Proceedingsof the 1999 IEEEBSJ

International Conference on
Intelligent Robots and Systems

Three or More Dimensional Sensor-Based Path-Planning

Algorithm H D - I
Hiroshi Noborio+, Yasuhiro Maeda++, and Kenji Urakawa+
+ Graduate School of Engineering, Osaka Electro-Communication Univ.
++ Department of Eng. Informatics, Osaka Electro-Communication Univ.

Abstract Secondly, the latter algorithm runs in an unknown

map. Thus, a robot finds a path until a destina-
To design an on-line or sensor-based path-planning tion while investigating a part of the unknown map
algorithm, we should decrease its search cost (the num- by sensing. Therefore, whether an algorithm is the
ber N of visited points of a robot an a known map or sensor-based algorithm or not is evaluated by the sum
the sum S of moving distances of it in an uncertain S of moving distances of a robot. If and only if S
map). For this purpose, the optimal algorithm, e.g. a n is small enough, the algorithm is used for the sensor-
A* algorithm is not suitable. Because in order t o se- based algorithm. Since S is exactly larger than N ,
lect an optimal path between start and goal points, the an off-line (model-based) algorithm is not absolutely
algorithm requests a huge set of visited points or driv- to be a sensor-based algorithm, and also an on-line
ing distances. O n the observation, b y decreasing the algorithm is not always to be it.
sum S , we design a new sensor-based path-planning In the past decade, many sensor-based path-
algorithm H D - I . planning algorithms have been proposed [l].In the
I n almost all the sensor-based algorithms, e.g. sensor-based path-planning, a robot always selects one
C l a s s l , convergence of a robot to its goal is ensured of two behaviors, that is, going straight to its desti-
b y decreasing their distance. Also in a famous algo- nation, and avoiding an encountered obstacle. The
rithm Best - F i r s t , the convergence is also kept by behaviors are switched if a robot hits or leaves an
visiting a node whose distance to its goal is the small- unknown obstacle. Here, as long as leave points
est. They are similar because a robot does not take a monotonously or asymptotically come close to a des-
distance from a start into account. Therefore, their tination, convergence of a robot to its destination is
modified versions can be mixed. Especially, B F can completely kept in an uncertain environment. This is
investigate around an unknown obstacle whose dimen- called a metric condition. In general, a leave point is
sions are over two. A s a result, by mixing modified always around an encountered obstacle. Therefore, if
B F and C l a s s l , we obtain three or more dimensional the dimension of an obstacle is two, a robot always
algorithm H D - I . reaches the leave point while tracing an uncertain ob-
stacle by one of the clockwise and counter-clockwise
directions. However, the characteristic cannot be un-
1 Introduction fortunately maintained on a three or more dimensional
obstacle because the number of following directions
The robot path-planning is covered by three types is infinite. As a result, all the previous sensor-based
of algorithms, that is, off-line (called model-based) path-planning algorithms cannot be applicable for a
algorithm, on-line algorithm, and sensor-based algo- robotic manipulator whose degrees of freedom are over
rithm. First of all, the former two algorithms run in two.
a known map. The difference of them is whether the To overcome this problem, we have considered two
number N of searching points is large or small. An ways: one is to restrict kinematics of a robotic ma-
algorithm is regarded as the off-line algorithm if its nipulator, and another is t o pick up a part of config-
N is large, otherwise, it is considered as the on-line uration space systematically. As a former example, a
algorithm. For example, in order to select an optimal manipulator with two degrees of freedom, or a scalar
(shortest) path between start and goal points, an A* manipulator without any revolution joint was consid-
algorithm investigates a huge set of points in a known ered [2], [3], [4]. As a latter example, a major pair of
map. Therefore, A* is to be the off-line (model-based) degrees of freedom was selected for a three-degrees-of-
path-planning algorithm. freedom manipulator [5], [6],or a 2-D part of config-

