Dynamic Task Allocation for Human-robot Teams

Tinka R. A. Giele

1,2

, Tina Mioch

, Mark A. Neerincx

and John-Jules C. Meyer

Utrecht University, Drift 10, 3512 BS Utrecht, The Netherlands

TNO, Kampweg 5, 3769 DE Soesterberg, The Netherlands

Keywords:

Human-robot Interaction, Agent Cooperation.

Abstract:

Artiﬁcial agents, such as robots, are increasingly deployed for teamwork in dynamic, high-demand environ-

ments. This paper presents a framework, which applies context information to establish task (re)allocations

that improve human-robot team’s performance. Based on the framework, a model for adaptive automation

was designed that takes the cognitive task load (CTL) of a human team member and the coordination costs of

switching to a new task allocation into account. Based on these two context factors, it tries to optimize the level

of autonomy of a robot for each task. The model was instantiated for a single human agent cooperating with

a single robot in the urban search and rescue domain. A ﬁrst experiment provided encouraging results: the

cognitive task load of participants mostly reacted to the model as intended. Recommendations for improving

the model are provided, such as adding more context information.

1 INTRODUCTION

Teams are groups consisting of two or more actors

that set out to achieve a joint goal. A good task al-

location is crucial for team performance, especially

when teams have to cope with high-demand situations

(e.g., at disaster responses). Task allocation should be

ﬂexible: when an environment is dynamic or states

of team members change, reallocating tasks could

be beneﬁcial for team performance (Brannick et al.,

1997). Making a (human) team member responsible

for dynamically allocating tasks, causes extra work-

load (Barnes et al., 2008). To avoid this, tasks should

be reallocated automatically. Such allocation is im-

portant for mixed human-robot teams (Burke et al.,

2004), for example rescue teams including an robot

to explore terrains unsafe for humans. The dynamic

allocation of tasks to human or robot is called adaptive

automation, distinguishing intermediate levels of au-

tonomy for each task in a joint effort to complete the

task. An example is way-point navigation, in which

the operator sets the way-points and the robot drives

along them. Recent research shows that dynamically

adapting autonomy levels of robots could help opti-

mizing team performance, when this process is auto-

mated (Calhoun et al., 2012).

An important challenge in adaptive automation is

deciding when to change the level of autonomy of the

robot, and to which level. This can be done based

on the cognitive task load of the operator (Neerincx,

2003), as cognitive task load has an inﬂuence on per-

formance (Neerincx et al., 2009). In addition, cog-

nitive task load itself is inﬂuenced by changing lev-

els of automation, as the level of autonomy and op-

erator task load are inversely correlated if other fac-

tors remain stable (Steinfeld et al., 2006). This does

not hold for the relation between autonomy levels and

operator performance. Setting robot autonomy very

high might cause human-out-of-the-loop problems,

whereas setting autonomy very low might cause task

overload for the operator; both decrease performance.

This study, ﬁrst, aims at the design and formaliza-

tion of a general dynamic task allocation framework

that speciﬁes concepts and their effect on team per-

formance, which can be used to dynamically allocate

tasks. Subsequently, this framework is used to design

a practical model for adaptive automation, based on

cognitive task load. Finally, the model is instantiated

for an experimental setting in the urban search and

rescue domain for a ﬁrst validation of the model.

2 BACKGROUND

Team Performance. Team performance is a mea-

sure of how well a common goal is achieved. Early

frameworks describing team performance commonly

follow the Input-Process-Output structure. For exam-

117

R. A. Giele T., Mioch T., A. Neerincx M. and C. Meyer J..

Dynamic Task Allocation for Human-robot Teams.

DOI: 10.5220/0005178001170124

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 117-124

ISBN: 978-989-758-073-4

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

ple, McGrath (McGrath, 1964) describes three input

concepts: individual level factors (e.g. cognitive abil-

ity), group level factors (e.g. communication) and en-

vironmental factors (e.g. resource availability, task

difﬁculty). These factors are input for the team’s in-

teraction processes; the output concept is team perfor-

mance. This framework has some downsides. Feed-

back loops are excluded, e.g., team performance it-

self cannot serve as an input for interaction processes.

Also, the Input-Process-Output structure suggests lin-

ear progression, but interactions between various in-

puts and processes or between different processes are

also possible (Ilgen et al., 2005). Outside McGrath’s

framework, a vast amount of research has focused on

the numerous factors that inﬂuence individual perfor-

mance (Matthew et al., 2000), for example cognitive

task load (Neerincx, 2003).

