International Journal of Control

ISSN: 0020-7179 (Print) 1366-5820 (Online) Journal homepage: http://www.tandfonline.com/loi/tcon20

Adaptive motion control of wheeled mobile robot

with unknown slippage

Haibo Gao, Xingguo Song, Liang Ding, Kerui Xia, Nan Li & Zongquan Deng

To cite this article: Haibo Gao, Xingguo Song, Liang Ding, Kerui Xia, Nan Li & Zongquan Deng
(2014) Adaptive motion control of wheeled mobile robot with unknown slippage, International
Journal of Control, 87:8, 1513-1522, DOI: 10.1080/00207179.2013.878038

To link to this article: https://doi.org/10.1080/00207179.2013.878038

Accepted author version posted online: 03

Jan 2014.
Published online: 26 Feb 2014.

Submit your article to this journal

Article views: 435

View related articles

View Crossmark data

Citing articles: 11 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tcon20
International Journal of Control, 2014
Vol. 87, No. 8, 1513–1522, http://dx.doi.org/10.1080/00207179.2013.878038

Adaptive motion control of wheeled mobile robot with unknown slippage

Haibo Gao, Xingguo Song∗ , Liang Ding, Kerui Xia, Nan Li and Zongquan Deng
State Key Laboratory of Robotics and System, Harbin Institute of Technology, Harbin 150001, China
(Received 3 March 2013; accepted 18 December 2013)

As a major representative nonholonomic system, wheeled mobile robot (WMR) is often used to travel across off-road
environments that could be unstructured environments. Slippage often occurs when WMR moves in slopes or uneven terrain,
and the slippage generates large accumulated position errors in the vehicle, compared with conventional wheeled mobile
robots. An estimation of the wheel slip ratio is essential to improve the accuracy of locomotion control. In this paper, we
propose an improved adaptive controller to allow WMR to track the desired trajectory under unknown longitudinal slip,
where the stabilisation of the closed-loop tracking system is guaranteed by the Lyapunov theory. All system states use
neural network online weight tuning algorithms, which ensure small tracking errors and no loss of stability in robot motion
with bounded input signals. We demonstrate superior tracking results using the proposed control method in various Matlab
simulations.
Keywords: nonholonomic systems; wheeled mobile robot (WMR); radial basis function (RBF); neural networks (NN); slip
ratio; Lyapunov theory

1. Introduction Fully autonomous WMR control systems need to cope

In the past decades, several researchers have investigated with dynamical robot uncertainties, unmodelled or unstruc-
the problem of controlling nonholonomic systems. As a tured disturbances, and nonlinear friction. But, it is difficult
major representative nonholonomic system, the wheeled to obtain an accurate mathematical model for applying com-
mobile robot (WMR) has attracted considerable attention. puted torque controllers or other model-based controllers in
Most control methods for WMR are based on this assump- practice. Following the neural network (NN) development,
tion that the wheels roll without slipping. the neural network-based control of mobile robots has been
Under this assumption, many modern control theories the subject of intense research in recent years (Fierro &
and algorithms have been proposed for the motion control Lewis, 1998; Jolly, Kumar, & Vijayakumar, 2009; Sun, Pei,
of WMR. A classical stable control scheme was proposed Pan, & Zhang, 2013; Wang, Ge, Lee, & Lai, 2006). These
(Kanayama, Kimura, Miyazaki, & Noguchi, 1990) to com- researches had produced new methods for solving the main
pute the vehicle control inputs for an autonomous mobile difficulties. A neural network-based model that combined
robot under the assumption of perfect velocity tracking. the backstepping technique with a torque controller was
This kind of control method solved the nonholonomic track- presented by Fierro and Lewis (1998), and this algorithm
ing problems based on over-simplified models in the initial was based on the application of multi-layered BP neural net-
periods of their development by neglecting the vehicle dy- works. The tracking control using an adaptive smart neural
namics and thus considering only the steering system, or network for WMR was investigated (Wang et al. 2006), and
considering a known mathematical model of dynamical sys- it produced fine motion control based on partially unknown
tems. Thereby, many robust nonlinear and adaptive control dynamics. Sun et al. (2013) proposed a robust adaptive
methods (Ge, Wang, Lee, & Zhou, 2001; Li & Xu, 2009; NN control for the nonholonomic mobile robot to track
Li, Li, & Kang, 2010) were designed to control the motion the desired environmental boundary. These control meth-
of mobile robots, these control performances are often de- ods stated above are designed under the constraint of pure
graded by modelling errors, information feedback errors, roll and no slip.
and external disturbances. However, there still exist some However, we know the applications of WMR usually
sorts of uncertainties in the dynamics of nonholonomic sys- require them to travel across off-road environments in task,
tems or environmental information, and their influences are such that some phenomena as slippage or sliding always
very important in practice. This means that we need a more exists between the wheels and ground, which cause rolling
intelligent controller to solve these problems. of wheels, are not perfect (Wong, 2001). When a WMR