0 1999 IEEE
0-7803-5184-3/99/$10.00 1699

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
 uration space of a 6-degrees-of-freedom manipulator to three, we can design a path for manipulation.
was systematically selected by using recursively sev- In this paper, we define a configuration space Cspace
eral dimensional hyper-planes [7], [8], [9]. As a sum- as a collection of cells. First of all, the configura-
mary, any sensor-based path-planning algorithm for a tion space Cspacewith m dimensions is built in a joint
manipulator with three or more revolution joints has space [e,, ..., Om]. All the cells are built by dividing an
not been presented. available range of each degree into an arbitrary reso-
In this paper, we propose a three or more dimen- lution: (el, ..., e,)/O,E[e;, 01”],..., eE
,O
[,: Om”], The
sional sensor-based path-planning algorithm H D - I differences Bi” - e:(i = 1, ...,rn) are available ranges,
for a robotic manipulator with three or more degrees and they are divided by 2‘ times. 1 is given by a hu-
of freedom. Especially in order to find a leave point man operator, and the arbitrary resolution is deter-
around three or more dimensional unknown object, mined by IOi” - / 2’(i = 1, ..., m). According to
we use a modified version of a Best - First algo- the division, a digital map D M is built by 21m cells
rithm. The reason why we select B F is the similarity (nodes). For example, 2-D and 3-D digital maps are
to sensor-based path-planning algorithm with the met- represented by arrangements [2’, 2’1 and [2’, 2‘, 2l], re-
ric condition. That is, both do not consider any cost spectively (Fig.1). A purpose of using a digital map
from a start point to a present point. In other words, is to express three or more dimensional configuration
both consider only a future cost from a present point space. However in this paper, for the simplification of
to a goal point, which is called heuristics. explanation, we basically use a 2-D digital map. Then,
A new B F used in H D - I differs from an original all cells (nodes) are classified into two categories Cf,.,,
B F in three following aspects. (1)The new B F starts (Free cells) and Cobs ( 0 6 s cells). If and only if a cell
(begins to input found points into the OPEN list) if a (node) does not intersect any configuration space ob-
robot hits an obstacle, and finishes (initializes OPEN stacle, the cell (node) is Free. Otherwise, the cell
and CLOSED lists) if the robot leaves the obstacle. (node) is 06s. Secondly, a robot is expressed by a
(2) A robot traces only boundary of an uncertain ob- cell (node), and is able to move one of neighbor cells
stacle from a hit point to a leave point. (3) A cost of (nodes) in the digital map. If the dimension of a con-
n, is evaluated by summing length of a shorter route figuration space is two and three, eight and twenty-six
between np and nu and length of the straight segment cells (nodes) are prepared for neighbors, respectively.
between n, and ng.The nodes np, nu, and ng are de- Two neighbor cells (nodes) are implicitly connected by
noted as a present point, a selected point, and a goal an edge. Furthermore, the distance (cost) of an edge
point. between horizontal and vertical neighbors is defined by
The paper is organized as follows: Section 2 de- one, and the distance (cost) of an edge between diag-
scribes a robot and its search space. In section 3, we onal neighbors is defined by a square route two ( d ) .
describe algorithms B F and A* as model-based path- A total distance between start and goal cells (nodes
planning algorithms [lo], [ll],and also explain the ns and ng)is calculated by summing these neighbor
algorithm Class1 as a sensor-based path-planning al- distances.
gorithm [12]. Then, we propose a high dimensional
sensor-based path-planning algorithm H D - I for se-
lecting a path from three or more dimensional search
L61
space. In section 4, we describe the number N of visit-
ing points and the sum S of moving distances by four
types of algorithms in 2-D and 3-D uncertain environ-
ments. Finally, section 5 presents a few concluding
remarks.

