Multi-Robot Task Allocation and Scheduling

Considering Cooperative Tasks and

Precedence Constraints
Esther Bischoff∗, Fabian Meyer† , Jairo Inga∗ and Sören Hohmann∗
∗ Institute
of Control Systems (IRS)
Karlsruhe Institute of Technology (KIT)
arXiv:2005.03902v1 [eess.SY] 8 May 2020

Email: [email protected]
† Research Center for Information Technology (FZI)

Email: [email protected]

©2020 IEEE. Personal use of this material is permitted.

Permission from IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional purposes,
creating new collective works, for resale or redistribution to servers or lists,
or reuse of any copyrighted component of this work in other works.

Abstract—In order to fully exploit the advantages inherent robots working together. In order to fully exploit these benefits,
to cooperating heterogeneous multi-robot teams, sophisticated sophisticated multi-robot task allocation (MRTA) algorithms
coordination algorithms are essential. Time-extended multi-robot are of great importance [5], [2, p. 1]. Given a set of tasks
task allocation approaches assign and schedule a set of tasks to
a group of robots such that certain objectives are optimized and to be performed by a known set of robots, these algorithms
operational constraints are met. This is particularly challenging assign each task to a capable robot or a team of robots and
if cooperative tasks, i.e. tasks that require two or more robots schedule the tasks such that an executable solution results
to work directly together, are considered. In this paper, we and an objective function is optimized. In many practical
present an easy-to-implement criterion to validate the feasibility, applications, this has to be done with respect to precedence
i.e. executability, of solutions to time-extended multi-robot task
allocation problems with cross schedule dependencies arising constraints that exist between tasks [6], e.g. if the outcome of
from the consideration of cooperative tasks and precedence one task is a prerequisite for the execution of another task.
constraints. Using the introduced feasibility criterion, we propose MRTA problems are often categorized using the taxonomy
a local improvement heuristic based on a neighborhood operator introduced by Gerkey and Matarić [5]. They differentiate
for the problem class under consideration. The initial solution is for one thing between single-task (ST) and multi-task (MT)
obtained by a greedy constructive heuristic. Both methods use
a generalized cost structure and are therefore able to handle robots, dependent on whether robots can execute only one task
various objective function instances. We evaluate the proposed at a time or multiple tasks simultaneously, and for another
approach using test scenarios of different problem sizes, all thing between single-robot (SR) and multi-robot (MR) tasks,
comprising the complexity aspects of the regarded problem. The dependent on whether tasks only require one robot for their
simulation results illustrate the improvement potential arising execution or also cooperative tasks are considered. Instanta-
from the application of the local improvement heuristic.
neous assignment (IA) problems are only concerned with the
assignment problem whereas time-extended assignment (TA)
I. I NTRODUCTION
problems additionally consider the scheduling aspect.
In recent years the deployment of multiple robots working MRTA approaches explicitly considering multi-robot tasks
together towards a common goal has gained increasing atten- have been proposed e.g. by Zhang and Parker [6] who intro-
tion in various application domains such as agriculture [1], duce a heuristic approach incorporating multi-robot tasks and
emergency assistance [2], cleaning work [3] and extraterres- precedence constraints. In [7], they investigate the question
trial exploration [4]. Multi-robot teams provide many advan- of coalition formation, i.e. dynamically finding a team of
tages compared to single-operating robots. Tasks can be per- robots capable of executing a specific task. Both approaches
formed in parallel and the robustness of the system as a whole are only concerned with task allocation and do not consider
increases since malfunctions of single robots can possibly be task scheduling. The time-extended problem is covered in [8]
compensated by the remaining robots. Furthermore, a team where they introduce heuristics to solve the ST-MR-TA prob-
of heterogeneous robots can create synergies that cannot be lem. The drawback of the presented approaches is that they
achieved by an individual robot or even a homogeneous team. only allow for a specific objective function and no precedence
This effect is intensified if also cooperative tasks are consid- constraints are considered.
ered, i.e. tasks which can only be performed by two or more Liu and Kroll [9] introduce a memetic algorithm with a
local search improvement heuristic for problems with single- (vs1 , t1 )1 (t1 , t (t3 , v1e )1
t1 2 )1 )1 v1e
robot and two-robot tasks also using a fixed objective function. v1s (t2 , t3 t3
They are the first to apply an improvement heuristic and t2
additionally give executability constraints inherent to two-
robot tasks. The criteria they state for detecting and repairing
infeasible solutions though are limited to the two-robot task Fig. 1. Schedule of a robot represented as directed path graph.
problem and do not include the consideration of precedence Task nodes are represent as circles, the pentagon illustrates the starting
constraints. node v1s and the triangle the final node v1e of the robot r1 . The arcs
(v1s , t1 )1 , (t1 , t2 )1 , (t2 , t3 )1 and (t3 , v1e )1 determine the task sequence for
Local improvement heuristics used to improve initial solu- robot r1 .
tions based on various existing neighborhood definitions are a
common approach in the field of vehicle routing [10], [11]. In
this field of research, similar kinds of problems like in MRTA II. M ODELING AND P ROBLEM F ORMULATION
emerge. The problem under consideration is to route a fleet We first introduce the notation and modeling used through-
of vehicles to serve distributed customer requests such that a out the paper before giving the formal problem statement.
given objective is optimized and certain constraints are met.
