Robotics and Autonomous Systems 34 (2001) 179202

Classification of the Dubins set

Andrei M. Shkel a, , Vladimir Lumelsky b
a

Department of Mechanical and Aerospace Engineering, University of California, 4208 Engineering Gateway Building,
Irvine, CA 92717-3975, USA
b Department of Mechanical Engineering, University of Wisconsin-Madison, Madison, WI 53706, USA
Received 14 August 1998; received in revised form 13 November 2000
Communicated by T.C. Henderson

Abstract
Given two points in a plane, each with a prescribed direction of motion in it, the question being asked is to find the
shortest smooth path of bounded curvature that joins them. The classical 1957 result by Dubins gives a sufficient set of paths
(each consisting of circular arcs and straight line segments) which always contains the shortest path. The latter is then found
by explicitly computing all paths on the list and then comparing them. This may become a problem in applications where
computation time is critical, such as in real-time robot motion planning. Instead, the logical classification scheme considered
in this work allows one to extract the shortest path from the Dubins set directly, without explicitly calculating the candidate
paths. The approach is demonstrated on one of two possible cases that appear here when the distance between the two
points is relatively large (the case with short distances can be treated similarly). Besides computational savings, this result
sheds light on the nature of factors affecting the length of paths in the Dubins problem, and is useful for further extensions,
e.g. for finding the shortest path between a point and a manifold in the corresponding configuration space. 2001 Published
by Elsevier Science B.V.
Keywords: Computational geometry; Geometric algorithms; Shortest path problems; Robotics; Nonholonomic motion

1. Introduction
Consider the problem of finding the shortest smooth path between two points in the plane, the initial and final
points, Pi and Pf . Each point is associated with its own orientation angle, and , respectively, which defines the
prescribed direction of motion in it (see Fig. 1). The combinations (Pi , ) and (Pf , ), called the initial and final
configurations, define two points in the corresponding configuration space (C-space), and present the problems
boundary conditions. Given (Pi , ) and (Pf , ), the task is to find the shortest smooth path from Pi to Pf , such
that it starts and ends with the directions of motion and , respectively, and the path curvature is limited by 1/,
where is the minimal radius of turning.
This kind of tasks appear in various applications, such as when joining pieces of railways [1] or planning two- and
three-dimensional pipe networks. In robotics, this problem plays a central role in most of the work on nonholonomic
motion planning [24].

This work was supported by the National Science Foundation Grant IRI-9220782 and the Sea Grant Program (National Oceanic and Atmospheric Administration, US Department of Commerce, Grant NA46RG048).
Corresponding author.
E-mail addresses: ashkel@uci.edu (A.M. Shkel), lumelsky@robios.me.wisc.edu (V. Lumelsky).
0921-8890/01/$ see front matter 2001 Published by Elsevier Science B.V.
PII: S 0 9 2 1 - 8 8 9 0 ( 0 0 ) 0 0 1 2 7 - 5

180

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 1. The coordinate system, the initial configuration (Pi , ) and the final configuration (Pf , ). Possible orientation angles are divided into
four quadrants.

The complete solution to this problem was first reported in an elegant paper by Dubins [5] in 1957. He showed
that any geodesic (i.e. the shortest path) consists of exactly three path segments and presents a sequence CCC or
CSC, where C (for circle) is an arc of radius , and S (for straight) is a line segment. Each arc C has two
options turning left or turning right. Denote those L and R, respectively, and the line segment by S. The Dubins
set, D, includes six admissible paths (or words), D = {LSL, RSR, RSL, LSR, RLR, LRL}. Furthermore, Dubins
theorem states that in order to be a candidate for the optimal path, each arc must be of the minimal allowed radius .
Using advanced calculus, this result of Dubins was later proved by Reeds and Shepp [6]. Also, Boissonnat et al.
[7] proved this result from the standpoint of optimal control, by making use of the powerful Pontryagins optimality
principle [8]. The more difficult case in which the path from (Pi , ) to (Pf , ) can be further shortened by allowing
reversals of motion (and thus introducing cusps) was first considered in the same work by Reeds and Shepp [6].
They showed that the initial and final configurations define a sufficient set of 48 paths which contains the optimal
path. The technique presented in [9] allows one to pick the optimal solution out of this set of 48 by partitioning the
C-space into multiple domains such that a single path type is associated with 150 elements and two path types are
associated with the other 11 elements.
An alternative approach to the problem with reversals was proposed by Soueres and Laumond [10]. They tie the
Pontryagins optimality principle with geometric reasoning, and arrive at the optimal solution via partitioning of
C-space into regions with uniform properties of path optimality.
In the context of robotics, the original Dubins problem of constructing a smooth path has a significance of its own.
In many motion planning tasks, such as in the aircraft control, motion reversals are not feasible. Or, if the shortest
time path, rather than the shortest path, is desired, the solution is likely to be a smooth path, because the deceleration,
stop, and acceleration at the reversal cusps add time to the path execution. Unfortunately, Dubins problem with
smooth paths is not a subset of the ReedsShepps problem the sufficient set of the former is not contained in the
sufficient set of the latter. Also, the techniques proposed in [9,10] are not directly applicable to the smooth path case.
To use Dubinss result for the shortest path calculation, one would need to explicitly calculate the lengths of
all arcs and straight line segments in the Dubins set, and then choose the shortest of the computed paths. The
time necessary for this calculation may become a bottleneck in time-constrained applications, as e.g. in real-time
robot motion planning which is one motivation for this work. Another motivation comes from problems where
one looks for the shortest path from a point to a manifold in the C-space. For example, in sensor-based obstacle
avoidance, when planning an arrival to some intermediate point P on the obstacle boundary, the current sensing
data may suggest that in order not to collide with the obstacle, the orientation angle at P must be within some
sector of angles (which may include, e.g. the tangent to the obstacle at P ). Finding the shortest path to P under this
constraint corresponds to finding the shortest path to a line in C-space.
In this work, we propose a scheme which allows one to select the shortest path from the Dubins set D directly,
without the usual exhaustive calculation of its elements. The scheme is based on a rather suggestive fact, developed
in Section 5, that the elements of the Dubins set can be classified into a small number of the so-called equivalency
groups, based on the angle quadrants of the corresponding pairs of the initial and final orientation angles. Each
equivalency group consists of a few classes of paths, such that any path in a group is equivalent, up to an orthogonal

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

181

transformation, to any other path in the same group. This means that the optimal path analysis can be reduced to
fewer terms. Further, a simple logical classification of the equivalency groups can be built which points directly to
the optimal path.
Below, d is the Euclidean distance between the initial and final points Pi and Pf . A rectangular coordinate
system (x, y) is chosen such that its origin is Pi = (0, 0) and the positive direction of x-axis is toward Pf = (d, 0)
(Fig. 1). The initial and final orientation angles, and , are measured counter-clockwise, with respect to the
positive direction of x-axis. Without loss of generality, assume a unit radius of the minimum turning circle, = 1
(any other can be reduced to 1 by the scaling d = D/, where D is the actual distance between Pi and Pf ). The
initial and final arc segments (of radius ) in the Dubins set are denoted Cil , Cir , Cfl , Cfr (where i and f stand for
initial and final, and r and l for right and left).
The analysis necessary for solving our classification problem turns out to become simpler if it is divided into two
cases, which can be called the long path case and the short path case. Our approach is equally applicable to both
cases, with minor differences between the resulting computational schemes. For the sake of example, we consider
here only one case, the long path case, which seems to be of more interest from the standpoint of applications and
the computational savings. More precisely, the long paths are those where the distance d between the points Pi
and Pf satisfies the condition of non-intersection of the four circles above, {Cil Cir } {Cfl Cfr } = . This covers
all cases when d > 4 and some cases when d < 4 (see Fig. 4 and Proposition 5).
We first develop, in Section 2, a proper specification scheme for admissible paths. The notion of an equivalency
group is then introduced in Section 3, and the scheme for classifying the Dubins set is fully developed in Section
4. This works main result which makes this classification possible and becomes the logical scheme for finding the
shortest path is summarized in Section 4.

