IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 31, NO.

3, MARCH 2023 799

Whole-Body Control of an Autonomous Mobile

Manipulator Using Model Predictive Control
and Adaptive Fuzzy Technique
Wang Yuan , Yong-Hua Liu , Chun-Yi Su , Senior Member, IEEE, and Feng Zhao , Member, IEEE

Abstract—Whole-body control (WBC) has emerged as an impor- expanded their operating space and exhibited better carrying
tant framework in manipulation for mobile manipulators. How- capacity, which can satisfy different application requirements.
ever, most existing WBC frameworks require known dynamics. When used in service scenarios, it exhibits excellent interac-
Considering whole-body manipulation and optimization with un-
known dynamics, this article presents the WBC of a nonholonomic tion capabilities with operating objects, such as transporta-
mobile manipulator using model predictive control (MPC) and tion [3], door opening [4], [5], cleaning tasks [6], and cart
fuzzy logic system. First, by constructing a dynamics-based feed- pushing [7].
back linearized robotic multi-input-multi-output (MIMO) system, Although scholars have carried out numerous research
an MPC-based WBC strategy is proposed for mobile manipula-
works [8], [9] on mobile manipulators in recent years, there still
tor. Such a strategy can provide the optimal control inputs with
the specified optimization index and constraints. Thereafter, a exist several challenges in current mobile manipulation tasks.
primal-dual neural network effectively addresses the constrained First of all, the mobile base and the fixed manipulator are two
quadratic programming (QP) problem over a finite receding hori- independent and different research topics. Therefore, how to
zon brought by the MPC. Then, in order to convert the interme- integrate both into a whole-body control (WBC) is a challenging
diate control signals into the optimal control torques that can be
issue. In addition, the system combined with the manipulator
executed by actuators, an adaptive FLS is employed to approximate
the unknown dynamics. The novel elements of the current design and the mobile chassis increases the difficulty of global optimal
control approach refer to the dynamics-based feedback linearized control. Considering motion safety and trajectory smoothing,
robotic MIMO system and the combination of an MPC module physical constraints, such as system inputs and states, also need
with an adaptive fuzzy controller. Finally, the trajectory tracking to be considered uniformly. If local constraints in the dynamics
experiments performed on a mobile dual-arm robot demonstrate
control are ignored, joint impacts or other unpredictable overall
the effectiveness of the proposed method.
responses may be generated. Moreover, the mobile manipulator
Index Terms—Fuzzy approximation, mobile manipulation, is a multi-input multi-output (MIMO) system with stronger non-
model predictive control (MPC), whole-body control (WBC). linearities and model uncertainties than the independent mobile
chassis or manipulator. Therefore, how to accurately identify the
I. INTRODUCTION parameters of the system becomes the premise of good control
effect.
HE mobile robotic manipulation inherits the vastness of the
T working space of the mobile platform and the flexibility
of the operating space of the robotic arm [1]–[3]. Compared
Regarding the mobile manipulative arm systems, there are two
ways for merging the mobile platform and the manipulator. One
is to separate the dexterous upper body and the platform, and
with fixed multijoint manipulators, mobile manipulators have
then independently control the aboveentioned two subsystems.
However, this requires a synchronization concept [10], [11].
Manuscript received 1 October 2021; revised 18 January 2022, 17 March Another approach is based on the overall dynamics model to
2022, and 6 June 2022; accepted 27 June 2022. Date of publication 11 July
2022; date of current version 1 March 2023. This work was supported by the control the whole body, known as WBC [12]–[15]. In such a case,
National Natural Science Foundation of China under Grant 62103111, Grant the synchronization is not relevant. Its main idea is to adopt op-
62173097, and Grant U2013601. (Corresponding author: Chun-Yi Su.) timization technology for solving the problems of operation and
Wang Yuan is with the School of Intelligent Manufacturing, Taizhou Univer-
sity, Taizhou 318000, China (e-mail: [email protected]). control. At present, the application of the WBC strategy in the
Yong-Hua Liu is with the School of Automation, Guangdong Province control of mobile manipulator has received extensive attention
Key Laboratory of Intelligent Decision and Cooperative Control, Guang- and made important progress [16]. The WBC performs multiple
dong University of Technology, Guangzhou 510006, China (e-mail:
[email protected]). tasks in an optimal way by exploiting the whole-body dynamics
Chun-Yi Su is with the School of Intelligent Manufacturing, Taizhou Univer- of the robot, taking consideration of the contact interaction and
sity, Taizhou 318000, China, and also with the Gina Cody School of Engineering the driving constraints. These tasks include interacting with the
and Computer Science, Concordia University, Montreal, QC H3G 1M8, Canada
(e-mail: [email protected]). environment, maintaining balance, and carrying out dynamic
Feng Zhao is with the Department of Automation, University of Science and locomotion on complex terrain. It should be noted that the tasks
Technology of China, Hefei 230027, China (e-mail: [email protected]). are not limited to the end-effector, but can also be performed
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TFUZZ.2022.3189808. elsewhere on the robot body [17], or even for a cooperative
Digital Object Identifier 10.1109/TFUZZ.2022.3189808 operational task between multiple robots [18].
1063-6706 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 800 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 31, NO. 3, MARCH 2023

