Sei sulla pagina 1di 8

Automatica 46 (2010) 13461353

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Brief paper

Robust adaptive motion/force control for wheeled inverted pendulumsI


Zhijun Li a, , Yunong Zhang b
a

Department of Automation, Shanghai Jiao Tong University, Shanghai, 200240, China

School of Information Science & Technology, Sun Yat-Sen University, Guangzhou, 510275, China

article

info

Article history:
Received 14 June 2009
Received in revised form
9 February 2010
Accepted 21 April 2010
Available online 7 June 2010
Keywords:
Wheeled inverted pendulums
Nonholonomic constraints
Unmodelled dynamics
Zero dynamics

abstract
Previous works for wheeled inverted pendulums usually eliminate nonholonomic constraint force in
order to make the control design easier, under the assumption that the friction force from the ground is
as large as needed. Nevertheless, such an assumption is unfeasible in practical applications. In this paper,
adaptive robust motion/force control for wheeled inverted pendulums is investigated with parametric
and functional uncertainties. The proposed robust adaptive controls based on physical properties of
wheeled inverted pendulums make use of online adaptation mechanism to cancel the unmodelled
dynamics. Based on Lyapunov synthesis, the proposed controls ensure that the system outputs track the
given bounded reference signals within a small neighborhood of zero, and guarantee the semi-global
uniform boundedness of all closed loop signals. The effectiveness of the proposed controls is verified
through extensive simulations.
2010 Elsevier Ltd. All rights reserved.

1. Introduction
Wheeled inverted pendulums have attracted a lot of attention
recently (Brooks et al., 2004; Gans & Hutchinson, 2006; Grasser,
Arrigo, Colombi, & Rufer, 2002; Jung & Kim, 2008; Li & Luo,
2009; Nasrallah, Michalska, & Angeles, 2007; Pathak, Franch, &
Agrawal, 2005) as shown in Fig. 1. Similar systems like the cart
and pendulums have been studied in the literature Ibanez, Frias,
and Castanon (2005) and Zhang and Tarn (2002). The differences
from these systems are that the inverted pendulums motion in the
present system is not planar and the motors driving the wheels are
directly mounted on the pendulum body (Pathak et al., 2005).
Motion of wheeled inverted pendulums is governed by underactuated configuration, i.e., the number of control inputs is less
than the number of degrees of freedom to be stabilized (Isidori,
Marconi, & Serrani, 2003), which makes it difficult to apply
the conventional robotics approach to control EulerLagrange
systems. Due to these reasons, increasing effort has been made
towards control design that guarantees stability and robustness for
mobile wheeled inverted pendulums.

I This work is supported by Shanghai Pujiang Program under Grant


No. 08PJ1407000 and the Natural Science Foundation of China under Grant
Nos. 60804003, 60935001 and the New Faculty Foundation under Grant
No. 200802481003. The material in this paper was not presented at any conference. This paper was recommended for publication in revised form by Associate
Editor Shuzhi Sam Ge under the direction of Editor Miroslav Krstic.
Corresponding author. Tel.: +86 21 34204616; fax: +86 21 34204616.
E-mail addresses: zjli@ieee.org, lzjuec@hotmail.com (Z. Li).

0005-1098/$ see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2010.05.015

Although wheeled inverted pendulums systems are intrinsically nonlinear and their dynamics are described by nonlinear differential equations, if the system operates around an operating
point, and the signals involved are small, we can obtain a linear
model approximating the nonlinear system in the region of operation. In Ha and Yuta (1996), motion control was proposed using a linear state-space model. In Grasser et al. (2002), dynamics
was derived using a Newtonian approach and the control was designed based on the dynamic equations linearized around an operating point. In Salerno and Angeles (2003), dynamic equations of
the inverted pendulum were studied involving pitch and rotation
angles of the two wheels as the variables of interest, and in Salerno
and Angeles (2004) a linear controller was designed for stabilization considering robustness as a condition. In Blankespoor and Roemer (2004), a linear stabilizing controller was derived by a planar
model without considering yaw. In Kim, Kim, and Kwak (2005), the
exact dynamics of a two-wheeled inverted pendulum was investigated, and linear feedback control was developed on the dynamic
model. In Pathak et al. (2005), a two-level velocity controller via
partial feedback linearization and a stabilizing position controller
were derived.
Based on the idea of linearization, a model-based approach
is generally utilized in dynamic control. If accurate knowledge
of the dynamic model is available, the model-based control
can provide an effective solution to the problem. However,
wheeled inverted pendulum control is characterized by unstable
balance and unmodelled dynamics, and subject to time varying
external disturbances, in the form of parametric and functional
uncertainties, which are generally difficult to model accurately.
Therefore, traditional model-based control may not be the ideal

