A Stable Tracking Control Method for an Autonomous Mobile Robot

Yutaka Kanayama
Dept of Computer Science Naval Postgraduate School Monterey, CA 93943 CORIDept of Computer Science University of California Santa Barbara, CA 93106

Yoshihiko Kimura
Center for Robotic Systems University of California Santa Barbara,CA 93106 Mitsubishi Metal Corp. Tokyo, Japan

Fumio Miyazaki
Osaka University Dept of Mechanical Engineering Toyonaka, Japan

Tetsuo Noguchi
Center for robotic Systems University of California Santa Barbara, CA 93106

ABSTRACT
This paper proposes a stable tracking control rule for non-bolonomic vehicles. Stability of the rule is proved through the use of a Liapunov function. Input to the vehicle are a reference posture (x,, y,, 8,)' and reference velocities (v,, ar)'. The major objective of this paper is to propose a control rule to find a reasonable target linear and rotational velocities (v, a ) ' . Linearizing the system's differential equation is useful to decide parameters for critical dumping for a small disturbance. In order to avoid any slippage, a velocity/acceleration limitation scheme was introduced. Several siniulation results are presented with or without the velocity/acceleration limiter. The control rule and limiting method proposed in this paper are robot-independent and hence can be applied to various kind of mobile robots with a dead reckoning ability. This method was implemented on the autonomous mobile robot Yamabico-11. Experimental results obtained are close to t h e results with the velocity/acceleration limiter. algorithm for lineadrotational velocity rules in an earlier locomotion control method on the Yamabico-11 robot [7]. Nelson proposed a locomotion control method for a cart with a front steering wheel, in which they also used the error coordinate system [8]. They adopted a linear function in control rules for steering and linear velocity. These two papers are regarded as pioneers of this paper. In this paper, a new control rule for determining vehicle's linear and rotational velocities are given, which are different from both of [7] and [8]. The stability of the control rule is proven using a Liapunov function [91[101[11]. The use of the trace function (I - cose) of orientation 0 is successful in finding an appropriate Liapunov function [Ill. One of t h e difficulties of this problem lies in the fact that ordinary vehicles possess only two degrees of freedom (linear velocity v and rotational velocity CO)for locomotion control, although vehicles have three degrees of freedom, x, y .and 8 in its positioning. Another difficulty is in the non-linearity of the i,6)'. The use of a Liapunov kinematic relation between (v, CO)'and (i, function resolves these difficulties. By linearizing the system's differential equation, we find a condition for critical dumping, which gives appropriate parameters for specilic control rules. The need of velocity/acceleration limitation is also discussed. After these analyses and discussions, abundant simulation results are presented. The method described so far is hardware independent and applicable to ordinary (not omni-directional) vehicles. This method is useful to the class of autonomous vehicles in which (a) a dead reckoning capability is provided, (b) reference path specification and current position estimation (through dead reckoning) are given separately, and (c) high precision in positional control is mandatory. This method was implemented on the autonomous mobile robot Yamabico-11 which has been developed at the University of Tsukuba, the University of Califomia at Santa Barbara, and Naval Postgraduate School. It was demonstrated that these algorithms are sound and provided precise tracking control. An extensive set of the experimental results are shown.

1. Introduction The purpose of this paper is to propose a stable tracking control method for a non-holonomic vehicle with abundant simulation results. Real experimental results on the autonomous mobile robot Yamabico-1I are also presented. Tsumura proposed a method in which the reference point sequence is stored in memory. In each cycle of the locomotion control, the reference point and the future position of the robot is compared for determining the next steering [2]. Kanayama proposed a method using straight line reference for the robot's locomotion instead of a sequence of points 131. Its velocity and steering control niethod has some similarities to the one proposed in this paper. Crowley developed a locomotion control system whose organization ha$ a three layered structure [4]. He defiles the concept of "virtual vehicle" which is useful for constructing a system wluch i s robot independent. In its command system, independent control of linear and rotational motion is possible, thus e n a b h g sniooth clothoid curves [SI. Sin& used a n inverse kinematic and a quintic polynomial metbod for compensating errors in vehicle tracking [6]. In the second method, he interpolates the current point and a future reference point with a smooth curve. Kanayama proposed the use of a reference and current postures for vehicle control, the use of a local error coordinate system, and a PI control

2. Problem Statements

Before stating the problem, we will give a few preliminary definitions.