However, the WBC causes complex MIMO systems, increas- r Compared with the traditional calculation method of MPC,
ing the difficulty of optimal control. Model predictive control the MPC in the present study consists of the LVI-PDNN
(MPC) is an optimal control method that considers constraints with low computational cost, which can satisfy the require-
and has been applied to numerous nonlinear MIMO systems ments of real-time computing.
[19]. Different from most traditional WBC strategies [12]–[15]
for optimization, MPC solves and obtains the optimal control II. MODELING OF MOBILE MANIPULATOR
signal within the prediction range of the limited horizon by defin-
ing a performance index that needs to be optimized, thus low- For n − 2 degrees of freedom manipulator and 2 wheel driven
ering the computational load without losing the performance. mobile platform, the whole robot has totally n joints. The coordi-
In [20] and [21], MPC obtains the optimal control input with the nates of the center point of the mobile platform are expressed as
dynamic or kinematic constraints of the robot. In [22] and [23], qs = [X, Y, θ]T ∈ R3×1 . According to the Lagrange formula of
considering that there exists inevitably external interference in the manipulator and the mobile platform, the dynamics equation
practical applications, the tube-based MPC achieves the sup- of the whole robot can be given as
pression of external interference by presenting a state feedback
controller. To apply MPC to the mobile manipulator system with M (qw )q̈w + V (qw , q̇w )q̇w + G(qw ) = Bτ + AT λ (1)
unknown dynamics, this study establishes a linearized MPC  
T T
model which can provide the dynamics control objective without where qw = qsT , qm ∈ R(3+n)×1 signifies the unified gener-
depending on the known dynamics. In addition, a primal-dual alized coordinates of the robot, qm = [θr , θl , θ1 , . . . , θn−2 ]T ∈
neural network algorithm (PDNN) is introduced to solve the Rn×1 represents the displacement of each joint including the left
MPC optimization problem. The PDNN is a parallel processing and right wheels, τ ∈ Rn×1 denotes the inputs of the manipula-
approach, which is a real-time solver without any pretraining, tor joints and mobile platform, M (qw ) ∈ R(3+n)×(3+n) denotes
and its computation errors feature superexponential convergent the inertia matrix of the robot, V (qw , q̇w )q˙w ∈ R(3+n)×1 signi-
rate and strong robustness [25]–[27], being especially beneficial fies nonlinear term matrix of Coriolis and Centrifugal forces
to reduce computational load for whole-body optimal control. and G(qw ) ∈ R(3+n)×1 denotes the gravitational forces vector,
Considering that the MPC approach depends on accurate B ∈ R(3+n)×n signifies a full rank input transformation matrix
forward predictions, it is challenging to handle unknown dy- denoted as
namics, such as uncertainties of parameters and functions, as ⎡ ⎤
cos θ cos θ 0 · · · 0
well as disturbances from the external environment. Fuzzy logic ⎢ sin θ sin θ 0 · · · 0 ⎥
systems (FLSs) are considered to be a powerful tool for handling ⎢ ⎥
⎢ 2l1 −2l1 0 · · · 0 ⎥
mathematically undefined systems that are subject to structural ⎢ ⎥
and unstructural uncertainties. Previously, studies have demon- 1⎢⎢ r 0 0 ··· 0 ⎥ ⎥
B= ⎢ 0 r 0 ··· 0 ⎥ (2)
strated the feasibility of using FLS to model unknown functions r⎢ ⎥
⎢ 0 0 r ··· 0 ⎥ ⎥
in robot control, thus ensuring the stability and robustness of the ⎢
⎢ .. .. .. . . . ⎥
whole control system [31]. ⎣ . . . . .. ⎦
In order to enhance the operating flexibility of the mobile ma- 0 0 0 ··· r
nipulator, a WBC framework of the mobile manipulator system
is introduced in this study, which combines the MPC, PDNN, and where r signifies the radius of the wheel, l1 refers to the distance
FLS to realize whole-body optimal dynamics control. Different from the center of the mobile platform to the driving wheels, and
from the separated control of mechanical manipulator and mo- l2 signifies the distance from the center of the mobile platform
bile chassis, the designed WBC based on the unified dynamics to the center of mass of the mobile platform. λ ∈ R3×1 signifies
model considers the coupling effect between manipulator and the Lagrange multiplier, and A ∈ R3×(3+n) refers to the matrix
chassis. Therefore, the optimal control on the whole model is the related to constraints. System constraints of the robot include
global optimal control in which all components simultaneously nonholonomic constraints and holonomic constraints, which can
participate in the coordinated motion. The main contributions be denoted as
can be summarized as follows.
r By constructing a dynamics-based feedback linearized A(qw )q̇w = 0 (3)
robotic MIMO system, an MPC-based WBC strategy for
mobile manipulator is proposed in this article. The WBC where
⎡ ⎤
strategy can provide the optimal control inputs based on the − sin θ cos θ −l2 0 0 0 · · · 0
specified optimization index and the specified dynamics A = ⎣ − cos θ − sin θ −l1 r 0 0 · · · 0 ⎦. (4)
constraints. − cos θ − sin θ l1 0 r 0 · · · 0
r A hierarchical multiloop control scheme is introduced to
combine the FLS with the MPC, the FLS involved in the By defining a (3 + n) × n dimensional matrix, and we have the
control loop compensates for the modeling uncertainties following:
and external disturbances, thus ensuring the stability of the
WBC. A(qw )S(qw ) = 0 (5)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 YUAN et al.: WHOLE-BODY CONTROL OF AN AUTONOMOUS MOBILE MANIPULATOR USING MODEL PREDICTIVE CONTROL 801

where Based on the abovementioned derivation, the following aug-

S(q⎡w ) mented system can be presented as:
⎤
c(l1 cos θ − l2 sin θ) c(l1 cos θ + l2 sin θ) 0 ··· 0 ė1 −κ−1 κ−1 e1
⎢c(l1 sin θ + l2 cos θ) c(l1 sin θ − l2 cos θ) 0 ··· 0 ⎥ ė = = (17)
⎢ ⎥ ė2 0 0 e2
⎢ c −c 0 ··· 0 ⎥
⎢ ⎥
⎢ 1 0 0 ··· 0 ⎥ 0
⎢ ⎥ + . (18)
=⎢ 0 1 0 ··· 0 ⎥ M−1 (μ1 − κBτ )
⎢ ⎥
⎢ 0 0 1 ··· 0 ⎥
⎢ ⎥ Given an auxiliary control input u for the system, the aug-
⎢ .. .. .. .. .. ⎥
⎣ . . . . . ⎦ mented closed-loop system can be further expressed as
0 0 0 ··· 1 e1
(6) ė = A + Bu (19)
e2
where c = r/2l1 , q̇w can, therefore, be expressed as where
q̇w = S(qw )q̇m . (7) −κ−1 κ−1
A= (20)
0 K1
Take the time derivative of both sides of (7)
0
q̈w = S(qw )q̈m + Ṡ(qw )q̇m . (8) B= (21)
I
The abovementioned expression for q̈w into (1) and multiplying u = M−1 (μ1 − κBτ ) − K1 e2 (22)
it by S T (qw ), we obtain the following expression:
where K1 ∈ R n×n
denotes a diagonal negative matrix, I ∈
S T (qw ) M (qw ) S(qw )q̈m + Ṡ(qw )q̇m Rn×n represents an identity matrix. We can rewrite (19) as

+ S T (qw )C(qw , q̇w )S(qw )q̇m ė = Ae + Bu. (23)

+ S T (qw )G(qw ) − S T (qw )Bτ = S T (qw )AT λ. (9) If the dynamics model is known beforehand, we have

By substituting (5), the dynamics of the whole body including κBτ = μ1 − M(K1 e2 + u). (24)
the mechanical arms and the mobile platform can be expressed We rewrite the model (23) in the discrete form into a standard
as linear state-space model for control design as
M(qw )q̈m + C(qw , q̇w )q̇m + G(qw ) = Bτ (10) x(k + 1) = a(x(k)) + b(x(k))u(k) (25)
where M(qw ) = S T (qw )M (qw )S(qw ), C(qw , q̇w ) = S T (qw ) where
M (qw )S(qw ) + S T (qw )C(qw , q˙w )S(qw ), G(qw ) = S T (qw )G
−κ−1 κ−1
(qw ), and B = S T (qw )B. a(x(k)) = x(k) + x(k)T
0 K1
III. PROBLEM FORMULATION 0
b(x(k)) = T (26)
I
The errors between the desired and actual trajectories can be
denoted as  T
and x(k) = e1 (k)T , e2 (k)T ∈ R2n is the system state vec-
e1 (t) = qd (t) − qm (t). (11) tor. T represents the sampling time.