2. Admissible paths and their specification

Given a path from the initial to the final configuration, the position of a point on the path is fully specified by
its Cartesian location x( ), y( ), where the parameterization variable can be interpreted as time or the length of
path traversed from Pi with unit velocity. Assume
i = 0, i.e. Pi = P (i ) = P (0);
the point on the path can move only forward, from Pi toward Pf ;
it moves with the unit speed;
the orientation angle (direction of motion) cannot change faster than 1/ radian per time unit,
where = 1 is the minimal turning radius.
Following [5], an admissible path is defined as a continuously differentiable curve which is either (i) an arc of a
circle of radius 1, followed by a line segment, followed by an arc of a circle of radius 1, or (ii) a sequence of three
arcs of circles of radius 1, or (iii) a subpath of a path of type (i) or (ii). A list of admissible paths forms a sufficient
set of optimal paths.
To specify admissible paths, we introduce three elementary motions: turning to the left, turning to the right
(both along a circle C of radius 1), and straight line motion S. Also needed will be three corresponding operators,
Lv (for left turn), Rv (for right turn), Sv (for straight), which transform an arbitrary point (x, y, ) R3 into its
corresponding image point in R3 ,
Lv (x, y, ) = (x + sin( + v) sin , y cos( + v) + cos , + v),
Rv (x, y, ) = (x sin( v) + sin , y + cos( v) cos , v),
Sv (x, y, ) = (x + v cos , y + v sin , ),

(1)

where index v indicates that the motion has been along the (C or S) segment of length v. With these elementary
transformations, any path in the Dubins set D = {LSL, RSR, RSL, LSR, RLR, LRL} can be expressed in terms of

182

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

the corresponding equations. In the coordinate system chosen, the initial configuration of each path is at (0, 0, )
and the final configuration at (d, 0, ). For example, a path made of segments L, R and L, of the lengths t, p, q,
respectively, which starts at point (0, 0, ), must end at Lq (Rp (Lt (0, 0, ))) = (d, 0, ). The length L of the path
can be defined as the sum of lengths t, p and q of its constituent segments,
L = t + p + q.

(2)

Our goal is to classify the elements of set D based on the boundary conditions, with the purpose of replacing the
explicit computation of all candidates for the shortest path with a simple logical procedure that would directly
produce the shortest path. To this end, we will now consider elements of D one-by-one and derive the operator
equations for the length of each path.
1. Lq (Sp (Lt (0, 0, ))) = (d, 0, ). By applying the corresponding operators (1), this first path in D can be
represented by a system of three scalar equations:
p cos( + t) sin + sin = d,
p sin( + t) + cos cos = 0,
+ t + q = {mod 2}.
The solution of this system with respect to the segments t, p and q is found as
tlsl = + arctan
plsl =

cos cos
{mod 2 },
d + sin sin

p
2 + d 2 2 cos( ) + 2d(sin sin ),

qlsl = arctan

(3)

cos cos
{mod 2}.
d + sin sin

Using definition (2), the length of the path LSL as a function of the boundary conditions can be now written as
Llsl = tlsl + plsl + qlsl = + + plsl .

(4)

2. Rq (Sp (Rt (0, 0, ))) = (d, 0, ). Using (1), we obtain the corresponding scalar equations:
p cos( t) + sin sin = d,
p sin( t) cos + cos = 0,
t q = {mod 2}.
The solution of this system, i.e. the lengths of the corresponding segments, is
trsr = arctan
prsr =

cos cos
{mod 2},
d sin + sin

p
2 + d 2 2 cos( ) + 2d(sin sin ),

qrsr = (mod 2) + arctan

(5)

cos cos
{mod 2},
d sin + sin

and the path length is given by

Lrsr = trsr + prsr + qrsr = + prsr .

(6)

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

183

3. Rq (Sp (Lt (0, 0, ))) = (d, 0, ). Using (1), we obtain the corresponding scalar equations:
p cos( + t) + 2 sin( + t) sin sin = d,
p sin( + t) 2 cos( + t) + cos + cos = 0,
+ t q = {mod 2}.
The solution of this system is


tlsr

cos cos
= + arctan
d + sin + sin

plsr =

2
arctan
plsr

{mod 2 },

p
2 + d 2 + 2 cos( ) + 2d(sin + sin ),


qlsr



cos cos
= (mod 2) + arctan
d + sin + sin

(7)


2
arctan
plsr


{mod 2},

and the path length is given by

Llsr = tlsr + plsr + qlsr = + 2tlsr + plsr .

(8)

4. Lq (Sp (Rt (0, 0, ))) = (d, 0, ). Using (1), we obtain the corresponding scalar equations:
p cos( t) 2 sin( t) + sin + sin = d,
p sin( t) + 2 cos( t) cos cos = 0,
t + q = {mod 2}.
The corresponding solution is

trsl = arctan
prsl =

cos + cos
d sin sin


+ arctan


{mod 2},

p
d 2 2 + 2 cos( ) 2d(sin + sin ),


qrsl

2
prsl

cos + cos
= (mod 2) arctan
d sin sin

(9)


2
+ arctan
prsl


{mod 2},

and the path length is given by

Lrsl = trsl + prsl + qrsl = + + 2trsl + prsl .
5. Rq (Lp (Rt (0, 0, ))) = (d, 0, ). Using (1), we obtain the corresponding scalar equations:
2 sin( t + p) 2 sin( t) = d sin + sin ,
2 cos( t + p) + 2 cos( t) = cos cos ,
t + p q = {mod 2}.

(10)

184

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

The solution of this system is



prlr
cos cos
+
{mod 2 },
trlr = arctan
d sin + sin
2
prlr = arccos 18 (6 d 2 + 2 cos( ) + 2d(sin sin )),

(11)

qrlr = trlr + prlr {mod 2 },

and the path length is obtained by substituting (11) into (2),
Lrlr = trlr + prlr + qrlr = + 2prlr .

(12)

6. Lq (Rp (Lt (0, 0, ))) = (d, 0, ). Using (1), we obtain the corresponding scalar equations:
2 sin( + t p) + 2 sin( + t) = d + sin sin ,
2 cos( + t p) 2 cos( + t) = cos + cos ,
+ t p + q = {mod 2}.
The corresponding solution is




cos + cos
plrl
+
{mod 2 },
tlrl = + arctan
d + sin sin
2
plrl = arccos 18 (6 d 2 + 2 cos( ) + 2d(sin sin )){mod 2 },

(13)

qlrl = (mod 2) + 2plrl {mod 2 },

and the path length is given by
Llrl = tlrl + plrl + qlrl = + + 2plrl .

(14)

3. Equivalency groups
We are now prepared to turn to the classification of the Dubins set D. Divide the range of possible orientation
angles (, ) into four quadrants; Fig. 1: quadrant 1 corresponds to the range [0, /2], quadrant 2 to [/2, ],
quadrant 3 to [, 3/2], and quadrant 4 to the range [3/2, 2 ]. Since each of or can be in any of the four
quadrants, together this produces 16 different combinations of possible quadrants. We represent those 16 by a 4 4
matrix, {aij }, where index i corresponds to the quadrant number of the initial, and index j that of the final orientation.
Element aij therefore describes the class of all paths whose initial and final orientation angles (, ) belong to the
quadrants i and j , respectively. For example, the case [0, /2], [/2, ] corresponds to the element a12
and covers all those paths whose orientation angles belong to the first and second quadrants, respectively.
It will be shown below that these 16 classes can be reduced to six independent clusters, called equivalency groups,
such that an orthogonal transformation of any path in a given group changes it into a path in the same or a different
class of the same group.
Dubins main theorem [5] says that each (non-degenerate) candidate for the optimal path in set D must start with
a piece of circle and end with a piece of circle (of radius = 1, see above). Depending on the path in D, the initial
and the final circle can turn either left or right; we denote those Cil , Cir and Cfl , Cfr , respectively. To proceed, we
will need the following definition.
Definition. Two paths are topologically equivalent (denoted by ') if there exists an orthogonal transformation
that maps one path into the other, with both paths sharing their respective initial and final points.