∗
Corresponding author. Email: [email protected]


C 2014 Taylor & Francis
1514 H. Gao et al.

moves on uneven terrain or slopes, the slippage will gen-

erate large numbers of accumulate position errors in the
vehicle compared with conventional WMR (Ding et al.,
2011). Therefore, the tracking control of mobile robot will
be considerably influenced by the condition of terrain (Ray,
Brande, & Lever, 2009). In order to solve this problem,
we propose an approach based on the estimation of each
wheel’s quantitative slip, and incorporate the slip ratio pa-
rameters into the kinematic model of the WMR. Many re-
searches have addressed the slip phenomenon in the WMR
(Endo, Okada, Nagatani, & Yoshida, 2007; Moosavian &
Kalantari, 2008). The slip ratios of all wheels could be esti-
mated by an experimental study (Ding, Gao, Deng, & Liu,
2010; Iossaqui, Camino, & Zampieri, 2010). These meth-
ods proved useful for estimating the slip ratios of wheels. Figure 1. A nonholonomic wheeled mobile robot.
In this paper, we propose an effective adaptive mo-
tion tracking control method based on NN and slip-
compensation for WMR systems. The control objective is is gravitational torques (or forces), τ d ∈ R n×1 denotes
to track a specified motion trajectory under the slip oc- bounded unknown disturbances, including unstructured un-
currence. We calculate the wheel-slips, and develop the modelled dynamics, B ∈ R n×r is the input transformation
proposed control method using NN to model the system dy- matrix, τ d ∈ R r×1 is the input vector of forces and torques
namics and nonlinearities. The radial basis function (RBF) supplied by actuators, A ∈ R m×n is the matrix associated
of neural networks is well suited to uncertain or nonlinear with the constraints, and λ ∈ R m×1 is the vector of con-
functions because of its rapid online learning ability and straint forces.
nonlinear characteristics. We use the control approach to In this paper, we use four-wheeled mobile robot as an
overcome unknown system parameters and slippage in the example. The mobile robot shown in Figure 1 is a typical
WMR system, and to find a suitable control input that sta- example of a nonholonomic mechanical system, and skid-
bilises the closed-loop system. This controller guarantees steered autonomous WMR. The motion and orientation are
perfect velocity tracking and the posture error converges controlled by independent actuators, i.e., DC motors pro-
to minimum. The proposed control method facilitates pre- vide the necessary torque to the wheels. The position of the
cise motion tracking performance, which was demonstrated robot in an inertial Cartesian frame {O, X, Y} is completely
using the Matlab simulation. specified by the vector, where xc , yc are the coordinates of
The remainder of this paper is organised as follows. the centre C of the mass and {C, Xc , Yc } is the local coor-
The basics of nonholonomic system and slip ratios, and dinate with an origin of (xc , yc ) with respect to the inertial
RBF neural networks are introduced in Section 2. Section 3 basis. Due to the symmetrical property of the platform, we
discusses the adaptive control method combined with neural assume C is also the geometrical centre.
networks and slip-compensation, and Section 4 covers the The nonholonomic constraint states that the robot can
stability analysis. The Matlab simulation is presented in only move in a direction normal to the axis of the driving
Section 5. Finally, Section 6 concludes the paper. wheels. Previous studies always assumed that the system
is subject to a ‘pure rolling without slipping’ constraint
(Fierro & Lewis, 1998):
2. Problem formulation
2.1 A nonholonomic WMR model ẏc cos θ − ẋc sin θ = 0.
A mobile robot system in an n-dimensional configuration
space C with the generalised coordinates (q1 , . . . ,qn ) that We consider that all kinematics equality constraints are
is subject to m constraints can be described by Fierro and independent of time and they can be expressed as follows:
Lewis (1998):
A(q)q̇ = 0. (2)
M(q)q̈ + V (q, q̇)q̇ + F(q̇) + G(q) + τ d
= B(q)τ − AT (q)λ, (1) Let S(q) be a full rank matrix, formed by a set of smooth
and linear independent vector fields spanning the null space
where M ∈ R n×n is a symmetric, positive definite iner- of A(q):
tia matrix, V ∈ R n×n denotes the centripetal and coriolis
forces, F ∈ R n×1 denotes the surface friction, G ∈ R n×1 A(q)S(q) = 0. (3)
 International Journal of Control 1515

