Professional Documents
Culture Documents
article info a b s t r a c t
Article history: Research in multi-robot systems is a rich field that has attracted much attention in recent decades.
Received 6 November 2020 However, robot coordination and task allocation to a correct mission accomplishment are still
Received in revised form 23 June 2021 challenging even with technological advances. Despite many proposals presented in the literature,
Accepted 24 June 2021
the applications and theories about the task allocation problem are not yet exhausted. Thus, this
Available online 26 June 2021
work proposes an axiomatic framework based on Social Choice Theory to analyze the task allocation
Keywords: problem in intentional cooperation multi-robot systems. It uses Kenneth J. Arrow’s framework of his
Multi-robot system famous Impossibility Theorem. The conditions imposed by Arrow aim to create an ideal for preference
Task allocation aggregation mechanisms through axiomatic analysis. This paper aims to transport this analysis to
Social choice the multi-robot domain. A behavior-based Multi-robot Task Allocation architecture is used to present
Arrow’s theorem simulation results and discuss two cases in the ordinal preference domain. The analysis results show
Preference aggregation that using the proposed framework to analyze, under the Arrovian perspective, implemented MRTA
architectures is feasible.
© 2021 Elsevier B.V. All rights reserved.
https://doi.org/10.1016/j.robot.2021.103839
0921-8890/© 2021 Elsevier B.V. All rights reserved.
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
are as many as the agents involved, i.e., each agent may have its 4.2. Arrovian conditions for the MRTA problem
own utility function, no ordinary meaning can be attributed to
utility levels or utility differences [49]. Thus, under the principles The present section defines analogous axioms and conditions
of ordinalism and non-comparability, robots’ individual prefer- for the robot society in the multi-robot domain considering Ar-
ences define society preference directly without generality loss. row’s axiomatization [6]. The number m of robots of the society is
On the other hand, scalar utility values allow an explicit search at least three, i.e., m ≥ 3. Besides, the number n of tasks presented
for an objective function maximization. to individuals is at least three, i.e., n ≥ 3. Such considerations
Thus, achieving social welfare in the task allocation problem is are essential so that theorem definitions are still valid. If n < 3,
to obtain a preference ordering that respects Completeness and the problem solution is trivial as the Arrow’s impossibility cannot
arise.
Transitivity Axioms, defined in Section 4.2, because, according
The Completeness Axiom 1 defines that the robots of society
to [50], ‘‘as with any type of behavior described by maximization,
must order the available alternatives according to their individual
the measurability of social welfare need not be assumed; all that
preference. With ties being allowed. We emphasize that prefer-
matters is the existence of a social ordering satisfying [those]
ence ordering considers the internal utility calculation performed
Axioms’’. by the robots. However, team members do not compare the
In this paper, the objective functions in which the maximiza- internal values with other robots.
tion is sought are called Arrovian social welfare and social choice
functions. The functions must satisfy a set of conditions similar Axiom 1 (Completeness Axiom for the MRTA Problem). Society must
to those established for Arrow’s Impossibility Theorem. The next order the given tasks, with {τ1 , τ2 } ∈ T , considering (τ1 ⪰ τ2 ) or
section presents these conditions translated into the domain of (τ1 ≻ τ2 ) or (τ2 ≻ τ1 ) or (τ1 ∼ τ2 ).
the MRTA problem.
Moreover, decisions must be rational. Thus, individuals pref-
erences ordering must be transitive, following the Transitivity
4.1. Social choice formal model of MRTA problems Axiom for the MRTA problem.
The word society means a group of individuals who live to- Axiom 2 (Transitivity Axiom for the MRTA Problem). Given a pref-
gether in an organized way, making decisions about doing things erence ordering over tasks, with {τ1 , τ2 , τ3 } ∈ T , if (τ1 ⪰ τ2 ) and
(τ2 ⪰ τ3 ), then (τ1 ⪰ τ3 ).
and sharing the work that needs to be done. Then, the term is
appropriate to denote a group of robots programmed to express The Transitivity Axiom is a reasonable requirement to avoid a
cooperative behavior on executing a mission task. Thus, R is the cyclic ranking. If an individual preference is non-transitive, then
set of robots, mobile or not, that compose society, subjecting itself the relationship between the alternatives may be as follows: x1 ⪰
to the same task allocation architecture. x2 , x2 ⪰ x3 , but x3 ⪰ x1 . Which is not a rational behavior [51].
Additionally, the paper considers only loosely coupled tasks The Unrestricted Domain condition imposes that a social or-
whose interdependence in execution is nil. This consideration dering results from the individual preference orderings without
simplifies the model. So, consider T the set of possibly weighted preference restrictions. In MRTA problems, this condition es-
independent tasks, requiring one or more robots to perform each tablishes that the preference relation can rank the alternatives
task. unrestrictedly, according to individual preference. However, these
Consider the binary relation ⪰. It is called weak preference orderings are weak since they do not exclude indifference re-
meaning ‘‘at least as preferable as’’ or ‘‘better than or equal in lations and encompass all the logically possible rankings. This
observation is essential when a robot is not able to perform two
value to’’. For robot ri , with i = (1, 2, . . . , m), if the execution of
or more tasks. In this case, if a robot cannot execute two or more
task τ1 is as preferable as the execution of task τ2 , then τ1 ⪰i τ2 .
tasks, its degree of preference over the execution of these tasks
Using it as a primitive relation, one can define the strict
must be the lowest. Moreover, the order of execution between
preference relation ≻ and indifference relation ∼. The execution
them must be indifferent.
of task τ1 is strictly preferable to running task τ2 , with notation
τ1 ≻i τ2 , if and only if (τ1 ⪰i τ2 ) ∧ ¬(τ2 ⪰i τ1 ). And if the Condition 1 (Unrestricted Domain). Regardless of the individual
agent ri is indifferent to the execution of both tasks, the notation preference ordering performed over T , there is a social ordering
τ1 ∼i τ2 is used if and only if (τ1 ⪰i τ2 ) ∧ (τ2 ⪰i τ1 ). Using the two resulting from individuals preferences aggregation.
previous established binary relations, the weak preference can be
denote by the reciprocity (τ1 ⪰i τ2 ) ↔ [(τ1 ≻i τ2 ) ∨ (τ1 ∼i τ2 )], Returning to social choice theory, the Independence of Irrel-
contrasting with the indifference relation. evant Alternatives (IIA) condition states that the judgment of
Still, considering only transitive preferences, the preference preferences should be pairwise, regardless of a third alternative.
