robotic task sequencing problem: a survey
TRANSCRIPT
Noname manuscript No.(will be inserted by the editor)
Robotic Task Sequencing Problem: A Survey
Sergey Alatartsev · Sebastian Stellmacher · Frank Ortmeier
Received: date / Accepted: date
Abstract Today, robotics is an important cornerstone
of modern industrial production. Robots are used for
numerous reasons including reliability and continuously
high quality of work. The main decision factor is the
overall efficiency of the robotic system in the produc-
tion line. One key issue for this is the optimality of the
whole set of robotic movements for industrial applica-
tions, which typically consist of multiple atomic tasks
such as welding seams, drilling holes, etc. Currently,
in many industrial scenarios such movements are opti-
mized manually. This is costly and error-prone. There-
fore, researchers have been working on algorithms for
automatic computation of optimal trajectories for sev-
eral years. This problem gets even more complicated
due to multiple additional factors like redundant kine-
matics, collision avoidance, possibilities of ambiguoustask performance, etc. This survey article summarizes
and categorizes the approaches for optimization of the
robotic task sequence. It provides an overview of exist-
ing combinatorial problems that are applied for robot
task sequencing. It also highlights the strengths and the
weaknesses of existing approaches as well as the chal-
lenges for future research in this domain. The article
S. Alatartsev,Otto-von-Guericke-Universitat,Universitatsplatz 2, D-39106, Magdeburg, GermanyTel.: +49 391-67-52662E-mail: [email protected]
S. StellmacherOtto-von-Guericke-Universitat,Universitatsplatz 2, D-39106, Magdeburg, GermanyE-mail: [email protected]
F. OrtmeierOtto-von-Guericke-Universitat,Universitatsplatz 2, D-39106, Magdeburg, GermanyTel.: +49 391-67-52804E-mail: [email protected]
is meant for both scientists and practitioners. For sci-
entists, it provides an overview on applied algorithmic
approaches. For practitioners, it presents existing solu-
tions, which are categorized according to the classes of
input and output parameters.
Keywords Robotic task optimization · Robotic task
scheduling · Multi-goal path planning · Traveling
Salesman Problem
1 Introduction
Improving the production efficiency is the key factor of
using robots in industry. The faster the robot moves,
the more it produces, the more efficient the produc-
tion process is. Therefore, a lot of research is concen-trated on making robots more productive: making effi-
cient control [72], optimizing trajectories by optimizing
velocity profiles [9] or end-effector path [4], planning
collision-free paths [60]. Mostly these approaches are
oriented on the movement from one point to another.
However, a vast majority of robot’s applications in in-
dustry consists of a set of tasks: taking a set of pictures
or deburring a set of edges. In that case, the sequence
of tasks and the way the robot is moving between them
greatly influence the overall performance. The easiest
way to solve the task sequencing problem is to model it
as the Traveling Salesman Problem (TSP) [6]. The goal
of the TSP is to find a minimal-cost cyclic tour through
a set of points such that every point is visited once. A
simple example of a task sequencing problem modeled
as TSP is shown in Fig. 1. In this scenario the robot
has to cut five holes and deburr the border of the detail.
Every cutting contour is approximated with a point and
a TSP tour is calculated. However, to meet industrial
requirements, multiple additional factors of the robotic
2 Sergey Alatartsev et al.
Fig. 1 Given a set of tasks {A,B,C,D,E,F} that robot has to visit with its end-effector. The goal is to find the optimal/near-optimal task sequence. Here, the optimal sequence is (A,C,E,D,B,F).
site have to be considered, e.g., obstacle avoidance, re-
dundant kinematics, partial order of the sequence, com-
plicated objective function (e.g., time or energy), a set
of possible robot base locations or complicated task ge-
ometry. All these factors make the optimization process
much more difficult and lead to a large search space.
In this survey, we observe papers that solve the
robotic task sequencing problem taking into account
constraints and freedom of the robot’s site. Researchers
referred to this problem differently. In general, there
are two common names: task sequencing/scheduling [5,
55, 56] and multi-goal path planning [49, 89]. Multi-
goal planning considers obstacles in the environment,
therefore, a collision-free path is the output. Task se-
quencing/scheduling may consider obstacles [95] or may
not [56, 94], therefore, the output could be either a se-
quence or a path. In this survey, we refer to the issue of
optimizing robot site in presence of multiple goals/tasks
as the task sequencing problem. This problem is rele-
vant to both service and industrial robotics. However, it
has greater impact for industrial robotics, as in produc-
tion environments the same tasks have to be repeated
multiple time.
The scope of this survey paper is limited to the task
sequencing problem, so that:
– Algorithms for offline sequencing with known envi-
ronment are considered.
– Tasks have no complex logic relations between each
other, e.g., such relations that are based on the task
geometry and task states.
– Robotic site with a single robot is observed.
– Sequencing is done for robot articulated arm. The
approaches for mobile robotics are included only if
they can be easily applied for robotic-arm scenarios.
– The core part of the survey is concentrated on se-
quencing approaches. Collision-free path planning is
considered to be an extra feature.
There are many other sequencing and planning do-
mains which are outside of the survey scope. We outline
them briefly in Section 2.
There are several recent surveys that cover the re-
lated domains, e.g., multi-robot patrolling algorithms
[74], assembly process planning [88], methods to solve
Traveling Salesman Problem [76] and coverage path
planning for robotics [33]. To the best of our knowl-
edge, there is no survey that covers the task sequenc-
ing problem for articulated robots. This survey provides
an overview of related combinatorial TSP-like problems
that are applied in robot task sequencing. We cover the
major milestones in the task sequence optimization for
the last 25 years. The article is useful for both scien-
tist and practitioners. For practitioners, the problems
are categorized by their input and output parameters.
It would allow them to choose the appropriate method
for a particular industrial task with a desired output:
sequence or path. For scientists, the survey allows to un-
derstand what problems exist, which of them are solved
and what perspective research directions exist.
The remainder of the paper is organized as follows.
Section 2 covers related task planning problems. Section
3 outlines the background in robotics and the TSP-like
problems applied in sequencing. The ways of catego-
rization are covered in Section 4. Section 5 outlines the
general ideas of the reviewed approaches. The conclu-
Robotic Task Sequencing Problem: A Survey 3
sion and outlook to the future work is presented in Sec-
tion 6. The acronyms used in this survey are listed in
the Appendix A.
2 Related Planning Problems in Robotics
The survey is concentrated on task sequencing algo-
rithms. In this chapter, we briefly outline other related
planning problems domains.
2.1 Collision-free path planning
In recent years a significant amount of research has been
focused on collision-free planning. The goal is to calcu-
late an optimal path between two robot configurations.
Optimality is expressed with multiple criteria: minimal
traveled distance, time or energy.
Graph-based: One of the most popular collision-free ap-
proaches is Probabilistic Roadmap (PRM) [54]. PRM
allows to construct a graph by random sampling of
collision-free vertices in space. After the graph is ob-
tained, it is used to navigate without collisions from one
configuration to another. PRM is an example of a multi-
request algorithm, as after the roadmap is obtained it
could be reused multiple times. Graph can also be con-
structed as a grid or with Voronoi diagrams [35]. The
navigation in the graph can be made with known graph-
search techniques, e.g., Dijkstra algorithm or A* [93].
Tree-based: Rapidly-growing Random Trees (RRT) [57]
is a single-request approach that builds a tree in a goal
direction. RRT checks for goal reachability from any
branch of the tree within the tree construction process.
Bidirectional RRT (BiRRT) is a variant of RRT, when
two trees are constructed: one is from the starting po-
sition to the goal and the second tree is from the goal
to the starting position.
Artificial potential fields The general idea of the arti-
ficial potential field is to apply two fields: one that at-
tracts to the goal and another that repels from the ob-
stacles [65]. One major drawback of the potential field
approach, is that it can easily stack in the local opti-
mum. Some methods were proposed to avoid this, e.g.,
application of deterministic annealing [26].
Sample based techniques, e.g., RRT/BiRRT or PRM
work well with high-dimensional spaces. Artificial po-
tential fields are more popular for mobile applications.
After the collision-free path is obtained, it is smoothed
and the desired velocity profile is applied in order to ob-
tain the robot trajectory [60]. There exist many path
planning libraries, e.g., MPK [40], MSL [59] or OMPL
[23]. Further information about collision-free path plan-
ners can be found in the surveys [47] and [52]. For
more details and comparison of the collision-free plan-
ners see [32].
The approaches stated above allow to calculate
collision-free paths between two robot configurations
automatically. However, they do not consider the
optimization of a task sequence. Collision-free path
planning is considered as one of the features of robotic
task sequencing problem.
2.2 Production scheduling
Note, that the problem we discuss here is different to
the scheduling cell optimization problem that is often
mentioned in the economics domain. This problem ad-
dresses optimization of a production facility, i.e., how
many robots should be used, how many parts a pro-
duced detail consist of and how to spread the work
evenly between the robots in order to increase the over-
all cell throughput. Due to the high computational com-
plexity of the problem, collision factors and kinematics
are ignored. As a consequence, solutions are obtained
based on Cartesian distance cost, but not cost of the
robot joints. This kind of research is mostly concen-
trated on how to organize production processes, rather
than how to construct optimal movements for a single
machine or robot [24].
This survey paper concentrates on the motion of a
robot, without involving the information on production
process.
2.3 Multi-robot task planning
In the scenarios with multiple robots, the task have to
be distributed among robots. One of the first who inves-
tigate this problem was Maimon [63]. He proposed the
clustering graph-based method based on cell layout and
robot specification. Later, Zhang et al. [97] proposed
the approach for task allocation for a team of robots.
