review articles

DOI:10.1145/ 2629535
Motion planning is about deciding
Exploring the distinction between an optimal on the existence of a motion to reach
a given goal and computing one if
robot motion and a robot motion resulting from this one exists. Robot motion control
the application of optimization techniques. arose from manufacturing and the
control of manipulators30 with rapid
BY JEAN-PAUL LAUMOND, NICOLAS MANSARD, effective applications in the automo-
AND JEAN-BERNARD LASSERRE tive industry. Motion control aims
at transforming a task defined in the

Optimality in
robot workspace into a set of control
functions defined in the robot motor
space: a typical instance of the prob-
lem is to find a way for the end-effec-

Robot Motion:
tor of a welding robot to follow a pre-
defined welding line.
What kind of optimality is about
in robot motion? Many facets of the

Optimal versus
question are treated independently
in different communities ranging
from control and computer science,
to numerical analysis and differen-

Optimized
tial geometry, with a large and di-
verse corpus of methods including,
for example, the maximum principle,
the applications of Hamilton-Jacobi-

Motion
Bellman equation, quadratic pro-
gramming, neural networks, simu-
lated annealing, genetic algorithms,
or Bayesian inference. The ultimate
goal of these methods is to compute
a so-called optimal solution whatever
the problem is. The objective of this
article is not to overview this entire
corpus that follows its own routes
independently from robotics, but
dedicated to robot motion was
T H E FIR ST B OOK
published in 1982 with the subtitle Planning and key insights
Control.5 The distinction between motion planning  C omputing an optimal robot motion is
and motion control has mainly historical roots. a challenging issue illustrated by more
than 20 years of research on wheeled
Sometimes motion planning refers to geometric path mobile robots.

planning, sometimes it refers to open loop control;  G eometric control theory and numerical
analysis highlight two complementary
sometimes motion control refers to open loop perspectives on optimal robot motion.

control, sometimes it refers to close loop control and  M ost of the time, robot algorithms
aimed at computing an optimal motion
stabilization; sometimes planning is considered as provide an optimized motion, which is not
optimal at all, but is the output of a given
an offline process whereas control is real time. From optimization method.
a historical perspective, robot motion planning arose W
 hen optimal motions exist, numerical
from the ambition to provide robots with motion algorithms mostly fail in accounting for
their combinatorial structure. Moreover,
autonomy: the domain was born in the computer optimization algorithms bypass (not
overcome) the question of the existence
science and artificial intelligence communities.22 of optimal motions.

82 COMMUNICATIO NS O F TH E ACM | S EPTEM BER 201 4 | VO L . 5 7 | N O. 9

 rather to emphasize the distinction example, as developed in the context problem: planning a robot motion is
between optimal motion and opti- of robot soccer or in task planning). equivalent to searching for connected
mized motion. Most of the time, ro- components in a configuration space.
bot algorithms aiming at computing What Is Optimal in Robot Motion Schwartz and Sharir then showed the
an optimal motion provide in fact an Planning and Control? problem is decidable as soon as we
optimized motion that is not optimal Motion planning explores the compu- can prove that the connected com-
at all, but is the output of a given opti- tational foundations of robot motion by ponents are semi-algebraic sets.35
mization method. Computing an op- facing the question of the existence of Even if a series of papers from com-
ILLUSTRATION BY A LIC IA KU BISTA /A NDRIJ BORYS ASSOCIAT ES

timal motion is mostly a challenging admissible motions for robots moving putational geometry explored various
issue as it can be illustrated by more in an environment populated with ob- instances of the problem, the gen-
than 20 years of research on wheeled stacles: how to transform the continu- eral piano mover problem remains
mobile robots (as we discuss later). ous problem into a combinatorial one? intractable.14 Finally by relaxing the
Note that the notion of optimality This research topic22,26 evolved in completeness exigence for the ben-
in robot motion as it is addressed in three main stages. In the early 1980s, efit of probabilistic completeness,
this article is far from covering all the Lozano-Perez first transformed the Barraquand and Latombe introduced
dimensions of robot motion.7 It does problem of moving bodies in the a new algorithmic paradigm3 in the
not account for low-level dynamical physical space into a problem of mov- early 1990s that gave rise to the popu-
control, nor for sensory-motor con- ing a point in some so-called configu- lar probabilistic roadmap20 and rapid
trol, nor for high level cognitive ap- ration space.28 In doing so, he initi- random trees31 algorithms.
proaches to motion generation (for ated a well-defined mathematical Motion planning solves a point-to-

