740 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 46, NO.

6, JUNE 2016

Trajectory-Tracking Control of Mobile Robot

Systems Incorporating Neural-Dynamic
Optimized Model Predictive Approach
Zhijun Li, Senior Member, IEEE, Jun Deng, Renquan Lu, Yong Xu, Jianjun Bai,
and Chun-Yi Su, Senior Member, IEEE

AbstractMobile robots tracking a reference trajectory are industry, and is a fundamentally important problem in the
constrained by the motion limits of their actuators, which impose robotics society [1][4]. The application can be found in
the requirement for high autonomy driving capabilities in robots. industry, domestic needs, dangerous areas, and entertainment.
This paper presents a model predictive control (MPC) scheme
incorporating neural-dynamic optimization to achieve trajectory Trajectory-tracking control of a mobile robot refers to the
tracking of nonholonomic mobile robots (NMRs). By using the actual position/orientation of a mobile robot converging to a
derived tracking-error kinematics of nonholonomic robots, the predefined reference path or a path produced by a moving
proposed MPC approach is iteratively transformed as a con- virtual target. From earlier studies, it is well known that an
strained quadratic programming (QP) problem, and then a artificial potential field can track the fixed target or moving vir-
primaldual neural network is used to solve this QP problem
over a finite receding horizon. The applied neural-dynamic opti- tual target [5][8]. The main idea is to attract the robot to the
mization can make the cost function of MPC converge to the target, and repel from obstacles. It is traditional that the robot
exact optimal values of the formulated constrained QP. Compared can be treated as a point mass and its physical limits ignored.
with the existing fast MPC, which requires repeatedly calculat- For example, in the potential field-based control approaches,
ing the Hessian matrix of the Langragian and then solves a the control outputs are normally proportional to the tracking
quadratic program. The computation complexity reaches O(n3 ),
errors; therefore, the control outputs can be very large, when
while the proposed neural-dynamic optimization contains O(n2 )
operations. Finally, extensive experiments are provided to illus- the robot keeps away from the desired positions. On the con-
trate that the MPC scheme has an effective performance on a trary, the robots velocities vary with large fluctuations, when
real mobile robot system. the robot is near the desired position.
Index TermsModel predictive controls (MPCs), neural- Recently, works have been done on the tracking control for
dynamic optimization, quadratic programming (QP) problem, mobile robots [9], [10]. In [11], a complete solution for mobile
trajectory tracking. robots incorporating the dynamical uncertainties was reported.
In [12], adaptive force and motion control were considered
I. I NTRODUCTION for force/motion tracking of mobile manipulators including
ONHOLONOMIC mobile robot (NMR) control has actuator dynamic uncertainties. However, the above reported
N attracted significant attention from both academia and works do not address the inside constraints, including actua-
tor saturation, velocity increment, and some of the states of
Manuscript received April 29, 2015; accepted June 13, 2015. Date of pub- the robots dynamics. To include these constraints in the con-
lication August 31, 2015; date of current version May 13, 2016. This work troller designs, model predictive control (MPC) is an ideal
was supported in part by the Natural Science Foundation of China under
Grant 61174045 and Grant 61573147, in part by the China National Funds for tool, also known as receding horizon control, because it is
Distinguished Young Scientists under Grant 61425009, in part by the Program used to handle constraints through optimization procedures. In
for New Century Excellent Talents in University under Grant NCET-12-0195, MPC, the defined objective function is optimized repeatedly
in part by the Guangzhou Research Collaborative Innovation Projects under
Grant 2014Y2-00507, and in part by the National High-Tech Research and over a finite moving prediction horizon. Utilizing the dynam-
Development Program of China (863 Program) under Grant 2015AA042303. ics model, the MPC can produce the control input for systems
This paper was recommended by Associate Editor S. Tong. within the system constraints. The objective of this paper is to
Z. Li and J. Deng are with the Key Laboratory of Autonomous System
and Network Control, College of Automation Science and Engineering, utilize the feature of MPC combined with the linear variational
South China University of Technology, Guangzhou 200240, China (e-mail: inequalities based primaldual neural network (LVI-PDNN)
zjli@ieee.org). optimization for the trajectory tracking of two-wheeled driven
R. Lu and Y. Xu are with the School of Automation, Guangdong
University of Technology, and Guangdong Key Laboratory of IoT Information robots taking into account the inside constraints. This gives a
Processing, Guangzhou 510006, China (e-mail: rqlu@hdu.edu.cn). major motivation for this paper.
J. Bai is with the Institute of Information and Control, Hangzhou Dianzi MPC can conduct online optimization of an objective func-
University, Hangzhou 310018, China (e-mail: jjbai@hdu.edu.cn).
C.-Y. Su is with the Department of Mechanical and Industrial tion through an inputoutput predictive model over a finite
Engineering, Concordia University, H3G 1M8 Montreal, QC, Canada, (e-mail: future horizon of sample times. An objective function based
cysu@alcor.concordia.ca). on the current controlled variables can be optimally computed
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. at each step. In the next horizon sampling interval, the opti-
Digital Object Identifier 10.1109/TSMC.2015.2465352 mization is reconducted with updated corresponding variables.
2168-2216 c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LI et al.: TRAJECTORY-TRACKING CONTROL OF MOBILE ROBOT SYSTEMS 741

