Man-Machine Teaming: AI’s Overload Management and Task Allocation

Weifei Xie

1,2,3 a

, C

edric Herpson

, Christophe Cambier

and Jean-Marie Dembelle

Sorbonne Universite, CNRS, LIP6, F-75005 Paris, France

Sorbonne Universite, IRD, UCAD,UGB UMI UMMISCO, F-75006 Paris, France

Universite Gaston Berger de Saint-Louis, Laboratoire d’Analyse Num

erique et Informatique, BP. 234 Saint Louis, Senegal

ﬁ ﬁ ﬁ

Keywords:

Task Planning, Mental Workload.

Abstract:

Sustained concentration can induce cognitive overload in demanding roles such as surgeons and soldiers,

prompting research into workload management via surveys, performance metrics, and physiological mea-

sures. However, these conventional methods face challenges like time-intensive survey processes, unreliable

metrics, and equipment dependency, complicating their application in projects like Recolte, which develops

partially autonomous drone ﬂeets for data collection.To overcome these challenges, we designed an AI-Human

task distribution algorithm that assesses mental workload considering task complexity, duration, and operator

competency. This algorithm also incorporates factors inﬂuencing recovery speed, including rest type, operator

state, and environmental conditions. We assessed four recovery models to determine their suitability for our

use case, employing AI for routine tasks to mitigate human workload. Our algorithm aims to optimize team

composition by determining the optimal timing for integrating additional human or advanced AI resources to

ensure mission success. Empirical ﬁndings from this study provide insights into the recovery models’ impact

on operational effectiveness, facilitating the analysis of success rates across different task conﬁgurations and

operator settings. This method ensures continuous human training, even with the presence of AI operators for

speciﬁc tasks.

1 INTRODUCTION

Maintaining a high level of concentration for ex-

tended periods can lead to cognitive overload,

whether for a surgeon, a soldier in combat, or a driver.

Cognitive overload can have a negative impact on

performance, employee productivity, and may even

lead to mission failure (Li et al., 2020). To prevent

overload and optimize performances, it’s necessary

to monitor or estimate worker’s mental workload and

to plan recovery periods. In the Recolte project, we

aim to develop a ﬂeet of semi-autonomous drones

for long-term data collection using different platforms

and both artiﬁcial and human pilots. In this context,

our goal in this paper is to measure the impact of the

various recovery models on a mission and to study the

evolution of a mission success rate according to spe-

ciﬁc tasks and operators conﬁgurations.

In the literature, methods for monitoring mental

workload can be divided into three categories: subjec-

tive measurement using surveys, physiological mea-

surement using equipment and performance measure-

https://orcid.org/0009-0008-8866-9855

ment. It is important to note that these categories

are not mutually exclusive. One disadvantage of sur-

veys is their limited applicability in continuous mis-

sions. Physiological measurements require special-

ized equipment not easily usable during a mission.

Performance measurement can be further divided into

primary task measurement and secondary task mea-

surement. Both have several disadvantages, including

insensitivity to variations in mental load, sensitivity

to differences between individuals, unreliability of the

secondary task limit and the addition of extra mental

load. The obstacles mentioned make it difﬁcult to di-

rectly apply these methods for long-term continuous

missions. Rather than monitoring mental workload,

we thus choose to estimate it using mental workload

and recovery models combined with a proposed task

allocation algorithm. Our algorithm also considers

the amount of training required for each mission to

be successful, which is not always a concern in tra-

ditional work-rest balancing planning. Additionally,

our task allocation algorithm is able to test whether

it is possible to execute the mission without cognitive

overload, using both human and AI pilots.

Xie, W., Herpson, C., Cambier, C. and Dembelle, J.

Man-Machine Teaming: AI’s Overload Management and Task Allocation.

DOI: 10.5220/0012984900004562

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Cognitive Aircraft Systems (ICCAS 2024), pages 95-103

ISBN: 978-989-758-724-5

In section 2 we present the general literature of

mental workloads and recovery models as well as as-

sociated scheduling algorithms. As workload can be

inﬂuenced by various factors, section 3 presents four

mental workload and recovery models grounded in

different strand of the literature and adapted to our

application context. We then present our scheduling

algorithm that combines both workload and recovery

criteria in section 4. In section 5 we discuss experi-

mental results comparing the behaviours of the differ-

ent models before concluding.