SE PT E MB E R 2 0 1 4 | VO L. 57 | N O. 9 | C OM M U N IC AT ION S OF T HE ACM 83
review articles

point problem in the configuration ration space induced by the presence ing among obstacles. For some right-
space. Whereas the problem is a dif- of obstacles and by the metric to be ful security reason, the robot cannot
ficult computational challenge that optimized. Time criterion is also dis- touch the obstacles. In mathemati-
is well understood, optimal motion cussed, as are practical approaches to cal language, the robot must move in
planning is a much more difficult optimize time along a predefined path. an open domain of the configuration
challenge. In addition to finding a so- Apart from finding a feasible so- space. Yet, an optimal motion to go
lution to the planning problem (that lution to a given problem, motion from one place to another one located
is, a path that accounts for collision- planning also wants to optimize this behind some obstacle will necessar-
avoidance and kinematic constraints solution once it has been found. The ily touch the obstacle. So this optimal
if any), optimal motion planning re- question is particularly critical for the motion is not a valid one. It appears
fers to finding a solution that optimiz- motions provided by probabilistic algo- as an ideal motion that cannot be
es some criterion. These can be the rithms that introduce random detours. reached. The best we can do is to get
length, the time or the energy (which The challenge here is to optimize no a collision-free motion whose length
are equivalent criteria under some more in the configuration space of the approaches the length of this ideal
assumption), or more sophisticated system, but in the motion space. shortest (but non-admissible) motion.
ones, as the number of maneuvers to In this article, optimal motion In other words, there is no optimal so-
park a car. planning is understood with the un- lution to the corresponding motion
In such a context many issues are derlying hypothesis that the entire ro- planning problem. The question here
concerned with optimization: bot environment is known and the op- is of topological nature: combinato-
For a given system, what are the timization criterion is given: the quest rial data structures (for example, visi-
motions optimizing some criteria? Do is to find a global optimum without bility graphs) may allow us to compute
such motions exist? The existence of considering any practical issue such solutions that are optimal in the clo-
optimal motion may depend either on as model uncertainties or local sen- sure of the free space, and that are not
the presence of obstacles or on the cri- sory feedback. solutions at all in the open free space.
terion to be optimized. Even without obstacle, the exis-
When optimal motions exist, are Optimal Motion Existence tence of an optimal motion is far
they computable? If so, how complex is Before trying to compute an optimal from being guaranteed. In determin-
their computation? How to relax exact- motion, the first question to ask is istic continuous-time optimal con-
ness constraints to compute approxi- about its existence. To give some in- trol problems we usually search for a
mated solutions? We will address the tuition about the importance of this time-dependent control function that
combinatorial structure of the configu- issue, consider a mobile robot mov- optimizes some integral functional
over some time interval. Addressing
Figure 1. A modern view of the piano mover problem: two characters have to move a the issue of existence requires us
shared piano while avoiding surrounding obstacles.
to resort to geometric control theo-
ry;18 for instance, Fillipovs theorem
proves the existence of minimum-
time trajectories,a whereas Prontry-
agin Maximum Principle (PMP) or
Boltyanskiis conditions give respec-
tively necessary and sufficient condi-
tions for a trajectory to be optimal.
However it is usually difficult to ex-
tract useful information from these
tools. If PMP may help to characterize
optimal trajectories locally, it gener-
ally fails to give their global structure.
Later, we show how subtle the ques-
tion may be in various instances of
wheeled mobile robots.
The class of optimal control prob-
lems for which the existence of an
optimal solution is guaranteed, is lim-
ited. The minimum time problems
for controllable linear systems with
bounded controls belong to this class:
optimal solutions exist and optimal
controls are of bang-bang type. How-

a Here and in the following, we treat trajectory

and motion as synonyms.

