Article

Maximum Task Reassignment Allocation (TRMaxAlloc)

Algorithm in Multi-Robot Systems
Rahim Ali Qamar1 , Mubashar Sarfraz2 , Sajjad A. Ghauri1 , Asad Mahmood3

1School of Engineering & Applied Sciences, ISRA University, Islamabad, Pakistan

2 Faculty
of Engineering and Computer Science, National University of Modern Languages, Islamabad, Pakistan
* Correspondence: ASAD SB

Abstract: This paper addresses the problem of maximum task allocation to a multi-robotic system 1

under strict time constraints. In order to achieve this objective, we propose some enhancements 2

to the existing performance impact (PI) algorithm, which is a task allocation algorithm that works 3

under strict time constraints and is based on the concept of minimizing the sum of the path cost for 4

all robots. The PI algorithm can be applied in search and in scenarios where robots are required to 5

provide support services, i.e. food and medicine, to the survivors. The proposed scheme works in 6

two phases, i.e., in the first phase, task allocation using the PI algorithm, and then in the second 7

phase, reassignment of the assigned task to other robots to create feasible time space for unassigned 8

tasks. The key to the task reassignment, which the PI algorithm lacks, is the selection of a task for 9

assignment to the other robots. This paper proposes three task selection methods along with a cost 10

mechanism for reassignment of tasks to other robots. Numerical simulation of the enhanced PI 11

algorithm with three task selection techniques and comparison with the parent PI algorithm confirms 12

an increase in the allocation of tasks under different constraints. 13

Keywords: Task Reassignment, Performance Impact, Multi-Robots System, TRMaxAlloc Algorithm 14

1. Introduction 15

1.1. Background 16

Multi-robotic systems are consistently gaining interest in many fields, including agri- 17

culture, manufacturing, forest firefighting, ocean and sea exploration [1], underwater 18

surveillance, search and rescue missions [2–4], power line inspection operations [5] etc. 19

Generally, cheaper and heterogeneous robots are employed and may have many limitations 20
Citation: Lastname, F.; Lastname, F.;
Lastname, F. Title. Journal Not Specified
like energy, communication range, sensor availability, etc. But these robots communicate 21

2022, 1, 0. https://doi.org/
and coordinate to overcome all the limitations and perform their assigned tasks [6,7]. 22

With the increase in the number of tasks, it becomes extremely difficult to allocate 23
Received: the tasks to robots to meet various objectives and constraints. Search and rescue is a time- 24
Accepted:
constrained scenario where tasks are subjected to deadlines as survivors may require food, 25
Published:
medicine, or recovery. Therefore, task allocation with time-limited scenarios is considered 26

Publisher’s Note: MDPI stays neutral as an NP-hard problem [8,9]. 27

with regard to jurisdictional claims in The key objective of a multi-robotic system for a search and rescue scenario is to 28

published maps and institutional affil- provide support services to the maximum number of survivors. A significant reduction in 29

iations. victims and economic losses is possible by efficiently allocating tasks to the multi-robotic 30

Copyright: © 2022 by the authors.

system in a search and rescue scenario. Due to the constraints, limited resources, and nature 31

Submitted to Journal Not Specified

of the problem, the search and rescue scenario is considered dynamic [10]. 32

for possible open access publication In order to solve the task allocation problem in a time-limited environment, authors in 33

under the terms and conditions [11], proposed a novel heuristic algorithm named the Performance Impact (PI) algorithm 34

of the Creative Commons Attri- with better performance than the well-known consensus-based bundle algorithm. The 35

bution (CC BY) license (https:// PI algorithm is based on minimizing the sum of path costs over all the vehicles. The 36

creativecommons.org/licenses/by/ PI algorithm uses the same task inclusion, communication, and conflict resolution and 37

4.0/).

Version August 24, 2022 submitted to Journal Not Specified https://www.mdpi.com/journal/notspecified

Version August 24, 2022 submitted to Journal Not Specified 2 of 20

consensus structure as the CBBA algorithm. The PI algorithm is proved to be better than 38

CBBA in time-critical applications and gives a lower mean time cost. 39

The PI algorithm gives sub-optimal solutions due to the problem of local minima and 40

introduces the concept of PI-MaxAss, which modifies the cost structure of the PI algorithm 41

during the task reassignment phase. Task reassignment creates feasible vacant slots in the 42

task list of vehicles to include unassigned tasks, thus enhancing the swap distance, which 43

limits the tasks to be exchanged between them [12]. 44

The PI algorithm was expanded by introducing task selection approaches such as PI- 45

MaxAss and soft-max task selection. With increased time cost, incorporating task selection 46

approaches improves PI task allocation performance [13]. The static structure problem is 47

resolved by introducing online rescheduling in real time and a soft-max action selection 48

procedure that increases the algorithm’s exploratory properties and addresses the local 49

minima problem. 50

1.2. Motivation and Contribution 51

In this paper, we extend the PI algorithm to maximize the task allocation and to avoid 52

the problem of local minima. The objective is to maximize task allocation to multi-robotic 53

systems, which is accomplished in two phases: 54

1. PI Phase 55

2. TRMaxAlloc Phase 56

In the PI phase, tasks are assigned to robots in a multi-robotic system considering all 57

constraints. If there are tasks that are still not assigned, then the TRMaxAlloc phase is 58

executed. In the TRMaxAlloc phase, some of the assigned tasks are reassigned to other 59

robots to create feasible space for the assignment of unassigned tasks. In the TRMaxAlloc 60

phase, we introduce the transformed inclusion performance impact (TrIPI) and transformed 61

removal performance impact Tr RPI concepts, which reduce the cost of already assigned 62

tasks by a factor dependent on computed IPI and RPI. The aim of the cost reduction 63

of a task is to make it re-assignable, as tasks assigned through the PI algorithm get their 64

minimum cost value. 65

Furthermore, three task selection approaches are proposed in the TRMaxAlloc phase:- 66

• Task selection (TS) based on maximum cost reduction (TSMCR) 67

• Task selection based on minimum distance (TSMD) 68

• Task selection based on the distributed bee algorithm (TSDBA) 69

1.3. Organization of the Article 70

The rest of the paper follows the following structure: The task assignment problem and 71

current approaches are presented in Section 2. In Section 3, the mathematical description 72

of the problem and architecture is presented in detail. The issues of the PI algorithm and 73

proposed task reassignment algorithm for maximum task allocation is presented in section 74