Also extensions to consider heterogeneous fleets of vehicles
A. Notation and Modeling
differing in velocity, capacity or the ability to serve certain
types of customer requests have already been made (cf. [12], We consider a set of tasks T = {t1 , . . . , tn }, n ∈ N, and
[13], [14]). Given the similarities to the properties of MRTA a set of robots R = {r1 , . . . , rm }, m ∈ N. A set of robot
problems, a direct application of the existing neighborhood alliances A = {a1 , . . . , ak }, k ∈ N, k ≥ m, with aj ⊆ R,
operators appears conceivable. Nevertheless, this might lead to ∀j ∈ {1, . . . , k}, specifies the possible robot coalitions.
infeasible solutions arising due to a non-explicit consideration For every robot rl ∈ R its sought schedule can be rep-
of cross-schedule dependencies. These comprise dependencies resented as a directed path graph Gl = (Vl , El ). The set of
between the schedules of individual robots which influence the vertices Vl = {vls , Tl , vle } contains all tasks Tl ⊆ T that are
objective function value [15]. They arise for example from the assigned to any alliance aj robot rl is part of as well as one
consideration of cooperative tasks or from waiting times due starting node vls and one end node vle . If a task ti is assigned
to precedence constraints. to an alliance aj of more than one robots, it is considered
In this paper, we propose a two-step solution approach as a vertex in the path graphs of all participating robots, i.e.
to heterogeneous multi-robot task allocation and scheduling ti ∈ Vl , ∀ l : rl ∈ aj . The nodes vls and vle can be related
problems with cross schedule dependencies arising from co- to a previously defined state or position of the robot rl at
operative tasks and precedence constraints. As a basis for our the beginning and the end of the plan execution, respectively.
solution approach we introduce an easy-to-verify criterion for Furthermore, each task ti ∈ T can be associated with specific
the feasibility, i.e. executability, of solutions to ST-MR-TA predefined properties, e.g. a position for its execution. The
problems with cross-schedule dependencies. This allows for sequence in which robot rl performs the assigned tasks is
an adaption of the relocate neighborhood operator well known determined by the edges (v, w)l ∈ El with v, w ∈ Vl . An
in vehicle routing [16] to make it applicable to the considered example of a directed path graph representing the schedule of
class of MRTA problems. Using the neighborhood operator, we a robot r1 is presented in Fig. 1.
apply an improvement heuristic to improve the initial solution The overall solution, denoted as mission plan M , is de-
found by a constructive heuristic. The constructive heuristic termined by the union of theS robots’ individual S schedules,
makes locally optimal choices and works similar to the Min- i.e. M = (V, E) with V = Vl , E = El , where M
l∈M l∈M
StepSum approach presented by Zhang and Parker [8], but denotes the index set of the robots M := {1, . . . , m}. We
is enhanced to handle precedence constraints. Both heuristics define the set Ein (v) to contain all incoming edges into a
use an introduced generalized objective function structure, vertex v ∈ V , i.e. Ein (v) includes all edges (w, v)l with w ∈ V
thus being applicable to many different objective function and l ∈ M.
instances. When generating the mission plan M , precedence con-
This paper is organized as follows: We introduce the model- straints between the tasks might have to be consid-
ing and notation used within this paper and give the formal ered. They are specified by means of a known function
problem statement in Section II. In Section III we introduce C : T × T → {0, 1} with C(ti , tj ) = 1 if task ti must be fin-
a feasibility definition for mission plans and state an easy- ished before the execution of task tj , with ti , tj ∈ T , ti 6= tj ,
to-verify criterion for its adherence. The solution approach and C(ti , tj ) = 0 if no such constraint exists. The precedence
including a greedy constructive heuristic and an improvement constraints can be included as directed edges (ti , tj )C into
heuristic for the heterogeneous multi-robot task allocation and the mission plan M , if Cij = 1. We denote the mission plan
scheduling problem with cooperative tasks and precedence extended by the set EC containing all precedence constraint
constraints is given in Section IV. We present simulation arcs as M + = (V, E + ) with E + = E ∪ EC . In Fig. 2 an
results in Section V and give a conclusion in Section VI. example of an extended mission plan with two robots, four
 (t1 , t3 )C (t3 , v e) C. Problem Statement
(v 1 , t1 )1
s
t1 t3 1 1
v1s v1e By means of the introduced notation, model and cost
(t1 )1
, t2 )
t2 , t3 structure we are able to state the key problem of this paper:
1 (t 2
(v2s , t4 )2 (t2 , e Problem 1. Let the sets of robots R and robot alliances A and
s , t 2) 2 v2 ) v2e
v2 t4 (t 4 2 the set of vertices V = {v1s , . . . , vm
s
, t1 , . . . , tn , v1e , . . . , vm
e
},
as well as the precedence constraint edges EC be given.
