IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO.

6, NOVEMBER 2012 1583

Neural-Adaptive Output Feedback Control of a Class of Transportation Vehicles

Based on Wheeled Inverted Pendulum Models
Zhijun Li, Senior Member, IEEE, and Chenguang Yang, Member, IEEE

Abstract—The wheeled inverted pendulum (WIP) models have number of degrees of freedom to be controlled [9]. This makes it
been widely applied in the transportation vehicles formed by difficult to apply the conventional robotics control approach to
a mobile wheeled inverted pendulum system with an operator control such an Euler-Lagrange system. Although WIP models
(demonstrated in Fig. 1). In this paper, we focus on the study of
nonlinear control design for the WIP model-based vehicles, for are intrinsically nonlinear, it is often possible to obtain a feed-
which accurate dynamics could not be obtained beforehand due to back linearized model of the system. If the system is operating
the presence of uncertainties caused by the human operator as well around an operating point, and the signals involved are small, a
as the vehicle. We develop an output feedback adaptive neural net- linear model that approximates the nonlinear system in the re-
work (NN) control incorporating a linear dynamic compensator gion of operation can be obtained. Thus, controller design and
to achieve stable dynamic balance and tracking of the desired
given trajectories. Comparison simulation studies demonstrate
analysis techniques for linear systems were applied. In [1], mo-
guaranteed tracking performance and stable dynamics balance in tion control was proposed based on linear state-space model.
the presence of uncertainties and thus verify the efficiency of the In [6], dynamics was derived using a Newtonian approach and
developed nonlinear controller. the control was designed based on the dynamic equations lin-
Index Terms—Adaptive control, neural network (NN), output earized around an operating point. In [8], dynamic equations of
feedback NN control, wheeled inverted pendulum (WIP). the inverted pendulum were studied involving pitch and rota-
tion angles of the two wheels as the variables of interest, and
in [13] a linear controller was designed for stabilization con-
I. INTRODUCTION sidering robustness as a condition. In [14], a linear stabilizing

W HEELED inverted pendulum (WIP) models have at-

tracted a lot of research attention recently [6], [10], [11],
[15], [17], and have been widely applied in autonomous robotics
controller was derived by a planar model without considering
yaw. In [19], the exact dynamics of two-wheeled inverted pen-
dulum was investigated, and then a linear feedback control was
and intelligent vehicles. Control design for the WIP models fea- developed on the dynamic model. In [15], a two-level velocity
ture both theoretical and practical significance. Many practical controller via partial feedback linearization and a stabilizing po-
systems based on WIP models have been developed, such as sition controller were derived.
JOE [6], B2 [7], Segway [12], etc. Among these systems, the Compared with many other robot control schemes, the model-
Segway PT has been proved to be a popular personal trans- based computed torque control method is effective and has a
porter. Such systems based on WIP models are characterized very good performance [22]. The underlying idea of this con-
by the dynamic balance using their two wheels. This maneuver- trol scheme is to use the existing knowledge of the WIP model
ability allows them to easily navigate on outdoor environments. to linearize and decouple the robot dynamics so that the motion
In addition, the compact structure design allows drivers to ac- of each joint can be individually controlled using other well de-
cess most places that can only be accessed by walkers in the veloped linear control strategies. However, there are two diffi-
past. Moreover, people can conveniently drive such vehicles to culties involved with this control scheme. The first is that the
travel short distances in a small area. fast computation of the dynamic model is required. The second
From the theoretical point of view, the transportation vehicles difficulty is that accurate dynamic of the model with different
modelled with MWIP models are governed by under-actuated operation conditions is required as the computed-torque method
configuration, i.e., the number of control inputs are less than the is not robust in the presence of modeling uncertainties. But in
practice the physical parameters of human operator in the trans-
portation vehicle could not be obtained beforehand, it is not pos-
Manuscript received July 04, 2011; accepted August 18, 2011. Manuscript sible to obtain a perfect, or even reasonably accurate, dynamics
received in final form September 07, 2011. Date of publication October 13, model of the transportation vehicle. Furthermore, the parame-
2011; date of current version August 09, 2012. Recommended by Associate
Editor A. Alessandri. This work was supported by the Natural Science Founda- ters of the robot dynamics also may be subject to change when
tion of China under Grants 60804003, 61174045, 61111130208, and 60935001, the vehicle changes its tasks.
by the International Science and Technology Cooperation Program of China It is well known that NN systems have been credited in
under 0102011DFA10950, and by the European Commission funded Marie
Curie International Incoming Fellowship H2R Project (FP7-PEOPLE-2010-IIF
robotics controls and applications as powerful tools capable
275078). of approximating any continuous function of systems that are
Z. Li is with the Department of Automation, Shanghai Jiao Tong University, subjected to structured and unstructured uncertainties [2], [16],
Shanghai 200240, China, and also with the School of Automation Science and
Engineering, South China of Technology, Guangzhou 510640, China (e-mail:
[20]. The feasibility of applying neural networks to model un-
[email protected]). known functions in dynamic systems has been demonstrated in
C. Yang is with the School of Computing and Mathematics, Plymouth Uni- our previous studies [16], [18]. In practical control applications,
versity, Plymouth PL4 8AA, U.K. (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available online
it is desirable to have systematic methods to ensure stability,
at http://ieeexplore.ieee.org. robustness, and performance of the overall system. Neural net-
Digital Object Identifier 10.1109/TCST.2011.2168224 work implicit control approaches have been developed in [4]