The approach uses Stochastic Clustering Auction tech-
nique along with Simulated Annealing. This allows to
make optimization not only based on the greedy prin-
cipal, but also make worse steps for purpose, to avoid
local minima. The general principal is to recluster the
tasks until termination condition is satisfied. If the clus-
ter is better than the previous, than it is accepted as
a current by following Simulated Annealing strategy.
Further, this approach was improved to be able to solve
4 Sergey Alatartsev et al.
heterogeneous scenarios. The clusters are calculated ac-
cording to different capabilities of different robot mod-
els. Further the efficiency of the approach was improved
by using a Swendsen-Wang method [96]. More details
on multi-robot planning can be found in the survey [74].
In the following of this survey paper task sequencing
approaches for a single robot are presented.
2.4 Task-level planning
In task-level planning, the robot actions are specified
by their interaction with objects. Often the final goal is
known, e.g, “put the object from the box to the table”.
The goal of the task level planning is to find a sequence
of actions that a robot has to perform to modify the en-
vironment from the initial state to the goal state [63].
For example, the solution for this example will consist
of the following actions: 1) “open the box”, 2) “grasp
the object”, 3) “move the object to the table”, 4) “re-
lease the object”. Every single action is then planned
with domain dependent planner. Cao et al. [17] pro-
posed a net called AND/OR used for reasoning about
geometrical task constraints. The general idea is to map
proposed net to Petri net. Then, the solution search
is performed by building a reachability tree from the
Petri net. Later, Chien et al. [18] [19] proposed the ef-
ficient way to incorporate the domain information into
the planner for the indoor robot scenario. The data is
represented with object-oriented model. This model in-
cludes relation between objects, categories and physical
laws. The case study that the solver is capable to solve
was to put all metal parts to the bench in the assembly
room. This type of planning ignores the kinematics and
collision planning. Task-level planning was also applied
to mobile robotics [34], where the hierarchical planning
technique was proposed to reduce the computational
expenses.
In this survey paper, the tasks have no complex logic
relations based on physical laws. In task sequencing,
the tasks are areas that have to be visited, but not
the objects with its handling features as in task-level
planning. However, task sequencing problem may allow
partial order of the sequence.
2.5 Combination of Task-level planning and path
planning
Often, researchers observe task-level and path planning
as a separate problems. However, in real life, the plan
created by the task-level planner often cannot be ex-
ecuted due to the fail of the path-planner. For exam-
ple, the task plan forces the robot to put the cup into
the dish washer, as there is an empty space for it, but
the path-planner cannot construct such a path that al-
lows to perform this motion without collision. In that
case, the task-level planner has to construct a new plan.
Therefore, it is critical to design a task-level planner
in combination with a collision-free planner. Bhatia et
al. [13] proposed the architecture that allows to combine
task-level planning with low-level motion planning by
introducing new synergy level. Gaschler et al. [36] pro-
posed the knowledge volume approach. The main idea
was to make an intermediate stage between continuous
robot motions and symbolic planning.
2.6 Online control-based planning
Sometimes, the constraints for the complex task execu-
tion might be changed online. Sensor might detect these
changes and the control-based approach should reorga-
nize the task execution. Mansard et al. [64] proposed
the approach to sequence tasks to reach the general
goal. The main idea is that a complex task is sepa-
rated into a set of subtasks. The sequence of these sub-
tasks is represented as a stack. The three level control-
architecture is proposed, which maintain the stack of
subtasks to satisfy the constraints. Due to environment
or robot restrictions, controller removes or add a new
task to the stack.
In this survey, the offline approaches for the known
environment are observed.
2.7 Manipulation planning
Another well-known robot related planning problem is
the manipulation planning problem. It occurs when a
robot has to move an object in an environment to the
specific location. The goal of manipulation planning is
to find a sequence of actions the robot has to make to
accomplish this task [60]. Often manipulation planning
is observed in combination with extra degrees of free-
dom or constraints, e.g., manipulation planning with
caging grasp [25] or manipulation of deformable ob-
jects [11]. The further details on manipulation plan-
ning of deformable objects can be found in the follow-
ing survey [50] and for dual-arm manipulation in the
survey [83].
3 Preliminaries
This section describes background on robotics and
related combinatorial problems for modeling task
sequencing.
Robotic Task Sequencing Problem: A Survey 5
3.1 Robotic background
Before starting with the optimization, it is required to
specify what the robot pose is and how it is defined. In
general, two spaces are involved in robotics to describe
the robot pose.
T-space – task space SE(3), is used to designate
the position R3 and the orientation of the end-effector
SO(3), i.e., SE(3) = R3 × SO(3). Normally, it is spec-
ified with homogeneous coordinates [60]:
M =
[R3×3 T3×1
01×3 1
]M is the T-space point that denotes the end-effector
pose, where R is a 3×3 rotation matrix that stands for
the end-effector orientation and T is a 3×1 translation
vector that stands for the end-effector position. As de-
scription the end-effector orientation with a matrix is
not intuitive for a human, often a shorter representa-
tion with Euler angles is used [22]. Euler angles depict a
sequence of rotations about the axes of coordinate sys-
tem. For example, to describe any rotation with Euler
angles in 3D space, three angles are required, i.e., one
angle denotes rotation about one axis.
C-space – configuration space C, is the set of
possible robot joint angles. Standard industrial
robots have 6 degrees of freedom (DOF), therefore,
C = R6. Sometimes it is also called joint or axis
space. The relation between T-space and C-space is
described with two mappings. Forward Kinematics
(FK) takes the robot joint angles and calculates the
corresponding end-effector position and orientation:
FK : C → SE(3). Inverse Kinematics (IK) takes the
end-effector position and orientation and calculates the
set of possible robot configurations: IK : SE(3)→ C.
Often, up to 8 solutions exist for a standard 6-DOF
industrial robot. An example of reaching one task
point with two robot configurations is shown in the
Fig. 2.
3.2 Task sequencing in industrial robotics
Today, there exist two different ways to program robots
in an industrial setting: either it is done online or offline.
In online programming, the robot is directly taught a
movement, which it has to replicate later in produc-
tion mode, e.g., with the lead-through method [82]. Ac-
cordingly, the complete movement (e.g., movement on
welding seams or movement between seams) must be
taught. This is a lengthy process, which requires a lot
of intuition and experience if high quality movements
are desired. The quality of the path depends only on
Fig. 2 The T-space point is (x0, y0, z0, α, β, γ). It is reachedwith two robot configurations: (Θ1, Θ2, Θ3, Θ4, Θ5, Θ6) and(Φ1, Φ2, Φ3, Φ4, Φ5, Φ6).
the experience of the programmer and almost no opti-
mization approaches are involved [14,71].
In contrast, an offline approach is typically based
on simulation. The trajectory that has to be executed
by the real robot later in production is calculated or
taught in a simulated environment. This approach has
some major benefits, e.g., the possibility of testing the
calculated path for potential collisions before real exe-
cution – and, therefore, preventing the expensive hard-
ware from being damaged.
Task sequence optimization is either fully ommited
in offline simulation environments (e.g., RobotWorks1)
or presented with limited functionality (e.g., customiza-
tion for DELMIA2 adapted for drilling applications3).
Customization for DELMIA allows automatic sequenc-
ing of drilling tasks, but applies a simple greedy algo-
rithm, i.e., the next drilling task is always the nearest.
This trivial algorithm does not provide good results, as
it often converges to a local minimum. In addition, the
tasks are specified with one entry robot configuration.
In summary, the task sequencing in industry is of-
ten done manually. In the following, we focus on the
sequencing approaches that are applied in an offline
mode.
3.3 Sequencing problems
Tour-searching combinatorial problems build the foun-
dation for task sequencing in robotics. In this section
the related combinatorial problems are described and
1 Compucraft Ltd, RobotWorks www.compucraftltd.com2 Dassault Systemes http://www.3ds.com3 DELMIA V5 Robotic Drilling Application
http://www.delfoi.com/web/products/delfoi products/en GB/drilling-app/
6 Sergey Alatartsev et al.
Inp
ut:
pri
mit
ive t
asks
GTSPTSP SSP++SSPSOP
"A before B"
ATSP
A
B
cost(j, i)
cost(i, j)
cost(j, i)
cost(i, j)
Output: sequence of the tasks
Output: path
Inp
ut:
com
ple
x t
asks
CETSP
Output: sequence of the task entry points
TPP
1
2
3
4
5
Traveling Salesman
ProblemAsymmetric TSP Sequential Ordering
Problem
Generalized TSPShortest Sequence
Problem
Shortest Sequence
Problem++
Output: path
MTP
Multi-Goal Path
Planning Problem
Touring-a-sequence-of-Polygons
ProblemClose-Enough TSP
TSPN
TSP with Neighborhoods
SRP
Safari Route Problem
ZRP
Zookeeper Route Problem
MTPGR
MTP for Goal Regions
GTSPN
Generalized TSPN
Fig. 3 Sequencing problems are categorized according to their input and output parameters. The tour/path is denoted withblue and obstacles are denoted with red. The dashed circles denotes the clusters.
categorized by their input and output values. These
problems are illustrated in Fig. 3.
The tasks can be represented as points or finite sets
of points. We refer to such type of tasks as primitive
tasks. The tasks can be also modeled as areas, in that
case we refer to them as complex tasks. For primitive
tasks the output can be a sequence or a path. The path
includes information about the sequence. For complex
tasks, knowledge of the sequence is not enough, as it
is also required to know from which entry point each
task has to be started. Therefore, the output is either
a sequence of entry points or a path.