There is a mobile robot which is located on a 2D plane in which a global Cartesian coordinate system is defined. The robot in the world

384

cH2876-1/90/oooO/0384$01.00 0 1990 IEEE

possesses three degrees of freedom in its positioning which are represented by aposture,
P"

I]

(1)

where the heading direction 0 is taken counterclockwise from the x-axis. Let 0 denote a null posture (0, 0, 2nx)', where n is an integer. Since the robot has a locomotion c a p a b ~ t y in the plane, the posture p is in fact a function of time r. The entire locus of the point (x ( I ) , y ( I ) ) is called a path or trujecrory. If the t i m e derivatives i and y exist, e(t) is not an independent variable any more, because

The vehicle's motion is controlled by its linear velocity v and rotational velocily o, which are also functions of time. The vehicle's kineniatics is defined by a Jacobian m m J : Fig. I

I'

Reference and Current Postures

where q = (v. w)'. This kinematics is common to all kinds of vehicles which are not omnidirectional. (For instance, an automobile, a bicycle, a vehicle with two parallel independent power wheels - power wheeled steering system, and a tricycle) The linear velocity v and rotational velocity w of t h i s kind of vehicle is controlled by its accelerator and s t e e h g wheel or handle respectively.
2.2. Error Posture

Y f

n
12

1+

In this control system, two postures are used; the reference posture p, = ( x r , y,, and the current posture pc = (xc, yc, A reference posture is a goal posture of the vehicle and a current posture is its "real" posture at this moment respectively (Fig. 1). We will deline an error posture p. of the two, which is a transformation of the reference posture p, in a local coordinate system with an origin of (xc, y , ) and an X-axisin the direction of Bc [7][8] (Fig. 2). T h i s is the "difference" between pr and pc:

e,)'

e=)'.

If p r = p c , the error posture pe =O. If pr is ahead of pc (the vehicle is behind of the god), x, > 0. For instance, if pc =(3/2, 1, d 6 ) and pr = (512, 1 + 6 , d4). p. = (6, 1, x / 12) (Figure 2 illustrates this case).

Fig. 2

Error Posture

2 . 3 . Problem
Now, we are able to state the architecture of a tracking control system for the vehicle (Fig. 3). The global input of the system is the reference poslure pr and reference velocities q , = (vr,or)', which are variables of time. The global output of the system is the current posture pc. The purpose of this tracking contcoller is to converge the error posture to 0. Let us describe each component in Figure 3 from left to right. The k t component calculate an error posture &om p, and pc using Eguation (4). The second box is a control rule for the vehicle, which calculates a target velocities q = ( v , o ) ' using the error posture pe and the reference velocities q , = ( v , cor):
(5)

)'(Pc

*(le)
+

PC(0

Fig. 3 The lhird box T stands for the vehicle hardware capability of transforming target velocities to vehicle's real current velocities. In Sections 3 a n d 4,

Architecture of Tracking Controller

385

specifically, we assume the identity !formation:

J:[
J

Soundness of this control rule (8) is established by the following proposition: Proposition 1 . If we use the control mle (8), p. = 0 is a stable equilibrium point if the reference velocity v, > 0. Proof. Let us propose a scalar function V as a Liapunov function candidate [9]:

= cl

This perfect velocio tracking assumption simplifies t h e forthcoming

T h e fourth box is the kinematics matrix M in Equation (3) to produce the derivative of a cumnt posture pc. The last box is for integration. Thus, only unknown component in this system is the control rule. Since the systems input pr is time-variable, it is called non-autonomous by the definition in the control theory [9].
3. A Control Scheme and Its Stability In this section, we will f i n d a stable control rule using a Liapunov function [9]. The following lemma follows the system depicted in Figure 1.

v=+x:+y:)+ci 2
Clearly, V 2 0. If pe = 0, V = 0. If p. 2:

-coSe.)iK,
# 0,

(10)

V > 0. Fwthermore, by Lemma

v =XJ,

+.ye

+ e.sine,/K,
1 ,

= [(w, + V,W,Y. + K e sine.))y. - KP,I

+ [+or + v,(K,y. + K e sine.)k + v,sine,l Y . + [-v,(K,y. + K e sine.)] sine,lK,

Then, V becomes a Liapunov function.