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

185

Fig. 2. Arcs Pi P2 and P3 Pf are obtained by orthogonal transformations of the arc Pi P1 : the first is a mirror reflection of Pi P1 with respect to
line (Pi Pf ), the second the central symmetry reflection of Pi P1 with respect to point O. Arc P4 Pf is obtained by applying a composition of
both transformations.

Note that two equivalent paths are of the same length and allow equivalent parameterization. The following
proposition relates the topological equivalency of paths and their initial/final configurations.
Proposition 1. For any path connecting two points, P (ti , ), P (tf , ), where (, ) are the initial and final
orientation angles, there exist another three paths which are topologically equivalent to it. Their corresponding
orientation angles are (, ), (, ), and (, ).
To see this, consider a path that starts at an initial configuration (Pi , ) and is of the form Pi P1 [, ];
here the ellipsis reflect our emphasis on the segment Pi P1 (Fig. 2). By applying a mirror reflection G(Pi Pf )
with respect to the line (Pi Pf ), this path is transformed into the path Pi P2 [, ]. Similarly, by applying the
central symmetry reflection G(O) with respect to the midpoint O of segment [Pi , Pf ], the same path is transformed
into P3 Pf [, ]. The composition of both transformations leads to
G(O) (G(Pi Pf ) (Pi P1 [, ])) = G(Pi Pf ) (G(O) (Pi P1 [, ])) = P4 Pf [, ].

(15)

This general fact will be used below in the analysis of paths defined by set D. Recall that those paths take a form
either CCC or CSC, where C is an arc of a circle of radius with options L and R (left and right), and S is a straight
line segment. To distinguish between the first and the second arc segments in the path CSC, subscripts will be used,
C1 SC2 .
Define the conjugate of C1 , denoted C 1 , as the complement of C, i.e. if C1 = R then its conjugate is C 1 = L,
and vice versa. The application of the orthogonal transformations G(Pi Pf ) and G(O) leads to
G(Pi Pf ) (C1 SC2 [, ]) = C 1 S C 2 [, ],

GO (C1 SC2 [, ]) = C 2 S C 1 [, ].

Notice that the mirror reflection G(Pi Pf ) reverses the signs of angles , and changes the arc segments to their
conjugates. The central symmetry reflection G(O) has a triple effect: it switches orientations and , switches segments C1 and C2 , and switches each segment to its conjugate. This relation can be proven rigorously by formalizing
the operators G(Pi Pf ) and G(O) and then applying them to the path presented in the general operator form, as e.g.
Lq (Sp (Lt (0, 0, ))) = (d, 0, ).
Independent of the order of transformations, the composition of G(Pi Pf ) and G(O) leads to
G(O) (G(Pi Pf ) (C1 SC2 [, ])) = G(Pi Pf ) (G(O) (C1 SC2 [, ])) = C2 SC1 [, ].
Using the above definition of topological equivalence, the following proposition defines the set of topologically
equivalent paths.

186

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 3. An illustration for Proposition 2.

Proposition 2. Given the path C1 SC1 [, ], the composition of orthogonal transformations G(Pi Pf ) and G(O) leads
to the topologically equivalent paths
C1 SC2 [, ] ' C 1 S C 2 [, ] ' C 2 S C 1 [, ] ' C2 SC1 [, ],
where , are the initial and final orientations, G(Pi Pf ) the mirror reflection with respect to line (Pi Pf ), G(O) the
central symmetry reflection with respect to the midpoint O of segment [Pi , Pf ], and ' the sign of equivalency.
Example. Consider the paths shown in Fig. 3. Initially there are two paths, RSR[, ] and RSL[, ] (Fig. 3(a)),
where [0, /2], [/2, ].
1. Consider first the path RSR[, ]. With the notation of Proposition 2, we have C1 = R and C2 = R. The
proposition gives the following set of topologically equivalent paths (see Fig. 3(a)(d)):
RSR[ [0, /2], [/2, ]] ' LSL[ [3/2, 2 ], [, 3/2]]
' LSL[ [/2, ], [0, /2]]
' RSR[ [, 3/2], [3/2, 2]].
2. Consider now an example C1 = R, C2 = L. By applying Proposition 2 to the path RSL[, ] with
[0, /2], [/2, ] (Fig. 3(a)), we obtain three other paths (Fig. 3(b)(d)):

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

187

RSL[ [0, /2], [/2, ]] ' LSR[ [3/2, 2], [, 3/2]]

' RSL[ [/2, ], [0, /2]]
' LSR[ [, 3/2], [3/2, 2 ]].
Proposition 3. Given the path C1 SC1 [, ], the individual orthogonal transformations G(Pi Pf ) , G(O) and
their composition lead to the topologically equivalent paths C1 SC2 [, ], C 1 S C 2 [, ], C 2 S C 1 [, ], and
C2 SC1 [, ], for which the following holds:
Lc1 sc2 = tc1 sc2 + pc1 sc2 + qc1 sc2 ,

Lc1 sc2 = tc1 sc2 + pc1 sc2 + qc1 sc2 ,

Lc2 sc1 = tc2 sc1 + pc2 sc1 + qc2 sc1 ,

Lc2 sc1 = tc2 sc1 + pc2 sc1 + qc2 sc1 ,

and
tc1 sc2 = tc1 sc2 = qc2 sc1 = qc2 sc1 ,

pc1 sc2 = pc1 sc2 = pc2 sc1 = pc2 sc1 ,

qc1 sc2 = qc1 sc2 = tc2 sc1 = tc2 sc1 .

It is convenient to combine Propositions 2 and 3 into one theorem.

Theorem 1 (Transformation Theorem). Given the path C1 SC1 [, ](t, p, q) with the lengths of the initial, middle and final segments equal to t, p, and q, respectively, the orthogonal transformations G(Pi Pf ) , G(O) and their
composition lead to the topologically equivalent paths
C1 SC2 [, ](t, p, q) ' C 1 S C 2 [, ](t, p, q) ' C 2 S C 1 [, ](q, p, t) ' C2 SC1 [, ](q, p, t).
The Transformation Theorem gives a linguistic rule for topologically equivalent transformations; it emphasizes
the structure of equivalent paths obtained as a result of these transformations. This theorem will be now used for
defining equivalency groups and thus reducing the amount of computations, namely the following statement holds.
Proposition 4. Matrix {aij } can be divided into six independent equivalency groups: (1) a11 ' a44 , (2) a12 ' a21 '
a34 ' a43 , (3) a13 ' a24 ' a31 ' a42 , (4) a14 ' a41 , (5) a22 ' a33 , and (6) a23 ' a32 .
Indeed, according to the Transformation Theorem, any path with [0, /2], [0, /2, ] (i.e. belonging
to class a11 ) is transformed into an equivalent path with [3/2, 2 ], [3/2, 2] (which is from class a44 ).
That is, the central symmetry reflection, G(O) , leads to a topologically equivalent path from class a11 , while the
composition of G(Pi Pf ) and G(O) leads to a topologically equivalent path from class a44 . In the case of the equivalency
group (2), for any path of class a12 there exists an equivalent path in each of the classes a21 , a34 , and a43 .
By choosing one representative from each equivalency group, we define a basis set B of matrix {aij } a list of six
mutually independent classes of orientation pairs , . This reduces from 16 to 6, the number of path classes to be analyzed for the optimal solution. Note that the basis set is not unique since its members can be chosen in various ways.