Then, a linear combination of system state variables qm and q̇m IV. MPC DESIGN
can be denoted as
By writing the continuous time model as a discrete time
e2 (t) = κė1 (t) + e1 (t) (12) model and considering the input and state constraints, the MPC
problem can be transformed into a closed-loop discrete time
where κ ∈ Rn×n denotes a positive diagonal matrix.
optimal control problem. First, we write the dynamics system
It is not difficult to have
of the robot as a discrete-time state-space system, where the
ė1 (t) = κ−1 (e2 − e1 ) (13) control input denotes uk ∈ U and the state input is xk ∈ X . The
U ⊂ Rm̄ signifies the set of control input vectors and X ⊂ Rn̄
ė2 (t) = κë1 + ė1 . (14)
denotes the set of admissible state. MPC can predict the future
Then, we have state of the system at each sampling time within a specified
preset time range, offering the current optimal control input
M(qw )ė2 (t) = μ1 − κBτ (15) according to the predicted state. For the simplicity of expression,
where the robot nonlinear function is xk+i and uk+i at each instant k can be simply expressed as xi
and ui , respectively, that is, the initial instant of the predicted
μ1= M(qw )(κq̈d + ė1 )+C(qw , q̇w )(κq̇m )+κG(qw ). (16) horizon. Subsequently, for the control problem, we define two

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 802 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 31, NO. 3, MARCH 2023

important horizons, namely the prediction horizon Np and the

control horizon Nu , which satisfy constraint 0 ≤ Nu ≤ Np . In ⎡ ⎤
a(xe (k|k − 1))
addition, we define the state space sequence X = x1 , . . . , xNp ⎢ a(xe (k + 1|k − 1)) ⎥
and the corresponding control input sequence U = u1 , . . . , uNu . ⎢ ⎥
ã = ⎢ .. ⎥
⎣ . ⎦
A. The Nominal MPC a(xe (k + Np − 1|k − 1))
⎡ ⎤
Since the MPC is an iterative optimization control method, b(xe (k|k − 1))u(k − 1)
its cost function J can be presented as follows: ⎢ b(xe (k + 1|k − 1))u(k − 1) ⎥
⎢ ⎥
b̃ = ⎢ .. ⎥.

Np u −1
N ⎣ . ⎦
J(k) = ||x̄T (k + j|k)||2P + ||ΔūT (k + j|k)||2Q b(xe (k + Np − 1|k − 1))u(k − 1)
j=1 j=0
Therefore, the optimization control problem (27) with the
(27)
multiple constraints (29)–(31) becomes
where x̄(k + j|k) refers to the predicted future horizon state, and
Δū(k + j|k) = ū(k + j|k) − ū(k − 1 + j|k) represents the in- min GΔū(k) + ã + b̃2P + Δū(k)2Q (38)
put increment. The Euclidean norm is denoted by  · . Besides,
P ∈ R2nNp ×2nNp and Q ∈ RnNu ×nNu indicate positive defi- subject to x̄min  ã + b̃ + G, Δū(k)  x̄max , ūmin  ū(k −
nite matrices. Obviously, the state variables of model (23) are 1)  ūmax , Δūmin  Δū(k) ⎡  Δūmax , ūmin
⎤  ū(k − 1) +
related to the tracking errors, and thus we call model (23) as I 0 ··· 0
⎢I I ··· 0⎥
the discrete tracking error state model, and further it can be ⎢ ⎥
˜
IΔū(k)  ūmax , where I˜ = ⎢ . . . . ⎥ ∈ RnNu ×nNu ,
expressed as ⎣ .. .. . . .. ⎦
I I ··· I
xe (k + 1) = a(xe (k)) + b(xe (k))ue (k) (28)
Δūmax , ūmax , and x̄max are the upper bounds with Δūmin , ūmin ,
which is subjected to the following constraints: and x̄min being the lower bounds.
Then, according to the objective of control optimization (38),
umin  ue (k)  umax (29) we can form the following optimization problem:
Δumin  Δu(k)  Δumax (30) 1
min ΔūT FΔū + cT Δū
xmin  xe (k)  xmax (31) 2
EΔū  d
where xe = [x1 , x2 ]T = [e1 , e2 ]T ∈ R2n is the state vector;
ue = M−1 (μ1 − Bτ ) − K1 e2 ∈ Rn is the input vector, and Δūmin  Δū  Δūmax (39)

x1 −κ−1 x1 + κ−1 x2 where the coefficients include

a(xe ) = + T ∈ R2n (32)  
x2 K1 x 2 F = 2 G T P G + Q ∈ RnNu ×nNu (40)
0
b(xe ) = T ∈ R2n×n (33) c = 2G T P g̃ + f˜ ∈ RnNu (41)
I
 T
where T is the period time of sampling, the lower and upper ˜ −G, G ∈ R(2nNu +4nNp )×nNu
˜ I,
E = −I, (42)
bound of input signals and states are umin , Δumin , xmin , and
⎡ ⎤
umax , Δumax , xmax , respectively. −ūmin + ū(k − 1)
Define the following vectors: ⎢ ūmax − ū(k − 1) ⎥
d= ⎢ ⎥
⎣ −x̄min + f˜ + g̃ ⎦ ∈ R
2nNu +4nNp
. (43)
¯ = [xe (k + 1|k), . . . , xe (k + Np |k)]T
x(k) (34)
x̄max − f˜ − g̃
ū(k) = [ue (k + 1|k), . . . , ue (k + Np |k)]T (35)
B. Neural-Dynamic Optimization
Δū(k) = [Δu(k|k), . . . , Δu(k + Nu − 1|k)]T . (36)
For the quadratic programming (QP) problem of (39), the
The predicted output can be expressed as PDNN obtains the decision vector  ∈ R3nNu +4nNp and its
boundary ± by defining a auxiliary vector β ∈ R2nNu +4nNp .
x̄(k) = GΔū(k) + ã + b̃ (37)
We have
     
with G = Δū − Δūmin + Δūmax
⎡ ⎤  := ,  := ,  := (44)
b(xe (k|k − 1)) ... 0 β −β − +β +
⎢ b(xe (k + 1|k − 1)) ... ⎥ 0
⎢ ⎥
⎢ .. .. ⎥ .. where − is the lower bound and + is the upper bound. For
⎣ . ⎦
. . any i, the elements βi +  0 in β + are sufficiently to represent
b(xe (k + Np − 1)|k −1)) . . . b(xe (k+Np −1)|k−1)) positive infinity. Similarly, for any i, the elements βi −  0 in

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 YUAN et al.: WHOLE-BODY CONTROL OF AN AUTONOMOUS MOBILE MANIPULATOR USING MODEL PREDICTIVE CONTROL 803

Fig. 1. Block diagram of the PDNN neural-dynamic system.