3.3.1 Input: primitive tasks
Output: sequence of the tasks The most well known
problem of searching for the optimal sequence is the
Traveling Salesman Problem (TSP) [6]. The goal of the
TSP is to find a minimal-cost cyclic tour through a
set of points such that every point is visited once. It
is a NP-hard combinatorial optimization problem and
has been well known in industrial robotics for many
years. The cost between two points in TSP is the same
for both directions. If the distance between two points
depends on the direction, then such problem is called
Asymmetric TSP (ATSP) [42]. Important extension to
ATSP is the constraints on order, e.g., when a point A
has to be visited before B in the tour. This problem is
referred to as Sequential Ordering Problem (SOP) [68].
Shortest Sequence Problem (SSP) is similar to TSP,
but with a difference that the salesman does not have
to return to a starting point [90]. If the points from SSP
are substituted with sets of points, then this problem is
called SSP++. The goal of SSP++ is to find a minimal-
Robotic Task Sequencing Problem: A Survey 7
cost tour that contains a point from every set with no
need to return to the starting point. A generalization
similar to the SSP++ exist for the TSP. This problem is
called Generalized TSP (GTSP) [85]. The goal of GTSP
is to find the minimal-cost cyclic tour that contains a
point from every set. GTSP is also known as TSP++,
set TSP or One-of-a-Set TSP.
Output: path Multi-Goal Path Planning Problem
(MTP), a variant of GTSP, was introduced by Wurll et
al. [90] especially for robotics. Each set here is a set of
the inverse kinematics solutions for a T-space goal. The
objective of the MTP is to calculate a minimal-cost
cyclic collision-free path that visit one point from each
set.
3.3.2 Input: complex tasks
Output: sequence of the task entry points In case the
goals are not points but polygons and the sequence
is given, then this problem is known as a Touring-a-
sequence-of-Polygons Problem (TPP) [27]. The goal of
the TPP is to find the minimal-cost cyclic tour that vis-
its a predefined sequence of regions. If the points should
be visited within a certain radius, then this TSP exten-
sion is known as Close-Enough TSP (CETSP) [66]. In
other words, every point in 2D space is represented as
a disk. One of the most general problems is the TSP
with Neighborhoods (TSPN) [7]. The goal of the TSPN
is to find the minimal-cost cyclic tour through a set of
regions such that every region is visited once. If the
neighborhoods are clustered then this problem is called
Generalized Traveling Salesman Problem with Neigh-
borhoods (GTSPN) [87]. The goal of the GTSPN is to
visit at least one neighborhood from each cluster.
Output: path If the TSPN tour must lay inside a bound-
ary then this problem is referred to the Safari Route
Problem (SRP) [69]. If the tour is not allowed to be
within given convex areas but must only visit their
borders, then this problem is referred to as Zookeeper
Route Problem (ZRP) [20]. In case obstacles are inte-
grated into TSPN and areas are convex polygons, then
this problem is referred to as MTP for Goal Regions
(MTPGR) [49]. In particular, this problem is the most
general and computationally difficult to be solved.
3.3.3 Solving methods
One possible way to solve these problems is to for-
mulate them as the Mixed-Integer NonLinear Program
(MINLP) (e.g., solution for TSPN [38]). However, the
exact methods are efficient only for a small number of
tasks. To be able to solve problems with large num-
ber of tasks, heuristic approaches are applied. In con-
trast to exact methods, heuristics do not guarantee
the optimum but often find a near-optimal solution
in a feasible time. Johnson et al. [51] classified TSP
heuristics into two groups: tour-construction (e.g., In-
sertion Heuristic [45]) and tour-improvement heuristics
(e.g., 2-Opt, 3-Opt [15, 77], Tabu Search [41]). While
tour-construction heuristics calculate a feasible solu-
tion, tour-improvement heuristics start with an initial
tour and reduce the cost.
The problems stated above with no appliance
to robotics are well researched and several efficient
solving-methods exist. To name a few: particle
swarm optimization-based algorithms for TSP and
GTSP [81], Lin–Kernighan heuristic for GTSP [53],
tour-construction and tour improvements heuristics for
CETSP [66], Rubber-band algorithm for TPP [27, 70],
approximation algorithms for TSPN [8,30,67].
The presented TSP-like problems and stated
approaches are the foundation for the optimization
of the robotics task sequence. However, applying
only the stated solving methods is not enough for
task sequencing in robotics, as multiple additional
robot specific features have to be considered: inverse
kinematics, multiple degrees of the task specification
freedom, collision-free planners, etc.
4 Categorization
We split the observed approaches for the robotic task
sequencing problem into the following four groups by
their input and output parameters:
Group 1: Input: primitive tasks, Output: sequence
Group 2: Input: primitive tasks, Output: path
Group 3: Input: complex tasks, Output: sequence
Group 4: Input: complex tasks, Output: path
The approaches from the Groups 1 and 3 might be
useful for environments with none or few obstacles that
have small impact on the robot motion. As Group 1 in-
volves less extra parameters to optimize, its approaches
could be more efficient for a large number of tasks. The
approaches from the Groups 2 and 4 are important for
highly-cluttered environments, however the calculation
of the collision-free paths needs more computational ef-
forts.
The summary of the approaches considered in this
study is provided in Table 1. The details of the par-
ticular realization of the observed approaches can be
found in Section 5. The overall optimization process
can involve large number of factors that influence the
sequence. Further, these factors are described in detail.
8 Sergey Alatartsev et al.
Table 1 Overview of the observed approaches. The “T-point” and “C-point” stand for the point from T-space and C-spacerespectively. “Cart. path” and “C-path” is the Cartesian and C-space path distance respectively.
Inp
ut
Ou
tpu
t
Ap
pro
ach
Collis
ion
free
Mu
ltip
leIK
Base
layou
t
Part
ial
ord
er
Taskspecification
ObjectiveActual
problemSolved
asApplied
algorithms
Sim
ple
task
s
Seq
uen
ce
[28] − − − + T-point cycle time TSP, SOP ATSP, SOP BB[1] − +a + − T-point cycle time GTSP TSP TSA[29] − + − − T-point cycle time TSP TSP NN[2] − +a − − T-point cycle time GTSP TSP TSA[94] − + − − T-point cycle time GTSP TSP GA[78] − − − − T-point Cart. path TSP TSP Random swaps[10] − + + − T-point cycle time GTSP TSP GA[62] − + − − T-point C-path GTSP TSP GA, SA[92] − − − − T-point complexb TSP TSP ENM, GA[55] − + − − T-point complexc GTSP GTSP Constraint solver
Path
[90] + + − − T-point C-path MTP GTSP, SSP++ GA[89] + + − − T-point C-path MTP GTSP, SSP++ GA, A*[73] + + − − T-point cycle time MTP TSP ENM[84] + − − − C-point C-path MTP TSP IH[80] + − − − C-point cycle time MTP TSP PA[79] + + − − T-point cycle time MTP GTSP PA[44] +d + − − T-point cycle time MTP TSP 2-Opt, LK[16] + + + + T-point cycle time MTP severale ACA[91] + + − − T-point cycle time MTP TSP GA[95] + + − − T-point cycle time MTP TSP GA[48] + + + − T-point complexf MTP TSP MOPSO, PSO, NNA[58] +g + − − T-point cycle time MTP TSP LK
Com
ple
xta
sks
Seq
uen
ce [37] − − − − 2D/3D areas Cart. path TSPN TSPN MINLP[3] − − − − 2D areas Cart. path TSPN TSP, TPP IH, 3-Opt[87] − − − − Rn areas Cart. path GTSPN GTSPN HRKGA[56] −h −h − − 3D volumes complexi TSPN TSP FIH, TS, 2-Opt
Path
[49] + − − − Polygons Cart. path MTPGR MTPGR SOM[43] + + − − T-point+2Dj cycle time MTPGR TSP SA[39] + − − − 7D curves cycle time MTPGR TSPN HRKGA[31] +k −h − − 3D volumes complexi MTPGR TSP FIH, TS, 2-Opt
a two IK solutions are observedb cycle time and welding distortionc cycle time and painting quality metricd using the redundancy of the systeme fixed, unfixed(i.e., TSP) or mixed sequence (i.e., SOP)f cycle time and cost of the candidate manipulatorg only for the end-effectorh applied afterwardsi Lexicographical order of three criteria: minimal cycle time / minimal path length (scanner head) / minimal path length
(laser end point)j certain freedom along z-axis and rotation around z-axisk only for sensor-workpiece, for the whole robot it is applied afterwards.
Collision-free planning It is computationally cheap
to calculate the cost of the point-to-point movement
(PTP) for the robot. However, when the environment
is cluttered with obstacles, often the PTP movement
is not possible to perform and, therefore, a collision-
free path has to be calculated. Integration of the
collision-free planner into the sequencing algorithm is
a challenging problem. The column “Collision-free” in
the Table 1 indicates the sequencing approaches that
consider obstacle avoidance.
Multiple inverse kinematics solutions Often the ar-
ticulated robot could reach a T-space point (e.g.,
drilling point) with several C-space points (e.g., several
robot configurations). Although this greatly increases
the search space and makes the sequencing problem
more difficult to solve, it brings a large potential
Robotic Task Sequencing Problem: A Survey 9
for optimization. The freedom of choosing a robot
configuration is used to avoid collisions and reduce the
duration of the movement. The column “Multiple IK”
in the Table 1 depicts the presence of this feature in
the considered approaches.
Robot base layout The location of the robot in the envi-
ronment in relation to the task locations greatly influ-
ences the sequence and vice versa. The main challenge
is that in the worse case the change of the robot loca-
tion requires the recalculation of the whole optimiza-
tion problem. This feature is presented in the column
“Base-layout” in the Table 1.
Task specification Robot tasks vary by their complex-
ity. Often they are modeled in a simple way as T-space
or C-space points. At a first glance, a simple task such
as a hole drilling could be represented with the ap-
propriate robot configuration, i.e., C-space point. How-
ever, this definition is too explicit, as the task could
be reached with multiple configurations. In that case
a T-space point might be a sufficient representation.