We want to find directed edges E such that the resulting
Fig. 2. Example for an extended mission plan M + with four tasks, two robots mission plan M = (V, E) is connected and feasible and an
and a fulfilled precedence constraint. Task t2 is performed by a coalition of the objective function J dependent on the static and dynamic cost
robots r1 and r2 . The black arc (t1 , t3 )C represents a precedence constraint components, i.e.
specifying that task t1 must be performed before task t3 .
J(M + ) = J cstat (ti , Ein (ti )), cdyn (v, Ein+ (v))


(1)
tasks (of which one is performed by a coalition of both robots) for all ti ∈ T, v ∈ V
and a precedence constraint is depicted. is minimized.
Analogous to Ein (v) we define Ein+ (v) to be the augmented
As stated in Problem 1, only feasible mission plans are
set of incoming edges to a vertex v ∈ V , additionally
sought. We define the feasibility of mission plans in the
considering the precedence constraint edges in EC , i.e. Ein+ (v)
following section.
includes Ein (v) and all edges (w, v)C with w ∈ V . Note that
with the knowledge of Ein+ (v) all predecessor nodes of v are III. F EASIBILITY OF MISSION PLANS
known. The feasibility of mission plans is defined as follows:
B. Generalized Cost Structure Definition 1 (Feasibility of a mission plan). A mission plan M
is feasible, if it can be conducted in finite time. The feasibility
The objective of this paper is to optimize mission plans
of a mission plan comprises the following aspects:
while being able to consider cooperative tasks and precedence
constraints. Therefore, an evaluation criterion is required. We D1.1 The alliance aj ∈ A assigned to any task ti ∈ T by the
present a generic cost structure considering both costs that are mission plan M , must be capable of its execution.
static and dynamic with respect to the optimization problem. D1.2 The mission plan M must represent a topological order.
This allows for the application of the presented solution D1.3 The precedence constraints defined by C
approach on a vast number of different individual objective a) must be consistent with one another and
functions which might be preferable for different problem b) must be fulfilled by the mission plan M .
instances. The necessity of aspect D1.1 is obvious, since the execution
Static costs cstat : T × A → R+ ∪ {∞}, cstat = cstat (ti , aj ) time of a task will be never-ending if the alliance assigned to it
are associated with an alliance aj ∈ A executing task ti ∈ T is incapable of its accomplishment. We present the following
and can be determined for every task-alliance pair prior to assumption as an easy-to-implement method to check for the
the optimization. For example, static costs may consider the first feasibility aspect D1.1.
execution duration or quality of alliance aj performing task ti .
Since the edges Ein (ti ) incoming to vertex ti determine the Assumption 1. We assume the static cost components for all
alliance aj assigned to task ti , the static costs can also be ti ∈ T , aj ∈ A, to be of the form
stated as cstat (ti , Ein (ti )).

∞ if alliance aj is incapable to
Cost components cdyn that are dynamic with respect to cstat (ti , aj ) : T × A → execute task ti ,
the optimization allow for the consideration of additional R+ else.
costs which might depend on the task sequence within the
We furthermore present Lemma 1 as an important insight to
mission plan or on the precedence constraints. Examples for
examine the aspects D1.2 and D1.3 of Definition 1 necessary
dynamic costs include moving durations and transport energy
for the feasibility of a mission plan.
as well as idle times resulting from waiting on the coalition
partners or on precedence constraints to be fulfilled. For a Lemma 1. If the mission plan M is feasible, then the directed
specific vertex v ∈ V the dynamic costs depend on the graph M + of the feasible mission plan M extended by the
incoming augmented edges, i.e. cdyn : V × E + → R+ , precedence constraint edges EC is acyclic.
cdyn = cdyn (v, Ein+ (v)).
Proof: A topological ordering of a directed graph is
Remark. The dynamic costs can easily be augmented to possible if and only if the graph is acyclic (cf. [17, Ch. 4.2]).