β − are sufficient to represent negative infinity. Based on the Fig. 2. Network architecture of the PDNN neural-dynamic system.
definition of (44), we introduce a convex set Φ which is denoted
as Φ = − ≤  ≤ + .
contains Δū into the system. Subsequently, the output signal ∗
In addition, we define a vector p and a matrix L
    of the PDNN system is obtained, and the first n elements of ∗
c F −E T are taken to form the optimal solution for the stabilization control
p= L= . (45) of the robot. The structure of the PDNN is depicted in Fig. 2,
−d E 0
where Λi denotes the ith row of the scaling matrix Λ = I + LT .
Then, for the reasonably equivalent transformation of QP prob- Pφ is a piecewise linear limiter. F () is a vector-valued function
lem (39), we have the following lemma. and F () =  − (L + ρ), C() = . If the dimensions of input
Lemma IV.1 (See [24]): The effect of the QP problem (39) is  is k, the neural network consists of 4k summers, k integra-
equivalent to finding a vector ∗ ∈ Φ that satisfies the following tors, k limiter-operations as well as 2k 2 multiplications, and
linear variational inequalities as: thus the PDNN has O(6k + 2k 2 ) operations per iteration [29].
( − ∗ )T (L∗ + p) ≥ 0 ∀ ∈ Φ (46) Compared with the traditional gradient descent method which
requires repeated calculation of the Hessian matrix [30], the
where , L, p, and Φ have been defined previously. proposed PDNN approach is computationally less complex and
Furthermore, the linear variational inequality (46) can be thus involves less computational cost.
equivalently transformed into the following piecewise linear
equation [25]: V. FUNCTIONAL UNIVERSAL APPROXIMATION USING FLSS
PΦ ( − (L + p)) −  = 0 (47) FLS is widely used in the control of nonlinear systems
where PΦ (·) refers to the projection operator onto Φ and because of its powerful ability to approximate complex systems
defined as PΦ () = [PΦ (1 ), . . . , PΦ (2nNu +4nNp )]T , ∀i ∈ without the prior experience of system dynamics [31]–[34].
{1, . . . , 2nNu + 4nNp }, with Considering an h-inputs FLS, single-output with the singleton
⎧ − fuzzifier, center average defuzzifier, product-inference rule,
⎨ i , i < i − and Gaussian membership function given by L fuzzy rules
PΦ (i ) = i , i − ≤ i ≤ i + (48) H (j) : If x1 is D1j and . . . and xh is Dhj , then y is W j , j =
⎩ +
i , i > i + . 1, . . . , L, where H j represents the jth rule, 1 ≤ j ≤ L,
To solve the abovementioned linear (47) based on piecewise (x1 , x2 , . . . , xh )T ∈ P ⊂ H h . Dij and W j indicate the fuzzy
projection, we employ the dual dynamical system design method sets in P and H. y ∈ H are the linguistic variables associated
to build a dynamical system. However, the matrix L is asymmet- with the inputs and output of the FLS, respectively. It can be
ric, leading to such an unstable system. Based on the design found that FLSs perform nonlinear mapping from P to H [35].
experience [28], the solution of (47) can be obtained by an The FLS can be presented as
improved dynamic neural network L h
  j=1 yj i=1 μDij (xi )
˙ = γ I + LT {PΦ ( − (L + p)) − } (49) y(x) = L n (50)
j=1 i=1 μD j (xi )
i
where γ denotes an adjustable positive parameter that affects
T
the convergence rate of the system. Obviously, the PDNN based where x = [x1 , x2 , . . ., xh ] , μDj (xi ) is the member-
i
on the linear variational inequality model (49) can quickly solve ship function of linguistic variable xi with μDj (xi ) =
strict convex QP problems.   i
(x − )2
Remark IV.1: The structure of the neural-dynamics system exp − i 2σ2ij , where = ( ij )h×L ∈ Rh×L and
ij

(49) is shown in Fig. 1. In the control process of dynamic system σ = (σij )h×L ∈ Rh×L . For brevity and clarity, we can
(49), we first construct coefficient vector and matrix, such as define the weight vector W = [y1 , y2 , . . ., yL ]T and function
F, d, E, c, Δūmin , and Δūmax , and then feed the vector  that vector S(x, , σ) = [s1 , s2 , . . ., sL ]T , respectively, where

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 804 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 31, NO. 3, MARCH 2023

  
L h ˙
sj = hi=1 μDj (xi )/ j=1 i=1 μ D k (xi ) , Then, (50) can
i
where
 Ŵ ∈ RL×n is positive constant matrices, Ŵ =
i
˙ ˙
be written as Ŵ1 , . . . , Ŵn ∈ RL×n is the real-time update rates of fuzzy
y = W T S(x, , σ). (51) ˙
weights, Ŵi ∈ RL is respectively ith column of Ŵ . Γi ∈ RL×L
It is demonstrated that the FLS (51) can uniformly approximate is positive definite matrices. φi is the chosen constants. Consider
any given real continuous function. In theory, it can converge the following Lyapunov function:
from a compact set to any degree of accuracy. Therefore, for 1 T 1
a given unknown nonlinear functions fi (xi ), i = 1, . . . , h, the V = e Meu + tr W̃ T Γ−1 W̃ (57)
2 u 2
approximation can be expressed as
where W̃ T S = W ∗T S − Ŵ T S. The time derivative of V̇ Lya-
fi (xi ) = Wi∗T S(xi ) + i (xi ) ∀xi ∈ Ei ⊂ H i (52)
punov function becomes
where Ei refers to the compact sets, S(xi ) is the fuzzy basis
1 ˙ .
vector, Wi∗ is an unknown constant parameter vector, and i (xi ) V̇ = eTu Mėu + eTu Ṁeu + tr W̃ T Γ−1 W̃ (58)
2
is the approximation error.
Theorem V.1: For arbitrary > 0 and any given real continu- Considering (53) and defining μ2 = M(−ud − K1 e2 + κq̈d +
ous function g(x) on the compact set P ∈ H h , a function f (x) ė1 ) + C(κq̇m ) + κG for brevity, it is easy to have
in (51) could be found to satisfy supx∈P g(x) − f (x) ≤ .
1 ˙
Remark V.1: The optimal weight vector Wi∗ in (52) is V̇ = eTu (μ2 − κBτ ) + eTu Ṁeu + tr W̃ T Γ−1 W̃
an artificial quantity required only for the purpose of the 2
analysis. Specifically, Wi∗ := arg minc∈H h {supx∈Ei |fi (xi ) − 1
= eTu (μ2 + Ceu − κBτ ) + eTu Ṁ − 2C eu
WiT S(xi )|}, i.e., Wi∗ is considered as the optimal weight vector 2
if it minimizes i (xi ) for all xi ∈ Ei where Ei ⊂ H i is a compact T −1 ˙
+ tr W̃ Γ W̃
set.
Assumption V.1: On a compact region Ei ∈ H i , | i (xi )| ≤
∗ ∗ where Ṁ − 2C is skew symmetry [36], therefore, eTu
i , where i > 0 remains unknown bound. Based on the above
mentioned analysis, it is not difficult to know that the uncer- Ṁ − 2C eu = 0. The above mentioned equation can be fur-
tainty relation of the system is reasonably transformed into the ther written as
estimation of the unknown parameters Wi∗ and the unknown ˙ .
bounds ∗i . V̇ = eTu (μ2 + Ceu − κBτ ) + tr W̃ T Γ−1 W̃ (59)

