=Paper=
{{Paper
|id=Vol-3776/paper07
|storemode=property
|title=Design and knowledge requirements for human-machine hybrid intelligence in autonomous driving systems
|pdfUrl=https://ceur-ws.org/Vol-3776/paper07.pdf
|volume=Vol-3776
|authors=Nada Elgendy,Anna Teern,Pertti Seppänen,Tero Päivärinta
|dblpUrl=https://dblp.org/rec/conf/tktp/ElgendyTSP24
}}
==Design and knowledge requirements for human-machine hybrid intelligence in autonomous driving systems==
https://ceur-ws.org/Vol-3776/paper07.pdf
Design and Knowledge Requirements for Human-Machine
Hybrid Intelligence in Autonomous Driving Systems
Nada Elgendy1,∗ , Anna Teern1 , Pertti Seppänen1 and Tero Päivärinta1
1
M3S, University of Oulu, Pentti Kaiteran katu 1, 90570 Oulu, Finland
Abstract
Hybrid intelligence in autonomous driving systems can potentially augment human-machine capabilities and lead to better data-driven
decision-making. Hence, various decision scenarios, such as route planning and change, can benefit from increased optimization.
However, while advancements in research and practice focus on developing technologies for enhancing vehicle autonomy, performance,
and algorithms, the requirements for designing such complex hybrid intelligence systems were not found to be elaborated in the
existing literature. Accordingly, as part of the 6G Visible project and as a result of expert interviews, this paper proposes a set of
design requirements for developing autonomous driving systems with hybrid intelligence. The design requirements cover a range
of multi-faceted attributes that should be reflected in the system, including the decision, decision-making process, knowledge, data,
human, machine, and decision evaluation considerations. Consequently, a set of knowledge requirements for the system is proposed,
from which an ontology can be developed, and the required data can be further determined. Therefore, the design and knowledge
requirements contribute to theory by establishing the initial objectives of a design science artifact, which can be developed in future
research. Furthermore, they support a more comprehensive and sociotechnical view for the practical development and implementation
of hybrid intelligence in autonomous driving systems beyond the prevailing focus on vehicular capabilities.
Keywords
Hybrid intelligence, autonomous driving systems, data-driven decision-making, design requirements, knowledge requirements
1. Introduction nologies, methods, techniques, and algorithms for enhanc-
ing vehicle autonomy and performance, research on hybrid
Autonomous driving has evolved with the advancements in intelligence autonomous driving systems (HI-ADS) is lim-
AI, sensor technology, and intelligent control in a multidisci- ited. Within such research, there is a further gap on how
plinary intersection of research [1]. While the concept was to design HI-ADS in practice and the relevant requirements
first implemented decades ago, technological developments and necessary knowledge. Therefore, the research question
have led to increased attention and research in autonomous of this paper is as follows:
driving systems (ADS) [2]. “What are the design and knowledge requirements for
An ADS generally has six stages or layers, with each stage human-machine hybrid intelligence in autonomous driving
feeding new information into the next stages: (1) the sen- systems?”
sor and hardware layer for gathering data about static and Expert interviews were conducted within the scope of
dynamic objects from the environment, (2) the perception the 6G Visible project to answer the research question and
layer for performing object tracking and detection tasks, (3) propose a set of design requirements and knowledge re-
the localization and mapping layer for getting the vehicle’s quirements for HI-ADS. The paper is structured as follows.
position in the environment, (4) the assessment layer for es- Section 2 provides the theoretical background on ADS and
timating the overall risk and performing predictions to avoid human-machine HI. Section 3 covers the research method
accidents, (5) the path planning and decision-making layer and design, including the research process, a description of
for getting the shortest path from start to end point without the 6G Visible project and the route planning and change de-
collisions, and (6) the control layer which uses actions to cision use case, and a summary of the interviews conducted.
control the vehicle to perform the path [3]. Subsequently, the results are presented and discussed in Sec-
The main focus of this paper is on the decision-making tion 4. Finally, Section 5 concludes the paper and suggests
layer; however, the aim is not to attain the shortest path future work.
but rather to utilize hybrid intelligence (HI), or hybrid aug-
mented intelligence (HAI) as it is referred to in some pa-
pers [1], for planning the best route according to multiple 2. Theoretical Background
human-machine criteria. This is because decision-making
and planning can evolve by augmenting human and machine 2.1. Autonomous Driving Systems
intelligence into HI. Human drivers have the capability to
ADSs are defined by the Society of Automotive Engineers
improve the performance of autonomous driving in real-
(SAE) as “vehicle driving automation systems that perform
world traffic, and their feedback should be introduced into
part or all of the dynamic driving task on a sustained basis”
the learning process of ADSs [1].