Then, a sequence of control inputs can be obtained at each of redundant manipulators and its physical constraints of
step. Therefore, it is obvious that MPC consists of control joint torques and the generalized forces on the object were
and planning. investigated.
For real-time implementation, online optimization for MPC Therefore, to avoid the shortcomings of significant compu-
implementation is a fundamentally important issue for MPC tational power, this paper proposes a novel model predictive
as applied to applications with high sampling rates. Since controller that incorporates neurodynamics optimization. Since
this optimization requires significant computational power, computation power is severely limited due to the onboard hard-
more computational burden imposes limits on MPC, which ware constraints in mobile robots, only online optimization
is currently applied only to slower dynamic systems such as is allowed in this case. The utilization of the neurodynamics
chemical factories and oil refineries. Traditional optimization optimization needs parallel computation in hardware support-
approaches did not overcome the high-dimensional problems ing. Moreover, model predictive techniques are capable of
and satisfy the real-time computation requirement. In [13], improving the performance and avoid constraint violation, as
based on the receding horizon controller, tracking control they are robust under environmental disturbances and sen-
of NMRs had been investigated. However, the computa- sor errors. However, since the onboard processing capacity
tional complexity of the proposed approach was not analyzed. of mobile robot is limited, the MPC should be augmented by
In [14], tracking control of a mobile robot is achieved using neurodynamics optimization that is responsible for the non-
MPC, and in [15] tracking-error kinematic model for mobile linear optimization. The latter can satisfy real-time processing
robots is linearized as well as analyzed at the equilibrium requirements.
point and transferred to an exact discrete form including the In this paper, consider the physical constraints on mobile
time delay. However, both control schemes were built on robots motion, when it tracks a reference trajectory that
an objective quadratic function derived from the linearized imposes the requirements of high autonomy driving capabili-
dynamics. ties in robots, an MPC method incorporating PDNN optimiza-
In [16], a nonlinear model predictive formation controller tion is proposed for trajectory tracking control of NMRs. By
was used on multiple omnidirectional mobile robots for track- using the derived tracking-error kinematics of the robot, MPC
ing the desired targets, and in [17] a trajectory tracking approach is iteratively transformed as a constrained QP prob-
controller using MPC had been investigated for omnidirec- lem, and then a LVI-PDNN method can be used to solve this
tional mobile robots. However, in both the schemes the instinct QP problem over a finite receding horizon. Another issue of
nature of nonholonomic constraints is not included. In [18], the MPC controller is the high computation cost. In this paper,
tracking control and stabilization of an NMR were consid- a LVI-PDNN method is implemented to handle QP problem
ered and presented using a model predictive controller. In [43], over a finite receding horizon and reduce the computation time.
an approach called fast optimization for real-time control on The applied neural networks can make the cost function of
embedded system (FORCES) was utilized to compute the MPC converge to the exact optimal values of the formulated
control inputs for autonomous racing cars, where linear time constrained QP. Extensive experiments are provided to illus-
varying models were obtained to build local approximations trate that MPC scheme has an effective performance on a real
of nonlinear optimization in the form of convex quadratic mobile robot system.
programs [44], [45]. However, FORCES requires repeated cal- The contributions of this paper can be summarized as
culation of the Hessian matrix of the Langragian to solve a QP, follows.
which brings extra computational complexity. From the above 1) An MPC combined with neural-dynamics optimization
discussion, each scheme suffers from its own disadvantages. In is proposed to achieve state constraint satisfaction and
this paper, motivated by the above results, by considering the successive feasibility for mobile robots.
complete constraints, including nonholonomic and the inside 2) Compared with the standard MPC, where the onboard
constraints (actuator saturation, velocity increments, and some computation power is severely limited, the MPC is
of the variables of robotic dynamics), the trajectory tracking augmented by a LVI-PDNN that is responsible for
control of the NMRs combining the MPC with the computa- cost function optimization. The latter satisfies real-time
tion optimization is developed. The above mentioned works computation requirements.
can be considered as special examples of the development 3) The proposed MPC algorithm is successfully imple-
results. mented to mobile robot control in trajectory tracking
Recently, as increasingly faster computers are being devel- under the velocity constraints and velocity increment
oped, some highly efficient algorithms for MPC bring the constraints (acceleration constraints).
possibility of faster applications. For example, neural net-
work optimization approaches become promising algorithms
for addressing the problem of online heavy computational cost. II. K INEMATICS OF T WO -W HEEL D RIVEN
Some neurodynamics approaches for constrained optimization M OBILE ROBOTS
problems have been discussed for real-time applications The investigated NMR is a typical two-wheel driven mobile
in [19] and [20]. Compared with the traditional optimiza- robot. The velocity of the two driving wheels (l and r )
tion algorithms, these numerous results have shown many can result in linear velocity = (l + r )/2 and angu-
advantages, especially in real-time applications. In [21], lar velocity = (l r )/B with the distance between
based on a dual neural network method, the coordination two wheels being B. The positions of the robot are denoted
742 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 46, NO. 6, JUNE 2016

by (x, y) and the orientation is . Then, the kinematics can be will be lost. Since the proposed MPC does not need the per-
described by sistent excitation condition, we will give the convergence in
MPC framework.
x cos cos 0
x = y = sin = sin 0 u (1)
0 1 III. M ODEL P REDICTIVE C ONTROL D ESIGN
where x = (x, y, )T are state variables, and u = (, )T is One can describe a general form of discrete nonlinear affine
the control vector. system [46]
Similarly, a reference state vector is defined as xr =
x(k + 1) = f (x(k)) + g(x(k))u(k) (8)
(xr , yr , r )T and a reference control signal vector ur =
(r , r )T as follows: subject to the constraints