According to (2) and (3), it is possible to find an auxiliary Substituting (9) into (6), we can rewrite the kinematic
vector time function v(t) ∈ R 2 such that for all t equation by
⎡ ⎤ ⎡ ⎤
q̇ = S(q)v(t). (4) ẋc (1 − s1 )cθ (1 − s2 )cθ 
⎢ ⎥ 1⎢ ⎥ vR
q̇ = ⎣ ẏc ⎦ = ⎣ (1 − s1 )sθ (1 − s2 )sθ ⎦ ,
2 vL
It follows that S(q) can be given by: θ̇ 1 − s1 /L 1 − s2 /L
(11)
⎡ ⎤ where cθ = cos θ , sθ = sin θ ,
cos θ 0
⎢ ⎥ ⎡ ⎤
S(q) = ⎣ sin θ 0⎦. (5) (1 − s1 )cθ (1 − s2 )cθ
1⎣
0 1 S̄ = (1 − s1 )sθ (1 − s2 )sθ ⎦ . (12)
2 1 − s /L 1 − s2 /L
1

Under such a constraint condition, the vehicle is described

According to (11), we can obtain the slip ratios of the right
by the following kinematic model:
and left wheels as follows:
⎡ ⎤ ⎡ ⎤
ẋc cos θ 0  ẋc + θ̇ Lcθ ẋc − θ̇Lcθ
⎢ ⎥ ⎢ ⎥ v s1 = 1 − , s2 = 1 − . (13)
q̇ = ⎣ ẏc ⎦ = ⎣ sin θ 0⎦ , (6) vR cθ vL cθ
ω
θ 0 1
Slip ratios are determined by physical interactions between
the wheels and the ground; however, in general, the aspects
where v = [v ω]T , v and ω are the displacement/linear of wheel-terrain interaction that are needed for accurate
and the angular velocities of the mass centre of the vehicle models are neither well known nor easily measurable in
body, respectively. In addition, |v| ≤ vmax and |ω| ≤ ωmax , realistic situations. Therefore, we assume that the current
vmax and ωmax are the maximum linear and angular veloci- angular velocity of the vehicle body, ω, can be directly mea-
ties of the mobile robot. sured by a gyro-sensor. The driving velocities of wheels, vR
However, when tangential slippage occurs between the and vL can be detected by using encoders. Then, odometry
wheels and the ground we can rewrite v and ω to reflect the can be implemented with consideration of the slip, to detect
effects of the slippage: slip ratios of both wheels using model (13), quantitatively.

vR (1 − s1 ) + vL (1 − s2 )
v= , (7) 2.2 Structural properties of a mobile robot
2
vR (1 − s1 ) − vL (1 − s2 ) The dynamics of system (1) is now transformed into a
ω= , (8) more appropriate representation for control purposes. Using
2L
 ⎡ ⎤ Equation (12), the dynamics Equation (1) can be rewritten
v 1 − s1 1 − s2  v by (Fierro & Lewis, 1998):
1⎣
1 − s1 1 − s2 ⎦
R
= , (9)
ω 2 vL
L −L q̇ = S̄(q)υ(t)
M̄(q)υ̇ + V̄ (q, q̇)υ + F̄(υ) + Ḡ(υ) + τ̄ d = τ̄ ,
where vector υ = [vR vL ]T , vR and vL are the driving veloc- (14)
ities of the right and left wheels, respectively, which can be
measured by encoders. 2L is the right- and left-wheel tread,
where M̄ = S̄ M S̄, V̄ = S̄ (M S̄˙ + V S̄), F̄ = S̄ F, Ḡ =
T T T
and s1 and s2 denote the slip ratios defined as follows:
T T T
S̄ G, τ̄ d = S̄ τ d , τ̄ = S̄ Bτ and τ̄ d  ≤ bd with bd as an
upper bound of system disturbances.
vR − vR vL − vL
s1 = , s2 = , (10)
vR vL Property 1: M̄ is a symmetric positive definite, and the
˙ − 2 V̄ is skew-symmetric (Fierro & Lewis, 1998):
matrix M̄
where vR and vL are the factual/current velocity of the i.e.,
right and left wheels, respectively, while s1 and s2 are the
slip ratios of the right and left wheels, respectively. If no
slippage occurs, the current velocity is equal to the driving
˙ − 2 V̄ )x = 0, ∀x = 0.
x T ( M̄
velocity (vR = vR , vL = vL ), and slip ratios s1 = s2 = 0.
1516 H. Gao et al.