4. The extensive simulations followed by a discussion is comprehensively demonstrated in 75

section 5 and concluding remarks along with future direction is in Section 6. 76

2. Related Work 77

In this section, the overview of existing work for distributed task allocation to a 78

multi-robotic system under time constraint conditions is presented in a comprehensive 79

manner. Task allocation (TA) is a very important area in multi-robot systems and NP-hard 80

combinatorial optimization problems. One of the basic factors in TA to a multi-robotic 81

system is the travelled distance, which is formulated as a travelling salesman problem 82

[14,15]. TA can be categorised into central and distributed [16–18]. 83

In central coordination, task allocation to a multi-robotic system is performed from a 84

central server, and failure of the central server causes the whole task allocation process to 85

be halted [19–21]. In contrast, each robot in a multi-robotic system assigns its own tasks 86

in distributed coordination. Within its sensing limits, each robot observes the status of its 87
Version August 24, 2022 submitted to Journal Not Specified 3 of 20

neighbors [22]. It compares the competency of its task execution with that of its neighbors’ 88

tasks and allocates itself cost-effective tasks [23]-[24]. 89

The consensus-based bundle algorithm (CBBA) employs the single task, single robot 90

(ST-SR) task allocation method [25] . For a multi-robotic system, each robot iterates through 91

a bundle construction phase, where a single bundle is constructed and updated as a market- 92

based auction takes place, and a conflict resolution phase, which is used for consensus on 93

the winning bids. Each robot released outbids tasks after each auction. The task inclusion 94

principle of the CBBA algorithm is based on reducing the local cost generated by each robot 95

[26]. 96

Modified particle swarm optimization as a centralized algorithm (MCPSO) to solve 97

the task allocation problem and to provide a benchmark in the search and rescue domain. 98

In order to prove the hypothesis that centralized algorithms perform better than distributed 99

ones, MCPSO is compared with PI and CBBA algorithms. The authors proposed a method 100

to determine the priority of a task based on the task deadline. Furthermore, an insert 101

operation and an updated local search method based on 2-OPT and the exchange method 102

were also presented [27]. 103

The original PI algorithm, which was extended with the critical tasks, is presented 104

in [28]. The objective is to maximize the critical tasks and score of the whole task assignment 105

solution under the constraint of the limited capacity of an unmanned aerial vehicle (UAV). 106

The task inclusion and conflict resolution criteria are modified by maximizing the total 107

score instead of minimize the average waiting time of the final assignment as a global 108

objective. They also proposed a time-discounted reward strategy and incorporated it into 109

the solution process’s global function. A task list resizing phase is also proposed to meet 110

the constraint of limited capacity and maximise the assignment of critical tasks. 111

In [7] distributed bees algorithm (DBA) for the allocation of tasks in a swarm of robots 112

is presented. The DBA is based on the concept of cost, i.e., distance from the task and 113

quality, priority, or complexity of the target [29]. Task utility, or the likelihood that a robot 114

will choose a task, is calculated using task cost and quality. Task selection is based on the 115

roulette wheel mechanism, and the selected task is communicated to all the robots in the 116

range with the task quality. 117

3. Mathematical Description and Architecture 118

A search and rescue operation is a time-constrained scenario where survivors need 119

support services before their specific deadlines. A general search and rescue scenario in an 120

urban area is described in figure 1, where survivors are distributed in an area and robots, 121

i.e., UAVs, are providing support services to the survivors. In this scenario, robots are 122

required to cooperatively distribute the tasks among themselves, i.e., to provide support 123

services to the survivors. All robots have the capability to communicate with each other for 124

cooperative task assignment and conflict resolution. Survivors’ locations are transmitted 125

to all the robots through a radio link from the base station, and robots are not required to 126

return to their initial positions after finishing the mission. 127

Consider N homogeneous UAVs R = [r1 , r2 , ..., r N ] working together to support M 128

survivors in search and rescue missions. Here, robots and tasks T = [t1 , t2 , ...t M ] are 129

alternatively used as UAVs and survivors, respectively. For search and rescue scenarios, 130

UAVs are required to start providing support services before a certain time limit, after 131

which the task is completed. Let D = [d1 , d2 , ..., d M ] be the set of deadlines for the tasks 132

respectively. A UAV can perform one task at a time, and each task can be performed by 133

one UAV. Therefore, it is a case of a single-task single robot (ST-SR) problem. Each robot 134

can perform any type of task in the homogeneous robot case, whereas in the heterogeneous 135

robot case, a robot-task compatibility matrix H is defined, with entries hij indicating the 136

robot’s ability to perform a task. 137

Each UAV is limited to performing a maximum of L tasks and the set limit of tasks 138

can be due to the carrying capacity of support supplies. UAVs collaborate based on certain 139
Version August 24, 2022 submitted to Journal Not Specified 4 of 20

Figure 1. Proposed Architecture

criteria to provide support, and each UAV is assigned a number of tasks less than L such 140

that si ∩ s j = ∅, where si and s j are ordered lists of tasks for robots i and j, respectively. 141

Tasks assigned to a UAV are executed in a specific order so that each task is completed 142

before its respective deadline but at the lowest possible time cost of ciq , where ciq is the 143

time cost incurred by robot i to complete task tq . Each robot is assigned a number of tasks 144

denoted as S = [s1 , s2 , ..., si , ..., s N ]. A robot’s cost to perform a task is the time it takes to 145

travel to the task location from its current location plus the time it takes to complete the 146

task (Tdur ). The time required to perform each type of task can be different. Therefore, 147

under this scenario, the objective of the proposed research is to optimize the task allocation 148

process in such a way that the maximum number of survivors gets support services while 149

considering the above-stated conditions as presented in Eq. (1):- 150

N
J = max [∑i=1 |si |] (1)
151

3.1. Performance Impact Algorithm 152

The performance impact algorithm for the assignment of tasks to a multi-robotic 153

system with time constraints is shown in the figure ??. They verified their approach by 154

applying it to a search and rescue scenario. The PI algorithm iteratively performs task 155

inclusion, communication, and conflict resolution phases to build a list of tasks for each 156

vehicle.The above process flow iteratively runs on every robot such that the process stops 157

when all tasks have been allocated or consensus has been achieved against some criteria. 158