additionally take into account explicitly time dependent cost Therefore, D1.2 holds if and only if M is acyclic. For the same
components, i.e. cdyn (v, Ein+ (v), τ ) with τ being the time. Time reason, the precedence constraints fulfill D1.3.a if and only
dependent cost components might for example arise from the if the graph GC = (V, EC ) only containing the precedence
consideration of time window constraints. constraint edges is acyclic. When adding the precedence
constraint edges EC to M , which results in the augmented A. Constructive Heuristic
mission graph M + , two cases have to be considered regarding The constructive heuristic iteratively calculates the effect
the feasibility of M : every new assignment would have on the objective function
• The sets of robots assigned to the tasks ti , tj ∈ and chooses the one that increases the objective function the
T which are related by a precedence constraint least until all tasks ti ∈ T have been assigned. The idea is
edge (ti , tj )C ∈ EC are disjoint. In this case, the similar to the MinStepSum algorithm introduced by Zhang
edge (ti , tj )C does not add a cycle to the acyclic graph M and Parker [8], but we expand it to handle the generalized
and the precedence constraint defined by (ti , tj )C can al- objective function given by (1). Furthermore we augment the
ways be fulfilled if the alliance assigned to task tj ensures method to additionally allow for the direct consideration of
to wait with its execution until task ti is finished. Capable precedence constraints. To do so, we split the tasks into the
alliance-task-assignments (D1.1) ensure the potential time sets of executable tasks Λ and non executable tasks Λ with
increment to be bounded. Λ ∩ Λ = ∅. The set Λ contains all tasks with a nonempty set
• The robot alliances assigned to the tasks ti , tj ∈ T of unassigned precedence tasks. The two sets are initialized
which are related by a precedence constraint with elements ti , where
edge (ti , tj )C ∈ EC are not disjoint. The edge (ti , tj )C (
only closes a cycle in M + , if at least one robot rl ∈ R Λ if C(t, ti ) = 0 ∀ t ∈ T
assigned to both tasks ti and tj violates the precedence ti ∈ (2)
Λ else.
constraint, which means that D1.3.b would not be
fulfilled and M would be infeasible. The detailed constructive heuristic is given in Algorithm 1
and works as follows: The algorithm is given the sets R,
A, and V (divided into the sets V \T and Λ, Λ which are
Using the results of Lemma 1 and combining it with As-
initialized according to (2) such that Λ∪Λ = T ) as well as the
sumption 1, Lemma 2 gives necessary and sufficient conditions
precedence constraints EC . For every robot rl ∈ R, the path
for the feasibility of mission plans according to Definition 1.
graph Gl is initialized with an empty graph containing only
Lemma 2 (Feasibility of a mission plan). Let Assumption 1 the robots initial vertex vls . Using these initial path graphs, the
hold. Then, a mission plan M is feasible w.r.t. Definition 1 if mission plan is initialized and the initial objective function
and only if is calculated (lines 1.1 to 1.5). For every not yet assigned
executable task ti ∈ Λ and every robot alliance aj ∈ A,
L2.1 the static costs of all task vertexes ti ∈ T are finite, i.e.
the increment of the objective function resulting from the
respective assignment is calculated by adding ti as leaf to
cstat (ti , Ein (ti )) < ∞, ∀ti ∈ T,
the path graphs of the respective robots and calculating the
objective function value increment for the resulting interme-
L2.2 and the directed graph of the augmented mission plan diate augmented mission plan M̃ (lines 1.8 to 1.16). Out of
M + is acyclic. all possible assignments the one with the smallest objective
Proof: With the results of Lemma 1, L2.2 gives a function increment is chosen (lines 1.17 to 1.25) . The assigned
necessary and sufficient condition for the feasibility aspects task is deleted from the set of executable tasks Λ and all tasks
D1.2 and D1.3 to be fulfilled. Since Assumption 1 holds, from the set of non executable tasks Λ that became executable
L2.1 gives necessary and sufficient condition to ensure that with the most recent assignment are transferred to the set of
M fulfills D1.1. executable tasks Λ (lines 1.26 to 1.31). The procedure repeats
until all tasks have been assigned. The algorithm terminates
by adding the robots final nodes vle , ∀rl ∈ R, to the respective
IV. S OLUTION A PPROACH robots’ path graphs and determining the resulting mission
plan Minit and the respective objective function value Jinit
Before presenting our two-step solution approach compris- (lines 1.34 to 1.40).