4. Classes of paths and their equivalency groups

The above scheme for classifying the Dubins set will be fully developed in this section; the necessary analysis
involves the following steps:
1. Find the necessary and sufficient condition of non-intersection of the unions {Cil Cir } and {Cfl Cfr }. This
condition formally defines what is meant by the long paths in the case under study.
2. Show that the condition 1, when satisfied, leads to a further simplification of the set of candidates for the optimal
solution that need be considered.

188

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

3. For every element of matrix {aij }, find the minimum number of the optimal path candidates.
4. For those elements aij which, as found in step 3, allow more than one candidate for the optimal solution, derive
the corresponding switching functions which uniquely define the optimal path.
Since elements from the same equivalency group have similar properties, the development steps 3 and 4 are
combined below for each of the groups, forming six corresponding subsections.
When applying results of this analysis to a specific problem, one would proceed as follows:
Make sure that the task at hand satisfies the condition in step 1.
Associate the given initial and final orientations with a class aij . For class aij , use the uniquely defined optimal
path.
Below, after analyzing steps 1 and 2, for the sake of convenience we choose for step 3, a particular example of basis
set B, B = {a11 , a12 , a13 , a14 , a22 , a23 }. For each of the six elements of B, questions posed in steps 3 and 4 are
then addressed in the respective six sections.
The following additional notation is used below: unless stated otherwise, forms like AB and AB represent
straight line segments and circular arc segments, respectively, with A and B being the segments endpoints. When
in mathematical expressions, the same forms denote the lengths of the segments. When needed for clarity, the strings
may be longer: e.g. A1 B1C1 D1 is an arc with the endpoints A1 , D1 and two inner points B1 , C1 .
4.1. The long path case
Turning to the step 1 above, the condition of non-intersection of union {Cil Cir } with union {Cfl Cfr } is as
follows.
p
Proposition 5. {Cil Cir } {Cfl Cfr } = if d > 4 (| cos | + | cos |)2 + | sin | + | sin |. This condition
on d is a precise definition of the long path case.
To prove this necessary and sufficient condition, consider the case when the union {Cil Cir } is tangent to the union
{Cfl Cfr }, i.e. there exist only one point belonging to both unions. Take, for instance, the case a11 ( and are in
the first quadrant, Fig. 4). Assume a unit radius, = 1. Given a common tangent to both arcs, IF = IA + AB + BF.
From 4IAO1 : IA = sin and O1 A = cos . From 4FBO
p 2 : BF = sin and O2 B = cos . From 4O1 O2 C:
O1 O2 = 2, O2 C = O2 B + O1 A, and therefore O1 C = 4 (cos + cos )2 . Summing up for IA + AB + BF,
obtain the expression for distance IF, which
is a condition for a common tangent for the right initial circle and left
p
final circle. In general, the expression 4 (| cos | + | cos |)2 + | sin | + | sin | = d covers all possible cases
of paths consisting of circular arcs with a common tangent point.
Proposition 6. For the long path case, the path CCC cannot be the optimal solution.
To show this, consider the basis set B of independent orientation pairs (, ). We need to show that for any
element from B, there exists a path of type CSC that is shorter than the path CCC.
Since for the long path case, {Cil Cir } {Cfl Cfr } = , then Cil , Cir , Cfl , Cfr do not intersect. Fig. 5 illustrates
this for the general case: it is clear in this example that though the path CCC is physically realizable, it can be
excluded from the list of candidates for an optimal solution.
To prove this, assume that circle Cleft is one of the initial circles Cil or Cir , and Cright is either Cfl or Cfr . Two
other circles tangent to Cleft and Cright are Cup (upper tangent) and Cdown (lower tangent). The initial and final
orientations can be chosen either clockwise or counter-clockwise. Notice that the orientation of the initial and final
circles has to be the same, since the path of type CCC switches directions when passing from one circle to another.
If two switchings take place, the initial and final circles must have the same orientation.

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

189

Fig. 4. The case of a path of two circular arcs with a common tangent point; here both orientation angles and are in the first quadrant.

The area of possible initial and final positions is limited; the first must lie on the arc A1 B1C1 D1 of circle Cleft ,
and the second on the segment A2 B2C2 D2 of circle Cright (both segments are shown in solid line in Fig. 5). These
restrictions are dictated by Proposition 5.
Let us say, the initial and final positions lie within the arcs A1 B1C1 D1 and A2 B2C2 D2 , respectively, both with
the same counter-clockwise orientations. (For the clockwise case, the analysis is similar.) There are four options:
(i) the initial position P (ti ) belongs to segment A1 B 1 C1 and the final position P (tf ) to segment A2 B 2 C2 , (ii)
P (ti ) C1D1 and P (tf ) C2D2 , (iii) P (ti ) C1D1 and P (tf ) A2 B 2 C2 , and (iv) P (t0 ) A1 B 1 C1 and
P (tf ) C2D2 .
In case (i) the solution LSL is the shortest possible path, since the straight line segment connecting C1 and C2 is
shorter than the sum of segments C1D1 + D1D2 + D2C2 . In case (ii), the path LRL cannot be the optimal solution:
since the middle arc is less than /2, this path is of type CCC, which was shown above to disqualify it from being a

Fig. 5. An illustration to the fact that path CCC cannot be the optimal solution in the long path case.

190

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 6. The optimal solution for class a11 ; both and are in the first quadrant.

candidate for the optimal solution. For the same reason, path CCC cannot be the optimal solution in cases (iii) and
(iv).
4.2. The equivalency groups of {aij } classes
We now turn to defining the equivalency groups for each element of the matrix {aij }. Recall that each element
represents a class of paths.
4.2.1. Equivalency group {a11 , a44 }
According to Proposition 4, classes a11 and a44 belong to the same equivalency group. We first show for class
a11 that the corresponding Dubins set can be reduced directly to a unique optimal solution. Then, by applying an
orthogonal transformation to the optimal path for class a11 , the optimal solution for a44 will be obtained.
Proposition 7. For the long path case (see Proposition 5), the optimal solution corresponding to the element a11
is RSL.
Note that the number of candidate curves for the optimal path in set D is now reduced to LSL, LSR, RSL, RSR
the curves of type CCC are excluded from consideration (Proposition 6). For three of those paths, LSL, LSR, RSR,
the x-coordinate goes outside the range 0 6 x 6 d. Take the length of the curve LS as the lower bound on the length
of the paths LSL, LSR and RSR (LS is a subpath of paths LSL or LSR with = /2).
It is claimed that the upper bound on the length of paths for which the x-coordinate is in the range 0 6 x 6 d
is the path RSL with = = /2. Indeed, from Section 2, Lrsl / > 0 and Lrsl / > 0. Therefore, the
maximum of Lrsl in this region will occur when and are equal to /2.
To prove that the optimal solution for paths within the range 0 6 x 6 d is RSL, one needs to show that the lower
bound on the path length in the Dubins subset {LSL, LSR, RSR} is bigger than the upper bound on the path RSL.
This case is illustrated in Fig. 6.
Lemma 1. If and are in the first quadrant then the upper bound for Lrsl is limited by
p
max Lrsl 6 d 2 4d + 2.
,[0,/2]

Indeed, when [0, /2] and [0, /2], the gradient of the path RSL is a positive function. This means that
the function is monotonically increasing on that interval and reaches its maximum on the intervals boundary, i.e.
= = /2. Assuming, as usual, = 1, the upper bound for the path length is
Lrsl = IA + AB + BF,

(16)

where
IA + BF < 2.