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

approach since it generally works best only when the dynamic


model is known exactly. The presence of uncertainties and
disturbances would disrupt the function of the traditional modelbased feedback control and lead to unstable balance.
Moreover, the wheeled inverted pendulum is definitely different from other nonholonomic systems subject to (i) only kinematic
constraints which geometrically restrict the direction of mobility,
i.e., wheeled mobile robot (Ge, Wang, Lee, & Zhou, 2001; Ge, Wang,
& Lee, 2003); (ii) only dynamic constraints due to dynamic balance
at passive degrees of freedom where no force or torque is applied,
i.e., the manipulator with passive link (Arai & Tanie, 1998; Luca &
Oriolo, 2002); (iii) both kinematic constraints and dynamic constraints. It is obvious that the wheeled inverted pendulum is more
complex than the former two cases, therefore, the controls suitable
for (i) and (ii) cannot be directly applied for (iii).
A challenging problem is to control a mobile inverted pendulum system whose cart is not constrained by guide rail like
cartpendulum systems, but moves in its terrain while balancing
the pendulum. Therefore, the nonholonomic constraint force between the wheels and the ground should be considered in order to
avoid slipping or slippage. A similar application of interacting with
environments can be found in Dong (2002), where motion/force
control is considered for mobile manipulators under holonomic
constraints. Recent works, including Grasser et al. (2002), Jung and
Kim (2008), Li and Luo (2009), Nasrallah et al. (2007) and Pathak
et al. (2005), where the wheeled inverted pendulum moves on the
planar plane or on the incline plane, do not consider nonholonomic
constraint force, i. e. friction force, while assuming beforehand that
the ground can provide enough friction as needed, but in practical applications, this assumption is difficult to satisfy. When the
ground friction cannot support motion, the control performance
by these controllers will be degraded.
In this paper, by discovering and utilizing the unique physical
properties of wheeled inverted pendulums, we separate the
zero-dynamics subsystem to simplify the model. Then, we
propose a robust adaptive motion/force control for wheeled
inverted pendulums. Since the system except the zero-dynamics
subsystem is still a MIMO nonlinear system, we propose adaptive
robust controls to accommodate the presence of parametric and
functional uncertainties in the dynamics of wheeled inverted
pendulums.
The main contributions of this paper are that: (i) adaptive
robust motion/force control is developed for wheeled inverted
pendulums by using their physical properties with parametric
and functional uncertainties; (ii) the nonholonomic constraint
force between the wheels and the ground is considered in
order to avoid slipping or slippage; (iii) based on Lyapunov
synthesis, motion/force stabilities are achieved and the input-tostate stability properties of the zero dynamics are used to derive
bounds on the the tracking errors.
2. Preliminaries
In the following study,
q let k k denote the 2-norm, i.e. given

A = [aij ] Rmn , kAk =

Pm Pn
i=1

j =1

|aij |2 .

Lemma 2.1. Let e = H (s)r with H (s) representing a (n m)dimensional strictly proper exponentially stable transfer function,
T mr
and e denoting its input
and output, respectively. Then r Lm
L
2
T
implies that e, e Ln2
Ln , e is continuous, and e 0 as t .
If, in addition, r 0 as t , then e 0 (Ge, Lee, & Harris,
1998).
Lemma 2.2. For x 0 and = 1 + (1+1t )2 1 with t > 0, we have
ln(cosh(x)) + x.

1347

Fig. 1. Mobile wheeled inverted pendulum.

Rx

Rx
< 0 e22 d = 1 e2x <
Rx
1 + (1+1t )2 . Therefore, ln(cosh(x)) + ln(cosh(x)) + 0 e22+1 d .
Rx
Let f (x) = ln(cosh(x)) + 0 e22+1 d x, we have f (x) = tanh(x) +
Proof. If x 0, we have

2
d
0 e2 + 1

e
1 = eex
+ e2x2+1 1 = 0. From the Mean Value Theorem,
+ex
we have f (x) f (0) = f (x)(x
R x0). Since f (0) = 0, we have
f (x) = 0, that is, ln(cosh(x)) + 0 e22+1 d = x, then, we have
ln(cosh(x)) + x. This completes the proof. 
2
e2x +1

3. System description
3.1. Dynamics of mobile wheeled inverted pendulums
Consider the following wheeled inverted pendulum dynamics
described by Lagrangian formulation:
M (q)q + V (q, q )q + G(q) + F = B + f

(1)

where q = [x, y, , ]T Rn with n = 4 is the vector of


generalized coordinates with x, y as the position coordinates, as
the heading angle, and as the tilt angle as shown in Fig. 1. M (q)
Rnn is the inertia matrix, V (q, q )q Rn is the vector of Coriolis
and centrifugal forces, G(q) Rn is the vector of gravitational
forces, F Rn is the vector of the bounded external disturbances
from the environment, B Rnnl is a full rank input transformation
matrix and is assumed to be known because it is a function of fixed
geometry of the system; Rnl is the vector of control inputs,
f = J T Rn denotes the vector of constraint forces, where
J T Rn(nm) is Jacobian matrix, and Rnm are Lagrangian
multipliers corresponding to n m nonholonomic constraints,
respectively. If q is partitioned into qv = [x, y, ]T and , we obtain

Mv
M (q) =
Mv
G(q) =

f = fv

Gv
,
G
f

T

Mv
,
M

 
F =

Vv
V (q, q ) =
Vv
Fv
,
F


= v

B(q) =
0

Vv
,
V

Bv
0

0
,
B

T

where Mv and M describe the inertia matrices for the mobile


platform and the inverted pendulum, respectively, Mv and Mv
are the coupling inertia matrices of the mobile platform and
the inverted pendulum, and Vv and V denote the Centripetal
and Coriolis torques for the mobile platform and the inverted
pendulum, respectively. Vv and Vv are the coupling Centripetal
and Coriolis torques of the mobile platform and the inverted
pendulum. Gv and G are the gravitational torque vectors for
the mobile platform and the inverted pendulum, respectively. v
is the control input vector for the mobile platform, and fv and
f denote the external forces on the mobile platform and the
inverted pendulum, respectively, and Fv and F denote the external
disturbances on the mobile platform and the inverted pendulum,
respectively.
Remark 3.1. Since in the actual implementations, friction between the wheels of the mobile platform and the ground cannot
be predicted beforehand, in order to avoid slipping of the wheels,
we have to guarantee the boundedness of the nonholonomic constraint force and make it arbitrarily small within a compact set.