ing a constructive and an improvement heuristic in detail, we
assume the following assumptions to hold: Remark. Assumption 2 implies that for every task ti ∈ T
an assignment will be found for which D1.1 is fulfilled. Fur-
Assumption 2. For a given instance of Problem 1 for every thermore, the explicit consideration of precedence constraints
task ti ∈ T at least one capable alliance aj ∈ A with by means of the sets Λ and Λ guarantees their adherence
cstat (ti , aj ) ∈ R+ exists. according to D1.3.b and Assumption 3 ensures D1.3.a such
Assumption 3. For a given instance of Problem 1 the that also D1.3 will be fulfilled by the mission plan resulting
graph GC = (V, EC ) is acyclic. from the constructive heuristic. The adherence of D1.2 is
ensured by the fact that all newly assigned tasks are added as
Assumption 3 ensures D1.3.a to be fulfilled by the a priori leafs to the path graphs of the respective robots which means
given precedence constraints. Thus, Assumptions 2 and 3 are that tasks assigned to coalitions of several robots are assured
made to assure meaningful problem instances. to have the same sequence within the individual path graphs
Algorithm 1 Constructive Heuristic Algorithm 2 Improvement Heuristic
Require: R, A, Λ, Λ, V \T, EC Require: Minit , Jinit , R, A, C
1: for all rl ∈ R do ⊲ Initialization: path graphs of robots ⊲ Initialization:
2: El ← ∅, Vl ← {vls }, Gl = (Vl , El ) 1: Jbest ← Jinit
3: end for 2: Mbest ← Minit
4: M ← ∪l∈M Gl ⊲ Initialize: mission plan 3: while Stopping criterion not fulfilled do
5: J ← J(M + ) ⊲ Initialize: objective function 4: for all ti ∈ Mbest do
6: while Λ 6= ∅ do 5: for all al̃ ∈ A do
7: ∆min ← ∞ ⊲ Initialize: objective function increment 6: if cstat (ti , al̃ ) < ∞ then
⊲ For every executable tasks-alliance pair calculate 7: Determine all possible reassignments
cost increment of the assignment: of ti to the path graphs Gl̃ : rl̃ ∈ al̃
8: for all ti ∈ Λ do 8: for all possible reassignments M̃ do
9: for all aj ∈ A do 9: Check whether the resulting aug-
10: for all rl ∈ aj do mented mission plan M̃ + is acyclic
11: vleaf ← {v ∈ Vl : ∄w ∈ Vl : (v, w)l ∈ El } 10: if M̃ + is acyclic then
12: Ṽl ← {Vl , ti }, Ẽl ← {El , (vleaf , ti )l } 11: Calculate J(M̃ + )
13: G̃l = (Ṽl , Ẽl ) 12: if J(M̃ + ) < Jbest then
14: end for 13: Mbest ← M̃
15: M̃ ← ∪l∈M G̃l 14: Jbest ← J(M̃ + )
16: J∆ = J(M̃ + ) − J 15: end if
17: if J∆ ≤ ∆min then 16: end if
18: Mmin ← M̃ ⊲ Remember best assignment 17: end for
19: ∆min ← J∆ ⊲ Remember smallest objec- 18: end if
tive function increment 19: end for
20: tmin ← ti ⊲ Remember assigned task 20: end for
21: end if 21: end while
22: end for 22: return Mbest , Jbest
23: end for
24: M ← Mmin
25: J ← J + ∆min
B. Improvement Heuristic
⊲ Check if tmin was the only remaining precedence
constraint to any ti ∈ Λ: We apply a local search to further improve the initial
26: for all ti ∈ Λ with C(tmin , ti ) = 1 do mission plan Minit found by the constructive heuristic. In
27: if C(t, ti ) = 0 ∀t ∈ {Λ ∪ Λ}\{tmin } then every iteration the currently best solution is modified using a
28: Λ ← {Λ, ti } ⊲ Add ti to Λ neighborhood operator which is based on the relocate neigh-
29: Λ ← Λ\{ti } ⊲ Delete ti from Λ borhood first introduced by Savelsbergh and Goetschalckx [16]
30: end if for the routing problem of a homegeneous fleet of vehicles.
31: end for The neighborhood operator given in Definition 2 expands the
32: Λ ← Λ\{tmin } ⊲ Delete tmin from Λ original relocate operation to be applicable to mission plans M
33: end while for heterogeneous robotic teams with cooperative tasks and
34: for all rl ∈ R do precedence constraints.
35: vleaf ← {v ∈ Vl : ∄w ∈ Vl : (v, w)l ∈ El }
36: Vl ← {Vl , vle }, Ẽl ← {El , (vleaf , vle )l } Definition 2 (neighborhood of a mission plan). The neigh-
37: Gl = (Vl , El ) borhood of a mission plan M contains all feasible mission
38: end for plans M̃ that result from relocating one task ti ∈ T out of
39: Minit ← ∪l∈M Gl the path graphs of the alliance aj it is assigned to by M , to
+
40: Jinit ← J(Minit ) any position of the path graphs of the robots of any capable
41: return Minit , Jinit alliance al̃ ∈ A.