Dynamic Task Allocation. Dynamic task allocation

beneﬁts team performance (Brannick et al., 1997), it

can be effectuated in numerous ways. First, responsi-

bility can be distributed, or it can be centralized. Dis-

tributed responsibility for dynamic task allocation has

the disadvantage that it causes extra workload for (hu-

man) team members (Barnes et al., 2008). Disadvan-

tages of centralized coordination are that it might be

unfeasible to implement for very large teams, and that

task reallocations need to be clearly communicated to

the team members. Second, Inagaki (Inagaki, 2003)

argues that a dynamic form of comparison allocation

is the best strategy for task allocation. Comparison

allocation means tasks are allocated based on capa-

bilities of actors.

Adaptive Automation. Traditionally, tasks in mixed

human-robot teams are allocated either fully to a hu-

man or fully to a robot, e.g. based on a list of static

human versus robot capabilities (Fitts et al., 1951).

This way of allocating tasks has the problem that it

is overly coarse. In addition, static task allocation

is insufﬁcient for dynamic environments, as capabil-

ities needed for a task could change (Inagaki, 2003).

Adaptive task allocation addresses these issues.

Numerous studies have shown the positive effects

of dynamic task allocation via adaptive automation in

single human-single robot teams, e.g., improved per-

formance, enhanced situation awareness and reduced

cognitive workload (Greef et al., 2010), (Bailey et al.,

2006), (Calhoun et al., 2012). A few studies have

looked at adaptive automation in the context of sin-

gle human-multiple robot teams (Parasuraman et al.,

2009), (Kidwell et al., 2012). In these studies how-

ever, only the level of autonomy of a single robot or

of a separate system on a single task was adapted.

Different techniques for triggering reallocation are

possible, for example techniques based on perfor-

mance (Calhoun et al., 2012), psycho-physiological

measures (Bailey et al., 2006), operator cogni-

tion (Hilburn et al., 1993), environment (Moray et al.,

2000) or hybrid techniques (Greef et al., 2010). How-

ever, not all tasks allow for real-time performance

measurement, psycho-physiological measures are not

suitable for all settings, and environment-based tech-

niques in isolation fail to capture changing states of

team members. Hybrid techniques are more robust

as multiple factors can be used (Greef et al., 2010).

Only a limited amount of studies have used hybrid

techniques (Greef et al., 2010).

Cognitive Task Load. An important factor

for dynamic task allocation in teams, operating

in high-demand situations, is cognitive task load

(CTL) (Guzzo et al., 1995). A model of CTL was

proposed by Neerincx (Neerincx, 2003). The model

describes how task characteristics are of inﬂuence on

individual performance and mental effort. CTL can

be described as a function over three metrics. The

time occupied is the amount of time a person spends

performing a task, the number of task-set switches is

the number of times that a person has to switch be-

tween different tasks. The level of information pro-

cessing is the type of cognitive processes required by

recent tasks. When the values for the three metrics fall

into a certain range (corresponding to a certain region

in CTL-space), the operator is diagnosed to be in a

certain mental state, i.e., vigilance, underload, over-

load, and cognitive lock-up. Being in such a state

has a negative inﬂuence on performance. The CTL

model has been experimentally validated in the naval

domain (Neerincx et al., 2009).

3 DYNAMIC TASK ALLOCATION

FRAMEWORK

Dynamic task allocation can be seen as optimizing

a utility (evaluation) function. Firstly, possible role

assignments are generated from context information.

Role assignments are a combination of a robot and

a set of tasks this robot could execute. These role

assignments are then evaluated using context infor-

mation relevant to how well the robot is able to exe-

cute the set of tasks. Secondly, an optimization algo-

rithm is applied, which ﬁnds the collection of options

which has the highest utility and allocates every task

to a robot. This collection of options is a task alloca-

tion (Gerkey and Matari

c, 2004).

This approach has some limitations. The utility

of a robot-task pair is assumed not to be inﬂuenced

by other tasks the robot might be doing. Also, this

analysis does not include mixed human-robot teams.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

118

More importantly, multi-robot task allocation prob-

lems are reduced to optimization problems, but some

important steps that are needed to realize this reduc-

tion are underspeciﬁed: generating the feasible role

assignments and how to evaluate these. Our frame-

work builds on Gerkey and Matari

c’s analysis, and

improves it on these aspects. We speciﬁcally address

the issues of option generation and utility calculation.