The Independence of Irrelevant Tasks condition (Condition 2)
orderings will be considered non-strict or weak. The weak pref-
refers to the task ordering. Thus, the preference between the two
erence ordering allows ties, i.e., the robot is indifferent to the
tasks depends only on the binary relation between these two
execution order of two tasks when: (i) the calculated utility (cost,
tasks. Also, it is independent of the preference information of a
fitness, score) value for both tasks is equal, or (ii) it is unable to
third alternative. This ordering is also independent of the irrel-
perform the tasks, for instance. In doing so, ⪰r i = (τ1 ⪰ τ2 ⪰ evant alternative utility information. The term irrelevant in this
... ⪰ τn ) denotes robot ri preferences ordering. Likewise, the case means that the third task is unimportant to the preference
non-strict ordering admits ties in a robot preference over robots, ordering between any other two tasks and not that the task is
following the same analysis. irrelevant to the mission.
The set R denotes the set of individual non-strict possible
rational preferences. The set of preference orderings, here also Condition 2 (Independence of Irrelevant Tasks). Given {τ1 , τ2 , τ3 } ∈
called the preference profile, is denoted by [⪰]R ∈ Rn , represent- T , the social ordering take the tasks τ1 and τ2 ranking pairwise,
ing robots individual preference orderings concerning to the task despite the third alternative. And the same holds for the other
execution. pairwise preferences over {τ1 , τ2 , τ3 }.
4
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
Together with Axioms 1 and 2, the welfare maximization is preference ordering ⪰ri related to tasks execution order (if robot
attained with the Pareto Principle or Unanimity Principle. Before ri is not capable of performing a task, then this task is the least
defining the Unanimity Condition for MRTA problems, let us preferred). Then, the goal is to assign robots to tasks in order to
redefine Pareto-dominance and Pareto Optimality concepts for maximize social welfare, i.e., satisfy the Axioms 1, and 2; and the
the MRS domain. Consider uri (x) a utility measurement of robot Conditions 1, 2, 3, and 4, resulting a social preference ordering
ri individual welfare over its preference profile. [⪰]R that aggregates individual preferences.
Definition 3 (Dominance in the MRTA Problem). A social welfare The existence conditions of the welfare functions defined in
function is said to be dominant, or dictatorial, if there exists an Section 4.2 are based on that defined by [6,52,53] for the exis-
individual ri such that, for all τ1 and τ2 , (τ1 ≻ri τ2 ) implies that tence of a social welfare and social choice functions, respectively.
(τ1 ≻R τ2 ), independently of the orderings {⪰r 1 , . . . , ⪰r i } of the
other robots, where ≻R is the social strict preference relation 5. Arrovian analysis of MRTA problem taxonomy
corresponding to {⪰r 1 , . . . , ⪰r m }.
The characteristic of Definition 3 is not beneficial to a pref- The analysis of task allocation problems that Gerkey and
erence aggregation seeking social welfare maximization. Further- Matarić’s taxonomy [5] defines will be done individually, under
more, since the paper focuses on decentralized multi-robot archi- each characteristic axis, i.e., robot, task, and allocation types. An
tectures, a dictatorship perverts the decentralized characteristic initial structure was proposed by dos Reis and Bastos [3], which
by centralizing the preferences and, consequently, the decisions. this paper expands. Some assumptions are considered for this
Therefore, Condition 4 states that the social welfare function analysis:
should not allow a single member to govern the task allocation
i. All the robots tell the truth and are not envious.
based on their individual preferences.
ii. The number of tasks is greater than or equal to the number
of robots, so n ≥ m and n ≥ 3.
Condition 4 (Non-dictatorship). A social welfare function should not
be dominant or dictatorial. iii. All mission tasks are known in advance to the task alloca-
tion process. Even if the mission is not previously known,
Given the existence conditions for social choice, Problem 1 the robots are aware of the resources and capabilities re-
revises the multi-robot task allocation problem from the Arrovian quired to perform the tasks.
viewpoint. iv. All orderings are rational, i.e., complete and transitive.
v. Robots must order (strictly) all available alternatives.
Problem 1 (Arrovian MRTA Problem). Given a society R of m vi. If a robot cannot perform a task, then that task must have
robots, each one capable of performing one or more tasks si- the least preference in this robot preference ordering.
multaneously, and the set T of n tasks, requiring one or more vii. Robots pursue to maximize social welfare, not exclusively
robots to perform each task. Also, given for each robot ri an weak individual welfare.
viii. Robots can calculate their utility for each task. From this
1 Adapted from Shoham and Leyton-Brown [51]. value, they establish their preference ordering.
5
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
ix. Robots have no external strategies or incentives to lie about Bade [58] states that the mechanism is Pareto-optimal be-
their intentions or utility functions and, so, about their cause it maps each Pareto-optimal allocation profile to a Pareto-
preference ordering, as well as they are embedded in a optimal allocation in this profile and is strategy-proof since there
cooperative and non-competitive environment. is no preference profile or agent such that a false preference
communication benefits it. However, the order in which the
The following analysis considers the following conditions are
agents are processed influences the final result (a path-dependent
satisfied to all situations [3]: mechanism). Abdulkadiroğlu and Sönmez [56] propose the Ran-
• Unrestricted Domain: Considering the task independence, the dom Serial Dictatorship (RSD) allocation mechanism to avoid the
preference ordering over these alternatives has any transi- problem of sorting agents. The mechanism establishes a uniform
tive and complete possible form. probability distribution that orders agents. Thus agent makes its
• Non-dictatorship: There is no robot ri whose preferences are choice according to its individual preferences.