[4]. There are six levels of driving automation from Lev-
Nevertheless, although research extensively covers tech-
els 0 to 5. Level 0 systems exhibit no driving automation.
Level 1 systems include primitive driver assistance, while
TKTP 2024: Annual Symposium for Computer Science 2024, June 10–11, Level 2 systems have partial driving automation into which
2024, Vaasa, Finland
∗ advanced driver assistance systems are integrated. Level
Corresponding author.
Envelope-Open nada.sanad@oulu.fi (N. Elgendy); anna.teern@oulu.fi (A. Teern); 3 systems exhibit conditional driving automation where
pertti.seppanen@oulu.fi (P. Seppänen); tero.paivarinta@oulu.fi drivers should be ready to respond and take over. On the
(T. Päivärinta) other hand, Level 4 systems have high driving automation
Orcid 0000-0001-6765-017X (N. Elgendy); 0000-0002-8214-3181 (A. Teern); and do not require human attention but can only operate in
0000-0002-4289-2487 (P. Seppänen); 0000-0002-7477-0783
limited domains and infrastructures. Finally, Level 5 systems
(T. Päivärinta)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License are those with full driving automation [4][5].
Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
The scope of this paper focuses on Level 1, 2, and 3 ADSs, such as the integration of complicated systems, ensuring
which require some form or degree of human supervision robustness and reliability in complex scenarios and weather
and intervention. In the future, this research can be devel- conditions, incorporating human cognition and knowledge
oped for Level 4 ADSs. However, Level 0 and 5 ADSs, which with existing information for decision-making, and data
either require full or no driver control and intervention, are privacy protection [12].
outside the scope of this research.
The ability of autonomous driving to handle dynamic driv- 2.3. Research Gap
ing environments is based on two fundamental components:
decision-making and motion planning. In decision-making, While some studies focus on augmenting human feedback
information is acquired on traffic perception and vehicle to reinforcement learning in the decision-making and plan-
states, and driving tasks are accomplished by generating ning of autonomous vehicles [1], human input is only ob-
desired driving behavior. Next, motion planning occurs by served in ex-post evaluation of the vehicle’s decisions, and
calculating the desired trajectory or vehicle actuator com- human knowledge is not augmented ex-ante into the deci-
mands [1]. sion. Other studies aim to model human psychology and
Higher levels of automation are more complex, although cognition into intelligent decision-making systems for au-
still lacking in outperforming human drivers, and require a tonomous driving [6]. However, such research mainly fo-
perception system (which perceives data through sensors cuses on enhancing the autonomy, algorithms, and perfor-
from the environment), traffic rules interpreter, and a ve- mance of the vehicle (e.g. [13][2][3]) without harnessing
hicle controller in addition to the decision system which the capabilities and knowledge of human drivers during the
controls vehicle behavior and what reaction the system collaborative driving and decision-making process.
should perform [6]. Another study emphasized the importance of HI for au-
tonomous urban vehicle control through a theoretical exam-
ple, with no practical guidelines on the design requirements
2.2. Human-Machine Hybrid Intelligence in
of such systems [11]. A theoretical architecture for HI in
Autonomous Driving autonomous driving was also proposed, along with several
Despite recent technological advancements, many tasks are research challenges [12]. Nonetheless, the requirements
yet unsolvable by machines alone and require sociotechni- for designing such systems in practice are not elaborated.
cal ensembles of humans and machines, or HI systems [7]. Hence, this research aims to define the design and knowl-
HI emphasizes human-AI complementarity by utilizing the edge requirements for HI in ADSs.
senses, perceptions, emotional intelligence, and social skills
of humans with the advanced processing, computational,
and pattern detection capabilities of AI [8].
3. Research Method and Design
Furthermore, such human-machine collaboration leads to
collaborative rationality, which is no longer bounded by the
3.1. Research Process
limitations of each and can lead to higher quality and more A design science research (DSR) methodology is planned to
informed data-driven decisions [9]. Accordingly, a human be adopted for designing a HI-ADS. This paper constitutes
and machine agent collaborate together, informed by data the initial stages of defining the requirements of the solu-
from various sources. Collaborative rationality and/or HI tion artifact [14] according to the practical needs derived
then facilitate data-driven decision-making, and the results from the environment and the knowledge gained from the
of the decisions and their evaluation provide feedback that knowledge base [15].
can be fed back into the system to enable new types of First, the 6G Visible project within which this research
learning [10]. is conducted is explained. Next, the data-driven decision-
Many application areas for HI have been discussed making use case, which was determined to narrow the scope
throughout the literature. For example, HI can be useful for of the research, is elaborated. Consequently, data was col-
autonomous vehicles, which require accurate and up-to-date lected through several semi-structured interviews, which
comprehensive maps to provide real-time environmental are covered in subsection 3.5.
