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

net/publication/3802458

Range-sensor based navigation in three dimensions

Conference Paper in Proceedings - IEEE International Conference on Robotics and Automation · February 1999
DOI: 10.1109/ROBOT.1999.769955 · Source: IEEE Xplore

CITATIONS READS

48 85

3 authors:

Ishay Kamon Elon Rimon

Technion - Israel Institute of Technology Technion - Israel Institute of Technology
12 PUBLICATIONS 760 CITATIONS 127 PUBLICATIONS 5,409 CITATIONS

SEE PROFILE SEE PROFILE

Ehud Rivlin
Technion - Israel Institute of Technology
248 PUBLICATIONS 9,924 CITATIONS

SEE PROFILE

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

Walking Robots View project

Robotic grasping system 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.

 Proceedings of the 1999 EEE
International Conference on Robotics & Automation
Detroit, Michigan May 1999

Range-Sensor Based Navigation in Three Dimensions

Ishay Kamon Elon Rimon Ehud Rivlin
CS Department, Technion ME Department, Technion CS Department, Technion
ishayQcs.technion. ac.il elonQrobby.technion.ac.il [email protected]

Abstract which monitors a globally convergent criterion, holds.

If the robot completes a loop around the obstacle with-
This paper presents a new globally convergent range- out satisfying the leaving condition, the robot concludes
sensor based navigation algorithm in three-dimensions, that the target is unreachable.
called 3DBug. The 3DBug algorithm navigates a point Currently all the B u g algorithms plan paths in two-
robot in a three-dimensional unknown environment us- dimensional configuration spaces. Extending these al-
ing position and range sensors. The algorithm strives gorithms to three-dimensional spaces is difficult since
to process the sensory data in the most reactive way the obstacle boundaries are surfaces, while the robot’s
possible, without sacrificing the global convergence guar- path is a one-dimensional curve. Thus, to conclude tar-
antee. Moreover, unlike previous reactive-like algo- get unreachability the robot cannot merely complete a
rithms, 3DBug uses three-dimensional range data and loop around the obstacle. Rather, the robot must verify
plans three-dimensional motion throughout the naviga- that the leaving condition is not satisfied on the entire
tion process. The algorithm alternates between two surface of the obstacle.
modes of motion. During motion towards the target, Recently, Kutulakos and coworkers studied the prob-
which is the first motion mode of the algorithm, the lem of Bug-type three-dimensional navigation [5]. They
robot follows the locally shortest path in a purely reactive suggest a scheme for three-dimensional path planning
fashion. During traversal of an obstacle surface, which which combines a 2D Bug algorithm with a 3D surface-
is the second mode of motion, the robot incrementally exploration algorithm. They argue that the reactive be-
constructs a reduced data structure of an obstacle, while havior of the planar algorithms must be relaxed during
performing local shortcuts based on range data. W e the three-dimensional obstacle surface exploration stage.
present preliminary simulation results of the algorithm, In particular, a three-dimensional algorithm must in-
which show that 3DBug generates paths that resemble clude a data structure which supports surface explo-
the globally shortest path in simple scenarios. Moreover, ration and allows a conclusion that the entire obstacle
the algorithm generates reasonably short paths even in surface has been explored. We propose a new and re-
concave, room-like environments. duced data structure of that type. Moreover, as dis-
cussed below, our algorithm is fully three-dimensional.
1 Introduction
The work presented here has two objectives. First,
we present basic results concerning sensor-based ob-
Autonomous robots which navigate in realistic settings stacle surface coverage, and results concerning compu-
must use sensors to perceive the environment, and plan tation of a locally shortest path in three-dimensions.
accordingly. Some approaches to sensor-based naviga- We believe that these results are of independent in-
tion are the works of Chatila [ l ]Foux
, [2],Taylor [14], terest to researchers in motion planning. Second, we
Rao [lo],Stentz [13], and their respective coworkers. incorporate these basic results into the new globally
One particular approach, called the Bug paradigm, was convergent navigation algorithm 3DBug. The algo-
originated by Lumelsky and Stepanov [8] and subse- rithm navigates a point robot in a three-dimensional
quently studied by [3, 7, 121. The Bug algorithms are unknown environment using position and range sen-
simple to implement since they act mainly in a reac- sors. 3DBug uses three-dimensional range data and
tive fashion, while augmenting the local planning with a plans three-dimensional motion throughout the naviga-
globally convergent criterion which influences the robot tion process. This is in contrast with Ref. [5], where the
decisions. In the basic Bug algorithms, the robot ini- motion towards the target and the convergence mech-
tially moves towards the target until it hits an obstacle. anism are restricted to a series of planes embedded in
Then the robot switches to motion along the obstacle R3. Thus 3DBug provides a new and more effective
boundary. The robot leaves the obstacle and resumes Bug-type navigation algorithm for three-dimensions.
its motion towards the target when a leaving condition, The paper is organized as follows. First we discuss