where θ ∈ R p is the input vector, W ∈ R L+1 contains the

ideal threshold w0 , and weights w1 · · · wL of the neural net-
work, φ(θ) ∈ R (L+1) is the activation vector comprising the
RBF, and ε(θ ) ∈ R is the neural network function approxi-
mation error, which is assumed to be uncorrelated with the
regressor φ(θ).
Using the neural network shown in Figure 2 with a
sufficiently large number (L) of RBF in the hidden layer
to approximate the smooth function g(θ ) described by (9),
there exists a positive number bf , such that |ε(θ )| ≤ bf , ∀θ .
Figure 2. Structure of RBF neural network Then, an approximation of g(θ ) can be given by

ĝ(θ ) = Ŵ T φ(θ), (18)

2.3 RBF neural networks
Many different models of neural networks have been estab- where Ŵ is an estimate of W.
lished (Er & Gao, 2003; Jung & Yum, 2011; Li & Yang, Here, the NN is trained to match specified exemplar
2012; Liaw, Farrell, & Schaal, 2009; Theodoridis, Boutalis, pairs (θ , g(θ )), with θ the ideal training system states. The
& Christodoulou, 2011) for specific objectives. Of these corresponding actual output is ĝ(θ ), while the target output
models, we find the RBF neural network is well suited to is g(θ ). For RBF neural networks, a common weight tuning
modelling uncertain or nonlinear functions. A typical RBF algorithm is the gradient algorithm based on the backprop-
neural network is shown in Figure 2. It is a two-layer net- agated error (Haykin, 2009).
work composed of a hidden layer and an output layer. Using
this simple structure, the RBF neural network facilitates a
more effective weight updating procedure compared with 3. Adaptive tracking controller design
other, more complex multilayer networks.
3.1 Motion tracking problem
A schematic for the RBF network is illustrated in
Figure 2. In this paper, the Gaussian type function is con- In this control system, we use two postures: a reference
sidered as posture qr = (xr , yr , θr )T that is a goal posture of the vehi-
cle, and a current posture q = (xc , yc , θ )T that is its ‘real’


1 posture at any given moment. The tracking position error
φi (θ) = exp − 2 θ − μi 2 , (15) vector is expressed on the basis of a frame linked to the
2σi
mobile robot (Kanayama et al., 1990) as:
where μi , σi are the centre and width of the ith neuron, i = T
1, . . . , L in (15). The activation of a neuron in the output em = xe ye θe = T e e, (19)
layer is determined by a linear combination of the fixed
nonlinear basis functions, i.e. where
⎡ ⎤ ⎡ ⎤

m cos θ sin θ 0 xr − xc
g(θ ) = wi φi (θ ), T e = ⎣ − sin θ cos θ 0 ⎦ , e = ⎣ yr − yc ⎦ .
i=1 0 0 1 θr − θ

where wi are the adjustable weights that link the output An auxiliary velocity control law input that achieves track-
nodes with the appropriate hidden neurons. These weights ing for (3) is given by (Kanayama et al., 1990):
in the output layer can be learnt using the least-squares
method.

