See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3298783

Sensory-based motion planning with global proofs

Article in IEEE Transactions on Robotics and Automation · January 1998

DOI: 10.1109/70.650160 · Source: IEEE Xplore

CITATIONS READS

203 57

2 authors:

Ishay Kamon Ehud Rivlin

Technion - Israel Institute of Technology Technion - Israel Institute of Technology
12 PUBLICATIONS 757 CITATIONS 248 PUBLICATIONS 9,895 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A comparison of Gaussian and mean curvature estimation methods View project

Mutual Awareness Activity Analysis View project

All content following this page was uploaded by Ishay Kamon on 14 February 2015.

The user has requested enhancement of the downloaded file.

 Sensory Based Motion Planning with Global Proofs
Ishay Kamon Ehud Rivlin

Computer Science Departme-nt Computer Science Department

Technion, Technological Institute of Israel Technion, Technological Institute of Israel
Haifa 32000, Israel Haifa 32000, Israel

Abstract robot can adjust its operation in changing and dy-

A sensory based algorithm DistBug, that is guar- namic environments.
anteed t o reach the target in an unknown environment Many reactive algorithms for sensory-based path
or report that the target is unreachable, is presented. planning are based on the idea of “navigation func-
The algorithm is reactive in the sense that it relies tion” that maps sensor readings to actions. Different
on range d a t a t o make local decisions, and does not methods are used to choose or to learn this mapping,
create a world model. The algorithm consists of two including potential fields and its variations and fuzzy
behaviors (modes of motion): straight motion between logic (see for example [5],[1],[11],[4]). Local poten-
obstacles and obstacle boundary following. Simulation tial field approaches usually suffer from local minima
results as well as experiments with a real robot are pre- problems. Another problem of the navigation function
sented. approaches is falling into a loop. The sensory infor-
The condition for leaving obstacle boundary is based mation is inherently local, and the same action will
on the free range in the direction to the target. This be always taken for the same local situation, without
condition allows the robot to leave the obstacle as any global considerations. Therefore once trapped in
soon as the local conditions guarantee global conver- a loop, there is no way t o get out of it.
gence. Range data is utilized for choosing the turn- A different approach, which guarantees reaching
ing direction when the robot approaches an obstacle. the target if possible, was presented by Lumelsky and
A criterion for reversing the boundary following di- Stepanov [7]. The robot, equipped only with contact
rection when it seems t o be the wrong direction is sensors, goes straight to the target until hitting an
also introduced. As a direct result of these local de- obstacle. It then follows the object boundary. The
cisions, a significant improvement in the performance algorithms define “leaving conditions” that determine
was achieved. where to leave the obstacle boundary and go directly
towards the target again. The convergence is based
on the fact that the distance from every leave point
11 Introduction t o the target is strictly smaller than the distance from
Finding a path from a given location t o a given des- the corresponding hit point. Algorithms that use only
tination is a fundamental problem in mobile robots the direction t o the target and abandon obstacles as
research. Traditional methods assume perfect knowl- soon as possible were presented in [lo] and [8]. Range
edge about the environment. In a more realistic set- sensing was used in [6] t o find “shortcuts” t o the path
ting the robot should rely on its sensors in order t o that would be planned using contact sensors.
perceive its environment and plan accordinp;ly. A pos- We will present a new algorithm DistBug that uses
sible approach is t o use the sensors to create a model of range information more effectively. The main contri-
the environment (world reconstruction). Traditional bution is a new “leaving condition” that allows the
path planning methods could then be applied to the robot t o abandon obstacles as soon as global conver-
world model. Reconstructing a world model based on gence is guaranteed, based on the free range in the di-
sensory information is not an easy task because of the rection to the target. Range d a t a is utilized for choos-
inherent uncertainty of the data. Various methods ing the turning direction when the robot approaches
were DroDosed for data/sensor fusion (see for e x a m d e an obstacle. A criterion for reversing the boundary
PI, [3j>[ill. following direction when it seems t o be the wrong di-
Sensorv-based algorithms use local sensory infor- rection is also introduced. As a direct result of these
rnation foi feedbacklontrol of the robot motion. This local decisions, a significant improvement in the per-
information enables the robot t o operate in an un- formance was achieved.
known environment. At every time instance, the robot We consider a point robot that moves among arbi-
uses its sensors t o locate close obstacles, and then trary obstacles in a planar environment. We assume
plans the next part of its path. Reactive strategies that the workspace is bounded. It follows directly that
emphasize the direct and immediate relations between the perimeter of any obstacle is finite, and that the
perception (input) and action (output), and does not number of obstacles is finite.
rely on world models. Using reactive navigation a The rest of the paper is organized as follows: Sec-