dominant related to the other’s preferences. Consider a uniform probability 1/m that sequences the m
• Independence of irrelevant tasks: The preference ordering of choosing robots, i.e., the probability of a first robot to choose its
each robot is pairwise, and a third task do not influence top-ranked preference, then a second robot choose its top-ranked
the preference ordering of any pair. However, all the mis- – or second-ranked, depending on available tasks – preference,
sion tasks are considered by the robots and relevant to the and so on, based on [56] proof.
mission execution.
Allocation Mechanism 1 (Task-centered MRTA-RSD).
The next sections introduce characteristic conditions to each sit-
uation. 1. With uniform probability 1/m, select a robot ri still available
in R and assign it the task of its highest preference.
2. Remove the robot from the list of available robots and the task
5.1. ST-SR-IA
from the list of possible alternatives.
3. Repeat steps 1. and 2. until all robots have been assigned to
According to [54], ST-SR-IA is the most straightforward task
some task.
allocation problem. The literature presents different solution ap-
4. If n > m, when robot ri finishes its task execution, return ri to
proaches of this allocation type, such as Greedy Heuristics, Linear
the list of available robots.
Programming, and Combinatorial Optimization [54,55]. From an
5. Repeat steps 3. and 4. until there are no more available tasks.
Arrovian view of the problem, the robot society’s social welfare
level increases as the task assignment respect the robot’s highest Allocation Mechanism 1 satisfies the conditions of an SWF
preferences. Thus, in the following analysis, we will first consider existence with a slight relaxation in Non-dictatorship condition,
the problems that fulfill the assumption (iii). with consideration of the Pareto ex-post Efficiency, and consid-
Assigning a task to a single robot is instantaneous, having ering it strategy-proof. Kato [59] considers serial dictatorship
no future allocation. Just one alternative is a winner (SCF) for mechanism a ‘‘dictator-maker’’, in a sense that it elects dicta-
each allocation. Although each iteration needs only one result, the tors (agents whose individual preference will be the only one
mechanism performs complete preferences ordering (SWF) as the recognized) serially. In the Allocation Mechanism 1, the first
mission tasks are known. robot chosen is the dictator to the others, who are affected by
The problem analysis considers Condition 2. Thus, the task or- it regarding the available alternatives (tasks) but cannot affect
der must be carried out pairwise, with comparisons independent it back. However, he is not a dictator in the broader sense of
of each other. Also, the Pareto Principle (Condition 3) means that task allocation since he does not impose any choice on the other
the social choice must follow the optimization principle. So, the agents, in line with [60] affirmation. As well, the second robot
solution is Pareto-optimal if robot ri welfare increases without chosen is the dictator for the others, and so on.
decreasing the other group members’ welfare. Bade [58] states that any two agents who submit the same
An SWF applies to the problem since it maps the social pref- preference face the same lottery over the alternatives, i.e., they
erence from the robot team individual preferences. Besides, one are both likely to choose their highest preference alternative. ‘‘So
can discuss the relative simplification in applying a centralized random serial dictatorship treats equals equally’’ [58]. Consider-
allocation mechanism, either in a robot of the team or a central ing the difficulty of meeting the non-dictatorship criterion, [60]
device. As the present study’s objective is to present a methodol- introduced the non-bossiness concept. A mechanism is non-
ogy for an actual application in an MRS, it always considers the bossiness if there is no agent capable of changing the choice
advantages of distributed systems. In this way, it introduces an of another agent without changing its own [58]. As mentioned,
allocation mechanism that meets the conditions imposed for SWF the mechanism is strategy-proof and meets the Pareto Principle.
existence. Reinforcing the consideration that the mechanism is strategy-
Gerkey [54] proposes a variation of the Greedy Algorithm for proof, the assumption (i) states that robots tell the truth and are
iterative task allocation. It presents a very close form to the not envious, so there is no possibility of manipulating individual
simple serial dictatorship (SD) mechanism. In the SD mechanism, preferences for higher individual welfare.
for a preset ordering of agents, the top-ranked agent chooses its The ex-post efficiency means the mechanism is efficient af-
most preferred alternative; the secondly ranked agent chooses ter the agents announce their preferences. According to [61],
its most preferred alternative among the remaining alternatives, this feature causes the mechanism to randomize over the set
and so on [56]. According to Brandt et al. [57], the mechanisms of Pareto-optimal alternatives. Formally, Definition 6 describes a
present Pareto-optimal solutions. The term Allocation Mechanism Pareto ex-post efficient mechanism. Consider π , called policy, the
used in this paper was borrowed from the Mechanism Design sequence in which robots make their task choices, i.e., their task
research area, studied by Social Choice and Computational Social allocation.
Choice researchers. A mechanism is not necessarily an algorithm,
but the ‘‘optimal rules for collective decision making and/or the Definition 6 (ex-post Pareto-efficient Mechanism). A mechanism
allocation of scarce resources by the convergence of a collection is ex-post Pareto-efficient if, ∀⪰r i ∈ [⪰]R , there is no policy π
of such principles’’ [57]. such that π ⪰ri π ∗ for all ri ∈ R and π ≻rj π ∗ for some rj ∈ R.