vr cos θe + k1 xe
In this study, the RBF neural network is employed for vc = (20)
ωr + k2 vr ye + k3 vr sin θe
nonlinear function approximation. We assume a special case
of the linear regression model as follows:
where (k1 , k2 , k2 ) > 0 are the feedback gains of xe , ye , and
θ e , respectively.
g(θ) = W φ(θ ) + ε(θ ), (16) However, we need to compensate for the loss of velocity
caused by wheel slippage while maintaining the desired
W = [w0 , w1 , . . . , wL ] , velocity of the WMR when we design the controller.
Therefore, according to (7), (8) and (20), based on
φ(θ) = [1, φ1 (θ), . . . , φL (θ)]T , (17) the estimation of slip ratios, we can recover an auxiliary
 International Journal of Control 1517

velocity control law input υ c = [v1 , v2 ]T by where w ∈ R (L+1)×2 is the vector of the ideal threshold and
 v +ω L their weights. The bounds for w and ε(θ ) are expressed as

 c c
v 1−ŝ1
υc = 1 = vc −ωc L
, (21) w ≤ bw and |ε(θ )| ≤ bε ∀θ . (28)
v2
1−ŝ2

where v1 and v2 are the auxiliary velocity control law input Then, an estimate of g(θ ) can be given by
of the right and left wheels, respectively; ŝ1 and ŝ2 are the
estimated slip ratios of the right and left wheels, which can ĝ(θ) = ŵ T φ(θ ) (29)
be detected by equality (13). From literature (Ding et al.,
2010), we know slip ratio is defined by s ∈ [−1, 1], when where ŵ ∈ R (L+1)×2 is the vector of the estimated threshold
wheels are slipping on original location and not able to and weights.
move forward/backward, in this case, slip ratio s = 1 or −1. The control law is given by
Therefore, for tracking object, we should assume ŝ = ±1.
After slip-compensation, the velocity tracking errors of τ̄ = ĝ + K ec + lsgn(ec ), (30)
the wheels are defined as follows:
where ĝ is regarded as the output torque of NN controller;


e Kec is the torque controller, and, K = diag[k4 , k5 ] is a
ec = 1 = υ c − υ. (22) positive definite matrix; lsgn(ec ) is a robust term to suppress
e2
the effect of disturbance and approximate errors.
Differentiating (22), using (14), the mobile robot dynamics Let the NN weights be further adjusted to minimise the
may be written in terms of the tracking errors as velocity tracking error. The adaptive law of ŵ is designated
as:
M̄(q)ėc = − V̄ (q, q̇)ec − τ̄ + g(x) + τ̄ d , (23)
ŵ˙ i = φeTc − κ ec  ŵi , i = 1, 2, . . . , L, (31)
where the important nonlinear mobile robot function is de-
fined as where ŵ = [ŵ1 . . . ŵL ], L is the number of NN hidden
layer, and learning law κ is a positive constant. The simul-
g(x) = M̄ υ̇ c + V̄ (q, q̇)υ c + F̄(υ) + Ḡ(υ). (24) taneous updates of the adaptive law may be suitable for
non-stationary conditions or online settings.
Here, the vector x required to compute g(x) can be defined The control law (30) can be rewritten as:
as
 τ̄ = ŵ T φ(θ) + K ec + lsgn(ec ) (32)
T T
x≡ υ Tc υ̇ Tc υ . (25)
and the parameter l is defined as
Function g(x) contains all the mobile robot parameters, such
as mass, moments of inertia, and friction coefficients. How- 1
ever, nonlinear function g(x) is often imperfectly known in l ≥ bε + bd + κbw2 +
, (33)
4
applications and it is difficult to determine. In this paper,
RBFNN is used to approximate it. which is related to the bounds described by (28), the pa-
rameter κ in (31), and a strictly positive constant ε.

3.2 Adaptive NN control scheme

In this paper, the unknown system function g(x) is approx- 4. Stability analysis
imated using the RBF neural network described by (15). A In this section, we perform system stability analysis of
major advantage is that this can always be accomplished, the closed-loop behaviour in the proposed control method.
due to the RBF neural network approximation property. To Thus, we derive and analyse the closed-loop dynamics.
consider the function g(x) given by (24), the vector in (15) Substituting the control input (32) into the mobile robot
is defined as: dynamics system described by (23) yields:
T
θ ≡ υ Tc υ̇ Tc υ T . (26) M̄ ėc = − (K + V̄ )ec +g̃+τ̄ d − lsgn(ec ), (34)

The neural network (15) for (24) is redefined as g(θ ): R6 → where g̃ = g − ĝ is the function estimation error. This es-
R2 , which is rewritten as: timation error is expressed according to (27) and (29) as

g(θ ) = wT φ(θ) + ε(θ ), (27) g̃(θ) = w̃T φ(θ ) + ε(θ ), (35)

