Beruflich Dokumente
Kultur Dokumente
Abstract—A mobile robot is one of the well-known nonholonomic sys- q + V (q; q_ )q_ + G(q ) = B (q ) + AT (q )
M (q ) (1)
tems. The integration of a kinematic controller and a torque controller for
the dynamic model of a nonholonomic mobile robot has been presented. In where q 2 Rn is generalized coordinates, 2 Rr is the input vector,
this paper, an adaptive extension of the controller is proposed. If an adap- 2 Rm is the vector of constraint forces, M (q ) 2 Rn2n is a sym-
metric and positive-definite inertia matrix, V (q; q_ ) 2 Rn2n is the
tive tracking controller for the kinematic model with unknown parame-
I. INTRODUCTION With respect to the dynamics of mobile robot (1), the following prop-
erties are known [10].
A mobile robot is one of the well-known systems with nonholonomic Property 1: M (q ) is a symmetric and positive-definite matrix.
Property 2: There is a parameter vector p0 2 Rl on dynamics that
constraints, and there are many works on its tracking control [1]–[4].
Their objects are mostly kinematic models, but recently one method
satisfies the following equation [11]:
for dynamic models has been proposed [5]. This method integrates a
kinematic controller and a torque controller for the dynamic model of
M (q )v_ + V (q; q_ )v + G(q ) = Y0 (q; q;
_ v; v_ )p0 (3)
a nonholonomic mobile robot by using backstepping [6].
where v 2 Rn and Y0 is an n 2 l0 matrix whose elements consist of
The control input of the controller for the kinematic model is gen-
erally velocity, but it is more realistic that the input is torque. In [5], a
known functions.
Property 3: The matrix M_ 0 2V is skew-symmetric [12], that is,
kinematic controller is designed first so that the tracking error between
a real robot and a reference robot converges to zero, and secondly a
xT (M _ 0 2V )x = 0, 8 x 2 Rn .
torque controller is designed by using backstepping so that the veloc-
The nonholonomic mobile robot (1) is transformed to and divided
ities of a mobile robot converge to the desired velocities, which are
into the following two equations [5]:
given by the kinematic controller designed at the first step.
In this paper, we present a method to design an adaptive tracking
q_ = S (q ) (t) (4)
controller for the dynamic model of a nonholonomic mobile robot with
unknown parameters by adaptive backstepping. The adaptive control M (q )_ + V (q; q_ ) + G(q ) = B (q ) (5)
where we let = [1 ; . . . ; n0m ], i = [i1 ; . . . ; il ]T , 1 i technique [6]. The adaptive control technique for the dynamic part (5)
n 0 m are parametric vectors, and ij (q ), 1 i n 0 m, 0 j li is based on [10].
are vectors whose elements consist of known functions. Theorem III.1: If Assumptions II.1 and Assumptions III.1–III.3 are
Furthermore, the following Property 2″ is satisfied. satisfied for a nonholonomic mobile robot (1), the following adap-
Property 2″: There is a parametric vector p 2 Rl on kinematics and tive tracking controller (13)–(16) achieves the control objective: q !
dynamics which satisfies qr (t ! 1) and the boundedness of the signals included in V2 , which
is defined in (17).
M (q )_ + V (q; q_) + G(q) = Y (q; q;_ ; _ )p (8)
T
= B 01 0Kd ~ + Yc p^ 0 @V1 ^
S
where Y is (n 0 m) 2 l matrix whose elements consist of known func-
(13)
@q
tions and p is a parametric vector which is composed of the elements ^_ i
a = Ti (q; qr ; a^); 1ik (14)
of p1 and i . T
_
^i = 3i
@V1
i ~i ; 1in0m (15)
@q
III. ADAPTIVE TRACKING CONTROL OF A NONHOLONOMIC
MOBILE ROBOT 0
p^_ = 0YcT ~ (16)
In [5], (t) is considered as a control input for the kinematic part where ^ = [^1 ; . . . ; ^n0m ] is the estimate of , Yc Y (q; q;
_ c ; _ c ),
(4), and an ideal control input c (t) is designed to track a reference
S^ S (q; ^), ~ = c = [~ 0 1 ; . . . ; ~n0m ], i = [i1 ; . . . ; il ],
trajectory. Since the real input of the mobile robot (1) is , is designed
to make (t) 0 c (t) ! 0 as t ! 1 by using backstepping [6]. But if
0
1 i n m, and Kd , 0, 3i , 1 i n m are symmetric and 0
positive-definite matrices with appropriate dimensions.
there exist some unknown parameters in a mobile robot, that is, S (q )
V2 is defined as
has unknown parameters or p is unknown in (8), we cannot design a
n0m
= V1 + 1 ~T M ~ + 1 p~T 001p~ + 1 ~T 301~
tracking controller according to [5].
