Proceedings of the 33rd Chinese Control Conference

July 28-30, 2014, Nanjing, China

Trajectory Tracking of Mobile Robots Based on Model Predictive

Control Using Primal Dual Neural Network
DENG Jun , LI Zhijun , SU Chun-Yi
The Key Lab of Autonomous System and Network Control
College of Automation Science and Engineering
South China University of Technology, Guangzhou, China
E-mail: zjli@ieee.org.

Abstract: Mobile robots motion is constrained by the maximum velocity its actuators can provide, when it tracks a reference
trajectory which imposes demanding requirements on the robots driving capabilities. In this paper, a model predictive control
(MPC) scheme is proposed for trajectory tracking control of two-wheel mobile robots. Based on the derived tracking-error kine-
matics of the robot, the proposed MPC approach can be iteratively formulated as a quadratic programming (QP) problem, which
can be solved using a linear variable inequality based primal-dual neural network (LVI-PDNN) over a nite receding horizon.
The applied neural networks are both stable in the sense of Lyapunov and globally convergent to the exact optimal solutions of
reformulated convex programming problems. The smoothness of the robot motion is improved, a reasonable magnitudes of the
robot velocities and a better tracking performance are achieved. Simulation and experimental results are provided to demonstrate
the effectiveness and characteristics of the proposed LVI-PDNN based MPC approaches to trajectory tracking control.
Key Words: Model Predictive Control (MPC), Quadratic Programming (QP) Problem, LVI-based Primal-dual Neural Network
(LVI-PDNN), Trajectory Tracking

1 Introduction bile robots is designed by using infrared sensors, and it uti-

lizes the fuzzy sliding-mode control scheme to realize re-
The control of mobile robots is a fundamentally impor- lated tracking behavior. In [7], the tracking control of an
tant issue in robotics, which has attracted signicant atten- electrically driven nonholonomic mobile robot with model
tion from both academia and industry. The applications uncertainties in the robot kinematics, the robot dynamics,
can be found in industry, domestic needs, dangerous areas and the wheel actuator dynamics was investigated, fuzzy
and entertainment. According to Brocketts condition, non- logic systems were utilized to learn the behavior of the un-
holonomic mobile robots cannot be asymptotically stabilized known dynamics of the robot and the wheel actuators. In
around an equilibrium using smooth time-invariant feedback [8], a simple adaptive control approach was proposed for
[5]. Completely nonholonomic, driftless systems are con- path tracking for uncertain nonholonomic mobile robots in-
trollable in a nonlinear sense; therefore, asymptotic stabi- corporating actuator dynamics. In [9], an observer-based tra-
lization can be obtained using time-varying, discontinuous jectory tracking controller is proposed for a wheeled mobile
or hybrid control laws [6]. robot model.
Tracking control refers to forcing the position and orienta- The methods above try to solve the problem of trajectory
tion of a mobile robot to follow a predened reference path tracking in different ways. However, there still exist some
or a path produced by a moving virtual target. It is well problems to be resolved in different degrees. For example,
known that articial potential elds, can achieve the purpose in the above algorithms, the constraint limits of the robot ac-
[14]. The main idea is to attract the robot to the target and tuators have not be considered in the control design, so it is
to repulse it from obstacles. It is traditional that the robot always assumed that the actuators can provide the desired
is modelled as a point mass and physical limit of the robots torques as needed. Thus, further research is needed to nd
velocity is ignored. Much research efforts have been done to a satisfactory solution. MPC can conduct online optimiza-
avoid local minima where the robot is trapped before reach- tion of an objective function through a inputoutput predic-
ing the target, and extend it for the tasks involving multi- tor model over a nite horizon of future time.
targets or multi-obstacles. The outputs of the potential eld- On the other hand, neural network optimization ap-
based controller are normally proportional to the tracking er- proaches emerge as a promising approach for addressing
rors. Very large control efforts are generated when the robot the issue of heavy online computational burden. Neural
is far away from the target, or large uctuations of the robot networks for constrained optimization problems have been
velocities occur when the robot is near the target. widely explored for real-time applications [15, 16]. Re-
In recent years, there have been many works on tracking ported results of numerous investigations have shown many
control for mobile robots. For example, in [10], a real-time advantages over the traditional optimization algorithms. In
fuzzy target tracking control scheme for autonomous mo- [13], a dual neural-network method for the coordination of
kinematically redundant robots was proposed, the physical
This work is supported in part by National Natural Science Founda- limits of both joint torques and the generalized forces ap-
tion of China grants 61174045, the Fundamental Research Funds for the
plied on the object are considered, which makes the origi-
Central Universities under Grant 2013ZG0035, the Program for New Cen-
tury Excellent Talents in University under Grant NCET-12-0195, and the nal coordination problem become a complicated optimiza-
Ph.D. Programs Foundation of Ministry of Education of China under Grant tion problem subject to both equality and inequality con-
20130172110026. To whom all correspondences should be addressed: Zhi- straints. In [11], a three-layer neural network was applied to
jun Li.