Once we have dealt with these issues, we reduce the

task allocation problem to the set-partitioning prob-

lem (SPP). Although the SPP is strongly NP-hard, it

has been studied extensively and many heuristic algo-

rithms that give good approximations have been de-

veloped (Gerkey and Matari

c, 2004).

An overview of the proposed framework is shown

in Figure 1. Three categories of factors that inﬂuence

task allocation (individual, environmental, and task

factors) are represented by the three input concepts

in the top of the ﬁgure.

Task models represents task factors: S

(T,t)

where S

is a name of a property or state, T is a task

and t is a time point. Task models contains functions

from a speciﬁc task to its properties (static) and states

(dynamic) at a certain point in time. Examples in-

clude location and resource requirement.

Environment models represent environmental fac-

tors: S

(E,t) where E is an environment. Environ-

ment models are functions that describe states and

properties of the environment that are dependent on

the location and possibly the time (e.g. resource avail-

ability and weather conditions).

Actor models represent individual factors. Actor

models are functions that describe for each actor their

relevant abilities and states, associated with a certain

point in time: S

(A,t) where A is an actor. Abili-

ties are static, for example IQ, personality traits and

skills. The dynamic counterpart of actor abilities are

actor states, for example emotion, location and fa-

tigue. An important inﬂuence on task allocation is the

cost caused by the reallocation of tasks (Barnes et al.,

2008); for that reason, our framework includes a feed-

back loop for the task allocation itself (denoted by the

dashed arrow). The current task allocation itself thus

is an actor state.

Some factors inﬂuencing task allocation can only

be described by combining factors from the categories

mentioned above. These factors are represented by

the concept of situation models in our framework:

(hA,T i,t) where T is a set of tasks. Situation

model functions are always described using functions

from actor models, environment models and/or task

factors. An example is the distance between an ac-

tor and a task, a function that is described using both

actor location and task location.

To come to an optimal task allocation, three pro-

cesses are identiﬁed, namely option generation and

pruning, utility calculation, and determining the opti-

mal task allocation (see colored boxes in Figure 1).

The ﬁrst process is option generation and pruning.

An option is a actor-task set combination, O = hA,T i.

Options are generated from the set of actors (input)

and the set of tasks (input). Then, restrictive factors

are used to prune the set of possibilities. For example,

an actor might lack the proper sensors to execute a

task.

The second process is utility calculation. For this

process, preference factors are used. Preference fac-

tors give an indication of how well the task set can be

executed by the actor. For example, if an actor has

been assigned a single, but difﬁcult task, he might do

better on this task than if he has also been assigned to

do several other tasks. All actor-task set combina-

tions are mapped to a utility value using some func-

tion that combines the outcomes of all the preference

factors.

The ﬁnal process is determining the optimal task

allocation. With the utility function and the set of

possible actor-task set pairs, we can use a SPP solving

algorithm (Gerkey and Matari

c, 2004) to arrive at the

best task allocation for a speciﬁc time.

Solving the task allocation problem by using the

SPP introduces the assumption that all tasks need

to be allocated to an actor. This excludes scenarios

where it might not be possible or preferable to allo-

cate all tasks. We relieve this assumption by intro-

ducing a placeholder for tasks that are not executed,

a dummy actor. Tasks allocated to the dummy ac-

tor are not executed. We can now model mandatory

tasks by deﬁning a restrictive factor that prunes role

assignments that assign the dummy actor to manda-

tory tasks. Also, the costs of not executing certain

tasks can be easily modeled using a preference factor,

since the set of tasks that are not executed is the set of

tasks assigned to the dummy actor.

4 MODEL FOR ADAPTIVE

AUTOMATION

In adaptive automation, tasks are dynamically allo-

cated at a speciﬁc level of autonomy. Based on the

framework, we build the model by deﬁning the factors

to be included as inﬂuence on adaptive automation.

As argued in Section 2, cognitive task load is a good

candidate as it affects performance and is inﬂuenced

by the tasks an actor has. Speciﬁcally, it is likely to be

inﬂuenced by at which level of autonomy an allocated

task is. We will include the predicted cognitive task

DynamicTaskAllocationforHuman-robotTeams

119

Situation models

Option

generation

and pruning

Utility

calculation

Determining

optimal task

allocation

Task

models

Environment

models

Actor

models

Restrictive

factors

Preference

factors

*Options are combinations of an actor and a set of tasks (allocated to that actor), i.e. a role assignment.