1518 H. Gao et al.

Figure 3. Structure of the proposed motion control method.

where w̃ is the vector of the threshold and weight estimation ≤ −K ec 2 − l ec  + ec  (ε + τ̄ d )
errors, defined as − κ ec  w̃2 + κ ec  w̃ bw
= −K ec 2 − l ec  + ec  (bε + bd )
w̃ = w − ŵ. − κ ec  (w̃2 − w̃ bw )
1 1
Therefore, (33) can be written as follows: − κbw2 ec  + κbw2 ec 
4 4
= −K ec 2 − l ec  + ec  (bε + bd )
M̄ėc = − (K + V̄ )ec +w̃ T φ(θ )
(36) 1 1
+ε(θ )+τ̄ d − lsgn(ec ). − κ ec  (w̃2 − w̃ bw + bw2 ) + κbw2 ec 
 4 4
It is necessary to develop the adaptive NN control (NNC)
= −K ec 2 − ec  l − bε − bd
given in Equation (32) by considering the system stability,
so that both the position and the velocity tracking errors  2 
converge to zero. Consider the following Lyapunov func- 1 1 2
+ κ w̃ − bw − bw
tional candidate for the proposed controller: 2 4

 
1
V = V1 (ec , t) + V2 (em , t) ≤ −K ec 2 − ec 
+κ w̃ − bw )2 , (38)
2
and ˙ − 2 V̄ )x = 0,
where inequality (33) and property 2, x T ( M̄
1 T 1 ∀x = 0 are used, and tr[] denotes the trace of matrix (Horn
V1 (ec , t) = ec M̄ ec + [(w − ŵ)T (w − ŵ)], (37) & Johnson, 1985).
2 2
With (38), V̇1 ≤ 0 is guaranteed to be negative. This
where matrix M̄ is a symmetric positive definite, so that shows V1 →0 and implies ec (t) → 0 and ėc (t) → 0 as
V1 ≥ 0. Differentiating Equation (37) with respect to time t → ∞. The tracking velocity convergence is guaranteed by
as the control law (32), which implies υ → υ c and υ̇ → υ̇ c
as t → ∞.
V̇1 (ec , t) We consider the other positive definite Lyapunov func-
1 ˙ tional candidate:
= eTc M̄ ėc + eTc M̄ ˙
ec +tr[(w − ŵ)T ŵ]
2
1 2
= eTc (− V̄ ec + w̃T φ − K ec − lsgn(ec ) + ε + τ̄ d ) V2 (em , t) = (x + ye2 ) + (1 − cos θe )/k2 . (39)
2 e
1 ˙  
+ eTc M̄ ec − tr w̃T φeTc − κ ec  ŵ
2
1 T ˙
= ec ( M̄ − 2 V̄ )ec + eTc w̃T φ − eTc K ec
2
− leTc sgn(ec ) + eTc (ε + τ̄ d )
− tr(w̃T φeTc − κ w̃T ec  ŵ)
= eTc w̃T φ − K ec 2 − l ec  + eTc (ε + τ̄ d )
− tr(eTc w̃T φ) + tr[κ w̃ ec  (w − w̃)] Figure 4. Reference circle trajectory.
 International Journal of Control 1519

Differentiation yields (Kanayama et al., 1990): 1991), both the system stability and tracking convergence
are guaranteed by the control law (32) driving the system
V̇2 (em , t) = −k1 xe2 − k3 sin2 θe /k2 ≤ 0. (40) (14), which closely tracks the desired motion trajectories.

This shows that the tracking position convergence is guar-

anteed by the control law (32). According to (38) and (40), 5. Simulation results
we can obtain V̇ = V̇1 + V̇2 ≤ 0, so that V→0 as t → ∞. The slip ratio of wheel is an important state variable. When
According to the standard Lyapunov theory (Slotine & Li, slippage occurs between the wheels and the ground, the

Figure 5. Artificial NNC without compensation (left). Adaptive NNC with compensation (right).
1520 H. Gao et al.