information. Human agents could provide information and
knowledge about the streets, such as rush hour situations, 3.2. The 6G Visible Project: Looking Around
detours in various times and conditions, and their experi-
ences and recommendations for navigation. On the other the Corner
hand, real-world information could be provided by traffic The 6G Visible project aims to develop advanced solutions
cameras and monitoring sensors, as well as satellites and for autonomous and semi-autonomous driving while consid-
global maps. Advanced algorithms can then suggest navi- ering the capabilities of 6G network technologies to support
gation controls based on fusing both human and machine real-time, safety-critical services. These solutions aim to
intelligence [11]. provide visibility for objects not visible to drivers or de-
HI can lead to a semi-autonomous driving architecture tectable by existing vehicle sensors, even in various weather
through which human and machine agents can coopera- conditions. This involves creating dynamic, real-time mod-
tively achieve driving tasks with better system performance els of the environment, traffic, and weather conditions, as
than would be achieved by each on their own [12]. Further- well as detecting obstacles for autonomous driving.
more, data from human guidance and real-time evaluation Furthermore, it involves augmenting human rationality
reflects the driving preferences of humans and can improve and knowledge with machine rationality and artificial intel-
the learning quality of machine agents [1]. Nevertheless, ligence while utilizing various data and knowledge sources
such systems are very complex, require high levels of infor- (e.g. vehicle, sensor, traffic, weather, feedback) to enhance
mation fusion, and are encumbered by several challenges, data-driven decision-making in autonomous driving. The
research is conducted through a collaboration between soft-
ware engineering/information systems, the Finnish Meteo-
rological Institute (FMI), and leading industry collaborators
in AI, digital twins, and autonomous driving.
Current autonomous vehicle solutions rely on static en-
vironmental models (such as maps of city streets and build-
ings, etc.) and integrated sensors on the vehicle. While
they work well for many applications, they are limited in
providing real-time information and rapidly changing situa-
tions beyond the sensor’s range. Furthermore, icy and harsh
weather conditions can affect driving and road conditions
in various ways that may not traditionally be accounted for.
Hence, real-time dynamic environmental data, including
weather adaptation, is essential to enhance autonomous
systems.
3.3. Data-Driven Decision-Making Use Case:
Route Planning and Change
To narrow the scope of the research, three data-driven
decision-making use cases were determined based on their
dependence on the changing conditions of the external en-
vironment. First, in the route planning and change use case,
the HI system should determine the optimal route and adapt
to complex road situations, changes in the environment,
and unexpected weather conditions. Second, in the parking Figure 1: An iterative top-down approach for defining knowledge
use case, the system should support automating the parking and data requirements based on the data-driven decision scenario
process according to the relevant external conditions. Third,
in the weather and lighting adaptation use case, the system
should adjust to various weather and lighting conditions Table 1
based on data and the environment. Interviewee details
The first data-driven decision-making use case, route No. of Interviewees Organization
planning and change, was selected as the most comprehen-
2 Finnish Meteorological Institute
sive and highest-priority use case to start with. 1 Siili Auto
Finding the route on the road network from the origin 1 Elektrobit
to the final destination is the responsibility of the global 1 Driving school teacher in Finland
planner in the ADS, where high performance has become 1 Local taxi company
an industry standard through advancements in GPS and
offline map navigation systems. However, depending on the
country or location, road hierarchies differ, and the shortest context and drilling down to the knowledge and then the
path may not be the fastest or most desirable. Furthermore, data required to support such decisions. By following this
there are other considerations, such as avoiding obstacles, approach, we can focus our resources to utilize only the
satisfying optimization criteria, and understanding road necessary knowledge and data.
semantics, lanes, and drivable surfaces [5]. The interview questions were structured into themes
In many countries, such as Finland, weather conditions according to the components of the framework proposed
complicate autonomous driving. Winter conditions that in Section 3.1. and the elements of data-driven decision-
cause slippery and icy road surfaces and low visibility due making [1]. Hence, there were questions about the deci-
to fog, heavy snowfall, or blizzards and blowing snow are sion, decision-making process, knowledge, data, human
challenging for ADSs and may require making sudden or decision-maker, machine, and decision evaluation as sepa-
precautionary changes to particular routes. Precise and rate attributes.
real-time weather information is necessary, along with in- Deductive thematic analysis [17] was then applied to
formation about road conditions, obstacles, flooding, or group the requirements discussed during the interviews
other unexpected situations. Additionally, instruments that according to the predefined themes.
detect the driving circumstances, such as sensors, cameras, A total of 6 interviews, each approximately 2 hours long,
LiDAR, and radar, are crucial for autonomous driving and were conducted as summarized in Table 1. Accordingly, the
yet face many vulnerabilities in harsh weather and winter interviewees were provided with the interview questions,
conditions [16]. and open discussions were held.