xr r cos r cos r 0
x(k) X , k = 1, 2, . . . , N (9)
xr = yr = r sin r = sin r 0 ur . (2)
r r 0 1 u(k) U, k = 1, 2, . . . , Nu (10)
The kinematic errors xe can be defined as follows: where x Rn denotes the state vector, u Rm is the
input vector, f () and g() are continuous nonlinear functions,
xe cos sin 0 xr x
xe = ye = sin cos 0 yr y . (3) f (0) = 0, X Rn denotes compact sets, U Rm denotes com-
e 0 0 1 r pact sets containing its internal origin, N denotes prediction
horizon, N 1, and Nu denotes control horizon, N Nu 0.
Based on the error state and its dynamic model, the error MPC can be formulated as an iterative optimization pro-
state dynamics can be written as follows: cedure. Through measuring or estimating the current control
xe = ye + r cos e variables at each step, a cost function is optimized such that
one can get an optimal control input vector; then repeated
ye = xe + r sin e
online calculation of an optimization function is conducted.
e = r . (4) Consider MPC to steer the control variables to the origin; the
Redefine the control signals MPC cost function can be described by
   
u v cos e v 
k+N1
ue = 1 = r . (5) J(x, u) = L(x(j), u(j)) + F(x(k + N)) (11)
u2 r
j=k
Then, the dynamic model (4) can be rewritten as follows:
where L(x, u) denotes the stage cost and satisfies L(0, 0) = 0;
xe
and L(x, u) (|x, u|)2 denotes the terminal cost.
xe = ye
For control stability, the prediction and control horizon
e
N = Nu is sufficiently large, i.e., N = Nu = in stage cost L.
0 0 xe u1 Therefore, one can combine the terminal cost F(x(k + N)) and
= 0 0 ye + r sin e . (6) the state constraint of x(K + N)  with a region in the
0 0 0 e u2 neighborhood of origin  into MPC optimization to achieve
To analyze the local stability for the error state dynamic (4), closed-loop performance.
one can obtain a linearized error state model of (6) Then, the standard quadratic form of J(x, u) can be defined
as follows: as follows,   denotes the Euclidean norm, and Q and R are
positive-definite matrices:
0 r 0 1 0
x e = r 0 r xe + 0 0 ue . (7) 
N u 1
N
0 0 0 0 1 J(k) = x(k + j|k)2Q + u(k + j|k)2R (12)
j=1 j=0
The tracking control objective for the mobile robot (7) has
been transformed into a stabilization issue based on the track- where the predicted state in future horizon is x(k + j|k), the
ing error dynamics model (7). Therefore, consider steering increment of u(k + j|k) is u(k + j|k), with u(k + j|k) =
the position errors to the origin, so that the control inputs u(k + j|k) u(k 1 + j|k).
u1 and u2 need to be designed. It should be noted that the The tracking error dynamics (8) can be rewritten in the form
error dynamics (7) has a form of nonlinear affine system.
Remark 1: Although the error state model (7) is control- xe (k + 1) = f (xe (k)) + g(xe (k))ue (k) (13)
lable, and some local asymptotic stable controllers have been
proposed [22][24], the local linear controllable property subject to the constraints
would be lost when the linear speed and angular speed con-
umin ue (k) umax (14)
verge to zero. Therefore, many controllers developed so far
require this persistent excitation condition [25], [26], i.e., the umin u(k) umax (15)
controlled robot cannot be stopped; otherwise, the stability xmin xe (k) xmax (16)
LI et al.: TRAJECTORY-TRACKING CONTROL OF MOBILE ROBOT SYSTEMS 743

where xe = [x1 x2 x3 ]T = [xe ye e ]T R3 is the state vector; Problem (21) can be rewritten as a QP problem
u = [u1 u2 ]T R2 with u1 = r cos e , u2 = r is the
1
input vector; and min uT Wu + cT u (26)
2
x1 r x2
subject to
f (xe ) = x2 + r x1 + r x3 R3
x3 0 Eu b (27)

1 0 umin u umax (28)
g(xe ) = 0 0 R32
0 1 where the coefficients are
where denotes the sampling time, umin , umin , and xmin  
W = 2 GT QG + R R2Nu 2Nu
denote the lower bound, and umax , umax , and xmax denote  
the upper bound. c = 2GT Q g + f R2Nu
Define the following vectors:  T
E = I, I, G, G R(4Nu +6N)2Nu
x = [xe (k + 1|k), . . . , xe (k + N|k)]T R3N
(17) umin + u(k 1)
u(k) = [ue (k|k), . . . , ue (k + Nu 1|k)]T R2Nu (18) umax u(k 1)
b=
xmin + f + g R
4Nu +6N
.
u(k) = [u(k|k), . . . , u(k + Nu 1|k)] R . (19)
T 2Nu
xmax f g
The predicted output can be written in the following form:
x(k) = Gu(k) + f + g (20) IV. N EURODYNAMICS O PTIMIZATION
From (20), a unified QP formulation for the MPC is devel-
with
oped as (26). The online efficient solution to the QP problem
g(xe (k|k 1)) 0 is needed. In [19] and [20], gradient PDNNs were developed
g(xe (k + 1|k 1)) 0
for robotic manipulators and optimal feet forces distribution
G= .. .. .. and control of quadruped robots, respectively. In this paper,
. . .
the gradient neural-dynamics optimization is adopted to solve
g(xe (k + N 1|k 1)) g(xe (k + N 1|k 1))
the QP problem in the MPC.
R3N2Nu
f (xe (k|k 1))
f (xe (k + 1|k 1)) A. Model Description

f = .. R3N Denoting d as the uper/lower bounds of primaldual
.
decision vector d, one has
f (xe (k + N 1|k 1))      
u umax umin
g(xe (k|k 1)u(k 1)) d= +
, d =
, d = R6Nu +6N
g(xe (k + 1|k 1))u(k 1) y +y+ y