2 A Point Robot and Its Digital Map

IJFree Hobs
In this paper, we adopt a cell decomposition ap-
proach to path planning, which has a flexibility for
representing free shaped objects with three or more Figure 1 : A cell robot and its 2-D digital map with Free and Obs
cells (nodes).
dimensions [13]. If the dimension is two, we can plan
a path for navigation, and if it is larger than or equals

1700

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
3 Four Kinds of Algorithms 3. Remove from OPEN a node np at which h(n,)
is minimum (break ties arbitrary), and place it on a
In this section, we explain three classic path- list called CLOSED to be used for expanded nodes
planning algorithms B F , A* [ll],and Classl [12]. (cells).
Then by mixing modified B F and Classl, we get an 4. Expand node (cell) np for selecting all Free and
algorithm HD - I as a 3-D on-line and sensor-based unvisited nodes (cells) n, neighboring np,and attach
path-planning algorithm. to them pointers back to np.
5 . If any nt is a goal node n g ,exit successfully with
3.1 An algorithm B F a shorter path obtained by tracing back the pointers
from ng to n,.
In this paragraph, we explain a classic B F algo- 6. For every neighbor n,:
rithm whose search space is a digital map shown in (sa) Calculate h(n,) as the pecewzseL2 metric
Fig.1. The difference against a state-space graph is as from n, to ng.
follows: For a node (cell) np in the digital map, there (6b)If nzwas neither on OPEN nor on CLOSED,
are neighbor nodes (cells) n,(z = 1,...,3m - 1). The put it on OPEN. Assign the newly computed h(n,)
distances (costs c(np,n,)) are horixontally and verti- to n,.
cally defined by one and diagonally defined by a square (6c) If n, already resided on OPEN or CLOSED,
route two. Furthermore, since a digital map is a fea- compare the newly computed h(n,)with the value pre-
tureless general state-space graph, a robot node (cell) viously assigned t o n,. If the old value is lower, discard
np is estimated numerically by a heuristic evalua- the newly generated node, otherwise, substitute it for
tion function h(np)which is the piecewiseL2 metric the old (n, now points back to np instead of to its pre-
between np and n g . As shown in Fig.2, a classic B F vious predecessor). If the matching node n, resided
algorithm cannot select the optimal (shortest) path on CLOSED, move it back to OPEN.
because it only considers a future cost h(np)from a 7. Go to step 2.
present node n p to a goal node ng (does not consider An algorithm B F always selects a sequence of Free
a past cost g(np)from a present node np to a start nodes (cells) from n, to ng since at the worst all Free
node n,). In B F , the node selected by expansion is nodes (cells) including ng are investigated in order of
the one having the lowest h ( ) ,and when two paths h ( ) .The combination search leads a huge set of visited
lead to the same node, the one with the higher h ( ) points or moving distances. For example, the set of
is discarded. However, because of no consideration visited nodes is described as gray area in Fig.2. There-
of the past cost, the number of nodes visited by B F , fore, this algorithm cannot be absolutely applied to an
which are described by a set of gray nodes (cells) in on-line or sensor-based path-planning algorithm.
Fig.2, is relatively smaller than the number of nodes
investigated by A*. 3.2 An algorithm A*
In this paragraph, we explain a classic A* algorithm
when the search space is a digital map described in
Fig.1. The difference against the B F algorithm is to
consider a past cost g(np)from a present node np to
a start node n,. An algorithm A* evaluates a node n p
by summing a past cost g(np)and a future cost h(np)
as the sum f(np),and consequently it always gets an
optimal (shortest) path between start and goal nodes.
However, the number of nodes visited by A*,which are
described by gray nodes in Fig.3, is relatively larger
than the number of nodes investigated by B F .
Figure 2 : A selected path and its visited nodes (cells) in B F ,
1. Put a start node n, on OPEN.
which are colored by black and gray in a 2-D unknown digital map, 2. If OPEN is empty, exit with failure.
respectively.
3. Remove from OPEN and place on CLOSED
1. Put a start node n, on a list called OPEN of a node np which f ( n p )is minimum (break ties arbi-
unexpanded nodes (cells). trary).
2. If OPEN is empty, exit with failure; no solution 4. If n p is a goal node n g ,exit successfully with a
exists. shortest (optimal) path obtained by tracing back the

