IEEE/RSJ International Workshop on Intelligent

Robots and Systems lROS 91,Nov. 3-5. 1991,

Osaka, Japan. IEEE Cat. No. 91TH0375d

A Method of Indoor Mobile Robot Navigation by Using Fuzzy Control

Shigeki ISHIKAWA (Regular Member)

IBM Research, Tokyo Research Laboratory

5-19 Samban-cho, Chiyoda-ku, Tokyo 102, Japan

Abstract environment. One solution is to use an expert system that holds many
In the design of an autonomous mobile robot (AMR), it is necessary kinds of navigation procedures in a database. They can be classified
to describe schemes of monitoring t h e status of t h e robot and suitable into two categories: top-down-types and bottom-up-types. The top-
procedures of handling various situations. This paper presents a down approaches analyze the abilities of the robot actions and arrange
sensor-based navigation method using fuzzy control, whose purpose them hierarchi~ally(~.~.~,~).For example, Crowley et a1.o created a
is to construct expert knowledge for efficient and better piloting of the surveillance robot, and Burks et a1.W dealt with a maintenance robot
AMAs. This method provides a function for tracing a planned path by for a contaminated nuclear plant. Bottom-up type approaches attempt
sensing the distance of an AMR from its planned path and the differ- to construct control methodologies in order to achieve primitive types
ence between its angle and that of the planned path, and another of behavior in the robot. Mori et al.@), have tried to construct primitive
function that allows an AMR to avoid stationary and moving obstacles behavior by using a kind of robot language. Brooks and his group@)
by sensing how far an open area extends ahead of It. Fuzzy control tried to implement such behavior by using state transition machines.
is also used to select suitable rules for tracing a path or avoiding However, this is not an easy way to achieve complex tasks, because
obstacles according to t h e situation, which is determined from sensor the complex networks of the state transition could prevent the control
information by using fuzzy control. The effectiveness of t h e estab- procedures of the robot from being described.
lished rules and the effect of fuzzy control on AMR navigation are There are two fundamental dificulties in constructing such control
discussed on t h e basis of simulations. expert knowledge for an intelligent mobile robot. The first is the pre-
viously mentioned limitation of sensory processing ability. The second
is that environments always change, so it is difficult to ascertain which
1. Introduction information is important for the robot. Some time sequence changes
could contain very important information for monitoring the environ-
In order to realize autonomous navigation of a mobile robot, it is ment and robot states. Furthermore, we need to clarify what kinds of
necessary to describe schemes for monitoring the status of a mobile procedure are required in order to achieve a certain task, and then to
robot and its environment, and to present suitable procedures for han- describe suitable procedures according to the states of the robot or its
dling various situations. In the field of mobile robots, much attention environment.
has been devoted to obstacle avoidance problems, which can be clas- Recently, some methods of mobile robot navigation for indoor en-
sified into three types: (1) those in which the world model is precisely vironments have been rep0rted(lo~l~.*~~3). An interesting point of a
known, (2) those in which the environmental information is extracted fuzzy control system is that it is easy to implement a system that deal
from sensors and the world model is not known precisely, and (3) those with many s ituations without defining an analytical control model, by
in which the environment in unknown. In the first case (I), environ- representing relations between inputs and outputs in an if-then manner
mental information can be obtained so precisely, that it is easy to and constructing a knowledge base. In navigation control of the mobile
define a problem analytically; examples are the configuration space, robots, fuzzy control methodology could overcome the difficulty of
and potential field approach(l.J). These approaches are effective for defining a simple control model and the necessity for a system to be
off-line types of obstacle avoidance problem (such as those involving flexible enough to handle many situations. However, the fuzzy system
manipulators and path planning). However, they are not so effective so far reported for navigation of mobile robots define an ideal inputs
for navigation and piloting of mobile robots, because, in many cases, and output model. In(1o),a mobile robot and a moving obstacle are
not all the environmental information is known precisely, and the avail- represented as points, and the distance between them is supposed to
able sensory data are limited. Situation (2) is natural for mobile robot be sensed ideally. In this approach, the number of obstacles is limited
navigation, in which the robot must monitor the environment and de- to one. This severely limits its applicability in a real environment.
termine its moving status from sensory data. The present paper deal Takeuchi(ll) presents a hallway following method using fuzzy control.
with this case. For situation (3), many people have tried to generate In this approach, the vehicle selects the widest free space extracted by
an environmental map by exploring some area in which the robot can simple image processing of the hallway, and moves ahead. In(12),the
move and sensing environments in various ways. The difficulties in vehicle follows a wall of the hallway, turns right or left when it comes
situation (2) are: to a comer, and moves into a garage space by using ultrasonic sensors.
Both of the last two studies used fuzzy controls to implement the macro
0 When a mobile robot encounters complex situations, such as when actions. The environments of the mobile robots are supposed to be
it must avoid not only stationary obstacle but also moving ones, it completely stationary, and the authors did not consider the problem
should behave accordingly. of avoiding moving obstacles.
This paper presents a sensor-based navigation method using fuzzy
0 Sensory data does not actually give very useful and precise informa- control in an indoor environment that includes stationary and moving
tion on the environment. obstacles. The purpose of using fuzzy control here is to construct an
efticient control knowledge base for autonomous mobile robot naviga-
It is necessary for a mobile robot to judge various situations exactly tion. This method provides a function for tracing a planned path by
and to select a suitable method of handling them. However, there is sensing the distance of an AMR fromits planned path and the difference
no simple method of dealing with a complex and dynamically changing between its angle and that of the planned path, and another function