2 STATE OF THE ART

Xie and Salvendy (2000) suggests that the evaluation

of mental workload should consider several aspects,

including instantaneous workload, peak workload,

average workload, accumulated workload, and overall

workload. Instantaneous workload reﬂects workload

ﬂuctuations and is typically measured through phys-

iological metrics. Peak workload, which represents

the highest level of instantaneous workload, is criti-

cal for identifying potential mental overload. Accu-

mulated workload is the total workload experienced

during a task, calculated from instantaneous work-

load measurements. Average workload is the mean of

all instantaneous workload values. Overall workload

can be assessed using a mapping function based on

instantaneous workload or another function that con-

siders both average and accumulated workload. These

mapping functions consider the complexity of the task

and individual differences, providing a comprehen-

sive view of the mental demands placed on profes-

sionals in high-stakes environments.

peak

= MAX(W

(t)) (1)

accumulated

(t) =

(u)du (2)

average

= W

accumulated

(t)/t (3)

(4)

overall

= f

(t)) = f

average

(t),W

accumulated

(t))

(5)

In the context of multitasking, it is assumed that

the cumulative workload associated with simultane-

ous tasks is additive in nature (Xie and Salvendy,

2000), as represented by the equation:

overall-multi

∑

i=0

overall-i

(6)

This formulation, however, does not account for

the potential mitigating effects of interspersed breaks

between tasks. Extensive research has demonstrated

that breaks can signiﬁcantly facilitate recovery from

high workload levels. Consequently, a revised model

incorporating the impact of breaks is proposed:

overall

∑

i=0

overall-i

∑

j=0

break-j

(7)

In this revised model, R

break-j

represents the reduc-

tion in mental workload attributed to the jth break.

The efﬁcacy of these breaks in reducing mental work-

load is inﬂuenced by various factors, including the

nature of the relaxation activity undertaken during

breaks, the surrounding environment, and the dura-

tion of the break, among others.

For our project, ”Recolte,” the practical challenges

of deploying physiological measurement equipment

for each pilot to directly assess workload necessitate

an alternative approach. As such, we aim to approxi-

mate the workload through an estimation based on the

workload associated with individual tasks.

2.1 Mental Workload Accumulation

In the study conducted by Xie and Salvendy (2000),

it was articulated that the mental workload is predom-

inantly inﬂuenced by the task at hand and the individ-

ual involved. Further elaboration by Hancock et al.

(2021) elucidated that the mental workload is modu-

lated by the chosen work procedure, which is inﬂu-

enced by four pivotal factors: the nature and com-

plexity of the task, the degree of autonomy and the

scope for decision-making allowed by the situational

context, the operator’s level of expertise, and the op-

erator’s current mental state. In the cognitive domain,

processing can be categorized into three distinct lev-

els (Drenth, 1998):

1. Skill-Based Processing: Automatic and highly

practiced actions requiring minimal thought.

They are fast and efﬁcient but susceptible to

attention-related errors.

2. Rule-Based Processing: Applies learned rules to

somewhat familiar situations. They are more

thoughtful than skill-based; useful for semi-

routine situations.

3. Knowledge-Based Processing Engages in novel

situations requiring deep thought and problem-

solving. They are highly ﬂexible and powerful but

slow and demands signiﬁcant mental resources.

We can thus conclude that with different expertise

the same task can conduct to different workload.

ICCAS 2024 - International Conference on Cognitive Aircraft Systems

2.2 Mental Workload Recovery

Numerous studies have demonstrated that incorpo-

rating breaks into the work schedule can signiﬁ-

cantly enhance both employee health and perfor-

mance. This practice can lead to increased produc-

tivity with reduced investment, ultimately beneﬁting

organizational efﬁciency. Albulescu et al. (2022) dis-

cusses how work demands deplete psychological re-

sources, which can be replenished during periods of

rest such as sleep, weekends, and vacations. The au-

thor also notes that recovery can occur during shorter

periods of downtime, including lunch breaks, work

breaks, and even brief micro-breaks. Caldwell et al.