1701

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
 3.3 The algorithm Classl

In this paragraph, we explain the algorithm C l a s s l

whose search space is a 2-D digital map. In the algo-
rithm, a point robot n pfirstly goes straight to its goal
ng if it does not encounter any obstacle (a set of Obs
nodes), otherwise, it sets a hit node hj (a closest node
c j ) by np. Then, np traces a set of Obs nodes by
the clockwise or counter-clockwiseorder fixed already.
For example, in Fig.4, the counter-clockwise direction
is always selected a t a hit node hj. If np is closer to
ng than the closest node c j , cj is always reset by np.
Figure 3 : A selected path and its visited nodes (cells) in A * ,
which are colored by black and gray in a 2-Dunknown digital map, In addition t o this, if the direction from np to ng is
respectively. Free (that is, a few nodes from np to ng are F r e e ) ,
np leaves the encountered obstacle (the encountered
pointers from np to n,. set of Obs nodes) from a leave node l j . Then, np goes
5. Otherwise, expand np for selecting all Free and straight to ng again.
unvisited nodes ni neighboring np,and attach to them
pointers back to np. For every neighbor node ni:
(sa)If ni was not already on O P E N or C L O S E D ,
estimate h(ni) by the piecewiseL2 metric from ni to
ng (a lower bound of the shortest route between ni
+
and n g ) ,and calculate f ( n i ) = g(ni) h(ni) where
+
g(n4 = g(np) c(np,ni)and dn,> =O.
(5b) If n, was already on O P E N or C L O S E D ,
direct its pointer along the path yielding the lowest
g(ni).
(5c) If ni required pointer adjustment and was
found on C L O S E D , remove from C L O S E D and
Figure 4 : A selected path (a sequence of visited nodes) in Classl
place on O P E N a node ni. with the counter-clockwise order, which is colored by black in a 2-D
unknown digital map.
6. Go to step 2.
Whenever h(n,) is an optimistic estimate of
h*(n,), an algorithm A* always returns an optimal
(shortest) path. h*(np)is length of an optimal route
between np and n g ,and h(n,) is an estimated length
of the route. This motivates naming this class of esti-
mates: Admissible heuristics [lo].
Definition: A heuristic function h is said to be
admissible if h(np)<h*
(np)Vnp.
In A * ,we always get an optimal (shortest) sequence
of Free nodes from n, to ng since h(np)is calculated
on the straight segment between np and n g ,which is a Figure 5 : How to find a neighbor Free node ni around a set
lower bound of their shortest route. It is measured by of neighbor Ob, nodes by the clockwise or counter-clockwise order
based on the chain coding.
the piecewiseL2 metric. In pursuit of the optimality,
A* needs the number of visited nodes and the sum of 1. Set j = 1, and set a robot node np and a leave
moving distances, which are larger than those of B F . node l j - 1 and a closest node cj-1 by a start node n,.
For example, a larger set of visited nodes is shown as 2. Select a neighbor node ni of n p ,which the seg-
gray region in Fig.3. Therefore, the algorithm cannot ment between centers of np and ng intersects in a dig-
be absolutely applicable for an on-line or sensor-based ital map.
path-planning algorithm. (2a) If ni is a goal node n g ,exit with success.

1702

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
 (2b) If n, is Free, set np by n, and then be con-
tinued, otherwise, set a hit node h, and a closest node
C,(J = J + 1) by np and go to step 3.
3. Select a neighbor Free node n, of np by the
clockwise or counter-clockwise order based on the
chain coding, which is adjacent to a neighbor Obs node
until one of the following occurs (Fig.5).
(3a) If n* is a goal node ng,exit with success.
(3b) If n, is closer to ng than the closest node c,,
reset c, by n,. Then, if the direction from n, to n9 is
Free (that is, a few nodes from n, to ng are F r e e ) ,
regard n, as a leave node 1, and return to step 2.
(3c) If nt equals to the last hit node h,, exit with Figure 6 : A selected path and its visited nodes in HD - I with
Q = 0.03, which are colored by black and gray in a 2-D unknown
failure. digital map, respectively.
In the algorithm Classl, leave and hit nodes (1,
robot to go straight to a goal. Another procedure is
and h3) monotonously come close to n9, and conse-
quently np arrives at ng finally if a path (a sequence for a robot to avoid an uncertain obstacle. In the algo-
rithm H D - I , the former procedure and the switching
of Free nodes) between n, and ng exists in a given
map. Otherwise, a point robot np automatically rec- are designed by the algorithm Classl, and the latter
procedure is designed by the modified B F .
ognixes no solution (path) between ns and ng in an
unknown map by returning the last hit node h,. Especially in the latter procedure, by using the
In every previous sensor-based algorithm with the modified B F , a point robot can trace 3-D boundary of
metric condition, e.g. Classl, np makes only efforts an encountered obstacle systematically without gener-
to find a closer node n, to ng than c, . In other words, ating any limit cycle and overlooking a leave ,node l j
np does not consider any past cost from npto n,. For which is closer to ng than a closest node c j . B F used
this reason, similar to B F , every sensor-based path- in H D - I differs from an original B F in three follow-
planning algorithm including Classl cannot select an ing aspects.
optimal path in an unknown map. However, compar-
ing with A* or B F , the number of visited nodes or the
sum of moving distances is small enough. For exam-
ple, a set of visited nodes is described as a sequence of
black nodes in Fig.4. In every sensor-based algorithm,
the set corresponds to a selected path itself. Therefore,
every previous algorithm can be undoubtedly used for
an on-line and sensor-based algorithm. '\

