Access to this full-text is provided by MDPI.
Content available from Smart Cities
This content is subject to copyright.
Academic Editor: Pierluigi Siano
Received: 1 December 2024
Revised: 8 February 2025
Accepted: 12 February 2025
Published: 17 February 2025
Citation: De Silva, D.; Mills, N.;
Moraliyage, H.; Rathnayaka, P.;
Wishart, S.; Jennings, A. Responsible
Artificial Intelligence Hyper-
Automation with Generative AI
Agents for Sustainable Cities of the
Future. Smart Cities 2025,8, 34.
https://doi.org/10.3390/
smartcities8010034
Copyright: © 2025 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license
(https://creativecommons.org/
licenses/by/4.0/).
Article
Responsible Artificial Intelligence Hyper-Automation with
Generative AI Agents for Sustainable Cities of the Future
Daswin De Silva * , Nishan Mills , Harsha Moraliyage , Prabod Rathnayaka , Sam Wishart
and Andrew Jennings
Centre for Data Analytics and Cognition, La Trobe University, Bundoora, VIC 3086, Australia
*Correspondence: d.desilva@latrobe.edu.au
Highlights:
What are the main findings?
•
Smart Cities as Hyper-Connected Digital Environments generate large and di-
verse data streams and repositories that do not consistently translate into insights
and decisions.
•
A Responsible AI Hyper-Automation framework with Generative AI agents is devel-
oped and evaluated to address these complex challenges.
What are the implications of the main findings?
•
The developed AI framework is effective when grounded on five core technical ca-
pabilities with an independent cognitive engine for hyper-automated agentic AI that
feeds into human-in-the-loop processes.
•
The framework provides a prototypical setting for university cities of the future to
provide direction, guidance, and standards for sustainable and safe smart cities of
the future.
Abstract: Smart cities are Hyper-Connected Digital Environments (HCDEs) that transcend
the boundaries of natural, human-made, social, virtual, and artificial environments. Human
activities are no longer confined to a single environment as our presence and interactions
are represented and interconnected across HCDEs. The data streams and repositories of
HCDEs provide opportunities for the responsible application of Artificial Intelligence (AI)
that generates unique insights into the constituent environments and the interplay across
constituents. The translation of data into insights poses several complex challenges originat-
ing in data generation and then propagating through the computational layers to decision
outcomes. To address these challenges, this article presents the design and development
of a Hyper-Automated AI framework with Generative AI agents for sustainable smart
cities. The framework is empirically evaluated in the living lab setting of a ‘University
City of the Future’. The developed AI framework is grounded on the core capabilities of
acquisition, preparation, orchestration, dissemination, and retrospection, with an indepen-
dent cognitive engine for hyper-automation of these AI capabilities using Generative AI.
Hyper-automation output feeds into a human-in-the-loop process prior to decision-making
outcomes. More broadly, this framework aims to provide a validated pathway for uni-
versity cities of the future to take up the role of prototypes that deliver evidence-based
guidelines for the development and management of sustainable smart cities.
Keywords: smart cities; artificial intelligence; generative AI; hyper-automation; university
city of the future
Smart Cities 2025,8, 34 https://doi.org/10.3390/smartcities8010034
Smart Cities 2025,8, 34 2 of 14
1. Introduction
Increasing populations and burgeoning economies are driving the dense urbanisation
of cities. This exponential growth has overwhelmed conventional city infrastructure
resulting in dire social, economic, and environmental consequences. Safe and sustainable
living are Global Goals in the 17 interlinked Sustainable Development Goals (SDG) adopted
by the United Nations (UN) General Assembly as a UN resolution to be achieved by
2030 [1].
It is a universal call to action adopted by all member states in 2015 to end poverty,
protect the planet, and ensure peace and prosperity by 2030 [
2
]. Despite investment and
ambition, by the midpoint year of 2023, only 15% of the targets are on track while many
others are in reverse [3].
In response to this slow progress, numerous governments across the world have com-
mitted to the development of transformative smart city infrastructure as a national priority
that delivers the UN SDGs. Smart cities are typically defined as the use of technology-based
systems for service innovations that improve the quality of life for citizens, increase en-
gagement with organisational entities, and enable sustainable development [
4
,
5
]. However,
this definition varies across locations, capability, and structure, with a primary distinc-
tion between the ‘hard’ domain and the ‘soft’ domain, where ‘hard’ describes the urban
technical infrastructure that can be enhanced through technology applications and pol-
icy intervention, while ’soft’ focuses on cultural and societal aspects such as education,
governance, and equity that are dependent on technology applications [
6
]. Originating
in the six smart city dimensions—economy, environment, governance, living, mobility,
and people—several smart city assessment tools and indexes have also been developed.