1348

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

3.2. Reduced dynamics and state transformation

3.3. Physical properties

When the system is subjected to nonholonomic constraints, the

(n m) nonintegrable and independent velocity constraints can be


expressed as
Jv (q)qv = 0.

(2)

The constraint (2) is referred to as the classical nonholonomic


constraint when it is not integrable. In the paper, constraint (2) is
assumed to be completely nonholonomic and exactly known.
Since Jv (q) R(nv m)n , introduce J Rn n , and J =
[Jv , J ]T R(nm)n with nv + n = n, such that it is possible
to find a (m + n ) rank matrix R(q) Rn(m+n ) formed by a set
of smooth and linearly independent vector fields spanning the null
space of J (q), i.e.,
RT (q)J T (q) = 0.

(3)

Denote R(q) = [r1 (q), . . . , rm (q), rm+1 (q), . . . , rm+n ] and define
an auxiliary time function z (t )
Rm+n , and z (t ) =
T
[z1 (t ), . . . , zm (t ), zm+1 , . . . , zm+n ] such that
q = R(q)z (t ) = r1 (q)z1 (t ) + + rm (q)zm (t )

+ rm+1 (q)zm+1 (t ) + + rm+n (q)zm+n (t ).

(4)

Eq. (4) is the kinematic model for the wheeled inverted pendulums.
Usually, z (t ) has physical meaning, consisting of the angular
velocity , the linear velocity v , and the tilt angle , i.e., z (t ) =
[v ]
T . Eq. (4) describes the kinematic relationship between the
motion vector q and the velocity vector z (t ).
Differentiating (4) yields
q = R (q)z + R(q)z .

z = [R (q)R(q)]

"
M1 =

m12
m22
m32

g1
g2 ,
g3

m13
m23 ,
m33
d1
d2 ,
d3

" #

G1 =

F1 =

" #
v1
V1 z = v2
v3

z1
z2 ,
z3

" #

z=

" #
1
R a = 2
T

(13)

where m11 , . . . , m33 , v1 , v2 , v3 , g1 , g2 , g3 , d1 , d2 , d3 , are unknown


functions.
From (1) and (13), we know the system is under-actuated,
control design directly based on (1) is difficult because there is zero
dynamics within the system. We can observe that z1 and z3 are
coupled. Making use of the physical properties just mentioned plus
a simple manipulation of (13), we can obtain three subsystems:
z1 -subsystem, z2 -subsystem, and z3 -subsystem, respectively, as
follows
m11 z1 = 1 v1 g1 d1 m12 z2 m13 z3
1
1
(m22 m21 m
z2 + (m23 m21 m
z3
11 m12 )
11 m13 )
1
1
(14)
+ v2 + g2 + d2 m21 m11 v1 m21 m11 g1
1
1
m21 m
11 d1 = 2 m21 m11 1
1
1
(m32 m31 m
z2 + (m33 m31 m
z3
11 m12 )
11 m13 )
1
1
+ v3 + g3 + d3 m31 m
11 v1 m31 m11 g1

(15)

1
1
m31 m
11 d1 = m31 m11 1

R (q)q.

1 T

(6)

The dynamic equation (1), which satisfies the nonholonomic


constraint (2), can be rewritten in terms of the internal state
variable z as
M (q)R(q)z + V z + G(q) + F = B(q) + J T (q)

(7)

with V = [M (q)R (q) + V (q, q )R(q)]. Substituting (4) and (5) into
(1), and then pre-multiplying (1) by RT (q), the constraint matrix
J T (q) can be eliminated by virtue of (3). As a consequence, we
have the transformed nonholonomic system

M1 (q)z + V1 (q, q )z + G1 (q) + F1 = RT u

(8)

where M1 (q) = R M (q)R, V1 (q, q ) = R [M (q)R + V (q, q )R],


G1 (q) = RT G(q), u = B(q) , F1 = RT F . which is more appropriate
for the controller design as the constraint has been eliminated
from the dynamic equation.
The force multipliers can be obtained by (7)
T

m11
m21
m31

" #

(5)

From (4), z can be obtained from q and q as


T

Considering (8), z (t ) has physical meaning, for example,


consisting of the linear velocity v and the angular velocity , and
the tilt angle , i.e., z (t ) = [z1 z2 z3 ]T = [ ]
T , we denote the
variable z3 for the inverted pendulum, and select z1 coupled with
z3 from the remaining variables such that the dimension of z1 and
z3 is same, and the left variables are denoted with z2 , we have

= Z [M (q)R (q) + V (q, q )R(q)]z + G + F u

(10)

Then, (8) and (9) can be changed to


M1 z + V1 z + G1 + F1 = RT a

(11)

= Z [M (q)R (q) + V (q, q )R(q)]z + G + F a + b .