The process flow of the main program of the PI algorithm running on each robot is also 159

shown in the figure ??. 160

3.1.1. Task List 161

Each robot maintains an ordered list of tasks si assigned to it. Tasks are included in 162

the task list of a robot using a PI costing mechanism and task selection criteria. Each robot 163

executes tasks assigned to it as per their order in the task list. The addition of a new task t7 164

at two different positions is shown in figure 2. 165

Version August 24, 2022 submitted to Journal Not Specified 5 of 20

Table 1. Notation & Description

Description Notation
Set of N robots R = [r1 , r2 , ..., r N ]
Set of M tasks T = [t1 , t2 , ...t M ]
Deadline of each task D = [d1 , d2 , ..., d M ]
Compatibility Matrix H
Maximum number of tasks L
Ordered list of tasks si
Time cost of task tq ciq (si )
Time to perform a task tq Tdur
Number of tasks assigned to ri | si |
Robot ID Vector βi
Inclusion Performance of Task v⊕
q si , t q )
(
RPI of Task v⊖
q ( si , t q )
Transformed IPI of each Task ℘iq ∧

Transformed RPI of each Task ℘iq ∼

Visibility of Task λiq

Quality of Task ζq

Algorithm 1: PI Task Allocation Algorithm

1 Initialize ⇐ N, M, L, RPI
2 Consensus ⇐ false
3 while Consensus is false do
4 Task Inclusion Phase
5 Communication & Conflict Resolution Phase
6 end
7 return
8 Status ⇒ Consensus

3.1.2. PI Costing Mechanism 166

A task can be added at any position in the task list but adding a task at different 167

positions results in different overall time cost. It is due to the execution order of tasks and 168

also tasks at later position in si are also shifted. Therefore with the addition of task tq in si , 169

overall cost of tasks as per task list is sum of time cost of task tq in si plus cost effect due to 170

shifting of later tasks. 171

First of all, assume that a task t7 is conformable for inclusion into the task list of r1 at 172

its all positions. Then the first step is to find the appropriate position in task list, where its 173

inclusion gives minimum increase in time cost. In figure 2, t7 is included at two positions 174

in task list. The overall cost to perform all tasks in the task list before insertion of t7 is given 175

as:- 176

c1 (s1 ) = c1,2 (s1 ) + c1,11 (s1 ) + c1,9 (s1 ) (2)

where, 177

c1,2 (s1 ) = f r1 , t2 + Tdur

c1,11 (s1 ) = c1,2 (s1 ) + f t2 ,t11 + Tdur
c1,9 (s1 ) = c1,11 (s1 ) + f t11 ,t9 + Tdur

Similarly, overall cost to perform all tasks in the task list after insertion of t7 between 178

t11 and t9 is given as:- 179

Version August 24, 2022 submitted to Journal Not Specified 6 of 20

ci (s1 ⊕3, t7 ) = c1,2 (s1 ) + c1,11 (s1 ) + c1,7 (s1 ) + c1,9 (s1 ) (3)
where,
c1,7 (s1 ) = c1,11 (s1 ) + f t11 ,t7 + Tdur
and (s1 ⊕3, t7 ) is the addition of task t7 in task list of r1 at position 3. Similarly, increase in
time cost by adding t7 at the end of task list is given as:-

c1 (s1 ⊕4, t7 ) = c1,2 (s1 ) + c1,11 (s1 ) + c1,7 (s1 ) + c1,9 (s1 ) (4)

Position in task list for inclusion of t7 is selected based on minimum increase in total cost
i.e.,
l = min(c1 (s1 ⊕3, t7 ) − c1 (s1 ), c1 (s1 ⊕4, t7 ) − c1 (s1 )) (5)
where, l is the position of the task inserted in the task list.

Figure 2. Task List

180

3.1.3. Inclusion Performance Impact (IPI) 181

The IPI of task tq is defined as overall increase in time cost of tasks in si by adding 182

task tq at position l in si . If there is no task in the task list then IPI of task tq is equal to the 183

time cost of task tq . Mathematically, IPI of task tq in si is computed as:- 184

|s |+1
v⊕ i ∗
q ( si , tq ) = minl =1 vq,l ( si , tq ) (6)
where,
|si |+1 | si |
v∗q,l (si , tq ) = ∑ ci,x (si ⊕ l, tq ) − ∑ ci,x (si ) (7)
x =l x =l

and si ⊕ l, tq denotes the inclusion of task tq in task list si at robot ri at position l. Equation 185

(6) computes the minimum IPI. In this way, the IPI of tasks is computed and stored in the 186

IPI vector ℘i⊕ = [v1⊕ , v2⊕ , ..., v⊕ ⊕ ⊕

q , ..., vm ], vq is the IPI of task tq in task list si of robot ri . The 187

IPI of task tq at position l in si is computed by equation (7). 188

3.1.4. Removal Performance Impact (RPI) 189

Each robot will calculate the RPI of tasks in the task list after it is full or no further 190

tasks can be added. RPI is a measure of the cost of performing a removed task and is given 191

as:- 192

| si | | si |
v⊖
q ( si , t q ) = ∑ ci,x (si ) − ∑ ci,x (si ⊖ tq ) (8)
x =b x = b +1
Version August 24, 2022 submitted to Journal Not Specified 7 of 20

v⊖
q ( si , tq ) shows the PI of a task being removed from its location in si , where b is the position 193

of task tq in si . 194

3.1.5. PI Task Inclusion Phase 195

In this phase, each robot builds a list of tasks that it can execute while considering all 196

the constraints. The task inclusion process as per figure 3 is explained as follows:-

Figure 3. Task Inclusion Steps

197

(i) If the task list is not full, each robot computes the IPI of tasks that are included in 198

its task list. Eq. (7) computes IPI of task tq at position l in si and Eq. (6) computes 199

minimum IPI. In this way, IPI of tasks is computed and stored in IPI vector ℘i⊕ . 200

(ii) Once the IPI of tasks is computed, we now decide which tasks to include in the
task list based on the IPI vector of tasks. For task inclusion in si , each value in the
IPI vector is compared with the corresponding value in the RPI vector. A task tq
is selected for inclusion in si , which gives the maximum cost reduction at position
l, given in Eq. (9):-
⊕
g = maxqM=1 (℘i,q − ℘i,q ) (9)