The following proposition demonstrates that the uniformly asymptotically stability around p. = 0 under some conditions: Proposition 2 . Assume that (a) v, and w, are conthiuous, (b) vr, a,., K,,and K e are bounded, and (c) i , and w, are suf6ciently small. Under these conditions, p. = 0 is uniformly asymptotically stable over LO, -). Proof. By linearizing the differential Equation (9) around p. = 0
P. = A P.

(11)

Then, A (.) is continuously differentiable and is bounded. The characteristic equation forA is:
ugS3+u2s2+u,s+a,=O

(13)

where

e. =er -ec = W, - W,
Substituting vc and w, by v(p,, q,.) and v(p,, (I) respectively (cf. Equations (6) and (5)). we obtain the lemma. 0 Let us propose a specific instance of the control rule (5) for the target velocities as follows:

(13)

Since all coefficients ai are positive and U 1 u 2- uoa3> 0, the real parts of aU roots are negative through the Routh-Hurwitz Criterion. Therefore, by Corollary 41 on page 223 in [91, the Proposition was proved.

where K,,Ky and K e are positive constants. The first term in each velocity is a feedfonuard part. By Lemma I:
Lemma 2

4. E f f e c t 8 of Control Parameters

In the previous section, we demonstrated that the system is stable for any combination of parameter values of K,, Ky, and K e . However, since we need a non-oscillatory, but not too slow response of the robot, we have to 6nd an optimal parameter set. In order to simplify the analysis, we consider only situations in which the reference posture is moving on the x axis to t h e positive direction at a constant velocity V,:

386

This condition is called the linear reference morion. lo addition, we

le,l a1 and l6,l a1

(15)

< = 1.25

Fig.4
froqf By substituting Equation (11) by Equation (4),

Transient Respmes (Simulatiou)

P, = T ; (A T .
By

- ~,XP,- P,) + P,

(17)

Equatiou (4), Condition (14) a n d (19, le,( = le,( al. 16, I = 16, I 1. Therefore T . and Te in Equation (17) can be considemi as the identity matrix and null matrix respectively. Therefore,
Pc

= A (Pc - Pr) + Pr

By substituting the previous equation by Equation (4) and condition (14). we obtain Equation (16). U Equation (16) shows the behavior x, is independent of ys and 8, is this small pemrbation case. l/Kx corresponds to the time constant of the exponential decay.

t
Ay = lOcm

Corollary I

When x, = Ax at r = 0,
x, = v,r +AXe+

By cancelling 8, in Equation(16).
- _ _

without limiter

I1

y,
Where,

+ 2 < 5 y, + 5 2 = 0

iI

r , , I , ,
0
Fig. 5

- with
!

limiter
, I

100

sc (cm)

Lateral Discontinuity

Corollary 2 The condition for critical damping is Kg2 = 4 Ky

Corollary 3 In critical damping and if y, =Ay and 0 , = 0a t r = 0,

(21)

ters are v, = 3Ocm/sec, Ky = 6.4 x IV3/cm, 5 = 2.4/sec, and K, = IOlsec. A0 over dumping case = 0.75 and Kg = 0.12/cm), a critical dumping case ( < = I and Ko=0.16/cm), and an oscillatory case. (<= 1.25 and Ke = 0.20/cm) a~ used. With the small Ay perturbation of 5an, the result of simulation and analysis matches. In the fdowing experiments and in our real implementation, we adopt the c r i t i c a l damping condition, (= 1, since the convergence is fastestunder non-oscillatory coeditioa

(c

5. Velocity/Acceleration Limiting

In this motion, the m r ratio of yJAy i s reducedto 9.2% when x, becomes

Simulation results on three distinct convergence characteristics are shown in Figure 4. Here, the robots p, and pc were moving on the x axis to the positive direction, when y, suddenly jumps up with Ay = 5cm while continuing a parallel horizontal reference motion. The conimon parame-

For t h i s tracking controller system, reference paths designated by p,(t) and e(!) should satisfy the following conditions for smoothness (Preposition 2); (a) the path itself is continuous, (d) the path has tangent direction continuity, (c) the path curvature is continuous,(d) the derivative G , is bounded, and (e) the derivative is bounded (and hence, the derivas curvature continuity requirement tive of curvatwe is also bounded). % (c) is the reason why clorltoid curves, cubic spirals, and polar polynomials