Using the neighborhood of Definition 2 and the results of
Lemma 2, the local search improvement heuristic is given in
Algorithm 2. The mission plan and the objective function
are initialized with the results of the constructive heuristic
(lines 2.1 to 2.2). In every iteration, the neighborhood of the
of the robots. Therefore the solution found by the constructive currently best mission plan is determined and evaluated and
heuristic given in Algorithm 1 will always be feasible w.r.t the best neighboring mission plan is chosen. To determine and
Definition 1. evaluate the neighborhood, for all tasks ti ∈ T , all possible
reassignment positions within the path graphs of the robots TABLE I
of every alliance al̃ ∈ A that is capable of the execution TASK DURATION IN ( S ) FOR EVERY TASK - ALLIANCE PAIR
of task ti (i.e. the static costs cstat (ti , al̃ ) are bounded) are Alliance Type A Type B Type C Type D
determined (lines 2.4 to 2.7). To assure the feasibility of {r1 } 100 ∞ ∞ ∞
the resulting new mission plan M̃ , Lemma 1 is applied and {r2 } 100 ∞ ∞ ∞
{r3 } 100 ∞ ∞ 200
it is determined whether the augmented mission plan M̃ + {r1 , r2 } ∞ 110 ∞ ∞
contains cycles (line 2.9). If M̃ + is acyclic and M̃ therefore a {r1 , r3 } ∞ 100 100 ∞
feasible neighboring mission plan, its objective function value {r2 , r3 } ∞ ∞ ∞ 100
is determined and assessed in comparison to the currently
best plan found (lines 2.10 to 2.16). This procedure repeats
until a stopping criterion is fulfilled, e.g. the improvement in To generate specific problem instances for each problem
the objective function J between to iterations falls below a class, we consider every task to be associated with a position
previously determined threshold or a previously determined for its execution. The individual task positions depend on the
maximum number of iterations is reached. task indices i and are located in the Cartesian plane at
To conduct the acyclicity check (line 2.9), any cycle search
for digraphs can be applied. Since only the statement about x(ti ) = L0 cos(θ0 (ti )) + L1 cos(θ1 ) (3)
acyclicity and not the potentially existing cycles are of interest, y(ti ) = L0 sin(θ0 (ti )) + L1 sin(θ1 ) (4)
we implemented an algorithm based on iteratively removing
leafs from the augmented mission plan M + , similar to the with
algorithm of Kahn [18]. L0 = 50m, L1 ∈ [0m, 10m],
V. S IMULATION R ESULTS i π
θ0 (ti ) = 2π + , θ1 ∈ [0, 2π].
The experimental setup and the structure of the objective |T | |T |
function used to evaluate the introduced MRTA approach are The values for L1 and θ1 are equally distributed with respect
described in the following section followed by the presentation to the given intervals over all problem instances of a certain
and discussion of the simulation results. problem class. Thus, for each task of a problem class an area
is defined in which the corresponding task must be located.
A. Experimental Setup
By these definitions we aim to ensure a high comparability
To evaluate the proposed solution approach we set up a of all the problem instances of a given problem class. For all
generalized simulation framework for MRTA problems with robots rl , l ∈ {1, 2, 3}, their starting node vls is associated
precedence constraints and cooperative tasks. It consists of with the initial position (0, 0) at the origin of the coordinate
four different types of tasks and three different mobile robots. system whereas the position of their end node vle is set to be
Each task type can be processed by a subset of the considered arbitrary.
alliances A = {{r1 }, {r2 }, {r3 }, {r1 , r2 }, {r1 , r3 }, {r2 , r3 }}.
Based on the different task types we define six different B. Structure of the Objective Function
problem classes: 3A1BCD, 3A2BCD, 3A3BCD, 6A1BCD, For every task-alliance pair the static cost component
6A3BCD and 6A3BCD. The first number in these coded cstat (ti , aj ) represents the duration needed by alliance aj ,
problem classes describes the number of tasks of type A j ∈ {1, . . . , 6} to perform task ti ∈ T . The respective values
whereas the second number denotes the number of tasks of dependent on the task types are given in Table I.
each type B, C and D. To indicate the type an individual task The dynamic cost components include:
belongs to, we add a superscript to the tasks ti . The index i 1 + +
• cdyn (v, Ein (v)): For every edge (w, v)l ∈ Ein (v) we
starting at 1 in ascending order is first assigned to all task of calculate the driving time τd,v (rl ) needed by robot rl
type A followed be the tasks of type B, C and D. This results to travel from the position of its previous node w to
for example in problem class 3A2BCD having the set of tasks the position of v. To calculate the individual traveling
T3A2BCD = {tA A A B B C C D D
1 , t2 , t3 , t4 , t5 , t6 , t7 , t8 , t9 }. times we use the euclidean distance between the positions
Let |A|, |B| and |C| denote the number of tasks of type A, of w and v and the robot’s individual velocities v(rl ),
B and C, respectively, within a given problem class. For all which are set to be v(r1 ) = v(r2 ) = 2 m s−1 and
problem classes we assume the following types of precedence v(r3 ) = 1 m s−1 .
constraints to exist exactly once. 2 + +
• cdyn (ti , Ein (ti )): For every edge (w, ti )l ∈ Ein (ti ) the
A A
• (t1 , t2 )C : The first type A task must be completed waiting time τw,ti (rl ) of robot rl resulting from waiting
before processing of the second type A task can begin. on coalition partners to reach the position of ti or on
A B
• (t3 , t|A|+1 )C : The third type A task must be completed precedence constraints for ti to be fulfilled is determined.