0-7803-5180-0-5/99 $10.00 0 1999 IEEE 163

some basic results necessary for sensor based naviga- Moreover, it can be verified that these edges belong to
tion in three-dimensions. Next we present the 3 D B u g the set S. Hence every point on the obstacle surface is
algorithm, and describe preliminary simulation results, directly visible from some convex edge in S,as summa-
showing that 3 D B u g generates paths which often re- rized in the following theorem.
semble the globally shortest path to the goal. Finally,
the concluding discussion outlines future work and ex- Theorem 1 The entire surface of a polyhedron 23 is vis-
tensions of 3DBug. ible from the convex obstacle edges which intersect its
convex hull.
2 Basic Results in 3D Sensor-Based Navigation
Thus a robot equipped with a range sensor can visually
In this section we define the sensor model, then discuss explore the entire surface of a polyhedral obstacle by
some basic results necessary for sensor based motion tracing all the convex obstacle edges in its convex hull.
planning in three-dimensions. We assume an environ- It is not hard to see that S is the minimal collection of
ment populated by polyhedral obstacles. First we show edges which guarantees visual coverage. Hence we use
that the robot can visually explore the entire surface of this compact representation both for surface exploration
a given polyhedron by tracing the convex obstacle edges and shortest path computation. We define a novel data
which intersect the polyhedron’s convex hull. Then we structure called the Convex Edges Graph, or CEG. The
define a novel data structure which supports both sur- CEG nodes represent convex obstacle edges which lie
face exploration and shortest path computation. Last in the obstacle’s convex hull and have been seen by the
we focus on the notion of local information, and present robot during the exploration. The CEG edges repre-
the locally shortest path and a technique for efficiently sent paths between the respective convex obstacle edges.
computing this path. A detailed analysis of the results They are constructed such that the connectivity of the
described here appears in Ref. [4]. CEG is maintained. To make the CEG also useful for
Let x be the current robot location. We assume a motion planning, we add the current robot location x
range sensor with infinite detection range, which pro- and the target location T as special CEG nodes, and
vides perfect readings of the minimal distance of the add special edges from x and T t o CEG nodes. Thus it
robot from the obstacles along rays which emanate from is possible to compute a path from x to T on the CEG.
x in all directions. The resulting visible set is a three- A detailed description of the CEG appears in [4].
dimensional star-shaped set centered at x and contained
in the free space. It can be verified that the obstacles’
visible surfaces are planar polygons [ll].Moreover, each 2.2 Locally Shortest Path in Three-Dimensions
visible surface is bounded by edges of two types-convex We define the locally shortest path from the robot cur-
obstacle edges’ and edges generated by occlusion. In the rent location x to the target T as the shortest collision-
following we assume that the range data is processed free path, based only on the currently visible obstacles.
and transformed into the three-dimensional coordinates We now present a technique for efficiently estimating
of the vertices and edges of the visible obstacles. this path. First let us introduce some terminology. We
model each visible surface as a polyhedral two-sided
2.1 Sensor Based Surface Exploration
thin wall or shell. If the target is not visible from x ,
In this section we show that the entire surface of a given there is some blocking obstacle between the robot and
polyhedron 23 can be visually explored while tracing only the target. In this case the line segment [.,TI crosses
convex edges. To investigate the visibility properties in the blocking obstacle, and we refer to the visible surface
three-dimensions, we use the following characteristic of which blocks [ x ,T]as the blocking surface. The blocking
the shortest path in three-dimensions. surface is bounded by a piecewise linear curve, termed
the blocking contour (Figure 2). By construction, the
Lemma 2.1 The shortest path in a three-dimensional blocking contour lies on the blocking obstacle, and its
polyhedral environment is piecewise linear, and the path edges are of two types-occluding and occluded edges.
vertices lie only on convex obstacle edges. Occluding edges are convex edges of the blocking ob-
stacle. Occluded edges are generated from occlusion by
Let S denote the set of convex edges contained in the some other obstacle. Each occluded edge of the block-
convex hull of the polyhedron B. It follows from the ing contour has a corresponding occluding edge, which
lemma that the shortest path between any two points on is a portion of a convex obstacle edge of some other ob-
the surface of B passes through convex obstacle edges. stacle (Figure 2(b)). The following proposition asserts
that the locally shortest path always passes through the
An obstacle edge is conves if there exists a plane which passes
through the edge such that the obstacle locally lies in one half blocking contour.
space only. An edge is concave if it is not convex.