&

have been developed for vehicle path planning [5][13][14]. From a vehicle navigators viewpoint, however, it is convenient if non-smooth paths am allowed to use. Although a path consisting of a line segment and a chcular an: does not possess curvature continuity, that kind of paths are widely used [71[151[161[17]. In the MML language on the Yamabico-11 mobile robot, a function called set-current(&p) is provided to compensate the robots positional error dynamically, and hence, it is frequently used in real-time navigation experiments [19]. However, if we allow these non-smooth paths, (i) either or both of the target velocities (v, w) by Equation (8) might become too large to be attained by a real vehicle, and (ii) the linear/rotational acceleration might become too large causing the robots slippage (Any slippage is a cause of a severe error in deadreckoning). Therefore, in order to handle those non-smooth reference paths, we need some limiter for velocities and accelerations. We adopt a simple ^ ^ algorithm of limiting the target velocitie: :( w) by constants (v, w) and the target accelerations (a, a)by constants (a, a), where a = V is a h e a r target acceleration and a = w a rotational wrget acceleration. This modification is implemented in the box Tin Figure 3. Figure 5 shows simulation results for various values of Ays with and without the velocity/acceleration limiter. Notice that the responses with a limiter are slower than that without a limiter. Hereafter, all simulations are done using the Critical damping parameter set. Figure 6 shows simulation results for A0 discontinuous jumps without limitation (AB = x/4, x I 2 and 3x14). Figure 7 shows simulation results for A0 discontinuous jumps with velocity/acceleration limitation. 6. Implementation
The results presented in Sections 3 , 4 and 5 were hardware independent. In this Section, we will describe how the theory was implemented on the robot Yamabico-11.

i
A ^

+) & ! 2

< 65cm/sec,

where W is the m a d (52.4cm). Through t h i s relation, we chose the maxb u m velocities as ( v , 0) = (4Ocm/sec, 0.8rdsec). We detemline the values of and by experiments with which the robot never slips: (a, a)= (50cm/sec2,5rad/sec).
1 . -

6.3. Experimental Results We conducted a few experiments to make sure that these values of Ky and K O are reasonable. Figure 8 shows experimental results with three distinct values of 5, which corresponds to Figure 4. Figure 9 shows results on Ay, which corresponds to Figure 5 . Figure 10 shows results on AB, which corresponds to Figure 7. (As shown here, the results on the real vehicle are close to that of simulation with a velocity/acceleration limiter.) In Figures 8-10, the trajectories are plotted using the current posture pc which is obtained by the vehicles dead reckoning.

6.1. Determining Control Parameters

Fig 6

Directional Discontinuity (Shnulation without Limiter)

1
I

the sampling time of the robots hardware. With a larger K x , the control system tends to be oscillatofy and instable even in its stop state (where p, = constant). An oscillation is observed at K, = 30/sec, when t h e constant IIK, = 33ms is compatible to the robots sampling time T, = IOms. Balancing these factors, K, = 10Isec was chosen. Adopting the critical damping condition (6 = 1 ) in Section 4, we also to determine a value of 6 for appropriate response of current posture pc. A larger 5 makes convergence faster. However, a too large 5 demands the robot an excessive rotational velocity. We decided to adjust lhe parameters so that the robot will reduce the error ycIAy into 9.2% during a 50cni run after a small perturbation of Ay. Therefore, by Corollaries 2 and 3, value K, = 6 . 4 ~ 1 0 - ~ / tand n ~K ~ e = 0.16/cm are detemiined. In this case, the time constant I / \ = 112.4sec is sufficiently larger than the sampling time Ts = I0nts. With these K,, K,. aid K e , no oscillations were seen. The emors .re and ye at constant reference velocity of 30cnilsec are about 2mm a d less than l r m respectively.
6.2. Determining Maximum VelwitylAcceleration

> - 0 ~-

. -

100

-50

xc

(cm)

Fig. 7

Directional Discontinuity (Simulation with Limiter)

The maximum linear velocity of the Yamabico-11 is known as 65cm/sec. We must consider the condition that even when the robot runs at and rotates at w at the same time, the velocity of the outer wheel should not exceed that maximum velocity, 65cm/sec. Thus,

-~

iii

cl

-15

xc ( c m )

388

Acknowledgement

4 = 0.75

' cc = = 1.25
0
t

Tk a u t h m thank Dr. Shin'ichi Yuta, Michiyuki Shindo aud Teijiro Kajiwara for the design and construction of Yamabico-11, and Amir Nilipour and Tony Lelm for their design and implementation of the earlier version of the Yamabico software system including the tracking controller. The authors also are thankfid to Dr. Daaiel Koditschek for his helpful comments on the stability theory.