Let a = m22 m21 m11 m12 , b = m23 m21 m11 m13 , l = m32
1
1
1
m31 m
z2 +
11 m12 , j = m33 m31 m11 m13 , e = (m22 m21 m11 m12 )
1
1
(m23 m21 m11 m13 )z3 , f = (m32 m31 m11 m12 )z2 + (m33
1
1
1
m31 m
z3 , h = (m21 m21 m
z1 + g2 + d2 m21 m
11 m13 )
11 m11 )
11 g1
1
1
1
m21 m11 d1 , k = (m31 m31 m11 m11 )z1 + g3 + d3 m31 m11 g1
1
T
m31 m
11 d1 . Then, let = [z3 , z2 ] , we can rewrite (15) and (16) as

M + V + D = B1 U
where M =

1
m33 m31 m11 m13
1
m23 m21 m11 m13

j
b

l
a

, D =

(17)

hi
k
h

1
m32 m31 m11 m12
1
m22 m21 m11 m12


, B1 =

1
m31 m11
1
m21 m11

0
1

, V =

= diag[1, 1], U =

[1 , 2 ]T .
Property 3.1. The inertia matrix M is symmetric and positive
definite.

2V is skew-symmetric.
Property 3.2. The matrix M

(9)

where Z = (JM 1 J T )1 JM 1 . Consider the control input u in (8)


decoupled into the motion control a and the force control b in
the following form:
u = a J T b .

(16)

(12)

Property 3.3. The eigenvalues of the inertia matrix B1 are positive.


Remark 3.2. There exist the minimum and maximum eigenvalues
min (B1 ) and max (B1 ), such that x R(m+n ) , xT min (B1 )Ix
xT B1 x xT max (B1 )Ix with the unit matrix I, and the known
positive parameter b satisfying 0 < b < min (B1 ), that is, xT bIx
xT min (B1 )Ix.
Remark 3.3. The dynamics model in (13) is the general form for
the mechanical system, but for the wheeled inverted pendulum,
the detained dynamical model is presented in (56).

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

4. Problem statement
4.1. Control objectives
By appropriate selection of a set of vector z (t ) Rm+n , since
the control input is only m dimension, and m > n , only m
variables of z (t ) can be controlled, the control objective can be
specified as: design a controller that ensures the tracking errors
of zi (1 i m) from their respective desired trajectories zid (t ) to
be within a small neighborhood of zero, i.e.,

|zi (t ) zid (t )| i ,

i = 2, 3

(18)

where i > 0. Ideally, i should be the threshold of measurable


noise, while the constraint force error ( d ) is bounded in a
certain region. At the same time, n variables of z (t ) are to be kept
bounded. The variables zi , (1 i m), can be thought of as the
output equation of the system.
Assumption 4.1. The desired reference trajectory zid (t ), (1 i
m) is assumed to be bounded and uniformly continuous, and has
bounded and uniformly continuous derivatives up to the second
order. The desired d (t ) is bounded and uniformly continuous.
Remark 4.1. Since we can plan and design the desired trajectory
for zid (t ) before implementing control, it is reasonable and feasible
that we give the trajectory satisfying the Assumption 4.1.

1349

1
4 )) + ),
u24 = sgn(s)(ln(cosh(
b
1
5 )) + ),
u25 = sgn(s)(ln(cosh(
b
1
6 )) + ),
u26 = sgn(s)(ln(cosh(
b

(27)
(28)
(29)

with sgn(s) = kssk , which are adaptively tuned according to


c 1 = 1 c1 + 1 kskkr k,

c1 (0) > 0
c 2 = 2 c2 + 2 kskkr k, c2 (0) > 0
c 3 = 3 c3 + 3 kskkr kkqk, c3 (0) > 0

(30)

c 4 = 4 c4 + 4 ksk,

(33)

(31)
(32)

c4 (0) > 0
c 5 = 5 c5 + 5 kskkqk, c5 (0) > 0
c 6 = 6 c6 + 6 kskku1 k, c6 (0) > 0

(34)
(35)

1 = c1 kr k,
2 = c2 kr k,
3 = c3 kqkkr k,
4 = c4 ,
with

5 = c5 kqk, 6 = c6 ku1 k, and i > 0 being design parameters


and satisfying
lim i = 0

(36)

i (t )dt = % <

(37)

5. Robust adaptive motion/force control

with finite constant % , and i > 0 as design parameter.

5.1. z2 and z3 -subsystems motion control

Remark 5.1. As we discuss in Section 1, a wheeled inverted pendulum subject to kinematic constraints and dynamic constraints is
apparently different from generalized nonholonomic systems, the
control (23) is proposed based on the obtained reduced model (17),
for the generalized nonholonomic system, if we can obtain the corresponding reduced model similar to (17), the control (23) is also
applicable.

Let us define the following notations as e = d , e = d ,


r = d 1 e , s = e +1 e , where d = [z3d , z2d ]T , r = [z3r , z2r ]T
is the reference signal described in internal state space, and 1 is
positive diagonal. Apparently, we have

= r + s.

(19)

From the dynamic equation (17) together with (19), we have

Ms = V s Mr V r D + B1 U.