(17)

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Segment AB can be found as AB = 2AD or, expressing AD in terms of O1 D and O1 A,

q
p
AB = 2 ( 21 (d 2))2 1 = d 2 4d.

191

(18)

Substituting (17) and (18) into (16), we obtain

p
max Lrsl 6 d 2 4d + 2.
,[0,/2]

Lemma 2. If and are in the first quadrant then the lower bound on Llsl , Llsr , Lrsr is
p
min {Llsl , Llsr , Lrsr } > d 2 + 2d + 3/2.
,[0,/2]

The lower bound on LSL, LSR and RSR can be obtained by taking = /2 and as shown in Fig. 6. The minimum
path length can then be estimated as
p
p
+ CF > 3/2 + (d + 1)2 1 > 3/2 + d 2 + 2d.
(19)
min {Llsl , Llsr , Lrsr } = I EC
,[0,/2]

In order to prove that the optimal solution corresponding to class a11 is RSL, we need to show that
min{Llsl , Llsr , Lrsr } max Lrsl > 0.
,

(20)

It is easy to see that (20) holds if d 2 + 2d d 2 4d /2> 0: move /2 to the right sideof the inequality
and multiply both sides by the positive expression d 2 + 2d + d 2 4d. The result is 6d > d 2 + 2d, which
is true if d 2 4d > 0 precisely the case we are interested in. This completes the proof of inequality (20) and of
the claim that the optimal solution for class a11 is RSL. By applying Proposition 2 to the path RSL, obtain a similar
statement for class a44 .
Proposition 8. Given that a11 and a44 are in the same equivalency group and the optimal solution for a11 is RSL,
the optimal solution for a44 is LSR.
4.2.2. Equivalency group {a12 , a21 , a34 , a43 }
According to Proposition 4, path classes a12 ' a21 ' a34 ' a43 are in the same equivalency group. We first
show how to extract the optimal solution for class a12 : it turns out that class a12 defines two (rather than one as with
class a11 ) elements of the Dubins set as candidates for the optimal solution. Accordingly, a switching function S12
will be derived whose sign will uniquely determine which of the two is the optimal solution. Then, by applying the
orthogonal transformation to the paths of class a12 (see Proposition 2), optimal solutions for path classes a21 , a34 ,
and a43 will be obtained.
Proposition 9. For the long path case, the optimal solution corresponding to the class a12 is either RSL or RSR.
It follows from Proposition 6 that the path of type CCC can be excluded from consideration. This leaves four
candidates, RSR, RSL, LSR, and LSL. Define the critical initial orientation as one where orientation coincides
Note that the set (, d) uniquely defines the critical initial
with the tangent to the circle ORF ; denote it = .
orientation, = (,

d). If > then the path LSR is not feasible and can be excluded from consideration;
otherwise, path RSR should be excluded. Consider the case when path LSR is feasible ( < ).
If = 0 and =
then the length of LSR is equal to that of RSL. Analysis of gradients of functions Lrsl and Llsr shows that if is
increasing or is decreasing then path RSL becomes shorter than LSR. This is true until reaches the critical initial
orientation = .
Comparing the lower bound on the length of path LSL and the upper bound on the length of path

192

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 7. Choosing the switching functions for the equivalency group (a12 , a21 , a34 , a43 ) an illustration.

RSL, observe that path LSL can be excluded from consideration. This argument plus the fact that Llsr > Lrsl for
< end the proof of the proposition.
To find now the optimal solution for class a12 , we define a switching function, S12 . The following proposition
holds.
Proposition 10. For class a12 (i.e. 0 < 6 /2, /2 < 6 ), the optimal solution is RSR if S12 (prsr , prsl , qrsl ) <
0, and it is RSL if S12 (prsr , prsl , qrsl ) > 0, with
S12 (prsr , prsl , qrsl ) = prsr prsl 2(qrsl ),

(21)

where prsr , prsl , and qrsl are defined by (5) and (9).
Consider an example in Fig. 7(a); ORI , ORF , and OLF are the centers of circles Cri , Crf , and Clf , respectively. The
realizable paths here are RSR and RSL. Line EH connects the origins OLF and ORF and intersects the circles Clf and
Crf in points E and H . Since line ORI G is parallel to line EH, arc AG is equal to arc BH defined by the angle , and
arc GC is equal to arc DE defined by the angle . For segment AB, denote srsr to be the length of the straight line
segment of path RSR; similarly for segment CD, srsl is the length of the straight line segment of path RSL. Then the
path lengths of RSR and RSL are given by Lrsr = SA + prsr + + and Lrsl = SA + + + prsl + + . Therefore,
the sign of the difference (Lrsr Lrsl ) defines the bigger of the lengths of RSR and RSL. Expanding the difference
(Lrsr Lrsl ) and substituting = (qrsl ), obtain expression (21) for the switching function of class a12 .

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

193

Remark. For class a12 , the region of , where RSR is the optimal solution is much smaller than the corresponding
region for RSL. To save on computations, divide the whole region into two subregions: one where RSL is the optimal
solution, and the other (much smaller), where the solution is RSL or RSR and depends on the condition of Proposition
9. The first subregion (where RSL is the optimal solution) is defined by
d cos 3 sin() cos + sin( ) + cos sin > 0.
This case occurs if and are such that the last arc in the path RSL, qrsl , is equal to . If qrsl < then path RSL
is shorter than path RSR. If qrsl > , then the switching function S12 needs be checked.
We now turn to the classes a21 , a34 , and a43 which are in the same equivalency group as a12 . By applying
the orthogonal transformation (see Proposition 2), the set of path candidates for a12 is transformed into the set of
candidates for the remaining elements of the equivalency group a21 , a34 , a43 (see Fig. 7). That leads to the following
result.
Proposition 11. Since a12 ' a21 ' a34 ' a43 and the optimal solution for a12 is {RSL or RSR}, then the optimal
solutions for the remaining elements of this equivalency group are a21 7 {RSL or LSL}; a34 7 {LSR or RSR};
a43 7 {LSR or LSL}.
The respective switching functions for classes a21 , a34 , and a43 (see the next three propositions) are obtained from
the switching function for class a12 , by replacing t and q segments as prescribed by the transformation theorem.
Proposition 12. For class a21 (i.e. /2 < 6 , 0 < 6 /2), the optimal solution is LSL if S21 < 0, and it is
RSL if S21 > 0, with
S21 (plsl , prsl , trsl ) = plsl prsl 2(trsl ).

(22)

To check, apply Theorem 1 to the switching function S12 = Lrsr Lrsl = prsr prsl 2 , where = qrsl ;
obtain Lrsr = SA + prsr + + and Lrsl = SA + + + prsr + + . According to Proposition 3, the orthogonal
transformation implies
Lrsr ( [0, /2], [/2, ]) = Llsl ( [/2, ], [0, /2]),
and, additionally, prsr ( [0, /2], [/2, ]) = plsl ( [/2, ], [0, /2]) and prsl (
[0, /2], [/2, ]) = prsl ( [/2, ], [0, /2]), and qrsl ( [0, /2], [/2, ]) = trsl (
[/2, ], [0, /2]). The switching function for class a21 can be simply obtained from the function S12 by
changing the segments from prsr , prsl , qrsl to plsl , prsl , trsl , respectively. The validity of this procedure follows from
the transformation theorem, and using again the same theorem, obtain directly the switching functions for classes
a34 and a43 .
Proposition 13. For class a34 (i.e. < 6 3/2, 3/2 < 6 2), the optimal solution is RSR if S34 < 0, and
it is LSR if S34 > 0, with
S34 (prsr , plsr , tlsr ) = prsr plsr 2(tlsr ).