8353
linear MPC to compute the exact optimal solution. In [12], Redening the control signals
a dual neural network was applied for multivariable general-    
ized predictive control. In [14], the echo state network was u1 vr cos e v
ue = = (5)
used to model the unknown nonlinear autoregressive exoge- u2 r
nous (NARX) systems and the simplied dual network was Then, the error state dynamic model (4) becomes
applied for solving the reformulated quadratic optimization

problems in MPC. xe
In this paper, considering the physical constraints on mo- xe = ye
bile robots motion, when it tracks a reference trajectory e
which imposes demanding requirements on the robots driv-
0 0 xe u1
ing capabilities, a model predictive control (MPC) scheme is
= 0 0 ye + r sin e (6)
proposed for trajectory tracking control of two-wheel mo-
0 0 0 e u2
bile robots. The proposed MPC approach can be itera-
tively formulated as a quadratic programming (QP) problem, To analyze the local stability for the error state dynamic
which can be solved using a linear variable inequality based model (4), the linearized error state model of (6) can be ob-
primal-dual neural network (LVI-PDNN) over a nite reced- tained as follows:
ing horizon. Simulation and experimental results are pro-
vided to demonstrate the effectiveness and characteristics of 0 r 0 1 0
the proposed LVI-PDNN based MPC approaches to trajec- x e = r 0 r xe + 0 0 ue (7)
tory tracking control. 0 0 0 0 1

2 Kinematics of Nonholonomic Mobile Robots The tracking problem for the mobile robot (1) has been
converted to a stabilization problem for the tracking error
The investigated mobile robot is a typical nonholonomic dynamics model (7). The control objective is to steer the po-
mobile robot, which has two driving wheels and two castors, sition errors to the origin with inputs u1 and u2 . It is worth
which is shown in Fig. 4. The speed control of the two pointing out that the error dynamics equation (7) is a nonlin-
driving wheels (l and r ) leads to the control of linear speed ear afne system.
= (l + r )/2, and angular speed = (l r )/B, where
B is the distance of two wheels. The state of the robot can 3 Model Predictive Control Scheme
be described by its position (x, y), and its orientation . The Consider a discrete nonlinear afne system in the general
kinematics equation is written as follows: form [17]

x cos cos 0 x(k + 1) = f (x(k)) + g(x(k))u(k) (8)
x = y = sin = sin 0 u (1)
0 1 subject to constraints

The state and control signal vectors are denoted as x = x(k) X , k = 1, 2, . . . , N (9)
(x, y, )T and u = (, )T . u(k) U , k = 1, 2, . . . , Nu (10)
The reference trajectories can also be described by a refer-
ence state vector xr = (xr , yr , r )T and a reference control where x Rn is the state vector, u Rm is the input vector,
signal vector ur = (r , r )T and have the same constraints f () and g() are continuous nonlinear functions, f (0) = 0,
as (1) X Rn and U Rm are compact sets and contain the
origin in their interiors, and N and Nu are prediction horizon