g = .. R3N .
. where the dual decision vector is defined as y R4Nu +6N ,
g(xe (k + N 1|k 1))u(k 1) for any i, the element y+ +
i  0 in y denotes +. Let the
coefficient matrix M and vector p being
Hence, the original optimization problem (12) subject to the
 
constraints (14)(16) becomes W ET
 2 M= R(6Nu +6N)(6Nu +6N)
  E 0
min Gu(k) + f + g + u(k)R 2
(21)  
c
Q
p= R(6Nu +6N) . (29)
b
subject to
Then, the following lemma is need for the subsequence
umin u(k) umax (22) derivation.
umin u(k 1) umax (23) Lemma 1 [19], [20]: Consider the constraints (14)(16),
umin u(k 1) + Iu(k) umax (24) then the QP problem (26)(28) can be converted to the fol-
xmin f + g + Gu(k) xmax (25) lowing LVI problem, and there exists a vector d  =
{d|d  d  d+ } satisfying
I 0 0
I I 0 (d d )T (Md + p) 0, d . (30)

where I = . . .. . R2Nu 2Nu , and umin , umin ,
.. .. . .. From [20], linear variational inequality (30) can be trans-
I I I formed to the piecewise linear equation as follows:
umin , and xmin are the lower bound and umax , umax , umax ,
and xmax denote the upper bound. P (d (Md + p)) d = 0 (31)
744 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 46, NO. 6, JUNE 2016

functions), 2k2 multipliers, and 5k summers, so the proposed

PDNN methods contain O(7(6Nu + 6N) + 2(6Nu + 6N)2 )
operations. To solve QP (26)(28), a traditional sequential
quadratic programming (SQP) using gradient descent meth-
ods is usually adopted, where the computational complexity
requires repeated calculation of the Hessian matrix to solve a
quadratic program [29][31], such as the MATLAB optimiza-
tion functionQUADPROG or LINPROG. Traditional
QP solution needs O(N 4 +N +(6Nu +6N)Nu2 +(7Nu +6N)3 )
operations; it is impossible to be conducted online for mobile
robot systems, owing to inefficient numerical algorithm. When
compared with the computational complexity, the proposed
PDNN approach can reduce the computational cost. For solv-
ing the QP problem (26)(28), the computational time of the
traditional SQP approach costs 0.3 s, while for the PDNN
Fig. 1. Network architecture of PDNN. approach, it is only 0.038 s [note that the experiments are run
on a personal computer with a CPU of Inter Pentium E5700 @
where P () is the piecewise-linear projection operator onto , 3.00 GHz, 2 GB memory], which is smaller when compared
or called piecewise-linear activation function, which can be with the sampling time 0.1 s; so, PDNN method can realize
defined as P (d) = [P (d1 ), . . ., P (u6Nu +6N )]T with real-time purpose.

Theorem 1: Consider nonlinear affine system (8), and
di if di < di
design the MPC input u(k), and the corresponding control
P (di ) = di if di  di  di+ , i {1, . . . , 6Nu + 6N}

+ increment u. Consider any initial variable state, the vector
di if di > di+ . d(t) = [uT , yT ]T of neurodynamics (32) would converge to
To solve linear projection equation (31), one can develop the the optimal results d , among which the first mNu elements
following modified dynamical system for (31) as follows: are the optimal control inputs u in (26)(28). Moreover, the
  convergence is exponential and one can find a constant > 0
d = I + M T {P (d (Md + p)) d} (32) satisfying d P (d (Md + p))22 d d 22 .
with a positive parameter used to scale the convergence rate Proof: See the Appendix.
of neural network model. Remark 3: Consider suitable N, Nu , and Q, the closed-loop
stability can be assured, which can be found in [32] and [33].
B. Network Architecture For example, in [34], considering constrained linear systems,
if horizons are sufficiently large enough, the terminal stability
The PDNN structure is shown in Fig. 1, where i repre- constraints are satisfied.
sents the ith row of the matrix = I + M T , Fi (d) is the ith
row of matrix F = (I M)d, and k = 6Nu + 6N is the dimen- V. E XPERIMENT R ESULTS
sions of d. In Fig. 1, the PDNN belongs to Hopfield neural
network, it is a single-feedback nonlinear network with a sin- The proposed control is implemented on the developed
gle layer structure, and adopts unsupervised learning methods mobile robot in the laboratory. The developed mobile robot
which are based on the Hebb rule. The PDNN is composed of platform is shown in Fig. 2. The aluminum chassis of the
6Nu + 6N neurons, the size is the dimension sum of inequality robot is 45 45 85 cm and contains two driving wheels,
constraint (27), and primal decision vector u. two supporting wheels for balance, a 24-V battery power, a
The ith neuron form is shown as follows: SICK LMS111 Laser Finderm and a Kinect sensor. The diam-
   k   eter of wheels is 19.5 cm and the distance between two wheels
di k
(i)
 is 45 cm. The wheels are driven by motors with rated torque
= i P Fi (d) pi di
dt 72.1 mNm/A at 5200 rpm. The feedback of each motor uti-
i=1 i=1
lizes incremental encoder with 2048 pulses/turn and a gear
i = 1, 2 . . . , k = 6Nu + 6N (33)
box with a factor of 85. The control computer uses Intel(R)
where i = 1, . . . , k, P () is the piecewise-linear activation Pentium(R) E5700 processor (3.0 GHz).
function, from [37] and [38], it can be implemented using The control structure consists of two levels as shown in
operational amplifiers. Fig. 3. In the high-level control, the algorithms are devel-
Remark 2: PDNN does not depend on penalty or analog oped in Visual C++ environments and run on an on-board
parameters, matrix inverses, or high-order nonlinear terms, computer with a sampling interval of 100 ms. The on-board
only with simple vector or matrix augmentation and operation. computer communicates with the Elmo driver through CAN
Consequently, the architecture of PDNN to be implemented bus, which controls the servo motor through computed torque
by analog circuits could be much simpler than those of the control. Wheel encoders are sent back to the Elmo driver and
existing recurrent neural network methods [39], [40]. In view on-board computers. The lower level control realizes computer
of (33) and Fig. 1, the circuit for PDNN (32) consists of k torque of the high-level commands. The Elmo driver con-
integrators, k limiters (being the piecewise-linear activation troller performs three basic tasks: 1) communicating with the
LI et al.: TRAJECTORY-TRACKING CONTROL OF MOBILE ROBOT SYSTEMS 745

Fig. 5. Errors of tracking circle.