Set of actors

Set of tasks

Set of options*

with utility

Optimal

task allocation

Figure 1: Overview of the proposed framework. Boxes denote processes, arrows represent ﬂow of information. Opposite to

Gerkey and Matari

c’s (Gerkey and Matari

c, 2004) focus, we focus on the process of pruning generated options and calculating

utility of options (darker boxes) and less on the process of optimization (lighter box).

load of an actor on a set of tasks as a preference fac-

tor in our model. Cognitive task load encompasses the

metric task switching. We deﬁne this metric to only

cover task switches that are not caused by task reallo-

cations, but only by an actor switching between tasks

he is both assigned to (for example switching between

driving and looking around while exploring an area).

We deﬁne costs that are caused by task reallocations

as coordination costs and include this as a separate

preference factor. Team performance could beneﬁt

from an actor switching between different (levels of

autonomy of) tasks if it reduces the negative effect on

performance of the cognitive state he is in, but only

if the coordination costs do not outweigh the cost of

the negative effect on performance of the cognitive

state (Inagaki, 2003).

Levels of Autonomy. Tasks that have multiple possi-

ble levels of automation are replaced in the task model

by a separate version of the task for each different

level of autonomy, T becomes {T

,..., T

}. Tasks

at intermediate levels of autonomy (for example way-

point driving) are divided into two subtasks, one for

an operator (setting way-points) and one for a robot

(driving along the way-points). The separate versions

all need to be described in terms of task state con-

cepts. The same task at several different levels of

autonomy can be modeled as several mutually exclu-

sive subtasks. All but one of the mutually exclusive

tasks (which could consist of two subtasks) should be

forcibly allocated to the dummy actor, ensuring a task

is only allocated at a single level of autonomy to a real

actor.

Cognitive Task Load. We use the predicted CTL

level of an actor on a task set to help decide how well

this task set is suited to be executed by the actor (rel-

ative to other tasks sets). All three metrics of CTL are

situation state concepts, they are some function over

an option (actor-task set), using the properties of the

tasks in the task set. Using the three metrics, we can

estimate whether the CTL level of an actor will be in a

problem region given a set of tasks. Task allocations

that keep actors out of CTL problem regions should

be preferred. Timing is also an important aspect in

CTL. The longer a person’s CTL is in a problem re-

gion, the more negative the effect on performance will

be. Typically, vigilance and underload problems oc-

cur only after some time (900 seconds), while over-

load and cognitive lock-up problems can occur even

if the CTL has only been in the problem region for a

short time (300 seconds) (Neerincx, 2003). Cognitive

task load as proposed by Neerincx only makes sense

in the context of humans, not for robots. For example,

robots cannot suffer from vigilance problems if they

are bored, because generally robots cannot be bored.

The formal description of the preference concept

CTL can be seen in Equation 1. Preference based on

CTL ranges from 1 (most preferred) to 0 (least pre-

ferred). The ’isHuman’ function describes whether

an actor is a human, the ’cognitiveState’ functions

describe whether an actor is in a certain cognitive

state and the ’cognitiveStatePast’ functions describe

for how long (seconds) an actor has been in a certain

cognitive state.

The ﬁrst line of the equation describes that prefer-

ence of a actor-task set pair based on CTL is 1 if the

actor is not human or the actor’s CTL is not is a prob-

lem region. The second to ﬁfth line describe the pref-

erence to be in between 0.7/0.5 and 0.2/0, depending

on how long an actor has been in the corresponding

problem region (preference decreasing faster for over-

load and cognitive lock-up as they can occur faster

than other problem states). As cognitive lockup is

slightly less problematic than the other states the per-

son can be in, the preference associated therewith is

set somewhat higher.

Coordination Costs. The coordination costs have

to take into account two aspects of switching between

tasks, namely how much attention is needed to switch

to a new task set, and how often task reallocations

take place. The ﬁrst aspect covers how much atten-

tion is needed to switch to a new task set. The formal

description of this aspect is seen in Equation 2. If the

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

120

ctl

(hA,T i,t) =











1 if ¬isHuman

(A,t) or neutral

(hA,T i,t)

0.7 − 0.5 ∗ (min(300, cognitiveLockUpPast

(A,t))/300) if cognitiveLockUp

(hA,T i,t)

0.5 − 0.5 ∗ (min(300, overloadPast

(A,t))/300) if overload

(hA,T i,t)