The Cities in Motion Index, Digital City Index, Cities Survey, and European Digital City In-
dex are some of the primary indexes used to evaluate the progression of the six dimensions.
Although academia, industry, and the public sector have initiated research endeavours to
investigate smart cities, the opportunities and challenges of leveraging diverse digital data
streams and repositories for the transformation of urban infrastructure have not been fully
explored [
7
]. The prime opportunity lies in the big data representations of its constituents,
the unstructured, unlabelled, and non-deterministic datasets and data streams that are
generated at high volume and high velocity, within a modern urban environment [8].
Artificial Intelligence (AI) is primed to address these gaps by providing a transforma-
tional capability to generate insights, predictions, and other aggregates from large volumes
of data through learning, reasoning, and optimisation. Despite AI being studied and prac-
tised for several decades, in 2023, the OECD presented a renewed definition of AI as, “An
AI system is a machine-based system that for explicit or implicit objectives, infers, from the
input it receives, how to generate outputs such as predictions, content, recommendations,
or decisions that can influence physical or virtual environments. Different AI systems
vary in their levels of autonomy and adaptiveness after deployment” [
9
]. This renewed
definition can be exemplified in the following two smart city scenarios. In a sustainable
energy management setting, the equilibrium between energy consumption, renewables
generation, and grid stability maintenance cannot be achieved by merely utilising control
systems implemented upon these model-based solutions. The access and availability of
digital source systems that monitor and control supply and demand is indicative of the real-
time high volume, high velocity, unstructured data that can be used to build and deliver
AI-based sustainable smart city infrastructure. Secondly, in ensuring the safety of urban
environments, a multitude of closed-circuit television cameras cannot be monitored round
the clock by human operators for equilibrium of movements and activity. Control systems
will not predict or detect beyond the predefined variables even when the expectation is to
detect unseen, anomalous activity. Much of the under-utilised voluminous, continuously
streamed video data can be leveraged by AI algorithms to predict, detect, and optimise
Smart Cities 2025,8, 34 3 of 14
operations in such settings. AI has been applied in diverse settings with many practical
benefits, such as industrial systems [10], digital health [11,12], and energy systems [13,14].
Despite these effective applications, AI is still impacted by risks and potential harms from
bias, inaccuracies, lack of transparency, accountability, and adverse effects on physical and
mental health [
15
,
16
]. Responsible AI adoption and AI regulation are making headway
into addressing some of these risks [17–19].
Hyper-automation was originally described by Gartner as a disciplined approach to
rapidly identify and automate organisational and technology processes using AI and other
emerging foundational technologies [
20
]. A typical hyper-automation effort is focused
on processes and tasks that are interactive, decision-making, and optimisation-focused,
instead of sequential and repetition that can be managed within an automation only effort.
Following the recent Gen AI advancements, hyper-automation now extends beyond the
original scope to include Generative AI agents with a broad range of ‘agency’, starting
from simple prompt and response to fully autonomous [
21
,
22
]. Agentic AI extends beyond
task automation to include decision-making processes and direct actions, creating systems
capable of unsupervised and self-supervised adapting and learning [
23
,
24
]. Inventors of
the GPT family of Gen AI models, OpenAI, provide the following definition, “Agentic AI
systems are characterized by the ability to take action which consistently contribute to-
wards achieving goals over an extended period of time, without their behavior having been
specified in advance”. Following through from this definition, the key attributes of agentic
AI are well-aligned with the decision-making needs of an HCDE, they are, goal-oriented be-
haviours, operations in complex environments, ability to adapt, and independent execution.
In order to take actions, agentic AI receive data streams about the operational environment
through sensing, past data, and human input, which is then followed by a deep analysis of
processes/task prior to automation, action, or decision, which can also enable learning and
a self-organising mechanism for subsequent AI workflows. By grounding agentic AI in
thorough process understanding, organisations can navigate the complexities of HCDE
environments responsibly and effectively, so that technology and deep process knowledge
converge to optimise and innovate HCDE operations.
Drawing on this context, we present the design and development of a responsible
AI framework for hyper-automation of sustainable smart cities, which is also empirically
evaluated in the living lab setting of a University City of the Future.
The selection of a university living lab setting is based on its structured yet dynamic
nature, which represents the primary aspects of urban complexity, including energy con-
sumption across multiple facilities, human mobility patterns, and data integration from dis-
parate data sources. While the campus does not capture the full extent of non-determinism
of a city, it provides a controlled environment for the proposed responsible AI framework
with agentic AI for automation and decision-making.