1063-6536/$26.00 © 2011 IEEE

1584 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012

and [18] for controlling nonaffine systems, and they guarantee In Assumption 2.1, we can obtain an upper bound for the
the stability of the closed-loop system. However, the above neural network parametrization, from which there is only one
mentioned work concern little on the under-actuated systems positive parameter to be adapted.
such as wheeled inverted pendulum transportation systems From (1) and Assumption 2.1, it shows
with nonholonomic constraints, which is to be investigated in
the paper.
In this paper, we study the implicit control for dynamic bal-
ance and stable tracking of desired trajectories for a class of
(3)
transportation vehicles based on WIP model, in which both the
dynamics and the dimension of the regulated system can be un-
known. However, the relative degree of the regulated output is where , and
assumed to be known. System transformation is first carried .
out to make the vehicle model is feedback linearizable, then Remark 2.1: The 2-norm is used to obtain so that it is
output feedback adaptive neural network incorporating a linear differentiable, which is important for removing chattering phe-
dynamic compensator is introduced to approximate the inver- nomenon since it will be used in the following sliding mode
sion dynamics in order to achieve stable dynamic balance and control.
desired tracking performance. Remark 2.2: It should be noted that the derivative of the
NN basis function is bounded, thus the approximation by NN
II. PRELIMINARIES can naturally result in a semi-global Lipschitz condition. On the
other side, using inequality (3) would greatly reduce the com-
Neural networks (NNs) have been widely used in modelling plexity and computation burden because the parameter needs
and controlling of nonlinear systems because of their capabili- to be adapted becomes only a scalar rather than a vector with
ties of nonlinear function approximation, learning, and fault tol- many elements which can increase computation burden when
erance. As a linear-in-parameter neural network (LPNN), radial large neural nodes are chosen.
basis function (RBF) NN is often used in practical control engi-
neering due to its simple structure and universal approximation
properties. In this paper, the Gaussian RBF NNs will be em- III. SYSTEM STRUCTURE AND DYNAMICS
ployed to approximate a continuous function [2]
A. WIP Dynamics
, as , where is the input
vector, is the weight vector and the The transportation system studied in this paper, i.e., the WIP-
kernel vector is with active func- based vehicle, is illustrated in Figs. 1 and 2 which show the prin-
tion being chosen as the commonly used Gaussian func- ciple of the WIP-based vehicle system. The following variables
tions , [17], where are used to describe the system (refer to Figs. 1 and 2):
is the center of the receptive field and is denote the torques of the left and right wheels; denotes the tilt
the width of the Gaussian function. By choosing enough nodes, angle of the pendulum; denotes the direction angle of the mo-
neural network can approximate any continuous function over a bile platform; denotes the radius of the wheels; denotes the
compact region with arbitrary accuracy, for every distance between the two wheels; denotes the length of the
pendulum; denotes the mass of the pendulum; denotes
the mass of the chair and human; denotes the mass of each
(1) wheel; denotes the moment of inertia of mobile pendulum;
denotes the moment of inertia of each wheel; and denotes
The ideal weight vector is an “artificial” quantity re- gravity acceleration.
quired for analytical purposes. It is defined as As illustrated in Fig. 1, the forward/backward motion of the
vehicle can be controlled by pendulum tilt angle. Acceleration
can be produced by the positive (forward) tilt angle of the pen-
dulum such that the velocity of the WIP-car can be increased,
For an unknown continuous function vector , it while deceleration can be produced by the negative tile angle
can be approximated by NN in the following manner, for every of the pendulum such that the velocity can be reduced. We as-
: sume the title angle can be detected and obtained in the control
design.
(2) Consider the following dynamics description of the WIP
model based vehicle using Lagrangian formulation as follows:
where are the desired weights and
basis function; and is the collective NN reconstruction errors. (4)
Since the functions approximated are assumed to be contin-
uous, we have the following assumption. where is the vector of generalized coor-
Assumption 2.1: For every , there exist ideal con- dinates with representing the component of the angular ve-
stant weights vector such that , and locity of the platform perpendicular to the line of wheel cen-
with bounds . ters and representing the magnitude of the velocity of the
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012 1585