Nevertheless, the rotation along the drilling axis is not
important for the manufacturing process, thus it could
be also used for optimization, as it greatly influences the
robot motion behavior. Therefore, even simple robotics
tasks usually bring multiple possibilities for optimiza-
tion. Thus, there is also a group of approaches that con-
sider the task geometry and represent them as 2D-3D
areas, closed-contours, etc.
Partial order Sometimes, the certain task cannot be
performed before another task is accomplished. It is
strongly motivated by the industrial applications, e.g.,
when the robot has to pick up the detail from the con-
veyor and then drill it [28]. The possibility to make the
constraints on the sequence provides industrial robot
programmer with a control on optimization process.
The column “Partial order” in the Table 1 depicts ei-
ther the approach involves partial order feature or not.
Objective The optimization objective in robotic task
sequencing can be minimization of the time, length of
the Cartesian path of the end-effector or the C-space
path as well as specific industrial objectives. The mo-
tion time in the PTP movement is dictated by the slow-
est joint velocity and acceleration. In the approaches,
where the kinematics is not considered, often the Carte-
sian path length has to be minimized. The objective
could be also dictated by the specificity of the man-
ufacturing process, e.g., minimization of the material
distortion caused by welding [92] or minimization of
the laser path length [56].
Problem models When taking several features into ac-
count the researchers might obtain a too complicated
problem that is difficult or impossible to solve within
a feasible amount of time even using the heuristic ap-
proaches. Therefore, often complicated problems are re-
duced to simpler problem models. That brings errors
but makes the optimization feasible. For example, when
solving MTPGR, first TSPN can be solved and then
collision-free paths can be calculated [56]. When solv-
ing TSPN, at first the sequence could be found (i.e.,
TSP) and after that the position of the points (i.e.,
TPP) [66]. Therefore, the actual problem obtained by
combining several features is often different to the se-
quencing problem that is solved at the end. See column
“Actual Problem” and “Solved as” in the Table 1 re-
spectively.
5 Description of the approaches
5.1 Input: primitive tasks, Output: sequence
One of the first research works that discovered the ap-
plication of TSP-like problems in industrial robotics for
PTP movements was done by Dubowsky et al. [28].
ATSP was applied instead of TSP, as in robotics the
cost between two points depends on the direction of
the movement due to gravity and kinematics. The gen-
eral idea of the proposed method is to manipulate the
weights in the cost matrix in order to satisfy the de-
sired constraints on the sequence. Scenarios with no
constraints or with partial constraints on the sequence
are considered. Application of the sequence constraints
was motivated by the industrial scenario, when at first
the robot has to pick up the detail from the conveyor
and after that drill it. The cases when the robot has to
change its tool during the sequence execution were inte-
grated into the planning. The branch-and-bound tech-
nique [61] was applied to obtain the exact ATSP solu-
tion. Nevertheless, the method proposed by Dubowsky
et al. is independent from the particular ATSP solv-
ing algorithm. The main limitation is that no multiple
inverse kinematics solutions were taken into account.
Therefore, an obtained sequence is far from being opti-
mal in real-life applications.
Abdel-Malek et al. [1] improved a manufacturing
cell, in which a 3-DOF robot has to solve a TSP problem
while considering the multiple solutions of the inverse
kinematics. TSP was solved with Traveling Salesman
Algorithm (TSA) [46]. Optimization of the robot base
location was performed to minimize the cycle time. For
that a grid search scheme was applied. It limits the
problem to the X and Y coordinates, which means that
the robot location can not move vertically in the cell.
10 Sergey Alatartsev et al.
The robot is placed on every point of the grid and the
cycle time is calculated. The position with the shortest
cycle time is considered to be the optimal location for
that grid. However, there is a trade-off between accu-
racy and computation time when choosing the granu-
larity of the grid. Later Abdel-Malek et al. [2] extended
the approach with support of several types of robots:
cartesian, cylindrical, spherical, articulated.
Edan et al. [29] proposed the approach for robot
task sequencing. It was motivated by the fruit picking
scenario. The main idea of the approach is straight-
forward: calculate the cost matrix and then apply NN
algorithm to solve TSP. The goal of the approach is to
minimize the cycle time while considering both robot
kinematics and dynamics. It was suggested than the
approach can be also applied for the task-based robot
choice. As far as the algorithm requires the full cost
matrix, it makes the integration of collision-free plan-
ning unfeasible for real life scenarios, due to the need
of extremely large computational time.
Zacharia et al. [94] presented a method based on the
Genetic Algorithm (GA) and involved multiple inverse
kinematic solutions in the optimization process. The
method is capable of searching the sequence for a 6-
DOF robot in a feasible time. In order to take multiple
inverse kinematics solutions into account, an innovative
encoding for the GA was introduced. The GA chromo-
some consists of two parts. The first part encodes the
sequence and the second part encodes the robot config-
urations for the correspondent sequence.
Later Baizid et al. [10] extended the approach by
Zacharia et al. [94] with a robot base layout optimiza-
tion. It was done by adding a third component into the
chromosome that denotes the position of the base.
Although applied heuristic approaches in [94] [10]
allows to efficient search space exploration it does no
guarantee optimal solution to be found. Such iterative
search heuristics can be stopped by the exceeding the
maximum number of iterations, therefore, it does not
reflect how far is the obtained solution from optimum.
Reinhart et al. [78] proposed an algorithm for
solving the sequencing problem for open-end curves
in remote-laser-welding applications. At first, every
curve is represented with one point and a random tour
is constructed. Then random swaps are performed to
improve the tour. Later the direction on the curves are
randomly chosen. The algorithm randomly changes the
direction to obtain a better tour. Although, a certain
improvement of the solution can be obtained, random
swaps efficiency should drop down with computation
time.
Loredo-Flores et al. [62] developed a GUI that al-
lows users to manually create the robot’s path. Later,
the constructed sequence is optimized by GA and Sim-
ulated Annealing (SA) approaches. The output is the
sequence of robot configurations with the objective to
minimize the robot joint displacement, i.e., C-space dis-
tance.
Yang et al. [92] considered the task sequencing prob-
lem for welding applications with the objective to mini-
mize the time and the distortion of the welding process.
The general idea is to split the tasks into two categories
by their influence on the product quality, i.e., welding
deformation. The sequence of tasks from the category
with minimal influence is optimized by an Elastic Net
Method (ENM) [73] to minimize the cycle time. The
category with high influence on deformation is opti-
mized with a GA to minimize the product distortion.
The main idea of ENM is to apply two forces on the
net: one is trying to keep the points closer to each other
and another force attracts them to the goal positions.
When two optimized sequences are obtained, they are
merged with the nearest neighbor criterion. The pro-
posed approach was evaluated in a 2D environment, as
the calculation of the distortion in 3D is much more
complicated problem. However, the general ideas are
domain-independent and, therefore, can be reused in
other domains.
Kolakowska et al. [55] motivated robot task
sequencing with painting scenario. The proposed
approach is strictly aim at solving task scheduling
problem and completely independent from the task
and motion planning stages. The main contribution
of the paper is the mathematical constraint model,
that can be further used in any constraint solver.
Proposed approach is efficient for scenarios with up
to 10 goals. The downside of solver application is the
large computational time for larger case studies. Due
to the large search space, the solver was not able to
finish calculation of the solution in some scenarios
within 40 hours.
5.2 Input: primitive tasks, Output: path
Wurll et al. [90] introduced Multi-Goal Path Planning
(MTP). The cost matrix for MTP is almost impossible
to pre-calculate because distances should be obtained
by taking obstacles into account and collision-free plan-
ners are computationally expensive algorithms. There-
fore, the costs are unknown and have to be obtained
on the solving stage. The solution process was split
into three stages: controlling, path planning and short-
est sequence planning. Controlling is the stage where
the decision is made, which pair of start and goal con-
figurations should be sent to the path planning stage.
Robotic Task Sequencing Problem: A Survey 11
They proposed four different strategies for the control-
ling stage. The collision-free path planning between two
configurations is calculated in a C-space grid by using
the best first search method. In the third stage, the
approach uses a GA to solve the TSP. The proposed
approach loops in these three stages and guarantees
that every newly obtained tour will be not worse than
the previous one. Later Wurll et al. [89] extended the
approach by parallelization of a bidirectional search to
reduce computation time and used the A* search for
collision-avoidance method.
Petiot et al. [73] solved the task scheduling problem
for industrial robots with an ENM. The collision avoid-
ance feature was integrated as another repulsive force
into the ENM, i.e., a penalty term in the ENM function.
The limitation of this approach is that obstacles must
be convex and easily representable in the C-space. The
algorithm performs well on robots with small number of
DOFs (e.g., with 2 or 3). But it is difficult to generalize
the approach for robots with high number of DOFs, as
it becomes computationally expensive to map obstacles
to the robot configuration space.
Spitz et al. [84] proposed a general framework suit-
able for a variety of robots. Domain specific knowledge
could be integrated into the planner, as a local plan-
ner or an extension of the roadmap created with PRM.
Based on this roadmap, another graph is created, which
consists of only target points and all superfluous points
are deleted. It is then used for tour optimization with
known TSP solvers. The method assumes that task is
reachable with one robot configuration, that limits po-
tential of optimization due to committing the kinemat-
ics redundancy.