The framework is grounded on the core capabilities of acquisition, preparation, or-
chestration, dissemination, and retrospection, with an independent cognitive engine for
hyper-automation. The rest of the paper is organised as follows: Section 2presents related
work and recent literature on the use of AI for smart cities; Section 3presents the design
and development of the AI framework; Section 4presents the practical application of this
framework in the living lab setting of a University City of the Future across
four empirical
studies; Section 5concludes the article.
2. Related Work
Smart cities are being developed across the world in response to the increase in more
than half of the global population currently residing in urban areas [
25
]. Smart cities
are defined as, “urban environments that deploy and leverage innovative technology
Smart Cities 2025,8, 34 4 of 14
solutions and platforms to enhance the quality of life through improved urban services,
reduced costs, and increased engagement across all stakeholders” [
4
,
26
]. As an overarching
framework [
27
,
28
], smart cities are impacted by the safety, privacy, and confidentiality of
people, interoperability of new technologies with legacy systems, regulatory frameworks for
engagement in decision-making, and disaster recovery and emergencies [
29
]. In response
to these challenges, Artificial Intelligence (AI) is enabling efficient processes, innovative
programs, productivity, and resilience within smart city environments [
30
,
31
]. Generative
AI (Gen AI) is a further enabler through contributions in increasing engagement, improving
workflows, and supporting growth and innovation within smart city environments [
32
,
33
].
Several studies have presented frameworks for smart cities, starting with the overarch-
ing sustainable, livable, and efficient communities based on the pillars of collaboration, data,
citizen engagement, sustainability, technology, and ongoing evaluation [
27
]. The authors
of [
34
] discuss the diverse business frameworks of smart cities, public–private partnerships,
build–operate–transfer arrangements, performance-based contracts, community-centric
models, innovation hubs, etc., while in [
35
], the focus narrows down specifically towards AI,
an open-source AI framework for robust decision support that assists policymakers. Frame-
works have also been proposed for specific technical areas, such as sustainable hydropower
plant operations [
36
], asthma attack prediction [
37
], machine learning on the edge [
38
],
security, privacy and risks [
39
], and planning and management [
40
]. Focusing more closely
on AI framework, several reviews present broad evidence supporting the effectiveness and
practical value of AI frameworks [
31
,
41
,
42
]. In terms of applications, the most impact from
AI is reported in healthcare, mobility, privacy, and energy, while the notion of an ‘urban
artificial intelligence’ has also been proposed through developments in autonomous cars,
robots, and the built environment. Despite a number of recent studies, an AI framework
that encompasses the contribution of agentic AI for the complex challenges originating in
data generation through to decision outcomes has not been investigated. The following
section presents the AI framework proposed in this paper to address this gap.
3. The Proposed AI Framework
The proposed AI framework is structured across five overarching themes: Acquisition,
Preparation, Orchestration, Dissemination, and Retrospection. These thematic categories
aim to encapsulate the full lifecycle of AI projects in smart city environments, offering
a holistic view of the factors that contribute to the effective implementation of such AI
systems. The themes can be described as follows: Acquisition, where the platform gathers
relevant data; Preparation, which involves processing the data for analysis; and Orches-
tration, where the platform process including AI algorithms applies the platform’s core
objectives and ethical principles on the inputs flowing through the system. The processed
information then transitions into outputs through the Dissemination stage, where the
platform’s actions are deployed back into the environment. This cycle is underpinned
by the Knowledge Reservoir, a comprehensive repository that records all data, decisions,
and outcomes. Leveraging this reservoir, the platform engages in Retrospection, a reflective
process that allows it to learn from past actions, refine its algorithms, and continuously
adapt to the evolving environment. This feedback loop ensures that the framework is
capable of responding and adapting to change in the environment. Across the five themes,
the framework ensures the ethical and responsible application of AI, as each transaction,
data flow, data record, decision outcome, and feedback are captured by the Knowledge
Reservoir. The Retrospection theme is a further innovation for Responsible AI as the
framework itself can navigate and unpack the complexities of AI from data to decisions,
through the ethical indicators of transparency, accountability, fairness, privacy, security,
and reliability.
Smart Cities 2025,8, 34 5 of 14
The five themes of the framework are articulated through sequential stages: Inputs,
Discover, Enlist, Marshal, Design, Build, Evaluate, Explainability, Optimise, and Deliv-
ery/Deploy. Each stage serves a unique purpose and presents a scheme of what takes place
in the overall cycle. Collectively, they form a platform conceptualisation that addresses
the multifaceted nature of AI platform initiatives, from concept to deployment. Figure 1
presents the composition of this framework, distinguishing between the AI platform,
external environment, and downstream and upstream integrations.