0.5 − 0.5 ∗ (min(900, vigilancePast

(A,t))/900) if vigilance

(hA,T i,t)

0.5 − 0.5 ∗ (min(900, underloadPast

(A,t))/900) otherwise (if underload

(hA,T i,t))

Eq. 1: Formal description of the preference concept CTL. The function min(x,y) returns the lesser of its two arguments. All

parameters used here and in other formulas are based on relevant literature and were tweaked using data from pilot studies.

task set of an actor does not change, there are no co-

ordination costs, which is preferable (fourth line of

Eq. 2). If a task gets assigned to an actor that was not

previously assigned to this actor at all, this has a rel-

atively high cost (ﬁrst line). If a task gets assigned to

an actor that was previously assigned to this actor, but

at a different level of autonomy, there are two scenar-

ios. The level of autonomy of a robot could increase,

in this case the coordination costs for the human actor

are small (third line). If the level of autonomy of a

robot decreases, the cost is a bit higher as the human

actor has increased responsibilities (second line).

The second aspect that coordination costs have

to take into account is how often task reallocations

take place. Changing the level of autonomy too of-

ten could cause extra workload (Inagaki, 2003). The

formal description of this aspect is seen in Equation

3. The ﬁrst line describes that there is no effect if

the last task reallocation is more than 300 seconds

ago or if the task was already assigned to the actor at

the same autonomy level. The second line describes

that a task reallocation in the last 300 seconds gives a

penalty to the preference (the longer ago, the smaller

the penalty).

The full preference function for coordination costs

is seen in Equation 4. It deﬁnes preference based on

coordination costs of a actor-task set pair to be the

average preference based on coordination costs for all

separate tasks in the task set.

Utility Function. The utility function maps role

assignments at a certain point in time to their utility.

The utility of a role assignment is some combination

of all preference concepts, in this case the preference

based on CTL and the preference based on coordina-

tion costs (CC). Team performance beneﬁts from an

actor switching between different (levels of autonomy

of) tasks if the the negative effect on performance

of the cognitive state he is in outweighs the costs of

switching. The utility of a role assignment thus is

the preference of the role assignment based on CTL

minus the coordination costs. The preference concept

CC is high if the coordination costs are low (because

this is preferred) and vice versa. Therefore the

utility of a role assignment is the addition of the two

preference concepts CTL and CC. We deﬁne that the

lowest utility equals 0 and the highest utility equals 1.

To ﬁt this range, we scale the sum of the preference

concepts CTL and CC (which also both range from 0

to 1) by dividing it by two. More formally, the utility

of a role assignment (an option) O = hA,T i at time t

is: utility(O,t) = (ctl

(O,t) + cc

(O,t))/2

5 EXPERIMENT

An experiment was set-up to test if the model reallo-

cates tasks at the right moment and if it chooses the

appropriate reallocations. We instantiated the model

to be used for a single operator-single robot team in

the urban search and rescue domain. This involved

specifying tasks, possible levels of autonomy of these

tasks and task properties. Furthermore, we used an

existing model that calculates CTL speciﬁcally for the

urban search and rescue domain (Colin et al., 2014).

Experimental Method. Twelve participants (aged

21 to 38) completed three ﬁfteen minute sessions and

one participant performed a single session. Partici-

pants were given the role of robot operator and asked

to execute a typical urban search and rescue task. The

task was to explore a virtual ofﬁce building with a

virtual robot after an earthquake, and to map the situ-

ation in the building. This was done by navigating the

robot through the building and adding ﬁndings (large

obstacles and victims) to a tactical map, a screen shot

of the interface is seen in Figure 2. Sometimes infor-

mation appeared on the map (e.g., “We think there are

two people in this room.”). As there might be victims

in the building in need of medical attention, partici-

pants were told to hurry. The tasks were allocated to

the participants by the task allocation model: the op-

timal level of autonomy for the robot, as calculated

by the model, was chosen. Four tasks were speciﬁed:

navigation, obstacle recognition & avoidance, victim

recognition and information processing. The level of

autonomy of the robot could change separately for

each of these four tasks. During task execution, the

CTL of the participant was calculated. When the CTL

was in a problem region, the task allocation model

was run. If the task allocation model determined that

a task reallocation was needed, this new task alloca-

tion was communicated to the robot and its operator.