Saha et al. [80] observed the problem of multi-goal
planning among the robot configurations. The use of
C-space points instead of T-space points limited the
potential freedom for optimization. Therefore, the ap-
proach was improved by specifying the input points in
T-space [79]. Lower-bound approximation is used for
computing the optimal path (normally the shortest dis-
tance) and the call of the collision-free path planner is
delayed. A bidirectional tree-expansion PRM was used
for collision-free planning. For TSP solving the Prim’s
algorithm was applied [21]. The main assumption of this
approach is that calculation of a sequence is computa-
tionally less expensive than calculation of the collision-
free movements. Therefore, collision-free paths are cal-
culated after the sequence is obtained. If the collision-
free path cost is significantly higher than the estimated
cost, the sequence should be recalculated taking the
newly obtained path cost into account.
Gueta et al. [44] proposed a method for solving
the TSP problem for goals located on a revolvable ta-
ble. The table and the robot are seen as one redun-
dant system. The goals are put into clusters and a se-
quence of these clusters is calculated with the 2-Opt
algorithm. Then for each cluster two goals are deter-
mined as connectors for the previous and following clus-
ters in the sequence. After that the TSP is solved with
Lin–Kernighan (LK) heuristic in each cluster. The dis-
tance between the goals is considered as a straight line
in C-space. The system redundancy is used to select
such configurations of the robot that allow to obtain a
collision-free path.
Bu et al. [16] presented a method for the optimiza-
tion of the robot base layout and robotic task sequence.
The objective is the minimization of the cycle time. The
input sequence could be either fixed (linear topology),
unfixed (complete directed graph topology) or fixed and
unfixed sequence (mixed topology). The first step of
the solution process is to calculate the 5D space of the
possible base locations, where 3D stands for position,
1D stands for orientation around the base axis and 1D
shows if the base axis is upward or downward. Then
this 5D space is divided into discrete grid cells. Opti-
mization of the sequence is done with an Ant Colony
Algorithm (ACA) for each cell. The next step is to find
the local optimum for the robot base position and ori-
entation with the pattern search for each cell in rela-
tion to the costs of the obtained sequences. Finally, the
global optimum is the result of comparing all local op-
tima. The approach does not provide automatic smart
technique for collision avoidance. Obstacle-avoidance is
organized in a way that when the collision occurs be-
tween the robot and a workpiece, an intermediate frame
should be added between the two target frames.
Xidias et al. [91] investigated the problem of de-
termining the optimal sequence of task points for an
articulated robot with the obstacle avoidance feature
in a 2D environment. The problem is divided into the
optimal sequencing problem – TSP and the path plan-
ning problem. The objective is to obtain the shortest
cycle time. The proposed concept is an adaptation of
the GA, proposed by Zacharia et al. [94]. Furthermore,
they introduce the Bump Surface concept for collision
checking. The general idea of the Bump-surface is to
represent the 2D working environment in a unified way
by mapping it into a 3D space. It is done by discretizing
the 2D map into a uniform grid of points and assigning
a value from [0, 1] to each point. If the point is out-
side the obstacle then the value is 0 and if it is inside
then the value is from the range (0, 1]. The search space
is represented as one mathematical entity, a B-spline.
Therefore, the complexity of the problem does not de-
pend on the complexity of the environment (shape and
12 Sergey Alatartsev et al.
location of obstacles). The limitation of this approach
is that the working environment is considered to be 2D.
Further, Zacharia et al. [95] extended two ap-
proaches: [94] by adding obstacle avoidance and [91]
by adapting it to the real-world 3D environment.
The optimization problem was solved with the GA.
The chromosome was extended with the information
about the intermediate configurations required for
collision-free movement. The search space is repre-
sented with the Bump-surface as a single mathematical
entity. The innovative characteristic of the proposed
approach is that sequencing, motion planning and
obstacle avoidance are incorporated into the objective
function. It leads to the situation, when all paths
(both collision and collision-free) are considered to be
feasible but collision-free paths are preferred by the
search algorithm.
Huang et al. [48] observed the task of selecting a
specific manipulator for the particular multi-goal task
planning in a system that consists of a robot arm and
a position table. The input parameters for the algo-
rithm are the list of points and the candidate manip-
ulator specification (Denavit–Hartenberg parameters).
As it is not possible to explore the whole search space
within feasible time, the authors decomposed the prob-
lem into the three nested stages: 1) manipulator se-
lection (solved by Multiple-Objective Particle Swarm
Optimization (MOPSO)), 2) layout design (solved by
Particle Swarm Optimization (PSO)) and 3) motion
planning (solved by Nearest Neighborhood Algorithm
(NNA)). Less computationally expensive algorithms are
chosen for the inner loops (i.e., stage 2 and 3) due to the
fact that stages are nested in each other and, therefore,
the inner loops have to be repeated more often than
outer loops. The output is the manipulator structural
configuration as well as positions of the manipulator
and rotational table.
Lattanzi et al. [58] proposed the extension to the
LK heuristic by integrating collision-free planning into
the cost function. As it is computationally expensive
to obtain a collision-free path for the whole robot, the
collision-free path was calculated only for the robot’s
end-effector. The robot end-effector path between two
Cartesian points is the shortest (in terms of robot mo-
tion time) when the robot performs PTP movement. It
was necessary to make sure that this PTP movement
does not cause collision of the end-effector. PTP joint
path is transferred into the end-effector path with FK
and represented as a poly-line. In case of a collision, an
intermediate path-point is inserted and PTP movement
is recalculated. As LK is a tour improvement heuristic,
it was evaluated with different tour construction heuris-
tics: the shortest/farthest insertion and the greedy ap-
proach. The work was motivated by a visual inspection
task in an industrial setup.
5.3 Input: complex tasks, Output: sequence
Within the last five years, the use of task shape poten-
tial for optimization attracted a lot of attention. Fur-
ther these approaches are discussed in detail.
Gentilini et al. [37] used TSPN for modeling the se-
quencing problem of camera inspection tasks. The robot
has to take a picture with a camera mounted on its
end-effector. The problem does not require a precise
position but rather an area from which the pictures
have to be taken. TSPN problem was formulated as
Mixed-Integer Non-Linear Program (MINLP). It was
shown that searching for an exact solution requires un-
reasonable amount of time. Therefore, a heuristic was
introduced to a MINLP solver to speed up the calcula-
tion time. The method was evaluated on test instances
with up to 16 areas. Although the obtained solutions
are very close to the optimum (the average error is less
than 0.6% on tests with up to 16 areas), the approach
omits collision-free path planning. Furthermore, the op-
timization process is based on the Euclidean distance
metrics, thus the robot kinematics flexibility is not con-
sidered.
Alatartsev et al. [3] used TSPN for modeling the
problem of sequencing closed-contour tasks. The work
was illustrated with the industrial use case when the
robot has to cut out a set of holes from a toboggan.
A tour construction method – Constricting Insertion
Heuristic was proposed to solve TSPN. The main idea
was to slit TSPN into TSP and TPP and solve them si-multaneously. The evaluation showed that the solution
could be obtained within a short period of time for a
large number of contours (e.g., 30s. for 150 contours).
However, the approach did not make use of the robot
kinematics, and applied only Cartesian distance as a
metric. Later, the quality of the solution was improved
by developing a tour improvement heuristic based on
3-Opt [5].
Sequencing of tasks with extra freedom was investi-
gated by Kovacs [56]. This problem was motivated by a
specific industrial application – robot remote laser weld-
ing. An efficient combination of the TSP solver (Far-
thest Insertion Heuristic, Tabu search and 2-Opt) and
the path planning method was proposed. Collision-free
path planning is not considered during the sequencing,
but rather done after the sequence is calculated. The
planning is done in Cartesian space and after that the
solution is converted into the robot configuration space
by using inverse kinematics transformation. The min-
imization objective is the lexicographical order of the
Robotic Task Sequencing Problem: A Survey 13
three criteria: minimal cycle time / minimal path length
(scanner head) / minimal path length (laser end point).
Vicencio et al [87] proposed GTSPN problem. It is
the extension of the well-known TSPN problem, where
the neighborhoods can be clustered and each cluster
should be visited one. The A Hybrid Random-Key Ge-
netic Algorithm (HRKGA) was proposed to solve the
GTSPN. It was shown that HRKGA is capable of effec-
tive and efficient solving of scenarios up to 300 neigh-
borhoods. The algorithm was tested on 3D and 7D in-
stances. The author stressed the importance of the dy-
namics integration into sequencing. Due to the problem
complexity, the collision-free planning was omitted.
5.4 Input: complex tasks, Output: path
Faigl et al. [49] presented the multi-goal path planning
problem for goal regions (MTPGR) in the polygonal
domain. The areas could be arbitrary shapes and are
allowed to overlap. The proposed approach is based on
self-organizing map (SOM) algorithm and was evalu-
ated in a 2D environment. The presented method ap-
pears to be very efficient and is capable of solving test
cases with up to 106 areas in less than 6 seconds.
Later Gueta et al. [43] extended the previously
described approach [44] by representing the goals
as T-space points with two parameters that provide
additional freedom: 1) distance of the camera from
the goal point along the approach vector and 2)
rotation about the approach vector. In addition, the
tool attachment was optimized using a Simulated
Annealing (SA) method.
Gentilini et al. [39] used the TSPN to model the task
sequencing problem of the camera inspection tasks for
the case when the inspected object is located on the
rotated table. The neighborhoods (sets of robot con-
figurations for each goal placement) are represented as
curves in the seven-dimensional configuration space (6D
is the robot configuration space and 1D is added by the
table rotation). The paper concentrates on exploring
the redundancy of the system, rather than kinematics,
therefore, only one configuration was chosen among the
set of possible inverse kinematics solutions. Due to the
search space explosion, the exact methods are limited to
a small amount of goals. Therefore, the heuristic Hybrid
Random-key Genetic Algorithm was applied. A single
query BiRRT was used to construct the roadmap that
is further used for collision-free planning.