Remark 3.1: The friction force vector considered is Coulomb

and Viscous friction given by

(7)

where the function is defined element-wise and the un-

known diagonal matrices and are bounded and positive.

B. Problem Formulation
It is noted that the dynamics of the wheeled inverted pen-
dulum based vehicle (4) can be represented by the nonlinear
non-affine multiple-input multiple-output (MIMO) form as
follows:

(8)

where , and function

Fig. 1. Transportation system based on wheeled inverted pendulum.
is a partially unknown vector field with respect to input .
System (8) is a general description of the dynamics of the
nonlinear vehicle in nonaffine form. We see that affine nonlinear
systems and linear systems are special cases of (8), such that by
designing a controller for (8), we actually include more general
systems as well.
Define for 1, 2, where denotes the
Lie derivative of with respect to the vector field .
We see that the system is input-output linearizable with strong
relative degree such that there exist a function
, , independent of
and that the mapping has a Jacobian
matrix which is nonsingular for all . Thus, is a
diffeomorphism on . Therefore, then system (8) can be ex-
pressed in the normal form
Fig. 2. Mobile wheeled inverted pendulum in tilt.
(9)
(10)
mid-point of the wheel centers, in other words, the heading ve- (11)
locity of the platform, and as the tilt angle as shown in Fig. 1.
is the inertia matrix, is the vector where is for
of Coriolis and centrifugal forces, is the vector of .
gravitational forces, is a full rank input transforma- Then, we can choose an invertible as approxima-
tion matrix and is assumed to be known because it is a function tion of . The feedback linearization control input
of fixed geometry of the system; is the vector of control can be defined as an inverse of as follows [5]:
inputs, is the actuator friction force vector.
According to the structure of the dynamics of the wheeled in- (12)
verted pendulum based vehicles, using Lagrangian formulation
we have where is commonly referred to as a pseudo control.
The control objective is to design an adaptive controller for
system (10) such that the output follows the desired
(5) trajectory , i.e., , while all the signals
in the closed-loop systems are bounded.
Assumption 3.1: The reference trajectory and its
derivatives up to third order remain bounded.
Remark 3.2: For the transportation vehicles, the desired ref-
(6) erence trajectory at a final time , , is set by the operator
in advance. Then reference can be formulated as interpola-
where are unknown constants, tion functions, for examples, ,
are unknown function of states. where , , and
1586 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012