xr r cos r cos r 0 1 N and control horizon 0 Nu N , respectively.
xr = yr = r sin r = sin r 0 ur (2) MPC is an iterative optimization technique: at each sam-
r r 0 1 pling time, measure or estimate the current state, and opti-
mize a cost function to obtain an optimal input vector. It-
The kinematic errors xe can be dened as follows: eratively solving an optimization problem online makes the
key difference between MPC and standard control strategies
xe cos sin 0 xr x where the control laws are obtained a priori. Assume that the
xe = ye = sin cos 0 yr y (3) control objective for system (8) in MPC is to steer the state
e 0 0 1 r to the origin, then the cost function for MPC can be dened
as
By using the error state and its dynamic model, the track-
k+N
1
ing control problem is converted into a regulation problem.
The error state chosen in a rotated coordinate frame (3) can J(x, u) = L(x(j), u(j)) + F (x(k + N )) (11)
simplify the dynamic model. The error state dynamic model j=k

is derived as follows: where L(x, u) is the stage cost and is assumed to satisfy
L(0, 0) = 0. From a theoretical point of view, an innite
xe = ye + r cos e
prediction and control horizon, i.e., N = Nu = in stage
ye = xe + r sin e cost L is desirable as it will guarantee stability. As a re-
e = r (4) sult, the terminal cost F (x(k +N )) and a terminal constraint

8354

x(K + N ) , where is a region in the neighborhood g(xe (k|k 1)u(k 1))
g(xe (k + 1|k 1))u(k 1)
of origin, are incorporated to MPC optimization problem to
have better closed-loop performance. g = .. R3N .
.
As a result, the following standard quadratic form of
g(xe (k + N 1|k 1))u(k 1)
J(x, u) without terminal constraint and terminal region is
most often used in practice: Hence, the original optimization problem (12) subject to
constraint (14)(16) becomes
N
 N
u 1

J(k) = x(k + j|k)2Q + u(k + j|k)2R (12) minGu(k) + f + g2Q + u(k)2R (23)
j=1 j=0

where x(k + j|k) denotes the predicted state in future subject to

horizon; u(k + j|k) denotes the input increment, where
u(k + j|k) = u(k + j|k) u(k 1 + j|k); Q and R umin u(k) umax (24)
are appropriate weighting matrices; and   denotes the Eu- umin u(k 1) umax (25)
clidean norm of the corresponding vector. The major benet umin u(k 1) + Iu(k) umax (26)
of using cost function (12) is that it results in a quadratic xmin f + g + Gu(k) xmax (27)
optimization problem for system (8) whose optimal solution
can be obtained reliably and efciently.
I 0 0
The equation (8) can be rewritten in the form I I 0

where I = .. .. .. .. R2Nu 2Nu
xe (k + 1) = f (xe (k)) + g(xe (k))ue (k) (13) . . . .
I I I
subject to constraints
Problem (23) can be rewritten as a QP problem
umin ue (k) umax (14)
umin u(k) umax (15) 1
min uT M u + pT u (28)
xmin xe (k) xmax (16) 2

where xe = [x1 , x2 , x3 ]T = [xe , ye , e ]T R3 is the state subject to

vector, u = [u1 , u2 ]T = [r cos e , r ]T R2 is the
F u q (29)
input vector, and from (7) and (13), we can get
umin u umax (30)

x1 r x2 where the coefcients are
f (xe ) = x2 + r x1 + r x3 R(17)
3

x3 0 M = 2(GT QG + R) R2Nu 2Nu

1 0 p = 2GT Q(g + f) R2Nu
g(xe ) = 0 0 R32 (18) I,
F = [I, G, G]T R(4Nu +6N )2Nu
0 1
umin + u(k 1)
where is the sampling period. umax u(k 1)
q=
xmin + f + g R
4Nu +6N
Dene the following vectors:
xmax f g
x = [xe (k + 1|k), . . . , xe (k + N |k)]T R3N (19)
u(k) = [ue (k|k), . . . , ue (k + Nu 1|k)]T R2Nu (20) 4 LVI-Based Primal-dual Neural Network Opti-
u(k) = [u(k|k), . . . , u(k + Nu 1|k)]T R2Nu (21) mization
The predicted output is then expressed in the following Consider (22), a unied QP formulation for the model
form: predictive control has been developed for missile guidance.
However, the online efcient solution to the QP problem is
x(k) = Gu(k) + f + g (22) needed to be developed. In [15] and [16], LVI (linear vari-
ational inequalities)-based primal-dual neural network was
G
where
g(xe (k|k 1)) 0 developed for robotic manipulator and optimal feet forces
g(xe (k + 1|k 1)) 0 distribution and control of quadruped robots. In this paper,