Refe"e8
[I] C. M. Wang, "Locatim Estimation a i d Unceflaicity Analysis for Mobjle Robots", Proc. IEEEIntemational Conference on Robotics and Automation, pp. 12341235,1988.

yc
Fig. 9
Laleral Discontinuity (Experiments)

eo

1
II

Ay =20cm Ay = lOcm

1 i
I

-0 I

[Z] T. Tsumura. N. Fbjiwam. T. Shirakawa and M. Hashinloto, "An Experimental Systeni for AutomaticGuidance of Robot Vehicle. following the route stored in Memory," Proceeding 11tIi Intemational Symposium on Industrial Robots, October 1981, pp. 187-193. [3] Y. Kannyama and S.Yuta. "VehiclePath Speci6cation by a Sequenceof Straight f Robotics and Autonialion,vol. 4, no. 3, pp. 265-276, Lines". IEEEJoumal o 1988. [4] J. Crowley. "Asynchronous ConIml of Orientation and Displacement in a Robotic Vehicle", Pmc. IEEE Cdcrencc on Robotics slid Aulomalion. pp. 12n-1282,1989. [5] Y. Kanayama and N. Miynke. 'Trajectory Generntim for Mobile Robots", Robotics RcUCWh, vol. 3, pp. 333-340, The MIT P ~ ~ E 1986. s, [6] S. Singh and D. H.Shin, "Position Based Path Tracking for Wllceled Mohilc Robots". Proc. IEEEIntemational Wodrshop on Intelligent Robots and Systuns, in Tsukuba, Japan, pp. 386-391, September 1989. fl] Y. b a y m a . A. Nilipour and C. Lelm, "A Locomotion Control Method for Autonomous Vehicles", Y m .IEBE Conferenceon Robolics slid Autanation, pp. 1315-1317.1988. [8] W. Nelson and L Cox. "Local Path Control for an Autonomous Vehicles'', P m . lEEE Conferenceon Robotics and Automation,pp. 1504-1510,1988. [9] M. Vidyasagar, "Nonlinear Systems Analysis", Prentice-Hall Inc. Englewood Cliffs, N.J.. 1978. [IO] F. Miyozaki, Y Masutani. C. Lclm and Y. Kanayama, "Precise Trajectory Following Contml for Autaiomous Vehicles", Proc. Annual Cod. of kislitute of Systems, Control and Infomation Engineers,in Kyoto. Japan. May 1989. [I21 D. Koditshck, "Application of a New Lyapunov Function: Global Adaptive Inverse Dynamics for a Single Rigid Body", Repot of Center for Systems Science - 8806. Yale University, 1988. [I31 Y. Kanayama and B. Hartnian. "Smooth Local Path Planning for Autonomous Vehicles", Proc. IEEE Intemational Conferenceon Robotics aid Autanation, pp. 1265-1270.1989.
[ 141

100

W.Nelson, "Continuous-CurvaturePaths for Autonomous Vehicles", Proc.

t
-io

c
'

Fig. 10

'

-100

-50

'

Directional Disconliauity (Experiments)

1989. [I51 T. Lozano-Perez. and M. A. Wesley. "An Algorithm for Planning CollisionFree Paths Among Polyhedral Obstacles," Co". ACM, vol. 22. No. 10. pp. 560-570. October 1979. [I61 K. Konioriyo, S. Tachi and K. Tanie, "A Method for Autonomous Locomotion of Mobile Robots." Joumalof Robotics Society of Japan, vol. 2, pp. 222-231, 1984. (171 T. Hongo and H. W a w a , G. Suginloto, K. Tange, Y. Yamnmoto, "An Automatic Guidance System of a Self-Controlled Vehicle - The Command Systeni and the Control Algorithm -," Proc. ECON. 1985. [I81 Y. Kanaynma and T. Noguchi. "Locomotion Functions for a Mobile Rohot Language", Proc. Intemational Workshop on Advanced Robots and Intelligent Systems, pp. 542-549, Septeniber 1989.
1191 B. Hiutman and Y. Kanayomn, "Model and Sensor Rased Precise Navigationby an Aulonumous Mobile Robot". Proc. ICAR, 1989.

,
-

L[-~-.

10 xc ( c m )
I

389