(2009) highlights napping as a highly effective non-

pharmacological method for enhancing alertness in

ﬂight pilots. The research also points out the time

required for falling asleep and waking up within des-

ignated nap durations, citing a NASA study which de-

termined that an actual sleep time of 26 minutes out

of a 40-minute nap period is typical. This observation

indicates that napping may not be practical for micro-

breaks because of the associated time constraints, thus

rendering it unsuitable for short breaks. Additionally,

research by Brazaitis and Satas (2023) demonstrates

that a 10-minute break following 50 minutes of work

does not sufﬁciently mitigate cognitive fatigue, sug-

gesting that such intervals are inadequate for prevent-

ing cognitive decline during longer tasks. Albulescu

et al. (2022) demonstrated that shorter, more frequent

breaks can effectively reduce fatigue levels without

adversely affecting productivity.

Tucker (2003) suggests that the type of relaxation

activity undertaken during a break can inﬂuence the

extent of recovery achieved. For instance, engaging in

activities such as taking a short nap (less than 15 min-

utes) or drinking coffee may be more effective at com-

bating sleepiness compared to simply taking a break.

Hoover et al. (2022) conducted an experimental com-

parison of two types of break activities: physical (sta-

tionary biking) and relaxation (progressive muscle re-

laxation). Their ﬁndings indicate that relaxation ac-

tivities are more effective in promoting psychological

detachment and relaxation, while physical activities

are superior in replenishing energy levels.

2.3 Work-Rest Balance Scheduling

Another critical approach to preventing overload in-

volves strategic task scheduling. It is essential to de-

velop a plan that avoids overload while simultane-

ously not compromising work efﬁciency. In Jaber

and Neumann (2010) introduced a mixed-integer lin-

ear programming (MILP) model designed to address

task scheduling issues, incorporating the impact of

human fatigue into the modeling process. The goal

of this MILP model is to enhance productivity while

reducing the physical strain on workers. However,

solving such a MILP problem remains a challenging

task. Zhang et al. (2022) studied task scheduling in

a human-robot collaborative assembly cell with the

goal of balancing job cycle time and human fatigue.

They developed an MILP model for this issue, using

job cycle time as the objective function and imposing

a constraint on maximum human fatigue. To address

the complexity of solving this MILP model, they sug-

gested the use of a genetic algorithm. But to our

knowledge there is currently no task scheduling algo-

rithm taking into account both task difﬁculty and the

AI ability for the scheduling. Studies such as Zhang

et al. (2022) and Jaber and Neumann (2010) propose

that speciﬁc types of work consistently impact hu-

mans in the same manner; however, their approach

does not align with the Project Recolte. For example,

in this project, the complexity of aerial tasks is inﬂu-

enced by factors like wind conditions, the presence

of obstacles, and the number of turns required. Fur-

thermore, whereas other studies may prescribe ﬁxed

tasks for robots, the Recolte project focuses on adapt-

ing tasks based on their difﬁculty rather than deﬁning

them by type.

To meet this need, we propose an algorithm that

takes into account the type of rest activity, operator

capabilities and task complexity. Our algorithm also

must also take into account the fact that the interven-

tion of AI shoud not negatively affect human learning.

3 PROPOSED MENTAL

WORKLOAD ACCUMULATION

AND RECOVERY MODELS

Drenth (1998) highlighted that mental workload accu-

mulation is affected by the environment, the task, and

the operator. In the context of Recolte, where the op-

erator’s environment remains relatively constant dur-

ing the mission, the primary factors inﬂuencing men-

tal workload accumulation are the task and the op-

erator. Consequently, we categorize the ﬂying task

into three levels of difﬁculty—easy, medium, and dif-

ﬁcult—based on criteria such as the number of turns

relative to the task area, the presence of a guiding rou-

tine, and whether the map is known or unknown. In

addressing the operator dimension of Recolte’s mis-

sion to make drones widely accessible, we recognize

the signiﬁcant variability in operator expertise. Ac-

cordingly, we classify operators into three distinct cat-

Man-Machine Teaming: AI’s Overload Management and Task Allocation

egories based on their skill level: novice, interme-

diate, and advanced. This stratiﬁcation allows us to

appropriately match task difﬁculty with operator ca-

pability, facilitating both efﬁcient and effective drone

operation.