3.4. Interview Data Collection
An iterative top-down approach was adopted for designing
semi-structured, qualitative interview questions and for data
collection, as shown in Figure 1. Rather than defining the
available data that can be used to support the data-driven
decision, we started by defining the decision scenario and
4. Results and Discussion While route planning and optimization is an important
aspect for autonomous driving, both for individuals and
4.1. Dynamic Route Planning and Change public personal transportation, such as taxis, it is currently
Decision Considerations mainly managed by the personal knowledge and experience
of the drivers.
The dynamic route planning and change decision scenario
is a complex data-driven decision with various internal and 4.1.2. Decision-Making Process
external factors. The considerations that must be accounted
for within the decision scenario that resulted from the in- The decision-making process requires environmental aware-
terviews are synthesized below. They are categorized into ness and starts with the collection of data regarding available
themes based on the seven attributes upon which the inter- routes, estimated travel times, and current and predicted
views were structured. road and weather conditions. This data is then evaluated
against a set of criteria that may include safety consider-
4.1.1. Decision ations, efficiency metrics, and personal preferences. Ac-
cordingly, the alternatives may be prioritized, and a route
The decisions involve selecting and adapting the optimal choice is made. The process should be iterative and allow
route based on the combination of a variety of factors, in- reassessing and modifying the decision and plans based on
cluding weather conditions, road friction, road conditions, incoming information and changing conditions.
visibility, accident alerts, traffic information, speed limits, The process is highly dependent on the level of automa-
familiarity with the route, road maintenance, traffic jams, tion of the ADS. Lower levels require more human involve-
construction or roadwork, rest points, personal preferences ment in the decision-making process, while higher levels
such as scenery and facilities, how busy the route is, and the would necessitate more automated processes with less hu-
driver’s urgency. Additionally, considerations like vehicle man intervention.
type (personal vs. public transport), driver experience, and
road safety are important to account for. 4.1.3. Knowledge
The necessity for route adaptation or change usually
arises from emergent conditions such as changes in weather, Informed decision-making in the context of route planning
mostly based on road friction from icy conditions, sudden and change requires access to and understanding of var-
traffic changes, accidents or emergencies, or alterations in ious knowledge areas. Essential knowledge includes ac-
driver preferences or requests. cess to and interpretation of reliable traffic, road condition,
Human preferences, experience, and input play a large and weather information, as well as their implications on
role in route planning and change. For instance, drivers the route, safety, and driving speed. Furthermore, while
may prefer particular routes based on services (e.g. cer- knowledge about other vehicles is crucial, current vehicle-
tain gas stations), familiarity with the route, avoidance of to-vehicle communication is limited.
roadwork, or other preferences, regardless of the expected Additionally, the application of real-time data from
time for taking the route or other variables. Additionally, weather stations and the utilization of algorithms for rapid
icy roads complicate driving and are difficult to account for. data analysis and interpretation are critical for making in-
They require consideration of the road lanes, which are not formed decisions about route adaptations. Human intuition
visible in such conditions, friction, speed, traffic, and other and experience remain to be important sources of knowl-
variables. edge and understanding of the environmental context.
For example, one interviewee remarked, “We have tested
the automatic vehicle with machine vision that works during 4.1.4. Data
summer somehow, but not very reliably even then, and
winter would be impossible because it’s trying to search for The decision is heavily reliant on multiple data sets, which
the white [lane] lines.” Additionally, sensors are not always include accurate GPS and location data, sensor, camera, and
accurate (e.g. may be blocked and sense a normal road as LiDAR data, detailed road weather information, traffic con-
icy). This usually requires an experienced driver who can gestion reports, construction and accident updates, visibility
realize whether or not the road is actually icy. (e.g. fog or increased sunshine), and precise road surface
As another example, one interviewee stated, “I know I temperatures and conditions (e.g. ice, snow, or water pres-
don’t change the route because basically, I know that there’s ence). Nevertheless, GPS routes, and maps are not always
no traffic jam there. There’s, for some reason, a slow car precise. One interviewee commented, “We need accurate
driving there, and it’s giving false readings to that server, GPS routes, but the maps aren’t usually so accurate. You
and the server is sending to my GPS in my own vehicle data need to go and measure them yourself with [real time kine-
that there’s a traffic jam in front, because there are slow cars matic] RTK GPS, usually.”
there…it’s not really necessary to drive 20-30 kilometers Regarding weather data, one interviewee said, “We need
more because of that.” to be wary in situations where it is below 0 degrees [C] or
Hence, traffic congestion may be due to a few cars which close to 0 degrees, and then we have observations, and we
are slow (e.g. due to weather or road conditions) and may also model the road conditions; which means if there is ice
not require a route change. However, the system might or snow or water on the road.”