. It is obvious that the reference trajec- and the pseudo control is chosen
tory satisfies Assumption 3.1. to be
Assumption 3.2: The zero dynamics of system (10) is estab-
lished by and they are exponentially stable. In (28)
addition, there exist Lipschitz constants and for
such that , for every where is diagonal positive, is the output of a linear dy-
. namic compensator, and is the adaptive control signal de-
According to Assumption 3.2 and the converse Lyapunov the- signed to cancel . Substituting (28) into the output dynamics
orem, there exists a Lyapunov function which satisfies the (26), we have
following inequalities:
(29)

(13) The following linear dynamic compensator is introduced:

(14)
(30)
(15)
where . Define , and then we can obtain the
where are positive constants. augmented tracking error dynamics as follows:
Define the tracking error as and let us
introduce and as
(31)
(16)
(17) where is a identity matrix and , , , and in (31)
should be designed such that is Hurwitz.
Consider the following signals defined as: Differentiating (23) and considering (24), one has

(18) (32)
(19) which leads to
(20)
(33)
(21)
(22) Considering (33), we can rewrite (29) as

where , , is a diagonal positive ma- (34)

trix, and is generated from the following linear observer [23],
Then, the linear dynamic compensator (30) is chosen as
which is used for velocity estimation
(35)
(23) (36)
(24)
where . Define , and consider (33), we can
with a positive constant. rewrite (35) and (36) as
Then, the time derivative of the filtered tracking error can be
written as

(25)

where . (37)
Add and subtract on the right-hand side of equa-
tion (25), we obtain Theorem 3.1: [3] The LTI system is asymptoti-
cally stable if and only if, given any symmetric positive-definite
(26) matrix , there exists a symmetric positive-definite matrix ,
which is a unique solution of the so-called Lyapunov equation
where is the difference between the unknown function
and its approximation defined as (38)

Lemma 3.1: For , and , the following

(27) inequality holds .
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012 1587

Lemma 3.2: Consider a first-order dynamical system to specify the parameters for the observer and controller will
, where is a bounded and continuous non-neg- also be discussed.
ative function and is a positive parameter. If a non-negative Theorem 4.1: Consider the adaptation and control law (28),
initial value is chosen, then the solution is non-nega- (30) and (43) over the compact set in which the NN
tive for all . approximation is valid. For initial , , and starting
in any compact set , then all closed-loop signals are
IV. CONTROL DESIGN semi-globally uniformly bounded over the following compact
A. Output Feedback Control sets. In addition,
Remark 4.1: The matrices and are Hurwitz and satisfy (i) for every , the transient boundedness of the
closed-loop system is depicted as
(39) ,
(40) ,
,
where , , , ,
. ;
The smooth function vector is unknown and can be (ii) for , the steady-state stability of the closed-loop
approximated by NN over a compact set as follows: system is depicted as
,
(41) ,
where is the ideal NN weight matrix, ,
and is the approximating error. Both and satisfy As- ,
sumption 2.1. , where ,
According to (19) and (20) in [21] as well as the NN ,
parametrization, it can be shown from (41) that and ,

, .
Proof: Using (37), (39), (40), and (43), one could obtain
(42) the first derivative of defined in (45) has
in which ,
, with NN basis (46)
function .
Define as the estimation of the unknown constant scalar Considering , from (19) and
, then we employ the following adaptive control law for : (21), noting the definition of in (44), we can rewrite (46) as

(47)
(43)
For the third item , it shows
where is diagonal definite, the matrix is defined

(44)

with positive definite and , , and are positive design

parameters and satisfying ,
with finite constant .