3.1 Mental Workload Accumulation

Model

We propose a mathematical model to quantify the

mental workload induced by task i on operator j as

follows:

= duration

∗ coe f

i j

(8)

• mw

represents the mental workload generated by

task i for operator j.

• duration

denotes the estimated duration of task i

• coe f

i j

is a predeﬁned coefﬁcient that reﬂects the

difﬁculty of task i adjusted for the skill level of op-

erator j. This coefﬁcient ensures that the level of

the operator is appropriately matched to the task

difﬁculty, such as assigning only difﬁcult tasks to

advanced operators.

3.2 Mental Workload Recovery Model

As introduced in section 2, recovery from mental

workload can be inﬂuenced by various factors includ-

ing the operator’s current state, the type of relaxation

activity engaged in, and the duration of the recovery

period. Based on this understanding, we developed

four models, each grounded in different strands of lit-

erature:

1. Linear Model: This model posits a direct, propor-

tional relationship between the intensity and du-

ration of the workload and the required recovery

time. It is based on the principle that recovery

time increases linearly with increases in the cogni-

tive demands of the task. This model can be espe-

cially useful for tasks with predictable and consis-

tent cognitive loads.(Asadayoobi et al., 2023) The

model is deﬁned by the equation R = ε ∗ duration

where ε denotes the recovery speed, duration is

the length of the break , R measures the recovery

of mental workload.

2. Linear Model With Delays: Inspired by Cald-

well et al. (2009), our model recognizes that the

body requires time to initiate the recovery process,

such as the time needed to fall asleep when using

napping as a recovery activity.The model is de-

ﬁned by the equation R = ε ∗ (duration − delay) if

duration > delay or R = 0 if duration <= delay

(delay is a predeﬁned constant).

3. Exponential Model: An exponential model inﬂu-

enced by pre-break mental workload.(Jaber and

Neumann, 2010) The model is deﬁned by the

equation R = F ∗ exp

µ∗duration

, F is pre-break

mental workload level.µ is constant.

4. Quadratic Model: Albulescu et al. (2022)sug-

gested that micro-breaks might be more efﬁcient

than extended breaks. In response, we propose a

quadratic model at

+ bt + c to explore this rela-

tionship. Here, a,b,c are coefﬁcients that shape

the recovery function. We set a negative to model

scenarios where shorter breaks yield better recov-

ery, implying a decrease in recovery beneﬁt as

break duration increases. Conversely, to exam-

ine conditions where longer breaks could be more

advantageous, we make a positive, indicating in-

creased recovery efﬁciency with longer break du-

rations.

4 PROPOSED SCHEDULING

ALGORITHM

Having developed models for both mental workload

accumulation and recovery, it is crucial to consider

Recolte’s objective of enabling novice operators to ef-

fectively execute missions. To achieve this, integrat-

ing artiﬁcial intelligence (AI) into the drone systems

to provide necessary assistance becomes essential.

For safety concerns, our algorithm should be ca-

pable of assessing the feasibility of missions given

the human operator’s expertise and the support of AI.

Speciﬁcally, the algorithm needs to evaluate:

1. Mission Feasibility: Assess if the human operator

can complete the mission with AI support without

suffering cognitive overload.

2. Risk of Overload: Identify potential mental over-

load risks during the mission.

3. Recovery Model Selection: Recommend an ap-

propriate recovery model based on observed stress

levels and task nature, choosing suitable relax-

ation activities that align with the mission timeline

and operator needs.

4. Dynamic Adjustments: Enable the system to dy-

namically adjust risk assessments and recovery

plans as mission parameters change, such as vari-

ations in task complexity or assignment.