-/ 0 / 3 -
for avoiding stationary and moving obstacles by sensing the object-free process separately, and then explain how to combine them. The nav-
distances from an AMR. Fuzzy control is also used to select suitable igation overall structure is shown in Figure 3.
rules for tracing a path or avoiding obstacles according to the situation,
which is determined from sensor information by using fuzzy control. 3.1. Tracinp a planned path
This paper discusses the effectiveness of established rules and the effect A planned path is assumed to be transferred as a list of coordinates
of fuzzy control on AMR navigation on the basis of simulations. This of reference points. It is easy to trace the path by using conventional
paper also reports the development of a simulator system to assist in numerical control. However, in t h s paper, a fuzzy control method of
construction of the rule base, because one of the disadvantages of using tracing a path based on(13) is used. Input parameters for tracing a path
fuzzy control is that it involves the acquisition of a rule base that are the distance 'D' between the robot position and the path, and the
consists of many if-then rules. difference angle 'A' between the vehicle direction and the path direction,
as shown in Figure 4(a). The output parameters are the steering angle
2. Hvpothetical environment of an indoor mobile robot 'S' and the speed ' V of the robot. In fuzzy control, parameters are
categorized into seven memberships: PE, PM, PS, ZZ, NS, NM, and
A hypothetical indoor environment of a mobile robot is shown in NB (eg. NE means negative big). The membership functionss used here
Figure 1, including block walls and stationary and moving obstacles. are shown in Figure 5(a), (b), (e) and (0, respectively. As for input
The problem to be solved here is how to navigate the robot along with parameter 'D' and 'A', negative value implies right side and positive
a planned path, avoiding obstacles on the basis of sensor information. value implies left side.
Figure 2 shows the entire system structure of an mobile robot. Path
planning is processed by a path planner before the robot moves. A
navigator pilots the robot along the planned path to a destination,
avoiding obstacles if necessary. We assume that there are two kinds
I Rules for Speed - V ' Output I I Rules for 'Steer - S' Output 1
of sensory data: one for positioning the robot, and the other for mon-
itoring the environment in order to avoid obstacles. In this paper,
sensors for the former are assumed to be wheel encoders and those for
the later are assumed to be range sensors (such as sonars or infrared
range sensors). These sensors are not at all sophisticated, but are re-
alizable in practical use.