Figure 5. (Continued)

wheel velocities are influenced by the slippage, necessi- NN controller with slip-compensation under slippage. We
tating the analysis using an estimated slip ratio for move- adopt vehicle parameters (Figure 1) as follows: m = 5 kg,
ment control. According to the definition of the slip ratio I = 4 kgm2 , L = 0.2 m, r = 0.15 m, under the time varying
(Ding et al., 2010), we know that the range of slip ratios is external disturbance τ d = (sin t, cos t, 1)T N, vr = 1 m/s,
from −1 to 1, the change of slip ratios can resemble a sine ωr = 1/3rad/s. The objective is to track the trajectory such
curve, and the slip ratios will increase with increasing slope. that the errors in the position and velocity tend to zero. The
Therefore, we can assume as follows. controller gains were selected so that the closed-loop sys-
Let the desired trajectory be a space circular ring, xr = tem exhibits a critical damping behaviour: k1 = 2, k2 =9,
3 cos(t/3), yr = 3 sin(t/3), zr = −0.4 cos(2t/3), and t ∈ k3 = 6, k4 = 20, k5 = 10. In the neural network, we selected
[0, 30]s and where the initial position is Q = (2, −1, −0.4) a radial basis function with Nh = 6 hidden-layer neurons
(Figure 4). and κ = 0.35.
The slip ratio could be considered as the time variable
function ‘0.2 sin (2t/3)’, which increases or decreases with
the slope angle of the reference trajectory. 5.1 Artificial NN control under slippage
We now demonstrate the adaptive NN control shown The response of artificial NN control (NNC) is described
in Figure 3. Here, we compare its performance with slip- under the tangential slip of the wheels, as shown (a1-g1)
compensation and without slip-compensation when slip in Figure 5. If we do not compensate the slippage or even
exits between the wheels and the ground. Two control neglect the effect of slippage in velocity control, these re-
performances were implemented and tested using Mat- sults of the tracking trajectory present large position and
lab Simulink models: A. Artificial NN controller (Jolly, velocity error changes. Obviously, these results are not the
Kumar, & Vijayakumar, 2009) under slippage; B. Adaptive desired precision for tracking control.
 International Journal of Control 1521

Figure 6. Adaptive NNC with compensation. (i1) Real slip ratio of left wheel. (i2) Real slip ratio of right wheel.

5.2 Adaptive NN control with slip-compensation Funding

under slippage This work was supported by the National Natural Science Foun-
The response of this controller is shown (a2–h2) in dation of China [grant number 50975059/61005080]; China Post-
doctoral Science Foundation [grant number 20100480994]; Spe-
Figure 5. Lost velocity has been compensated due to wheel cial Foundation [grant number 201104431]; ‘111’ Project [grant
slippage. And it is obvious that both the posture and velocity number B07018].
tracking errors converge to the smaller values. Thus, when
slippage occurred between the wheels and the ground, the
control method with slip-compensation and adaptive NNC
References
laws performed the better tracking effect during the motion
Ding, L., Gao, H.B., Deng, Z.Q., Song, J., Liu, Y., Liu, G., &
control. Through this comparison, we can know that slip- Lagnemma, K. (2011). Experimental study and analysis on
compensation plays a significant role during the feedback driving wheels’ performance for planetary exploration rovers
of estimated slip ratios. Figure 6 (i1, i2) showed that the moving in deformable soil. Journal of Terramechanics, 48,
real slip ratios with slip-compensation is converging to 27–45.
the smaller scopes than the prior designed ones based on Ding, L., Gao, H.B., Deng, Z.Q., & Liu, Z. (2010). Slip-ratio-
coordinated control of planetary exploration robots traversing
the terrain and trajectory. over deformable rough terrain. IEEE International Confer-
ence on Intelligent Robots and Systems, 10, 4958–4963.
Endo, D., Okada, Y., Nagatani, K., & Yoshida, K. (2007).
Path following control for tracked vehicles based on slip-
6. Conclusion and future work compensating odometry. In Proc. IEEE/RSJ International
In this paper, an adaptive tracking control method has been Conference on Intelligent Robots and Systems (pp. 2871–
2876). San Diego, CA.
proposed and demonstrated via neural networks and slip- Er, M.J., & Gao, Y. (2003). Robust adaptive control of robot
compensation. This control scheme has been designed for manipulators using generalized fuzzy neural networks. IEEE
tracking the desired motion trajectory in a WMR system Transactions on Industrial Electronics, 50(3), 116–129.
under slip condition. In addition, we have developed a neu- Fierro, R., & Lewis, F.L. (1998). Control of a nonholonomic mo-
ral network learning procedure to enhance the performance bile robot using neural networks. IEEE Transactions on Neu-
ral Networks, 9, 589–600.
of the proposed control scheme, even when tracking on Haykin, S. (2009). Neural networks and learning machines
uneven terrain. The stabilisation of the inner closed-loop (3rd ed). Ontario: McMaster University Press.
system has been analysed. It has been shown that the con- Horn, R.A., & Johnson, C.R. (1985). Matrix analysis. New York:
vergence of the velocity and position tracking errors to Cambridge Univ. Press.
zero is guaranteed by the proposed adaptive NNC law (32) Ge, S.S., Wang, J., Lee, T.H., & Zhou, G.Y. (2001). Adaptive ro-
bust stabilization of dynamic nonholonomic chained systems.
based on slip-compensation. Even though there exist ex- Journal of Robotic Systems, 18, 119–133.
ternal disturbances or unknown system parameters, such Iossaqui, J.G., Camino, J.F., & Zampieri, D.E. (2010). Adaptive
as friction and slippage, which are very difficult to model tracking control of tracked mobile robots with unknown slip
using conventional techniques. Finally, it has demonstrated parameter. XVIII Congresso Brasileiro de Automática, 12 a
the results of precise tracking performance using a Matlab 16-setembro, Bonito-MS, 1846–1851.
Jolly, K.G., Kumar, R.S., & Vijayakumar, R. (2009). An artificial
simulation. neural network based dynamic controller for a robot in a
This paper considers the compensation of calculating multi-agent system. Neurocomputing, 73, 283–294.
slip ratios, but not the interaction forces between wheels Jung, J.R., & Yum, B.J. (2011). Artificial neural network based
and ground. The future work is hoped to apply neural com- approach for dynamic parameter design. Expert System with
puting methods to terramechanics theory for terrain param- Applications, 38, 504–510.
Kanayama, Y., Kimura, Y., Miyazaki, F., & Noguchi, T. (1990).
eter estimation. The experiments need to be performed to A stable tracking control method for an autonomous mobile
compare with the system simulations under different tra- robot. Proc. IEEE International Conference on Robotics and
jectories and terrain. Automation, 1, 384–389
1522 H. Gao et al.