84 COMMUNICATIO NS O F TH E AC M | S EPTEM BER 201 4 | VO L . 5 7 | N O. 9

 review articles

ever, the so-called Fuller problem may cal limitations of the system or the
arise: it makes the optimal solution cost functional is not appropriate to
not practical at all as it is of bang-bang guarantee the existence of an opti-
type with infinitely many switches. mal solution in that function space.
Other examples include the famous
linear-quadratic-Gaussian problem Even without To the best of our knowledge, this is-
sue is rarely discussed in textbooks
(the cost is quadratic and the dynam-
ics is linear in both control and state
an obstacle, or courses in optimal control.

variables), and systems with a bound- the existence of Optimal Path Planning
ed input and with a dynamics that is
affine in the control variables. In the
an optimal motion Considering a motion is a continu-
ous function of time in the configu-
former a closed loop optimal solution is far from being ration (or working) space, the image
can be computed by solving algebraic
Riccati equations, whereas in the lat-
guaranteed. of a motion is a path in that space.
The piano mover problem refers to
ter the existence of an optimal control the path-planning problem, that is,
trajectory is guaranteed under some the geometric instance of robot mo-
appropriate assumptions. tion planning (see Figure 1). The con-
In more general cases, we can only straint of obstacle avoidance is taken
hope to approximate as closely as de- into account. In that context, optimal-
sired the optimal value via a sequence ity deals with the length of the path
of control trajectories. There is in- without considering time and control.
deed no optimal solution in the too The issue is to find the shortest path
restricted space of considered con- between two points.
trol functions. This has already been Depending on the metric that
realized since the 1960s. The limit of equips the configuration space, a
such a sequence can be given a pre- shortest path may be unique (for
cise meaning as soon as we enlarge example, for the Euclidean metric)
the space of functions under consid- or not unique (for example, for the
eration. For instance, in the class of Manhattan metric). All configuration
problems in which the control is af- space metrics are equivalent from a
fine and the integral functional is topological point of view (that is, if
the L1-norm, the optimal control is there is a sequence of Euclidean paths
a finite series of impulses and not a linking two points, then there is also a
function of time (for example, see sequence of Manhattan paths linking
Neustadt29). In some problems such these two points). However, different
as the control of satellites, such a metrics induce different combinato-
solution makes sense as it can ap- rial properties in the configuration
proximately be implemented by gas space. For instance, for a same ob-
jets. However, in general, it cannot be stacle arrangement, two points may
implemented because of the physical be linked by a Manhattan collision-
limitations of the actuators. free path, while they cannot by a
Changing the mathematical for- collision-free straight-line segment:
mulation of the problem (for ex- both points are mutually visible in a
ample, considering a larger space of Manhattan metric, while they are not
control candidates) may allow the in the Euclidean one. So, according
existence of an optimal solution. In to a given metric, there may or may
the former case of satellite control, not exist a finite number of points
the initial formulation is coherent that watch the entire space.25 These
as an ideal impulse solution can combinatorial issues are particularly
be practically approximated by gas critical to devise sampling-based mo-
jets. However, in other cases the tion planning algorithms.
initial problem formulation may be Now, consider the usual case of a
incorrect as an ideal impulse solu- configuration space equipped with a
tion is not implementable. Indeed, Euclidean metric. Exploring visibil-
if we feel a smooth optimal solu- ity graph data structures easily solves
tion should exist in the initial func- the problem of finding a bound on
tion space considered and if in fact the length of the shortest path among
it does not exist, then either the dy- polygonal obstacles. This is nice, but
namics and/or the constraints do it is no longer true if we consider
not reflect appropriately the physi- three-dimensional spaces populated

SE PT E MB E R 2 0 1 4 | VO L. 57 | N O. 9 | C OM M U N IC AT ION S OF T HE ACM 85
review articles