In this paper, it is shown that an adaptive tracking controller can be V2 i i i (17)
designed for the dynamic model with unknown parameters if it is pos-
2 2 i=1 2
sible to design an adaptive tracking controller for the kinematic model
with unknown parameters. with p~ = p^ 0 p, ~ = ^ 0 .
Control Objective: Design an adaptive tracking controller for a non- Proof: The derivative of V2 is
holonomic mobile robot (1), in order that
k
V_ 2 = @V1 S (c + ~) + @V1 q_r + @V1
T
lim (q(t) 0 qr (t)) = 0 (9) @q @qr ^i i
i=1 @ a
t!1
Assumption III.2 is easily relaxed by the existing adaptive control From Assumption III.1, (17), and (18), the signals included in V2
technique, if B is constant and the sign of each elements is known. are bounded. Because ~_ is proved to be bounded, V2 2 L1 . From
Assumption III.3 can be always satisfied by the appropriate selection Barbalat’s lemma [14], [6], we can show (t) ! c (t) as t ! 1.
of V1 . Therefore, the equation q_ = S = S (~ + c ) = Sc + S ~ shows
Now, we provide the following theorem to design an adaptive q (t) ! qr (t) as t ! 1, since ~ ! 0 as t ! 1 and the kinematic
tracking controller of a mobile robot using adaptive backstepping model of a mobile robot satisfies Assumption III.1
IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 16, NO. 5, OCTOBER 2000 611
where
sin 0 cos 0 0 0
A(q ) = cos sin b 0r 0 : (25)
cos sin 0 b 0 0r
Equations (4) and (5) can be written as the following:
q_ = S (q ) (t) (26)
M (q )_ + V (q; q_ ) = B (q ) (27)
where S (q ) is selected as
r r
cos cos
2 2
r r
sin sin
2 2
r
S (q ) =
2b
0 2rb (28)
1 0
Fig. 1. Mobile robot with two actuated wheels.
0 1
Remark III.1: Because adaptive control is applied to treat unknown and M , V , B are expressed as
r2 r2
(mb 0 I )
parameters in the kinematic part, it is more important to consider the 2 2
2 ( mb + I ) + Iw
4b2
dynamic part properly, that is, our proposed model-based controller M = 4b 2
r r2
(mb 0 I )
is better than a high-gain feedback controller to treat the dynamics. 2 2
2 (mb + I ) + Iw
As is generally known, the adaptive control system designed for the 4b 4b2
2
kinematics may be unstable if there exists the error ~ [15]. r
0 mc d_
2b
V = 2
IV. MOBILE ROBOT WITH TWO ACTUATED WHEELS 0 r2b mc d_ 0
In this section, we consider a mobile robot with two actuated 1 0
wheels as an example which the theorem can be applied to. An B= (29)
0 1
adaptive tracking controller is designed for the kinematic model and
the dynamic model, and some simulation results are provided. and = [r ; l ]T consists of motors’ torques r and l , which act
on the right and left wheels, respectively, and let m = mc + 2mw ,
A. Model of a Mobile Robot with Two Actuated Wheels I = mc d2 + 2mw b2 +Ic + 2Im .
We consider the mobile robot with two actuated wheels, which is B. Adaptive Control of the Kinematic Model
shown in Fig. 1 [16].
We design an adaptive tracking controller for the kinematic part (26)
With regard to the mobile robot shown in Fig. 1, 2b is the width of
the mobile robot and r is the radius of the wheel. O 0 xy is the world
modifying the method proposed by Kanayama et al. [9].
coordinate system and P0 0 XY is the coordinate system fixed to the
First, we consider as a control input and construct the adaptive
mobile robot. P0 is the origin of the coordinate system P0 0XY and the
control system for the following kinematic model:
r r
middle between the right and left driving wheels. The center of mass of cos cos
x 2 2
the mobile robot is Pc , which is on the X -axis, and the distance from r r
y sin sin
P0 to Pc is d. For the later description, mc and mw are the mass of d 2 2 1
= r r
the body and wheel with a motor, Ic , Iw , and Im are the moment of dt
r 2b
0 2b
2
(30)
inertia of the body about the vertical axis through Pc , the wheel with 1 0
a motor about the wheel axis, and the wheel with a motor about the l
0 1
wheel diameter, respectively.