6
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
5.2. ST-SR-TA complications. With this design, an MRTA architecture has a low
robustness level since all tasks need all robots. Even a partial
ST-SR-TA problem class is similar to the previous one, how- failure of a robot dooms the whole mission. Still, this arrangement
ever, with the time-extended task assignment. According to does not apply one of the main advantages of a multi-robot
Gerkey [54], this problem reflects knowing the future utilities system, the task division.
(preferences) of robots. Also, it requires a thorough understanding Another possible arrangement is a centralized task allocator,
of mission tasks or a model of how the system includes new tasks. which holds all environment information, including individual
These considerations reflect the number of tasks must be greater preferences of each robot on the task execution and possible
than the number of robots, that is, n > m. coalitions. It is a combinatorial problem since the optimum allo-
Because it is a problem that requires future planning, the func- cation is a combinatorial solution from the complete environment
tion applied will be SWF over the set T . However, the function information. The number of possible states grows exponentially
outputs a possible subset Bri ⊆ T of tasks allocated to each with increasing tasks and/or robots in the system. This solu-
robot ri . This subset of tasks is the planning for each robot task tion likewise does not address the MRTA problem in pragmatic
execution, following its preferences ordering. And, after assigning calls. The centralized feature is not beneficial in MRS with mo-
tasks, Br1 ∩ Br2 = ∅. bile robots, and such a mechanism requires high computational
Brandt et al. [57] call this a picking sequence situation. Objects capacity.
or resources are allocated to the agents involved incrementally in It demands a more general approach. The MR characteristic
a predetermined sequence. The author also points out that this of the problem generates a need to form coalitions between
mechanism has two desirable properties: ‘‘it is effortless to im- robots and assign tasks to these coalitions. To the framework
plement and explain, and it also frees the central authority from proposed in this work, both the coalition formation and the task
the burden of provoking agent preferences’’ [57]. Fundamentally, allocation need to consider individual preferences. Thus, one can
agents sequentially choose the alternative of their highest pref- understand the two processes as distinct processes, not being
erence that is still available. The question is how to determine executed by merely one mechanism.
the sequence to meet the conditions for social choice presented
in this paper. 5.4. ST-MR-TA
For the case where each agent chooses more than one object,
there are two possibilities: it reveals one alternative at a time or As discussed by dos Reis and Bastos [3], such a problem
a bundle of alternatives, with all its preferences. In this paper, we includes coalition formation and time-extended task allocation.
consider the simplest case: one alternative at a time. The combi- Nonetheless, [5] indicates a relaxation of the problem by ignoring
natorial nature of the problem and the simplification regarding
the assignment component in time, approximating the ST-MR-IA
the preferences profile and, consequently, the choice mechanism
problem. Consider the individual ordering ⪰r i of the robot ri . dos
justify this proposal.
Reis and Bastos [3] present the robot rm preferences ordering as
Bouveret et al. [62] point out that the number of bundles
a pairwise comparison on the set of alternative tasks T , ⪰rm =
grows exponentially with the increase of alternatives. For exam-
{τ1 ⪰ τ2 ⪰ τ3 . . . ⪰ τn }. Also, the work presents the set R = {⪰r1
ple, with a set of twenty alternatives, the agent can classify more
, ⪰r2 , . . . , ⪰rm } of individual preferences orderings.
than one million bundles, considering an unrestricted domain for
Consider the simplest case: the number of robots required to
the orderings. It is considered, for simplicity, that if the agent
perform each task is equal to m. That is the same number of mem-
presents the preference ordering (τ1 ⪰ τ2 ⪰ τ3 ⪰ τ4 ), indi-
bers in the robot society. The SWF has a preference profile [⪰]R
rectly, it establishes the preference ordering of bundles (τ1 , τ2 ) ⪰
input and maps the complete task ordering to society execution.
(τ3 , τ4 ), without loss of generalization. Aziz et al. [63] point out
In addition to the coalitions task assignment, it also schedules
that the ‘‘well-known mechanisms like serial dictatorship fall un-
coalitions for the time-extended task execution. Therefore, robot
der the umbrella of sequential mechanisms’’. Thus, time-extended
ri has a bundle of allocated tasks similar to the ST-SR-TA problem.
is an extension of the instantaneous assignment problem.
However, task execution depends on coalitions with different
robots.
5.3. ST-MR-IA
ST-MR-IA issue requires the joint effort of several robots to do 5.5. The multi-task robots problem
the same task. According to Gerkey [54], this case must consider
the associated utility of robots, but not necessarily the sum of This class of MRTA problems involves multi-task robots (MT)
these utilities. With the Arrovian problem, the robots’ prefer- and allows the execution of at least two tasks simultaneously by
ences must be aggregated to result in a coalition. Gerkey and the same robot. Gerkey and Matarić [5] describe that the solutions
Matarić [5] further emphasize that this problem is somewhat of the MT–SR–IA and MT–SR–TA problems are equivalent to the
more difficult than those presented above, restricted to single mathematical solutions described for the ST-MR-IA and ST-MR-
robot tasks. The problem is equivalent to the Partitioning Prob- TA problems, respectively. Korsah et al. [37] affirm they are
lem. Furthermore, although literature proposes heuristics for its unaware of MRTA architecture work that encompasses MT-SR-IA,
solution, no MRTA solution immediately derives from such propo- MT-SR-TA, and MT-MR-TA problems. Likewise, the research for
sitions. Some works propose MRTA solutions as ASyMTRe [14] this article found no proposition for these problems.
and RACHNA [64] architectures.