435
0-8186-7108-4/95 $4.00 0 1995 IEEE
tion 2 presents the algorithm DistBug and gives a
convergence proof for it. In Section 3 the experimen-
tal setting and results are presented. Finally, t h e con-
clusions are presented in Section 4.
2 DistBug algorithm
The algorithm DistBug consists of two modes of
motion: moving directly towards the target between
obstacles and following obstacle boundaries. T h e
robot goes straight t o the target until hitting an ob-
stacle. T h e Obstacle is then followed until a condition
for leaving the obstacle boundary holds. The leaving
condition, described in section 2.1, guarantees the con-
vergence of the algorithm. T h e mechanism for choos-
ing boundary following direction, described in section
2 . 2 , influences the path length. T h e algorithm is as Figure 1: An example of a path planned using the
follows: distance-based leaving condition (dashed line), compared
with VisBug21 algorithm from [5] (dotted line). In point
1. Go directly t o the target, until one of the follow- L1 we have ( d i s t ( L 1 , T ) - Freed,,t < B e s i d e s t ) and the
ing occurs: robot goes straight to T (note t h e “radius of vision” R).

1
a T h e target is reached. Stop.
b An obstacle is reached. Go t o step 2 .
However VisBug21 algorithm would drive the robot near
the obstacle boundary until the point V from which the
line (S,T) is “visible”.
2. Choose the boundary following direction. Follow
the obstacle boundary untll one of the following
occurs: F T e e d i s t 5 0) captures t h e special case where the tar-
a T h e target is reached. Stop.
get is within t h e sensed freespace, and can be reached
b] T h e free range in the direction to the target directly. It is necessary for scenarios where the target
guarantees t h a t the next hit point will be closer is close t o the hit point (Hitd;st < Step), thus forc-
t o the target than the last hit point.Go t o step 1. ing the minimal improvement Step would prevent the
c) T h e robot completed a loop around the obsta-
robot from leaving the obstacle.
cle. T h e target is unreachable. Stop.
T h e condition (Currdi,t - Freedist 5 BeStdi,t)
guarantees t h a t whenever the robot leaves an ob-
2.1 Leaving boundaries as soon as conver- stacle and hits another obstacle, the distance from
gence is guaranteed its next hit point t o the target NeztHitdi,t will
Motivated by the fact t h a t straight motion is faster be strictly smaller than the distance from the last
and safer than boundary following, the condition for hit point Hitdist. Furthermore it guarantees t h a t
leaving an obstacle boundary is designed t o abandon (Hitdist - NextHitdiSt > Step) for every hit point,
t h e boundary as soon as convergence is guaranteed. where Step is fixed (note t h a t BeStdist is initialized
Range information in the direction t o the target is t o HitdiSt - Step). T h e up-to-date shortest distances
used t o guarantee t h a t the next hit point will be closer along the obstacle boundary Bestdist is considered in-
t o the target relative to the last hit point. stead of (Hitdist - Step) in order t o avoid undesir-
We will use the following notations and definitions: able behaviors t h a t may occur if t h e the distance from
CurrdiJt - distance from the current location t o the the last hit point was considered (see figure 2). T h e
target; Hitdi,, - distance from the last hit point H to
the target; Step - a constant t h a t will be the minimal
improvement in the distance t o t h e target between hit
robot hits on its way t o N = w,
parameter S t e p bounds t h e number of obstacles the
where dist(S,T )
is the distance from t h e starting point S t o the tar-
points. Bestdiat - t h e up-to-date shortest distance t o get T. T h e algorithm will terminate after at most N
the target since the last hit point. It is initialized
hit points. Given t h a t the perimeter of every obstacle
t o BeStdist + (HitdiJt - Step). Freedist - distance
is finite the path t h a t t h e robot will traverse will be
in freespace from the current ocation t o the nearest
finite.
obstacle in the direction t o the target. We consider a Increasing the sensor range R increases Freedist
limited sensor range R. If no obstacles are detected
we set Freedist c R; (when no close obstacles are present) and makes it pos-
sible to leave the obstacle boundary earlier. Therefore
We define the following leaving condition:
increasing the sensor range has a direct effect on the
If F T e e d i s t > 0) and improvement of the path (see figure 1). This behavior
[ ( c ‘ U ’ P T d i s t - F T e e d i s t 5 0) OF
differs from the algorithms presented in [GI where the
( C ‘ U ‘ P T d i s t - F r e e d i n t 5 Bestdi,t)) current obstacle would be left only if t h e line ( S , T )
then leave the obstacle. is within the “radius of vision”. DistBug uses range
The condition Freedis$ > 0) checks whether it is d a t a directly, in contrast t o modeling the local envi-
i,
possible t o leave t e obstacle boundary and move di-
rectly towards the target. T h e condition (Currdi,t -
ronment in [ G ] algorithms.
We will first present a convergence proof for the al-