Results. In the experiment, we evaluated whether

DynamicTaskAllocationforHuman-robotTeams

121

attention

(hA,T

i,t) =











0 if ¬∃w : T

∈ currentTasks

(A,t)

0.2 if ∃w : T

∈ currentTasks

(A,t) ∧ v < w

0.5 if ∃w : T

∈ currentTasks

(A,t) ∧ v > w

1 otherwise (if ∃w : T

∈ currentTasks

(A,t) ∧ v = w )

Eq. 2: The function describing preference based on how much attention is needed for switching between tasks. The ’current-

Tasks’ function describes the set of tasks currently allocated to an actor.

time

(hA,T

i,t) =



attention

(hA,T

i,t) if reallocation(hA,T

i,t) ≥ 300 or cc

attention

(hA,T

i,t) = 1

max(0,cc

attention

(hA,T

i,t) − penalty) otherwise

where penalty = ((300 − reallocation(hA, T

i,t))/300) ∗ 0.25

Eq. 3: The preference function also taking into account how often task reallocations take place. The ’reallocation’ function

describes how long ago the last reallocation of a task was (in seconds).

(hA,T i,t) =











∑

∀T

∈T

time

(hA,T

i,t)

/|T | if isHuman

(A,t)

1 otherwise (A is a robotic or dummy actor)

Eq. 4: The full preference function describing preference based on the cost of switching between tasks.

Figure 2: A screen shot of the practice level. The left screen shows the building through the camera mounted on the robot. A

victim can be seen, accompanied by a number that could be used to look up information about the victim. The right screen

shows the tactical map. The circle on the left corresponds to the location of the robot, the trail to the driven route. Other items

shown on the map are (from left to right) a point of interest, a remark, a waypoint, an obstacle, a victim and a picture.

the participants thought that the task reallocations of

the model were done at the right time, whether the

task reallocations were thought to be appropriate, and

whether, after a task reallocation, the CTL of the par-

ticipants changed as predicted by the model.

Six statements about timing of reallocations were

given to participants after the experiment. Cronbach’s

alpha was used to check the internal consistency of

these six statements, which yielded 0,607. This is

quite low, but expected as the concept of timing is

rather broad and we use only six statements. The

average response over all six statements describes if

participants think the model reallocated tasks at the

right moment ranging from 1 (strongly disagree) to 5

(strongly agree). The average value over all partici-

pants is 2,65 (standard deviation 0,68). Participants

are thus quite neutral about the timing of the model.

We cannot say, based on this data, that the model re-

allocates tasks at the right moment. Conversely, we

also cannot say the timing of the model was fully off.

Five statements about the appropriateness of re-

allocations were given to participants after the exper-

iment. Cronbach’s alpha yielded 0,694. The aver-

age response over the ﬁve statements describes if par-

ticipants think the model chose appropriate task re-

allocations, ranging from 1 (strongly disagree) to 5

(strongly agree). Averaged out over all participants,

this value is 2,10 (standard deviation 0,39). Partic-

ipants are thus quite negative about the appropriate-

ness of the reallocations. We cannot say, based on

this data, that the model chooses appropriate reallo-

cations. Conversely, we can say participants think the

model does not choose appropriate reallocations.

The real shift in CTL was compared to the pre-

dicted shift in CTL for each task reallocation. This

comparison was done separately for the three metrics.

We checked whether the difference between the pre-

dicted CTL for the old and new task allocation is the

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

122

same as the difference between the average real CTL

in the two minutes before and after the reallocation.

This difference is calculated by subtracting the value

for the new task allocation from the value for the old

one. The correlation coefﬁcients are 0,32 (p < 0, 05)

for LIP, 0,43 (p < 0,01) for TO, and 0,29 (p = 0,06)

for TSS. The correlations for LIP and TO are signiﬁ-

cant (p < 0,05), the correlation for TSS is not. Based

on this data, we can validate that the LIP and TO re-

spond to the task reallocations as the model predicts.

6 DISCUSSION

Trust in Model. During the experiment, partici-

pants found it hard to trust the model and to have no

control over the task allocation. Making work agree-

ments could help improve trust as they give an op-

erator room to restrict which tasks can be done by

the robot(s) and when. Work agreements can also

give insight into what tasks actors can expect to be

reallocated and when reallocations occur. To further

give actors insight and even some inﬂuence, we could

adapt the level of automation of the task reallocation

model itself. A hybrid approach might be most suit-

able. The model could decide for high workloads

and suggest for low workloads (operator decides).