From the social choice point of view, the most straightforward 6. Behavior-based MRTA architecture simulation
case occurs when the number of robots required to complete
each task is equal to m, the number of robots in society. Here, The section presents an MRTA architecture simulation based
the preference aggregation must result from an SWF over T on Alliance implemented in Robot Operational System (ROS)
that elects a single task at a time through an SCF over T . From to illustrate the proposed ordinal axiomatic analysis of MRS. It
individual preferences, the m society robots decide which task analyzes two cases: (a) when the individuals have the same
τ ∈ T to execute. At each execution term, society makes a preference profile and (b) when individuals have a different pref-
new choice. However, in pragmatic terms, this condition presents erence profile. In the last case, the simulations aim to show that a
7
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
τ4 Report: the robot must go to a report area on the map and policies are ex-post Pareto-efficient, depending entirely on the
publishes the mission status in a defined ROS topic. random robot startup.
The architecture satisfies Condition 1, considering robots can
Each task has a motivation function with several parameters,
order tasks in any logically possible way.2 It also satisfies Con-
as explained by [66,67], which set the motivation level that
dition 2, as shown in the analysis of the transitivity and com-
increases over time. The motivation function is the individual
pleteness of the preference ordering. As the individual preference
utility function that the robot uses to order tasks according to its
orderings, Conditions 3 and 4 are automatically met. Therefore,
preference. The robot allocates a task when the related behavior
the architecture performs a Social Welfare Function over the set
set motivation level reaches its activation threshold level. Each
T , mapping the individual preferences into a complete social
task has its threshold, but the behavior set activation threshold
preference ordering.
level is 2000 units in the simulations for simplicity.
As the robot chooses the task, it communicates to other society
6.2. Robots have the same preference ordering
members that suppress the task execution. While the robot exe-
cutes the task, the others maintain a growing impatience level. If
In this situation, the preference ordering of all individuals is
the first robot cannot complete the task, it gives up, and another
equal, as Eq. (2) shows.
one, if idle, can activate the same behavior set. Robots broadcast
task-related messages over a ROS topic, in which all society ⪰r 1 = ⪰r 2 = ⪰r 3 = (τ3 ≻ τ1 ∼ τ2 ≻ τ4 ) (2)
members are publishers and subscribers. The message content
is simply the active behavior set and no further information on From Eq. (2) one can infer the pairwise preferences: if (τ3 ≻
individual motivation levels. Neither negotiation occurs. τ1 ) and (τ1 ∼ τ2 ), then (τ3 ≻ τ2 ); if (τ1 ∼ τ2 ) and (τ2 ≻ τ4 ), then
Although the simulated architecture was not developed under (τ1 ≻ τ4 ); and, therefore, (τ3 ≻ τ4 ). So the preference ordering
an Arrovian point of view, we show it follows the Arrovian is complete and transitive. It is not hard to conclude that the
conditions the paper established and operates as described in emerging social ordering of this setting the same as the above
Allocation Mechanism 1. Step 1 requires the robots to have the preference orderings. And, as Fig. 4 shows, the sequence as the
same probability of choosing their top-ranked task. As the sim- robots allocate the tasks follows the social ordering. Fig. 4 shows
ulation launch starts the robots’ ROS nodes in different instants, the task allocation over time. The vertical axis is the mission task.
we added a random delay – in the range from 0 to 1 s – in each So, the graph shows the system works and which robot executes
robot initialization to emulate a real robot delay initialization, each task. All the robots start at idle behavior. As each behavior
for instance. It prevents the first initialized robot node from set motivation grows, as shown in Fig. 5, the task allocation
always allocates its top-ranked task first. In the second step, a occurs. The graph Detail on task allocation presents the behavior
robot with an allocated task is not available to allocate another sets activation within the first seconds of simulation.
one. In Alliance, when a robot activates a behavior set, the Considering the robots have the same preferences, the random
other motivation levels remain null, meaning the other behavior delay of initialization implemented in the simulation defines the
sets are inhibited. While the robots motivation levels increase, task allocation order. The first instantiated robot node starts the
robots choose their top-ranked available tasks, i.e., behavior sets motivation levels calculation first, then reaches the threshold
not inhibited by other society members. Thus, the architecture first. Fig. 5 shows the robots’ motivation evolution over time. The
accomplishes step 3. When a robot finishes a task, the motivation plot x-axis presents a time scale for the first 10 s of simulation
levels of inhibited behavior sets increase overtime. So the robot and another time scale for the remaining. It aims to highlight the
can choose an available task. And so on, until accomplishing period of task allocation. Looking at the robot r1 plot, it has a
the mission. It is equivalent to steps 4 and 5 of the Allocation delay of 1 s. So, it took longer than the other robots to start the
Mechanism 1. motivation calculation. Despite the difference in the initialization
Both situations implement the random delay on robots startup, time, observe the motivation levels of the behavior sets have the
meeting the Allocation Mechanism 1 first step. However, the same growth rate in all the society.
random startup impacts similar preference profiles. The second In this case, according to the graphs in Fig. 5, r3 is the first
case preference profiles and motivation level settings do not instantiated robot, so it chooses the top-ranked task first, the
impact the task assignment. Besides, the architecture is ex-post boundary patrol task. After its allocation, the motivation of the
Pareto-efficient, Definition 6. No other policy π increases the
welfare of a robot without decreasing from another, i.e., no other 2 Here, as a logically possible way, we mean an order that considers tasks
task allocation sequence improves the social welfare. Further, ordering dependencies as defined in Alliance. For instance, a task like find the
with the same preference orderings, in the first case, all possible box must come first, then a task such push the box to position A.
9
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
Fig. 4. Task allocation and execution over time in a situation that robots have the same preference ordering.
⪰r 1 τ3 ≻ τ2 ≻ τ1 ≻ τ4
[ ] [ ] 3 We simulate a robot failure by closing its control node in ROS, killing
[⪰b ]R = ⪰r 2 = τ3 ≻ τ1 ≻ τ2 ≻ τ4 (4) its process. As an Alliance characteristic, robot r3 gave up the task execution
⪰r 3 τ1 ≻ τ2 ≻ τ3 ≻ τ4 because of the robot acquiescence, and robot r2 finished it.