3. Indoor navigation using fuzzy control for a mobile

-
robot

There are two kinds of navigation for a mobile robot: (1) navigating
to a destination, and (2) avoiding obstacles. This paper describes each
Table 1: Rules for tracing a planned path

Block

\
\
j

- Pilotting Description
- Updating Model
I R o t s t i o n / S t e e r i n 8 An8 l e

sec

Figure 1: Movable environment of an AMR Figure 2: AMR system structure

 A Planned Path

t
D,A
yracing a planned
*
path

* Path side

a Path
1 -K
II I I I
W1-5, Dwl-5 -?vodinig obstacle ( a ) lnput p a r a a e t e r s f o r t r a c i n g a P a t h

t DWJ
Figure 3; Overall structure of Navigator

The developed fuzzy rule base for tracing a path is shown in Table
1. There are two rule bases: one for steering and one for speed. Each
has 49 rules. These rule bases specify, for example, that if the robot
has missed a path, it should slow down and then steer in the correct
direction.

3.2. Fuzzv control for avoiahg obstricles

To avoid obstacles in front of the vehicle, we defined twelve input
parametes. Those are the obstacle-free distances (Wl-W5) in five di-
rections, their time differences (Dwl-Dw5) as shown in Figure 4(b). ,
and STATE/SIDE parameters, which show the robot btatuses of path ( b ) lnput parameters
- for a v o i d i n g obstacles
tracing.
T h e object-free distances are very important to know the degree of
danger of the robot. On the real mobile robot system, such infomation Figure 4: Input parameters of fuzzy control
can be acquired actually by using active range sensors, such as sonars
and infrared PSD range sensors. The object-free distances are catego- ating a proper control instruction. The developed rule base for obstacle
rized into four memberships: PB, PM,PS, and ZZ (cg. ZZ means the avoidance navigation consists of 155 rules. These rules are classified
distance between the robot and some object is almost zero). Figure into following the categories:
5(c). The time differences are categorized into seven memberships, as
shown in Figure 5(d). 0 Rules for cases in which any obstacle ahead of the mobile robot is
STATE/SIDE parameters are evaluated by comparing the robot's nearly touched.
position and the current path of the robot. STATE has three values:
0 Rules for dead alleys
0 If STATE is 1, it shows that the robot runs away from the path,
0 Rules for steering in an obstacle-free direction
0 If STATE is 0, it shows that the robot runs on the path.
0 Rules for tracing a narrow crank hallway
0 If STATE is -1, it shows that the robot runs toward the path.
0 Rules for moving straight along a right/left wall (on the planned path)
As for SIDE, it also has three values:
0 Rules for moving straight between walls
0 If SIDE is 1, it shows that the robot runs right side of the path,
0 Rules for cases in which any obstacle exists ahead of the robot when
If SIDE is 0, it shows that the robot runs almost on the path. it is moving along a right/left wall
0 If SIDE is -1, i; shows that the robot runs left sind of the path. 0 Rules for moving nearby obstacles (off the planned path)
The output parameters of the robot contorl are the same as those for 0 Rules for following a rightjeft object (off the planned path)
tracing a path.
In order to avoid collision with obstacles, the system needs to judge
the direction in which the robot should move and to generate proper 3.3. Combininp inferences for tracina a path and avoiding
instructions to low-level drivers, by considering the obstacle-free dis- obstacles
tances and their time differences. To judge a correct situation, the The fuzzy inferences mentioned above are processed separately. It is
system should take all account of the parameters. F u u y control pro- necessary to combine them properly according to the situation, because,
vides a useful mechanism for judging a situation correctly and gener- for example, the robot could not avoid obstacles on the planned path