(20)

Let M0 , V0 , D0 and B10 be nominal parameter vectors which


give the corresponding nominal functions M0 r + V0 r + D0 and
(B10 )1 , respectively. There exist some finite positive constants
ci > 0 (1 i 6), such that q, q Rn , kM M0 k c1 ,
kV V0 k c2 + c3 kqk, kD D0 k c4 + c5 kqk, kB1 B10 k c6 .
The proposed control for the system is given as

U = u1 + u2

1
1 )) + ),
u21 = sgn(s)(ln(cosh(
b
1
2 )) + ),
u22 = sgn(s)(ln(cosh(
b
1
3 )) + ),
u23 = sgn(s)(ln(cosh(
b

1
2

sT M s +

6
1X

2 i=1

ci i1 ci

(38)

where ci = ci ci , i = 1, . . . , 6, therefore, c i = c i . Its time


derivative is given by

1 = sT
V


M s +

1
2

M s +

6
X

ci i1 c i .

(39)

i =1

Considering Property 3.2, and substituting (20) into (39), we have


(22)

where Kp is a diagonal positive constant, and u2 is designed


to compensate for the parameter errors and the function
approximation errors arising from approximating the unknown
function as
u2 = u21 + u22 + u23 + u24 + u25 + u26

V1 =

(21)

where u1 is the nominal control


1
1
u1 = B10
Kp s + B10
(M0 r + V0 r + D0 )

To analyze closed loop stability for the z2 and z3 subsystems,


consider the following Lyapunov function candidate

(23)
(24)

1 = sT B1 U Mr + V r + D
V


(26)

6
X

ci i1 c i .

(40)

i =1

Integrating (22) and (23) into (40), we have

1 = sT (B1 B10 )u1 + B1 u2 + B10 u1 Mr V r D


V


6
X

ci i1 c i

i=1

= sT [(B1 B10 ) u1 + B1
(25)



6
X

u2i Kp s (M M0 )r

i=1

(V V0 )r (D D0 )] +

6
X
i =1

ci i1 c i

1350

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

= sT Kp s + s B1 u21 (M M0 )r


+ sT B1 (u22 + u23 ) (V V0 )r

T

6
6
=
c6 c6 =
6
6

+ sT [B1 (u24 + u25 ) (D D0 )]


6
X
1
ci c i .
+ s [B1 u26 + (B1 B10 )u1 ] +

i
i=1

(41)

B1 u21 (M M0 )r +

1 )) + ) + c1 c 1
c1 kskkr k ksk(ln(cosh(
1


1
c1 kskkr k c1 kskkr k + c1
c1 kskkr k
1
2

1
1 2
1
1
+
c1 c1 =
c1 c1
c1
=
1
1
2
41

c2 kskkr k + c3 kskkqkkr k
3
X
i
i=2

(42)

i )) + )
ksk(ln(cosh(

c2 kskkr k + c3 kskkqkkr k c2 kskkr k c3 kskkqkkr k






2
3
+ c2
c2 kskkr k + c3
c3 kskkqkkr k
2
3

2
3
3
X i
X i
1
i 2
=
ci ci =
ci ci
+
ci
(43)
i
i
2
4i
i=2
i=2
Similarly, considering (27), (28), (33), (34) and Lemma 2.2, the
fourth right-hand term of (41) is bounded by

c4 ksk + c5 kskkqkk k

5
X

5
X
1
ci c i

i
i=4

i )) + )
ksk(ln(cosh(

5
X
1
+
ci c i

i
i=4

c4 ksk + c5 kskkqk c4 ksk c5 kskkqk






4
5
+ c4
c4 ksk + c5
c5 kskkqk
4
5

2
5
5
X i
X i
1
i 2
=
ci ci =
ci ci
+
ci .
i
i
2
4i
i=4
i=4

c6 c 6


c6 kskku1 k c6 kskku1 k + c6

(45)


6
c6 kskku1 k
6

ci

1
2

2
ci

6
X
i 2
c .
4
i i
i=1

6
X
i 2
ci .
4i
i =1
P6

i 2
i=1 4i ci

(46)
is bounded, there exists

%1 with the finite constant %1 , when

as T . Integrating both sides of (46) gives V1 (t )


RT
R T P6 i 2
V1 (0) 0 min (Kp )ksk2 dt + 0
i=1 4 ci dt. Since ci and i

i dt = %i , we can rewrite (46) as


0
RT
P6 %
V1 (t ) V1 (0) 0 min (Kp )ksk2 dt + i=1 4i ci2 < . Thus V1
i
is bounded, which implies that s L . From the above equation,
RT
P
6
%
we have 0 min (Kp )ksk2 dt V1 (0) V1 (t ) + i=1 4i ci2 , which
i
leads to s L2 . From s = e + 1 e , it can be obtained that e , e
L . As we have established e , e L , from Assumption 4.1, we
conclude that , r L .

are constants, moreover,

Therefore, all the signals on the right hand side of (20) are
bounded, and we can conclude that s and therefore is bounded.
Thus, s 0 as t can be obtained. Consequently, we have
e 0, e 0 as t . It follows that e , e 0 as t .
5.2. z1 -subsystem stability
Finally, for system (15)(16) under control laws (21), apparently, the z1 -subsystem (14) can be rewritten as

= f (, , U)

(47)

where = [z1 , z1 ] , = [ , T ]T = [z3 , z2 , z3 , z2 ]T .