3 + +
before processing of the first type B task can begin. • cdyn (v, Ein (v)): For every edge (w, v)l ∈ Ein (v) we
C D
• (t|A|+|B|+1 , t|A|+|B|+|C|+1 )C : The first type C task must calculate the euclidean distance dv (rl ) driven by robot rl
be completed before processing of the first type D task to travel from the position of its previous node w to the
can begin. position of v.
 In the objective function we consider the total mission
60 tA
2 r1
duration given by latest finishing time over all robots rl ∈ R, r2
tA
1
i.e. 40 r3
tA

y-coordinate in m
( 3 prec.
X
J1 (M + ) = max (cstat (ti , Ein (ti )) + τw,ti (rl )) 20 tD
l∈{1,2,3} 9
ti ∈Vl
0
)
+
X
τd,v (rl ) , tB
4

v∈Vl −20 tD
8
(5)
as well as the average finishing time of the robots −40 tC6
( tB
5 tC7
3
+ 1X X −60
J2 (M ) = (cstat (ti , Ein (ti )) + τw,ti (rl )) + −60 −40 −20 0 20 40 60
3
l=1 ti ∈Vl
) x-coordinate in m
X
τd,v (rl )
v∈Vl Fig. 3. Resulting mission graph representing a locally optimal solution of an
(6) instance of problem class 3A2BCD. The colored paths are associated to a
and the sum over all driven distances divided by the number robots individual schedule. Black edges illustrate precedence constraints.
of robots
3
1XX tD
J3 (M + ) = dv (rl ). (7) ↓ 9 ↓ tC6 ↓ tC7 ↓ tD
8
3 r3
l=1 v∈Vl
Robots
This choice of weighting factors reflects the fact that in ↓ tA
3 ↓ tB
4 ↓ tB
5 ↓ tA
2 ↓ tD
8
r2
many practical relevant scenarios the minimization of the total
mission duration comes with the highest priority.
To test and validate our approach we evaluated 100 problem ↓ tA
1 ↓ tB
4 ↓ tB
5 ↓ tC6 ↓ tC7
r1
instances for each problem class on an Intel(R) Core(TM) i5-
8250U CPU at 1.6GHz and 8GB RAM with a Windows 10 150 300 450 600 750
operating system. Our optimization method was implemented Mission time in s
using MATLAB R2017b.

C. Results Fig. 4. Gantt chart corresponding to the mission graph depicted in Fig. 3.
For all tested instances of every problem class, our pro- Tasks of the same type are identically colored. Traveling times are represented
by thin lines and waiting times as dashed black lines. Gray arrows pointing
posed construction and improvement heuristic yield feasible downwards indicate the starting time of each task.
mission plans. Fig. 3 shows the resulting mission graph of an
optimized solution of an instance of problem class 3A2BCD.
Additionally to the requirements for assigning cooperative heuristic finds valid solutions in 0.01s to 0.05s. The computa-
tasks to capable alliances, which are denoted in Table I, this tional effort for the improvement heuristic increases noticeably
problem class requires the precedence constraints (tA A
1 , t2 )C , with increasing problem sizes.
A B C D
(t3 , t4 )C and (t6 , t8 )C to be fulfilled. The temporal behavior
corresponding to the mission graph depicted in Fig. 3 is D. Discussion
visualized in the Gantt chart in Fig. 4. Herein tasks of the The simulation results show that the constructive heuristic
same type are colored identically. Waiting times are marked yields feasible initial solutions independent of the problem
with dashed black lines, while traveling times are represented size. The assessment of the improvement heuristic depicted
by thin lines in the same color as their succeeding task.
An assessment of the proposed improvement heuristic based
on the evaluation of 100 problem instances of each problem Problem Class Constr. Heuristic Impr. Heuristic
3A1BCD 0.01s 0.29s
class is given in Fig. 5. It can be seen that applying the 3A2BCD 0.02s 4.10s
improvement heuristic leads to an average improvement of 3A3BCD 0.04s 19.58s
around 10 % and a maximum improvement of almost 30 %. 6A1BCD 0.02s 2.00s
6A2BCD 0.04s 12.04s
For the smallest problem class 3A1BCD the average improve- 6A3BCD 0.05s 36.58s
ment drops to approximately 1.5 %.
TABLE II
The computation times for the developed constructive AVERAGE COMPUTATION TIMES OF CONSTRUCTION AND IMPROVEMENT
heuristic and the neighborhood-based improvement heuristic HEURISTIC
are given in Table II. It can be seen, that the constructive
 Improvement in % sults. Furthermore, we will put focus on improving calculation
time to allow for solving larger problem instances.
R EFERENCES
[1] C. Zhang and N. Noguchi, “Development of a multi-robot tractor system
for agriculture field work,” Computers and Electronics in Agriculture,
vol. 142, pp. 79–90, 2017.
[2] G. A. Korsah, “Exploring bounded optimal coordination for heteroge-
neous teams with cross-schedule dependencies,” Dissertation, Carnegie
C D BC D BC D BC D BC D BC D
1B 2 3 1 2 3 Mellon University, Pittsburgh, USA, 2011.