Considering (54), we have

VI. FUZZY CONTROL DESIGN AND ITS STABILITY ANALYSIS
t
Define eu = 0 (u − ud )dt, and ėu = u − ud where ud rep- V̇ = eTu μ2 + Ceu − K2 eu − Ŵ T S
resents the intermediate term optimized by MPC and u can be
obtained from the sensor data combined with (19). Considering ˙ .
+ tr W̃ T Γ−1 W̃ (60)
(22), it is easy to have
By integrating the approximation, we can rewrite (60) as
Mėu = M(u − ud )
 
= − Mud + μ1 − κBτ − MK1 e2 V̇ = eTu W ∗T S + − K2 eu − Ŵ T S
= M (−ud − K1 e2 +κq̈d + ė1 ) ˙
+ tr W̃ T Γ−1 W̃
+ C(κq̇m ) + κG − κBτ. (53)
˙
= eTu W̃ T S + − K2 eu + tr W̃ T Γ−1 W̃
The external torques is given by
κBτ (t) = K2 eu + Ŵ T S (54) ˙
= eTu ( − K2 eu ) + tr eTu W̃ T S + W̃ T Γ−1 W̃
with
= eTu ( − K2 eu )
∗T
W S + = M (−ud − K1 e2 + κq̈d + ė1 )

n
˙
+ C (κq̇m + eu ) + κG (55) + eu,i W̃iT Si + W̃iT Γ−1
i W̃i . (61)
i=1
where W ∗ ∈ RL×n , S(x, , σ) ∈ RL , L is numbers of fuzzy
 T ˙
T
rules, x = qm T
, q̇m , (−ud − K1 e2 +κq̈d + ė1 )T ,(κq̇m + eu )T Considering Ŵi = −Γi Si eu,i + φi Ŵi , it is easy to have

∈ R4n . , and σ ∈ R4n×L , ∈ Rn is approximation errors and V̇ = eTu ( − K2 eu )

  ≤ ∗ . The updating laws can be expressed as follows:

n
˙
Ŵi = Γi Si eu,i + φi Ŵi (56) + eu,i W̃iT Si − W̃iT Γ−1
i Γi Si eu,i + φi Ŵi
i=1

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 YUAN et al.: WHOLE-BODY CONTROL OF AN AUTONOMOUS MOBILE MANIPULATOR USING MODEL PREDICTIVE CONTROL 805

of the shoulder, allowing independent motion control. A single

arm weighing 4.9 kg has a total of five joints. Each joint is driven
by a Maxon dc flat brushless motor fitted with an incremental
encoder and a harmonic reducer, equipped with an Elmo Solo
Whistle servo driver. In the current experiments, the elbow joint
and the shoulder joint were selected as the test joints due to
the heavy weight bearing and obvious coupling of motion. The
motor in the shoulder joint of the manipulator is EC 90 flat
160 W (reduction ratio 160:1, 2048 lines). The motor in the
elbow joint of the manipulator is EC 45 flat 30 W (reduction
ratio 100:1, 1000 lines). Besides, the wheeled mobile platform
is equipped with two driving wheels and two universal wheels for
balancing and supporting. The wheel is driven by a saegmotor
(48SYK68) equipped with an incremental encoder with 1024
lines and a planetary gear reducer with reduction ratio 64:1,
similar to the joint motors, the wheel motors are also driven by
the Elmo Solo Whistle servo drivers. Due to the symmetrical
Fig. 3. Developed redundant mobile manipulator. design, the geometric center of the mobile robot is the center
of mass, that is, l2 = 0.0 m in (4). To obtain a wider range of

n movement and more flexible operation of the manipulator, the
= eTu ( − K2 eu ) + −W̃iT φi Ŵi . (62) shoulder width L4 of the manipulator is designed to exceed the
i=1 distance between the two wheels of the mechanical platform.
Correspondingly, this design also brings more motion coupling
Because −W̃iT Ŵi ≤ − u1 W̃iT W̃i + 12 Wi∗ Wi∗ and eu T
≤ 1
2 between manipulator and platform, reflecting the importance of
eTu eu + 12 T we have the WBC strategy proposed in the present study. The important
V̇ ≤ eTu ( − K2 eu ) dimension variables of the robot structure marked in Fig. 3
are measured as: L1 = 0.27 m, L2 = 0.28 m, L3 = 1.28 m,

n
φi 
n
φi L4 = 0.80 m, r = 0.10 m, and l1 = 0.39 m.
+ Wi∗T Wi∗ − W̃iT W̃i
2 2 The whole control system is developed based on Windows 10
i=1 i=1
operation system and runs on the industrial computer inside the
1 1 robot. The program developed in C++ programming language in
≤ − K2 eTu eu + eTu eu + T
2 2 visual studio 2019 communicates with the motor drivers through
n
φi ∗T ∗  φi T
n the controller area network bus, which periodically executes
+ Wi Wi − W̃ W̃i the following three steps: reading motor position and speed
i=1
2 i=1
2 i
information from serve drivers, computing the output of the
 
I 1 controller as well as sending torque control signal to the drivers.
= eTu − K2 eu + T
2 2

n
φi 
n
φi
+ Wi∗T Wi∗ − W̃iT W̃i B. Experiment Setup
i=1
2 i=1
2
In order to verify the effect of the proposed WBC
≤ − κ1 V + C 1 (63) scheme by combining MPC and adaptive fuzzy technology,
where we carried out comparative experiments between the
     proposed controller and a fuzzy controller given as:
2λmin I2 − K φi κBτ (t) = K2 e2 + Ŵ T S(x) where W ∗T S(x) + = μ1 + Ce2
κ1 = min , 2 min , (64)
λmax (M) 2Γ−1
i with x = qm T T
, q̇m , (κq̈d + e˙1 )T , (e2 + κq̇m )T ∈ R16 . We
set the parameter values of the comparative experiments, as
1 ∗ 2  φi ∗T ∗
n
C1 =   + W Wi . (65) shown in Table I. In more details, P = pI ∈ R24×24 and
2 i=1
2 i Q = qI ∈ R8×8 . Considering the compromise and balance of
convergence speed and accuracy, the parameter of the neural
VII. EXPERIMENT DESCRIPTION network is set as γ = 0.001. It is of note that the comparison of
controllers is performed on the premise that the parameters of
A. Robotic System Description
the FLS in the two experiments are completely consistent.
The developed mobile manipulator mainly consists of two Experiments in the present study verify the effect of the WBC
parts: the double arms and the differential driven mobile plat- strategy, which is based on the known given clear trajectory
form, as shown in Fig. 3. The arms of the mobile manipulator including the trajectory of the manipulator and the trajectory
are symmetrical in design and they are arranged on both sides of the mobile platform. We define the motion trajectory of the

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 806 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 31, NO. 3, MARCH 2023

TABLE I
PARAMETER SETTING OF THE COMPARATIVE EXPERIMENTS

Fig. 5. Desired and actual trajectories of the elbow joint.