Furthermore, it beneﬁts trust if the actor has insight

into how the model chooses a task reallocation, e.g.,

through showing how options are rated. It needs to be

further speciﬁed and evaluated how the internal pro-

cesses of the model can be made visually available to

the user to improve his understanding and trust of the

model. In addition, future research on work agree-

ments and hybrid models is needed to investigate how

trust affects the effectiveness of the model.

Factors in Choosing a Task Allocation. CTL

is a very important factor in choosing a task alloca-

tion, but two possible additional factors were identi-

ﬁed during the experiment. The ﬁrst factor is the ca-

pability of an actor to do a task. A second factor is

the preference for particular tasks of the actor. Tak-

ing this into account could greatly beneﬁt actor trust

towards the model and reduce reluctance to accept its

decisions. Also, the actor is probably more likely to

execute a task well that he likes. Future research is

needed to explore the effects of including additional

factors such as capability and preference, both on the

trust and on the performance of the tasks.

Conﬁguration. The exact moment of a task allo-

cation relies on the conﬁguration of the CTL model.

Participants’ opinion about the timing of the task al-

location model will likely beneﬁt from personaliz-

ing conﬁguration of CTL problem region boundaries,

which was not done in the current experiment. Fu-

ture research should be executed to determine these

boundaries and to explore the effects of personal con-

ﬁguration. Conﬁguration poses additional challenges:

Results of experiments using task allocation models

with different conﬁgurations are hard to generalize

and conﬁguration takes a lot of time and effort. Ide-

ally, models will need to become self-learning, adapt-

ing themselves to novel tasks and actors when needed.

Representation and Notiﬁcation. This study did

not address how to communicate this task allocation

to the actors using the model. More research is needed

to investigate how to keep all actors aware of which

tasks are allocated to themand how to do this in the

most intuitive and understandable way.

7 CONCLUSION

A high-level framework for dynamic task alloca-

tion, aimed at improving team performance in mixed

human-robot teams, was presented. The framework

describes important concepts that inﬂuence team per-

formance and can be used to dynamically allocate

tasks. The framework applies to a wide array of prob-

lems, including heterogeneous teams that might in-

clude multiple human actors and multiple robots or

agents, a variety of tasks that might change over time

and complex and dynamic environments.

We used the framework as a basis for designing

a model for adaptive automation triggered by cogni-

tive task load. The framework was general and ﬂex-

ible enough to cover all aspects needed to formalize

the model, mainly cognitive task load (as a preference

factor) and adaptive automation (as dynamic task allo-

cation). We noticed that although cognitive task load

is an important factor, some other factors are also im-

portant, such as capability, preference and trust or per-

ceived capability. As the adaptive automation model

is based on the framework, it can be quite easily ex-

tended to include other factors, which will be done in

future work. The model addresses a wider range of

problems than most current adaptive automation re-

search, as it focuses on multiple tasks each with their

own variable level of autonomy.

We designed an experiment using the model, to

explore the effects of the resulting adaptive automa-

tion. The model was instantiated for a single hu-

man agent cooperating with a single robot in the ur-

ban search and rescue domain. An experiment was

conducted aimed at testing the model. The experi-

ment did not result in conclusive evidence that the

model worked as it should, but encouraging results

were found. Two of the three cognitive task load met-

DynamicTaskAllocationforHuman-robotTeams

123

rics (both the level of information processing and the

time occupied) of participants could be managed us-

ing the model. Furthermore, important focus points

for improving the model and furthering research on

adaptive automation in general were identiﬁed.

ACKNOWLEDGEMENTS

This research is supported by the EU-FP7 ICT Pro-

grammes project 247870 (NIFTi) and project 609763

(TRADR).

REFERENCES

Bailey, N. R., Scerbo, M. W., Freeman, F. G., Mikulka,

P. J., and Scott, L. A. (2006). Comparison of a brain-

based adaptive system and a manual adaptable system

for invoking automation. Human Factors: The Jour-

nal of the Human Factors and Ergonomics Society,

48(4):693–709.

Barnes, C. M., Hollenbeck, J. R., Wagner, D. T., DeRue,

D. S., Nahrgang, J. D., and Schwind, K. M. (2008).

Harmful help: The costs of backing-up behavior in

teams. Journal of Applied Psychology, 93(3):529.

Brannick, M. T., Salas, E., and Prince, C. (1997). Team

Performance Assessment and Measurement: Theory,

Methods, and Applications. Series in Applied Psy-

chology. Lawrence Erlbaum Associates.