The algorithm should also have the capability to

train human operators, for example, by enabling them

to execute tasks proactively while minimizing the risk

of overload. From there we propose Algorithm (1)

ICCAS 2024 - International Conference on Cognitive Aircraft Systems

: Distribution

below and introduce the associated

notations :

• W

i j

: the mental workload at the start of task i of

the operator j

• W

i j

: the mental workload at the end of task i of

the operator j

• W n

i j

: the mental workload will be needed for

complete the rest task attribute from task i for the

operator j. If there is no task assign after task i,

W n

i j

= W

i j

. If there are tasks assigned to operator

j after task i, identify the task c that is the closest

subsequent task to task i, then W n

i j

= W

i j

+W n

c j

• n

number of preset difﬁculty levels

• d

difﬁculty of task i(easy, medium, and difﬁcult

for 1,2,3)

• l

level of operator j (novice, intermediate, and

advanced for 1,2,3)

• W

i j

mental workload generate by task i for the op-

erator j

• W max

the maximum mental workload an oper-

ator can tolerate without negatively inﬂuencing

their performance

In this algorithm, we initially assign tasks to oper-

ators such that the difﬁculty of each task corresponds

with the operator’s level of expertise—easy tasks to

novice operators, medium tasks to intermediate oper-

ators, and difﬁcult tasks to advanced operators, ensur-

ing not to exceed their maximum workload capacities.

For tasks that remain unassigned, we proceed to allo-

cate these to higher-level operators. However, such

assignments must comply with the speciﬁed equation

W max

− W

> W n

(where the a present operator,

tpresent the task) to ensure that the addition of any

new task does not cause the operator to surpass their

tolerable workload threshold. This approach guar-

antees that all tasks are appropriately assigned while

maintaining operational efﬁciency and preventing op-

erator overload.

5 EXPERIENCE

5.1 Common Parameter

In this section we explore our model through a case

study inspired by the Project Recolte. Recolte aims to

develop a ﬂeet of partially autonomous solar-powered

Source code : https://github.com/xiewf2019/distribut

ion

Algorithm 1: Distribution.

1: Input: agentPool taskPool, recovryModel

2: Output: tabDistribution, riskOverload

3: divide both the agent pool and the task pool based

on their respective levels and difﬁculty such that

task level i is in taskPool[i]

4: riskOverload = False

5: for i in rangen

6: for task t in taskPool[i] do

7: for agent a in agentPool[i] do

8: calculate W

9: if W

< W max

assign the task

to agent a, delete the task in taskPool[i], update

break

10: end for

11: end for

12: end for

13: for i in rangen

14: update workload necessary list for agent in

agentPool[i]

15: end for

16: for i in rangen

17: for task t in taskPool[i] do

18: for j in rangei + 1,n

19: for agent a in agentPool[ j] do

20: calculate W

21: calculate W n

22: if thenW max

− W

> Wn

23: assigning the task to agent a,

delete the task in taskPool[i], update W

break

24: end if

25: end for

26: end for

27: end for

28: for i in rangen

29: update workload necessary list for agent

in agentPool[i]

30: end for

31: end for

32: for i in rangen

33: if exist task in taskPool[i] then

riskOverload = True

34: end for

35: return tabDistribution,riskOverload

drones designed for prolonged data collection involv-

ing a combination of various platforms and human pi-

lots. Speciﬁcally, we analyze the performance of a

drone undertaking 100 consecutive ﬂying tasks, each

lasting approximately 15 minutes without interrup-

tion. The difﬁculty of each task is randomly assigned

in each trial according to a predeﬁned probability dis-

tribution.

For the operator component, we have examined

Man-Machine Teaming: AI’s Overload Management and Task Allocation

various conﬁgurations to address the following ques-

tions with application of our scheduling algorithm:

• What are the beneﬁts of integrating an AI, even

at a novice level, compared to relying solely on a

human operator?

• Which recovery model is most suitable for this

scenario?

• If an AI and a human operator with equivalent

skill levels are both involved, will the human op-

erator receive sufﬁcient training?

In the calculations that follow, the units utilized

within the accumulation and recovery model are ex-

pressed as percentages. For instance, if the opera-

tor accumulates a value of 50, this indicates that they

have reached 50% of their maximum capacity. For the

mental workload accumulation coefﬁcients, inspired

by the work in Drenth (1998), we have considered

scenarios where the operator’s skill level is lower than

the task difﬁculty, rendering the operator unable to

perform the task. In USA Air Carrier Operations Sur-

vey, they mentioned that after 2 hour operation needs

a break, so we consider that for a novice pilot we need

break after 1 hour operation. The coefﬁcients are de-

termined based on the indices i and j, as speciﬁed be-

low.