Figure 1. The Proposed Responsible AI Framework for Hyper-automation.
The framework provides agentic orchestration of components that work in synergy
to automate complex HCDE processes. The External Environment, System Objectives,
and System Principles, establish the fundamental guidelines. The Data Reservoir acts as a
centralised repository for all the information gathered and utilised by the system, ensuring
that data is readily available for analytics and further learning. Each of the stages allow for
the framework to direct the Acquisition of raw data, the Preparation of that data for analysis,
the Orchestration of workflows, and the Dissemination of insights or actions derived from
the processed data. These stages ensure that the system operates with precision, aligns with
organisational goals, and adapts to the external factors that might influence its performance.
Supporting this lifecycle is the Retrospection stage, where the system’s performance is
continuously analyzed to derive insights for improvement. The framework is rounded out
by Supporting Services/Deployment Slots that provide the necessary infrastructure for the
deployment and operation of the system, and by Downstream and Upstream Integrations
that enable the exchange of data and functionalities with external and internal systems.
As depicted in Figure 2, the Cognitive Engine provides the hyper-automation ca-
pability of this AI framework. It allows for the Gen AI capabilities of (1) chat, (2) open
Q&A, (3) closed Q&A, (4) text classification, (5) text summarisation, (6) text generation,
(7) information
extraction, (8) text rewrite, and (9) brainstorming as discussed in the In-
structGPT model proposed by OpenAI [
43
]. The engine also constitutes an agent factory
which is the production line for autonomous agents, each tailored for specific tasks within
Smart Cities 2025,8, 34 6 of 14
the framework. These agents are autonomous units that take actions which consistently
contribute towards achieving goals over an extended period of time. The creation of these
agents involves configuring them with the necessary functional capabilities to achieve
the task assigned. The agent council implements the Socrates model framework, which
allows for more complex tasks to be broken down and disseminated through the system
using agents. The council allows the engine to monitor performance, manages resource
allocation, and oversees the interaction between agents to maintain system integrity and
efficiency. It is also responsible for decommissioning agents that are no longer needed
or updating them in line with new requirements, much like a council would oversee the
roles and responsibilities of its constituents in a community. The model also allows for
the integration of Human Intervention to direct the tasks. The cognitive engine contin-
ually evolves through learning; as it is exposed to new data and different scenarios, it
refines its algorithms for better accuracy and efficiency in decision-making processes. This
adaptive learning ensures that the system remains effective in a dynamic environment, ac-
commodating new challenges and opportunities that arise within the external environment.
The engine has access to the functional manifest and the data reservoir of the system, this
allows for different functionalities to be addressed and made more efficient through the
processing in the engine.
Figure 2. Schematic representation of the structure and function of the cognitive engine. The arrows
in red are indicative of bi-directional information and instruction flows, for instance the human agent
engages the Council with information and instructions on complex tasks that are deconstructed and
assigned to agents, with feedback loops to the human following execution and delivery.
The cognitive engine adopts a Functional Codification process to codify repetitive
functions which are routed through the framework. This is akin to motor learning in
human physiology which is developed through consistent repetition of specific movements.
This process involves the brain and spinal cord creating efficient neural pathways. When
muscle memory is fully established, a person enters an autonomous stage where the
movement can be triggered without higher cognitive activation. Similarly, in the initial
stages, the framework uses generative AI to plan responses to presented problems. The plan
could be the calling of a function or the writing of some code; this is then executed and an
answer is returned to the calling component and entity. The framework uses a feedback
mechanism to ensure fit for the purpose of the result. If the schematic is inconclusive it
asks follow-up questions to clarify. Over time, the framework analyses multiple calls which
use the similar pathway in terms of the schematic, on a threshold it establishes the most
Smart Cities 2025,8, 34 7 of 14
efficient schematic to execute a task, similar to how the brain and spinal cord work to
establish efficient movement patterns. The framework would then codify this process in
source code which is deployed to the function manifest, embedding it into its ecosystem.
The AI framework shifts from actively computing each step of a task to retrieving and
executing pre-established, optimised code. This transition is illustrated in Figure 3.
Figure 3. Functional Codification: from active computing to retrieval and execution of pre-
established code.
The framework is designed to be adaptable across different smart city contexts, which
necessitates a level of generalization in its structure. However, its implementation involves
specific input parameters depending on the smart city context. For instance, data sources
such as smart energy meters, mobility tracking systems, environmental monitoring sta-
tions, and administrative policies are key inputs that drive AI-driven decision-making.
The experiments in the following section demonstrate the application of a subset of these
inputs in a university setting, but the same framework can integrate additional parameters
when applied to larger urban environments.