Fig. 2. Developed mobile service robot.

Fig. 6. Convergent behaviors of the PDNN for tracking circle.

Fig. 3. Control architecture of the mobile robot.

In these experiments, the two tracking trajectories are pro-
vided. First, the chosen reference trajectory is a circle, which
can be described as follows:
xr = 2 cos( )
yr = 2 sin( ).
The parameters of MPC are chosen as N = 3, Nu = 2, Q = I,
R = 0.1I, umin = [1, . . . , 1] R4 , umax = [1, . . . , 1] R4 ,
xmin = [ 1, . . . , 1] R9 , xmax = [1, . . . , 1] R9 , umin =
[ 0.1, . . . , 0.1] R4 , and umin = [0.1, . . . , 0.1] R4 .
The optimization problem (26)(28) can be computed by using
the PDNN (32) during each 100 ms sampling interval. The
experiment results are listed in Figs. 46, where Figs. 4 and 5
show the trajectory tracking and tracking errors by the mobile
robot with respect to the reference trajectory, respectively.
Fig. 6 shows the output u in each sample time.
Second, the chosen reference trajectory is a lane-change
curve, whose equation can be described as follows:
Fig. 4. Tracking a circle.
2
yr = .
higher-level on-board computer through Kvaser CAN device; 1 + e2(xr 2)
2) reading encoder feedback; and 3) generating computed The parameters of MPC are chosen as N = 3, Nu = 2, Q = 5I,
torques. R = I, umin = [ 1, . . . , 1] R4 , umax = [1, . . . , 1] R4 ,
746 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 46, NO. 6, JUNE 2016

Fig. 7. Tracking a lane-change curve.

Fig. 10. Tracking a circle using SQP-based MPC.

Fig. 8. Errors of tracking lane-change curve.

Fig. 11. Errors of tracking circle using SQP-based MPC.

track the desired lane-change curve trajectory in real time.

Fig. 9 shows the output u.
In contrast, since SQP has been widely utilized in the MPC.
To highlight the performance of our presented approach, the
traditional SQP for comparison is chosen. Under the same cho-
sen MPC parameters and the sampling step, the control results
Fig. 9. Convergent behaviors of the PDNN for tracking lane-change curve. are listed in Figs. 1015. Figs. 10 and 13 show the tracking
trajectories of the mobile robot, and the tracking errors are
shown in Figs. 11 and 14. The produced inputs of SQP are
xmin = [ 1, . . . , 1] R9 , xmax = [1, . . . , 1] R9 , umin = shown in Figs. 12 and 15. In Fig. 10, we can see the track-
[ 0.1, . . . , 0.1] R4 , and umin = [0.1, . . . , 0.1] R4 . ing errors are larger and in Fig. 13, we can see the tracking
The optimization problem (26)(28) can be computed by using errors become larger with time. In Fig. 13, it is obvious that
PDNN (32) during each 100 ms sampling step. The experiment the actual trajectories deviates the desired one.
results are listed in Figs. 79, where the tracking trajectory of After comparing these experimental results, it can be con-
the mobile robot is shown in Fig. 7, and the tracking errors are cluded that the proposed method has the better performance
depicted in Fig. 8. Through implementing the proposed MPC and more suitable for practice implementation. From track-
approach, we can see that the mobile robot can successfully ing performance, we can see the proposed control is with
LI et al.: TRAJECTORY-TRACKING CONTROL OF MOBILE ROBOT SYSTEMS 747

Fig. 12. Control inputs of SQP-based MPC. Fig. 15. Control inputs of SQP-based MPC.

lots of computation source and is more dependent on initial

values.

VI. C ONCLUSION
This paper proposes an MPC scheme incorporating PDNN
optimization to achieve trajectory tracking of NMR. By using
the derived tracking-error kinematics of nonholonomic robots,
the proposed MPC approach is iteratively transformed as a
constrained QP, then a LVI-PDNN over a finite receding hori-
zon is employed. Finally, extensive experiments are conducted
to illustrate the performance of the proposed MPC scheme on
a real mobile robot system.