Erdos et al. [31] proposed the improvement of the
previous work of Kovacs [56]. The objective was the
same – minimization of the cycle time of the remote
laser welding operations. During the planning stage
only the scanner-workpiece was checked for collision,
but not the whole robot. Due to the search space fast
growth, collision-free path planners for the whole robot
have to be applied only after the sequence is obtained.
6 Conclusion
6.1 Summary
This paper surveys the approaches for the task sequenc-
ing problem in robotics. The sequencing of tasks with
primitive geometry attracted a lot of research atten-
tion, in contrast to the sequencing of tasks with com-
plex geometry. However, complex task are more com-
mon in industry. There is still no approach that is able
to efficiently optimize task sequence and consider all
degrees of freedom and constraints, e.g., collision-free
planning, multiple-IK, robot base layout optimization,
making use of the task specification fleixibily (see Sec-
tion 4). The fast growth of the search space makes the
optimization a very challenging problem. Thus, appli-
cation of the exact methods is problematic. Therefore,
the use of heuristics is preferred.
Mainly, two general concepts were applied: de-
composition and metaheuristics. In the first case,
researchers decomposed the complicated problems into
simple ones that can be solved efficiently: consecutive
solving of subproblems (e.g., [56, 79, 89]), solving of
nested subproblems (e.g., [3, 48]) or parallel solving of
subproblems and merging afterwards (e.g., [92].) Meta-
heuristics are problem-independent algorithms that are
capable of efficient search space exploration and finding
near-optimal solution. Metaheuristics are proved to
be efficient for tour-searching problems and as a
consequence suit well for task sequencing in robotics.
Researchers preferred to use GA (e.g., [39, 94]), ENM
(e.g., [73, 92]) and SA (e.g., [43, 62]). General trends
to obtain collision-free paths are: grid decomposi-
tion [90,91], sample based planning (e.g., PRM or/and
RRT) [39,79,84], redundancy of the system [44].
MTPGR (see Fig.3) is the most general problem
that is able to describe the majority of tasks. However,
it is too complicated to be solved for articulated robots.
In some industrial scenarios its use can be excessive. For
example, for many welding scenarios the probability to
collide with obstacles is small because tasks are located
roughly in one 2D plane. Therefore, there is no need
to make excessive collision-free planning, as the move-
ments between seams can be performed as “up - move
horizontally - down”. When there is no need to return
to the current position SSP or SSP++ might provide
good results. These problems in particular suit better
to service robotics, where jobs are rapidly changing and
14 Sergey Alatartsev et al.
after performing one task sequence the robot has to im-
mediately start with the next one without returning to
the starting configuration.
6.2 Future Challenges
Further research is clearly needed in the task sequenc-
ing domain. In this section we highlight the possible
directions.
High DOF robots It was already demonstrated that re-
dundancy, i.e., high DOF systems, could be used for
collision avoidance [43] [44] and in general leads to more
effective solutions. One future direction is to observe the
case when robots have a large or even infinite number
of inverse kinematics solutions. It greatly increases the
complexity of the problem, but brings a huge potential
for optimization.
Moving goals and obstacles Mainly, stated approaches
observe static goals and obstacles. In presence of a con-
veyor or other moving manufacturing devices the task
sequencing becomes more challenging. Wurll et al. [89]
stated importance of the moving goals scenario and sug-
gested it as a future research direction. This requires to
include the time dimension into the solution process.
Additional research is needed in this direction as it is
not clear how the moving laws will influence the com-
plexity of the problem.
Kinodynamic task sequencing Dubowsky et al. [28] in
1989 proposed ATSP to be the most suitable modeling
problem for robotic task sequencing, because the costs
of the movements “A to B” and “B to A” are differ-
ent due to the robot dynamics and gravity. However,
after that, the researchers considered robot kinemat-
ics but ignored robot dynamics. Till now TSP is the
most popular combinatorial problem for modeling the
robotic task sequencing problem. The exception is Edan
et al. [29] who included robot dynamics into the se-
quencing. The importance of the dynamics integration
was stated in [16] and [87]. Although, integration of the
robot dynamics increases the search space and compu-
tational time, it allows to utilize robots in production
much more efficiently. Involving the dynamics into the
task sequencing problem and, therefore, making use of
ATSP is still an open research question.
Constraints on sequence order The majority of the ob-
served approaches assume that tasks are independent
from each other within a sequence. Nevertheless, many
industrial robotic applications impose constraints on
order. There are cases, when a certain task has to be
performed before or after another one in the sequence.
Among the observed approaches, only two methods
(i.e., [28] [16]) considered the possibility to set sequence
constraints. This issue was also referred in [79] as a
future problem to be solved. The research in this
direction is needed as it will allow to broaden the
application domains to the scenarios that require the
constraints on order. It will also allow industrial robot
programmers to partially control the optimization
process.
Evaluation and benchmarks There are test instances
for benchmarking TSP-like problems, e.g., TSP4,
TSPN5,6, CETSP7. There is some research on evalua-
tion of collision-free planners [32]. However, there are
still no benchmarks for the task sequence optimization
problem, though it is known for many years. It is
difficult to compare existing approaches, as researchers
use different industrial case studies, robots, included
features and constraints. Nowadays, such platforms
as ROS [75] allow high re-usability of the code and,
therefore, might be used as a base for creating bench-
marks. Nevertheless, it is difficult to create a general
benchmark, therefore, domain dependent test instances
can be the first step, e.g., instances for spot-welding,
palletizing or deburring domains. Developing of the
benchmark for the task sequencing in robotics is still
an open research question.
Considering more freedom in task description Within
the last 5 years, sequencing of tasks with complicated
shapes attracted a lot of research attention. However,
there is still a lot of potential for improvement. The
observed approaches provide the freedom for a particu-
lar industrial tasks, e.g., closed-contours cutting [3] or
laser-welding [56]. But still, there is no general frame-
work that allows a general task description and opti-
mization techniques that are independent from the par-
ticular industrial task. At the same time, approaches
for collision-free “goal-to-goal” planning allow specifi-
cation of goals with a certain freedom, e.g., [12], [86].
This experience might be reused for creating a general
framework for the robot task sequencing problem.
“Super” TSP problem It is clear that TSP problem is
not enough for effective modeling real-life applications.
Therefore, researchers formalized TSP-like generaliza-
tions. We outlined TSP-like problems for robotic task
4 http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/5 https://cse.cs.ovgu.de/cse/robotics/tspn/6 http://wpweb2.tepper.cmu.edu/fmargot/ampl.html7 http://drum.lib.umd.edu/handle/1903/9822
Robotic Task Sequencing Problem: A Survey 15
sequencing in 3. The most complex and general problem
is MTPGR. However, even such a wide range problem is
not enough to describe real-world problems in robotics.
This problem clearly misses three features introduced in
other TSP-like problems: 1) sequence constraints from
SOP, 2) asymmetrical costs from ATSP and 3) areas
clustering from GTSPN. The generalization that would
incorporate the MTPGR+SOP+ATSP+GTSPN in one
“Super” TSP problem is a very challenging issue. How-
ever, it would allow to model the robot task sequencing
problem much more precise.