(23)

Proposition 14. For class a43 (i.e. 3/2 < 6 2, < 6 3/2), the optimal solution is LSL if S43 < 0, and
it is LSR if S43 > 0, with
S43 (plsl , plsr , qlsr ) = plsl plsr 2(qlsr ).

(24)

194

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 8. Choosing the switching functions for the equivalency group (a13 , a24 , a31 , a42 ) an illustration.

4.2.3. Equivalency group {a13 , a24 , a31 , a42 }

According to Proposition 4, classes a13 , a24 , a31 , a42 belong to the same equivalency group. First we show that
for class a13 the Dubins set can be reduced to two path candidates, and then further reduced to a unique optimal
solution using an appropriate switching function, S13 . Then, by applying the transformation theorem, the switching
function S13 will be modified to produce the corresponding switching functions for classes a24 , a31 , a42 . For a13 ,
the following holds.
Proposition 15. For the long path case, the optimal solution corresponding to the element a13 is either RSR or
LSR.
Similar to the argument for class a12 above, one can see that path RSL can be excluded from the set of candidates
considered for the optimal path. Indeed, if = 0 and = , then Llsr = Lrsl . If or are increasing then path
LSR becomes shorter than path RSL. To see this, compare the line segment SD, which is a tangent to circle OLF ,
with SD, a tangent to circle ORF (see Fig. 8(a)). Now, SA < SD for any (, 3/2], and AB < DC; also,
SD < S G + GE + E D. This leads to
SA + AB < S G + GE + E C.
Therefore, if [0, ]
then Llsr < Lrsl . Similarly, if > then Lrsr < Lrsl . Here is the critical angle
= (d,

), where paths LSR and RSR degenerate to SR. This implies that only paths LSR and RSR are candidates
for the optimal solution.
Using Proposition 2, a similar argument extends this result to the remaining three elements, as follows.

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

195

Proposition 16. Since a13 ' a24 ' a31 ' a42 and the optimal solution for a13 is either RSR or LSR, the optimal
solutions for a24 , a31 , a42 are, respectively,
a24 7 {RSR or RSL}; a31 7 {LSL or LSR}; a42 7 {LSL or RSL}.
Turning back to class a13 , we can verify that the critical angle defines switching of the optimal path from LSR
to RSR; i.e. if > the optimal solution is RSR, otherwise it is LSR. Another indicator of the switch from LSR to
RSR is the length of arc trsr . Notice that if RSR is the optimal solution for class a13 then trsr cannot exceed .
This observation leads to the corresponding switching functions necessary for obtaining the optimal solution.
Proposition 17. For class a13 (i.e. 0 < 6 /2, 3/2 < 6 2), if
1. S13 (trsr ) < 0, then the optimal solution is RSR,
2. S13 (trsr ) > 0, then the optimal solution is LSR,
where
S13 (trsr ) = trsr .

(25)

The transformation theorem extends the results obtained for class a13 to the accompanying classes a24 , a31 and a42 .
Proposition 18. For class a24 (i.e. /2 < 6 , 3/2 < 6 2), the optimal solution is RSR if S24 < 0, and it
is RSL otherwise, where
S24 (qrsr ) = qrsr .

(26)

Proposition 19. For class a31 (i.e. < 6 3/2, 0 < 6 /2), the optimal solution is LSL if S31 < 0, and it
is LSR otherwise, where
S31 (qlsl ) = qlsl .

(27)

Proposition 20. For class a42 (i.e. 3/2 < 6 2, /2 < 6 ), the optimal solution is RSL if S42 < 0, and it
is LSL otherwise, where
S42 (tlsl ) = tlsl .

(28)

4.2.4. Equivalency group {a14 , a41 }

According to Proposition 4, path classes a14 ' a41 are in the same equivalency group, and so, by the rules
established in the transformation theorem, the classification of optimal solutions for class a14 leads to the optimal
solutions for class a41 .
Proposition 21. For the long path case, the optimal solution corresponding to the element a14 is {RSR or LSR or RSL}.
Recall that by now the total number of candidates for the optimal solution is reduced to four LSL, LSR, RSL,
RSR because the curves of type CCC are excluded from consideration due to Proposition 4. As before, define
the critical orientations, which for class a14 appear for both and (see Fig. 9): namely, the values d and
determine the critical orientation for , = (d,

), and the values d and determine the critical orientation for

) (notice that 6= (d, )).

, = (d,
Among the four path candidates mentioned, path LSL can be excluded from consideration since the lower bound
for this path is bigger than the upper bound for path RSR. The critical angle gives the optimal solution SR, and

196

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 9. Example of critical orientations of the path candidates for class a14 .

thus defines the switch condition for the optimal path from LSR to RSR. Similarly, the critical angle indicates the
switch of the optimal path from RSR to RSL.

) the
The proof of these facts is similar to those for classes a11 and a13 . Define two critical angles: = (d,
) the angle at which paths RSR
angle at which paths LSR and RSR degenerate into the path SR, and = (d,
and RSL degenerate into the path RS. First consider the case when [,
/2] and [3/2, (d, )].
It is
claimed that if [,
/2] and [3/2, (d, )]
then the upper bound for Lrsr is limited by
Lrsr 6 d 2 + .

max

[,/2],[3/2,(d,

)]

In the
/2] and strictly negative when [3/2, ].
Indeed, the gradient of Lrsr is strictly positive when [,

regions [,
/2] and [3/2, ], the maximum length of path RSR occurs when = /2 and = 3/2
(this follows from monotonicity of the function Lrsr ), i.e. when
max

Lrsr 6 d 2 + .

(29)

[,/2],[

,2]

then the lower bound on Llsl , Lrsl , and Llsr is

One can conclude that if [,
/2] and [3/2, ]
p
{Llsl , Lrsl , Llsr } > d 2 + 2d + 3/2.
min

[,/2],[3/2,

These critical cases are shown in Fig. 10. The resulting bounds are
min

Llsl > 3 + d + 2,

min

Lrsl > 3/2 +

p
d 2 + 2d,

(31)

min

Llsr > 3/2 +

p
d 2 + 2d.

(32)

(30)

[,/2],[3/2,

[,/2],[3/2,

[,/2],[3/2,

This leads to the condition

min

{Llsl , Lrsl , Llsr } > 3/2 +

[,/2],[3/2,

p
d 2 + 2d.

(33)

Notice that the lower bound on paths {Llsl , Lrsl , Llsr } is larger than the upper bound on Lrsr (Eqs. (29) and (33)).
This completes the proof of the fact that if [,
/2] and [3/2, (d, )],
then the path RSR is the optimal

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

197

Fig. 10. Example of critical orientations of the path candidates for class a14 .

solution. For class a14 two more candidates are possible, i.e. LSR and RSL. The proposition below relates the set of
candidates for class a14 to the accompanying class a41 .
Proposition 22. Since a14 ' a41 and the optimal solution for class a14 is {LSR or RSL or RSR}, the optimal solution
for class a41 is {RSL or LSR or LSL}.
Similar to the class a13 , the critical angles and define the switch of the optimal solution from RSR to LSR
and from RSR to RSL, respectively. To simplify the calculations, we choose to build our classification based on the
length of segments trsr and qrsr . Following the same logic as in the case of class a13 , one notices that trsr and qrsr
cannot exceed if RSR is an optimal solution for class a14 . If either of the two segments trsr or qrsr is bigger than
, then there is always another path that is shorter than RSR; i.e. if trsr > then the optimal solution is LSR, and if
qrsr > then the optimal solution is RSL. This observation leads to the following classification rule.
Proposition 23. For class a14 (i.e. 0 < 6 /2, 3/2 < 6 2), if
1 (t ) > 0, then the optimal solution is RSR,
1. S14
rsr
2 (q ) > 0, then the optimal solution is RSL,
2. S14
rsr
3. if neither (1) or (2) holds, then the optimal solution is RSR,

where
1
(trsr ) = trsr ,
S14

(34)