The configuration of the mobile robot can be described by five gen- where 1 and 2 represent the angular velocities of right and left
eralized coordinates wheels.
We focus on only three states x, y , , except r and l . The relation-
q = [x; y; ; r ; l ]T (20) ship between v , w , and 1 , 2 is the following:
1 b
where (x; y ) are the coordinates of P0 , is the heading angle of the 1 r r v
mobile robot, and r ; l are the angles of the right and left driving = (31)
b
wheels.
2 1
r
0 r
w
We assume the wheels roll and do not slip. Then, there exist three where v is the straight line velocity and w is the angular velocity of the
constraints; the velocity of P0 must be in the direction of the axis of mobile robot at the point P0 .
symmetry and the wheels must not slip Substituting (31) for (30), we get the ordinary form of a mobile robot
y_ cos 0 x_ sin = 0 (21)
with two actuated wheels
x cos 0
x_ cos + y_ sin + b_ = r_r (22) d v
y = sin 0 : (32)
x_ cos + y_ sin 0 b_ = r_l : (23)
dt
0 1
w
These constraints can be rewritten in the form The various design methods for this system (32) have already been
proposed. Our method is based on the method [9] whose objective is
A(q )q_ = 0 (24) tracking on a reference robot shown in Fig. 2.
612 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 16, NO. 5, OCTOBER 2000
1 a
^1 a^2 vf
2
=
a
^1 0a^2 wf
(41)
a1 + a~1 a2 + a
~2 vf
:
=
a1 + a~1 0a2 0 a~2 wf
(42)
Therefore
d
e1
a
~1
01
e2 = 1+ vf 0
dt a1
e3 0
e2 vr cos e3
Fig. 2. Reference robot and real robot. a
+ 1+
~2
a2
wf 0e1 + vr sin e3 : (43)
The kinematics of the reference robot is given as
01 wr
xr cos r 0 We define V1 as
d vr
yr = sin r 0 (33)
dt wr 1 2 1 2
r 0 1 V1 = V0 + a
~1 + a
~2 (44)
2
1 a1 2
2 a2
where xr , yr , and r are the configure of the reference robot, and vr ,
wr are its reference inputs. with positive constants
1 ,
2 .
We define e1 , e2 , e3 as following: The derivative of V1 is
e1 cos sin 0 xr 0 x
e_ 3 sin e3 a a~2 _
e2 = 0 sin cos 0 yr 0 y :
~1 _
(34) V_ 1 = e1 e_ 1 + e2 e_ 2 + + a
^1 + a
^2
K2
1 a1
2 a2
e3 0 0 1 r 0
a
= e1 0K1 e1 0
~1
e1 , e2 , e3 describe the difference of position and direction of the vf + e2 vr sin e3
a1
reference robot from the real robot. The inputs v , w , which make e1 ,
a sin e3
+ 0K2 e2 vr 0 K3 sin e3 0
~2
e2 , e3 converge to zero, are given by the following [9], [5]: wf
a2 K2
vf = vr cos e3 + K1 e1 a
~1 _ a
~2 _
+ a^1 + a
^2
wf = wr + vr K2 e2 + K3 sin e3 (35)
1 a1
2 a2
a a wf sin e3
a^_ 1 0
1 e1 vf + a^_ 2 0
2
where K1 ; K2 ; K3 are positive constants. ~1 ~2
= V_ 0 + :
We can easily confirm that e1 , e2 , e3 satisfy
1 a1
2 a2 K2
d
e1 01 e2 vr cos e3
(45)
e2 = v 0 + w 0e1 + vr sin e3 : (36)
dt
e3 0 01 wr
Now, the parameter update rules are chosen as
We define V0 as ^_ 1
a =
1 e1 vf
V0
1 2 2
(e1 + e2 ) +
1 0 cos e3 wf sin e3
= (37) ^_ 2
a =
2 + f (^
a2 ) (46)
2 K2 K2
then, the derivative of V0 satisfies the following inequality:
2 where
sin e3
V_ 0 = e1 e_ 1 + e2 e_ 2 + e_ 3
K2
= 0K1 e12 0 K3 sin
K2
e3
0: 0; a^2 >
f (^
a2 ) = 2
a^2
0
(38) 2 (47)
1 (f0 + 1); a^2
If the parameters in kinematics (30), r and b, are unknown, we cannot
choose the inputs as (35) because of the relationship (31) between v , w with f0 = (
2 wf sin e3 )=K2 .
and 1 , 2 . Hence, we design an adaptive controller to attain the control Then
objective by using the estimates of r and b.