Finally, we should note that all the previous sensor- ne

based path-planning algorithms including Classl can-
not be applicable for three or more dimensional search
Figure 7 : How to find a neighbor Free node n, of a hit node h,,
space because a point robot np has the infinite num- which is adjacent to its neighbor Obs node in a 2-D digital map.
ber of directions to trace an encountered obstacle. To
overcome this drawback, we propose a sensor-based (1)H D - I uses B F in order for np to trace an en-
path-planning algorithm H D - I for searching three
or more dimensional space in the next.
9
untere obs acle w ose dimensio is t h e or more
%at is, a robot noie npmeets a k t node around
an uncertain set of Obs nodes, B F starts (begins
to input a found or expanded node into OPEN or
3.4 The algorithm H D -I CLOSED), and if np reaches a leave node 1, around
it, B F ends (cleans O P E N and CLOSED).
In this paragraph, we focus on B F and Classl (2) HD- I uses B F in order for npto trace bound-
which do not consider a past cost g(n,) from a start ary of an unknown obstacle, i.e., an uncertain set of
node n, to a present node np. Then, we design a three
Obs nodes. First of all, in 2-D and 3-D digital maps,
the number of nodes neighboring np is denoted by
or more dimensional sensor-based path-planning algo- 32 - 1 and 33 - 1, respectively. Secondly, the tracing
rithm H D - I by mixing modified B F and Classl. As starts if a robot node np finds a Obs node as its neigh-
mentioned previously, the sensor-based path-planning bor ni on the segment between np and ng. Therefore,
consists of two procedures. One procedure is for a npis Free and n, is Obs. In this case, there is at least

1703

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
 BF A' Class1 HD-I
N S P N S P N S P N S P
Clockwise 30633 2359953 1308 186384 22497804 922 2965 3155 3155 1307 2466 1365
C-Clockwise - - - - - - 2114 2313 2313 - - -

Table 1 : N , S,and P of B F , A * , Classl, and H D - I in a 2-D uncertain digital map.