164
 point within the visible set. First we describe the global
structure of the algorithm and then discuss its detailed
14 operation. A detailed convergence proof for the 3DBug
algorithm appears in Ref. [4].

3.1 Algorithm Description

Figure 1. The blocking-contour graph of (a) a convex ob- The 3 D B W algorithm uses two basic motion-modes:
stacle (b) a concave obstacle. motion towards the target and motion along an obsta-
cle surface. During motion towards the target the robot
Proposition 2.2 I n a polyhedral three-dimensional en- moves along the locally shortest path based on the cur-
vironment, let a blocking obstacle lie between the robot rently visible obstacles. At each step of this motion
location x and the target T. Then the shortest path mode the robot senses the environment, chooses a fo-
from x to T , considering only the currently visible cus point F , and moves directly to F without perform-
obstacles, passes through the blocking contour. ing any sensing or replanning until it reaches the focus
point. Let w be a point in the free space, and let the
In principal, it is possible to compute the locally function d(w,T)be the Euclidean distance of w to the
shortest path using €-optimal algorithms (e.g. [91) on target T . The robot keeps moving towards the target
the thin-wall model of the currently visible obstacles. until it becomes trapped in the basin of attraction of a
But these algorithms are computationally intensive, and local minimum of d ( w , T ) . The local minimum is gener-
we now present an alternative method for efficientlyesti- ated by an obstacle which blocks the robot's path to the
mating the locally shortest Path. The resulting estimate target, and the robot switches to traversing the surface
is called the blocking-contour path. For each point y on of this obstacle.
the blocking contour, consider a path consisting of two D~~~~~ the surface-traversing mode, the robot
line segments: the visible Part [x,Y]and the (OPtimisti- searches for a suitable leave point on the surface from
callY) expected part b q The length Of this path is which it can resume its motion towards the target. At
Lblock(Y) = (1% - 911 4- 119 - For each line segment li the Same time the robot expands its knowledge of the
of the blocking contour, we Compute the Point V i which obstacle surface and stores this information in the Con-
minimizes L b k k ( Y ) . This Computation Can be done in vex Edges Graph, CEG. At each step during the surface-
constant time per line segment 1%- Then we construct a traversing mode, the robot computes the shortest path
local graph, called the blocking-contour graph, consisting to the target based on the current CEG, chooses the
of edges from x to each and (optimistic) edges from next focus point F on this path, and moves to F . The
each vi t o T , as &own in Figure 1. The blocking-contour robot then Senses the environment, updates the CEG,
path is the shortest path on the blocking-contour graph, and records the closest point to the target observed SO
and it can be found in time linear in the number of line far on the surface of the followed obstacle. The closest
segments in the blocking contour. observed point is denoted p,,, and its distance to the
w e may ask, what is the relation of the blocking- target is denoted drnin(T).
contour Path to the exact locally shortest Path? If the After updating the CEG, the robot tests the leav-
blocking obstacle is convex, the blocking-contour path ing condition as ~ollows~It inspects Zlleave, which is
is precisely the locally shortest path, for the the closest point to the target along the visible par-
Let Y be a Point On the contour. If tion of the line segment [$,TI,where x is the current
the line segment [Y,TI intersects the 'IUface at
robot location. The robot leaves the obstacle surface
some visible point, the blocking obstacle must have a when d(wleave,T)< dmzn(T).After leaving the obsta-
visible concavity. Hence if the obstacle is convex, the cle, the robot performs a transition phase before it re-
line [!I, T ] never intersects the blocking surface, and the sumes its motion towards the target. In this phase the
blocking-contour Path is Precisely the loCa1lY shortest robot moves directly towards wleave until it reaches a
path. But in general IIY - iS merely an optimistic point where d(z,T) < dmin(T). At this point the
estimate of the path length from y to T . robot resumes its motion towards the target.
3 The 3DBug Algorithm While searching for a suitable leave point, the robot
accumulates data on the obstacle surface in the CEG. If
The 3DBug algorithm navigates a point robot in a the robot finishes tracing all the convex obstacle edges
three-dimensional unknown environment populated by in the CEG before finding a leave point, it has com-
stationary polyhedral obstacles. The sensory informa- pleted the exploration of the obstacle surface. On that
tion available to the robot consists of the robot cur- occasion, the robot performs the following final target-
rent position x, and range data from x to every obstacle reachability test. The robot moves to the closest point

