1 State of the art in UAV planning

In this chapter, the fundamentals of the path planning of one UAV are discussed. We
will distinguish between continuous and discrete planner, pointing out their differences,
advantages and drawbacks. One example of each type of planner is presented and
tested with several simulation results. In order to fix ideas, a a collision-free discrete
trajectory planning algorithm for an UAV in a common airspace with other aerial
vehicles and static obstacles is presented; and a continuous RRT* trajectory planner
is also presented.

1.1 Statement of the problem

1.2 Graph search method
Historically, the first path planning algorithms consisted on algorithms that found a
path to connect two nodes in a graph. In these early algorithms, it was assumed that
a graph G =< G, V > that unites points in the configuration space in such a way
that G ∈ Cf ree was either introduced by the designer or obtained by discretization.
That is, dividing the configuration space into cells that could be or not collision free
as shown in Figure 1.1.
The first developed algorithm that finds whether two vertices in a graph G =<
V, E > are or not connected is the Depth First Search algorithm. It was introduced
in 1889 by French mathematician Charles Pierre Trmaux as a strategy for solving
mazes. The main idea was to explore one of the paths of the graph until the solution
was found, an already visited node is found or no further paths exist. In such case, a
backtracking process until the first node with unexplored path is carried out exploring
another path of this node.
Another popular algorithm to find connectivity between graphs is the Breadth First
Search algorithm (BFS). This algorithm instead of greedily following a path like DFS,
explores first all the nodes connected to the starting node, then reproduces this proce-
dure in each of the children nodes until the desired node is found or all nodes connected
to the starting node are found. This algorithm has two main advantages to the DFS:
it can obtain the goal node in infinite graphs and in addition it returns the path with
lowest number of edges. However, it is much more memory consuming than DFS.
Note that in path planning problems, BFS algorithm returns the path with minimum
length if all the edges have equal length.
Djikstra algorithm [1] is the first algorithm that efficiently solved the problem of
translating from one cell to another with lowest cost in a graph with weighted edges.
Thus, it can be used for obtaining the lowest cost path that connects the starting and

1
1 State of the art in UAV planning

Figure 1.1: Discretization cell-map obtained in a 2-dimensional space with a triangular

obstacle. In red line the path obtained by using the A* algorithm is shown.

goal nodes. It can be considered an extension of the BFS algorithm to weighted graphs
(i.e. the edges of the graph have associated a cost). This algorithm will explore first
the nodes with lowest distance to the starting node, updating the cost of the successor
of the current node.
One of the main problems of Djisktra algorithm is that it does not take into account
the position of the goal in order to explore the graph. This can yield to unnecessary
node exploring in most situations. In order to overcome this, it was found that some
heuristics regarding to the distance to the goal node could be introduced in the al-
gorithm in order make the exploration of the graph more directed to the goal node
generating the greedy algorithms explore first the nodes which are closer to the goal
node. These considerations would be mixed in the A* algorithm [2] in which the nodes
are first explored not only taking into account the distance to the starting node but
also the estimated distance to the goal node. It has been demonstrated that if the
heuristic function fulfills some requirements, the A* algorithm returns the optimal
path between the starting and goal nodes. This kind of algorithms are often named
as best first search algorithms.
Whenever these algorithms are applied, if not more details are given, the description
of the environment is assumed to be given by means of a matrix (which can be 2-
dimensional or 3-dimensional) where each cell can be blocked or not. Additionally,
it can contain a certain number that indicates how convenient is to travel the cell.
These matrices are usually called cell-maps. They are usually obtained by performing
a discretization of the space with a given resolution. Obviously, this kind of algorithms
are not complete in a strict sense: depending on the resolution of the matrix it will
find or not existing paths from a start to a goal. The classical matrix representation is
also expensive in terms of memory allocation. However, some methods such as quad
and oct-trees [3] can be a solution for generating more memory-efficient cell-maps.
A* algorithm is the one that is most often used in classical planning and has yield
many variations. Hence, it has been adapted to handle relevant problem that where
not taken into account in the original procedure. Next, the most important algorithms

2
 1.3 Obtaining the graph representation

that have been developed taking A* as basis are listed as well as their motivation:

• More directed search. Some works [4] demonstrate that pondering more the
heuristic function than the distance to the starting node results on great run-time
benefits while obtaining paths not noticeably longer.