= .. .. .. we utilize the LVI primal-dual neural network for perform-
. . .
ing the optimization of the error dynamics. It is straightfor-
g(xe (k + N 1|k 1)) g(xe (k + N 1|k 1)) ward to obtain the optimization solution u
R3N
2Nu
,
f (xe (k|k 1)) d = (I + Z T ){J ((I Z)d a) d}. (31)
f (x (k + 1|k 1))
e
f = .. R3N ,
.   d is the primal-dual decision vector, it dened as d :=
where
u
f (xe (k + N 1|k 1)) , with h R4Nu +6N being the corresponding dual
h

8355
 1 0.6
desired trajectory error of x(m)
actual trajectory error of y(m)
0.8 0.4 error of angle(rad)

0.6
0.2

0.4
0
y coordinate(m)

0.2
0.2

error
0
0.4
0.2

0.6
0.4

0.8
0.6

0.8 1

1.2
1 0.5 0 0.5 1 0 20 40 60 80 100 120 140
x coordinate(m) time(s)

Fig. 1: Tracking a eight-shaped reference trajectory. Fig. 2: Error of tracking a eight-shaped trajectory.

decision vector for (29), and its upper/lower bounds d are 0.25
the current moment optimal control input increment

dened respectively as the current moment optimal control input increment

the next moment optimal control input increment

the vector of optimal control input increment

0.2 the next moment optimal control input increment
   
umax umin
d+ = , d
= R6Nu +6N 0.15

+h+ h
0.1

where for any i, the elements h+

 0 in h denotes +.
i
+
0.05

Thus, the convex set can be presented as = {d  d 

d+ }, is a strictly-positive design parameter used to scale 0

the convergence rate of the system, I is an unit matrix. The 0.05

coefcient matrix Z and vector a being
  0.1

M F T
0 20 40 60 80 100 120 140

R(6Nu +6N )(6Nu +6N )

time(s)
Z=
F 0
Fig. 3: the primal-dual neural network output.
 
p
a := R(6Nu +6N )
q
tracking trajectory of the mobile robot are depicted in Fig. 1,
and J () is the projection operator onto and dened as
and the tracking errors are shown in Fig. 2. It can be seen
J (d) = [J (d1 ), , J (d6Nu +6N )]T with
that the simulated mobile robot is able to track the desired
eight-shaped trajectory. The tracking errors asymptotically


di if di < di converge to zero. Fig. 3 shows the output PDNN u in each
J (di ) = di if di  di  d+

i

+ sample time.
di if di > d+ i
5.2 Experimental Results
for i {1, , 6Nu +6N }. By the dual dynamical system The proposed control algorithm is implemented on mo-
(31), a sequence of control actions can be obtained at each bile robot system implemented in the laboratory. The devel-
step, and then the state information can be updated by (13). oped mobile robot platform is illustrated in Fig. 4. The alu-
5 Simulation and Experimental Results minum chassis of the robot measures 45 cm45 cm85 cm
and contains two motors, two free wheel for balance, trans-
5.1 Simulation Results mission elements, 24-V battery, and SICK LMS111 Laser
In this section, we show the design process and effective- Finder and Kinect. The wheels have a radius 9.5 cm and are
ness of the predictive control developed in previous sections mounted on an axle of length 45 cm. The wheels are driven
using of a two-wheel mobile robot tracking desired trajecto- by motors having rated torque 72.1 mNm/A at 5200 rpm and
ries. at 24-V rated voltage. Each motor is equipped with incre-
Considering the mobile robot in Fig. 4, we choose a eight- mental encoder counting 2048 pulses/turn and a gear driven
shaped reference trajectory, which is dened as follows: assembly, which reduces the speed by a factor of 85.
The control structure is shown in Fig. 5, it has a two-level
xr = sin(t/10) control architecture. High-level control algorithms (includ-
yr = sin(t/20) ing reference motion generation) are written in VC++ and
run with a sampling time of 100 ms on a host computer (Intel
The parameters of model predictive control are chosen as 2-core processor). The PC communicates through CAN bus
N = 3, Nu = 2, Q = 0.1I, R = 0.1I. The optimization with the Elmo driver, which controls the servo motor through
problem (28)(30) can solved by using the primal-dual neu- computed torque control. A wheel encoder measures is re-
ral network (31) during each 100 ms sampling interval. The ceived for odometric computation. The lower level control