usually suggest changing to another, longer route due to The sources of data range from the vehicle and its sensors
traffic congestion. An experienced driver would know that and systems to external weather stations or other services
this is not actual congestion, but due to a slow car which and data sources. Derived or calculated variables are also
will pass soon. Accordingly, it would not be necessary to necessary, such as friction levels, driving condition indices,
change to a less desirable route. and precipitation information. One interviewee informed,
“We also calculate these kinds of more advanced variables
like friction that tell us if the friction value is low, then and efficiency, while drawing on quality data and updated
it’s slippery, and if it’s high, then it’s not slippery and also maps. However, current systems remain subpar in terms of
gives this kind of three-level index that is the driving condi- data and recommendation reliability, availability of more
tion: normal, bad, or very bad. Precipitation information is comprehensive information and environmental awareness,
probably important, like if there’s a storm or something.” consideration of a variety of human preferences, and effec-
Additionally, for route changes, new information ac- tive and efficient communication with the human driver.
quired en route is crucial. The availability of traffic ac- Particularly in route change scenarios, such decisions
cident data, changes in weather forecasts (according to an need to be made on the spot and communicated to the driver
interviewee, “They [weather data] can also change because in the fastest, simplest, and least disturbing way possible.
sometimes the forecast is wrong and then we suddenly no- Additionally, there should be some degree of transparency
tice…sometimes there can be these kinds of exceptional for the human driver to understand why the route is being
changes.”), and real-time traffic flow data further inform changed, and to be given the chance to override the decision
decision-making. Metadata and decision variables can be as long as it does not compromise safety requirements.
recorded in logs to monitor the decision and inform future On the other hand, vehicles need more situational aware-
decisions. ness about the environment and decision context, as is em-
phasized by Siili Auto. While not all available data is used
4.1.5. Human Decision-Maker by the vehicle, such data may be crucial to other vehicles.
Accordingly, there should be vehicle-to-vehicle communica-
The role of the human is highly dependent on the level of tion, and data can be shared across a network.
ADS automation. Within the scope of this research, humans Some companies, such as Elektrobit, have developed au-
play a critical role in overseeing and finalizing decisions, tomated driving software that aims to make the vehicle
with the ability to adjust or override automated suggestions aware of the surroundings and road ahead, see beyond the
based on their personal judgment, experience, and real-time range of sensors through predictions, provide precise po-
observations. The human decision-maker generally evalu- sitioning, and model the vehicle environment. With such
ates automated recommendations, verifies machine obser- features, vehicle capabilities may be enhanced and provide
vations (e.g. from sensors), ensures safety, and considers more knowledge to the HI-ADS for supporting data-driven
personal preferences or knowledge of local conditions. decision making. Nevertheless, renewing vehicle-level in-
Some limitations to human drivers exist. For example, not strumentation, which deploys the latest technologies, takes
all drivers are effective at predicted driving, or predicting time.
and reacting to driving scenarios, due to limitations in cog-
nitive and memory functions. Furthermore, emotions may
4.1.7. Decision Evaluation
also lead to poor decisions. For example, if there is a traffic
delay, a driver’s mood may become negatively affected and After the data-driven decision is made, it should be eval-
lead to rash decisions that wouldn’t have otherwise been uated, either during the route or after the destination is
made. reached. It can be evaluated through several metrics or
Nevertheless, while such human errors and biases exist, criteria, which include passenger satisfaction, safety per-
these can be reduced through the provision of simplified ceptions, and the accuracy of estimated versus actual travel
information and reliable data to support decision-making. times. Manual mechanisms for feedback, such as user rat-
For example, knowing what is causing the traffic delay and ings or direct input on satisfaction levels, could help in
possible predictions may prevent negative moods and lead refining future decision-making processes and should be
to better decisions. Additionally, more minor errors can supported by the system. Additionally, more automated
automatically be corrected or alerted to by the vehicle, such feedback mechanisms should also be considered.
as lane keeping. For example, an interviewee added, “I Moreover, although it is not generally considered, the im-
would say in normal cars, there are cameras monitoring pact of the decision on the external environment and other
humans, so if they fall asleep or something like that, then vehicles should be accounted for. For example, actions such
their vehicle can try to stop at least some of them…Of course, as unnecessary overtaking may also delay other vehicles
the lane keeping and stuff like that, there’s a lot of that and negatively impact traffic and should be evaluated as
already in modern cars.” such. However, humans and current technologies may not
Furthermore, humans may have more experience as well consider the effect or consequences of the decision on others
as environmental and situational awareness than machines. or the external environment.