- /0/5-
 safely by combining 50% of each of the outputs, if the rule base for
tracing a planned path generated instructions to follow the path. T o
solve this problem, another fuzzy rule base inferences which rule base
should be weighted. The final control outputs for steering 'S' and ve-
locity 'V' are the weighted sums of both outputs by equation below:
s = S f * K + sa * (1.0 - K )
V = V f * K + Va*(l.O-K)
-3O.Oe. 30.OE. where S I is the steering angle output for tracing a path and Sa is that
l a 1 D : D i r l i s c e b e l v e e n I p l a n n e d ~ a l ha n d A U K
for avoiding obstacles, Vt is the velocity output for tracing a path, and
Va is that for avoiding obstacles, and K is the weighted coefficient.
The rule base has the same input parameters and conditions as the
NU HH NS dZ PS PI PU
rule base for avoiding obstacles. That is, if any dangerous situation
occurs, the weight coefficient 'K' will be getting larger. The f u u y values
for 'K' takes ZZjPS/PM/PB.

4. Acsuisition of navigation knowledge for f u z n control

III A : Direction Dillrrence bslreen a plroned path a n d A U K
Generally, the procedures for constructing a fuzzy rule base are as
follows:

1. Define the input and output parameters of the control system and
membership functions.

2. Describe rules in an if-then manner by predicting the situations of

the target system.

3. Then tune the rule base and/or membership functions on a simu-

lator and/or the target system.

We developed a simulation system to assist in constructing a rule base,

as shown in Figure 6. This simulator is based on the mobile robot
system shown in Figure 2, by which a motor driver and sensor proces-
sors are replaced. A planned path is passed to a simulator with graphical
data. The simulator also holds a world model represented in graphical
data, which is edited by a graphic editor. We also developed a f u u y
production language and a fuzzy rule compiler. A fuzzy rule base is a
normal text file, and is translated into a rule basz object by the compiler.
A fuzzy system inferences the fuzzy logic according to a rule base
object. The propriety of the rule base is evaluated by simulation of
mobile robot navigation. The rule is then tuned manually.

5. Simulation Experiments
The experimental results are shown in Figures 6 to 9. The interval
of control is set to 0.1 msec. The results show the trajectories every 1
sec. The maximum speed of the robot is assumed to be 30 cm/sec.
Figure 7 shows a companson between the rule base for avoiding
obstacles in a novice version and the rule base in an advanced version.
The former does not describe the rules for following a right/left object.
This resulted in the robot failing to avoid an obstacle, because the
system generated instructions for the robot to resume tracing the
planned path after the rules were found to be insuffcient for the situ-
ation. With the advanced version, the robot avoided the obstacle
smoothly, and then resumed tracing the planned path. This behavior
was achieved by combining the two rule bases with an inferred weight.
Figure 8 shows an example in which a robot moved among stationary
obstacles. In this case, the rule base was enhanced and rules added for
moving through a narrow crank course. When the robot passed the
crank course, it returned to a planned path.
Figure 9 shows almost the same example as in Figure 8 except that
a moving obstacle was encountered after the last stationary obstacle.
Because the robot moved faster than the moving obstacle, it passed
that obstacle.
The final experimental result shown in Figure IO, shows how a robot
0.0 1.0 avoided collisions with two moving obstacles, the first moving in the
IL) 11 Y.isblio1 Cc.eIriciev.1 same direction as the robot, but more slowly, and the second moving
Figure 5: Membership functions in the same way as in Figure 9.
 Block

Planned Path
I
G r a p h i c datal
1
I

I
YotorfSteerinr Common

I
I

Figure 6 Navigation rule acquisition system vanced rules

6: Conclusion

This paper has presented a sensor-based method of navigating a

mobile robot in an indoor environment that includes stationary and
moving obstacles. The method uses fuzzy control to judge the status Block
of the robot and the environment and to generate proper instructions.
The rule base that we have developed needs to be improved to handle
all situations; however, we confirmed the effectivenessof this approach
in a hypothetical environment and the expandability of the rule base.
Two types of rule are combined to trace path behavior and avoid
collisions with obstacles. Problems to be solved in the next step are as
follows:

0 There are ten input parameters for the avoiding task. This makes it
hard work to develop a rule base, because the number of rules and
iterations for tuning rules must be expanded. We must improve the
way of developing rule bases.