Acknowledgements The final publication is available atSpringer via http://dx.doi.org/10.1007/s10846-015-0190-6.Expected to be published in: Journal of Intelligent & RoboticSystems. DOI: 10.1007/s10846-015-0190-6
References
1. Abdel-Malek, L.L., Li, Z.: Robot location for minimumcycle time. Engineering Costs and Production Economics17(1), 29–34 (1989)
2. Abdel-Malek, L.L., Li, Z.: The application of inverse kine-matics in the optimum sequencing of robot tasks. The In-ternational Journal Of Production Research 28(1), 75–90(1990)
3. Alatartsev, S., Augustine, M., Ortmeier, F.: Constrictinginsertion heuristic for traveling salesman problem withneighborhoods. In: 23rd International Conference on Au-tomated Planning and Scheduling (ICAPS) (2013)
4. Alatartsev, S., Belov, A., Nykolaychuk, M., Ortmeier,F.: Robot trajectory optimization for the relaxed end-effector path. In: 12th International Conference on Infor-matics in Control, Automation and Robotics (ICINCO)(2014)
5. Alatartsev, S., Mersheeva, V., Augustine, M., Ort-meier, F.: On optimizing a sequence of robotic tasks.In: IEEE/RSJ International Conference on IntelligentRobots and Systems (IROS) (2013)
6. Applegate, D.L., Bixby, R.E., Chvatal, V., Cook, W.J.:The traveling salesman problem: a computational study.Princeton University Press (2007)
7. Arkin, E.M., Hassin, R.: Approximation algorithms forthe geometric covering salesman problem. Discrete Ap-plied Mathematics 55, 197–218 (1995)
8. Arora, S.: Approximation schemes for NP-hard geomet-ric optimization problems: A survey. Mathematical Pro-gramming 97, 43–69 (2003)
9. Ata, A.A.: Optimal trajectory planning of manipulators:A review. Journal of Engineering Science and Technology1, 32 (2007)
10. Baizid, K., Chellali, R., Yousnadj, A., Meddahi, A., Ben-taleb, T.: Genetic algorithms based method for time op-timization in robotized site. In: IEEE/RSJ InternationalConference on Intelligent Robots and Systems (IROS).pp. 1359–1364 (2010)
11. Berenson, D.: Manipulation of deformable objects with-out modeling and simulating deformation. In: IEEE/RSJInternational Conference on Intelligent Robots and Sys-tems (IROS). pp. 4525–4532. IEEE (2013)
12. Berenson, D., Srinivasa, S., Kuffner, J.: Task space re-gions: A framework for pose-constrained manipulation
planning. International Journal of Robotics Research(IJRR) 30(12), 1435–1460 (October 2011)
13. Bhatia, A., Maly, M.R., Kavraki, E., Vardi, M.Y.: Motionplanning with complex goals. Robotics & AutomationMagazine, IEEE 18(3), 55–64 (2011)
14. Biggs, G., MacDonald, B.: A survey of robot program-ming systems. In: Proceedings of the Australasian Con-ference on Robotics and Automation (2003)
15. Bock, F.: An algorithm for solving traveling-salesmanand related network optimization problems. In: Unpub-lished manuscript associated with talk presented at the14th ORSA National Meeting (1958)
16. Bu, W., Liu, Z., Tan, J.: Industrial robot layout based onoperation sequence optimisation. International Journal ofProduction Research 41, 4125–4145 (2009)
17. Cao, T., Sanderson, A.C.: Task sequence planning in arobot workcell using and/or nets. In: Intelligent Control,1991., Proceedings of the 1991 IEEE International Sym-posium on. pp. 239–244. IEEE (1991)
18. Chien, S.Y., Xue, L.Q., Palakal, M.: Task planning fora mobile robot in an indoor environment using object-oriented domain information. IEEE Transactions on Sys-tems, Man, and Cybernetics, Part B: Cybernetics 27(6),1007–1016 (1997)
19. Chien, Y., Hudli, A., Palakal, M.: Using many-sortedlogic in the object-oriented data model for fast robottask planning. Journal of Intelligent and Robotic Sys-tems 23(1), 1–25 (1998)
20. Chin, W.P., Ntafos, S.: The zookeeper route problem.Information Sciences: an International Journal 63, 245–259 (1992)
21. Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.,et al.: Introduction to algorithms, vol. 2. MIT press Cam-bridge (2001)
22. Craig, J.: Introduction to Robotics: Mechanics and Con-trol. Pearson (2005)
23. Sucan, I.A., Moll, M., Kavraki, L.E.: The Open MotionPlanning Library. IEEE Robotics & Automation Maga-zine (2012), http://ompl.kavrakilab.org
24. Dawande, M.W., Geismar, H.N., Sethi, S.P., Sriskandara-jah, C.: Throughput Optimization in Robotic Cells, In-ternational Series in Operations Research & ManagementScience, vol. 101. Springer US (2007)
25. Diankov, R., Srinivasa, S.S., Ferguson, D., Kuffner, J.:Manipulation planning with caging grasps. In: HumanoidRobots, 2008. Humanoids 2008. 8th IEEE-RAS Interna-tional Conference on. pp. 285–292. IEEE (2008)
26. Doria, N.S.F., Freire, E.O., Basilio, J.C.: An algorithminspired by the deterministic annealing approach to avoidlocal minima in artificial potential fields. In: AdvancedRobotics (ICAR), 2013 16th International Conference on.pp. 1–6. IEEE (2013)
27. Dror, M., Efrat, A., Lubiw, A., Mitchell, J.S.B.: Touringa sequence of polygons. In: 35th annual ACM symposiumon Theory of Computing. pp. 473–482. ACM Press (2003)
28. Dubowsky, S., Blubaugh, T.: Planning time-optimalrobotic manipulator motions and work places for point-to-point tasks. IEEE Transactions on Robotics and Au-tomation 5, 377–381 (1989)
29. Edan, Y., Flash, T., Peiper, U.M., Shmulevich, I., Sarig,Y.: Near-minimum-time task planning for fruit-pickingrobots. IEEE Transactions on Robotics and Automation7(1), 48–56 (1991)
30. Elbassioni, K.M., Fishkin, A.V., Sitters, R.: Approxima-tion algorithms for the euclidean traveling salesman prob-lem with discrete and continuous neighborhoods. Inter-national Journal of Computational Geometry and Appli-cations pp. 173–193 (2009)
16 Sergey Alatartsev et al.
31. Erdos, G., Kemeny, Z., Kovacs, A., Vancza, J.: Planningof remote laser welding processes. In: Forty Sixth CIRPConference on Manufacturing Systems (2013)
32. Fragkopoulos, C., Abbas, K., Eldeep, A., Graeser, A.:Comparison of sampling based motion planning algo-rithms specialized for robot manipulators. In: 7th Ger-man Conference on Robotics (ROBOTIK) (2012)
33. Galceran, E., Carreras, M.: A survey on coverage pathplanning for robotics. Robotics and Autonomous Systems61(12), 1258–1276 (2013)
34. Galindo, C., Fernandez-Madrigal, J.A., Gonzalez, J.: Im-proving efficiency in mobile robot task planning throughworld abstraction. Robotics, IEEE Transactions on 20(4),677–690 (2004)
35. Garrido, S., Moreno, L., Abderrahim, M., Martin, F.:Path planning for mobile robot navigation using voronoidiagram and fast marching. In: 2006 IEEE/RSJ Inter-national Conference on Intelligent Robots and Systems,.pp. 2376–2381 (2006)
36. Gaschler, A., Petrick, R., Giuliani, M., Rickert, M., Knoll,A.: Kvp: A knowledge of volumes approach to robot taskplanning. In: IEEE/RSJ International Conference on In-telligent Robots and Systems (IROS). pp. 202–208. IEEE(2013)
37. Gentilini, I., Margot, F., Shimada, K.: The travellingsalesman problem with neighbourhoods: MINLP so-lution. Optimization Methods and Software pp. 1–15(2011)
38. Gentilini, I., Margot, F., Shimada, K.: The travellingsalesman problem with neighbourhoods: MINLP solu-tion. Optimization Methods and Software 28(2), 364–378(2013)
39. Gentilini, I., Nagamatsu, K., Shimada, K.: Cycle timebased multi-goal path optimization for redundant roboticsystems. In: IEEE/RSJ International Conference on In-telligent Robots and Systems (IROS) (2013)
40. Gipson, I., Gupta, K., Greenspan, M.: MPK: An OpenExtensible Motion Planning Kernel. Journal of RoboticSystems 18(8), 433–443 (2001)
41. Glover, F.: Tabu search-part i. ORSA Journal on com-puting 1(3), 190–206 (1989)
42. Glover, F., Gutin, G., Yeo, A., Zverovich, A.: Construc-tion heuristics for the asymmetric TSP. European Jour-nal of Operational Research 129, 555–568 (2001)
43. Gueta, L.B., Cheng, J., Chiba, R., Arai, T., Ueyama, T.,Ota, J.: Multiple-goal task realization utilizing redundantdegrees of freedom of task and tool attachment optimiza-tion. In: Robotics and Automation (ICRA), 2011 IEEEInternational Conference on. pp. 1714–1719 (2011)
44. Gueta, L.B., Chiba, R., Ota, J., Ueyama, T., Arai, T.:Coordinated motion control of a robot arm and a po-sitioning table with arrangement of multiple goals. In:IEEE International Conference on Robotics and Automa-tion, ICRA. pp. 2252–2258 (2008)
45. Hassin, R., Keinan, A.: Greedy heuristics with regret,with application to the cheapest insertion algorithm forthe TSP. Operations Research Letters 36(2), 243–246(Mar 2008)
46. Held, M., Karp, R.M.: The traveling-salesman problemand minimum spanning trees. Operations Research 18(6),1138–1162 (1970)
47. Hoy, M., Matveev, A.S., Savkin, A.V.: Algorithms forcollision-free navigation of mobile robots in complex clut-tered environments: a survey. Robotica pp. 1–35 (2013)
48. Huang, Y., Gueta, L.B., Chiba, R., Arai, T., Ueyama,T., Ota, J.: Selection of manipulator system for multiple-goal task by evaluating task completion time and cost
with computational time constraints. Advanced Robotics27(4), 233–245 (2013)
49. Jan Faigl, Vojtech Vonsek, L.P.: A multi-goal path plan-ning for goal regions in the polygonal domain. In: Euro-pean Conference on Mobile Robots (ECMR) (2011)
50. Jimenez, P.: Survey on model-based manipulation plan-ning of deformable objects. Robotics and computer-integrated manufacturing 28(2), 154–163 (2012)
51. Johnson, D.S., McGeoch, L.A.: Local search in combina-torial optimization. John Wiley and Sons, London (1997)
52. Karaman, S., Frazzoli, E.: Sampling-based algorithms foroptimal motion planning. The International Journal ofRobotics Research 30(7), 846–894 (2011)