with polyhedral obstacles. Indeed, usually addressed by means of nu- constraints on robot controls or robot
finding the shortest path in that case merical techniques. Among the most dynamics add another level of diffi-
becomes a NP-hard problem.14 So, in popular ones are the discrete search culties. The goal here is to compute a
general, there is no hope to get an algorithms operating on bitmap rep- minimal-time motion that goes from
algorithm that computes an optimal resentations of work or configuration a starting state (configuration and ve-
path in presence of obstacles, even spaces.3 The outputs we only obtain locity) to a target state while avoiding
if the problem of computing an opti- are approximately optimal paths, that obstacles and respecting constraints
mal path in the absence of obstacle is is, paths that are not so far from a on velocities and acceleration. This
solved and even if we allow the piano- hypothetical (or ideal) estimated op- is the so-called kinodynamic motion
robot to touch the obstacles. timal path. Another type of methods planning problem.12 The seminal al-
As a consequence of such poor consists in modeling the obstacles by gorithm is based on discretizations
results, optimal path planning is repulsive potential. In doing so, the of both the state space and the work-
goal is expressed by an attractive po- space.
Figure 2. A car (logo of the European Project tential, and the system tends to reach It gave rise to many variants includ-
Esprit 3 PRO-Motion in the 1990s) together
with the unicycle model equations.
it by following a gradient descent.21 ing nonuniform discretization, ran-
The solution is only locally optimal. domized techniques, and extensions
Moreover, the method may get stuck of A* algorithms (see LaValle26). Today,
in a local minimum without finding they are the best algorithms to com-
a solution, or that a solution actually pute approximately optimal motions.
exists or not. So it is not complete. Less popular in the robot motion
Some extensions may be considered. planning community are numerical
For instance, exploring harmonic po- approaches to optimal robot con-
tential fields8 or devising clever navi- trol.11 Numerical methods to solve op-
gation functions34 allow providing timal control problems fall into three
globally optimal solutions; unfortu- main classes. Dynamic programming
nately, these methods require an ex- implements the Bellman optimality
plicit representation of obstacles in principle saying that any sub-motion
the configuration space, which is in- of an optimal motion is optimal. This
formation that is generally not avail- leads to a partial differential equation
able. At this stage, we can see how the (the so-called Hamilton-Jacobi-Bell-
presence of obstacles makes optimal man equation in continuous time)
path planning a difficult problem. whose solutions may sometimes be
computed numerically. However, dy-
Optimal Motion Planning namic programming suffers from the
In addition to obstacle avoidance, well-known curse of the dimensional-
ity bottleneck. Direct methods consti-
Figure 3. Dubins car. Figure 4. Reeds-Shepp car. tute a second class. They discretize in
time both control and state trajecto-
ries so that the initial optimal control
problem becomes a standard static
non-linear programming (optimiza-
tion) problem of potentially large
size, for which a large variety of meth-
ods can be applied. However, local op-
timality is generally the best one can
hope for. Moreover, potential chatter-
ing effects may appear hidden in the
obtained optimal solution when there
= 1 and 1 1 = 1 and 1 1 is no optimal solution in the initial
function space. Finally, in the third
category are indirect methods based
on optimality conditions provided by
the PMP and for which, ultimately,
the resulting two-point boundary
value problem to solve (for example,
by shooting techniques) may be ex-
Reachable set in the (x, y, )- Reachable set in the (x, y, )- tremely difficult. In addition, the pres-
configuration space configuration space ence of singular arcs requires special-
ized treatments. So direct methods
are usually simpler than indirect ones