Li, Z.J., Li, J.X., & Kang, Y. (2010). Adaptive robust coordinated Ray, L.R., Brande, D.C., & Lever, J.H. (2009). Estimation of net
control of multiple mobile manipulators interacting with rigid traction for differential-steered wheeled robots. Journal of
environments. Automatica, 46, 2028–2034. Terramechanics, 46, 75–87.
Li, Z.J., & Xu, C.Q. (2009). Adaptive fuzzy logic control of dy- Slotine, J-J.E., & Li, W. (1991). Applied nonlinear control. Engle-
namic balance and motion for wheeled inverted pendulums. wood Cliffs, NJ: Prentice-Hall.
Fuzzy Sets and Systems, 160, 1787–1803. Sun, T., Pei, H.L., Pan, Y.P., & Zhang, C.H. (2013). Ro-
Li, Z.J., & Yang, C.G. (2012). Neural-adaptive output feedback bust adaptive neural network control for environmen-
control of a class of transportation vehicles based on wheeled tal boundary tracking by mobile robots. International
inverted pendulum models. IEEE Transactions on Control Journal of Robust and Nonlinear Control, 23, 123–
Systems Technology, 20, 1583–1591. 136.
Liaw, H.C., Farrell, J.A., & Schaal, S. (2009). Robust neural Theodoridis, D., Boutalis, Y., & Christodoulou, M. (2011). A
network motion tracking control of piezoelectric actuation new adaptive Neuro-Fuzzy controller for trajectory tracking
systems for Micro/Nanomanipulation. IEEE Transactions on of robot manipulators. International Journal of Robotics and
Neural Networks, 20, 356–367. Automation, 26, 1–12.
Moosavian, S.A.A., & Kalantari, A. (2008). Experimental slip Wang, Z.P., Ge, S.S. Lee, T.H., & Lai, X.C. (2006). Adaptive smart
estimation for exact kinematics modeling and control of a neural network tracking control of wheeled mobile robots.
tracked mobile robot. Proc. IEEE/RSJ International Confer- ICARCV.
ence on Intelligent Robots and Systems, Nice, France, Sept Wong, J.Y. (2001). Theory of ground vehicles. Hoboken, NJ: John
22-26, 95–100. Wiley & Sons.