165
to the target on the obstacle surface, pmin, and checks
the leaving condition from there. If the leaving con-
dition is not satisfied at pmin, the target is necessarily
unreachable and the robot halts its motion. A summary
of the algorithm now follows.
1. Move towards T along the locally shortest path,
until one of the following events occurs: Figure 2. Motion towards the target. (a) From x = S, the
0 The target is reached. Stop. locally shortest path passes through the focus point 9 . (b)
0 A local minimum is detected. Go to step 2. At x = F1, the point G lies on an occluded edge. Hence the
2. Traverse the obstacle's surface, searching for next focus point, F2, is set on the occluding edge. (c) From
2 = Fz, the locally shortest path passes through F3, from
a suitable leave point, while updating the CEG
and recording pmin and dmin(T), which the target can be reached directly.
until one of the following events occurs:
0 The target is reached. Stop.
The robot terminates its motion towards the target
0 The leaving condition holds:
and switches to the surface-traversing mode after de-
3vleaue d(Uleouet T ) < dmin(T). Go to step 4- tecting that it is trapped in the basin of attraction of a
0 The entire surface has been sensed. Go to step 3.
local minimum of the function d(w, T ) . The correspond-
3. Perform the final target reachability test: ing sensor-based termination condition is that the fea-
GOto pmin. sible sub-contour becomes empty, and it can be verified
If the leaving condition holds at pmin, go to step 4. that this event is always associated with the presence of
Otherwise the target is unreachable. Stop. a local minimum of d(w, T ) [4].
4. Perform the transition phase.
Move directly towards uleoue until reaching
a point z where d ( z , T ) < dmin(T). Go to step 1. 3.3 Traversing an Obstacle Surface
This motion mode has two simultaneous objectives-to
3.2 Motion Towards the Target
find a suitable leave point and t o explore the obsta-
During motion towards the target, the robot moves be- cle surface. Let P denote the point where the robot
tween successive focus points along the locally shortest switches to surface-traversing mode. It can be verified
path to the target, based on the currently sensed ob- that the local minimum of d(w, 2') which terminated the
stacles. If the target is visible to the robot, the robot motion towards the target is visible from P , and lies on
moves directly towards the target. the robot moves di- the surface of the obstacle which blocks the direct path
rectly to it. Otherwise, the locally shortest path passes from P to the target (the blocking obstacle). The robot
through the blocking contour (Proposition 2.2). In or- traverses the surface of this obstacle until either a leav-
der to guarantee convergence to the target, we wish to ing condition is satisfied or the entire obstacle surface is
ensure that the distance of the robot to the target de- explored. Upon starting a new surface traversing seg-
creases monotonically between successive focus points. ment, the robot moves into the convex hull of the block-
To achieve this objective, the algorithm computes the ing obstacle, senses the environment, and generates the
locally shortest path based only on the points y of the initial CEG of the blocking obstacle.
blocking contour satisfying d ( y , T ) 5 d(z, T ) , where x At each step after the initial one, the robot computes
is the current robot location. We call this subset of the shortest path to the target, y,on the current CEG.
the blocking contour the feasible sub-contour (Figure Given this path, the robot chooses the next focus point
2(a)). Once the feasible sub-contour is computed, the F as the last vertex along y which lies on an obsta-
algorithm constructs the blocking-contour graph based cle edge. The robot then moves to F by repeatedly
on the feasible sub-contour and the target node, and performing the following procedure. The robot chooses
searches this graph for the shortest path to T . the furthest visible point v along y, and moves directly
Once the locally shortest path is computed,.the next towards v without performing any sensing. After reach-
focus point F is chosen on this path as follows. Let G ing U , the robot senses the environment, and repeats the
be the point on the feasible sub-contour through which same procedure of moving t o the furthest visible point
the locally shortest path passes. If G lies on a convex along y. After finitely many such steps the robot reaches
edge of the blocking obstacle, F is set to G (FI in Fig. the focus point F.
2(a)). If G lies on an edge generated from occlusion, Upon reaching F , the robot traces a small portion
F is chosen on the occluding edge, at the point where of the convex edge on which F is located while contin-
the line segment [x,G] crosses the occluding edge (Fi in uously sensing the environment. The accumulative ef-
Fig. 2(b)). fect of tracing small edge segments each time the robot