436
 2. There is an obstacle on the straight line from
Closest to T . It is not the current obstacle 0
because Closest is closest to the target on the
boundary of 0 . In this case (Freedi,t 2 Minda,t).
The leaving condition (Currdi,t - FTeedist 5
Hitdi,t - Step) holds, because (Currdist _<
Hitdist) and (MindiSt 2 Step).

We showed that if the target is reachable then the

leaving conditions will enable the robot to leave the
obstacle. qed.
Proposition 1: If the target is reachable from the
Figure 2: A problematic behavior that could occur if the starting point then the robot will reach the target.
leaving condition was ( C u t T d r s t - F T e e d r s t 5 Hatdi,t -
Step): the robot leaves the obstacle in L l and goes towards
the target because ( d i s t ( H 1 , T ) - d i s t ( H 2 , T : ) > Step).
hit at most No obstacles, where No = w.
Proof: Leaving the starting point S the robot may

robot moves directly towards the target until reaching

The

it or hitting an obstacle. If an obstacle is hit, Lemma 1

guarantees that the robot will leave it and proceed to-
gorithm, in which the parameter Step will be set using
knowledge about the environment. When no knowl-
edge about the environment is available setting a too may hit at most Ni obstacles, where Ni =
<
The leaving condition (CUrrdist-Freedist Bestdist)
w.
wards the target. From every leave point Li the robot