T

kv1 + g1 + d1 k L1 k k + L1f

(48)

kf + kk L2 k k + L2f

(49)

where = [z3 , z2 , z3 , z2 ]T , moreover, from the stability analysis


of 2 and 3 subsystems, converges to a small neighborhood of
d = [z3d , z2d , z3d , z2d ]T .

(44)

Similarly, considering (29) and (35) and Lemma 2.2, the fifth righthand term of (41) is bounded by
1

c62 .

Assumption 5.1. From (15) and (16), the reference signal satisfies
Assumption 4.1, and the following functions are Lipschitz in ,
i.e., there exist Lipschitz positive constants L1 , L2 , and L1f , L2f ,
such that

i =4

sT [B1 u26 + (B1 B10 )u1 ] +

6
46

0. From all the above, s converges


q
%1
to a small set : ksk
containing the origin
min (Kp )

ci c i

sT [B1 (u24 + u25 ) (D D0 )] +

c6

Therefore

i 2
T > t1 ,
c
q i=1 4i i
%1
ksk min (Kp ) , then V 1

i=2

2


6
X
i

P6

3
X
1
ci c i

i
i=2

3
X

Considering (36) and (37),

Considering (25), (26), (31), (32), and Lemma 2.2, the third
right-hand term of (41) is bounded by

1 min (Kp )ksk2


V

1 min (Kp )ksk2 +


V

c1 c 1

sT B1 (u22 + u23 ) (V V0 )r +

c6

i =1

Considering (24) and (30), and Lemma 2.2, the second righthand term of (41) is bounded by

Combining (42)(45), we could obtain


T

Remark 5.2. In (48), v1 might be function of z including z1 and z1 ,


however, for the dynamics presented in Section 6, we can see v1 is
a function only including z3 , therefore, Assumption 5.1 is satisfied.
Remark 5.3. Under the stability of z2 and z3 subsystems, and
considering (18), we have k d k 1 , it is easy to obtain k k
kd k + 1 , and similarly, let = [z3 , z2 ]T , and d = [z3d , z2d ]T ,
kk kd k + 2 , where 1 and 2 are small bounded errors.
Lemma 5.1. If z2 -subsystem and z3 -subsystem are stable, the z1 subsystem (14) is globally asymptotically stable too.

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

Proof. From (14)(16), we choose the following function as


V2 = V1 + ln(cosh(z1 )).

(50)

Differentiating (50) along (14) gives

2 = V 1 + tanh(z1 )z1
V
1
2 m13 z3 ). (51)
= V 1 + tanh(z1 )m
11 (1 v1 g1 d1 m12 z

From (16), we have


1
2 + jz3 + f + k).
1 = m11 m
31 (lz

(52)

Integrating (52) into (51), we have


1
2 = V 1 + tanh(z1 )(m
2 + jz3 + f + k)
V
31 (lz
1
2 m13 z3 ))
+ m
11 (v1 g1 d1 m12 z
T  

1
1
tanh(z1 )(m
z2
31 l + m11 m12 )
= V 1
1
1
z3
tanh(z1 )(m
31 j + m11 m13 )
1
tanh(z1 )(f + k) tanh(z1 )m
11 (v1 + g1 + d1 ).
1
Since
k tanh(z1 )k 1, m
31 , m12 and m13 are all bounded. Let

T
 1
m31 l + m111 m12
1
1 , and km
m1 j + m1 m
11 k 2 , where 1 and 2 are

31
11 13
bounded constants. Considering Assumption 5.1 and Remark 5.3,
we have

X i
2 1 min (Kp )ksk2 +
ci2
V
2
4

i
i =1
+ 1 (kd k + 2 ) + L2 (kd k + 2 ) + L2f
6

Kp such that s can be arbitrarily small. Therefore, we can obtain


that the internal dynamics is stable with respect to the output
z1 . Therefore, the z1 -subsystem (14) is globally asymptotically
stable. 
Theorem 5.1. Consider the system (15)(16) with Assumption 4.1,
under the action of control laws (21). For compact set , where
(z2 (0), z3 (0), z2 (0), z3 (0)) , the tracking errors converges to the
compact set defined in (18), and all the signals in the closed loop
system are bounded.
Proof. From the results (18), it is clear that the tracking error s
converges to the compact set defined by (18). From Lemma 2.1,
we can know e , e are also bounded. From the boundedness of
z2d , z3d in Assumption 4.1, we know that z2 , z3 are bounded. Since
z2d , z3d are also bounded, it follows that z2 , z3 are bounded. From
Lemma 5.1, we know that the z1 -subsystem (15) is stable, and z1 , z1
are bounded. This completes the proof. 
5.3. Force control
The force control input b is designed as
T

(54)

where e = d , Kf is a constant matrix of proportional control


feedback gains. Substituting the control (21) and (54) into the
reduced order dynamics (12) yields

(I + Kf )e = Z [(M (q)R (q) + C (q, q )R(q)z ) + G + F a ] + b


T

= Z R+ M1 z + c1 Z R+ M1 zd .