For example, if the ADS suggests a route change based on Furthermore, evaluation could be conducted by incorpo-
a traffic jam, a human who has more experience with that rating external data sources, such as weather or accident
particular route may know that the traffic jam is due to a data, to perform a more comprehensive evaluation. Local
slow car, which will clear up quickly, and that the current forecasts can be compared to observation data (e.g., roads
route remains more optimal than a route change. that were predicted to be slippery but weren’t, or vice versa)
to enhance future decisions, not only for the vehicle but
4.1.6. Machine for other vehicles as well. However, more global evaluation
would require communication with other servers, weather
The role of the machine is also dependent on the level of stations, and other vehicles.
the ADS automation, with higher levels requiring larger ma-
chine roles. Nevertheless, the current status of technology
requires human oversight, validation, and control. Thus, 4.2. Design Requirements
humans generally make the final decisions. Machines can Based on the results of the interviews and considerations
support (or make) the decision by suggesting (or taking) discussed in the previous subsection, the requirements that
routes based on programmed criteria, preferences, safety,
h
Table 3
Table 2 Knowledge Requirements for HI-ADS
Design Requirements for HI-ADS
No. Knowledge Requirement
Attribute Requirements to be Supported by the Sys-
tem 1 Internal knowledge from the vehicle, including
vehicle and tire conditions, fuel status, vehicle
Decision Define the decision context and goals. De- functions, speed, location, time, other measure-
termine the factors affecting the decision ments, and internal sensors.
context and its change across time. 2 External knowledge from the vehicle, including
Decision- Define the decision-making process accord- information from sensors and cameras about
Making ing to the level of ADS automation. actual road surfaces and types, road friction,
Process proximity, and the surrounding (immediate) en-
Knowledge Determine various human and machine vironment.
knowledge sources to support the decision. 3 Knowledge from the human, including driver
Map the knowledge sources to an ontology. preferences, experience, and continuous feed-
Data Determine accurate data sources required back.
for the decision. Determine how the data 4 Knowledge from external weather services and
is represented and affects the decision. stations, including current weather information,
Human Determine the role of the human in station and road forecasts, road conditions, road
decision-making according to the level of condition severity, and resulting crash risk lev-
ADS automation. Define areas where hu- els.
man judgment, environmental awareness, 5 Knowledge from map and positioning services
and experience should have a higher im- about roads, including road segments and loca-
pact on the decision. tions, and road types.
Machine Determine the role of the machine in 6 Knowledge about the road environment and
decision-making according to the level of semantics that may affect human preferences,
ADS automation. Define areas where ma- including roadside facilities and stops, road
chine capabilities should have a higher im- scenery, road surface, and speed limits.
pact on the decision. Provide the capability 7 Knowledge about the current and predicted traf-
to consider human experience and prefer- fic situations from traffic or city services, includ-
ences. Determine efficient and effective ing traffic density, emergencies, accidents, and
means for human-machine communication road maintenance.
and interaction. Enhance situational aware- 8 Knowledge incoming from other vehicles and
ness of the vehicle. Determine available external sources.
software and toolkits which may support 9 Knowledge from predictions incoming from ad-
the system. vanced systems.
Decision Define evaluation criteria and feedback
Evaluation mechanisms. Incorporate various aspects
for evaluation, including impact and conse-
quences on the internal and external envi- determined from the interviews are summarized in Table 3
ronments. Provide the capability to detect, and can subsequently be considered for building the ontol-
reduce, and learn from both human and ogy.
machine decision-making and errors. Based on the knowledge requirements and through fur-
ther research and work, an initial ontology is developed.
However, the ontology is not presented within the scope of
should be supported when designing a HI-ADS are summa- this paper and will be developed further in the future for
rized in Table 2. constructing the KGs within the HI-ADS.
4.3. Knowledge Requirements 4.4. Theoretical and Practical Contributions
For representing human and machine knowledge in the sys- The main contributions of this paper are two-fold: the de-
tem, knowledge graphs (KG) can be used to structure the sign requirements in Table 2 and the knowledge require-
knowledge and enable HI systems. A KG is a large-scale ments in Table 3.
knowledge base commonly used for intelligent applications, The design requirements cover a range of multi-faceted
which comprises a large number of entities and the rela- attributes that should be reflected in the system, as opposed
tionships between them [18]. The KG construction process to focusing only on vehicular requirements, technological
is as follows. Starting with the knowledge acquisition, the advancements, or planning and prediction algorithms (e.g.
data for building the ontology is acquired. Knowledge fu- [13][3][5]). While such focused research is necessary for
sion tasks are then performed. The knowledge is stored advancing ADSs, it generally does not consider the human
and represented, and then processed to create and update aspect or additional decision-making factors. This paper
an ontology. Consequently, after the ontology is built, the contributes to such a gap by proposing a more comprehen-
knowledge can be utilized, and further iterations can occur sive viewpoint to the data-driven decision scenario.