Burke, J. L., Murphy, R. R., Rogers, E., Lumelsky, V. J., and

Scholtz, J. (2004). Final report for the DARPA/NSF

interdisciplinary study on human-robot interaction.

Systems, Man, and Cybernetics, Part C: Applications

and Reviews, IEEE Transactions on, 34(2):103–112.

Calhoun, G. L., Ruff, H. A., Spriggs, S., and Murray, C.

(2012). Tailored performance-based adaptive levels

of automation. Proceedings of the Human Factors and

Ergonomics Society, 56(1):413–417.

Colin, T., Mioch, T., Smets, N., and Neerincx, M. (2014).

Real time modeling of the cognitive load of an urban

search and rescue robot operator. In The 23rd IEEE

International Symposium on Robot and Human Inter-

active Communication (RO-MAN2014).

Fitts, P. M., Viteles, M. S., Barr, N. L., Brimhall, D. R.,

Finch, G., Gardner, E., Grether, W. F., Kellum,

W. E., and Stevens, S. S. (1951). Human engineering

for an effective air-navigation and trafﬁc-control sys-

tem. Technical report, Ohio State University Research

Foundation Columbus.

Gerkey, B. P. and Matari

c, M. J. (2004). A formal analysis

and taxonomy of task allocation in multi-robot sys-

tems. The International Journal of Robotics Research,

23(9):939–954.

Greef, T. E. d., Arciszewski, H. F. R., and Neerincx, M. A.

(2010). Adaptive automation based on an object-

oriented task model: Implementation and evaluation

in a realistic C2 environment. Journal of Cognitive

Engineering and Decision Making, 4(2):152–182.

Guzzo, R. A., Salas, E., and Goldstein, I. L. (1995). Team

Effectiveness and Decision Making in Organizations.

Jossey-Bass San Francisco.

Hilburn, B., Molloy, R., Wong, D., and Parasuraman, R.

(1993). Operator versus computer control of adaptive

automation. In The Adaptive Function Allocation for

Intelligent Cockpits (AFAIC) Program: Interim Re-

search Guidelines for the Application of Adaptive Au-

tomation, pages 31–36. DTIC Document.

Ilgen, D. R., Hollenbeck, J. R., Johnson, M., and Jundt, D.

(2005). Teams in organizations: From input-process-

output models to IMOI models. Annual Review of

Psychology, 56:517–543.

Inagaki, T. (2003). Adaptive automation: Sharing and trad-

ing of control. In Handbook of Cognitive Task Design,

pages 147–169. Mahwah, NJ: Lawrence Erlbaum As-

sociates.

Kidwell, B., Calhoun, G. L., Ruff, H. A., and Parasuraman,

R. (2012). Adaptable and adaptive automation for su-

pervisory control of multiple autonomous vehicles. In

Proceedings of the Human Factors and Ergonomics

Society, pages 428–432.

Matthew, G., Davies, D. R., Westerman, S. J., and Stam-

mers, R. B. (2000). Human Performance: Cognition,

Stress, and Individual Differences. Psychology Press.

McGrath, J. E. (1964). Social Psychology: A Brief Intro-

duction. Holt, Rinehart and Winston.

Moray, N., Inagaki, T., and Itoh, M. (2000). Adaptive au-

tomation, trust, and self-conﬁdence in fault manage-

ment of time-critical tasks. Journal of Experimental

Psychology: Applied, 6(1):44.

Neerincx, M. A. (2003). Cognitive task load design: Model,

methods and examples. In Handbook of Cognitive

Task Design, pages 283–305. Mahwah, NJ: Lawrence

Erlbaum Associates.

Neerincx, M. A., Kennedie, S., Grootjen, M., and Grootjen,

F. (2009). Modeling the cognitive task load and per-

formance of naval operators. In Schmorrow, D., Es-

tabrooke, I., and Grootjen, M., editors, Foundations of

Augmented Cognition, volume 5638 of LNCS, pages

260–269. Springer Berlin / Heidelberg.

Parasuraman, R., Cosenzo, K. A., and Visser, E. d. (2009).

Adaptive automation for human supervision of multi-

ple uninhabited vehicles: Effects on change detection,

situation awareness, and mental workload. Military

Psychology, 21(2):270–297.

Steinfeld, A., Fong, T., Kaber, D., Lewis, M., Scholtz, J.,

Schultz, A., and Goodrich, M. (2006). Common met-

rics for human-robot interaction. In Proceedings of the

Conference on Human-Robot Interaction, pages

33–40. ACM.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

124