6. Simulation
Let us consider a mobile wheeled inverted pendulum as shown
in Fig. 1. The following variables have been chosen to describe the
vehicle (see also Fig. 1): l , r : the torques of the left and right
wheels; : the tilt angle of the pendulum; : the direction angle
of the mobile platform; r: the radius of the wheels; d: the distance
between the two wheels; 2l: the length of the pendulum; m: the
mass of the mobile pendulum; Mw : the mass of each wheel; Im :
the moment of inertia of the mobile pendulum; Iw : the inertia
moment of each wheel; g: gravity acceleration; : the motion
friction coefficient of the ground.
The wheeled inverted pendulum is subject to the following
constraints x sin y cos = 0. Using the Lagrangian approach,
we can obtain the reduced dynamics for qv = [x, y, ]T , q = ,
J = [sin , cos , 0, 0], and z = [, , ]
T in (13) as

M1 =

m11

(55)

m22

ml cos
0

ml2 sin 2
2
1
ml2 sin 2
2

2 cos t

where m11 =

1
2Mw r 2
r2
2 2
4 mrd2 l sin2

4r 2
I
d2 m

0
0
mgl sin

"

G1 =

ml + Im

ml cos
2

F1 = 1 sin t

2Iw +

0
1

0
V1 =

min (Kp )

b = c1 Z R+ zd + d Kf e

Since z2 z2d , z3 z3d , z1 is bounded from Theorem 5.1,


T
T
Z R+ M1 z + c1 Z R+ M1 zd is also bounded, therefore, the size of
e can be adjusted by choosing the proper gain matrix Kf . Since s,
z, z , r , r , and e are all bounded, it is easy to conclude that is
bounded from (21) and (54).

2
0

(53)
+ 2 (L1 (kd k + 1 ) + L1f ).
P6 i 2
Let =
i=1 4i i + 1 (kd k + 2 ) + L2 (kd k + 2 ) + L2f +
2 (L1 (kd k + 1 ) + L1f ) and q
it is apparently bounded positive, we
2
2 0, when ksk
have V
, we can choose the proper

1351

ml sin

ml2 sin 2 ,
2

3 sin t

T

+ 2Iw + mr 2 ) and m22 =

(56)
d2
4r 2

(2Mw r 2 +

).

In the simulation, we assume the parameters Iw = 0.5 kg m2 ,


Mw = 0.2 kg, Im = 1.0 kg m2 , m = 10.0 kg, l = 1.0 m,
d = 1.0 m, r = 0.5 m, 1 = 2 = 3 = 1.0,
z (0) = [0, 0.2, /180]T , z (0) = [0.1, 0.0, 0.0]T . The desired
trajectories are chosen as d = 0.05t rad, d = 0.0 rad,
initial velocity is 0.1 m/s. The external disturbances are set as
1.0 sin(t ) and 1.0 cos(t ). The desired nonholonomic constraint
force is set as 10 N. The system state is observed through the
noisy linear measurement channel, zero-mean Gaussian noises
are added to the state information. All noises are assumed to be
mutually independent. The noises have variances corresponding
to a 5% noise to signal radio. The design parameters of the above
controllers are: Kp = diag[10.0, 800.0], 1 = diag[1.0, 5.0], i =
0.5, i = = 1/(t + 1)2 , B10 = [{1.0, 0}; {1.0, 1.0}], b =
0.9, [c1 (0), . . . , c6 (0)]T = [1.0, 1.0, 10.0, 20.0, 10.0, 10.0]T , Kf =
10.0.
The trajectory tracking by the proposed control approach is
shown in Fig. 8, and direction angle, the tilt angles for the
dynamic balance and the stable velocities are shown in Figs. 2
5, respectively. The input torques are shown respectively in
Fig. 6. The nonholonomic constraint force error with the desired
force is shown in Fig. 7. Therefore, the slipping and slippage
cannot happen since the friction force converges within the
region of 10 N. From these figures, even if without the prior
knowledge of the system, we can obtain good performance by
the proposed adaptive robust control. Robust adaptive approach is
tolerant of modeling errors because accurate modeling of wheeled
inverted pendulums dynamics is difficult, and time-consuming

1352

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

Fig. 2. Tracking the direction angle.

Fig. 5. The stable velocity.

Fig. 3. The direction angle error.


Fig. 6. Input torques.

Fig. 4. The tilt angle tracking error.

and uncertain. The presence of parametric errors is a common


problem since the identification of dynamic parameters is errorprone. The robust adaptive control presented in this paper is not
susceptible to this problem, since the unknown parameters are
learned during the wheeled inverted pendulum operation in actual
conditions. Although the parametric uncertainties and the external
disturbances are both introduced into the simulation model, the
motion/force control performance of system, under the proposed
control, is not degraded. The simulation results demonstrate the
effectiveness of the proposed adaptive control in the presence of
unknown nonlinear dynamic systems and environments.

Fig. 7. The nonholonomic constraint force error with d = 10 N.

7. Conclusions
In this paper, robust adaptive motion/force control design is
carried out for dynamic balance and stable tracking of desired trajectories of a mobile wheeled inverted pendulum, in the presence
of unmodelled dynamics, or parametric/functional uncertainties

Fig. 8. The produced trajectory.