8356
 desired trajectory
actual trajectory
8.5

7.5

y coordinate(m)
7

6.5

5.5

5
2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5
x coordinate(m)

Fig. 6: Tracking a circle.

Fig. 4: The developed mobile service robot. 0.025

error of x(m)
error of y(m)
error of angle(rad)
0.02

0.015

error
0.01

0.005

0.005
0 20 40 60 80 100 120
time(s)

Fig. 7: Error of tracking circle.

Fig. 5: Control architecture of the mobile robot.

trajectory of the mobile robot are depicted in Fig. 9, the
layer is in charge of the execution of the high-level veloc- tracking errors are shown in Fig. 10, and Fig.11 shown the
ity commands. It consists of Elmo driver controller. The output PDNN u in each sample time.
Elmo driver controller performs three basis tasks: 1) to com- 6 Conclusion
municate with the higher-level controller through the Kvaser
CAN device; 2) reading encoder counts interrupt driven; and In this paper, a model predictive control (MPC) scheme
3) generating the computed input torques. is proposed for target tracking control of two-wheel mobile
In this experiments, we have realized two tracking trajec- robots. Based on the tracking-error kinematics, the proposed
tories. First, the chosen reference trajectory is a circle, which MPC approach iteratively solves a formulated quadratic pro-
can be described as: gramming (QP) problem using a LVI-based primal-dual neu-
ral network (LVI-PDNN) over a nite receding horizon. The
xr = 2 cos() applied neural networks are both stable in the sense of Lya-
yr = 2 sin() punov and globally convergent to the exact optimal solu-
tions of reformulated convex programming problems. The
The parameters of model predictive control are chosen as smoothness of the robot motion is improved, the reason-
N = 3, Nu = 2, Q = 0.1I, R = I, and sampling interval able magnitudes of the robot velocities and a better target
T = 100ms. The experimental results are shown in Fig. 6 tracking performance are achieved. Simulation and exper-
8. Fig. 6 shows the trajectory tracking by the mobile robot iment results are provided to demonstrate the effectiveness
with respect to the reference trajectory. The tracking errors and characteristics of the proposed LVI-PDNN based MPC
are shown in Fig. 7, and Fig. 8 shows the output PDNN u approaches to target tracking control.
in each sample time.
Second, the chosen reference trajectory is a lane-change References
curve, whose equation can be described as: [1] L. Huang, Velocity planning for a mobile robot to track a
2 moving target-a potential eld approach, Robotics and Au-
yr = tonomous Systems, 57: 5563, 2009.
1 + e2(xr 2)
[2] J. Yan, X.-P. Guan, F.-X. Tan, Target Tracking and Obstacle
The parameters are chosen the same as tracking circle. The Avoidance for Multi-agent Systems, International Journal
experimental results are shown in Fig. 911. The tracking of Automation and Computing, 7(4): 550556, 2010.

8357
 4
x 10
2 0.03
error of x(m)
error of y(m)
1 0.025
error of angle(rad)
the vector of optimal control input increment

0 0.02

1 0.015

2 0.01

error
3 0.005

4 0

5 0.005

6 the current moment optimal control input increment 0.01

the current moment optimal control input increment
7 the next moment optimal control input increment 0.015
the next moment optimal control input increment
8 0.02
0 20 40 60 80 100 120 0 5 10 15 20 25 30
time(s) time(s)

Fig. 8: Convergent behaviors of the primal-dual neural net- Fig. 10: Error of tracking lane-change curve.
work for tracking circle.
4 the current moment optimal control input increment
x 10
1 the current moment optimal control input increment
the next moment optimal control input increment
0 the next moment optimal control input increment

the vector of optimal control input increment

7.5
desired trajectory
actual trajectory 1

7 2

3
6.5
y coordinate(m)

6 5

5.5
7

8
5

9
0 5 10 15 20 25 30

4.5
time(s)
5 5.5 6 6.5 7 7.5 8 8.5
x coordinate(m)
Fig. 11: Convergent behaviors of the primal-dual neural net-
Fig. 9: Tracking a lane-change curve. work for tracking lane-change curve.