• Unknown terrain. The above algorithms assume that perfect knowledge of

the environment is given by means of a graph or a cell matrix. However, this
complete knowledge is whether difficult to obtain or can change over the time.
One possible approach is to distinguish between global and local planners. Global
planners will give a complete path between starting and goal nodes while local
planners will avoid collision with unexpected obstacles and could invoke again
the global planner if the original path is completely blocked [5]. The development
of D* algorithm [6] proposes to fuse the two planners into a common procedure
and thus simplifying the planning process. This idea has proved to be successful
and some variants such as Focused D* [7] and D* Lite [8] (it is important to note
that is not developed from D* but from A*) are still used in present days.

• Any Angle approaches. The main drawback of A* algorithm when used in

cell-maps is that it returns a path which is not smooth and their directions are
restricted to the directions of the cell map (see figure 1.1. So the resulting path
is not realistic. A typical solution is to apply a post-processing algorithm to the
returned path where some smoothing techniques are applied, but the procedure
can fail and no guaranties of finding optimal smoothed paths are given. In
order to overcome this, any angle approaches have been proposed like Theta*
[9], its lazy variant [10] and Block A* [11] which can give paths whose headings
are not defined by the directions of the cells and can be smoother more easily.
In addition, an adaptation of Theta* algorithm to handle unknown terrain is
proposed with the incremental Phi* planner [12].

1.3 Obtaining the graph representation

In section 1.2 it was assumed that the graph that unites collision free configurations of
the robot is given in advance. In addition, a naive implementation for obtaining this
graph is given by means of a cell map. In this case, the space is discretized in cells
and each cell can be free or obstaculized.
However, obtaining and storing these cell maps in 3D environments and in large areas
can be time and memory consumpting. In addition, the time required to perform an
A* search grows with the number of nodes expanded and with the branching factor
(average number of successors per node) of the graph. A tradeoff decision between the
resolution of the grid and the completeness of the algorithm is needed.
In this section, other methods for generating the graph that describes the config-
uration space that are collision-free are described. Firstly, we will explain the exact
methods that give compact representation of a configuration space with the presence
of polygonal obstacles. These techniques present some drawbacks that yielded to the

3
1 State of the art in UAV planning

development of probabilistic methods. Last, an extension of the probabilistic methods

for optimal planning is given.

1.3.1 Exact graph generation methods

Alongside with the naive grid discretitation method there are some methods for ob-
taining a graph in configuration spaces with polygonal obstacles. These methods are
listed below:

• Exact Cell Decomposition Methods. In these methods the common idea is

to divide the free space into areas which are called cells. The centroid of each
cell, as well as the middle point of each division, are inserted as nodes in the
graph and then nodes belonging to neighbouring nodes are connected. These
methods include vertical and horizontal cell decomposition, see Figure 1.2a.

• Visibility Graph(VG). This method can be used for finding shortest path in
a configuration space with polygonal obstacles. In this case, each vertex of each
obstacle is added to the graph and connected with other nodes if they have line
of sight as shown in Figure 1.2b.

• Voronoi Diagrams (VD). The visibility graph method can find the shortest
path but its main drawback is that can make the robot move close to the ob-
stacles. One technique to avoid this is to artificially expand the obstacles to a
desired clearance (minimum distance to the obstacles) taking into account the
shape of the robot or a safety area in which the robot is wrapped, as shown in
Figure 1.3. If maximum clearance paths are necessary, they can be found by gen-
erating the Voronoi Diagram of the obstacles and adding to the graph the points
where three or more obstacles are at the same distance (see figure 1.2c). This
can be of great use if only uncertain knowledge of the environment is available.
The main drawback of the method based in Voronoi Diagrams is that staying as
close as possible from obstacles can result on too conservative paths.