big Step may prevent the robot from leaving obstacles,

and thus from reaching the target. We will present a guarantees that Ni 5 Ni-1- 1 for every i > 0. There-
modified version of the leaving condition that can be <
fore after k hit points (k No) there will be Nk = 0,
used in this scenario. and the target will be reachable directly from the leave
point Lk. qed.
Convergence proof. We assume that the Proposition 2: If the robot completes a loop around
minimal distance between obstacles is Mindist. an obstacle then the target is unreachable.
The value of Step is set accordingly: Step t Proof Follows directly from Lemma 1. If the tar-
minimum(Mindi,t , R ) , where R is the maximal sen- get is reachable then the robot will leave any obstacle
sor range. We will show that under these conditions, before completing a loop around it. Therefore if the
the robot will reach the target if possible, and halt robot completes a loop around an obstacle it means
otherwise. that the target is unreachable. qed.
Lemma 1: If the target is reachable from a hit point Proposition 3: The algorithm terminates.
H , then the leaving conditions will enable ithe robot to Proof: Given that the workspace is bounded, it is suf-
leave the obstacle boundary and proceed towards the ficient to show that the robot may hit only a bounded
target. number of obstacles on its way. The leaving condi-
Proof: Consider a hit point H , where the robot tions enable the robot to leave obstacle boundaries
reaches the obstacle 0. The robot can reach any only a limited number of times. The first condition
point on 0 ’ s boundary in “boundary following” mode. (CurrdiSt- Freedjst 5 0) can be used only once, be-
:If the target T is reachable from H , then there ex- cause after it is used the target will be reached directly,
ist at least one point on the boundary of 0 from and the algorithm will halt. The second condition
(CurrdiSt- FreediSt 5 BestdiSt) can be used only N
which the robot can leave the obstacle and proceed
towards the target. In particular we can look at a
point Closest on the boundary of 0 that is closest
times, where N =
+
w. After the robot hits the
N lthobstacle there are two possibilities: either the
to the target. When the robot reaches the point
Closest we have (Freedist > 0) because T is reach- target can be reached directly using the first condition,
able from Closest and the straight line [Closest,T ] or the robot will not be able to leave the obstacle. In
does not intersect 0. Currdist is the distance from the second case the robot will complete a loop around
Closest to T . If (Currdi,t 5 Hitdist - Step) the obstacle and halt. qed.
t,hen (BestdiSt = Currdi,t) and the leaving condi-
tion (Currdist - Freedi,t 5 Bestdi,t) holds. Other- Assuming no knowledge about the environ-
wise BeStdi.,t has its initialization value (Bestdi,t = ment. In the convergence proof described above the
H i t d i s t - Step). There are two possibilities for the parameter Step was set using some knowledge about
way from Closest to the target T : the environment. When no knowledge about the en-
vironment is available setting a too big Step may pre-
1. There are no obstacles between Closest and T . vent the robot from leaving obstacles, and thus from
If the target is within the sensor range R then reaching the target (note that BeStdi,t is initialized
the leaving condition (CUrrdist - Freedist 5 0 ) to Hitd;,t - Step). To overcome this problem we add
holds. If the target is not within the seinsor range a version of the leaving condition of Bug2 algorithm
then (Freed;,t +- E ) . The leavin condition from [7] to our distance based leaving condition, us-
(Currdi,t - Freedist 5 Hitdist - Step? holds, be- ing a boolean O R relation. In this way the robot can
cause (currdi,t <
Hitdirt) and ( R Step). > always leave an obstacle. Using the modified leaving

437
condition it is reasonable t o set the parameter Step are gathered as the robot approaches the obstacle,
for typical scenarios. For a real robot the improve- from a limited range of angles in front of the robot.
ment in the distance t o the target between hit point’s The largest readings from the left and right sides
will usually be larger t h a n the robot size. are st,ored in the variables L e f t , Right. The variable
We define CROSS (line crossing) as a boolean con- Dir is initialized t o zero, and updated in every cycle
dition that holds if t h e robot meets the straight line Dir = Dir + ( L e f t - Right). The value of Dir is
( H , T ) between t h e last hit point H and the target T ; accumulated up t o a constant. T h e turning direction
We define the following leaving ” condition: is defined by t h e sign of Dir. If Dir is positive then
If (Freedist > 0) and ” t,he turning direction is left (there is more freespace
(((CROSS holds) and (CurrdiSt < Hitdist))
\ \ \ , , OR
in t,he left side). Otherwise the turning direction is
(CaTTdi,t - Freedis; 5 Bestdist)) right. In this way information of several cycles is con-
then leave the obstacle. sidered, and several cycles are usually necessary t o re-
The condition C1 G ((CROSS holds) and verse the turning direction. Big differences between
(Currd;,t < Hitdist)) is equivalent t o the leav- range readings, indicating more global information,
ing condition of Bug2 algorithm. Our reference have a stzonger effect t h a n smaller differences t h a t are
straight line is drawn from the last hit, point H in- usually more local.
stead of the starting point S. The condition C2 E The initial boundary following direction is chosen

i Currdist - Freedist 5 Bestdis2) guarantees that

Hitdist - NextHitdist 2 S t e p ) for every hit point.
based on local information near the hit point, and may
be the “wrong” direction in the sense t h a t it will result

tivations of this leaving condition t o N = w.