B. Stability Analysis
Define the estimation error of as , and a
Lyapunov candidate as (48)

where is denoted as .
In (48), is positive with positive initial value according to
(45)
Lemma 3.2 and (42) and (43). With Lemma 3.1, it has
where is diagonal positive, then we are ready to present
the main result of this paper in the following theorem, for
which the stability analysis for the closed-loop system under
(49)
the proposed control will be provided, and the analysis on how
1588 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012

In view of (49), (48) becomes Multiplying both sides of (56) by and integrating over
, it can be shown that
(50)
(57)
For the fourth item of (47), we have
which can be written into
, for every ,
, . Similarly, we can obtain
(51)
the rest result as follows: ,
for every , ,
For the fifth item of (47), we have
, , for
every , ,
, for every ,
, ,
for every , , . This completes the
proof.
(52)
V. SIMULATION
For the sixth item of (47), we have
To demonstrate the effectiveness of the proposed con-
trol algorithm, let us consider the mobile wheeled in-
verted pendulum-based autonomous vehicle as shown
(53) in Fig. 1. The vehicle is subjected to the following con-
straints: . Using Lagrangian approach,
With (43), (50), (51), (52), and (53), we can rewrite (47) as we can obtain the reduced dynamics for ,
, and as [11]

(54) where , ,
and
By completing the squares .
In the simulation, we choose the parameters from [25] as
(55) 0.5 kgm , 4 kg, 9.0407 kgm , 87 kg,
90 kg, 1.0 m, 0.5 m, 0.25 m. The initial values
Let , then are set as , . The
, , then (47) disturbances from environments on the system are introduced as
becomes , in the simulation model. We assume that
The operator choose the desired angle velocity ras s
and tilt angle . From above, we can obtain the desired
trajectories as rad, rad. It is obvious that
the produced trajectories satisfy Assumption 3.1. We assume the
initial velocity is 0.1 m/s.
The system state is observed through the noisy linear mea-
surement channel, zero-mean Gaussian noises are added to the
state information. All noises are assumed to be mutually inde-
pendent. The noises have variances corresponding to a 5% noise
to signal ratio. We designed the following dynamic compensator
, with , and
, with . Ac-
cording to (23), the observer gain is selected as ,
(56) , .
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012 1589

Fig. 4. Tracking the desired tilt angle by neural-adaptive network.

Fig. 3. Tracking the direction angle by the neural-adaptive network.

Because there are two input torques, i.e., , the neural

network for controller consists of two sub-NNs, with common
basis function of input . The centers of radial
basis function evenly spaced in
. The design parameters of
the above controllers are: , and ,
, , ,
, .
The friction vector considered is Coulomb and Viscous fric-
tion, given by (7). However, since the discontinuity of the fric-
tion characteristics at zero velocity is required for testing zero
velocity, and when the velocity is zero, or the system is sta- Fig. 5. Stable velocity by the neural-adaptive network control.
tionary, the friction is indefinite and depends on the controlled
torque. In the simulation, to improve the numerical efficiency,
we adopt a revised friction model from [24]. The structure for
the revised friction model described by the following mathemat-
ical model:
, where is the revised friction and is a small
positive scalar, is given by
if
if
if
where is the motor angular velocity, in the simula-
tion, we choose 12 Nm, 1.2 Nm/rad/s, and
0.0088 Nm/rad/s. When the velocity is near zero, defined
by , the friction is equal to the applied torque .
When the velocity is greater than this, the third term in the Fig. 6. Input torques by neural-adaptive network control.

above expression vanishes and the friction given by this revised

model is equal to the friction given by (7).
The direction angles tracking by three control approaches are
shown in Fig. 3, and the input torques are shown respectively
in Fig. 6, the tilt angles for the dynamic balance and the stable
velocities under three control approaches are shown in Figs. 4
and 5, respectively. From these figures, even if without the prior
knowledge of the system, we can obtain good performance by
the proposed neural network control.
From these figures, even if the dynamics parameters of
the system is unknown, we can get good performance by the
proposed control. The performance of model-based approaches
is sensitive to the accuracy of the model representation. Neural
network control approaches are tolerant of modelling errors,
and can be viewed as a key advantage over model-based Fig. 7. Tracking the desired direction angle by the model-based control.
1590 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012

these figures, the simulation results show the proposed control

is with the better performance and is more realistic in practice.