10
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
Fig. 6. Task allocation and execution over time for three preference profiles.
Fig. 7. Evolution of motivation values over time in the case of preference profile b, with a fail of robot r3 .
τ4 and closes the mission, in Fig. 8(f) and (g). See Fig. 8 for more The consideration of independent tasks simplifies the propo-
details. Nevertheless, the fault-tolerant characteristic of Alliance sitions of this article. In the ordinal preference, a greater depen-
does not affect the social ordering and Arrovian analysis. Even dence relation among alternatives demands a conditional prefer-
with a robot failure, social ordering remains the same. ence definition among them. Consider the tasks as alternatives.
The utility calculation performed on each robot must incorporate
a conditional function to evaluate the influence of a new state
7. Discussion
of the system on the other alternatives. For example, robot rm
allocates the task τk . Does this new state change preferences
This paper’s allocation mechanisms do not consider certain among other alternatives? If the answer is yes, then the tasks
practical aspects of task allocation architectures in MRS. From the have a dependency the allocation mechanism must consider.
hardware point of view, some of these aspects are communication Another critical point is the number of robots and tasks in real
means, computational power, actuators, and sensors’ limitations. MRS. The number of robots does not exceed a dozen, in general.
The simulation discussion shows the MRTA problem can be an- Likewise, situated systems, i.e., not simulated, usually do not
alyzed from the Arrovian point of view even in a task allocation present a high number of tasks, even on account of the execution
architecture implemented in another domain. time. So preference orders do not involve a large number of
11
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
Fig. 8. Simulation of preference profile b: r1 has a gray trail, r2 has a white trail, and r3 has a black one.
alternatives, being it tasks or robots. Nevertheless, future work an analysis framework based on Social Choice Theory and not a
must verify the scalability of the algorithms when applied to new MRTA architecture. It addresses the single-task robot prob-
large-scale MRS. lem and presents the results in a behavior-based architecture.
However, we can extend the same framework for other MRTA
8. Conclusion taxonomy problems and other architectures.
Although it is not the intent of this work since it introduces
This paper proposes a Social Choice-based Arrovian struc- the research, a scheme that obeys the axioms of social choice can
ture to evaluate task allocation mechanisms in multi-robot sys- focus on preference communication and aggregation rather than
tems. Moreover, it addresses the use of individual ordinal prefer- individual preference ordering. The hypothesis for future research
ences and preference aggregation mechanisms for social decision- is that an individual’s utility function can be as simple as possible,
making. The proposed Arrovian view is another perspective to using its metrics, and it will not impact the social decision. Also as
address the MRTA problem analysis. Applied to MRTA problems, it future work, we can cite the implementation of a market-based
contributes to the theoretical and practical study of such problem and optimization-based architecture and their simulation and the
solutions. Like Arrow’s axiomatic method, the paper proposes Arrovian analysis of other architectures.
12
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
Declaration of competing interest [27] A.B. Bahgat, O.M. Shehata, I.M. El Sayed, A multi-level architecture for
solving the multi-robot task allocation problem using a market-based
approach, Int. J. Mech. Eng. Robot. Res. 9 (2) (2020).
The authors declare that they have no known competing finan- [28] M. Rodríguez, A. Al-Kaff, Á. Madridano, D. Martín, A. de la Escalera,
cial interests or personal relationships that could have appeared Wilderness search and rescue with heterogeneous multi-robot systems,
to influence the work reported in this paper. in: 2020 International Conference on Unmanned Aircraft Systems (ICUAS),
IEEE, 2020, pp. 110–116.
[29] X. Chen, P. Zhang, G. Du, F. Li, A distributed method for dynamic multi-
References robot task allocation problems with critical time constraints, Robot. Auton.
Syst. 118 (2019) 31–46.
[1] K.J. Arrow, A. Sen, K. Suzumura, Handbook of Social Choice and Welfare, [30] X. Zhou, H. Wang, B. Ding, T. Hu, S. Shang, Balanced connected task
Vol. 2, Elsevier, Amsterdam, 2011. allocations for multi-robot systems: An exact flow-based integer program
[2] W. Gaertner, A Primer in Social Choice Theory: Revised Edition, Oxford and an approximate tree-based genetic algorithm, Expert Syst. Appl. 116
University Press, Oxford, 2009. (2019) 10–20.
[3] W.P.N. dos Reis, G.S. Bastos, An arrovian view on the multi-robot task [31] M. Shelkamy, C.M. Elias, D.M. Mahfouz, O.M. Shehata, Comparative analysis
allocation problem, in: Advanced Robotics (ICAR), 2017 18th International of various optimization techniques for solving multi-robot task allocation
Conference on, IEEE, 2017, pp. 290–295. problem, in: 2020 2nd Novel Intelligent and Leading Emerging Sciences
[4] L.E. Parker, Heterogeneous Multi-Robot Cooperation (Ph.D. thesis), Conference (NILES), IEEE, 2020, pp. 538–543.
Massachusetts Institute of Technology, 1994. [32] F. Zitouni, S. Harous, R. Maamri, A distributed solution to the multi-robot
[5] B.P. Gerkey, M.J. Matarić, A formal analysis and taxonomy of task allocation task allocation problem using ant colony optimization and bat algo-
in multi-robot systems, Int. J. Robot. Res. 23 (9) (2004) 939–954. rithm, in: Advances in Machine Learning and Computational Intelligence,
[6] K.J. Arrow, Social Choice and Individual Values, Vol. 12, Yale university Springer, 2021, pp. 477–490.
press, New York, 1951. [33] N. Seenu, K.C. RM, M. Ramya, M.N. Janardhanan, Review on state-of-the-art
[7] L. Yliniemi, A.K. Agogino, K. Tumer, Multirobot coordination for space dynamic task allocation strategies for multiple-robot systems, Ind. Robot
exploration, AI Mag. 35 (4) (2014) 61–74. Int. J. Robot. Res. Appl. (2020).