86 COMMUNICATIO NS O F TH E AC M | S EPTEM BER 201 4 | VO L . 5 7 | N O. 9

 review articles

even though the resulting problems to according to a given metric. The sim- Recently, successful improvements
solve may be very large. Indeed, their plest technique consists in picking have been introduced by following the
structural inherent sparsity can be pairs of points on the path and linking same basic principle of optimizing an
taken into account efficiently. them by a shortest path: if the shortest initial guess in motion space. Zucker
At this stage, we can conclude that path is collision-free, it replaces the et al. take advantage of a simple func-
exact solutions for optimal motion corresponding portion of the initial tional expressing a combination of
planning remain out of reach. Only path. Doing so iteratively, the path be- smoothness and clearance to obstacles
numerical approximate solutions are comes shorter and shorter. The itera- to apply gradient descent in the trajec-
conceivable. tive process stops as soon as it does not tory space.40 A key point of the method
significantly improve the quality of the is to model a trajectory as a geometric
Optimal Motion Planning path. The technique gives good results object, invariant to parametrization.
Along a Path in practice. In the same framework, Kalakrishman
A pragmatic way to bypass (not over- Beside this simple technique, sev- et al. propose to replace the gradient
come) the intrinsic complexity of the eral variational methods operating descent with a derivative-free stochas-
kinodynamic and numerical approach- in the trajectory space have been in- tic optimization technique allowing us
es is to introduce a decoupled ap- troduced. Among the very first ones, to consider non-smooth costs.19
proach that solves the problem in two Barraquand and Ferbach2 propose to
stages: first, an (optimal) path plan- replace a constrained problem by a What We Know and What We Do
ning generates a collision-free-path; convergent series of less constrained Not Know About Optimal Motion
then a time-optimal trajectory along subproblems increasingly penalizing for Wheeled Mobile Robots
the path is computed while taking into motions that do not satisfy the con- Mobile robots constitute a unique
account robot dynamics and control straints. Each sub-problem is then class of systems for which the ques-
constraints. The resulting trajectory is solved using a standard motion plan- tion of optimal motion is best un-
of course not time-optimal in a global ner. This principle has been success- derstood. Since the seminal work by
sense; it is just the best trajectory for fully extended recently to humanoid Dubins in the 1950s,13 optimal mo-
the predefined path. From a compu- robot motion planning.9 tion planning and control for mobile
tational point of view, the problem is Another method introduced by robots has attracted a lot of interest.
much simpler than the original global Quinlan and Khatib consists in mod- We briefly review how some challeng-
one because the search space (named eling the motion as a mass-spring sys- ing optimal control problems have
phase plane) is reduced to two dimen- tem.32 The motion then appears as an been solved and which problems still
sions: the curvilinear abscissa along elastic band that is reshaped according remain open.
the path and its time-derivative. Many to the application of an energy func- Let us consider four control models
methods have been developed since tion optimizer. The method applies for of mobile robots based on the model
the introduction of dynamic program- nonholonomic systems as soon as the of a car (Figure 2). Two of them are
ming approaches by Shin and McKay36 nonholonomic metric is known16 as simplified models of a car: the so-
in configuration space and simultane- well as for real-time obstacle avoidance called Dubins (Figure 3) and Reeds-
ously by Bobrow et al.4 in the Cartesian in dynamic environments.6 Shepp (Figure 4) cars respectively. The
space. Many variants have been con-
sidered including the improvement by Figure 5. The Hilare robot at LAAS-CNRS in the 1990s.
Pfeiffer and Johanni31 that combines
forward and backward integrations,
and the recent work by Verscheure et
al.39 who transform the problem into a
convex optimization one.

Optimization in Motion Space

Extensions of path-tracking methods
may be considered as soon as we al-
low the deformations of the support-
ing paths. Here, we assume some mo-
tion planner provides a first path (or
trajectory). Depending on the motion
planner, the path may be far from be-
ing optimal. For instance, probabilistic
motion planners introduce many use-
less detours. This is the price to pay for
their effectiveness. So, the initial path
must be reshaped, that is, optimized
with respect to certain criteria. Geo-
metric paths require to be shortened

SE PT E MB E R 2 0 1 4 | VO L. 57 | N O. 9 | C OM M U N IC AT ION S OF T HE ACM 87
review articles