A PPENDIX
Proof of Theorem 1: Consider nonlinear affine system (8),
we need to design the MPC u(k) and the corresponding con-
Fig. 13. Tracking a lane-change curve using SQP-based MPC. trol increment u. For any initial variable state, we define
the primaldual decision vector d(t) = [uT , yT ]T . To prove
exponential convergence of the system, we firstly introduce
projection related inequalities (P ()  )T ( P ())
0( Rn+m+dim(b) ,  ), we can obtain
(P (d (Md + p) d )T
(d (Md + p) P (d (Md + p))) 0 (34)
it can be written as
(d P (d (Md + p)))T
((Md + p) d + P (d (Md + p))) 0. (35)
Combining (35) and (31), we have
(d P (d (Md + p)))T (Md + p) 0. (36)
Then, combining (35) and (36), we can readily have
(d P (d (Md + p)))T
Fig. 14. Errors of tracking lane-change curve using SQP-based MPC. (M(d d ) d + P (d (Md + p))) 0. (37)
The above (37) can be further written as follows:
better performance in comparison with MPC using SQP. The
main difference is that the proposed MPC is more suitable for (d d + d P (d (Md + p)))T
the on-board computation, while the SQP-based MPC needs (M(d d ) d + P (d (Md + p))) 0. (38)
748 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 46, NO. 6, JUNE 2016

Then, the above (38) can be expanded as follows: [5] L. Huang, Velocity planning for a mobile robot to track a moving
   targetA potential field approach, Robot. Auton. Syst., vol. 57, no. 1,
d d I + M T (d P (d (Md + p))) pp. 5563, 2009.
    [6] J. Yan, X.-P. Guan, and F.-X. Tan, Target tracking and obstacle avoid-
d P (d (Md + p))22 + d d M d d (39) ance for multi-agent systems, Int. J. Autom. Comput., vol. 7, no. 4,
pp. 550556, Nov. 2010.
due to [7] Y. Zhu, T. Zhang, J. Song, X. Li, and M. Nakamura, A new method
  for mobile robots to avoid collision with moving obstacle, Artif. Life
M + MT T W 0
d Md = d
T T
d=d d0 (40) Robot., vol. 16, no. 4, pp. 507510, 2012.
2 0 0 [8] M. Defoort, A. Doniec, and N. Bouraqadi, Decentralized robust colli-
sion avoidance based on receding horizon planning and potential field
so M is required a positive semi-definite, but not required to for multi-robots systems, in Informatics in Control Automation and
be symmetric [42]. Then we have Robotics (LNEE 85), J. A. Cetto, J. Filipe, and J.-L. Ferrier, Eds. Berlin,
   Germany: Springer, 2011, pp. 201215.
d d I + M T (d P (d (Md + p))) [9] B. S. Park, S. J. Yoo, J. B. Park, and Y. H. Choi, A simple adaptive
 2 control approach for trajectory tracking of electrically driven nonholo-
d P (d (Md + p))2 + d d  0. (41)
2 2
nomic mobile robots, IEEE Trans. Control Syst. Technol., vol. 18, no. 5,
pp. 11991206, Sep. 2010.
We select a Lyapunov function V(d) = d d 22
0, where [10] K. Shojaei and A. M. Shahri, Output feedback tracking control of
uncertain non-holonomic wheeled mobile robots: A dynamic surface
V(d) = 0 if and only if d = d . With the PDNN in (32), its control approach, IET Control Theory Appl., vol. 6, no. 2, pp. 216228,
derivative is Jan. 2012.
 [11] M. Oya, C.-Y. Su, and R. Katoh, Robust adaptive motion/force tracking
dV(d) V(d) T dd control of uncertain nonholonomic mechanical systems, IEEE Trans.
=
dt d dt Robot. Autom., vol. 19, no. 1, pp. 175181, Feb. 2003.

   [12] Z. Li, S. S. Ge, M. Adams, and W. S. Wijesoma, Adaptive robust
= d d I + M T (P (d (Md + p)) d) output-feedback motion/force control of electrically driven nonholo-
  