Note that the above algorithms are formulated in the Cartesian space and can be
solved easily in a 2D Cartesian space. However, the 3-dimensional version of the
Euclidean shortest path problem is much harder. In this case, the graph shortest path
may not transverse the edges of the polyhedral obstacles but rather any point in the
edges of the obstacles. This problem has been demonstrated to be NP-hard [13].
However, in some cases we are interested in obtaining the graph in the Configuration
space, i.e. taking into account not the position in the space but the configuration of
the system. For example, a configuration of a non-circular robot moving in a 2D space
can be defined by three coordinates: (x, y, θ), where θ represents the orientation of
the robot. Hence, the configuration space of this type of robots is of dimension three.
The things become much harder when planning with robots moving in a 3D space. In
this case the configuration can be determined by the vector (x, y, z, φ, θ, ψ), where φ, θ
and ψ represent zyx-Euler angles which are called roll, pitch and yaw respectively (see
figure 1.4). Note that extending this to 3-dimensional multi-robot motion planning

4
 1.3 Obtaining the graph representation

Figure 1.2: (a) Exact cell decomposition in 2-D environment with polygonal obstacles.
(b) Visibility graph. (c) Voronoi graph, note that the graph vertices are
not straight segments in general.

Figure 1.3: Expansion of the obstacles according to the shape of the robot and visibility
graph generated with the expanded obstacles.

5
1 State of the art in UAV planning

Figure 1.4: zyx-Euler angles for a robot moving in a 3-dimensional space.

yields to systems with configuration spaces with tenths of dimensions in which the
above listed exact methods cannot be applied.

1.3.2 Probabilistic Roadmaps

In order to overcome the limitations of exact methods in systems with high dimen-
sional Configuration Spaces, probabilistic methods for generating the graph that de-
scribes Cf ree where introduced in [14]. The basic method is called Probabilistic
Roadmaps(PRM) and is divided in two steps: the graph generation step and the
query step.
The first stage is the graph generation step, in which the graph G is computed by
randomly generating nodes in Cf ree and connecting them with close enough existing
nodes in G. Note that in this case, the nodes are not generated taking into account the
geometry of the problem with the exception that the nodes and edges must be collision
free. So the problem is splitted into two blocks: the graph generator block which do
not have any geometric description of the environment and the collision detector block
that reports to the graph generator whether a configuration or a path between two
configurations is or not collision-free.
The second stage is the query step. In this step, the initial and goal configurations
(qinit and qgoal ) of the problem to be solved are added and connected to G. Then,
the shortest path is obtaining by applying one of the methods described in section
1.2 to G. It is important to note that this query step can be performed in tenths of
seconds or seconds depending on the size of G, while the computation of G is much
more computationally demanding.
This approach has been modified in order to reduce the complexity of the generated
graph in the Visibility-based Probabilistic Roadmaps (v-PRM)[15]. Gaussian sampling
has been proposed in order to encourage connectivity in scenarios with narrow corridors
[16]. Also, a lazy-PRM approach in which edge collision checking is delayed until
the query step is introduced with interesting runtime improvements when performing
collision checks is computationally expensive [17].

6
 1.3 Obtaining the graph representation

1.3.3 Rapidly-exploring Random Trees

The above mentioned methods, Probabilistic Roadmaps, perform an exahustive ex-
ploration of the configuration space in order to generate a Graph that represents this
space. This graph will be used in the query phase each time a path planning problem
is being solved. This indicates that PRM is a multiple query algorithm. In contrast,
simple query algorithms generates the graph starting from the starting configuration
of the problem and the exploration is usually stopped when the goal configuration is
reached or is sufficiently close. This graph will be generated whenever a path planning
problem is formulated and will be discarded when the problem is solved.
Rapidly-exploring Random Trees (RRT) [18] was the first simple query path plan-
ner that generates a tree shaped graph G ∈ Cf ree . This tree will rapidly cover the
configuration space until the goal configuration is sufficiently close from the tree.
The basic RRT algorithm is summarized in algorithm 1. It starts a tree by creating
the root in qinit (step 1) and extends the tree by generating random samples (qrand )
of the configuration space and by making the tree extend towards qrand (steps 3-5).
This extend procedure is usually referred is usually done by interpolation in the basic
RRT algorithm: the new vertex (qnew ) will be located at a distance ∆q in the direction
towards qrandom . Matematically:
q new − q near
qnew = ∆q (1.1)
||q new − q near ||
If the path between qnear and qnew is collision-free, this node is added to the tree
(steps 6-10). This procedure is repeated until the distance between the new node and
final state qgoal goes below dmin (step 11). Note that some procedures are necessary
for the algorithm to be run. Below you can find the list of procedures.

• Nearest(G, q). Searches for the closest vertex in the graph G to the configura-
tion q.
• Extend(q1 , q2 ). Obtains a configuration q3 from q1 towards q2 . This can be
done by interpolation (basic RRT) or by integrating the model of the robot in
Control Space RRT (CS-RRT).
• CollisionFree(q1 , q2 ). Returns true if the path that unites q1 and q2 is collision
free.
• qrand =SampleFree(). Returns a configuration qrand ∈ Cf ree .