3A 3A 3A 6A 6A 6A [3] P. García, P. Caamaño, R. J. Duro, and F. Bellas, “Scalable task
assignment for heterogeneous multi-robot teams,” International Journal
of Advanced Robotic Systems, vol. 10, no. 2, p. 105, 2013.
[4] J. Schneider, D. Apfelbaum, D. Bagnell, and R. Simmons, “Learning
Fig. 5. Average improvement of the optimized solution compared to the inital opportunity costs in multi-robot market based planners,” in 2005 IEEE
solution depicted as blue dots. The minimum and maximum improvement are International Conference on Robotics and Automation (ICRA). Piscat-
represented as lower and upper bound of the blue lines. away, N.J: IEEE, 2005, pp. 1151–1156.
[5] B. P. Gerkey and M. J. Matarić, “A formal analysis and taxonomy of
task allocation in multi-robot systems,” The International Journal of
Robotics Research, vol. 23, no. 9, pp. 939–954, 2004.
in Fig. 5 reveals that applying the improvement heuristic [6] Y. Zhang and L. E. Parker, “Considering inter-task resource con-
has a high potential to further improve the initial solution straints in task allocation,” Autonomous Agents and Multi-Agent Systems,
especially for larger problem sizes. Nevertheless, there is a big vol. 26, no. 3, pp. 389–419, 2013.
[7] ——, “Iq-asymtre: Forming executable coalitions for tightly coupled
gap between the maximum and the minimum improvement of multirobot tasks,” IEEE Transactions on Robotics, vol. 29, no. 2, pp.
different instances within a certain problem class. This is most 400–416, 2013.
likely due to the fact that some initial solutions are close to a [8] ——, “Multi-robot task scheduling,” IEEE International Conference on
Robotics and Automation (ICRA), pp. 2992–2998, 2013.
local optimum, while others are not. This in turn influences the [9] C. Liu and A. Kroll, “Memetic algorithms for optimal task allocation
improvement that can be achieved by subsequent local search. in multi-robot systems for inspection problems with cooperative tasks,”
To further improve our solution approach modifications to the Soft Computing, vol. 19, no. 3, pp. 567–584, 2015.
[10] S. Ropke and D. Pisinger, “An adaptive large neighborhood search
neighborhood operator based on our feasibility criterion and heuristic for the pickup and delivery problem with time windows,”
the application of other improvement heuristics, that are able Transportation Science, vol. 40, no. 4, pp. 455–472, 2006.
to escape local optima, are conceivable. [11] F. Ferrucci, Pro-active Dynamic Vehicle Routing. Berlin, Heidelberg:
Springer, 2013.
VI. C ONCLUSION [12] R. Baldacci, M. Battarra, and D. Vigo, “Routing a heterogeneous fleet
of vehicles,” in The Vehicle Routing Problem: Latest Advances and New
In this paper we presented new insights to the feasibility Challenges. Springer, 2008, vol. 43, pp. 3–27.
of mission plans for time-extended multi-robot task allocation [13] J. J. De la Cruz, C. D. Paternina-Arboleda, V. Cantillo, and J. R.
Montoya-Torres, “A two-pheromone trail ant colony system—tabu
and scheduling problems with cooperative tasks and prece- search approach for the heterogeneous vehicle routing problem with
dence constraints. We gave an easy-to-implement criterion time windows and multiple products,” Journal of Heuristics, vol. 19,
to verify the feasibility of mission plans and proposed a no. 2, pp. 233–252, 2013.
[14] C. D. Tarantilis and C. T. Kiranoudis, “A flexible adaptive memory-
constructive and an improvement heuristic working with a based algorithm for real-life transportation operations: Two case studies
generalized objective function structure. We demonstrated from dairy and construction sector,” European Journal of Operational
the effectiveness of the proposed method by evaluating it Research, vol. 179, no. 3, pp. 806–822, 2007.
[15] G. A. Korsah, A. Stentz, and M. B. Dias, “A comprehensive taxonomy
using several generalized problem classes of different size. for multi-robot task allocation,” The International Journal of Robotics
The results show that both the constructive as well as the Research, vol. 32, no. 12, pp. 1495–1512, 2013.
improvement heuristic yield feasible mission plans and that [16] M. W. P. Savelsbergh and M. Goetschalckx, “An efficient approximation
algorithm for the fixed routes problem,” University of Georgia, Atlanta,
the local search in average yields significant improvements. In USA, 1992.
future research we will focus on more neighborhood-operators [17] R. Sedgewick and K. Wayne, Algorithms, 4th ed. Addison-Wesley
based on the introduced feasibility criterion and apply more Professional, 2011.
[18] A. B. Kahn, “Topological sorting of large networks,” Communications
sophisticated improvement heuristics to further improve the re- of the ACM, vol. 5, no. 11, pp. 558–562, 1962.