= d d I + M T (d P (d (Md + p)))
nomic mobile manipulators, IEEE Trans. Control Syst. Technol., vol. 16,
no. 6, pp. 13081315, Nov. 2008.
 2
= d P (d (Md + p))2 d d  0.
2 M
[13] D. Gu and H. Hu, Receding horizon tracking control of wheeled mobile
robots, IEEE Trans. Control Syst. Technol., vol. 14, no. 4, pp. 743749,
(42) Jul. 2006.
[14] G. Klancar and I. Skrjanc, Tracking-error model-based predictive con-
Based on Lyapunov theory [41], we can see that the state trol for mobile robots in real time, Robot. Auton. Syst., vol. 55, no. 6,
pp. 460469, 2007.
vector d(t) of the PDNN in (32) is stable, and it is also globally [15] Y. Gao, C. G. Lee, and K. T. Chong, Receding horizon tracking con-
convergent to an equilibrium d in the light of V(d) = 0 with trol for wheeled mobile robots with time-delay, J. Mech. Sci. Technol.,
d = 0 and d = d . Based on (30) and (31), we can get vol. 22, no. 12, pp. 24032416, 2008.
that the solution of the LVI problem in (30) is d , and the [16] T. P. Nascimentoa, A. P. Moreirab, and A. S. Conceicao, Multi-robot
nonlinear model predictive formation control: Moving target and target
optimal solution of QP problem in (26)(28) is the first 2Nu absence, Robot. Auton. Syst., vol. 61, no. 12, pp. 15021515, 2013.
elements of d . [17] Z. Zeng, H. Lu, and Z. Zheng, High-speed trajectory tracking based
Then, by using V(d), V(d), and the extra condition, it can on model predictive control for omni-directional mobile robots, in
Proc. 25th Chin. Control Decis. Conf. (CCDC), Guiyang, China, 2013,
be obtained as follows: pp. 31793184.
dV(d) [18] M.-M. Ma, S. Li, and X.-J. Liu, Tracking control and stabilization of
d P (d (Md + p))22 d d 2M (43) wheeled mobile robots by nonlinear model predictive control, in Proc.
dt
 2  2 31st Chin. Control Conf., Hefei, China, Jul. 2012, pp. 40564061.
d d 2 d d M (44) [19] Z. Li, S. S. Ge, and S. Liu, Contact-force distribution optimization
    and control for quadruped robots using both gradient and adaptive neu-
d d (I + M) d d
T
(45) ral networks, IEEE Trans. Neural Netw. Learn. Syst., vol. 25, no. 8,
pp. 14601473, Aug. 2014.
V(d) (46) [20] Y. Zhang, S. S. Ge, and T. H. Lee, A unified quadratic-programming-
based dynamical system approach to joint torque optimization of
where = > 0, it is easy to have V(d) = physically constrained redundant manipulators, IEEE Trans. Syst., Man,
O(e(tt0 ) ), t t0 , then we have d d 2 = Cybern. B, Cybern., vol. 34, no. 5, pp. 21262132, Oct. 2004.
O(e((tt0 ))/2 ), t t0 , with /2 (i.e., /2) is the expo- [21] Z. Hou, L. Cheng, and M. Tan, Multicriteria optimization for coordina-
tion of redundant robots using a dual neural network, IEEE Trans. Syst.,
nential convergence rate of PDNN in (32). The proof is end. Man, Cybern. B, Cybern., vol. 40, no. 4, pp. 10751087, Aug. 2010.
[22] C. Samson and K. Ait-Abderrahim, Feedback control of a nonholo-
nomic wheeled cart in cartesian space, in Proc. IEEE Int. Conf. Robot.
R EFERENCES Autom., Sacramento, CA, USA, 1991, pp. 11361141.
[1] A. Mora, D. F. Glas, T. Kanda, and N. Hagita, A teleoperation approach [23] A. M. Bloch, M. Reyhanoglu, and N. H. McClamroch, Control and
for mobile social robots incorporating automatic gaze control and three- stabilization of nonholonomic dynamic systems, IEEE Trans. Autom.
dimensional spatial visualization, IEEE Trans. Syst., Man, Cybern., Control, vol. 37, no. 11, pp. 17461757, Nov. 1992.
Syst., vol. 43, no. 1, pp. 630642, Feb. 2013. [24] A. M. Bloch and S. Drakunov, Stabilization of a nonholonomic sys-
[2] K. Zheng, D. F. Glas, T. Kanda, and H. Ishiguro, Designing and imple- tem via sliding modes, in Proc. 33rd IEEE Conf. Decis. Control,
menting a humanrobot team for social interactions, IEEE Trans. Syst., Lake Buena Vista, FL, USA, 1994, pp. 29612963.
Man, Cybern., Syst., vol. 43, no. 4, pp. 843859, Feb. 2013. [25] T. Fukao, H. Nakagawa, and N. Adachi, Adaptive tracking control of a
[3] A. Konar, I. G. Chakraborty, S. J. Singh, L. C. Jain, and A. K. Nagar, A nonholonomic mobile robot, IEEE Trans. Robot. Autom., vol. 16, no. 5,
deterministic improved Q-Learning for path planning of a mobile robot, pp. 609615, Oct. 2000.
IEEE Trans. Syst., Man, Cybern., Syst., vol. 43, no. 5, pp. 11411153, [26] Z. P. Jiang and H. Nijmeijer, Tracking control of mobile robots: A
Aug. 2013. case study in backstepping, Automatica, vol. 33, no. 7, pp. 13931399,
[4] J. Su, Representation and inference of user intention for Internet robot, 1997.
IEEE Trans. Syst., Man, Cybern., Syst., vol. 44, no. 8, pp. 9951002, [27] L. Wang and F. Wan, Structured neural networks for constrained model
Jan. 2014. predictive control, Automatica, vol. 37, no. 8, pp. 12351243, 2001.
LI et al.: TRAJECTORY-TRACKING CONTROL OF MOBILE ROBOT SYSTEMS 749