(iii) The overall time reduction is improved by iteratively adding tasks to robot task 201

list. With the addition of tasks to the task list, the time cost vector ciq against each 202

task is also updated. Tasks are added in si of robot i until the task list is full or no 203

further tasks can be added. 204

(iv) After the task list is full or no further tasks can be added, each robot will update 205

its RPI vector against the tasks added in its task list. To update the RPI vector, 206

compute RPI values against the tasks in si using Eq. (9). 207

(v) In addition, the robot ID vector is also updated corresponding to the added tasks 208

in si . 209

3.1.6. Communication & Conflict Resolution Phase 210

During the communication and conflict resolution phases, each robot broadcasts its 211

RPI vector and robot ID vector to other robots, it is connected as per the topology. After 212

receiving RPI vectors from other robots, each robot updates its RPI and robot ID vectors 213

against the minimum RPI value achieved so far. A received RPI value of a task lower than 214

its computed RPI value results in the removal of that task from its own task list. After a task 215

is removed from a robot’s task list, its time cost vector is also updated. The algorithmic flow 216
Version August 24, 2022 submitted to Journal Not Specified 8 of 20

Algorithm 2: Communication & Conflict Resolution Steps

1 send Connected Robots ⇒ ℘i & β i
2 receive Connected Robots ⇒ ℘ j & β j
3 compare ℘i &℘ j
4 Find No. of Task to Remove ⇒ Remote Task List (RTL)
5 while All Tasks Removed do
6 Task List ⇒ Remove Task
7 ciq ⇒ U pdateTimeCost
8 β i ⇒ U pdateRobotID
9 end
10 return
11 Update ⇒ ℘i
12

of the communication and conflict resolution phases is shown in figure ??. Task inclusion, 217

communication, and conflict removal processes on each robot continue until all tasks have 218

been allocated or no further improvement can be made. 219

4. Reassignment Algorithm for Maximum Task Allocation 220

With tight deadlines and a high ratio of tasks, PI can fail to allocate more tasks 221

even though it is possible to do so because of the scoring strategy and absence of task 222

reassignment capability. Once a task has been included in the task list of a robot, it cannot 223

be removed from the task list unless some other robot includes this task in its task list with 224

a lower IPI. For example as shown in the figure 4, Robot r1 includes task t1 in its task list 225

due to the lowest IPI, whereas Robot r2 includes task t2 in its task list due to the lowest IPI. 226

Now task 3 cannot be included in the task list of any robot because no robot meets the task 227

deadline requirement. Robot r1 can serve t3 , but its IPI is greater than the IPI of task t1 , so 228

it cannot be included in the task list of Robot r1 . If robot r2 serves tasks t1 and t2 , which the 229

PI algorithm does not permit, all three tasks can be served before their deadlines.

Figure 4. Issue in PI Task Allocation

230
The maximum number of survivors who get support services before their respective 231

deadlines is achieved is the main objective and can be attained, if already assigned tasks 232

can be reassigned to create enough space to include un-assigned tasks in the task list of 233

any robot. This objective cannot be achieved with the existing PI cost mechanism without 234

devising a mechanism to lower the cost of already assigned tasks to be included in other 235

robot task lists. 236

The task reassignment for maximum task allocation (TRMaxAlloc) algorithm maxi- 237

mizes the task allocation in two phases, i.e., in the first phase, task allocation is done using 238

the PI algorithm until no more tasks can be added or the limit of task allocation is reached. 239
Version August 24, 2022 submitted to Journal Not Specified 9 of 20

Algorithm 3: Maximum Task Allocation Algorithm

1 Initialize
2 get Consensus & PI Consensus is False
3 while Consensus is False do
4 if PI consensus is False then
5 Task Inclusion PI Algorithm
6 end
7 else
8 Task Inclusion TRMaxAlloc Algorithm
9 end
10 Conflict and Communication Phase
11 if Consensus & PI Consensus is True then
12 break;
13 end
14 end
15 return
16 Status ⇒ PI Consensus
17 Status ⇒ Consensus

In the second phase, if there are still unallocated tasks, it reassigns tasks to other robots to 240

create space for unallocated tasks. TRMaxAlloc is based on two main concepts: first is the 241

selection of an appropriate task which is already assigned to some other robot, and second 242

is the task costing mechanism. The algorithm of the TRMaxAlloc for task inclusion is given 243

as algorithm 1:- 244

4.1. TRMaxAlloc Task Costing Mechanism (TRMaxAlloc) 245

In the TRMaxAlloc algorithm, the concept of transformed IPI ℘i∧

of tasks is introduced. 246

In the second phase of the TRMaxAlloc algorithm, first find the list of tasks which can be 247

included in si by computing the IPI of each task using equation (6). There are two types of 248

tasks in the IPI list: those that have already been assigned to other robots and those that 249

have yet to be assigned. As per the PI algorithm, unassigned tasks cannot be included in si 250

as the system has already achieved its consensus and tasks that are assigned to some other 251

robots are already at their minimum RPI value. So, any task assignment on the basis of 252

computed IPI is not possible. The transformed IPI ℘iq ∧ is the performance impact of tasks 253

but is computed differently for unassigned tasks and already assigned tasks. To compute 254

Tr IPI, the first IPI of tasks is computed as follows:- 255

⊕
(
∧
℘iq − ℘iq /℘iq if task is already assigned
℘iq = ⊕
(10)
℘iq if task is unassigned

For unassigned tasks, the value of IPI and Tr IPI is the same. For an assigned task, the 256