Table 1: Mental workload accumulation coefﬁcients based

on human level jand task difﬁcultyi.

1 2 3

1 1.5 1.125 0.75

2 - 2 1.5

3 - - 3

With this coefﬁcient we can assumed that after one

hour operation the novice operator will be consider

as overload. Building on the research presented in

Hoover et al. (2022), and recognizing that drone ﬂight

involves greater stress than the e-tasks referenced in

Hoover et al. (2022),according their result they con-

sider that 15 minute break is enough for 30 minute

e-task work, as drone ﬂight is more stress than e-task

so we consider that we need 45 minute to recover for

30 minute ﬂight we have accordingly established the

numerical values for the recovery models:

• µ for exponential model: 0.096 , we take the value

of work Jaber and Neumann (2010) (model 7 in

experience)

• a, b,c for quadratic model: 1/15,0, 0 (model 1

in experience) , 1/30, 0.5,0 (model 2 in expe-

rience), −1/30, 1.5,0 (model 4 in experience),

−1/15,2,0 (model 5 in experience)

• delay for delay model: 3 (model 6 in experience)

• ε for linear model: 1 (model 3 in experience)

For the aspect of mental workload capacity, unlike

other systems, we assume that everyone reaching their

limit will be overloaded. We acknowledge that main-

taining optimal capacity can be challenging, and for

safety reasons, we consider a person to be overloaded

if they reach 80% of their capacity. So we set all the

operator has 80 as mental workload capacity.

The simulation creates a scenario where operators

completes 100 consecutive tasks. In this model, each

task is performed by an operator whose capacity must

not be less than the requirements of the task. If the

assigned operator for a task is at risk of overload, i.e.,

the sum of the cumulative mental workload and the

expected workload exceeds a predetermined thresh-

old and no other operator is available to replace him

or no available operator is capable of completing the

task, the task is considered to be a failure due to the

potential risk of overload. To avoid the effects of ran-

domness and to assess the impact of task difﬁculty

distributions on the results, 100 tasks with speciﬁed

difﬁculty distributions were generated. The algorithm

was then applied to each set of tasks to assess the risk

of overload. Each distribution was repeated 100 times

to ensure that it was not affected by random varia-

tions.

5.2 Experience and Result

To answer aux questions, in the initial simulation, we

compared three conﬁgurations:

• An advanced-skilled individual paired with a

novice.

• An advanced-skilled individual paired with a

novice AI.

• An advanced-skilled individual, a beginner indi-

vidual, and a beginner AI.

The ﬁrst two conﬁgurations assess the impact of re-

placing a human with an AI, while the third conﬁg-

uration explores the effects of introducing an AI into

the scenario. In the referenced ﬁgure (1), the y-axis

Figure 1: The result with 100 simulation, one human oper-

ator level advanced, one human level novice.

ICCAS 2024 - International Conference on Cognitive Aircraft Systems

100

Figure 2: The result with 100 simulation, one human oper-

ator level advanced, one AI level novice.

represents the probability distribution across different

levels of task difﬁculty, categorized as novice, inter-

mediate, and advanced. For instance, the distribution

[0.9, 0.05, 0.05] signiﬁes a 90% probability of the task

being classiﬁed as novice level, a 5% probability for

intermediate, and a 5% probability for advanced level.

The x-axis delineates various types of recovery mod-

els. The numbers displayed on the grid indicate the

frequency, out of 100 simulations, at which the oper-

ator is unable to manage the tasks, signifying a risk

of overload. For example, a value of 54 means that

in 100 simulations, there were 54 instances where the

risk of overload was present.

To address the ﬁrst question, a comparison of

Figures 1 and 2 clearly indicates that integrating AI

enhances the likelihood of success in simulations.

Speciﬁcally, in a scenario where the task difﬁculty

distribution is [0.9,0.05,0.05] and Model 1 serves as

the recovery model, the AI-integrated conﬁguration

results in 30 failures, compared to 60 failures in the

non-AI conﬁguration. This trend is consistent across

various cases, with AI-assisted conﬁgurations outper-

forming those without AI. The beneﬁts of AI integra-

tion are particularly evident when the probability of

encountering difﬁcult tasks is low, signiﬁcantly reduc-

ing the rate of task failures.