166
 envl I env2 I env3 1 env4
1.00 I 1.02 I 1.06 I 1.03

Table 1. Average simulation results of 3DBug, relative to

the (approximate) globally shortest path.

reaches a focus point is the visual coverage of the entire

obstacle surface. During the tracing operation the robot
updates the CEG according to the sensed range data,
and continuously records the closest point to the target
observed so far on the obstacle surface, pmin.
After updating the CEG, the robot tests the leaving
condition as follows. The robot inspects vieave,the clos-
est point to the target along the visible portion of the Figure 3. 3DBug in env2. (a) The visible surfaces as seen
segment [z, TI,where z is the current robot location. from the start point S. The locally shortest path leads to F1
The leaving condition is satisfied when d ( v l e a v e , T ) < since the blocking obstacle 0 1 is only partially visible. (b)
d m i n ( T ) , where d,in(T) is the distance of pmin to T. The path generated by 3DBug, compared to the globally
If the entire surface has been explored without find- optimal path.
ing a leave point, the robot performs the following final
target-reachability test. The robot moves to pmin, and
checks the leaving condition at pmin. If the leaving con-
dition is not satisfied at pmjn,the target is unreachable.
This final test is necessary since the leaving condition is
previously tested only at discrete points on convex ob-
blocking contour
stacle edges. But in general these points do not suffice
to conclusively determine target unreachability.
Finally, after leaving the obstacle, the robot performs
a transition phase where it moves directly towards vleave
until it reaches a point z where d(z, T) < dmin(T).The
combination of the leaving condition and the transition
phase ensures that each local-minimum of d ( w ,T) is as-
sociated with at most one switch from motion-towards-
the-target to surface-traversing mode.

4 Simulation Results

In this section we present simulation results which com-

pare the path generated by 3DBug to the globally short-
est path. To simulate the 3DBug algorithm, we devel-
oped a three-dimensional range-sensor simulator, which
computes the blocking surface in environments popu-
lated by general polyhedra. We approximate the glob-
ally shortest path by constructing and searching a three-
dimensional generalized visibility graph [6].
bloc!& contour
ta two components of
blocking contour

We present simulation results of 3 D B u g in four sim-

ulated environments. The average results of the exper-
Figure 4. 3DBug in env3, as the robot moves into the room.
iments are expressed in Table 1, relative to the (ap-
(a) The entire path of 3DBug, compared to the globally
proximate) globally shortest path. The environment optimal path. (b) The blocking contour computed from S,
envl consists of a single box-like obstacle. In this en- shown in bold line. (c) The blocking contour from F I . (d)
vironment 3DBug's paths are almost identical to the The blocking contour from F2. The target is directly visible
visibility-graph paths in all of the runs. The environ- from F3.
ment env2 is more complex and consists of seven box-
like obstacles (Fig. 3 ) . The average path length of