Tr IPI, value is computed and from equation (10), it is clear that the Tr IPI, of an assigned 257

task is lower than its RPI of a task. Similarly, transformed RPI ℘i∼ is computed for the 258

communication and conflict resolution phases. To compute the Tr RPI, the first RPI of tasks 259

is computed and for unassigned tasks, the RPI and Tr RPI values are the same, whereas for 260

assigned tasks, the Tr RPI value is computed as:- 261

(
∼ ℘iq − v⊖ q ( si , tq ) / ℘iq if task is already assigned
℘i = ⊖
(11)
v q ( si , t q ) if task is unassigned
By using equation (11), a lower value of Tr RPI of a task than the RPI value results 262

in task assignment to the robot with the lowest value of Tr RPI. In this way, a vacancy 263

is created for the possible inclusion of an unassigned task. The TRMaxAlloc algorithm 264

includes tasks using the following steps: 265

Version August 24, 2022 submitted to Journal Not Specified 10 of 20

(i) The first step in the TRMaxAlloc algorithm is to find the list of tasks which can 266

be added to the task list of any robot such that all the constraints mentioned are 267

satisfied. 268

(ii) The second step is to select the tasks to include in the task list. Task selection can 269

be made using one of the following methods: 270

• Task selection (TS) based on maximum cost reduction (TSMCR) 271

• Task selection based on minimum distance (TSMD) 272

• Task selection based on the distributed bee algorithm (TSDBA) 273

4.1.1. TS based on Maximum Cost Reduction 274

The value of κ is calculated for assigned and unassigned tasks present in the IPI 275

list. In the event that the IPI list contains unassigned tasks, the task with the 276

maximum value of κ is selected. On the other hand, if there is no unassigned 277

task in the IPI list, the assigned task with the maximum value of κ is selected for 278

inclusion as in equation (12): 279

∼ ∧
κ = maxqM=1 (℘iq − ℘iq ) (12)

4.1.2. TS with a Minimum Distance 280

If there are unassigned tasks in the IPI list, then the unassigned task with the 281

minimum Tr IPI value is selected for inclusion in the task list. If there is no 282

unassigned task in the IPI list, then a task is selected on the basis of its distance 283

from an unassigned task. By selecting a task tq with a minimum distance from an 284

unassigned task, there is a greater chance that the unassigned task will be allocated 285

to the robot in place of task tq . 286

4.1.3. TS with Distributed Bee Algorithm 287

Task selection is based on the concept of utility, which considers both the cost and
the quality of a target. Utility is the probability that robot i selects a task tq to
include in si from the Tr IPI list. The cost is defined as Tr IPI of a task tq and the
inverse of Tr IPI of a task is defined as visibility:-
∧
λi,q = 1/℘iq (13)

Quality is represented as the priority of a target in a search and rescue scenario in

equation (14). Each target is associated with a task deadline, so the inverse of the
deadline is taken as the priority of a task. The closer a task’s deadline, the higher
its priority, and vice versa.
T
ζ k = dk / ∑ d j (14)
j =1

where, T is the number of tasks. The utility or probability to select a task tq from
Tr IPI list is given as:-
T
α ω
P(si , tq ) = ζ tq λtq / ∑ ζ αjq λω
jk (15)
j =1

where T is the number of tasks, which can be included in the task list si . α and ω 288

are control parameters that bias the decision either towards the task deadline. In 289

the Tr IPI list, there are two types of tasks: those that are already assigned to some 290

robot and those that are unassigned. As a result, it may return different results for 291

assigned and unassigned tasks . From probability distribution, tasks are selected 292

based on the roulette wheel rule. 293

294

(iii) Update the time cost and robot ID vectors. 295

Version August 24, 2022 submitted to Journal Not Specified 11 of 20

(iv) Update the RPI vector when the task limit is reached or no more tasks can be 296

added. 297

4.2. Task Inclusion Limit 298

It is necessary in time-critical scenarios that a task allocation system converge to a 299

consensus within certain time limit. Due to the IPI and RPI lowering mechanisms, tasks 300

keep on switching between robots. In order to avoid the system getting stuck in such 301

task switching, a generalized task inclusion limit (TIL) is defined, which restricts a task to 302

including a robot task list more than a defined limit. A task inclusion vector (TIV) equal to 303

the number of tasks is defined at the start of the simulation for each robot. At the start of 304

the simulation, all values in (TIV) are initialized to zero. Each time a task is included in the 305

task list of a robot ri , the value corresponding to that task tq in (TIV) is increased by 1. Once 306

the limit of inclusion for a task tq against (TIL) is reached for a robot ri , it cannot include tq 307

again in its task list. By setting a high value of (TIL), communication iterations increase 308

significantly. The optimum number of tasks included against the number of communication 309

iterations is TIL = 2.

Figure 5. Proposed Task Inclusion

310

Figure 5 shows all tasks are allocated using our proposed technique with three task 311

selection mechanisms. During phase-2 of our proposed algorithm, r1 , neither t2 nor t3 312

can be included in s1 because t1 is already included in its task list and adding any task 313

before t1 will violate the threshold condition. Similarly, adding t2 or t3 after t1 also does not 314

conform to their threshold requirements. Whereas, r2 can include t1 before t2 and satisfy 315

all constraints. 316

In the second step, since t1 is the only task which r2 can include in its task list, t1 317

is selected for inclusion in s2 by all three task selection mechanisms. The Tr IPI of t1 is 318

calculated and compared with its global RPI value. As the Tr IPI of t1 for r2 is less than its 319

global RPI value, t1 is included in the task list of r2 during the communication and conflict 320

resolution phase. 321

Accordingly, the robot ID vector is also updated because there is still an unassigned 322

task, t3 , the entire procedure will be repeated. In second iteration, as task list of r1 is empty, 323

all tasks t1 , t2 and t3 can be included in its task list. Since t3 is an unassigned task, it will 324

be selected using TSM, TrIPI and TSMDUT. On the other hand, task selection based on 325

TSDBA is done using a probabilistic function. However α and ω are adjusted in such a way 326

that unassigned tasks get higher probabilities. In doing so, after some task reassignment 327

iterations, r1 will select task t3 . In this way, all tasks are assigned to the robots. The task 328

allocation algorithm is given in algorithm 2. 329

5. Simulation Results 330

In this section, the simulation results of proposed task maximization approaches are 331

compared with the baseline PI approach. The algorithms are implemented in MATLAB 332
Version August 24, 2022 submitted to Journal Not Specified 12 of 20