Fig. 4. Desired and actual trajectories of the shoulder joint.

shoulder and elbow joints of the manipulator as follows: Fig. 6. Tracking errors of the shoulder joint.

1
qm1 =π sin(0.2πt)
9
1
qm2 = π sin(0.2πt). (66)
10
To highlight the approximation effect of the FLS, the initial con-
figuration of the manipulator was uniformly set as 30-degrees
swing arm rather than natural droop in the comparative experi-
ments, as displayed in Fig. 3. The expected trajectory [39]–[42]
of chassis movement for large steering in narrow space is set as
follows:
X = 0.8 sin(0.05t)
Y = −0.8 sin(0.10t) (67)
Fig. 7. Tracking errors of the elbow joint.
where X and Y are global coordinates of the center point of the
platform. The purpose of the experiments is to test the perfor-
proposed controller almost covers the desired trajectory even at
mance of the proposed controller for trajectory tracking. It is of
the time of joint movement turning. The tracking effect of the
note that the swing of the manipulator joints and the movement
controller is satisfactory. Figs. 6 and 7 display the evolution of
of the mobile platform were carried out simultaneously in the
the tracking errors of each joint. Because the FLS compensates
experiments. The experimental results show that the controller
for the control torque, the tracking error range of the system
combining MPC and fuzzy makes better tracking performance,
moves toward the X-axis. As can be observed from the blue
especially the robustness of anti-interference.
lines in Figs. 6 and 7, the error range of the fuzzy controller
keeps moving toward the X-axis in the first 100 s. Nevertheless,
C. Experimental Results and Analysis
during the next 100–125 s, that is, the last three cycles, the error
Figs. 4 and 5 show the expected and actual trajectories of the range no longer moves toward negative axis but maintains within
shoulder and elbow joints. Obviously, the actual trajectory of the a fixed small range. When compared with fuzzy controller, it can

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 YUAN et al.: WHOLE-BODY CONTROL OF AN AUTONOMOUS MOBILE MANIPULATOR USING MODEL PREDICTIVE CONTROL 807

Fig. 8. Control inputs for the shoulder and elbow joints. Fig. 10. Tracking error x of the mobile platform.

Fig. 9. Desired and actual trajectories of the mobile platform. Fig. 11. Tracking error y of the mobile platform.

be found that the proposed controller has a significantly faster

convergence rate. Based on the black lines presented in Figs. 6
and 7, the control error rapidly stabilizes in a smaller fixed region
after a short period. Fig. 8 presents the control input of the two
joints based on the proposed method. As presented in the figure,
the control input does not exceed the saturation value set by the
program and is within a reasonable range.
The expected trajectory and actual trajectories of the platform
are displayed in Fig. 9. According to the curve in the figure,
the proposed method still shows good tracking performance
even when the robot arm moves in a relatively large range, as
illustrated in Figs. 4 and 5. The tracking errors of the platform
in task space are shown in Figs. 10 and 11. Different from the
data curve in joint space, the tracking errors in task space can Fig. 12. Angular and linear velocities of the mobile platform.
directly evaluate the precision of robot task execution. It can
be observed from the figure that the proposed method shows
better task execution precision than fuzzy controller. The linear for evaluating our tracking effect. Here are the following three
and angular velocities are shown in Fig. 12. Since the expected important indicators.
angular and linear velocities of the robot do not start from zero, t 1/2
the actual angular and linear velocities of the robot change 1) L2 (e) = t1e 0 e |e|2 dt : The scalar value norm of
rapidly in the initial stage, and keep consistent with the expected tracking error can objectively evaluate the overall perfor-
velocity. Accordingly, Fig. 13 illustrates the evolution of the mance of tracking accuracy.
control inputs of the proposed method over time. 2) eF = max(te −2≤t≤te ) |e(t)|: The maximum absolute error
In order to further objectively explore the control perfor- of tracking in the last 2 s is regarded as the controller’s final
mance, we adopted the colleague approved indexes [37], [38] tracking error.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 808 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 31, NO. 3, MARCH 2023