[8] G.P. Das, T.M. McGinnity, S.A. Coleman, L. Behera, A distributed task [34] Y. Rizk, M. Awad, E.W. Tunstel, Cooperative heterogeneous multi-robot
allocation algorithm for a multi-robot system in healthcare facilities, J. systems: a survey, ACM Comput. Surv. 52 (2) (2019) 1–31.
Intell. Robot. Syst. (2014) 1–26. [35] D.A.S. Lantao Liu, Large-scale multi-robot task allocation via dynamic
[9] Y.U. Cao, A.S. Fukunaga, A.B. Kahng, F. Meng, Cooperative mobile robotics: partitioning and distribution, Auton. Robots 33 (2012).
Antecedents and directions, in: Intelligent Robots and Systems 95.’Human [36] A. Hooshangi, A.A. Navid, Agent-based task allocation under uncertainties
Robot Interaction and Cooperative Robots’, Proceedings. 1995 IEEE/RSJ in disaster environments: An approach to interval uncertainty, Int. J.
International Conference on, 1, IEEE, Pittsburgh, 1995, pp. 226–234. Disaster Risk Reduct. 24 (2017).
[10] L.E. Parker, Multiple mobile robot systems, in: Springer Handbook of [37] G.A. Korsah, A. Stentz, M.B. Dias, A comprehensive taxonomy for
Robotics, Springer, 2008, pp. 921–941. multi-robot task allocation, Int. J. Robot. Res. 32 (12) (2013) 1495–1512.
[11] L.E. Parker, ALLIANCE: An architecture for fault tolerant multirobot [38] M. Cristani, F. Olivieri, G. Governatori, Non-monotonic collective decisions,
cooperation, Roboti. Autom. IEEE Trans. 14 (2) (1998) 220–240. in: International Conference on Principles and Practice of Multi-Agent
Systems, Springer, 2019, pp. 387–404.
[12] B.B. Werger, M.J. Matarić, Broadcast of local eligibility for multi-target
[39] Z.A. Dodevska, Computational social choice and challenges of voting in
observation, in: Distributed Autonomous Robotic Systems 4, Springer,
multi-agent systems, Tehnika 74 (5) (2019) 724–730.
2000, pp. 347–356.
[40] C. Cornelio, M.S. Pini, F. Rossi, K.B. Venable, Multi-agent soft constraint
[13] T. Vu, J. Go, G. Kaminka, M. Veloso, B. Browning, MONAD: A flexible archi-
aggregation via sequential voting: theoretical and experimental results,
tecture for multi-agent control, in: Proceedings of the Second International
Auton. Agents Multi-Agent Syst. 33 (1–2) (2019) 159–191.
Joint Conference on Autonomous Agents and Multiagent Systems, ACM,
[41] J. Fisher, Efficient random assignment with cardinal and ordinal
2003, pp. 449–456.
preferences, J. Mech. Inst. Des. 3 (2018) 51–96.
[14] F. Tang, L.E. Parker, Asymtre: Automated synthesis of multi-robot task
[42] B. Greve, What is welfare? Central Eur. J. Public Policy 2 (1) (2008) 50–73.
solutions through software reconfiguration, in: Robotics and Automation,
[43] E. Allardt, To have, to love, to be–about welfare in the Nordic countries,
2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference
Argos, Lund (Swedish), 1975.
on, IEEE, 2005, pp. 1501–1508.
[44] S.K. Nath, A Perspective of Welfare Economics, Macmillan Press LTD,
[15] L.E. Parker, On the design of behavior-based multi-robot teams, Adv. Robot.
London, 1973.
10 (6) (1995) 547–578.
[45] B.M. Van Praag, P. Frijters, The measurement of welfare and well-being:
[16] Z. Wang, J. Zhu, X. Guo, Y. Ma, Z. Li, Distributed task allocation method
The leyden approach, Well-Being: Found. Hedonic Psychol. (1999) 413.
based on self-awareness of autonomous robots, J. Supercomput. 76 (2)
[46] D. Read, Utility theory from Jeremy Bentham to Daniel Kahneman,
(2020) 831–843.
Department of Operational Research, 2004.
[17] M. Vinagre, J. Aranda, A. Casals, Suitable task allocation in intelligent sys-
[47] S. Barberà, P. Hammond, C. Seidl, Handbook of Utility Theory: Volume 1
tems for assistive environments, in: Iberian Robotics Conference, Springer,
Principles, Springer Science & Business Media, Boston, 2004.
2019, pp. 497–509.
[48] C. List, Social choice theory, in: E.N. Zalta (Ed.), The Stanford Encyclopedia
[18] W. Dai, H. Lu, J. Xiao, Z. Zheng, Task allocation without communication of Philosophy, Winter 2013 edn., Metaphysics Research Lab, Stanford
based on incomplete information game theory for multi-robot systems, J. University, 2013.
Intell. Robot. Syst. 94 (3) (2019) 841–856. [49] M. Fleurbaey, F. Maniquet, A Theory of Fairness and Social Welfare, Vol.