4. Experiments and Results
This framework was implemented and operationalised at La Trobe University’s living
lab setting of the ‘University City of the Future’. It is a flagship responsible AI initiative
for the achievement of several university objectives in sustainability and safety, including,
student safety, optimised energy usage, and net zero carbon emissions targets. Building on
the technical details of this implementation that have been published in recent articles [
44
],
here, we present (Figure 4) the cognitive engine and agentic AI capabilities for autonomous
and independent actions and decision-making in support of human operators and third-
party systems. Positioned on top of Data Sources, Data Management, AI Capabilities,
and Applications, the AI agents provide a higher cognitive capacity of human operators to
interact with the data and AI layers of the framework.
La Trobe University is spread across five campuses, with approximately
40,000 students
and 3250 staff, and a total operating revenue of USD 770 million [
45
]. The main metropolitan
campus is La Trobe’s Bundoora campus, which is spread over 235 hectares, it is the largest
university campus in the southern hemisphere. The other four campuses are in regional
locations, Bendigo, Shepparton, Albury-Wodonga, and Mildura. All five campuses are
in the state of Victoria which has a temperate oceanic climate, and four seasons: summer
Smart Cities 2025,8, 34 8 of 14
from December to February, autumn from March to May, winter from June to August, and
spring from September to November.
Figure 4. Implementation of the framework with a cognitive engine and agentic AI capabilities.
Using this living lab setting, the technical capabilities of the framework are demon-
strated using four experiments in the following subsections: (1) human mobility prediction
for indoors and outdoor activities; (2) forecasting consumption and generation for en-
ergy usage optimisation; (3) energy savings profiling by building, campus, and time;
and
(4) human
physical activity monitoring. In all four experiments, we delineate how the
responsible AI framework with agentic AI operates through the layers of the framework
to generate the output and actions. Each experiment reports the AI model performance
against the actual data of the specific setting. This comparison highlights the performance
and decision-making impact and contribution of using this framework. Datasets and
related information have also been published separately [46].
4.1. Experiment 1: Human Mobility Prediction
For mobility prediction, the framework extracts data from several input sources,
consumption from smart meters, weather data from external feeds, schedules from internal
systems, and WiFi activity from real-time data feeds. The AI agents are deployed to evaluate
quality, integrity, and potential integration points for these data sources prior to further
processing and validation. Next, the agents are activated to determine suitability of AI
capabilities based on the profile of the data sources and the preset AI capabilities.
LSTM and RNN-based deep learning models for prediction and forecasting with
global and local training are run using the filtered data streams. Many machine learning
models with high accuracy are impacted by overfitting, where the model becomes specific
to the training data, resulting in poor performance on testing data and in real-world
applications. K-fold cross-validation is used to detect and assess overfitting. To further
optimize the model’s performance, hyperparameters, dropout, number of layers, and
sequence length. Since each hyperparameter can take a wide range of values and multiple
hyperparameters can be adjusted, the computational complexity of model development
increases. This challenge is typically handled using methods like Grid Search, which trains
models for all possible hyperparameter combinations to find the optimal configuration,
or Randomized Search, which randomly selects hyperparameter combinations to save on
computational time.
The results generated for indoor and outdoor activities are depicted in Figure 5.
The global model is more accurate than local models, with the highest accuracies being
for indoor activity, indicating a 25% increase in RMSE and a 20% MAE increase. In con-
trast, for outdoor activities, the local model is a lot more accurate than the global model,
Smart Cities 2025,8, 34 9 of 14
presumably due to the high variance of the data streams for outdoor activities. In terms of
the work week, Monday, Tuesday, and Wednesday are the largest peaks, with low indoor
activity from Thursday through the weekend. Outdoor activities sustain the same amount
of activity on weekends as weekdays. For agentic AI, these insights form the next steps
for decisions in planning and budgeting that drive energy savings, carbon reductions, and
optimised space configurations.
Figure 5. Human mobility prediction for indoor and outdoor activities.
4.2. Experiment 2: Energy Consumption and Generation Forecasting
In this experiment, consumption, generation, and tariff data streams are used to fore-
cast the short to long-term trend of demand, generation, and supply, so that decisions and
actions can be taken based on temporal needs. As shown in Figure 6, solar generation has a
direct contribution to the predicted decrease in metered consumption and thereby, the pre-
dicted decrease in the demand from ‘Max Peak Demand’ to ‘Max Current Demand’ and
the targeted ‘Max Metered Demand’ for future decision-making scenarios. The magnitude
of generation, consumption, and demand are not shown on this graph due to the privacy
and confidentiality of the university energy usage. Peaks and anomalies are also visualised
in Figure 6, where AI agents are able to identify, analyse, and provide an explanation
through access to a multimodality of data sources within the framework. Collectively,
these capabilities will inform decision-making for energy savings and emissions reduction,
as well as arbitrage, maintenance, and daily operations.