Dubins car moves only forward. The driving wheel mobile robot:b in the two examples of non-linear systems
Reeds-Shepp car can move forward first one (Hilare-1), the controls are for which we know exactly the struc-
and backward. Both of them have the linear velocities of the wheels; in ture of the optimal trajectories. Note
a constant velocity of unitary abso- the second one (Hilare-2), the controls that in both examples the norm of the
lute value. Such models account for are the accelerations (that is, the sec- linear velocity is assumed to be con-
a lower bound on the turning radius, ond system is a dynamic extension of stant. In those cases, time-optimal
that is, the typical constraint of a car. the first). trajectories are supported by the cor-
Such a constraint does not exist for a Time-optimal trajectories. The car- responding shortest paths. Dubins
two-wheel differentially driven mobile like robots of figures 3 and 4 represent solved the problem for the car moving
robot. This robot may turn on the spot only forward.13 More than 30 years lat-
while a car cannot. Let us consider b The distance between the wheels is supposed er, Reeds and Shepp33 solved the prob-
two simple control schemes of a two- to be 2. lem for the car moving both forward
and backward. The problem has been
Figure 6. Hilare-1: A different drive mobile robot. First model: The controls are the velocities completely revisited with the modern
of the wheels. The optimal controls are bang-bang. Optimal trajectories are made of pure
rotations and of straight-line segments.
perspective of geometric techniques
in optimal control theory:37,38 the appli-
cation of PMP shows that optimal tra-
2 2
jectories are made of arcs of a circle of
minimum turning radius (bang-bang
controls) and of straight-line segments
(singular trajectories). The complete
1 1
structure is then derived from geo-
metric arguments that characterize
the switching functions. The Dubins
Rotation Straight-line and Reeds-Shepp cars are among the
on a spot segment few examples of nonlinear systems for
which optimal control is fully under-
stood. The same methodology applies
for velocity-based controlled differen-
Figure 7. Hilare-2: A different drive mobile robot. Second model: The controls are the accel- tial drive vehicles (Hilare-1 in Figure
eration of the wheels. The optimal controls are bang-bang. Optimal trajectories are
made of arcs of clothoids and of arcs of involute of a circle.
6). In that case, optimal trajectories are
bang-bang, that is, made of pure rota-
tions and straight-line segments. The
switching functions are also fully char-
acterized.1 This is not the case for the
dynamic extension of the system, that
is, for acceleration-based controlled
differential drive vehicles (Hilare-2 as
shown in Figure 7). Only partial results
are known: optimal trajectories are
2 2 bang-bang (that is, no singular trajec-
tory appears) and are made of arcs of
clothoid and involutes of a circle.15
However, the switching functions are
1 1 unknown. The synthesis of optimal
control for the Hilare-2 system the re-
main an open problem.
While the existence of optimal
Clothoid Involute of a trajectories is proven for the four sys-
circle
tems shown here, a last result is worth
mentioning. If we consider the Reeds-
Shepp car optimal control problem in
presence of an obstacle, even if we al-
low the car touching the obstacles, it
has been proven that a shortest path
may not exist.10
Motion planning. These results are
very useful for motion planning in the
presence of obstacles. In figures 3 and
4 we display the reachable domain