In Figure 1 and Figure 2, we observe that Mod-

els 1 and 2 exhibit fewer instances of overload risk.

Speciﬁcally, Figure 1 shows that in 100 simulations,

Recovery Models 1 and 2 register 60 and 71 cases

of overload risk, respectively, while all other models

consistently reach 100 cases of overload. Similarly,

Figure 2 conﬁrms that the risk of overload for Models

1 and 2 is lower compared to other recovery models.

This indicates a potential advantage in the effective-

ness of Models 1 and 2 in managing workload to pre-

vent overload. we conclude that quadratic model with

a positive quadratic coefﬁcient is mostly efﬁcient in

our simulation case. And quadratic model with a neg-

ative a are not suitable for this situation.

Our preliminary ﬁndings suggest that the inclu-

sion of AI novices could potentially reduce the risk

of overload, particularly when the probability of en-

countering difﬁcult tasks is relatively low.

To answer the question if both an AI and a human

operator of equivalent skill levels are involved, will

the human receive adequate training, we conducted a

comparative analysis between Figure 1 and Figure 2.

This analysis focused on the proportion of tasks per-

formed by novice operators in two scenarios: one in-

volving both human novices and advanced operators,

and the other incorporating human advanced opera-

tors, and AI novices.

Figure 3: The portion of task assigned to human novice for

the case human novice + advanced.

Figure 4: The portion of task assigned to human novice for

the case human novice + advanced + AI novice.

In ﬁgures 3 and ﬁgure 4, both the y-axis and x-axis

are identical with ﬁgure 1. The numbers displayed on

the grid represent the proportion of tasks assigned to

a human novice in simulations where there is no risk

of overload. For instance, a value of 54 indicates that,

out of 100 tasks, 54 are assigned to the human novice.

In comparing Figure 3 with Figure 4, it becomes ev-

ident that integrating AI not only maintains the need

for training novice operators but also enhances their

capacity to handle more complex missions. This ob-

servation is further supported when analyzing Figure

1 and Figure 2, which demonstrate that the presence

of AI enables less experienced operators to undertake

more challenging tasks. This adaptation indicates a

substantial improvement in both operational capacity

and training efﬁciency, as AI assists in managing and

mitigating the potential challenges faced by novice

operators in high-demand scenarios.

Man-Machine Teaming: AI’s Overload Management and Task Allocation

101

5.3 Summary of Findings for Each

Question

1. Regarding the ﬁrst question, the results indicate

that integrating an AI system proves beneﬁcial in

reducing cognitive overload.

2. Concerning the second question, our analysis re-

veals that a quadratic model with positive coefﬁ-

cients is most appropriate for our case.

3. The results from the third question indicate that

integrating AI into the workﬂow does not change

the quantity of tasks assigned to trainees. This

demonstrates that the inclusion of AI has no ad-

verse effects on the volume of operational training

tasks.

5.4 Discussion

Our experimental ﬁndings afﬁrm the critical role

of AI integration in mitigating cognitive overload.

Nevertheless, addressing overload requires a multi-

dimensional approach that extends beyond AI facil-

itation to include ﬂexibility, learning capabilities, and

unforeseen event management. Crucially, maintain-

ing comprehensive training for human pilots empha-

sizes the imperative of human oversight.

We have identiﬁed a quadratic model with posi-

tive coefﬁcients as optimal for our experimental con-

text. Yet, real-world applications necessitate consid-

eration of environmental impacts, feasibility, and ben-

eﬁts associated with downtime activities. Our algo-

rithm not only supports AI assistance but also strate-

gically manages the workload on trainees—essential

in training-centric scenarios. To bolster security and

maintain operational readiness, an AI-ﬁrst strategy

is recommended, ensuring human operators are pre-

pared to intervene when AI limitations surface.

6 CONCLUSION

This paper studied the tasks scheduling of a Human-

AI collaborative drone-piloting mission. A novel task

scheduling algorithm integrating micro-breaks during

the mission was proposed to schedule necessary re-

covery time for human workers during their working

hours to deal with the human mental workload over-

load while while fulﬁlling the designated mission.