A route between
suming, and on the other side, the moving cost m will
be used as a rough criterion because a is usually less
1
than 0.5. For this reason, in this paper, the shortest
route is approximated by a shorter route selected by
BF
Comparing with Classl, H D - I can be straight-
forwardly applied for an unknown map with three
'Configuration
'Configuration
space obstacle space obstacle or more dimensions. As a result, we can get a
\ / \ / three or more dimensional sensor-based (on-line) path-
planning algorithm.
1. Set j = 1, and set a robot node np and a leave
node l j - 1 and a closest node cj-1 by a start node n,.
Figure 8 : (a) Two routes are independently selected by tracing 2. Select a neighbor node ni of a robot node np,
back the pointers from np and n* to h j . (b) A route is obtained by which the segment between np and n, intersects in a
comparing two routes and eliminating their common part and then
combining the rest parts. digital map.
(2a) If ni is a goal node n,, exit with success.
(2b) If ni is Free, set np by n, and then be con-
tinued, otherwise, set a hit node hj and a closest node
cj(j =j + 1) by np and go to step 3.
3. Put the node hj on a list called O P E N of un-
expanded nodes.
4. If O P E N is empty, exit with failure; no path
e x i p.
Remove from O P E N a node np a t which q is
minimum (break ties arbitrary), and place it on a list
called C L O S E D to be used for expanded nodes.
6 . Select each neighbor Free node ni of n p ,which
neighbors on the set of 06s nodes (Fig.7). For every
Figure 8 : Trial and error occurs in selection of clockwise and neighbor ni:
counter-clockwise orders around each hit node h, when CY in H D - I (sa) If any ni is a goal node n,, exit successfully.
is small enough (a= 0.03). A set of visited nodes is described by
light gray, and a selected path is colored by dark gray. (6b) If ni is closer to ng than the closest node cj,
reset cj by ni and if a few nodes from ni to ng are
one Free node neighboring n p , which neighbors the Free, regard ni as a leave node l j . Then, clear O P E N
8bgfl.b" nodes. All the Free nodes are inserted into and C L O S E D and return to step 2.
(3) To select a hopeful node n* from O P E N , we ( 6 c ) If ni was neither on O P E N nor on C L O S E D ,
evaluate each node n* by summing a moving cost register it to O P E N and generate ni with the pointers
m(n*)from np to n* and a future cost h(n*)from back to np.
n* to n,. The future cost is straightforwardly cal- (6d) Evaluate each node n* in O P E N as fol-
culated by the piecewiseLz metric along the segment lows: Firstly, calculate a future cost h(n*)by the
between n* and n,. On the other hand, the moving piecewiseL2 metric along the segment between n* and
cost is defined by multiplying cy and the length of a n,. Secondly, calculate a moving cost m(n*)by the
shorter route between np and n*. The route is se- piecewiseL2 metric along a selected route between np
lected in B F as follows: Firstly, two routes are inde- and n* (Fig.8). Finally, assign q(n*)= a x m ( n * ) +
pendently obtained by tracing back the pointers from h(n*)(O< a < 1) to each node n' in O P E N , and
np and n* to hj (Fig.8(a)). Secondly, a selected route then return to step 4.
is determined by comparing two routes and eliminat- In the algorithm H D - I , leave and hit nodes ( l j
ing their common part and then combining the rest and hj) monotonously converge to a goal node (n,)
parts (Fig.g(b)). Finally, the selected route is mea- in a 2-D or 3-D digital map. In addition, n p always
sured by the piecewiseL2 metric. The selected route finds an adequate leave node l j around an encountered
is not always to be an optimal (shortest) route between set of Obs nodes as long as a sequence of Free nodes
np and n*. To pick up the optimal route generally, we from n, to n, exists in a given digital map. For this
should use A*. However as you see, A* is time con- reason, if such a sequence exists, npcertainly arrives a t

1704

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
 BF A* HD-I
N S I P
I N I S I P N l S l P
1162 1 10563 I 187 77058 I 1528502 I 138 1082 I 1930 I 777

able 2 : N , S,and P of B F , A * , and H D - I in a 3-D uncertain digital map.

Figure 10 : (a), (b), (c) Selected paths and their sets of visited nodes in B F , A * , and H D -I. Each path and set are colored by dark and
light gray in a 3-D uncertain digital map, respectively.