ACKNOWLEDGMENT
The authors would like to thank Professor Zhijun Li of the
Institute of Advanced Technology, University of Science and
Technology of China for his kind assistance.
REFERENCES
[1] F. Wang, J. R. G. Olvera, and G. Cheng, “Optimal order pick-and-place of
objects in cluttered scene by a mobile manipulator,” IEEE Robot. Automat.
Lett., vol. 6, no. 4, pp. 6402–6409, Oct. 2021.
[2] M. Rigotti-Thompson, M. Torres-Torriti, F. A. A. Cheein, and G. Troni,
“H∞ -Based terrain disturbance rejection for hydraulically actuated mobile
manipulators with a nonrigid link,” IEEE/ASME Trans. Mechatronics,
vol. 25, no. 5, pp. 2523–2533, Oct. 2020.
[3] J. Jiao, Z. Cao, N. Gu, S. Nahavandi, Y. Yang, and M. Tan, “Transportation
by multiple mobile manipulators in unknown environments with obsta-
Fig. 13. Control inputs for the left and right wheels. cles,” IEEE Syst. J., vol. 11, no. 4, pp. 2894–2904, Dec. 2017.
[4] W. C. Rhee, Y. Shim, H. Lee, and S. Park, “Door-opening control of
a service robot using the multifingered robot hand,” IEEE Trans. Ind.
TABLE II Electron., vol. 56, no. 10, pp. 3975–3984, Oct. 2009.
COMPARISON OF PERFORMANCE INDEXES BETWEEN THE TWO CONTROL [5] S. Ahmad, H. Zhang, and G. Liu, “Multiple working mode control of door-
SCHEMES opening with a mobile modular and reconfigurable robot,” IEEE/ASME
Trans. Mechatronics, vol. 18, no. 3, pp. 833–844, Jun. 2013.
[6] D. Leidner, A. Dietrich, M. Beetz, and A. Albu-Schaffer, “Knowledge-
enabled parameterization of whole-body control strategies for compliant
service robots,” Auton. Robot, vol. 40, no. 3, pp. 519–536, 2016.
[7] X. Ren, Y. Liu, Y. Hu, and Z. Li, “Integrated task sensing and
whole body control for mobile manipulation with series elastic actu-
ators,” IEEE Trans. Automat. Sci. Eng., early access, Mar. 9 2022,
doi: 10.1109/TASE.2022.3156127.
[8] Y. Ren, S. Sosnowski, and S. Hirche, “Fully distributed cooperation for
networked uncertain mobile manipulators,” IEEE Trans. Robot., vol. 36,
no. 4, pp. 984–1003, Aug. 2020.
[9] J. Obregón-Flores, G. Arechavaleta, H. M. Becerra, and A. Morales-
t 1/2 Díaz, “Predefined-time robust hierarchical inverse dynamics on torque-
3) L2 (τ ) = t1e 0 e |τ |2 dt : The average control input controlled redundant manipulators,” IEEE Trans. Robot., vol. 37, no. 3,
pp. 962–978, Jun. 2021.
reflects the consumed energy, which can be applied as [10] L.-P. Ellekilde and H. I. Christensen, “Control of mobile manipulator
an index to evaluate the controller. using the dynamical systems approach,” in Proc. IEEE Int. Conf. Robot.
Automat., 2009, pp. 1370–1376.
The control performance is listed in Table II. From the pre- [11] D. H. Shin, B. S. Hamner, S. Singh, and M. Hwangbo, “Motion planning
vious analysis, it can be found that the proposed controller is for a mobile manipulator with imprecise locomotion,” in Proc. IEEE/RSJ
featured by good stability and robustness in spite of complex Int. Conf. Intell. Robots Syst., 2003, pp. 847–853.
[12] K. Nagasaka et al., “Whole-body cooperative force control for a two-armed
coupling between joints, model uncertainties, and unknown and two-wheeled mobile robot using generalized inverse dynamics and
external disturbances. idealized joint units,” in Proc. IEEE Int. Conf. Robot. Automat., 2010,
pp. 3377–3383.
[13] S. Hyon, Y. Ida, J. Ishikawa, and M. Hiraoka, “Whole-body locomotion
VIII. CONCLUSION and posture control on a torque-controlled hydraulic rover,” IEEE Robot.
Automat. Lett., vol. 4, no. 4, pp. 4587–4594, Oct. 2019.
During the whole-body movement of the mobile manipu- [14] V. Klemm et al., “LQR-assisted whole-body control of a wheeled bipedal
robot with kinematic loops,” IEEE Robot. Automat. Lett., vol. 5, no. 2,
lator, the strong coupling effect between the manipulator and pp. 3745–3752, Apr. 2020.
the mobile platform cannot be ignored. Meanwhile, uncertain [15] C. Mastalli, I. Havoutis, M. Focchi, D. G. Caldwell, and C. Semini,
dynamics and external disturbances also pose challenges to the “Motion planning for quadrupedal locomotion: Coupled planning, terrain
mapping, and whole-body control,” IEEE Trans. Robot., vol. 36, no. 6,
control. To address the above mentioned issues, we develop pp. 1635–1648, Dec. 2020.
a whole-body dynamics control strategy with the MPC and [16] M. V. Minniti, F. Farshidian, R. Grandia, and M. Hutter, “Whole-body
adaptive fuzzy technique. The MPC based on the PDNN obtains MPC for a dynamically stable mobile manipulator,” IEEE Robot. Automat.
Lett., vol. 4, no. 4, pp. 3687–3694, Oct. 2019.
the optimal control solution under the specified optimization [17] E. S. Neo, K. Yokoi, S. Kajita, and K. Tanie, “Whole-body motion
criteria and constraints, which makes the system robust in the generation integrating operator’s intention and robot’s autonomy in con-
presence of external disturbances. The adaptive control based trolling humanoid robots,” IEEE Trans. Robot., vol. 23, no. 4, pp. 763–775,
Aug. 2007.
on the FLS tracks the optimal control solution of the MPC by [18] M. Wu, A. Konno, S. Ogawa, and S. Komizunai, “Symmetry cooperative
approximating the system, ensuring the adaptability of the WBC. object transportation by multiple humanoid robots,” in Proc. IEEE Int.
Experiments on a mobile manipulator prove the effectiveness Conf. Robot. Automat., 2014, pp. 3446–3451.
[19] Y. Song, X. Huang, and C. Wen, “Tracking control for a class of unknown
and feasibility of the proposed WBC scheme. For future work, nonsquare MIMO nonaffine systems: A deep-rooted information based
we will concentrate on the development of trajectory planning robust adaptive approach,” IEEE Trans. Autom. Control, vol. 61, no. 10,
technology for the whole-body joints, and combine it with the pp. 3227–3233, Oct. 2016.
[20] Y. Hu et al., “Nonlinear model predictive control for mobile medical robot
developed controller to realize the coordinated and stable motion using neural optimization,” IEEE Trans. Ind. Electron., vol. 68, no. 12,
behavior of mobile manipulator in complex environment. pp. 12636–12645, Dec. 2021.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.
 YUAN et al.: WHOLE-BODY CONTROL OF AN AUTONOMOUS MOBILE MANIPULATOR USING MODEL PREDICTIVE CONTROL 809