VI. CONCLUSION

In this paper, we have developed an neural-adaptive output

control of a large class of transportation vehicles that consist
of a mobile WIP system and an operator, in the presence of
unmodelled dynamics, or parametric/functional uncertainties.
It is rigorously established that the controller is able to guar-
antee semi-globally uniformly bounded stability. Furthermore,
the steady state compact sets into which the closed loop error
signals converge are also given. Simulation results demonstrate
Fig. 8. Tracking the desired tilt angle by the model-based control.
satisfactorily tracking performance as well as boundedness of
all closed-loop signals.

ACKNOWLEDGMENT

The authors would like to thank N. Ding for help correcting

the typos and errors.

REFERENCES
[1] Y. S. Ha and S. Yuta, “Trajectory tracking control for navigation of the
inverse pendulum type self-contained mobile robot,” Robot. Autonom.
Syst., vol. 17, pp. 65–80, 1996.
[2] F. L. Lewis, A. Yesildirek, and K. Liu, “Multilayer neural network
robot controller with guaranteed tracking performance,” IEEE Trans.
Neural Netw., vol. 7, no. 2, pp. 388–399, Mar. 1996.
[3] S. S. Ge, T. H. Lee, and C. J. Harris, Adaptive Neural Network Control
Fig. 9. Stable velocity by the model based control.
of Robot Manipulators. London, U.K.: World Scientific, 1998.
[4] S. S. Ge, C. C. Hang, and T. Zhang, “Adaptive neural network control
of nonlinear systems by state and output feedback,” IEEE Trans. Syst.,
Man, Cybern., B, Cybern., vol. 29, no. 6, pp. 818–828, Dec. 1999.
control of wheeled inverted pendulums, for which accurate [5] N. Hovakimyan, F. Nardi, and A. J. Calise, “A novel error observer-
modelling of wheeled inverted pendulums dynamics is difficult, based adaptive output feedback approach for control of uncertain sys-
time-consuming and costly. Therefore, we will compare the tems,” IEEE Trans. Autom. Control, vol. 47, no. 8, pp. 1310–1314,
Aug. 2002.
performance of the wheeled inverted pendulums, in terms of [6] F. Grasser, A. Arrigo, S. Colombi, and A. C. Rufer, “JOE: A mobile,
tracking error, between the neural network control approaches inverted pendulum,” IEEE Trans. Ind. Electron., vol. 49, no. 1, pp.
and a model-based approach. In particular, we consider 107–114, Jan./Fe 2002.
[7] H. Tirmant, M. Baloh, L. Vermeiren, T. M. Guerra, and M. Parent, “B2,
model-based controller as an alternative two wheeled vehicle for an automated urban transporta-
and tion system,” in Proc. IEEE Intell. Veh. Symp., 2002, p. 594C603.
[8] A. Salerno and J. Angeles, “On the nonlinear controllability of a quasi
holonomic mobile robot,” in Proc. IEEE Int. Conf. Robot. Autom.,
with . 2003, pp. 3379–3384.
Considering the unmodelled dynamics, we introduce the effects [9] A. Isidori, L. Marconi, and A. Serrani, Robust Autonomous Guidance:
of 10% parametric uncertainties in the dynamic model. In An Internal Model Approach. New York: Springer, 2003.
[10] R. Brooks, L. Aryanada, A. Edsinger, P. Fitzpatrick, C. C. Kemp, U.
the simulation, we found that since the is very sensitive to O’Reilly, E. Torres-jara, P. Varshavskaya, and J. Weber, “Sensing and
the model, more than 10% model uncertainty would cause the manipulating built-for-human environments,” Int. J. Humanoid Robot.,
system unstable under the model-based controller. The presence vol. 1, no. 1, pp. 1–28, 2004.
[11] Z. Li and Y. Zhang, “Robust adaptive motion/force control for wheeled
of parametric errors is a common problem for model-based inverted pendulums,” Automatica, vol. 46, pp. 1346–1353, 2010.
controllers since the identification of dynamic parameters is [12] H. G. Nguyen, J. Morrell, K. Mullens, A. Burmeister, S. Miles, N. Far-
error-prone. For instance, the controlled conditions in the test rington, K. Thomas, and D. W. Gagee, “Segway robotic mobility plat-
form,” in Mobile Robots XVII, D. W. Gage, Ed. New York: Springer,
facility under which the parameters are identified are often very 2004, vol. 5609, Proceedings of the SPIE, pp. 207–220.
different from actual conditions, thus rendering the parameters [13] A. Salerno and J. Angeles, “The control of semi-autonomous
inaccurate for real operating conditions. Neural network control two-wheeled robots undergoing large payload-variations,” in Proc.
IEEE Int. Conf. Robot. Autom., 2004, pp. 1740–1745.
presented in this paper are not susceptible to this problem, since [14] A. Blankespoor and R. Roemer, “Experimental verification of the dy-
the unknown parameters are learn during the wheeled inverted namic model for a quarter size self-balancing wheelchair,” in Proc.
pendulum operation in actual conditions. The position tracking Amer. Control Conf., 2004, pp. 488–492.
[15] K. Pathak, J. Franch, and S. K. Agrawal, “Velocity and position con-
using the model-based control are shown in Figs. 7 and 8. The trol of a wheeled inverted pendulum by partial feedback linearization,”
final velocity is shown in Fig. 9, which fluctuates greatly. From IEEE Trans. Robot., vol. 21, no. 3, pp. 505–513, Jun. 2005.
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 6, NOVEMBER 2012 1591