ng,otherwise, np sees no path since O P E N becomes adequate direction globally. In conclusion, a set of
empty. visited nodes widely spreads out around boundary of
an unknown obstacle at hit nodes, which is colored by
light gray in Fig.9. As a result, even though an en-
vironment has complex shape, the algorithm H D - I
4 Simulation Results whose a is small enough ( a = 0.03 in Fig.6, 9, Tablel)
is the best in the on-line path-planning.
In this section, we compare the number N of vis- For the simplification, S of BF and A* is given
ited nodes, the sum S of moving distances, and the by a lower bound of the true moving distances. The
length P of a selected path in a 2-D digital map de- reason is that each distance between np and n* is
scribed in Fig.2, 3, 4 and 6 by B F , A*, Classl and easily calculated as the distance of their straight- seg-
H D - I , respectively (Table 1). Firstly, the 2-D dig- ment. The distance is always smaller than or equals
ital map is given by [2', 29] (1=9, m=2), and includes to the distance of their shortest route. This prop
262144 nodes totally, 32368 Obs nodes, and 229776 erty is true even in the piecewiseL2 metric. As con-
Free nodes. Now, our questions concerning to path- trasted with this, S of H D - I is calculated as an
planning are how many nodes a point robot visits in upper bound of the true moving distance. The reason
the map and how long the robot moves in it. If the is that each distance between np and n* is calculated
number N is small enough, the algorithm is regarded as the piecewiseL2 metric along a satisfactory route
as the on-line algorithm, otherwise, it is considered as selected in H D - I, which was explained in the last
the off-line algorithm. In addition, if and only if the section (Fig.8). The route is always longer than or
value S is small enough, the algorithm is regarded as equals to the shortest route between np and n*. In
the sensor-based algorithm. addition, S of Classl is calculated as the true moving
As shown in Table 1, N of B F or A* is extremely disfpce.
larger than N of Class1 or H D - I with a = 0.03, i.e., evertheless, S of BF or A* is extremely larger
NA* >> NBF >> Nclassl > N H D - I . Therefore, the than S of Classl or H D - I , i.e., SA+ > SBF >>
former two are off-line, and the latter two are on-line. S c l a s s l = S ~ ~ Therefore,
-~. we can see that the al-
Here, the inequality Nclassl > N H D - I is kept. The gorithms Classl and HD - I are better ones as the
reason is that a robot supervised by Classl fixes its 2-D sensor-based path-planning algorithm. Finally, as
direction following an encountered obstacle and conse- shown in Table 1, P obtained by A* is the shortest,
quently sometimes loses the way to its destination in and P selected by B F , Classl, or H D - I nearly
a complicated uncertain environment shown in Fig.4. equals, i.e., PA+ < P B F M P c ~ ~ ~ ~ ~ MFor P Hthis
D-I.
On the other hand, a robot supervised by H D - I with reason, A* is the best algorithm in the 2-D off-line
a small a always selects an adequate direction at a hit (model-based) path-planning.
node hj by trial and error concerning to selection of Secondly, we compare the number N of visited
clockwise or counter-clockwise direction (Fig.6 and 9). nodes, the sum S of moving distances, and the length
Especially in Fig.9, due to the trial and error, a point P of a selected path in a 3-D digital map (Table 2).
robot locally changes clockwise and counter-clockwise Selected paths and visited sets by B F , A*, and H D - I
directions around two hit nodes in order to find an with a = 0.5 are described in Fig.lO(a), (b), and (c).