[21] E. Daneshmand, M. Khadiv, F. Grimminger, and L. Righetti, “Variable Wang Yuan received the M.S. degree from the Col-
horizon MPC with swing foot dynamics for bipedal walking control,” lege of Engineering, South China Agricultural Uni-
IEEE Robot. Automat. Lett., vol. 6, no. 2, pp. 2349–2356, Apr. 2021. versity, Guangzhou, China, in 2014, and the Ph.D.
[22] M. Cannon, B. Kouvaritakis, S. V. Rakovic, and Q. Cheng, “Stochastic degree from the School of Automation Science and
tubes in model predictive control with probabilistic constraints,” IEEE Engineering, South China University of Technology,
Trans. Autom. Control, vol. 56, no. 1, pp. 194–200, Jan. 2011. Guangzhou, in 2019.
[23] A. J. Prado, M. Torres-Torriti, and F. A. Cheein, “Distributed tube- He is currently with Taizhou University, Taizhou,
based nonlinear MPC for motion control of skid-steer robots with terra- China. His research interests include robotics, intel-
mechanical constraints,” IEEE Robot. Automat. Lett., vol. 6, no. 4, ligent control, and human–robot interaction.
pp. 8045–8052, Oct. 2021.
[24] Z. Li, J. Deng, R. Lu, Y. Xu, J. Bai, and C.-Y. Su, “Trajectory tracking
control of mobile robot systems incorporating neural-dynamic optimized
model predictive approach,” IEEE Trans. Syst. Man Cybernetics: Syst.,
vol. 46, no. 6, pp. 740–749, Jun. 2016. Yong-Hua Liu received the B.Eng. degree in pro-
[25] Y. Zhang, “On the LVI-based primal-dual neural network for solving online cess equipment and control engineering from Yangtze
linear and quadratic programming problems,” in Proc. Amer. Control University, Jingzhou, China, in 2008, and the Ph.D.
Conf., 2005, pp. 1351–1356. degree in control theory and control engineering
[26] W. S. Tang and J. Wang, “A primal-dual neural network for joint torque op- from the South China University of Technology,
timization of redundant manipulators subject to torque limit constraints,” Guangzhou, China, in 2014.
in Proc. IEEE Int. Conf. Syst. Man, Cybern., 1999, pp. 782–787. He is currently an Associate Professor with the
[27] Y. Zhang, Z. Li, H. Tan, and Z. Fan, “On the simplified LVI-based primal- School of Automation, Guangdong University of
dual neural network for solving LP and QP problems,” in Proc. IEEE Int. Technology, Guangzhou, China. From July 2014 to
Conf. Control Automat., 2007, pp. 3129–3134. May 2017, he was a Lecturer with the School of
[28] Y. S. Xia, G. Feng, and J. Wang, “A primal-dual neural network for online Mechanical and Electrical Engineering, Hunan Uni-
resolving constrained kinematic redundancy in robot motion control,” versity of Science and Technology, Xiangtan, China. His research interests
IEEE Trans. Syst. Man Cybern. Part B Cybern., vol. 35, no. 1, pp. 54–64, include nonlinear control, universal control, and PID control.
Feb. 2005.
[29] H. Xiao et al., “Robust stabilization of a wheeled mobile robot using model
predictive control based on neurodynamics optimization,” IEEE Trans.
Ind. Electron., vol. 64, no. 1, pp. 505–516, Jan. 2017. Chun-Yi Su (Senior Member, IEEE) received the
[30] W. Li and J. Swetits, “A new algorithm for solving strictly convex quadratic Ph.D. degree in control engineering from the South
programs,” SIAM J. Optim., vol. 7, no. 3, pp. 595–619, 1997. China University of Technology, Guangzhou, China,
[31] Z. Li, X. Cao, and N. Ding, “Adaptive fuzzy control for synchronization in 1990. After a seven-year stint with the University of
of nonlinear teleoperators with stochastic time-varying communication Victoria, Victoria, BC, Canada, he joined Concordia
delays,” IEEE Trans. Fuzzy Syst., vol. 19, no. 4, pp. 745–757, Aug. 2011. University, Montreal, QC, Canada, in 1998.
[32] J. Sun, J. Yi, Z. Pu, and Z. Liu, “Adaptive fuzzy nonsmooth backstepping He is currently with Taizhou University, Taizhou,
output-feedback control for hypersonic vehicles with finite-time conver- China, on leave from the Gina Cody School of En-
gence,” IEEE Trans. Fuzzy Syst., vol. 28, no. 10, pp. 2320–2334, Oct. 2020. gineering and Computer Science, Concordia Univer-
[33] J. Huang, M. Ri, D. Wu, and S. Ri, “Interval type-2 fuzzy logic modeling sity. He has authored or coauthored more than 500
and control of a mobile two-wheeled inverted pendulum,” IEEE Trans. publications, which have appeared in journals, as
Fuzzy Syst., vol. 26, no. 4, pp. 2030–2038, Aug. 2018. book chapters, and in conference proceedings. His research focuses on theory
[34] X. Tao, J. Yi, Z. Pu, and T. Xiong, “Robust adaptive tracking control and its applications to various mechanical systems.
for hypersonic vehicle based on interval type-2 fuzzy logic system and Dr. Su was an Associate Editor for IEEE TRANSACTIONS ON AUTOMATIC
small-gain approach,” IEEE Trans. Cybern., vol. 51, no. 5, pp. 2504–2517, CONTROL and IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY. He
May 2021. has been on the editorial board of several journals, including Mechatronics
[35] H. Han, C. Su, and Y. Stepanenko, “Adaptive control of a class of nonlin- (IFAC). He is a Distinguished Lecturer of the IEEE RA Society and has served
ear systems with nonlinearly parameterized fuzzy approximators,” IEEE for many conferences as an organizing committee member in various capacities,
Trans. Fuzzy Syst., vol. 9, no. 2, pp. 315–323, Apr. 2001. including general chair and program chair.
[36] J. -J. E. Slotine and W. Li, “Adaptive manipulator control a case study: A
case study,” IEEE Trans. Autom. Control, vol. 33, no. 11, pp. 995–1003,
Nov. 1988. Feng Zhao (Member, IEEE) received the B.Eng.
[37] L. Xu and B. Yao, “Adaptive robust precision motion control of linear degree in electronic engineering from the University
motors with negligible electrical dynamics: Theory and experiments,” of Science and Technology of China (USTC), Hefei,
IEEE/ASME Trans. Mechatronics, vol. 6, no. 4, pp. 444–452, Dec. 2001. China, in 2000, and the M.Phil. and Ph.D. degrees
[38] Y. Liu, Z. Li, H. Su, and C. -Y. Su, “Whole-body control of an autonomous in computer vision from The Chinese University of
mobile manipulator using series elastic actuators,” IEEE/ASME Trans. Hong Kong, Hong Kong, in 2002 and 2006, respec-
Mechatronics, vol. 26, no. 2, pp. 657–667, Apr. 2021. tively.
[39] J. Deng, Z. Li, and C. -Y. Su, “Trajectory tracking of mobile robots based From 2006 to 2018, he was a Post-Doctoral Fellow,
on model predictive control using primal dual neural network,” in Proc. Research Fellow, and permanent Senior Lecturer in
33rd Chin. Control Conf., 2014, pp. 8353–8358. Hong Kong, Singapore, and the U.K. Since 2019, he
[40] Y. Chen, Z. Li, H. Kong, and F. Ke, “Model predictive tracking control of has been a Professor and Ph.D. Advisor with USTC.
nonholonomic mobile robots with coupled input constraints and unknown He has authored or coauthored more than 70 papers in refereed international jour-
dynamics,” IEEE Trans. Ind. Inform., vol. 15, no. 6, pp. 3196–3205, nals and conferences, including IEEE Transactions (e.g., IEEE TRANSACTIONS
Jun. 2019. ON MULTIMEDIA, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENS-
[41] Z. Miao, H. Zhong, Y. Wang, H. Zhang, H. Tan, and R. Fierro, “Low ING, IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I), Pattern Recognition,
complexity leader-following formation control of mobile robots using only CVPR, ECCV, IJCAI, and ACM MM. His research interests include image
FOV-constrained visual feedback,” IEEE Trans. Ind. Inform., vol. 18, no. 7, processing, biomedical image analysis, computer vision, pattern recognition,
pp. 4665–4673, Jul. 2022. machine learning, artificial intelligence, brain-inspired vision, and robotics.
[42] Z. Wang, L. Wang, H. Zhang, L. Vlacic, and Q. Chen, “Distributed Prof. Zhao was the recipient of the Best Paper Award in ICIDM 2018,
formation control of nonholonomic wheeled mobile robots subject to SenseTime Scholar 2019, Innovation Scholar 2019 and 2021, Champion of
longitudinal slippage constraints,” IEEE Trans. Syst., Man, Cybern. Syst., IJCAI-PRICAI 2020 3D AI Challenge, Runner-up of CVPR 2021 Streaming
vol. 51, no. 5, pp. 2992–3003, May 2021. Perception Challenge, and Runner-up (3rd) of CVPR 2021 4th UG2+ Challenge.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY CALICUT. Downloaded on January 27,2025 at 19:06:41 UTC from IEEE Xplore. Restrictions apply.