167
 the roof from FI, the robot moves along the shortest
path from Fl to T . Partial occlusion is another mani-
T (invisible) festation of the limited nature of local information, as
demonstrated in Figure 3. Using the motion-towards-
the-target mode, the robot moves from S to the focus
point Fl, which lies on an occluding edge (Fig. 3(a)).
The robot chooses this path from S since it does not see
the entire blocking obstacle, denoted 01, and occluded
portions of obstacles are considered as non-existent. An-
other reason for the difference between the two paths
is the incorporation of the global convergence require-
ment during motion-towards-the-target. Restricting the
computed shortest path to the feasible sub-contour may
prevent the robot from moving along the precise locally
shortest path, which may pass through any point on the
blocking contour. Thus there are several reasons which
cause 3DBug’s paths to differ from the globally optimal
ones.
3DBug as a search algorithm: Last we discuss the
search characteristics of 3DBug. In the graph-search
terminology, the motion-towards-the-target mode is
a hill-descending strategy, and the surface-traversing
mode is a mechanism for escaping local minima. For
Figure 5 . 3DBug in env4. (a,b) The robot leaves house1 comparison, we consider the classical A* algorithm
from the window, and enters house2 from the door. The which uses the generalized visibility graph as the un-
globally optimal path is almost identical to 3DBug’s path.
derlying search space. The 3DBug algorithm finds the
(c) The blocking contour from S. (d) The blocking contour
from FI (located at the internal window frame). target in fewer steps than A* for the following reasons.
First, 3DBug performs a depth search, thus it moves
faster towards the target. Second, the candidate loca-
tions for the next step in 3DBug are limited to a single
3DBug’s paths in e m 2 is 1.02. In both envl and env2
obstacle, which is the blocking obstacle in both modes
the algorithm used the motion-towards-the-target mode
of motion. In contrast, A* must consider all the nodes
along the entire path in over 99% of the runs.
which are visible from each node U in the generalized vis-
The environment env3 consists of a single concave ibility graph. In env3, for example, 3DBug reaches the
room-like obstacle (Fig. 4 ) . The average path length in target after 3.3 steps on average, while A* requires 32.4
this environment is 1.06 (relative to the generalized vis- steps to reach the target. The advantage of 3DBug is
ibility graph shortest path), and the surface-traversing even more pronounced when the target is unreachable.
mode was activated in 65% of the runs. The last envi- The 3DBug algorithm concludes target unreachability
ronment env4 consists of two room-like obstacles, sep- after exploring the entire surface of a single obstacle in
arated by a wall (Fig. 5). The start and target points which the target is trapped, while A* must expand all
were always placed inside or near the rooms, on differ- the nodes in its search space to conclude unreachability.
ent sides of the separating wall. In this environment the Another advantage of 3DBug is that it uses a compact
average path length of 3DBug is 1.03. The tested sce- data structure, since it uses only a limited amount of
narios constitute only a preliminary study. There are global information. In contrast, a data structure which
other environments, in which the locally optimal deci- represents the entire environment may be very large.
sions do not necessarily lead to globally optimal paths, For example, the generalized visibility graph of env2
and in these environments 3Dbug would be less effective. with resolution 0.1 consists of 620 nodes and 118912
The local characteristics of 3DBug: The paths edges, while 3DBug’s data structure consists on the av-
of 3DBug are distinct from the globally optimal ones erage of 7 nodes and 9 edges.
for several reasons. As demonstrated in Figure 4 , the
locally shortest path may differ from the globally op-
timal one due to the limited nature of local informa- 5 Concluding Discussion
tion. From S, the robot moves towards the ”roof” of We presented new basic results in sensor-based surface
the room, since the roof is not visible from S and thus exploration, and locally shortest path computation in
considered non-existent (Fig. 4 ( b ) ) . After observing three-dimensional polyhedral environments. Consider-