0 If opposite types of rules coexist in the same rule base, it may cause
faulty navigation leading to collision with obstacles. To avoid this,
we should separate a rule base into sub-rule bases and control them
from outside the rule base. This is an expansion of the approach
mentioned in Sec. 4.3. The important things tasks are to learn how
to separate the rule base (automatically, if possible), and how to
construct new rule bases structurally.

As the rule base is being created, a new simple control model may
happen to be extracted from it. It is not necessary to apply fuzzy con-
trol to all navigation processing. The important objective here is to
acquire efficient knowledge for mobile robot navigation. We are cur-
rently focusing on how to extract an efficient structure of knowledge,
automatically if possible. We are also testing the effectiveness of this
_approach by using a prototype mobile robot. Figure 8: Simulation result 2: avoiding stationary obstades
 Bibliography

1. 0. Khatib, Real-Time Obstacle Avoidance for Manipulators and

Mobile Robots, International Journal of Robotics Research, Vol. 5,
Block NO. 1, pp. 90-98, 1986.
2. E. Freund and H. Hoyer, Real Time Path-Finding in Multi-Robot
Systems Including Obstacle Avoidance, International Journal of
Robotics Research, Vol. 7, No. 1, pp. 42-70, Feb 1988.
3. K. Fujimura and H. Samet, Path Planning among Moving Obsta-
cles Using Spatial Indexing, Proc. IEEE Int. Conf. on Robotics and
Automation, pp. 1662-1667, 1988.
4. J. L. Crowley, Coordination of Action and Perception in a Sur-
veillance Robot, IEEE Expert, pp. 32-13, Winter 1987.
5 . B.L. Burks, G. de Saussure, C.R. Weisbin, et al., Autonomous
Navigation, Exploration, and Recognition Using the HERMIES-IIB
Robot, IEEE Expert, Winter 1987, pp. 18-27.
6 . C. Isik and A. Meystel, Pilot Level of a Hierarchical Controller
for an Unmanned Mobile Robot, IEEE Journal of Robotics and
Automation, Vol. 4, No. 3, pp. 241-255, Jun 1988.
7. D. Kriegman, E. Triendl andT. Binford, A Mobile Robot: Sensing,
Planning and Locomotion, Proc. IEEE Int. Conf. on Robotics and
Automation, pp. 402-408, 1987.
8. H. Mori, Y. Chino, K. Saitoh et al., Mobile Robot Harunobu-4
for Outdoor Environment - Stereotyped Motion by Sign Pattern,
Proc. IEEE Int. Workshop on Intelligent Robots and Systems, pp.
633-638, Oct 1988.
9. R. A. Brooks, A Hardware Retargetable Distributed Layered
Architecture for Mobile Robot Control, Proc. IEEE International
Conference on Robotics and Automation, pp. 106-110, 1987.
10. Yoichiro Maeda and Morikazu Takegami, Collision Avoidance
Conrrol among Moving Obstacles for a Mobile Robot on the Fuzzy
Reasoning, Journal of the Robotic Society of Japan, Vol. 6, NO. 6,
pp. 518-522, Dec. 1988.
1 I , Tomoyoshi Takeuchi, An Autonomous Fuzzy Mobile Robot,
Journal of the Robotic Society of Japan, Vol. 6, No. 6, pp. 549 - 556.
Dec. 1988.
12. Toshiaki Murofushi and Michio Sugeno, Fuzzy Control of a
Model Car, Journal of the Robotic Society of Japan, Vol. 6, No. 6,
pp. 536-541, Dec. 1988.
Block 13. Michio Sugeno, Fuzzy Control, pp. 84-87, Nikkan Kougyo
Shinbun Ltd., Japan 1988.

Figure 10: Simulation result 4: avoiding moving obstacles

--/0/8
-