The improvement size Step bounds the number of ac-

The convergence proof of the algorithm is as fol-

in a longer path. The following criterion attempts t o
reverse a wrong decision based on higher level of infor-
mation that is still local in nature (If t h e robot went
clockwise around t h e obstacle, it would turn back-
lows: (1) The condition C1 alone guarantees reaching wards and go counterclockwise, and vice versa). The
the target in a finite path if the target is reachable (t,he criterion for reversing the boundary following direc-
complete proof is presented in [7]). Considering any tion is trigerred when the robot moves in the opposite
leave point L; as a new start, convergence is guaran- direction t o the target (more precisely, when t h e robot
teed if only C1 will be used after L ; . (2) The condition heading momentarily points t o the opposite direction
C 2 can be activated at most N times during the path, t o the target). Changing the boundary following di-
thus defining a t most N leave points. Claim (1) im- rection can be done only once during the following of
plies t h a t after the last leave point defined by C 2 the the same obstacle, t o avoid local loops. Boundary fol-
algorithm will converge using the leaving condition C1 lowing direction should be reversed only if the robot
alone. did not go too far from the hit point. In this way
2.2 Choosing the boundary following di- t,he addition t o the path length due t o the reversed
rect ion following direction will be bounded.
The boundary following direction has a significant
effect on the path length, especially when the perime- 3 Experimental results
ter of the followed obstacle is long. It is desirable A Nomad200 mobile robot was used for testing. U1-
t h a t the robot would choose the boundary following trasonic, infrared and structured light sensors were
direction t h a t will result in a shorter path t o the tar- used for range sensing. There are 16 ultrasonic and
get. However, this requires global knowledge t h a t the infrared sensors, t h a t provide 360 degrees coverage
robot does not have. We address the problem of choos- around the robot. The structured light sensor gen-
ing the boundary following direction in two stages: the erates a horizontal plane of light in front of the robot,
initial direction is chosen based on local information and provides range readings using triangulation.
near the hit point. The boundary following direction Three simulated environments with decreasing
can be reversed if the robot concludes that it had cho- complexity were used t o test the algorithms (see figure
sen the wrong direction. 3). “World3” is a rough description of our laboratory.
Choosing the initial boundary following direction “World2” was constructed from “world3” by erasing
(clockwise/counterclockwise) based on local sensory parts of the obstacles. “Worldl” describes a simple
information is similar t o choosing the motion direction environment with a few separated obstacles. Nine
(left/right). A reasonable heuristic is t o choose the di- start/target points were chosen by hand for “world2”
rection where the robot can get closer t o the target. and “world3”. All the possible combinations between
If the decision is turn right, for example, the bound- those points were used t o generate 72 start/target
ary following direction will be counterclockwise. It is pairs for testing in each environment. One hundred
better t o make this decision before the robot reaches pairs were used for testing in “worldl” . The start and
the obstacle, because the robot “view area” shrinks as target location were chosen randomly: the a: coordi-
it gets closer t o the obstacle. The robot should decide nates were fixed (left and right sides of t h e environ-
while approaching an obstacle (moving in a straight ment) and the y coordinates were chosen randomly
line) which turning direction t o use when it reaches within a given range.
the obstacle. Three versions of DistBug algorithm are presented
Following a purposive approach we want t o choose in the following experiments. Choosing the initial
the turning direction without building a local model turning direction was used in version DzstBugl. The
of the environment. We use range readings, t h a t criterion for reversing the boundary following direc-

438
Figure 3: The simulated environments: left - “worldl”. Figure 5: Simulation results in “world2” environment.
The start/target points were chosen randomly from the Left - Bug2 algorithm. Middle - DistBug2, choosing the
vertical lines on both sides; middle - “world2”, the nine turning direction (path length is 0.94). Right - DistBug4,
start/target points are marked with +; right -- “world3”. using the leaving condition (path length is 0.61).