88 COM MUNICATIO NS O F TH E ACM | S EPTEM BER 201 4 | VO L . 5 7 | N O. 9

 review articles

for both the Dubins and Reeds-Shepp theoretical results. Int. J. of Control 74, 8 (2001), 776782.
18. Jurdjevic, V. Geometric Control Theory. Cambridge
cars. While the reachable set of the The distinction between optimal University Press, 1996.
Reeds-Shepp car is a neighborhood of and optimized motions as it is ad- 19. Kalakrishnan, M., Chitta, S., Theodorou, E., Pastor,
P. and Schaal, S. Stomp: Stochastic trajectory
the origin, it is not the case for the Du- dressed in this article does not cover all optimization for motion planning. In Proceedings of the
bins car. Stated differently, the Reeds- facets of optimality in robot motion. In IEEE Int. Conf. on Robotics and Automation (2011).
20. Kavraki, L., Svestka, P., Latombe, J.-C. and Overmars,
Shepp car is small-time controllable, a companion article,23 we consider the M. Probabilistic roadmaps for path planning in high-
while the Dubins car is only controlla- issue of motion optimal as an action dimensional configuration spaces. IEEE Trans. on
Robotics and Automation 12, 4 (1996), 566580.
ble. The consequence in terms of mo- selection principle and we discuss its 21. Khatib, O. Real-time obstacle avoidance for
tion planning is important. In the case links with machine learning and recent manipulators and mobile robots. The Int. J. of
Robotics Research 5, 1 (1986), 9098, 1986.
of the Reeds-Shepp car, any collision- approaches to inverse optimal control. 22. Latombe, J.-C. Robot Motion Planning. Kluwer
Academic Press, 1991.
freenot necessarily feasiblepath 23. Laumond, J., Mansard, N. and Lasserre, J. Robot
can be approximated by a sequence of Acknowledgments motion optimization as action selection principle.
Commun. ACM, (to appear).
collision-free feasible paths. Optimal The article benefits from comments 24. Laumond, J.-P., Jacobs, P., Taix, M. and Murray, R.
paths allow building the approxima- by Quang Cuong Pham, from a care- A motion planner for nonholonomic mobile robots.
IEEE Trans. on Robotics and Automation, 10, 5 (1994),
tion, giving rise to an efficient motion- ful reading by Joel Chavas, and above 577593.
planning algorithm.24 Not only does all, from the quality of the reviews. The 25. Laumond, J.-P. and Simon, T. Notes on visibility
roadmaps and path planning. New Directions in
such an algorithm not apply for the Du- work has been partly supported by ERC Algorithmic and Computational Robotics. B. Donald, K.
bins car, we still do not know whether Grant 340050 Actanthrope, by a grant Lynch, and D. Rus, eds. A.K. Peters, 2001.
26. LaValle, S. Planning Algorithms. Cambridge University
the motion-planning problem for Du- of the Gaspar Monge Program for Op- Press, 2006.
bins car is decidable or not. timization and Operations Research of 27. LaValle, S. and Kuffner, J. Rapidly-exploring random
trees: Progress and prospects. Algorithmic and
In Laumond et al.,24 we prove the the Fdration Mathmatique Jacques Computational Robotics: New Directions. B. Donald, K.
number of maneuvers to park a car var- Hadamard (FMJH) and by the grant Lynch and D. Rus, eds. A.K. Peters, 2001, 293308.
28. Lozano-Prez, T. Spatial planning: A configuration
ies as the inverse of the square of the ANR 13-CORD-002-01 Entracte. space approach. IEEE Trans. on Computer 32, 2
clearance. This result is a direct conse- (1983), 108120.
29. Neustadt, L. Optimization, a moment problem and non
quence of the shape of the reachable References linear programming. SIAM J. Control (1964), 3353.
1. Balkcom, D. and Mason, M. Time optimal trajectories 30. Paul, R. Robot Manipulators: Mathematics,
sets. So, the combinatorial complex- for differential drive vehicles. Int. J. Robotics Research Programming, and Control. MIT Press, Cambridge,
ity of (nonholonomic) motion plan- 21, 3 (2002), 199217. MA, 1st edition, 1982.
2. Barraquand, J. and Ferbach, P. A method of progressive 31. Pfeiffer, F. and Johanni, R. A concept for manipulator
ning problems is strongly related to constraints for manipulation planning. IEEE Trans. on trajectory planning. IEEE Journal of Robotics and
Robotics and Automation 13, 4 (1997) 473485.
optimal control and the shape of the 3. Barraquand, J. and Lacombe, J.-C. Robot motion
Automation 3, 2 (1987).
32. Quinlan, S. and Khatib, O. Elastic bands: Connecting
reachable sets in the underlying (sub- planning: A distributed representation approach. path planning and control. In Proceedings of the IEEE
Intern. J. of Robotics Research 10, 6 (1991), 628649.
Riemannian) geometry.17 4. Bobrow, J., Dubowsky, S. and Gibson, J. Time-optimal
Int. Conf. on Robotics and Automation (1993).
33. Reeds, J. and Shepp, L. Optimal paths for a car that
control of robotic manipulators along specified paths. goes both forwards and backwards. Pacific Journal of