2
(trsr ) = qrsr .
S14

(35)

By applying the transformation theorem, obtain the switching conditions for class a41 .
Proposition 24. For class a41 (i.e. 3/2 < 6 2, 0 < 6 /2), if
1 (t ) > 0, then the optimal solution is RSL,
1. S41
rsr

198

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

2 (q ) > 0, then the optimal solution is LSR,

2. S14
rsr
3. if neither (1) or (2) holds, then the optimal solution is LSL,

where
1
(tlsl ) = tlsl ,
S14

(36)

2
(tlsl ) = qlsl .
S14

(37)

4.2.5. Equivalency group {a22 , a33 }

According to Proposition 4, classes a22 ' a33 are in the same equivalency group, and so the classification of
optimal solutions for class a22 leads to that for class a33 .
Proposition 25. For the long path case, the optimal solution corresponding to the class a22 is {LSL or RSL or RSR}.
The path LSR is excluded from the consideration since Llsr is always larger than Lrsl . Indeed, for = = the
lengths Llsr = Lrsl , and path LSR has its minimum at = = , and path RSL its maximum within class a22 .
Proposition 26. Since a22 ' a33 and the optimal solution for a22 is {LSL or {RSL or RSR}}, then the optimal solution
for a33 is {RSR or {LSR or LSL}}.
The following proposition gives the condition for finding the optimal paths.
Proposition 27. For class a22 (i.e. /2 < 6 , /2 < 6 ), if
1.
2.
3.
4.

>
>
<
<

1
and S22
1
and S22
2
and S22
2
and S22

< 0, the optimal solution is LSL,

> 0, the optimal solution is RSL,
< 0, the optimal solution is RSR,
> 0, the optimal solution is RSL,

where the switching functions are

1
(plsl , prsl , trsl ) = plsl prsl 2(trsl ),
S22

(38)

2
(prsr , prsl , qrsl ) = prsr prsl 2(qrsl ),
S22

(39)

and plsl , prsr , prsl , trsl , and qrsl are defined by (3) and (9), respectively.
The proof of the proposition is similar to that for class a12 with the only difference that the set of candidates
depends upon the relations between and . Notice that if = then the length of path LSL is equal to that of RSR.
When starts increasing, path LSL becomes shorter than path RSR (see Fig. 11). Similarly, using the transformation
theorem, obtain the solutions for class a33 are as follows.
Proposition 28. For class a33 (i.e. < 6 3/2, < 6 3/2), if
1.
2.
3.
4.

<
>
<
>

1
and S33
1
and S33
2
and S33
2
and S33

< 0, the optimal solution is RSR,

> 0, the optimal solution is LSR,
< 0, the optimal solution is LSL,
> 0, the optimal solution is LSR,

where the switching functions are

1
(prsr , plsr , tlsr ) = prsr plsr 2(tlsr ),
S33

(40)

2
(plsl , plsr , qlsr ) = plsl plsr 2(qlsr ).
S33

(41)

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

199

Fig. 11. For class a22 , the set of candidates for the optimal solution depends much upon the relations between and : if > , the candidates
are LSL and RSL, otherwise they are RSR and RSL.

4.2.6. Equivalency group {a23 , a32 }

According to Proposition 4, classes a23 ' a32 are in the same equivalency group. The corresponding classification
of optimal solutions proceeds as follows.
Proposition 29. For the long path case, the optimal solution corresponding to the class a23 is RSR.
Four path candidates need be considered here RSR, RSL, LSR, LSL. Note that
max

Lrsr =

[/2, ],[,3/2]

min

{Lrsl Llsr Llsl }.

[/2,],[,3/2]

The maximum of the length of path RSR and the minimum of {Lrsl Llsr Llsl } occur when = = . By applying
the transformation theorem, the classification for a32 is given by the following proposition.
Proposition 30. Since a23 ' a32 and the optimal solution for a23 is RSR, the optimal solution for a32 is LSL.
The switching functions are logical functions of Boolean type. They are not uniquely defined. The choice of a
particular function will be usually guided by computational considerations.
5. The main result
We can now summarize the whole scheme developed above, which presents the main result of this work. Using
the scheme, the problem of finding the shortest smooth path between two configurations is solved without an explicit
calculation of the paths involved. Instead, a simple logical scheme is used based on the aggregation of all possible
paths into classes aij and on the equivalency groups as defined above.
The input to the scheme are the angular quadrants of the directional angles (, ); its output is the name of the
element of the Dubins set that presents the shortest path. The scheme forms a decision tree summarized in the table
in Fig. 12. Each block of the table represents one element aij of matrix {aij } and includes the corresponding options
for the optimal path. It may take only one step to obtain the solution, as e.g. the solution RSL for the quadrants
(1, 1), or it may take two steps, as for the quadrants (3, 1), or it may take at most three steps, as for the quadrants
(2, 2) (Fig. 12). When the second and third steps are necessary, their outcome is determined by the signs of one or
two switching functions {S}. The general form of functions {S} is
f (s1 , s2 , s3 ) = s1 s2 2(s3 ),

g(s) = s .

(42)

200

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

Fig. 12. The decision table for finding the shortest path.

The complete list of switching functions for all classes aij is as follows (note that for some classes, the unique
optimal solution is obtained directly, without switching functions):
Class a11 : unique solution,
Class a12 : S12 = f (prsr , prsl , qrsl ) = prsr prsl 2(qrsl ),
Class a13 : S13 = g(trsr ) = trsr ,
Class a14 :
Class a21 :
Class a22 :

1 = g(t ) = t
2 = g(q ) = q
S14
S14
rsr
rsr ,
rsr
rsr ,
S21 = f (plsl , prsl , trsl ) = plsl prsl 2(trsl ),
(
1 = f (p , p , t ) = p p 2(t ),
if > , then S22
lsl
rsl rsl
lsl
rsl
rsl

Class a34

2 = f (p , p , q ) = p
if < , then S22
rsr
rsr prsl 2(qrsl ),
rsl rsl
: unique solution,
: S24 = g(qrsr ) = qrsr ,
: S31 = g(qlsl ) = qlsl ,
: unique solution,
(
1 = f (p , p , t ) = p
if < , then S33
rsr
rsr plsr 2(tlsr ),
lsr lsr
:
2 = f (p , p , q ) = p p 2(q ),
if > , then S33
lsl
lsr lsr
lsl
lsr
lsr
: S34 = f (prsr , plsr , tlsr ) = prsr plsr 2(tlsr ),

Class a41
Class a42
Class a43
Class a44

1 = g(t ) = t ,
2 = g(q ) = q ,
: S41
S41
lsl
lsl
lsl
lsl
: S42 = g(tlsl ) = tlsl ,
: S43 = f (plsl , plsr , qlsr ) = plsl plsr 2(qlsr ),
: unique solution.