a wf sin e3
a^_ 2 0
2
By using 1 and 2 , (36) is transformed to ~2
V_ 1 = V_ 0 +
0 2r + 2rb e2 0 2r 0 2rb e2
2 a2 K2
e1 vr cos e3 a
~2
d r r
dt
e2 = 1 0 2b
e1 + 2
2b
e1 + vr sin e3
= V_ 0 +
2 a2
f (^ a2 )
e3
0 2rb r wr ^2 0 a2
a
= V_ 0 + f (^
a2 ): (48)
2b
2 a2
(39)
where we set When a ^2 > , the second term of the right-hand side of (48) is zero.
When a ^2 , the second term is less than zero because f (^ a2 ) 0,
b
^2 0 a2 0 a2 0. Therefore, it is shown that V_ 1 0. Also, from
1
a1 = and a2 = (40) a
r r (43), e_ 1 and e_ 3 are bounded since a
~1 and a~2 are bounded. After all, V1
and it is assumed that we know a positive constant which satisfies is bounded. Barbalat’s lemma shows that V1 ! 0 as t ! 1, that is,
_
a2 noticing a2 > 0. e1 ! 0 and sin e3 ! 0. Furthermore, (35) and (41) show that 1 and
IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 16, NO. 5, OCTOBER 2000 613
a _
e3 = 0 1+
~2
a2
w_ f 0 aa^22 wr + w_ r
a
= 0 1+
~2
a2
(w
_ r + v_ r K2 e2 + vr K2 e_ 2 + K3 e_ 3 cos e3 )
_
0 aa^22 wf + w_ r : (49)
REFERENCES
[1] Y. Kanayama, Y. Kimura, F. Miyazaki, and T. Noguchi, “A stable
tracking control method for a nonholonomic mobile robot,” in Proc.
IEEE/RSJ Int. Workshop Intelligent Robots and Systems, 1991, pp.
1236–1241.
[2] C. Samson and K. Ait-Abderrahim, “Feedback control of a nonholo-
nomic wheeled cart in cartesian space,” in Proc. IEEE Int. Conf.
Robotics and Automation, 1991, pp. 1136–1141.
[3] Y. Nakamura and S. Savant, “Nonholonomic motion control of an au-
tonomous underwater vehicle,” in Proc. IEEE/RSJ Int. Workshop Intel-
ligent Robots and Systems, 1991, pp. 1254–1259.
Fig. 6. Errors between ideal and real value: ~ , ~ . [4] M. Sampei, T. Tamura, T. Itoh, and M. Nakamichi, “Path tracking control
of trailer-like mobile robot,” in Proc. IEEE/RSJ Int. Workshop Intelligent
Robots and Systems, 1991, pp. 193–198.
The simulation results are shown in Figs. 4–8. From these simulation [5] R. Fierro and F. L. Lewis, “Control of a nonholonomic mobile robot:
results, we can confirm the usefulness of Theorem III.1 and the limi- backstepping kinematics into dynamics,” in Proc. 34th IEEE Conf. De-
tation of Theorem IV.1. The control performance is good when wr is cision Control, 1995, pp. 3805–3810.
close to zero, but the performance becomes bad as wr is far from zero.
[6] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adap-
tive Control Design. New York: Wiley, 1995.
[7] Y. Chang and B. Chen, “Adaptive tracking control design of nonholo-
V. CONCLUSION nomic mechanical systems,” in Proc. 35th IEEE Conf. Decision Control,
1996, pp. 4739–4744.
In this paper, we proposed a design method of an adaptive tracking [8] S. V. Gusev, I. A. Makarov, I. E. Paromtchik, V. A. Yakubovich, and C.
controller for a nonholonomic mobile robot with unknown parame- Laugier, “Adaptive motion control of a nonholonomic vehicle,” in Proc.
1998 IEEE Int. Conf. Robotics and Automation, 1998, pp. 3285–3290.
ters. It was proved that an adaptive tracking controller for the dynamic [9] Y. Kanayama, Y. Kimura, F. Miyazaki, and T. Noguchi, “A stable
model can be designed by using adaptive backstepping if an adaptive tracking control method for an autonomous mobile robot,” in Proc.
tracking controller for the kinematic model exists. As one example, IEEE Int. Conf. Robotics and Automation, 1990, pp. 384–389.
IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 16, NO. 5, OCTOBER 2000 615
[10] J. E. Slotine and W. Li, “On the adaptive control of robot manipulators,” obstacles nearby (GNRON). In most of the previous studies, the goal
Int. J. Robot. Res., vol. 6, no. 3, pp. 49–59, 1987. position is set relatively far away from obstacles. In these cases, when
[11] H. Mayeda, K. Osuka, and A. Kanagawa, “A new identification method the robot is near its goal position, the repulsive force due to obstacles
for serial manipulator arm,” in Proc. 9th IFAC World Congress, 1984,
pp. 2429–2434. is negligible, and the robot will be attracted to the goal position by the
[12] S. Arimoto and F. Miyazaki, “Stability and robustness of PID feedback attractive force. However, in many real-life implementations, the goal
control for robot manipulators of sensory capability,” in Robotics Re- position needs to be quite close to an obstacle. In such cases, when the
search, M. Brady and R. P. Paul, Eds. Cambridge, MA: MIT Press, robot approaches its goal, it also approaches the obstacle nearby. If
1984, pp. 783–799.
[13] C. Su and Y. Stepanenko, “Robust motion/force control of mechan-
the attractive and repulsive potentials are defined as commonly used
ical systems with classical nonholonomic constraints,” IEEE Trans. [2]–[4], the repulsive force will be much larger than the attractive
Automat. Contr., vol. 39, pp. 609–614, Mar. 1994. force, and the goal position is not the global minimum of the total
[14] W. Li and J. Slotine, Applied Nonlinear Control. Englewood Cliffs, potential. Therefore, the robot cannot reach its goal due to the obstacle
NJ: Prentice-Hall, 1991. nearby.
[15] P. A. Ioannou and J. Sun, Robust Adaptive Control. Englewood Cliffs,
NJ: Prentice-Hall, 1996. To overcome this problem, the repulsive potential functions for path
[16] N. Sarkar, X. Yun, and V. Kumar, “Control of mechanical systems with planning are modified by taking into account the relative distance be-
rolling constraints: Application to dynamic control of mobile robots,” tween the robot and the goal. The new repulsive potential function en-
Int. J. Robot. Res., vol. 13, no. 1, pp. 55–69, 1994. sures that the total potential has a global minimum at the goal position.
[17] H. Wang, T. Fukao, and N. Adachi, “Adaptive tracking control of
nonholonomic mobile robots: A backstepping approach,” in Proc. 1998
Therefore, the robot will reach the goal finally. Note that we are not
Japan–USA Symp. Flexible Automation, 1998, pp. 1093–1096. trying to tackle the common local minima problems due to obstacles
[18] , “An adaptive tracking control approach for nonholonomic mobile between the robot and the goal. We shall restrict our attention to the
robot,” in Proc. 1999 IFAC World Congress, 1999, pp. 509–515. formulation and solution of the GNRON problem only.
This paper is organized as follows. In Section II, the cause of the
GNRON problem is analyzed after the introduction of the potential
field methods. Section III presents the new repulsive potential func-
tion and its properties. In Section IV, the relationship between scaling
parameters of the potential functions is presented. In Section V, safety
New Potential Functions for Mobile Robot Path Planning issues of the new potential functions are discussed, and a control system
directly making use of the new potentials is also suggested. Simulation
S. S. Ge and Y. J. Cui results are presented in Section VI to show the problems of the conven-
tional potential field method and the effectiveness of the new method.
Abstract—This paper first describes the problem of goals nonreachable
with obstacles nearby when using potential field methods for mobile robot II. POTENTIAL FIELD METHOD AND GNRON PROBLEM
path planning. Then, new repulsive potential functions are presented by
taking the relative distance between the robot and the goal into considera- For simplicity, we assume that the robot is of point mass and moves
tion, which ensures that the goal position is the global minimum of the total in a two-dimensional (2-D) workspace. Its position in the workspace is
potential. denoted by q = [ x y ]T .
Index Terms—GNRON problem, new repulsive potential function, po- Different potential functions have been proposed in the literature.
tential field. The most commonly used attractive potential takes the form [1]–[3]
1
I. INTRODUCTION Uatt (q) = m (q; qgoal ) (1)
2
Manuscript received August 31, 1999; revised June 15, 2000. This paper 2
was recommended for publication by Associate Editor J. Ponce and Editor
Urep (q) =
1
2
1
01
(q; qobs ) 0
; if (q; qobs ) 0
V. Lumelsky upon evaluation of the reviewers’ comments. This paper was
presented in part at the Third Asian Control Conference, Shanghai, China, July 0; if (q; qobs ) > 0
4–7, 2000. (3)
The authors are with the Department of Electrical and Computer En-