Figure 6. Energy Consumption and Generation Forecasting for Time-based Decisions on Demand.
Smart Cities 2025,8, 34 10 of 14
4.3. Experiment 3: Energy Savings Profiling
Standard building profiling or segmentation is further developed in this experiment,
by constructing automated comparison of building energy profiles against consumption
norms as determined across building function, building location, time period, and locality
of building within pre-determined classifications. As illustrated in Figure 7, the selection
of the date and the building will generate the bar graph on the left which provides a
visualisation of the percentage opportunity and Kwh for energy savings compared to
median consumption. The bar graph presents such opportunity across the whole year
which is a significant insight for downstream actions and decisions conducted by the
agentic AI layer. The agentic layer can further funnel through the entire buildings or a
subset to determine energy savings on daily, weekly, monthly, or annual cycles, and this
information drives decisions related to operations, planning, budgeting, and maintenance.
Figure 7. Evaluating the impact of events in the building on energy consumption.
4.4. Experiment 4: Human Physical Activity Monitoring
For further validation, the framework was empirically evaluated on a benchmark
dataset for human physical activity monitoring, using the PAMAP2 (Physical Activity
Monitoring for Aging People V2) dataset for Physical Activity Monitoring [
47
,
48
]. This
dataset established benchmarking for physical activity monitoring of 12 activities (lie, sit,
stand, walk, run, cycle, Nordic walk, iron, vacuum clean, rope jump, and ascend and
descend stairs). The data was collected through a heart rate monitor and three inertial
measurement units (IMU) worn on the hand, chest, and ankle. The IMU measures spe-
cific force, angular rate, the magnetic field surrounding the body using a combination
of accelerometers, gyroscopes, and magnetometers. This multivariate time series dataset
consists of more than 3.8 million data records timestamp, 52 attributes of raw sensory data,
and the ground truth activity label.
The framework initiated data quality and integrity checks through its agentic AI
layer, leading to the removal of transient activities which are coded with class ‘0’, removal
of records with missing values due to the loss of wireless network connectivity, and re-
moval of acceleration data with the scale of ±6 g resolution during high impact activities.
Subsequently, the agents initiated the determination of new features calculated from the
triaxial signals of each IMU, including triaxial acceleration, acceleration magnitude, etc.,
so that records from one subject are processed as a single data stream is used in a single
instance.
Figure 8
presents the initial results from the framework, the segmentation into
high-intensity vs. low-intensity
activities and the granular breakdown into activity types.
This was followed by the incremental learning capability that determined the sequential
Smart Cities 2025,8, 34 11 of 14
flow of activities, as shown in Figure 9. These results can be interpreted as follows: Vacuum
cleaning has a high variation due to a gradual increase in heart rate which is recognised by
a higher maturity in the nodes: vacuum cleaning (w: 0.02); walking (w: 0.07); rope jumping
(w: 0.07); and running (w: 0.07). Similarly, activities with a smaller variance such as sitting
(w: 0.26), standing (w: 0.41), and ironing (w: 0.26) are learned faster. Ascending stairs (w:
0.21) activity has less variance compared to descending stairs (w: 0.17) while cycling and
Nordic walking achieved lower maturity as the activities were carried out only for a brief
period. For this experiment, we have only presented the visualisation of segments and
drift detection outcomes produced by the framework. The development of performance
metrics that can evaluate segments and drift detection accuracies are a work in progress at
this stage.
Figure 8. PAMAP 2 Results 1: Segmentation of High vs. Low Intensity Activities.
Figure 9. PAMAP2 Results 2: Incrementally Learned Sequential Flow of Activities.
5. Conclusions
This paper presents the development of a hyper-automated reponsible AI framework
with Generative AI agents for sustainable smart cities. The expanding challenges of HCDEs
within smart cities means the extraction, analysis, translation of data streams, and repos-
itories of HCDEs through to actionable insights poses several complex challenges of the
volume of data, disconnected source systems, frequency of decisions, privacy, confiden-
Smart Cities 2025,8, 34 12 of 14
tiality, and bias. This framework is built on the dimensions of acquisition, preparation,
orchestration, dissemination, retrospection, with an independent cognitive engine for
hyper-automation of these AI capabilities using Generative AI agents. Hyper-automation
outcomes are further reviewed by human-in-the-loop processes prior to decision-making
outcomes. The framework was evaluated in the smart city living lab setting of La Trobe’s
‘University City of the Future’. Across four experiments, human mobility prediction, hu-
man activity monitoring, energy forecasting for consumption and generation, and energy
profiling by buildings, the technical capabilities and practical value of the AI framework
are empirically validated. The framework establishes a prototypical setting for university
cities of the future to provide direction, guidance, and standards for the development and
management of sustainable and safe smart cities. Future research can extend this validation
by implementing the framework in broader, diverse, and non-deterministic urban contexts.