168
 ing surface exploration, we showed that the entire sur- I. Kamon, E. Rimon, and E. Rivlin. Tangentbug: A
face of a polyhedral obstacle is visible from the convex range-sensor based navigation algorithm. To appear
obstacle edges within the obstacle’s convex hull. Then in the International Journal on Robotic Research.
we introduced the notion of a locally shortest path in
three-dimensions, and showed that it must pass through I. Kamon, E. Rimon, and E. Rivlin. Range-sensor
the blocking contour. We used this property to for- based navigation in three dimensions. CIS 9712,
mulate an efficient technique for estimating the locally Center of Intelligent Systems, Dept. of Computer
shortest path in time linear in the number of edges in Science, Technion, Israel, 1997.
the blocking contour. K. N. Kutulakos, V. J. Lumelsky, and C. R. Dyer.
These results were incorporated into the new glob- Vision guided exploration: a step toward general
ally convergent 3 D B u g algorithm, which navigates a motion planning in three dimensions. I E E E Conf.
point robot equipped with position and range sensors on Robotics and Automation, 289-296, 1993.
in a three-dimensional unknown environment. The al-
gorithm falls within the general framework of the Bug T. Lozano-Perez and M. A. Wesley. An algo-
paradigm since it strives to process the sensory data rithm for planning collision free paths among poly-
in the most reactive way possible, without sacrificing hedral obstacles. Communications of the A C M ,
the global convergence guarantee. During motion to- 22(10):560-570, 1979.
wards the target, the robot follows the locally shortest
path in a purely reactive fashion. During traversal of V. J. Lumelsky. A comparative study on the path
an obstacle surface, the robot incrementally constructs length performance of maze-searching and robot
the CEG of the obstacle, while performing local short- motion planning algorithms. I E E E Transactions
cuts based on the local range data. Simulation results on on Robotics and Automation, 7(1):57-66, 1991.
show that 3 D B u g generates paths which resemble the V. J. Lumelsky and A. A. Stepanov. Path-planning
globally shortest path in simple scenarios, consisting of strategies for a point mobile automaton moving
disjoint convex obstacles. Moreover, the algorithm gen- amidst obstacles of arbitrary shape. Algoritmica,
erates reasonably short paths even in concave, room-like 2:403-430, 1987.
environments.
Let us mention several potential uses and extensions C. H. Papadimitriou. An algorithm for shortest
for the new algorithm. First, 3 D B u g is useful for nav- path motion in three dimensions. Information pro-
igating free-flying robots in either real tasks such as cessing letters, 20:259-263, 1985.
surveillance, or in simulated scenarios such as virtual re-
N. S. V. Rao and S. S. Iyengar. Autonomous robot
ality games. Second, 3 D B u g provides new insight into
navigation in unknown terrains: visibility graph
the important problem of sensor-based navigation in
based methods. I E E E transactions on Systems,
three-dimensions. Third, the algorithm can be extended
Man and Cybernetics, 20(6):1443-1449, 1990.
to other three-dimensional configuration spaces, such as
the ones associated with three degrees-of-freedom mo- J. H. Rieger. The geometry of view space of opaque
bile robots. Last, 3 D B u g is also useful as a search al- objects bounded by smooth surfaces. Artificial In-
gorithm in completely known three-dimensional envi- telligence, 44:l-40, 1990.
ronments. The main advantage of 3Dug over classical
search algorithms such as A* is that it takes into con- A. Sankaranarayanan and M. Vidyasagar. Path
sideration the geometric characteristics of the locally planning for moving a point object amidst unknown
shortest path. Consequently, 3 D B u g finds the target obstacles in a plane: the universal lower bound on
much faster than other, less informed, algorithms. worst case path lengths and a classification of algo-
rithms. I E E E Conf. on Robotics and Automation,
1734-1941,1991.
References
A. Stentz. Optimal and efficient path planning
R. Chatila. Deliberation and reactivity in au- for partially known environments. I E E E Conf. on
tonomous mobile robots. Robotics and Autonomous Robotics and Automation, 3310-3317, 1994.
Systems, 16:197-211, 1995. C. J. Taylor and D. J. Kriegman. Vision based mo-
tion planning and exploration algorithms for mobile
G. Foux, M. Heymann, and A. Bruckstein. Two robots. Workshop on Algorithmic Foundations of
dimensional robot navigation among unknown sta- Robotics, 69-83. A K Peters, 1995.
tionary polygonal obstacles. IEEE Transactions on
Robotics and Automation, 9(1):96-102, 1993.

169

View publication stats