[11] L. Wang and F. Wan, Structured neural networks for con-

[3] Y. Zhu, T. Zhang, J. Song, X. Li, M. Nakamura, A new
strained model predictive control, Automatica, 37(8): 1235
method for mobile robots to avoid collision with moving ob-
1243, 2001.
stacle, Articial Life Robotics, 16: 507510, 2012.
[12] L. Cheng, Z. Hou, and M. Tan, Constrained multi-variable
[4] M. Defoort, A. Doniec, and N. Bouraqadi, Decentralized ro-
generalized predictive control using a dual neural network,
bust collision avoidance based on receding horizon planning
Neural Compution and application, 16: 505512, 2007.
and potential eld for multi-robots systems, Informatics in
[13] Z. Hou, L. Cheng, and M. Tan, Multicriteria optimization
Control Automation and Robotics, LNEE , J.A. Cetto et al.
for coordination of redundant robots using a dual neural net-
(Eds.), 85: 201215, 2011.
work, IEEE Trans. System, Man, and Cybernectics, Part B:
[5] R. W. Brockett, Asymptotic stability andfeedback stabiliza-
Cybernectics, 40(4): 10751087, 2010.
tion, In Differential geometry control theory, 1983: 181208.
[14] Y. Pan and J. Wang, Model predictive control of unknown
[6] S. S. Ge, Z. P. Wang and T. H. Lee, Adaptive stabilization
nonlinear dynamical systems based on recurrent neural net-
of uncertain nonholonomic systems by state and output feed-
works, IEEE Trans. Industrial Electronics, 59(8): 3089
back, Automatica, 39(8): 1451-1460, 2003.
3101, 2012.
[7] Z.-G. Hou, A. Zou, L. Cheng, M. Tan, Adaptive Control
[15] Y. Zhang, S. S. Ge, and T. H. Lee, A unied quadratic-
of an Electrically Driven Nonholonomic Mobile Robot via
programming-based dynamical system approach to joint
Backstepping and Fuzzy Approach, IEEE Transactions on
torque optimization of physically constrained redundant ma-
Control Systems Technology, 17(4): 803815, 2009.
nipulators, IEEE Trans. System, Man and CyberneticsPart
[8] B. S. Park, S. J. Yoo, J. B. Park, and Y. H. Choi, A Simple
B, 34(5): 21262132, 2004.
Adaptive Control Approach for Trajectory Tracking of Elec-
[16] Z. Li, S. S. Ge, and S. Liu, Contact-Force Distribution Op-
trically Driven Nonholonomic Mobile Robots, IEEE Trans-
timization and Control for Quadruped Robots Using Both
actions on Control System Technology, 18(5): 11991206,
Gradient and Adaptive Neural Networks, IEEE Trans. Neu-
2010.
ral Network and Learning System, available online, DOI:
[9] K. Shojaei, A. M. Shahri, Output feedback tracking control
10.1109/TNNLS.2013.2293500.
of uncertain non-holonomic wheeled mobile robots: a dy-
[17] D. Q. Mayne, J. B. Rawlings, C. V. Rao, P. O. M. Scokaert,
namic surface control approach, IET Control Theory & Ap-
Constrained model predictive control: Stability and optimal-
plications, 6(2): 216228, 2012.
ity, Automatica, 36: 789814, 2000.
[10] T. Li, S. Chang, W. Tong, Fuzzy target tracking control of
autonomous mobile by using infrared sensors, IEEE Trans-
actions on Fuzzy Systems, 12: 491501, 2004.

8358