Author Contributions: Conceptualization, D.D.S. and A.J.; Methodology, D.D.S., N.M., H.M., S.W.
and A.J.; Software, N.M., H.M. and P.R.; Validation, N.M., H.M. and P.R.; Formal analysis, D.D.S.,
P.R., S.W. and A.J.; Investigation, D.D.S., N.M., P.R. and S.W.; Resources, S.W.; Writing—original draft,
D.D.S., N.M., H.M., P.R. and S.W.; Supervision, D.D.S. and A.J. All authors have read and agreed to
the published version of the manuscript.
Funding: This work is supported by the Department of Climate Change, Energy, the Environment
and Water of the Australian Federal Government, as part of the International Clean Innovation
Researcher Networks (ICIRN) program, grant number ICIRN000077.
Data Availability Statement: Data available upon request.
Conflicts of Interest: The authors declare no conflicts of interest.
References
1.
The United Nations General Assembly. Transforming Our World: The 2030 Agenda for Sustainable Development. 2015. Available
online: https://sdgs.un.org/2030agenda (accessed on 1 July 2024).
2.
Griggs, D.; Stafford-Smith, M.; Gaffney, O.; Rockström, J.; Öhman, M.C.; Shyamsundar, P.; Steffen, W.; Glaser, G.; Kanie, N.; Noble,
I. Sustainable development goals for people and planet. Nature 2013,495, 305–307. [CrossRef] [PubMed]
3.
Independent Group of Scientists Appointed by the Secretary-General. Global Sustainable Development Report 2023: Times
of Crisis, Times of Change: Science for Accelerating Transformations to Sustainable Development. 2023. Available online:
https://desapublications.un.org/file/1182/download (accessed on 1 July 2024).
4.
Ramaprasad, A.; Sánchez-Ortiz, A.; Syn, T. A unified definition of a smart city. In Proceedings of the Electronic Government:
16th IFIP WG 8.5 International Conference, EGOV 2017, Proceedings 16, St. Petersburg, Russia, 4–7 September 2017; Springer:
Berlin/Heidelberg, Germany, 2017; pp. 13–24.
5.
Blasi, S.; Ganzaroli, A.; De Noni, I. Smartening sustainable development in cities: Strengthening the theoretical linkage between
smart cities and SDGs. Sustain. Cities Soc. 2022,80, 103793. [CrossRef]
6.
Albino, V.; Berardi, U.; Dangelico, R.M. Smart cities: Definitions, dimensions, performance, and initiatives. J. Urban Technol. 2015,
22, 3–21. [CrossRef]
7.
Silva, B.N.; Khan, M.; Han, K. Towards sustainable smart cities: A review of trends, architectures, components, and open
challenges in smart cities. Sustain. Cities Soc. 2018,38, 697–713. [CrossRef]
8.
Lim, C.; Kim, K.J.; Maglio, P.P. Smart cities with big data: Reference models, challenges, and considerations. Cities 2018,82, 86–99.
[CrossRef]
9.
Organisation for Economic Co-Operation and Development. The State of Implementation of the OECD AI Principles Four Years
on 2023. Available online: https://www.oecd.org/en/publications/the-state-of-implementation-of-the-oecd-ai-principles-four-
years-on_835641c9-en.html (accessed on 1 July 2024). [CrossRef]
10.
Gamage, G.; Kahawala, S.; Mills, N.; De Silva, D.; Manic, M.; Alahakoon, D.; Jennings, A. Augmenting industrial chatbots
in energy systems using chatgpt generative ai. In Proceedings of the 2023 IEEE 32nd International Symposium on Industrial
Electronics (ISIE), Espoo, Finland, 19–21 June 2023; IEEE: Piscataway, NJ, USA, 2023; pp. 1–6.
11.
Adikari, A.; De Silva, D.; Ranasinghe, W.K.; Bandaragoda, T.; Alahakoon, O.; Persad, R.; Lawrentschuk, N.; Alahakoon, D.; Bolton,
D. Can online support groups address psychological morbidity of cancer patients? An artificial intelligence based investigation
of prostate cancer trajectories. PLoS ONE 2020,15, e0229361. [CrossRef] [PubMed]
Smart Cities 2025,8, 34 13 of 14
12.