Algorithm 4: TRMaxAlloc Maximum Task Inclsuion Algorithm

1 Initialize
2 while |si | ≤ L do
3 Compute Tr IPI i.e., ℘i∧
4 if Candidate Tasks → Inclusion then
5 if Task Selection ∈ TSMCR then
6 if Tr IPI → Unassigned Tasks then
7 Select →
8 Unassigned Task with minimum Tr IPI
9 end
10 else
11 Select →
12 Assigned Task with minimum Tr IPI
13 end
14 end
15 if Task Selection ∈ TSMDUT then
16 if Tr IPI → Unassigned Tasks then
17 Select →
18 Unassigned Task with minimum Tr IPI
19 end
20 else
21 Select →
22 Assigned Task with Minimum Distance
23 end
24 end
25 if Task Selection ∈ TSBDA then
26 Select Task→
27 Roulette Wheel Rule
28 end
29 Update Robot ID Vector β i
30 Update Time Cost ci (s j )
31 Update Task List si
32 else
33 break
34 end
35 end
36 end
37 return
38 Update → Tr IPI
39 Update → RPI

with an Intel Core i5-4310 and a CPU running at 2.0 GHz. The extensive results show the 333

effectiveness of the proposed algorithm. 334

The scenario used for the proposed method is the search and rescue scenario as 335

described in [11]-[12]. A disaster condition is considered when survivors are randomly 336

distributed in a defined area and heterogeneous UAVs are required to provide support 337

services, i.e. food and medicine, to the survivors. The speed of UAVs for providing food is 338

50 m/s and for medicine it is 30 m/s, and the velocities of UAVs are considered constant 339

during the whole operation. It takes 100s to provide food or medicine to the survivor, and 340

each UAV can serve a maximum number of survivors. 341

The tasks are bounded in 3D space by 10, 000 m × 10, 000 m × 1000 m, where tasks 342

are randomly generated in this area. The UAVs are placed randomly in a 10, 000 m x 343

10, 000 m × 0 m area. The deadline of the tasks is randomly selected between [0,2000s]. 344

Failure to comply with the deadline of any task is considered a failure. Network topologies 345

considered in the simulations are row topology, circular topology, mesh topology, and star 346

topology as shown in figure 6. For each set of parameters and scenario, fifty simulations 347

are performed. The simulation parameters are shown in table 2. The PI algorithm and 348
Version August 24, 2022 submitted to Journal Not Specified 13 of 20

(a) Mesh Topology

(b) Star Topology

(c) Circular Topology

(d) Row Topology

Figure 6. Network Topologies

the proposed TRMaxAlloc algorithm with three task selection schemes are compared for 349

the following four graph topologies: mesh, star, circular, and row. The basis for using four 350

graph topologies is to study graph topology effects on the number of allocated tasks and 351

average cost of baseline and proposed technique. 352

5.1. Simulation Scenario 1: Task Allocation 353

In this scenario, PI and the proposed TRMaxAlloc algorithm are compared for fixed 354

10 robots and 64 tasks for four different topologies: mesh, star, circular, and row. In the 355

TRMaxAlloc algorithm, tasks are allocated using three task selection techniques: task 356

selection using the distributed bee algorithm (TSDBA), minimum cost reduction (TSMCR), 357

and minimum distance from unassigned task (TSMDUT). The task inclusion limit is set to 358

2 for all topologies. 359

Table 2. Simulation Parameters

Medicine Supply Food Supply
Speed 30 m/s 50m/s
Task duration 100 m/s 100 m/s
Task deadline 0-2000 s 0-2000 s
Task area 10, 000 m × 10, 000 m × 1000 m
Version August 24, 2022 submitted to Journal Not Specified 14 of 20

In TRMaxAlloc with TSDBA, α and ω parameters are used to bias the decision either 360

toward cost or quality of a solution. In order to find the values of α and ω where the 361

performance of the TSDBA task selection method is best, keep one parameter fixed ω, 362

i.e., and change the value of α. The value of ω set as 1 and varied from 0.4 to 1.8 with 363

an increment of 0.2. A total of eight simulations, i.e., one for each value of α with each 364

simulation containing 50 iterations is performed. 365

The minimum increase in average percentage of tasks is achieved when ω=α , and get 366

the maximum value when α = 1.6. Since the difference in average percentage increase at all 367

values of α is not very significant, and similarly, the number of iterations to assign these 368

tasks is comparable except at α=1.8, which is slightly higher. On this basis, the value of α 369

and ω parameters are taken as 1.6 and 1. 370

In figure 7, the average number of tasks allocated by PI and TRMaxAlloc algorithms 371

for four topologies is shown, and figure 8 gives the corresponding average percentage 372

increase in tasks as compared to the PI algorithm. Figure 9 gives the average cost incurred to 373

perform these tasks.From figure 7, it is clear that PI allocates the maximum number of tasks 374

in mesh topology, which is due to the reason that all robots simultaneously communicate 375

with each other and the task is selected with the minimum increase in overall cost. Due 376

to the distributed positions of robots and tasks and in other topologies, robots do not 377

communicate simultaneously. Robots can allocate tasks, which may increase the cost. For 378

similar reasons, the number of allocated tasks by the PI algorithm in star, row, and circular 379

topologies is equal to or less than the mesh topology. In figures 7 and 8, the TRMaxAlloc 380

algorithm with TSMDUT task selections allocates the maximum number of average tasks 381

with a maximum percentage average increase in tasks for all topologies. 382

Figure 7. Average Allocated Tasks

Table 3 shows the maximum percentage increase in tasks during 50 iterations for the 383

TRMaxAlloc algorithm. The maximum percentage increase in tasks with task selection 384

method TSDBA is 22.41%, for TSMCR it is 19.64% and for TSMDUT it is 21.05% which 385

gives a very clear improvement in the number of allocated tasks. 386

Version August 24, 2022 submitted to Journal Not Specified 15 of 20

Figure 8. Average % Increase of Tasks

Figure 9. Average Cost of Tasks

Version August 24, 2022 submitted to Journal Not Specified 16 of 20

Table 3. Compatrison of TRMaxAlloc with PI Algorithm

Mesh Topology Star Topology Circular Topology Row Topology
A1 A2 A3 A1 A2 A3 A1 A2 A3 A1 A2 A3
PI 48 49 49 51 47 51 45 45 45 46 45 46
TRMaxAlloc 56 57 58 60 55 59 58 56 57 54 55 55
% Improvement 13.12 14.04 15.52 16.36 14.55 13.56 22.41 19.64 21.05 13.56 18.18 16.36
1 A1=TSDBA; 2 A2=TSMCR; 3 A3=TSMDUT