Like PRM algorithms, RRT are claimed to be probabilistically complete: the prob-
ability that the generated tree will be closer than a minimum distance of the goal
node with probability that converges to zero with increasing time [19]. RRT is also
capable to generate paths that meet kinodynamic contraints: in this case the tree will
be extended by integrating a model of the vehicle for a determinate amount of time
∆t . The control inputs will usually be generated randomly, however it is possible to
select randomly several control inputs, selecting the one that generates the nearest
node to qrandom . Planning in the Control Space of the robot (CS-RRT) requires more

7
1 State of the art in UAV planning

computational power than the basic RRT but it yields to feasible paths in vehicles
with kinodynamic constraints.
Several variations of the original RRT have been developed. In goal-biased RRT the
goal is selected as the qrandom with some probability, encouraging the growth of the
tree towards the goal. Bidirectional RRT (bi-RRT) keeps the same principle as RRT,
but it starts one extra tree from the goal configuration and then attempts to unite the
trees. Procedure to encourage the connection between trees have been proposed [20].

Algorithm 1 Basic RRT algorithm

Require: RRT(qinit , qgoal )
1: V ← {xinit }; E ← ∅
2: repeat
3: qrand ← SampleF ree()
4: qnearest ← N earest(G = (V, E), xrand )
5: qnew ← Extend(qnearest , qnew )
6: if CollisionF ree(qnearest , qnew ) then
7: // Add the new vertex and the connection
8: V ← V ∪ {qnew }
9: E ← E ∪ {(qnearest , qnew )}
10: end if
11: until Dist(qgoal , G) < dmin
12: return G = {V, E}

1.3.4 Optimal probabilistic methods

The main problem of the basic probabilistic methods such as PRM and RRT is that
even though they generate paths that unite qinit and qgoal without collisions, no consid-
erations regarding the quality of the path are introduced. Thus, they have been proved
useful when generating paths in problems with high dimensionality and in cluttered
environments. But their lack of consideration regarding to the quality of the path was
evident when they were applied to mobile robot path planning. Their generated path
yielded to randomized like motions that were not properly optimized and were difficult
to forecast.
Transition based RRT (t-RRT) [21] were proposed in order to generate paths with
better quality when planning on Configuration-space with an associated Costmap. Its
strategy is to sample first in the zones of Cf ree that have lowest cost (i.e. valleys of the
cost function) and gradually broaden the sampling region if until a solution is found.
This approach has been tested to obtain paths with more quality than the original
RRT procedure. However, once the path is obtained, no further improvements on it
are performed, so optimal convergence is not assured.
Optimal convergence of a probabilistic planner was first proved in [22], where RRT*
was proposed. This planner is found to be both probabilistic complete and asympoti-
cally optimal. RRT* planning algorithm makes two main modifications to the original

8
 1.4 Evolutionary and Swarm Optimization Applied to Path Planning

algorithm as shown in algorithm 2.

First, when a new sample is generated, the algorithm attempts to connect it not
only to the nearest neighbour but also to a set of neighbours that are close enough.
Only the connection that optimizes the path between the new sample and the starting
configuration is added to the tree (steps 9-16).
The other modification is called the rewiring step. In this phase, the current cost of
the neighbours of the new sample is compared to the cost that would be obtained by
traveling through the new sample. If this new cost is less than the current cost, the
graph is rewired (steps 20-23).
Some extra functions are necessary for RRT* algorithm to work. These are:

• Cost(n ∈ V ). Associates the node n with its calculated cost.

• c(Path). Gives a cost to a calculated past. In the basic version the cost is the
distance of the path.
• Near(G, q, d). Returns a set of vertices N = {n ∈ V \dist(n, q) < d}.

Note that the basic RRT* algorithms uses interpolation in the extend procedure.
The use of this planner in Control Space is not straightforward, one approach obtained
by linearizing the system in the surroundings of qnew is found in [23].
In recent years, the multi-core capabilities of new processors have encouraged the
paralelization of the path planning techniques. The most interesting procedure is found
by generating more than one tree at the same time; that is, generating a forest of trees
between qinit and qgoal . The approach proposed in [24] claims to achieve Superlinear
speedup to the original RRT* algorithm, i.e. the profit of using multiple cores is
greater than the number of cores used. This is achieved by sharing the best paths
between the different trees, focusing the sampling to the interesting regions when a
valid path has been found and performing tree pruning according to the length of the
best path.