[19] Z. Yan, N. Jouandeau, A.A. Cherif, A survey and analysis of multi-robot 48, Cambridge University Press, Cambridge, 2011.
coordination, Int. J. Adv. Robot. Syst. 10 (12) (2013) 399. [50] K.J. Arrow, A difficulty in the concept of social welfare, J. Polit. Econ. (1950)
[20] B.P. Gerkey, M.J. Matarić, Sold!: Auction methods for multirobot 328–346.
coordination, Robot. Autom. IEEE Trans. 18 (5) (2002) 758–768. [51] Y. Shoham, K. Leyton-Brown, Multiagent Systems: Algorithmic, Game-
[21] M.B. Dias, Traderbots: A new paradigm for robust and efficient multirobot Theoretic, and Logical Foundations, Cambridge University Press, 2008.
coordination in dynamic environments, Robot. Instit. (2004) 153. [52] A. Gibbard, Manipulation of voting schemes: a general result, Econometrica
[22] S. Botelho, R. Alami, Robots that cooperatively enhance their plans, in: (1973) 587–601.
Distributed Autonomous Robotic Systems 4, Springer, 2000, pp. 55–65. [53] M.A. Satterthwaite, Strategy-proofness and arrow’s conditions: Existence
[23] S. Botelho, R. Alami, Multi-robot cooperation through the common use of and correspondence theorems for voting procedures and social welfare
‘‘mechanisms’’, in: Intelligent Robots and Systems, 2001. Proceedings. 2001 functions, J. Econom. Theory 10 (2) (1975) 187–217.
IEEE/RSJ International Conference on, 1, IEEE, Maui, 2001, pp. 375–380. [54] B.P. Gerkey, On multi-robot task allocation (Ph.D. thesis), Center for
[24] S.C. Botelho, R. Alami, M+: a scheme for multi-robot cooperation through Robotics and Embedded Systems, University of Southern California, Los
negotiated task allocation and achievement, in: Robotics and Automation, Angeles, 2003, p. 2003. 117.
1999. Proceedings. 1999 IEEE International Conference on, 2, IEEE, Detroit, [55] B.B. Choudhury, Task Allocation Strategies in Multi-Robot Environ-
1999, pp. 1234–1239. ment (Ph.D. thesis), Department of Mechanical Engineering, National
[25] W. Dai, H. Lu, J. Xiao, Z. Zeng, Z. Zheng, Multi-robot dynamic task Institute of Technology Rourkela, Orissa, Índia, 2009, p. 2009. 236.
allocation for exploration and destruction, J. Intell. Robot. Syst. 98 (2) [56] A. Abdulkadiroğlu, T. Sönmez, Random serial dictatorship and the core
(2020) 455–479. from random endowments in house allocation problems, Econometrica 66
[26] F. Zitouni, R. Maamri, S. Harous, FA–QABC–MRTA: a solution for solving (3) (1998) 689–701.
the multi-robot task allocation problem, Intell. Serv. Robot. 12 (4) (2019) [57] F. Brandt, V. Conitzer, U. Endriss, A.D. Procaccia, J. Lang, Handbook of
407–418. Computational Social Choice, Cambridge University Press, Nova York, 2016.
13
W.P.N. dos Reis, G.L. Lopes and G.S. Bastos Robotics and Autonomous Systems 144 (2021) 103839
[58] S. Bade, Random serial dictatorship: the one and only, Royal Holloway [67] W.P.N. dos Reis, G.S. Bastos, Multi-robot task allocation approach using ros,
College, Unpublished Mimeo, 2014. in: 2015 12th Latin American Robotics Symposium and 2015 3rd Brazilian
[59] M. Kato, S. Ohseto, Toward general impossibility theorems in pure Symposium on Robotics (LARS-SBR), IEEE, 2015, pp. 163–168.
exchange economies, Soc. Choice Welf. 19 (3) (2002) 659–664.
[60] M.A. Satterthwaite, H. Sonnenschein, Strategy-proof allocation mechanisms
at differentiable points, Rev. Econom. Stud. 48 (4) (1981) 587–597.
Wallace Pereira Neves dos Reis: Graduated in Elec-
[61] H. Aziz, J. Mestre, Parametrized algorithms for random serial dictatorship,
trical Engineering with an emphasis on Electronics,
Math. Social Sci. 72 (2014) 1–6.
Specialization in New Educational Technologies, and
[62] S. Bouveret, U. Endriss, J. Lang, Fair division under ordinal preferences:
Master in Electrical Engineering, the last from the
Computing envy-free allocations of indivisible goods, in: Proceedings of
Federal University of Itajubá - UNIFEI. He is currently
the 2010 Conference on ECAI 2010: 19th European Conference on Artificial
a doctoral student in the Computer Science Program at
Intelligence, IOS Press, Düsseldorf, 2010, pp. 387–392.
the Federal University of São Carlos - UFSCar, research
[63] H. Aziz, T. Walsh, L. Xia, Possible and necessary allocations via sequential
line Artificial Intelligence, in Industrial Automation. He
mechanisms, in: Proceedings of the 24th International Conference on
is a Professor of Basic Technical and Technological
Artificial Intelligence, AAAI Press, Buenos Aires, 2015, pp. 468–474.
Education at the Federal Institute of Education, Science,
[64] L. Vig, J.A. Adams, Market-based multi-robot coalition formation, Distrib.
and Technology of Rio de Janeiro, IFRJ - Campus Volta
Autonomous Robot. Syst. 7 (2006) 227–236.
Redonda, since 2014, working mainly on technical courses in Electrotechnics
[65] R. Brooks, A robust layered control system for a mobile robot, IEEE J. Robot.
and Industrial Automation. He served as Director of Administration from
Autom. 2 (1) (1986) 14–23.
March 2016 to June 2018 at the same campus. He has experience in Control
[66] W.P.N. dos Reis, G.S. Bastos, Implementing and simulating an alliance-
Systems, Multirobot Systems, Social Robotics, Computational Social Choice, Em-
based multi-robot task allocation architecture using ROS, in: Robotics,
bedded Electronics, Image Processing, Educational Robotics. His current research
Springer, 2016, pp. 210–227.
interests are AGV Dispatching and AGV Position Control Design.
14