5.2. Simulation Scenario II: : Effect on Task Allocation with Increasing Task 387

In this scenario, the performance of the PI and TRMaxAlloc algorithms with different 388

numbers of robots and tasks and their effects on average allocated tasks and maximum 389

increase in tasks is analyzed. The number of robots is 6, 8, 10, 12, and 14 and the tasks are 390

multiple of the number of robots, i.e., 2, 4, is chosen and the rest of the parameters are the 391

same as for simulation scenario I. 392

It is evident from Table 4 that for task to robot ratio 4, PI gives an increasing trend 393

for percentage average allocation. A similar is trend is observed for TRMaxAlloc with 394

three task selection methods. However TRMaxAlloc with TSMDUT task selection method 395

gives better results in terms of percentage average allocation. When task to robot ratio is 2, 396

percentage average allocated task decreases from 94.2% to 87.9% at 14 number of robots. A 397

similar 15.8% percentage maximum increase in allocated tasks is achieved with increase of 398

3 tasks for 10 robots and TSDBA, TSMDUT methods. However, when task to robot ratio 399

is 4, 18.2% percentage maximum increase in allocated tasks is achieved with increase of 3 400

tasks for 6 robots and TSMDUT method. 401

Table 4. Comparison of Average Task Allocation with increasing No. of Tasks

No of Robots 6 8 10 12 14 6 8 10 12 14
Multiple of 2 Multiple of 4
No of Tasks
12 16 20 24 28 24 32 40 48 56
PI Avg. Alloc. (%) 94.2 93.1 94 93.3 87.9 86.3 97.8 88.5 90 91.3
Avg. Alloc. (%) 95 94.4 95.5 95.4 90.4 90.4 92.8 93 94.6 95.5
TSDBA
Max Increase(%) 10 15.4 15.8 9.1 12.5 13 12.5 13.5 15.9 11.5
Assigned
Avg. Alloc. (%) 95 94.4 95 95 90.4 89.6 92.2 92.8 94.2 95.4
Tasks TRMaxAlloc TSMCR
Max Increase(%) 10 15.4 11.1 9.1 12.5 13 10 13.5 15.9 9.1
Avg. Alloc. (%) 95 94.4 95.5 95 90.7 91.3 93.4 93.8 95.2 95.9
TSMDUT
Max Increase(%) 10 15.4 15.8 9.1 12.5 18.2 12.5 15.8 15.9 11.3

The average cost incurred to perform the allocated task is lower for the PI algorithm 402

than for the TRMaxAlloc algorithm with three task selection methods. The reason behind 403

the increased average cost is that tasks allocated by PI algorithm are based on minimum 404

path cost over all robots. However, TRMaxAlloc task allocation is based on maximum 405

assignment of tasks rather than minimum path cost over all robots, which results in an 406

increased average path cost as given in Table 5. 407

5.3. Simulation Scenario III: Task Inclusion Limit Selection 408

In the PI algorithm, a task is assigned to a robot when the cost incurred by it is less 409

than the corresponding RPI value. When the task allocation process by the PI algorithm 410

achieves its convergence, the RPI value of each task is at its minimum. In order to introduce 411

task reassignment, the IPI of the task is required to be less than the achieved minimum. 412

TRMaxAlloc proposes the concept of transformed IPI i.e. TrIPI and transformed RPI i.e. 413

TrRPI, which provide a method to reduce the task cost to less than the achieved minimum. 414

But this cost-lowering mechanism introduces an infinite state of task reassignment, which 415

has to be stopped at an appropriate level. 416

In an infinite state of task reassignment, a robot repeatedly includes a particular task 417

in its task list. In order to stop this recursive task inclusion, the task inclusion limit i.e. TIL 418

parameter is defined in TRMaxAlloc, which stops a task being included in the task list of 419
Version August 24, 2022 submitted to Journal Not Specified 17 of 20

Table 5. Average Cost with increasing No. of Tasks

No of Robots 6 8 10 12 14
No of Tasks (x of 2) 12 16 20 24 28
PI Algorithm 170.3 167.9 160.4 158.8 194.9
TSDBA 194.9 209.7 209 218.2 256.9
TRMaxAlloc TSMCR 199.3 208.5 206.6 208.4 260.2
TSMDUT 197.2 205.1 201.7 204.7 250.7
No of Robots 6 8 10 12 14
No of Tasks (x of 4) 12 16 20 24 28
PI Algorithm 167 159.5 157.3 151.8 150.2
TSDBA 209.1 209 207.6 204.8 206.6
TRMaxAlloc TSMCR 204.8 202.9 200.1 197.4 197.6
TSMDUT 210.8 207 206.7 202.1 202

a particular robot after a certain number of inclusions. TIL is defined as the maximum 420

number of times a task can be included in the task list of a robot. In order to find the 421

appropriate TIL value, we performed simulations with a total of 10 numbers of robots and 422

64 tasks using mesh topology are shown in figure 11-12.

Figure 10. Average Allocated Tasks vs TIL

423

6. Conclusion 424

This paper introduces task reassignment capability in the PI algorithm on the basis of 425

a modified costing mechanism and three task selection methods for task inclusion in the 426

task list. The simulations show that a maximum of 22.41% more tasks are allocated than 427

the baseline PI algorithm, which is a significant improvement in time-critical applications. 428

For search and rescue scenarios, if a greater number of tasks are executed within resource 429

limits, then task cost can be compromised. By analysing the results of three task selection 430

methods, it is evident that the selection of a task for inclusion in a task list is pivotal to 431

bringing down the number of iterations. In future work, the task selection method will be 432

improved to reduce the number of iterations while maintaining the solution quality. 433

References 434

1. Wu, X.; Gao, Z.; Yuan, S.; Hu, Q.; Dang, Z. A Dynamic Task Allocation Algorithm for Heteroge- 435

neous UUV Swarms. Sensors 2022, 22, 2122. 436

Version August 24, 2022 submitted to Journal Not Specified 18 of 20

Figure 11. Average % Increase of Tasks vs TIL

Figure 12. Average Cost of Tasks vs TIL

Version August 24, 2022 submitted to Journal Not Specified 19 of 20

2. Drew, D.S. Multi-agent systems for search and rescue applications. Current Robotics Reports 437

2021, 2, 189–200. 438

3. Rishwaraj, G.; Ponnambalam, S. Integrated trust based control system for multirobot systems: 439

development and experimentation in real environment. Expert Systems with Applications 2017, 440