[16] Z. Li, C. Yang, and J. Gu, “Neuro-adaptive compliant force/ motion [22] J. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs,
control for uncertain constrained wheeled mobile manipulator,” Int. J. NJ: Prentice-Hall, 1991.
Robot. Autom., vol. 22, no. 3, pp. 206–214, 2007. [23] Z. Li, S. S. Ge, M. Adams, and W. S. Wijesoma, “Adaptive robust
[17] D. S. Nasrallah, J. Angeles, and H. Michalska, “Modeling of an anti- output-feedback motion/force control of electrically driven nonholo-
tilting outdoor mobile robot,” in Proc. 5th Int. Conf. Multibody Syst., nomic mobile manipulators,” IEEE Trans. Control Syst. Technol., vol.
Nonlinear Dyn., Control, 2006, pp. 2275–2284. 16, no. 6, pp. 1308–1315, Nov. 2008.
[18] C. Yang, S. S. Ge, C. Xiang, T. Chai, and T. H. Lee, “Output feedback [24] D. Karnopp, “Computer simulation of strick-stip friction in mechanical
NN control for two classes of discrete-time systems with unknown con- dynamic systems,” ASME J. Dyn. Syst., Meas., Control, vol. 107, pp.
trol directions in a unified approach,” IEEE Trans. Neural Netw., vol. 100–103, 1985.
19, no. 11, pp. 873–1886, Nov. 2008. [25] A. R. S. McGuire, K. F. Jensen, B. S. M. Henriques, H. C. N. Nguyen,
[19] Y. Kim, S. H. Kim, and Y. K. Kwak, “Dynamic analysis of a nonholo- K. F. Jensen, K. Vinther, and R. Jespersen, “Trajectory planning and
nomic two-wheeled inverted pendulum robot,” J. Intell. Robot. Syst., control for a segway RMP,” Aalborg Univ.—Dept. Electron. Syst.,
vol. 44, pp. 25–46, 2005. 2009.
[20] G. Loreto and R. Garrido, “Stable neurovisual servoing for robot ma-
nipulators,” IEEE Trans. Neural Netw., vol. 17, no. 4, pp. 953–965, Jul.
2006.
[21] H. Du and X. Chen, “NN-based output feedback adaptive variable
structure control for a class of non-affine nonlinear systems: A non-
separation principle design,” Neurocomput., vol. 72, pp. 2009–2016,
2009.