Experiments show that our method allows to integrate

AI in order to reduce the risk of overload while pre-

serving the training of human pilots. Regarding the

recovery model, results require to be tested with ob-

servation from real-life situations.

REFERENCES

Albulescu, P., Macsinga, I., Rusu, A., Sulea, C., Bodnaru,

A., and Tulbure, B. T. (2022). ”Give me a break!”

A systematic review and meta-analysis on the efﬁcacy

of micro-breaks for increasing well-being and perfor-

mance. PLOS ONE, 17(8):e0272460.

Asadayoobi, N., Taghipour, S., and Jaber, M. (2023). Opti-

mising stochastic task allocation and scheduling plans

for mission workers subject to learning-forgetting,

fatigue-recovery, and stress-recovery effects. Expert

Systems with Applications, 229:120524.

Brazaitis, M. and Satas, A. (2023). Regular short-duration

breaks do not prevent mental fatigue and decline in

cognitive efﬁciency in healthy young men during an

ofﬁce-like simulated mental working day: An EEG

study. International Journal of Psychophysiology,

188:33–46.

Caldwell, J. A., Mallis, M. M., Caldwell, J. L., Paul, M. A.,

Miller, J. C., and Neri, D. F. (2009). Fatigue Counter-

measures in Aviation. Aviation, Space, and Environ-

mental Medicine, 80(1):29–59.

Drenth, P. J. D., editor (1998). Organizational psychology.

Number 4 in Handbook of work and organizational

psychology / ed. by Pieter J. D. Drenth. Psychology

Press, Hove.

Hancock, G., Longo, L., Young, M., and Hancock, P.

(2021). MENTAL WORKLOAD. In Salvendy, G. and

Karwowski, W., editors, HANDBOOK OF HUMAN

FACTORS AND ERGONOMICS, pages 203–226. Wi-

ley, 1 edition.

Hoover, C. S., Ragsdale, J. M., and Ayres, T. B. (2022).

An experimental test of resource recovery from phys-

ical and relaxation work break activities. Stress and

Health, 38(3):477–489.

Jaber, M. Y. and Neumann, W. P. (2010). Modelling worker

fatigue and recovery in dual-resource constrained sys-

tems. Computers & Industrial Engineering, 59(1):75–

84.

Li, K., Xu, S., and Fu, H. (2020). Work-break schedul-

ing with real-time fatigue effect and recovery. Inter-

national Journal of Production Research, 58(3):689–

702.

Tucker, P. (2003). The impact of rest breaks upon acci-

dent risk, fatigue and performance: A review. Work &

Stress, 17(2):123–137.

Xie, B. and Salvendy, G. (2000). Prediction of Men-

tal Workload in Single and Multiple Tasks Envi-

ronments. International Journal of Cognitive Er-

gonomics, 4(3):213–242.

Zhang, M., Li, C., Shang, Y., Huang, H., Zhu, W., and Liu,

Y. (2022). A task scheduling model integrating micro-

breaks for optimisation of job-cycle time in human-

robot collaborative assembly cells. International Jour-

nal of Production Research, 60(15):4766–4777.

ICCAS 2024 - International Conference on Cognitive Aircraft Systems

102

APPENDIX

Experience and Result 2

We also have compared other conﬁgurations:

• human advanced, human intermediate

• human advanced, AI intermediate

Figure 5: The result with 100 simulation, one human oper-

ator level advanced, one human operator level intermediate.

Figure 6: The result with 100 simulation, one human oper-

ator level advanced, one AI operator level intermediate.

Comparing Figure 2 and Figure 5, it is appar-

ent that a novice AI proves more beneﬁcial than an

intermediate human. Additionally, when assessing

Figure 2 and Figure 6, the beneﬁts of high level AI

in terms of risk avoidance become more pronounced

as the probability of encountering difﬁcult tasks in-

creases. While the improvement may not be signif-

icant at lower probabilities, the value of high level

AI in managing higher-difﬁculty tasks becomes in-

creasingly crucial as the likelihood of such challenges

grows. We don’t test scenarios involving the insertion

of an advanced AI because it’s clear that such an AI

could handle all tasks efﬁciently, eliminating the risk

of overload.

Man-Machine Teaming: AI’s Overload Management and Task Allocation

103