53. Karapetyan, D., Gutin, G.: Lin–Kernighan heuristicadaptations for the generalized traveling salesman prob-lem. European Journal of Operational Research 208(3),221–232 (2011)
54. Kavraki, L., Svestka, P., Latombe, J., Overmars, M.:Probabilistic roadmaps for path planning in high-dimensional configuration spaces. In: IEEE InternationalConference On Robotics and Automation. pp. 566–580(1996)
55. Kolakowska, E., Smith, S.F., Kristiansen, M.: Constraintoptimization model of a scheduling problem for a roboticarm in automatic systems. Robotics and AutonomousSystems 62(2), 267–280 (2014)
56. Kovacs, A.: Task sequencing for remote laserwelding inthe automotive industry. In: 23rd International Confer-ence on Automated Planning and Scheduling (ICAPS)(2013)
57. Kuffner, J.J., Lavalle, S.M.: RRT-Connect: An efficientapproach to single-query path planning. In: IEEE Inter-national Conference On Robotics and Automation. pp.995–1001 (2000)
58. Lattanzi, L., Cristalli, C.: An efficient motion planningalgorithm for robot multi-goal tasks. In: IEEE Interna-tional Symposium on Industrial Electronics (ISIE) (2013)
59. LaValle, S., Cheng, P., Kuffner, J., Lindemann, S.,Manohar, A., Tovar, B., Yang, L., Yershova, A.: MSL -Motion strategy library, http://msl.cs.uiuc.edu/msl/
60. LaValle, S.: Planning algorithms. Cambridge UniversityPress (2006)
61. Little, J.D., Murty, K.G., Sweeney, D.W., Karel, C.: Analgorithm for the traveling salesman problem. Operationsresearch 11(6), 972–989 (1963)
62. Loredo-Flores, A., Gonzalez-Galvan, E.J., Cervantes-Sanchez, J.J., Martinez-Soto, A.: Optimization of in-dustrial, vision-based, intuitively generated robot point-allocating tasks using genetic algorithms. Systems, Man,and Cybernetics, Part C: Applications and Reviews,IEEE Transactions on 38(4), 600–608 (2008)
63. Maimon, O.: The robot task-sequencing planning prob-lem. Robotics and Automation, IEEE Transactions on6(6), 760–765 (1990)
64. Mansard, N., Chaumette, F.: Task sequencing for high-level sensor-based control. Robotics, IEEE Transactionson 23(1), 60–72 (2007)
65. Masoud, A.A.: A harmonic potential approach for simul-taneous planning and control of a generic uav platform.Journal of Intelligent & Robotic Systems 65(1-4), 153–173 (2012)
66. Mennell, W.: Heuristics for solving three routing prob-lems: close-enough traveling salesman problem, close-enough vehicle routing problem, sequence-dependentteam orienteering problem. Ph.D. thesis, University ofMaryland (2009)
Robotic Task Sequencing Problem: A Survey 17
67. Mitchell, J.S.: A constant-factor approximation algo-rithm for TSP with pairwise-disjoint connected neigh-borhoods in the plane. In: Proceedings of the 2010 an-nual symposium on Computational geometry. pp. 183–191. SoCG ’10 (2010)
68. Montemanni, R., Smith, D.H., Gambardella, L.M.: Aheuristic manipulation technique for the sequential order-ing problem. Computers & Operations Research 35(12),3931 – 3944 (2008)
69. Ntafos, S.: Watchman routes under limited visibility.Computational Geometry 1, 149–170 (1992)
70. Pan, X., Li, F., Klette, R.: Approximate shortest pathalgorithms for sequences of pairwise disjoint simple poly-gons. In: Canadian Conference on Computational Geom-etry. pp. 175–178 (2010)
71. Pan, Z., Polden, J., Larkin, N., Duin, S.V., Norrish, J.:Recent progress on programming methods for industrialrobots. In: Proceedings for the joint conference of 41stInternationel Symposium on Robotics and 6th GermanConference on Robotics (2010)
72. Pchelkin, S., Shiriaev, A., Robertsson, A., Freidovich, L.:Integrated time-optimal trajectory planning and controldesign for industrial robot manipulator. In: IEEE/RSJInternational Conference on Intelligent Robots and Sys-tems (IROS) (2013)
73. Petiot, J.F., Chedmail, P., Hascoet, J.Y.: Contributionto the scheduling of trajectories in robotics. Roboticsand Computer-Integrated Manufacturing 14(3), 237–251(1998)
74. Portugal, D., Rocha, R.: A survey on multi-robot pa-trolling algorithms. In: Technological Innovation for Sus-tainability, pp. 139–146. Springer (2011)
75. Quigley, M., Gerkey, B., Conley, K., Faust, J., Foote, T.,Leibs, J., Berger, R.W.E., Ng, A.: ROS: an open-sourceRobot Operating System. In: Open-Source Softwareworkshop of the International Conference on Roboticsand Automation (ICRA) (2009)
76. Rego, C., Gamboa, D., Glover, F., Osterman, C.: Trav-eling salesman problem heuristics: Leading methods, im-plementations and latest advances. European Journal ofOperational Research 211(3), 427–441 (2011)
77. Rego, C., Glover, F.: The Traveling Salesman ProblemAnd Its Variations, chap. Local search and metaheuris-tics, pp. 309–368. Kluwer Academic Publishers (2002)
78. Reinhart, G., Munzert, U., Vogl, W.: A programmingsystem for robot-based remote-laser-welding with con-ventional optics. CIRP Annals-Manufacturing Technol-ogy 57, 37 – 40 (2008)
79. Saha, M., Roughgarden, T., Latombe, J.C., Sanchez-Ante, G.: Planning tours of robotic arms among parti-tioned goals. Robotics Research 25, 207 – 223 (2006)
80. Saha, M., Sanchez-Ante, G., Latombe, J.C.: Planningmulti-goal tours for robot arms. In: International Con-ference on Robotics and Automation (ICRA). vol. 3, pp.3797–3803. IEEE (2003)
81. Shi, X.H., Liang, Y.C., Lee, H.P., Lu, C., Wang, Q.X.:Particle swarm optimization-based algorithms for TSPand generalized TSP. Information Processing Letters103(5), 169–176 (2007)
82. Siciliano, B., Khatib, O. (eds.): Springer Handbook ofRobotics. Springer Verlag, Berlin (2008)
83. Smith, C., Karayiannidis, Y., Nalpantidis, L., Gratal, X.,Qi, P., Dimarogonas, D.V., Kragic, D.: Dual arm ma-nipulationa survey. Robotics and Autonomous Systems60(10), 1340–1353 (2012)
84. Spitz, S.N., Requicha, A.A.G.: Multiple-goals path plan-ning for coordinate measuring machines. In: IEEE Inter-national Conference on Robotics and Automation. pp.2322–2327 (2000)
85. Srivastava, S., Kumar, S., Garg, R., Sen, P.: Generalizedtraveling salesman problem through n sets of nodes. In:CORSE Journal. vol. 7, pp. 97–101 (1969)
86. Stilman, M.: Global manipulation planning in robot jointspace with task constraints. Robotics, IEEE Transactionson 26(3), 576–584 (2010)
87. Vicencio, K., Davis, B., Gentilini, I.: Multi-goal pathplanning based on the generalized traveling salesmanproblem with neighborhoods. In: IEEE/RSJ Interna-tional Conference on Intelligent Robots and Systems(IROS) (2014)
88. Wang, L., Keshavarzmanesh, S., Feng, H.Y., Buchal,R.O.: Assembly process planning and its future in collab-orative manufacturing: a review. The International Jour-nal of Advanced Manufacturing Technology 41(1-2), 132–144 (2009)
89. Wurll, C., Henrich, D.: Point-to-point and multi-goalpath planning for industrial robots. Journal of RoboticSystems 18(8), 445–461 (2001)
90. Wurll, C., Henrich, D., Worn, H.: Multi-Goal Path Plan-ning for Industrial Robots. In: International Conferenceon Robotics and Application (1999)
91. Xidias, E.K., Zacharia, P.T., Aspragathos, N.A.: Time-optimal task scheduling for articulated manipulatorsin environments cluttered with obstacles. In: Robotica.vol. 28, pp. 427–440. Cambridge Univ Press (2010)
92. Yang, H., Shao, H.: Distortion-oriented welding pathoptimization based on elastic net method and geneticalgorithm. Journal of Materials Processing Technology209(9), 4407–4412 (2009)
93. Yao, J., Lin, C., Xie, X., Wang, A., Hung, C.C.: Pathplanning for virtual human motion using improved a*star algorithm. In: Information Technology: New Gen-erations (ITNG), 2010 Seventh International Conferenceon. pp. 1154–1158. IEEE (2010)
94. Zacharia, P.T., Aspragathos, N.A.: Optimal robot taskscheduling based on genetic algorithms. Robotics andComputer-Integrated Manufacturing 21, 67–79 (2005)
95. Zacharia, P., Xidias, E., Aspragathos, N.: Task schedul-ing and motion planning for an industrial manipulator.Robotics and Computer-Integrated Manufacturing 29,449 – 462 (2013)
96. Zhang, K., Collins Jr, E.G., Barbu, A.: An efficientstochastic clustering auction for heterogeneous roboticcollaborative teams. Journal of Intelligent & Robotic Sys-tems 72(3-4), 541–558 (2013)
97. Zhang, K., Collins Jr, E.G., Shi, D.: Centralized anddistributed task allocation in multi-robot teams via astochastic clustering auction. ACM Transactions on Au-tonomous and Adaptive Systems (TAAS) 7(2), 21 (2012)
Appendix A
The list of acronyms used in this survey is stated in
Table A.1.
18 Sergey Alatartsev et al.
Table 1 List of acronyms
ACA Ant Colony AlgorithmATSP Asymmetric TSPBB Branch and BoundBiRRT Bidirectional RRTCETSP Close-Enough TSPDH DenavitHartenberg parametersDOF Degrees of FreedomENM Elastic Net MethodFIH Farthest Insertion HeuristicsGA Genetic AlgorithmGTSP Generalized TSPGTSPN Generalized TSPNHRKGA Hybrid Random-Key Genetic AlgorithmIH Insertion HeuristicLK Lin-Kernighan heuristicMINLP Mixed-Integer Non-Linear ProgramMOPSO Multiple-Objective PSOMTP Multi-Goal Path PlanningMTPGR MTP for Goal RegionsNNA Nearest Neighborhood AlgorithmPA Prims AlgorithmPRM Probabilistic RoadMapsPS Pattern SearchPSO Particle Swarm OptimizationPTP Point-to-pointRLW Remote Laser WeldingRRT Rapidly-growing Random TreesSA Simulated AnnealingSOM Self-organizing MapSOP Sequential Ordering ProblemSRP Safari Route ProblemSSP Shortest Sequence ProblemTPP Touring-a-sequence-of-Polygons ProblemTS Tabu SearchTSA Traveling Salesman AlgorithmTSP Traveling Salesman ProblemTSPN TSP with NeighborhoodsWRP Watchman Route ProblemZRP Zookeeper Route Problem