Class a23
Class a24
Class a31
Class a32
Class a33

(43)

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

201

Take, e.g. a set ( [0, /2], [/2, ]), which corresponds to the element a12 of matrix {aij }. Block (1, 2)
in the table in Fig. 12 suggests that the optimal solution is either RSL or RSR. The switching function S12 listed in
block (1, 2) is then calculated as in the list above (Eq. (42)); its elements prsr , prsl and qrsl come from the second
equation of (5) and the second and third equations of (9), respectively. The sign of S12 then determines which of
the two versions indicated in block (1, 2) is the unique solution.
To demonstrate efficiency of the proposed classification scheme, consider the task of finding the shortest smooth
path from Pi to Pf , located on the distance |Pi Pf | = 6, such that the path starts and ends with the directions of
motion = /6 and = /3, respectively, and the path radius of curvature is limited by = 1. According to
Dubins result, six paths candidates {LSL, RSR, RSL, LSR, RLR, LRL} have to be calculated and compared. Using
expressions given in Section 2
Llsl = tlsl + plsl + qlsl = + + plsl = 12.4526,
Lrsr = trsr + prsr + qrsr = + prsr = 12.1361,
Llsr = tlsr + plsr + qlsr = + 2tlsr + plsr = 18.3890,
Lrsl = trsl + prsl + qrsl = + + 2trsl + prsl = 6.2488,
Lrlr = trlr + prlr + qrlr = + 2prlr path is not feasible,
Llrl = tlrl + plrl + qlrl = + + 2plrl path is not feasible.

(44)

After all paths-candidates are calculated and compared, we conclude that the shortest path is RSL. In contrast, if
we use the approach proposed in this paper, then based only on information about the initial and final configuration
and without any calculations involved, we can make the same conclusion. Indeed, for the case studied, the initial
and final configurations are in the first quadrant, i.e. the path belongs to the class a11 . According to the look-up
table (Fig. 12), the optimal solution for all paths from this class have to have the topology RSL. Thus, using the
classification scheme we were able to obtain the optimal solution without any calculations involved. This example
is a good illustration of the computational efficiency of the proposed classification scheme.

6. Conclusion
The central idea developed in this work is that the problem of finding the shortest path between two configurations
can be reduced to a logical manipulation of the set of appropriate path candidates, without their explicit calculation.
This is in sharp departure from the direct computation and comparison of the candidate paths that the existing
techniques require. The candidate paths come from the sufficient set known as the Dubins set [5]. One direct
benefit of the suggested scheme is computational savings an important consideration in real-time control. For
example, when attempting to find the shortest path to a given position/orientation (configuration) for a driverless
car or a mobile robot, one would simply find in the table (Fig. 12), the element that corresponds to the initial and
final configurations, and then pinpoint the unique solution either immediately or using the sign of an appropriate
switching function of the form (42).
As mentioned in Section 1, the derivation of the approach is simplified if the problem at hand is divided into two
cases, called here the long path case and the short path case. To save space, the suggested logical classification
scheme is fully developed here only for the long path case. Situations with short paths require a roughly similar,
though a bit tedious, analysis, resulting in a computational procedure that is somewhat more complex and less
economical than the one presented here (see Proposition 5 for a formal definition of both cases).
The presented result also gives a new interesting insight into the nature of Dubins problem. It suggests that
partitioning of the appropriate C-space can be a powerful tool for analyzing the shortest path problem in more
general and complex cases, as e.g. in finding the shortest path between a point and a manifold.

202

A.M. Shkel, V. Lumelsky / Robotics and Autonomous Systems 34 (2001) 179202

References
[1] M. Krein, A. Nudelman, The Markov Moment Problem and Extremal Problems, The American Mathematical Society, Providence, RI,
1977.
[2] P. Jacobs, J. Canny, Planning smooth paths for mobile robots, in: Proceedings of the IEEE International Conference on Robotics and
Automation, Scottsdale, AZ, May 1989.
[3] P. Jacobs, J. Laumond, M. Taix, Efficient motion planners for nonholonomic mobile robots, in: Proceedings of the IEEE/RSJ International
Workshop on Intelligent Robots and Systems, Osaka, Japan, August 1991.
[4] J. Barraquand, J.-C. Latombe, On nonholonomic mobile robots and optimal maneuvering, in: Proceedings of the Fourth International
Symposium on Intelligent Control, Albany, NY, 1989.
[5] L.E. Dubins, On curves of minimal length with a constraint on average curvature, and with prescribed initial and terminal positions and
tangents, American Journal of Mathematics 79 (1957) 497516.
[6] J.A. Reeds, L.A. Shepp, Optimal paths for a car that goes both forwards and backwards, Pacific Journal of Mathematics 145 (1990) 367393.
[7] J. Boissonnat, A. Cerezo, J. Leblond, Shortest paths of bounded curvature in the plane, in: Proceedings of the IEEE International Conference
on Robotics and Automation, Nice, France, May 1992.
[8] L.S. Pontryagin, The Mathematical Theory of Optimal Processes, Interscience, New York, 1962.
[9] G. Desaulniers, F. Soumis, An efficient algorithm to find a shortest path for a car-like robot, IEEE Transactions on Robotics and Automation
11 (6) (1995) 819828.
[10] Ph. Soueres, J.-P. Laumond, Shortest paths synthesis for a car-like robot, IEEE Transactions on Automatic Control 41 (5) (1996) 672688.

Andrei M. Shkel is an Assistant Professor in the Department of Mechanical and Aerospace Engineering and an Assistant
Professor in the Department of Electrical and Computer Engineering at the University of California, Irvine. He is also
the Director of the UCI Microsystems Laboratory. Dr. Shkel received his BS degree (magna cum laude) in Applied
Mathematics in 1990 and MS degree in Applied Mechanics and Control in 1991, both from Lomonosovs Moscow
State University. In 1997, he received PhD degree from the University of Wisconsin-Madison with major in Mechanical
Engineering and allied in Electrical and Computer Engineering. The subject of his PhD was the sensor-based motion
planning with kinematic and dynamic constraints. After receiving his PhD, Dr. Shkel joined Berkeley Sensor and Actuator
Center (BSAC) as a Postdoctoral Researcher, where he was developing micro-machined monolithic gyroscopes. In 1999,
Dr. Shkel joined the MEMSolutions, Inc., as a Senior MEMS Design Engineer, where he was conducting development
and design of novel cost-effective sensors and optical MEMS-based products. He is currently a technical consultant
for the MEMSolutions, Inc. Dr. Shkel is active in robotics, sensor-based intelligence, MEMS, and MEMS-based sensor technology. He served
as a reviewer for more than a dozen major journals and international conferences in these areas, published 23 papers in archival journals and
international conferences, and presented his work on a number of technical meetings and industry advisory boards. During the last 2 years, Dr.
Shkel authored and co-authored five inventions (currently patents pending) on design of novel MEMS inertial sensors and optical MEMS-based
products. Dr. Shkel is a member of the Administrative Committee of the IEEE Sensors Council, he is also an associate member of the IEEE and
ASME.

Vladimir J. Lumelsky is on the faculty at the University of Wisconsin-Madison, where he is Professor of Mechanical Engineering, Professor of Electrical Engineering, Professor of Computer Science, and Professor of Mathematics.
He received his BS/MS in Electrical Engineering and Computer Science from the Institute of Precision Technology,
Leningrad, Russia, and his PhD in Applied Mathematics from the Institute of Control Sciences (ICS), USSR National
Academy of Sciences, Moscow, in 1970. He then held academic and research positions with the ICS, Moscow, Ford
Motor Research Laboratories, General Electric Research Center, and Yale University. His professional interests, reflected in over 200 publications, include robotics, computational geometry, sensor-based intelligence, control theory,
kinematics, industrial automation, computer vision, and pattern recognition. He has served on a number of editorial
boards, including the Editor-in-Chief of IEEE Sensors Journal and Senior Editor of the IEEE Transactions on Robotics
and Automation, on the Board of Governors of the IEEE Robotics and Automation Society, as Chairman of this Societys
Technical Committee on Robot Motion Planning, Chairman of IFAC Working Group on Robot Motion, Sensing, and Planning, Program Chair
of the 1989 IEEE International Conference on Intelligent Robots and Systems (IROS89), Tokyo, Guest Editor for two special issues of the
IEEE Transactions on Robotics and Automation. He is IEEE Fellow, and member of ACM and SME.