to evolve the KGs [19]. Accordingly, the attributes considered in the design re-
Such aspects can serve as a basis for augmenting human quirements include the decision, decision-making process,
and machine knowledge within the HI-ADS to support data- knowledge, data, human, machine, and decision evaluation
driven decision-making and enable knowledge sharing and considerations.
utilization. A set of knowledge requirements for the system is sub-
As an initial stage, the knowledge requirements that were sequently proposed, from which an ontology can be devel-
oped, and the data required by the system can further be on Intelligent Transportation Systems (2024) 1–13.
determined. By adopting a top-down approach, resources URL: https://ieeexplore.ieee.org/abstract/document/
can be focused on gathering only the necessary data and 10400976. doi:10.1109/TITS.2023.3349198 , confer-
knowledge that support the data-driven decision. ence Name: IEEE Transactions on Intelligent Trans-
Therefore, the design and knowledge requirements con- portation Systems.
tribute to theory by establishing the initial objectives of a [2] M. A. Khan, Intelligent Environment Enabling Au-
design science artifact, which can be developed in future tonomous Driving, IEEE Access 9 (2021) 32997–33017.
research. Furthermore, they support a more comprehensive URL: https://ieeexplore.ieee.org/abstract/document/
and sociotechnical view for the practical development and 9356463. doi:10.1109/ACCESS.2021.3059652 , con-
implementation of hybrid intelligence in autonomous driv- ference Name: IEEE Access.
ing systems beyond the prevailing and somewhat myopic [3] M. Reda, A. Onsy, A. Y. Haikal, A. Ghan-
focus on vehicular capabilities. bari, Path planning algorithms in the au-
Moreover, this paper supports existing research that em- tonomous driving system: A comprehensive
phasizes the need for improving situational awareness of review, Robotics and Autonomous Systems 174
the vehicle and its understanding of the environment [2], (2024) 104630. URL: https://www.sciencedirect.
incorporating human feedback into the learning process for com/science/article/pii/S0921889024000137.
HI-ADSs [1], and considering various factors from multiple doi:10.1016/j.robot.2024.104630 .
systems and utilizing human cognition [12]. Such needs [4] J3016_202104: Taxonomy and Definitions for Terms
are reflected in the design and knowledge requirements and Related to Driving Automation Systems for On-Road
extended upon with additional considerations regarding Motor Vehicles - SAE International, 2021. URL: https:
the data-driven decision context. Additionally, determin- //www.sae.org/standards/content/j3016_202104/.
ing the knowledge requirements is an important, although [5] E. Yurtsever, J. Lambert, A. Carballo, K. Takeda,
commonly overlooked, prerequisite for supporting the de- A Survey of Autonomous Driving: Common Prac-
velopment of knowledge graphs within the HI-ADSs. tices and Emerging Technologies, IEEE Access
8 (2020) 58443–58469. URL: https://ieeexplore.ieee.
org/document/9046805/. doi:10.1109/ACCESS.2020.
5. Conclusion 2983149 .
[6] M. Czubenko, Z. Kowalczuk, A. Ordys, Autonomous
In this paper, we aimed to determine the design and knowl-
Driver Based on an Intelligent System of Decision-
edge requirements for developing HI-ADSs. Six expert in-
Making, Cognitive Computation 7 (2015) 569–581.
terviews were conducted within the scope of the 6G Visi-
URL: https://doi.org/10.1007/s12559-015-9320-5.
ble project. Consequently, following a top-down approach,
doi:10.1007/s12559- 015- 9320- 5 .
considerations for a dynamic route planning and change
[7] D. Dellermann, A. Calma, N. Lipusch, T. Weber,
decision scenario were defined based on seven attributes:
S. Weigel, P. Ebel, The Future of Human-AI Collabora-
the decision, decision-making process, knowledge, data,
tion: A Taxonomy of Design Knowledge for Hybrid
human decision-maker, machine, and decision evaluation.