1.4 Evolutionary and Swarm Optimization Applied to

Path Planning
As seen in section ??, the problem of robot path planning for multiple vehicles is NP-
Hard. This implies that no exact algorithm with polynomial complexity can be found
to solve the problem. Some heuristic search algorithms such as evolutionary algorithms
and swarm intelligence based algorithms can return a quasi-optimal solution to these
problem. In this section, the state of the art of the application of this algorithms to
solve the path planning problem is discussed.
Evolutionary algorithms englobe a huge set of general purpose optimization algo-
rithms that are based on the evolution theory of the species that was first proposed in
[25]. This set includes Genetic Algorithms, Genetic Programming (GP), Evolutionary
Programming and many more. On the other hand, Swarm intelligence algorithms en-
globe a huge set of general purpose optimization algorithms that are inspired in the

9
1 State of the art in UAV planning

Algorithm 2 RRT* algorithm

Require: RRT(qinit , qgoal )
1: G = {V, E}
2: V ← {xinit }; E ← ∅
3: repeat
4: qrand ← SampleF ree()
5: qnearest ← N earest(G, qrand )
6: qnew ← Steer(qnearest , qnew )
7: if CollisionF ree(qnearest , qnew ) then
8: V ← V ∪ {qnew }
9: // Connect along a minimum-cost path
10: U ← N ear(G, mqnew , η)
11: qmin ← qnearest ;cmin ← Cost(qnearest ) + c(P ath(qnearest , qnew ));
12: for all u ∈ U do
13: if CollisionF ree(u, qnew ) and Cost(u) + c(P ath(u, qnew )) < cmin then
14: qmin ← u; cmin ← Cost(u) + c(P ath(qmin , qnew ))
15: end if
16: end for
17: E ← E ∪ {(vmin , xnew )}
18: // Rewire vertices
19: for all u ∈ U do
20: if CollisionF ree(qnew , u) and Cost(qnew ) + c(P ath(qnew , u)) < Cost(u)
then
21: qparent ← P arent(u)
22: E ← (E\{(qparent , u)}) ∪ {(qnew , u)}
23: end if
24: end for
25: end if
26: until Dist(qgoal , G) < dmin
27: return G

10
 1.4 Evolutionary and Swarm Optimization Applied to Path Planning

behaviour of natural swarms such as ants, birds, bees and many more. This set in-
cludes Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), Cuckoo
Search Algorithm (CSA) and Artificial Bee Colony (ABC) amongst other. All of these
algorithms share in common that are biologically inspired and that they have found
to be able to solve complex global optimization problems.
Some effort on developing genetic path planning algorithms has already been carried
out. One of the first approaches to the problem, as well as an almost mandatory
reference book, can be found in [26]. In [27], the convenience of the genetic algorithm
optimization to solve the path planning problem is deeply discussed. However, only
simple results on discretized 2D space with no experimental results are presented.
Besides, the computational power available in those days was not enough to make it
possible to develop a planner with reasonable running time. In [28] a novel 2D path
planner is presented. The aim of this approach is to generate a continuous path in
the search space that will be obtained by mixing sinusoidal and semi-sinusoidal paths.
However, no realistic simulations nor experimental work are presented in the paper.
Yet another planner is presented in [29] with the addition of a post processing step
that smooths the path with bezier curves. However, no extension to 3D trajectories is
shown in the paper and no relevant simulation studies nor experiments are given.
Last but not least, several mixed approach to GA-path planning can be found in
the literature. In [30] fuzzy logic is used in the representation of the genome. Also,
some indirect approaches can be found. For example, in [27] it is proposed to use the
GA in order to tune the parameters of a planner based on Artificial Fields.
Methods based on Ant Colony Optimization (ACO) algorithms have also been pro-
posed [31]. In [32], the application of a game theory approach to airborne conflict
resolution is presented. These techniques present a disadvantage: they are not well
suited for applications that require a high level of scalability for their application in
systems of many UAVs.
On the other hand, PSO application to path planning algorithm is still underdevel-
oped. An application for space vehicles path planning is presented in [33]. In addition,
some efforts a planner with efficient replanning capabilities is presented in [34] and ap-
plied to mobile robots in dynamic environments. However this approach is only shown
in simulation and no significant studies about its optimality and safety are given. One
mixed approach that uses PSO to optimize the graph generation of a standard PRM
algorithm can be found in [35].