1705

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.
 Figure, 11 . (a), (b), (c), (d) A sequence"of manipulator motions made in H D - I.
Firstly, the 3-D digital map is given by [a7,27, 27] (1=7, References .
m=3), and includes 2097152 nodes totally, 396579 Obs
"
nodes, and 1700573 Free nodes. [l]H.Noborio, "On, a sensor-based navigation for a mobile
As shown in Table 2, N of A* is extremely larger robot," Journal of Robotzcs Mechatronzcs, Vo1.8, No.1, pp.2-
than N of B F or H D - I, i.e., .NA* >> NBF > 14, 1996
N H D - I . Therefore, the former algorithm is off-line, [2] V.J.Lumelsky, "Effect of Kinematics on motion planning for
and the latter two algorithms are on-line. Shape of planar robot arms moving amidst unknown obstacles," IEEE
the unknown obstacle is not so complex, and there- Journal of Robotzcs and Automatzon, Vo1.3, No.3, pp.207-
fore B F is not so bad as an on-line algorithm, which 223, June 1987.
sequentially picks up a closest node to ng. On the [3]V.J.Lumelsky and K.Sun, "A unified methodology for m e
other hand, S of A* or B F is extremely larger than tion planning with uncertainty for 2D and 3D two-link robot
S of H D - I , i.e., SA' >> SBF >> S H D - I . There- arm manipulators," The Internatzonal Journal of Robotzcs
fore, we can see that the algorithm H D - I with a Research, Vo1.9, No.5, pp.89-104, October, 1990.
large a is the best algorithm in the 3-D sensor-based
[4] K.Sun and V.J.Lumelsky, "Path planning among unknown
path-planning (Table 2). Especially in an unknown
obstacles: The case of a three-dimensional Cartesian arm,"
map whose obstacles are simple shape, a robot super- IEEE Trans. o n Robotzcs and Automatzon, Vo1.8, No.G,
vised by H D - I with a large a selects a satisfactory pp. 77G-78G, December, 1992.
path globally without many trial and error at each
hit point h, locally ( a = 0.5 in Fig.lO(c) and Table [5] E.Cheung and V.J.Lumelsky, "Real-time path planning pro-
2). As a result, the algorithm H D -! with a large cedure for a whole-sensitive robot arm manipulator," Robot-
aca, Vol.10, pp.339-349, 1992.
a is the best sensor-based path-planning algorithm.
Finally, the length P given by A* or B F is the short- [GI V.J.Lumelsky and E.Cheung, "Real-time collision avoidance
<< P H D - I . For this
est or shorter, i.e., P A * ~ P B F in teleoperated whole-sensitive robot arm manipulators,"
reason, A* and B F are regarded as the 3-D off-line IEEE Trans. on System, Man, and Cybernetacs, Vo1.23,
path-planning algorithm in such a map with simple No.1, pp.194-203, January/February, 1993.
shape. [7]H.Noborio, "A hyperplane-based path planning algorithm
Finally, the used 3-D map is given as a configura- for a four-degrees-of-freedom robotic manipulator in a n un-
tion space of a three-degrees-of-freedom manipulator. certain<workspace," Proc. of the USA-Japan Symposzum o n
Therefore, a selected path corresponds to a sequence of Flexable Automatzon, pp.95-98, July 1992.
manipulator motions described in Fig.11, which is gen- [8]H.Noborio, T.Hamaguchi, and T.Yamamoto, "Experiments
erated by H D - I. All programs are running in a per- with a 3DD robotic manipulator operating by a sensor-based
sonal computer Gateway GPG-400 (Windows 98 with assembly," PTOC.of the IEEE Internatzonal Symposaum on
Open' GL, Pentium I1 400MHz, 25GMB main mem- Assembly and Task Plannzng, pp.249-256, August 1997.
ory). (91 H.Noborio, M.Kadowaki, and K.Urakawa, "A near-optimal
sensor-based motion-planning algorithm for parts mating,"
PTOC.of the IEEE/RSJ Int. Conf. on Intellagent Robots and
Systems, pp.510-517, October 1998.
5 Conclusions [lo] N.Nillson, "Principle of artificial intelligence", Tzoga Pub-
lzshang Co., 1980.
In this paper, we propose a sensor-based path-
planning algorithm for investigating an uncertain envi- [ll]J. Pearl, "Heuristics", Addzszon Wesley, October 1985.
ronment whose dimensions are three or more. The key [12] H.Noborio, "A sufficient condition for designing a family of
idea of our new algorithm is mixing two modified algo- sensor-based deadlock-free path-planning algorithms," Jour-
rithms B F and Classl. Especially based on a mod- nal of Advanced Robotzcs, Vo1.7, No.5, pp.413-433, January
ified B F algorithm, a point robot is flexibly able to 1993.
avoid boundary of an unknown obstacle whose direc- [13] J.-C. Latombe, "Robot motion planning", Chapter 6: Ap-
tions are over two. Therefore, the proposed algorithm prommate Cell Decomposataon, Kluwer Academzc Publzsh-
can be used for manipulation or assembly by a robotic ers., 1991.
manipulator. By several simulation results, usefulness
and efficiency of our algorithm are ascertained.

1706

Authorized licensed use limited to: Cranfield University. Downloaded on May 02,2021 at 02:46:55 UTC from IEEE Xplore. Restrictions apply.