Int. J. of Robotics Research 4, 3 (1985), 317.
Conclusion 5. Brady, M., Hollerbach, J., Johnson, T., Lozano-Prez, T.
Mathematics 145, 2 (1990), 367393.
34. Rimon, E. and Koditschek. Exact robot navigation using
When optimal solutions cannot be and Masson, M.T. Robot Motion: Planning and Control. artificial potential fields. IEEE Trans. on Robotics and
MIT Press, 1983.
obtained for theoretical reasons (for 6. Brock, O. and Khatib, O. Elastic strips: A framework for
Automation 8, 5 (1992), 501518.
35. Schwartz, J. and Sharir, M. On the piano movers
example, nonexistence) or for practi- motion generation in human environments. Int. J. of problem II: General techniques for computing
Robotics Research 21, 12 (2002), 10311052. topological properties of real algebraic manifolds.
cal ones (for example, untractability), 7. Choset, H., Lynch, K.M., Hutchinson, S., Kantor, A., Advances of Applied Mathematics 4 (1983), 298351.
we have seen how the problem can be Burgard, W., Kavraki, L.E. and Thrun, S. Principles of 36. Shin, K.G. and McKay, N.D. Minimum-time control of
Robot Motion: Theory, Algorithms, and Implementations. robotic manipulators with geometric path constraints.
reformulated either by considering a MIT Press, Cambridge, MA, June 2005. IEEE Trans. on Automatic Control 30, 6 (1985), 531541.
discrete representation of space and/ 8. Connolly, C. and Grupen, R. Applications of harmonic 37. Soures, P. and Laumond, J.-P. Shortest paths
functions to robotics. J. of Robotic Systems 10, 7 synthesis for a car-like robot. IEEE Trans. on
or time, or by slightly changing the op- (1992), 931946. Automatic Control 41, 5 (1996), 672688.
timization criterion, or by resorting to 9. Dalibard, S., Nakhaei, A., Lamiraux, F. and Laumond, 38. Sussmann, H. and Tang, G. Shortest paths for the
J.-P. Whole-body task planning for a humanoid robot: A Reeds-Shepp car: A worked out example of the use
numerical optimization algorithms. way to integrate collision avoidance. In Proceedings of of geometric techniques in nonlinear optimal control.
IEEE-RAS Int. Conference on Humanoid Robots, 2009.
In all these cases, the resulting solu- 10. Desaulniers, G. On shortest paths for a car-like
Rutgers Center for Systems and Control Technical
Report 91-10, 1991.
tions are only approximated solutions robot maneuvering around obstacles. Robotics and 39. Verscheure, D., Demeulenaere, B., Swevers, J.,
Autonomous Systems 17 (1996), 139148.
of the original problem. 11. Diehl, M. and Mombaur, K., eds. Fast Motions in
Schutter, J.D. and Diehlm M. Time-optimal path
tracking for robots: A convex optimization approach.
In conclusion, it appears the ex- Biomechanics and Robotics. Lecture Notes in Control IEEE Trans. on Automatic Control 54, 10 (2009),
and Information Sciences, vol 340. Springer, 2006.
istence of optimal robot motions is 12. Donald, B., Xavier, P., Canny, J. and Reif, J.
23182327.
40. Zucker, M., Ratliff, N., Dragan, A., Pivtoraiko, M.,
rarely guaranteed. When it is, finding Kinodynamic motion planning. JACM 40, 5 (1993), Klingensmith, M., Dellin, C., Bagnell, J. and Srinivasa,
10481066.
a solution has never been proven to be 13. Dubins, L. On curves of minimal length with a
S. Chomp: Covariant Hamiltonian optimization for
motion planning. Int. J. of Robotics Research 32, 9-10
a decidable problem as is the motion- constraint on average curvature and with prescribed (2013), 11641193.
initial and terminal positions and tangents. Amer. J. of
planning problem. So, optimal mo- Mathematics 79 (1957), 497516.
tion is most often an expression that 14. Hopcroft, J. Schwartz, J. and Sharir, M. Planning, Jean-Paul Laumond (jpl@laas.fr) is a CNRS director of
Geometry, and Complexity of Robot Motion. Ablex, 1987. research at LAAS, Toulouse, France.
should be understood as optimized 15. Jacobs, P., Laumond, J.-P. and Rege, A. Nonholonomic
motion, that is, the output of an op- motion planning for HILARE-like mobile robots. Nicolas Mansard (nmansard@laas.fr) is a CNRS
Intelligent Robotics. M. Vidyasagar and M.Trivedi, eds. researcher at LAAS, Toulouse, France.
timization numerical algorithm. How- McGraw Hill, 1991.
16. Jaouni, H., Khatib, M. and Laumond, J.-P. Elastic bands Jean-Bernard Lasserre (lasserre@laas.fr) is CNRS
ever, motion optimization techniques for nonholonomic car-like robots: algorithms and director of research at LAAS, Toulouse, France.
follow progress in numerical optimiza- combinatorial issues. In Robotics: The Algorithmic
Perspective. P. Agarwal, L. Kavraki, and M. Mason, eds.
tion with effective practical results on A.K. Peters, 1998.
real robotic platforms, if not with new 17. Jean, F. Complexity of nonholonomic motion planning. 2014 ACM 0001-0782/14/09 $15.00

SE PT E MB E R 2 0 1 4 | VO L. 57 | N O. 9 | C OM M U N IC AT ION S OF T HE ACM 89