Sevilla, 2014

11
Bibliography
[1] E. W. Dijkstra, “A note on two problems in connexion with graphs,” Numerische
Mathematik, vol. 1, pp. 269–271, 1959.

[2] N. J. R. B. Hart, P. E.; Nilsson, “A formal basis for the heuristic determination of
minimum cost paths”,” IEEE Transactions on Systems Science and Cybernetics,
vol. 4, no. 2, p. 100107, 1968.

[3] D. Meagher, “Geometric modeling using octree enconding,” vol. 19, pp. 129–148,
1982.

[4] J. Pearl, Heuristics: Intelligent Search Strategies for Computer Problem Solving.
Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc., 1984.

[5] Y. Goto and A. Stentz, “The cmu system for mobile robot navigation,” in Robotics
and Automation. Proceedings. 1987 IEEE International Conference on, vol. 4,
Mar 1987, pp. 99–105.

[6] A. T. Stentz, “Optimal and efficient path planning for partially-known environ-
ments,” in Proceedings of the IEEE International Conference on Robotics and
Automation (ICRA ’94), vol. 4, May 1994, pp. 3310 – 3317.

[7] A. Stentz, “The focussed d* algorithm for real-time replanning,” in

Proceedings of the 14th International Joint Conference on Artificial In-
telligence - Volume 2, ser. IJCAI’95. San Francisco, CA, USA: Mor-
gan Kaufmann Publishers Inc., 1995, pp. 1652–1659. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1643031.1643113

[8] S. Koenig and M. Likhachev, “D*lite,” in Eighteenth National Conference

on Artificial Intelligence. Menlo Park, CA, USA: American Associa-
tion for Artificial Intelligence, 2002, pp. 476–483. [Online]. Available:
http://dl.acm.org/citation.cfm?id=777092.777167

[9] K. Daniel, A. Nash, S. Koenig, and A. Felner, “Theta*: Any-angle path planning
on grids,” Journal of Artificial Intelligence Research, vol. 39, no. 1, pp. 533–579,
2010.

[10] A. Nash, S. Koenig, and C. Tovey, “Lazy theta*: Any-

angle path planning and path length analysis in 3d,” in
AAAI Conference on Artificial Intelligence, 2010. [Online]. Available:
https://www.aaai.org/ocs/index.php/AAAI/AAAI10/paper/view/1930/1945

13
Bibliography

[11] P. Yap, N. Burch, R. C. Holte, and J. Schaeffer, “Block a*: Database-driven

search with applications in any-angle path-planning,” in Proceedings of the
Twenty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2011, San
Francisco, California, USA, August 7-11, 2011, 2011. [Online]. Available:
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3449
[12] A. Nash, S. Koenig, and M. Likhachev, “Incremental phi*: Incremental any-angle
path planning on grids.” in IJCAI, 2009, pp. 1824–1830.
[13] J. Canny, The Complexity of Robot Motion Planning, ser. ACM
doctoral dissertation award. MIT Press, 1988. [Online]. Available:
http://books.google.es/books?id= VRM sczrKgC
[14] L. Kavraki, P. Svestka, J.-C. Latombe, and M. Overmars, “Probabilistic roadmaps
for path planning in high-dimensional configuration spaces,” Robotics and Au-
tomation, IEEE Transactions on, vol. 12, no. 4, pp. 566–580, Aug 1996.