Z. Li, Y. Zhang / Automatica 46 (2010) 13461353

and nonholonomic constraint force. The control is mathematically


shown to guarantee semi-global uniformly bounded stability, and
the steady state compact sets to which the closed loop error signals converge are derived. The size of compact sets for motion and
force can be made small through appropriate choice of control design parameters. Simulation results demonstrate that the system
is able to track reference signals satisfactorily, with all closed loop
signals uniformly bounded.
References
Arai, H., & Tanie, K. (1998). Nonholonomic control of a three-dof planar underactuated manipulator. IEEE Transactions on Robotics and Automation, 14(5), 681694.
Blankespoor, A., & Roemer, R. (2004). Experimental verification of the dynamic
model for a quarter size self-balancing wheelchair. In American control conference, Boston, MA (pp. 488492).
Brooks, R., Aryanada, L., Edsinger, A., Fitzpatrick, P., Kemp, C. C., & OReilly, U.
(2004). Sensing and manipulating built-for-human environments. International
Journal of Humanoid Robotics, 1(1), 128.
Dong, W. (2002). On trajectory and force tracking control of constrained mobile
manipulators with parameter uncertainty. Automatica, 38(8), 14751484.
Gans, N.R., & Hutchinson, S.A. (2006). Visual servo velocity and pose control of a
wheeled inverted pendulum through partial-feedback linearization. In Proc.
IEEE/RSJ int. conf. intelligent robots and systems (pp. 38233828).
Ge, S. S., Lee, T. H., & Harris, C. J. (1998). Adaptive neural network control of robot
manipulators. London: World Scientific.
Ge, S. S., Wang, J., Lee, T. H., & Zhou, G. Y. (2001). Adaptive robust stabilization
of dynamic nonholonomic chained systems. Journal of Robotic System, 18(3),
119133.
Ge, S. S., Wang, Z., & Lee, T. H. (2003). Adaptive stabilization of uncertain nonholonomic systems by state and output feedback. Automatica, 39(8), 14511460.
Grasser, F., Arrigo, A., Colombi, S., & Rufer, A. C. (2002). Joe: a mobile, inverted
pendulum. IEEE Transactions on Industrial Electronics, 49(1), 107114.
Ha, Y. S., & Yuta, S. (1996). Trajectory tracking control for navigation of the inverse
pendulum type self-contained mobile robot. Robotics and Autonomous System,
17, 6580.
Ibanez, C. A., Frias, O. G., & Castanon, M. S. (2005). Lyapunov-based controller for
the inverted pendulum cart system. Nonlinear Dynamics, 40(4), 367374.
Isidori, A., Marconi, L., & Serrani, A. (2003). Robust autonomous guidance: an internal
model approach. New York: Springer.
Jung, S., & Kim, S. S. (2008). Control experiment of a wheel-driven mobile inverted pendulum using neural network. IEEE Transactions on Control Systems
Technology, 16(2), 297303.
Kim, Y., Kim, S. H., & Kwak, Y. K. (2005). Dynamic analysis of a nonholonomic twowheeled inverted pendulum robot. Journal of Intelligent and Robotic Systems,
44, 2546.

1353

Li, Z., & Luo, J. (2009). Adaptive robust dynamic balance and motion controls of
mobile wheeled inverted pendulums. IEEE Transactions on Control Systems
Technology, 17(1), 233241.
Luca, A. D., & Oriolo, G. (2002). Trajectory planning and control for planar robots
with passive last joint. The International Journal of Robotics Research, 21(56),
575590.
Nasrallah, D. S., Michalska, H., & Angeles, J. (2007). Controllability and posture
control of a wheeled pendulum moving on an inclined plane. IEEE Transactions
on Robotics, 23(3), 564577.
Pathak, K., Franch, J., & Agrawal, S. K. (2005). Velocity and position control of a
wheeled inverted pendulum by partial feedback linearization. IEEE Transactions
on Robotics, 21(3), 505513.
Salerno, A., & Angeles, J. (2003). On the nonlinear controllability of a quasiholonomic
mobile robot. In Proc. IEEE int. conf. robotics and automation (pp. 33793384).
Salerno, A., & Angeles, J. (2004). The control of semi-autonomous two-wheeled
robots undergoing large payload-variations. In Proc. IEEE int. conf. robotics and
automation (pp. 17401745).
Zhang, M., & Tarn, T. (2002). Hybrid control of the pendubot. IEEE/ASME Trans.
Mechatronics, 7(1), 7986.

Zhijun Li received Dr. Eng. Degree in Mechatronics,


from Shanghai Jiao Tong University, PR China, in 2002.
From 2003 to 2005, he was a postdoctoral fellow in
Department of Mechanical Engineering and Intelligent
systems at the University of Electro-Communications,
Tokyo, Japan. From 2005 to 2006, he was a research fellow
in the Department of Electrical and Computer Engineering
at the National University of Singapore, and Nanyang
Technological University, Singapore. Currently, he is an
associate professor in the Department of Automation,
Shanghai Jiao Tong University, PR China. Dr. Li is IEEE
Senior Member and his current research interests are adaptive/robust control,
mobile manipulators, nonholonomic systems, etc.

Yunong Zhang received B.S., M.S. and Ph.D. degrees


respectively from Huazhong University of Science and
Technology (HUST), South China University of Technology
(SCUT) and the Chinese University of Hong Kong (CUHK),
respectively, in 1996, 1999 and 2003. He is currently
a professor at the School of Information Science and
Technology, Sun Yat-Sen University (SYSU), Guangzhou,
China. Before joining SYSU in 2006, he had been with
the National University of Ireland (NUI), University of
Strathclyde, and National University of Singapore (NUS)
since 2003. His main research interests include neural
networks, robotics and Gaussian processes. His web-page is now available at
http://www.ee.sysu.edu.cn/teacher/detail.asp?sn=129.

Potrebbero piacerti anche