[28] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, Jun Deng received the B.Eng. degree in control sci-
Constrained model predictive control: Stability and optimality, ence and engineering from the College of Electrical
Automatica, vol. 36, no. 6, pp. 789814, 2000. Engineering and Automation, Jiangxi University of
[29] M. S. Bazaraa, H. D. Sherali, and C. M. Shetty, Nonlinear Programming Science and Technology, Ganzhou, China, in 2012.
Theory and Algorithms. New York, NY, USA: Wiley, 1993. He is currently pursuing the masters degree from
[30] F. T. Cheng, R. J. Sheu, and T. H. Chen, The improved compact QP the College of Automation Science and Engineering,
method for resolving manipulator redundancy, IEEE Trans. Syst., Man, South China University of Technology, Guangzhou,
Cybern., vol. 25, no. 11, pp. 15211530, Nov. 1995. China.
[31] W. Li and J. Swetits, A new algorithm for solving strictly convex His current research interests include mobile
quadratic programs, SIAM J. Optim., vol. 7, no. 3, pp. 595619, 1997. robot, model predictive control, neural network
[32] L. Grune and J. Pannek, Nonlinear Model Predictive Control: Theory control, and optimization.
and Algorithms. London, U.K.: Springer, 2011.
[33] J. Garriga and M. Soroush, Model predictive control tuning methods:
A review, Ind. Eng. Chem. Res., vol. 49, no. 8, pp. 35053515, 2010. Renquan Lu received the Ph.D. degree in control
[34] P. Scokaert and J. Rawlings, Constrained linear quadratic regulation, science and engineering from Zhejiang University,
IEEE Trans. Autom. Control, vol. 43, no. 8, pp. 11631169, Aug. 1999. Hangzhou, China, in 2004.
[35] J. A. Primbs and V. Nevistic, Feasibility and stability of constrained He is currently a Full Professor with the Institute
finite receding horizon control, Automatica, vol. 36, no. 7, pp. 965971, of Information and Control, Hangzhou Dianzi
2000. University, Hangzhou, China. His current research
[36] A. Jadbabaie and J. Hauser, On the stability of receding horizon control interests include robust control, singular systems,
with a general terminal cost, IEEE Trans. Autom. Control, vol. 50, no. 5, and complex systems.
pp. 674678, May 2005.
[37] A. Cichocki and R. Unbehauen, Neural Networks for Optimization and
Signal Processing. Chichester, U.K.: Wiley, 1993.
[38] H. Hikawa, A digital hardware pulse-mode neuron with piecewise
linear activation function, IEEE Trans. Neural Netw., vol. 14, no. 5,
Yong Xu was born in Zhejiang Province, China,
pp. 10281037, Sep. 2003.
in 1983. He received the B.S. degree in information
[39] D. W. Tank and J. J. Hopfield, Simple neural optimization networks: An
engineering from Nanchang Hangkong University,
A/D converter, signal decision circuit, and a linear programming circuit,
Nanchang, China, in 2007, the M.S. degree in
IEEE Trans. Circuits Syst., vol. 33, no. 5, pp. 533541, May 1986.
information and control from Hangzhou Dianzi
[40] M. P. Kennedy and L. O. Chua, Neural networks for nonlinear pro-
University, Hangzhou, China, in 2010, and the Ph.D.
gramming, IEEE Trans. Circuits Syst., vol. 35, no. 5, pp. 554562,
degree in cyber-system and control from Zhejiang
May 1988.
University, Hangzhou.
[41] Y. Zhang, On the LVI-based primal-dual neural network for solving
He is currently a Lecturer with the Institute
online linear and quadratic programming problems, in Proc. Amer.
of Information and Control, Hangzhou Dianzi
Control Conf., Portland, OR, USA, 2005, pp. 13511356.
University. His current research interests include
[42] Y. Xia and J. Wang, A recurrent neural network for solving linear
networked control systems, estimation, and stochastic systems.
projection equations, Neural Netw., vol. 13, no. 3, pp. 337350, 2000.
[43] A. Liniger, A. Domahidi, and M. Morari, Optimization-based
autonomous racing of 1:43 scale RC cars, Optim. Control Appl. Method,
vol. 36, no. 5, pp. 628647, Sep./Oct. 2015. Jianjun Bai received the B.S. degree in control
[44] A. Domahidi, A. Zgraggen, M. N. Zeilinger, M. Morari, and C. N. Jones, science and engineering from the University of
Efficient interior point methods for multistage problems arising in Petroleum China, Beijing, China, in 2006, and the
receding horizon control, in Proc. IEEE Conf. Decis. Control (CDC), Ph.D. degree from Zhejiang University, Hangzhou,
Maui, HI, USA, Dec. 2012, pp. 668674. China.
[45] A. Domahidi. (2012). FORCES: Fast Optimization for Real-Time He is currently a Lecturer with the Institute
Control on Embedded Systems. [Online]. Available: http://forces.ethz.ch, of Information and Control, Hangzhou Dianzi
accessed Mar. 2014. University, Hangzhou. His current research interests
[46] C. T. Chen, Linear System Theory and Design. London, U.K.: Oxford include robust control, networked control systems,
Univ. Press, 1995. and multirobot system.

Zhijun Li (M07SM09) received the Dr.Eng.

degree in mechatronics from Shanghai Jiao Tong
University, Shanghai, China, in 2002. Chun-Yi Su (SM99) received the Ph.D. degree in
From 2003 to 2005, he was a Post-Doctoral control engineering from the South China University
Fellow with the Department of Mechanical of Technology, Guangzhou, China, in 1990.
Engineering and Intelligent Systems, University He is with the College of Automation Science and
of Electro-Communications, Tokyo, Japan. From Engineering, South China University of Technology,
2005 to 2006, he was a Research Fellow with on leave from Concordia University, Montreal, QC,
the Department of Electrical and Computer Canada. After a seven-year stint at the University of
Engineering, National University of Singapore, Victoria, Victoria, BC, Canada, he joined Concordia
Singapore, and Nanyang Technological University, University, Montreal, in 1998. He conducts research
Singapore. From 2007 to 2011, he was an Associate Professor with the in the application of automatic control theory to
Department of Automation, Shanghai Jiao Tong University. In 2008, he mechanical systems. He has authored or co-authored
was a Visiting Scholar with Microsoft Research Asia, Beijing, China. Since over 300 publications, which have appeared in journals, book chapters, and
2012, he has been a Professor with the College of Automation Science and conference proceedings. His current research interests include control of sys-
Engineering, South China university of Technology, Guangzhou, China. Since tems involving hysteresis nonlinearities.
2015, he has been a Visiting Professor with Faculty Science and Technology, Dr. Su has served as an Associate Editor for the IEEE T RANSACTIONS ON
University of Macau, Macau, China. He has coauthored three books AUTOMATIC C ONTROL, the IEEE T RANSACTIONS ON C ONTROL S YSTEMS
published in Springer and CRC Press, and over 100 international journal and T ECHNOLOGY, and the Journal of Control Theory and Applications. He
conference papers. His current research interests include nonlinear control, was on the Editorial Board of 18 journals, including the International
neural network optimization, service robotics, and tele-operation systems. Federation of Automatic Control journals of Control Engineering Practice and
Dr. Li is an Editor-at-Large of the Journal of Intelligent and Robotic Mechatronics. He has also severed for several conferences as an Organizing
Systems, an Associate Editor of the IEEE T RANSACTIONS ON N EURAL Committee Member, including the General Co-Chair of the 2012 IEEE
N ETWORKS AND L EARNING S YSTEMS, and an Editorial Board Member of International Conference on Mechatronics and Automation, and the Program
several international journals. Chair of the 2007 IEEE Conference on Control Applications.