[15] C. Nissoux, T. Simeon, and J.-P. Laumond, “Visibility based probabilistic

roadmaps,” in Intelligent Robots and Systems, 1999. IROS ’99. Proceedings. 1999
IEEE/RSJ International Conference on, vol. 3, 1999, pp. 1316–1321 vol.3.
[16] V. Boor, M. Overmars, and A. van der Stappen, “The gaussian sampling strategy
for probabilistic roadmap planners,” in Robotics and Automation, 1999. Proceed-
ings. 1999 IEEE International Conference on, vol. 2, 1999, pp. 1018–1023 vol.2.
[17] R. Bohlin and L. E. Kavraki, “Path planning using lazy prm,” in IEEE Interna-
tional Conference on Robotics and Automation, 2000, pp. 521–528.
[18] S. M. Lavalle, “Rapidly-exploring random trees: A new tool for path planning,”
Tech. Rep., 1998.
[19] S. M. Lavalle, J. J. Kuffner, and Jr., “Rapidly-Exploring Random Trees: Progress
and Prospects,” in Algorithmic and Computational Robotics: New Directions,
2000, pp. 293–308.
[20] J. Kuffner and S. LaValle, “Rrt-connect: An efficient approach to single-query
path planning,” in Robotics and Automation, 2000. Proceedings. ICRA ’00. IEEE
International Conference on, vol. 2, 2000, pp. 995–1001 vol.2.
[21] L. Jaillet, J. Cortes, and T. Simeon, “Transition-based rrt for path planning in
continuous cost spaces,” in Intelligent Robots and Systems, 2008. IROS 2008.
IEEE/RSJ International Conference on, Sept 2008, pp. 2145–2150.

[22] S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion plan-
ning,” International Journal of Robotics Research, vol. 30, pp. 1–76, 2011.
[23] A. Perez, R. Platt, G. Konidaris, L. Kaelbling, and T. Lozano-Pérez,
“Lqr-rrt*: Optimal sampling-based motion planning with automatically derived
extension heuristics,” in Proceedings of the IEEE International Conference

14
 Bibliography

on Robotics and Automation, May 2012, pp. 2537–2542. [Online]. Available:

http://lis.csail.mit.edu/pubs/perez-icra12.pdf

[24] M. Otte and N. Correll, “C-forest: Parallel shortest path planning with superlin-
ear speedup,” Robotics, IEEE Transactions on, vol. 29, no. 3, pp. 798–806, June
2013.

[25] C. Darwin, On the Origin of Species by Means of Natural Selection. London:

Murray, 1859, or the Preservation of Favored Races in the Struggle for Life.

[26] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learn-

ing, 1st ed. Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc.,
1989.

[27] S. Kent, “Evolutionary approaches to robot path planning,” Doctor of Philosophy,

Brunel University, 1999.

[28] M. Gerke, “Genetic path planning for mobile robots,” in American Control Con-
ference, 1999. Proceedings of the 1999, vol. 4, 1999, pp. 2424–2429 vol.4.

[29] O. Sahingoz, “Flyable path planning for a multi-uav system with genetic algo-
rithms and bezier curves,” in Unmanned Aircraft Systems (ICUAS), 2013 Inter-
national Conference on, May 2013, pp. 41–48.

[30] K. Deb, D. K. Pratihar, and A. Ghosh, “Learning to avoid moving obstacles

optimally for mobile robots using a genetic-fuzzy approach,” in Proceedings of
Fifth International Conference on Parallel Problems Solving from Nature (PPSN,
1998.

[31] N. Durand and J. Alliot, “Ant colony optimization for air traffic conflict resolu-
tion,” in Proceedings of the Eighth USA/Europe Air Traffic Management Research
and Development Seminar (ATM2009), Napa, (CA, USA), 2009.

[32] P. Masci and A. Tedeschi, “Modelling and evaluation of a game-theory approach

for airborne conflict resolution in omnet++,” in Second International Conference
on Dependability, June 2009.

[33] P. Huang, G. Liu, J. Yuan, and Y. Xu, “Multi-objective optimal trajectory

planning of space robot using particle swarm optimization,” in Advances in Neural
Networks - ISNN 2008, ser. Lecture Notes in Computer Science, F. Sun, J. Zhang,
Y. Tan, J. Cao, and W. Yu, Eds. Springer Berlin Heidelberg, 2008, vol. 5264, pp.
171–179. [Online]. Available: http://dx.doi.org/10.1007/978-3-540-87734-9 20

[34] A. Nasrollahy and H. Javadi, “Using particle swarm optimization for robot path
planning in dynamic environments with moving obstacles and target,” in Com-
puter Modeling and Simulation, 2009. EMS ’09. Third UKSim European Sympo-
sium on, Nov 2009, pp. 60–65.

15
Bibliography

[35] E. Masehian and D. Sedighizadeh, “A multi-objective pso-based algorithm for

robot path planning,” in Industrial Technology (ICIT), 2010 IEEE International
Conference on, March 2010, pp. 465–470.

16