I algorithm tvDe
, v ..- II, , worldl I world2 II world3 I I

Dist Bug1 )I 0.86 1 0.81 11 0.96

Dist Bua2 II 0.86 I 0.80 It 0.52
I Dist Bug3 11 0.79 I 0.70 11 0.45
Figure 4: Simulation results in ‘‘worldl” environment Table 1: Performance of DistBug algorithms. Choos-
(target on the left side). Left - Bug2 algorithm. Middle ing the initial turning direction was used in DistBugl.
- DistBugl, choosing the turning direction (path length
The criterion for reversing boundary following direction
is 0.84 relative t o Bug2). Right - DistBug3, using the was added in DistBug2. The condition for leaving obsta-
leaving condition (path length is 0.67 relative to Bug2). cle boundary was added in DistBug3. The average path
length over 7 2 runs is presented relative t o B11g2 perfor-
mance.
tion was added t o the version DistBug2. T h e condi-
tion for leaving obstacle boundary was added t o the
version DistBug3. Examples of simulated runs are ting). (2) working in unchanged laboratory environ-
presented in figures 4,5 and 6. The simulation results ment. T h e obstacles were walls, tables, chairs, cup-
are summed in table 1. We used Bug2 algorithm from boards etc. (see figure 7) (3) adding artificial obsta-
[7] as a base line for comparison. cles t o the laboratory environment. T h e algorithm
The results show a qualitative difference between was successful in most cases, driving the robot t o the
“simple” environments I(“world1” and “world2”) and target location. However several problems were no-
complex ones (“world3 ). In the simple scenarios ticed. T h e range sensors (sonar, infrared and struc-
many passages exist between separate obstacles, and tured light) have low reliability in the laboratory en-
those are “found” by t h e local turning direction mech- vironment. As a result t h e robot might take wrong
anism (see first line in table 1). T h e obstacles are rel- decisions and the convergence of the algorithm is not
atively simple in shape, and in most cases, the robot guaranteed (for example - leaving a n obstacle bound-
is not driven far away from the target area. Therefore ary assuming that there are no obstacles in the di-
reversing the following direction has a negligible effect rection to the target). Moving obstacles, i.e. people,

1see second line in table 1). In the complex scenario

“world3”) there is only one connected obstacle with
long perimeter and complicated shape. T h e local in-
disturb t h e robot motion if they are detected during
boundary following behavior. T h e robot tries t o fol-
low these obstacle, but can not find them as they go
fformation is not sufficient t o determine the globally away. T h e problem is partially solved by a procedure
optimal turning direction, thus choosing the direction t h a t looks for “lost obstacles”, and drives the robot
has a small effect on performance. Choosing the cor- back to the previously followed obstacle after such a
rect turning direction becomes a crucial issue in com- disturbance. However a better analysis t h a t will dis-
plex environments, where turning t o the wrong direc- tinguish between moving a n d static obstacles is nec-
tion may increase the path length significantly. There- essary in order t o react differently to stationary and
fore the possibility t o reverse the boundary following moving obstacles. Reaching a person while moving in
direction improved the performance significantly (see straight line did not cause a problem in most cases:
second line in table 1). Leaving t h e obstacle bound- t h e robot tried t o bypass t h e obstacle, and as the ob-
ary and going straight towards the target improved stacle disappeared it concluded t h a t it can go straight
the path length in all the environments (see third line t o the target again.
in table 1). It also improved significantly the aver-
age distance from obstacles, and hence produced safer 4 Summary and conclusions
paths. We presented a sensory based algorithm DistBug
The algorithm was tested in more than one hun- t h a t is guaranteed t o reach the target if possible,
dred runs of our Nomad200 robot in three experimen- and report if the target is unreachable. T h e algo-
tal scenarios: (1) using several artificial obstacles (car- rithm is reactive in the sense t h a t it relies on range
toon boxes). The obstacles were usually grouped into d a t a to make local decisions, and does not create a
2-3 simple shaped obstacles (similar t o “worldl” set- world model. T h e algorithm consists of two behav-

