory (2004).

[25] H. Khalil, Nonlinear Systems, Prentice Hall, Upper Saddle River, NJ, USA, 2002.

[26] S. Li, J. Shihua, J. Yang, W. Chen, X. Xisong, Generalized extended state ob-
server based control for systems with mismatched uncertainties, IEEE Transac-
465 tions on Industrial Electronics 59 (12) (2012) 4792–4802.

[27] H. Sira-Ramı́rez, C. López-Uribe, M. Velasco-Villa, Linear observer-based active


disturbance rejection control of the omnidirectional mobile robot, Asian Journal
of Control 15 (2013) 51–63, doi: 10.1002/asjc.523.

[28] J. Ackermann, V. Utkin, Sliding mode control design based on ackermann’s for-
470 mula, IEEE Transactions on Automatic Control 43 (2) (1998) 234–237.

[29] A. Poznyak, Advanced mathematical tools for automatic control engineers, Vol.
I. Deterministic Systems, Elsevier, Amsterdam, 2008.

[30] A. B. Khurzhanski, P. Varaiya, Ellipsoidal techniques for reachability under state


constraints, SIAM Journal on Control and Optimization 45 (2006) 1369–1394.

475 [31] B. T. Polyak, S. A. Nazin, C. Durieu, E. Walter, Ellipsoidal parameter or state


estimation under model uncertainty, Automatica 40 (7) (2004) 1171–1179.

[32] S. Cuccurullo, Physical Medicine and Rehabilitation Board Review, Demos Med-
ical Publishing, 2004.

[33] L. Luenberger, G. Colombo, R. Riener, V. Dietz, Biofeedback in gait training with


480 the robotic orthosis lokomat, in: IEMBS 26th Annual International Conference of
the IEEE Engineering in Medicine and Biology Society, Vol. 2, 2004, pp. 4888–
4891.

[34] J. Moreno, M. Osorio, Strict Lyapunov functions for the super-twisting algorithm,
IEEE Transactions on Automatic Control 57 (4) (2012) 1035–1040.

29
485 [35] J. Moreno, Lyapunov analysis of non homogeneous super-twisting algorithms, in:
Proceedings of the 11th International Workshop on Variable Structure Systems
(VSS), 2010, pp. 534–539.

[36] A. N. Atassi, H. Khalil, Separation results for the stabilization of nonlinear sys-
tems using different high-gain observer design, Systems and Control Letters 39
490 (2000) 183–191.

[37] J. Ahrens, H. Khalil, High-gain observers in the presence of measurement noise:


A switched-gain approach, Automatica 45 (4) (2009) 936–943.

30
LaTeX Source Files
Click here to download LaTeX Source Files: ADRC_biped_blinded_r3cbib.tex

You might also like