Intelligence Systems, 2019. URL: http://hdl.handle.net/
From these considerations, the design requirements were
10125/59468.
determined, followed by the knowledge requirements for
[8] P. Hemmer, M. Schemmer, M. Vössing, N. Kühl,
HI-ADSs
Human-AI Complementarity in Hybrid Intelligence
Future work involves building an ontology and knowl-
Systems: A Structured Literature Review, PACIS
edge graph based on the knowledge requirements, merging
2021 Proceedings (2021). URL: https://aisel.aisnet.org/
the main concepts from the data-driven decision-making
pacis2021/78.
and evaluation model in [10] and the KG construction pro-
[9] N. Elgendy, A. Elragal, T. Päivärinta, Evalu-
cess in [19] into a conceptual framework for augmenting
ating collaborative rationality-based decisions: a
human-machine intelligence and knowledge, and design-
literature review, Procedia Computer Science
ing a HI-ADS based on the design requirements. Finally,
219 (2023) 647–657. URL: https://linkinghub.elsevier.
data privacy, legal, and ethical regulations and requirements
com/retrieve/pii/S1877050923003447. doi:10.1016/j.
were not covered within the scope of this research and may
procs.2023.01.335 .
be considered in the future.
[10] N. Elgendy, Enhancing collaborative rationality be-
tween humans and machines through data-driven de-
Acknowledgments cision evaluation, in: CEUR Workshop Proceedings,
2022.
This research was partially funded by Business Finland. We [11] M. Moradi, M. Moradi, F. Bayat, A. N. Toosi,
would like to acknowledge the Finnish Meteorological In- Collective hybrid intelligence: towards a con-
stitute, Siili Auto, Elektrobit, and the interviewees from the ceptual framework, International Journal of
local taxi company and driving school in Finland for their Crowd Science 3 (2019) 198–220. URL: https:
cooperation. //ieeexplore.ieee.org/abstract/document/9826654.
doi:10.1108/IJCS- 03- 2019- 0012 , conference Name:
International Journal of Crowd Science.
References [12] H. Ning, R. Yin, A. Ullah, F. Shi, A Survey
on Hybrid Human-Artificial Intelligence for Au-
[1] K. Yuan, Y. Huang, S. Yang, M. Wu, D. Cao, Q. Chen,
tonomous Driving, IEEE Transactions on Intelli-
H. Chen, Evolutionary Decision-Making and Plan-
gent Transportation Systems 23 (2022) 6011–6026.
ning for Autonomous Driving: A Hybrid Aug-
URL: https://ieeexplore.ieee.org/abstract/document/
mented Intelligence Framework, IEEE Transactions
9423526. doi:10.1109/TITS.2021.3074695 , confer-
ence Name: IEEE Transactions on Intelligent Trans-
portation Systems.
[13] A. Benterki, M. Boukhnifer, V. Judalet, C. Maaoui,
Artificial Intelligence for Vehicle Behavior An-
ticipation: Hybrid Approach Based on Maneu-
ver Classification and Trajectory Prediction,
IEEE Access 8 (2020) 56992–57002. URL: https:
//ieeexplore.ieee.org/abstract/document/9043491.
doi:10.1109/ACCESS.2020.2982170 , conference
Name: IEEE Access.
[14] K. Peffers, T. Tuunanen, M. A. Rothenberger, S. Chat-
terjee, A Design Science Research Methodology
for Information Systems Research, Journal of Man-
agement Information Systems 24 (2007) 45–77. URL:
https://doi.org/10.2753/MIS0742-1222240302. doi:10.
2753/MIS0742- 1222240302 , number: 3 Publisher:
Routledge _eprint: https://doi.org/10.2753/MIS0742-
1222240302.
[15] A. R. Hevner, S. T. March, J. Park, S. Ram, Design Sci-
ence in Information Systems Research, MIS Quarterly
28 (2004) 75–105. URL: https://www.jstor.org/stable/
25148625. doi:10.2307/25148625 , number: 1 Pub-
lisher: Management Information Systems Research
Center, University of Minnesota.
[16] M. Hippi, T. Sukuvaara, K. Mäenpää, T. Perälä,
Weather Service to Support Autonomous Driving
in Adverse Weather Conditions, DRUSKININKAI,
LITHUANIA (14-16TH JUNE 2022) (????) 123.
URL: https://www.zurnalai.vu.lt/proceedings/article/
download/27645/26922/#page=124.
[17] J. A. Smith, Qualitative Psychology: A Practical Guide
to Research Methods, SAGE, 2015. Google-Books-ID:
lv0aCAAAQBAJ.
[18] Z. Chen, Y. Wang, B. Zhao, J. Cheng, X. Zhao, Z. Duan,
Knowledge Graph Completion: A Review, IEEE Access
8 (2020) 192435–192456. URL: https://ieeexplore.ieee.
org/document/9220143/. doi:10.1109/ACCESS.2020.
3030076 .
[19] A. Teern, M. Kelanti, T. Päivärinta, M. Karaila, Design
Objectives for Evolvable Knowledge Graphs, Com-
plex Systems Informatics and Modeling Quarterly
(2023) 1–15. URL: https://csimq-journals.rtu.lv/article/
view/csimq.2023-36.01. doi:10.7250/csimq.2023- 36.
01 , number: 36.