439
 Acknowledgments
We would like t o thank Michael Heymann and
Freddy Bruckstein for very helpful discussions. We
want t o express special thanks t o Elon Rimon for his
encouragement and helpful comments.
References
Figure 6: Simulation results in “world3” environment. R. C. Arkin. Motor schema based navigation for a
Left - B u g 2 algorithm. Middle - DistBug2, choosing the mobile robot: an approach for programming by be-
turning direction + reversing boundary following direction
havior. In Proceedings of the I E E E Conference o n
(path length is 0.54). Right - D i s t B u g 3 , using the leaving
condition (path length is 0.36). Robotics and Automation, pages 264-271, 1987.
James L. Crowley. World modeling and position esti-
mation for a mobile robot using ultrasonic ranging. In
Proceedings of the I E E E Conference o n Robotics and
Automation, pages 674-680, 1989.
James L. Crowley and Yves Demazeau. Principles and
techniques for sensor data fusion. Signal Processing,
32:5-27, 1993.
S.G. Goodridge and R. C. Luo. Fuzzy behavior fusion
for reactive control of an autonomous mobile robot:
Marge. In Proceedings of the I E E E Conference o n
Robotics and Automation, pages 1622-1627, 1994.
0. Khatib. Real-time obstacle avoidance for manipu-
Figure 7: The Nomad200 robot in a typical laboratory lators and mobile robots. In Proceedings of the I E E E
environment.
Conference o n Robotics and Automation, pages 500-
505, 1985.
iors (modes of motion): straight motion between ob- V. J. Lumelsky and T. Skewis. Incorporating range
stacles and obstacle boundary following. Simulation sensing in the robot navigation function. I E E E
results as well as experiments with a real robot were transactions o n Systems, M a n , and Cybernetics,
presented. 20 (5) :1058-1 068, 1990.
The condition for leaving obstacle boundary is
based on the free range in the direction t o the tar- [7] V. J. Lumelsky and A. A. Stepanov. Path-planning
get. This condition allows the robot t o leave the ob- strategies for a point mobile automaton moving
stacle as soon as the local conditions guarantee global amidst obstacles of arbitrary shape. Algoritmica,
convergence. Range d a t a is utilized for choosing t,he 2:403-430, 1987.
initial boundary following direction when the robot [8] V. J. Lumelsky and S.Tiwari. An algorithm for maze
approaches an obstacle. A criterion for reversing the
searching with azimuth input. In Proceedings of the
boundary following direction when it seems t o be the
wrong direction is also introduced. As a direct result I E E E Conference o n Robotics and Automation, pages
of these local decisions a significant improvement in 111-116, 1994.
the performance was achieved. [9] H. P. Moravec and A. Elfes. High resolution maps
T h e simulation results showed a qualitative differ- from angle sonar. In Proceedings of the I E E E Con-
ence between simple environments and complex ones. ference o n Robotics and Automation, pages 116-121,
T h e environment simplicity is related t o the obstacles 1985.
shape, that affects the usefulness of local decisions,
and t o the length of the obstacles perimeter, that af- [lo] H. Noborio and T. Yoshioka. An on-line and deadlock-
fects the penalty for choosing the wrong turning di- free path-planning algorithm based on world topology.
rection. Choosing the initial turning direction was In Proceedings of the I E E E / R S J Conference o n Intel-
useful in simple environment. Reversing t h e bound- lzgent Robots and Systems, pages 1425-1430, 1993.
ary following direction was useful in complex envi-
ronments. Leaving the obstacle as soon as the local [ll] P. Reignier. Molusc: an incremental approach of fuzzy
conditions guarantee global convergence improved the learning. In Proceedings of the International sympo-
performance in all the environments. In our simple sium o n Intelligent Robotics Systems, pages 178-186,
environment the average path length over 100 runs 1994.
was reduced t o 0.79 (relative t o Bug2 performance),
the average traveling time was reduce t o 0.67, and the
minimal distance from obstacles (safety measure) was
increased t o 1.6. In the complex environment the av-
erage path length was reduced t o 0.45.

440

View publication stats