86, 177–189. 441

4. Seenu, N.; RM, K.C.; Ramya, M.; Janardhanan, M.N. Review on state-of-the-art dynamic task 442

allocation strategies for multiple-robot systems. Industrial Robot: the international journal of 443

robotics research and application 2020. 444

5. Calvo, A.; Silano, G.; Capitán, J. Mission Planning and Execution in Heterogeneous Teams of 445

Aerial Robots supporting Power Line Inspection Operations. In Proceedings of the International 446

Conference on Unmanned Aircraft Systems, 2022. 447

6. Liu, S.; Kurniawan, E.; Tan, P.H.; Zhang, P.; Sun, S.; Ye, S. Dynamic scheduling for heterogeneous 448

resources with time windows and precedence relation. In Proceedings of the TENCON 2017- 449

2017 IEEE Region 10 Conference. IEEE, 2017, pp. 3045–3050. 450

7. Xie, T.; Guo, J.; Zhang, X.; Yu, J.; et al. Mathematical Problems in Engineering Improved 451

CNP-Method-Based Local Real-Time Cooperative Task Allocation of Heterogeneous Multi-UAV 452

in Communication-Constrained Environment. Mathematical Problems in Engineering 2022, 2022. 453

8. Yang, M.; Bi, W.; Zhang, A.; Gao, F. A distributed task reassignment method in dynamic 454

environment for multi-UAV system. Applied Intelligence 2022, 52, 1582–1601. 455

9. Wang, Z.; Liu, C.; Gombolay, M. Heterogeneous graph attention networks for scalable multi- 456

robot scheduling with temporospatial constraints. Autonomous Robots 2022, 46, 249–268. 457

10. Korsah, G.A.; Stentz, A.; Dias, M.B. A comprehensive taxonomy for multi-robot task allocation. 458

The International Journal of Robotics Research 2013, 32, 1495–1512. 459

11. Zhao, W.; Meng, Q.; Chung, P.W. A heuristic distributed task allocation method for multivehicle 460

multitask problems and its application to search and rescue scenario. IEEE transactions on 461

cybernetics 2015, 46, 902–915. 462

12. Turner, J.; Meng, Q.; Schaefer, G.; Whitbrook, A.; Soltoggio, A. Distributed task rescheduling 463

with time constraints for the optimization of total task allocations in a multirobot system. IEEE 464

transactions on cybernetics 2017, 48, 2583–2597. 465

13. Whitbrook, A.; Meng, Q.; Chung, P.W. Reliable, distributed scheduling and rescheduling for 466

time-critical, multiagent systems. IEEE Transactions on Automation Science and Engineering 2017, 467

15, 732–747. 468

14. Trigui, S.; Cheikhrouhou, O.; Koubaa, A.; Baroudi, U.; Youssef, H. FL-MTSP: a fuzzy logic 469

approach to solve the multi-objective multiple traveling salesman problem for multi-robot 470

systems. Soft Computing 2017, 21, 7351–7362. 471

15. Arif, M.U.; Haider, S. An Evolutionary Traveling Salesman Approach for Multi-Robot Task 472

Allocation. In Proceedings of the ICAART (2), 2017, pp. 567–574. 473

16. Johnson, L.B.; Choi, H.L.; How, J.P. The role of information assumptions in decentralized task 474

allocation: A tutorial. IEEE Control Systems Magazine 2016, 36, 45–58. 475

17. Semwal, T.; Jha, S.S.; Nair, S.B. On ordering multi-robot task executions within a cyber physical 476

system. ACM Transactions on Autonomous and Adaptive Systems (TAAS) 2017, 12, 1–27. 477

18. Al-Yafi, K.; Lee, H. Centralized versus market-based approaches to mobile task allocation 478

problem: State-of-the-art 2009. 479

19. Li, D.; Fan, Q.; Dai, X. Research status of multi-robot systems task allocation and uncertainty 480

treatment. In Proceedings of the Journal of Physics: Conference Series. IOP Publishing, 2017, 481

Vol. 887, p. 012081. 482

20. Hwang, N.E.; Kim, H.J.; Kim, J.G. Centralized Task Allocation and Alignment Based on 483

Constraint Table and Alignment Rules. Applied Sciences 2022, 12, 6780. 484

21. Skaltsis, G.M.; Shin, H.S.; Tsourdos, A. A survey of task allocation techniques in MAS. In 485

Proceedings of the 2021 International Conference on Unmanned Aircraft Systems (ICUAS). 486

IEEE, 2021, pp. 488–497. 487

22. Sung, Y.; Budhiraja, A.K.; Williams, R.K.; Tokekar, P. Distributed simultaneous action and target 488

assignment for multi-robot multi-target tracking. In Proceedings of the 2018 IEEE International 489

conference on robotics and automation (ICRA). IEEE, 2018, pp. 3724–3729. 490

23. Lerman, K.; Jones, C.; Galstyan, A.; Matarić, M.J. Analysis of dynamic task allocation in 491

multi-robot systems. The International Journal of Robotics Research 2006, 25, 225–241. 492

24. Luo, L.; Chakraborty, N.; Sycara, K. Provably-good distributed algorithm for constrained 493

multi-robot task assignment for grouped tasks. IEEE Transactions on Robotics 2014, 31, 19–30. 494
Version August 24, 2022 submitted to Journal Not Specified 20 of 20

25. Choi, H.L.; Brunet, L.; How, J.P. Consensus-based decentralized auctions for robust task 495

allocation. IEEE transactions on robotics 2009, 25, 912–926. 496

26. Zitouni, F.; Harous, S.; Maamri, R. A distributed approach to the multi-robot task allocation 497

problem using the consensus-based bundle algorithm and ant colony system. IEEE Access 2020, 498

8, 27479–27494. 499

27. Geng, N.; Meng, Q.; Gong, D.; Chung, P.W. How good are distributed allocation algorithms for 500

solving urban search and rescue problems? A comparative study with centralized algorithms. 501

IEEE Transactions on Automation Science and Engineering 2018, 16, 478–485. 502

28. Zhang, A.; Yang, M.; Bi, W.; Gao, F. Distributed task allocation with critical tasks and limited 503

capacity. Robotica 2021, 39, 2008–2032. 504

29. Senanayake, M.; Senthooran, I.; Barca, J.C.; Chung, H.; Kamruzzaman, J.; Murshed, M. Search 505

and tracking algorithms for swarms of robots: A survey. Robotics and Autonomous Systems 2016, 506

75, 422–434. 507