De Silva, D.; Burstein, F.; Jelinek, H.F.; Stranieri, A. Addressing the complexities of big data analytics in healthcare: The diabetes
screening case. Australas. J. Inf. Syst. 2015,19. [CrossRef]
13.
Sumanasena, V.; Gunasekara, L.; Kahawala, S.; Mills, N.; De Silva, D.; Jalili, M.; Sierla, S.; Jennings, A. Artificial Intelligence for
Electric Vehicle Infrastructure: Demand Profiling, Data Augmentation, Demand Forecasting, Demand Explainability and Charge
Optimisation. Energies 2023,16, 2245. [CrossRef]
14.
De Silva, D.; Yu, X.; Alahakoon, D.; Holmes, G. Semi-supervised classification of characterized patterns for demand forecasting
using smart electricity meters. In Proceedings of the 2011 International Conference on Electrical Machines and Systems, Beijing,
China, 20–23 August 2011; IEEE: Piscataway, NJ, USA, 2011; pp. 1–6.
15.
Wolniak, R.; Stecuła, K. Artificial Intelligence in Smart Cities—Applications, Barriers, and Future Directions: A Review. Smart
Cities 2024,7, 1346–1389. [CrossRef]
16.
Smuha, N.A. From a ‘race to AI’ to a ‘race to AI regulation’: Regulatory competition for artificial intelligence. Law Innov. Technol.
2021,13, 57–84. [CrossRef]
17.
Wang, W.; Chen, L.; Xiong, M.; Wang, Y. Accelerating AI adoption with responsible AI signals and employee engagement
mechanisms in health care. Inf. Syst. Front. 2023,25, 2239–2256. [CrossRef]
18.
Arrieta, A.B.; Díaz-Rodríguez, N.; Del Ser, J.; Bennetot, A.; Tabik, S.; Barbado, A.; García, S.; Gil-López, S.; Molina, D.; Benjamins,
R.; et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Inf.
Fusion 2020,58, 82–115. [CrossRef]
19. Finocchiaro, G. The regulation of artificial intelligence. AI Soc. 2024,39, 1961–1968. [CrossRef]
20. Ray, S.; Tornbohm, C.; Kerremans, M.; Miers, D. Move Beyond RPA to Deliver Hyperautomation. Gart. Dec. 2019, 1–16.
21.
Acharya, D.B.; Kuppan, K.; Divya, B. Agentic AI: Autonomous Intelligence for Complex Goals—A Comprehensive Survey. IEEE
Access 2025,13, 18912–18936. [CrossRef]
22. Brynjolfsson, E.; Li, D.; Raymond, L. Generative AI at work. Q. J. Econ. 2025, qjae044.. [CrossRef]
23. Xie, J.; Chen, Z.; Zhang, R.; Wan, X.; Li, G. Large multimodal agents: A survey. arXiv 2024, arXiv:2402.15116.
24.
Osipov, E.; Kahawala, S.; Haputhanthri, D.; Kempitiya, T.; De Silva, D.; Alahakoon, D.; Kleyko, D. Hyperseed: Unsupervised
learning with vector symbolic architectures. IEEE Trans. Neural Netw. Learn. Syst. 2022,35, 6583–6597. [CrossRef] [PubMed]
25.
Cremonini, L.; Carotenuto, F.; Famulari, D.; Fiorillo, E.; Nardino, M.; Neri, L.; Georgiadis, T. Urban Environment and Human
Health: Motivations for Urban Regeneration to Adapt. Environ. Sci. Proc. 2023,27, 28. [CrossRef]
26.
Lai, C.S.; Jia, Y.; Dong, Z.; Wang, D.; Tao, Y.; Lai, Q.H.; Wong, R.T.; Zobaa, A.F.; Wu, R.; Lai, L.L. A review of technical standards
for smart cities. Clean Technol. 2020,2, 290–310. [CrossRef]
27.
Gracias, J.S.; Parnell, G.S.; Specking, E.; Pohl, E.A.; Buchanan, R. Smart cities—A structured literature review. Smart Cities 2023,
6, 1719–1743. [CrossRef]
28.
Moreno, C.; Allam, Z.; Chabaud, D.; Gall, C.; Pratlong, F. Introducing the “15-Minute City”: Sustainability, resilience and place
identity in future post-pandemic cities. Smart Cities 2021,4, 93–111. [CrossRef]
29.
Syed, A.S.; Sierra-Sosa, D.; Kumar, A.; Elmaghraby, A. IoT in smart cities: A survey of technologies, practices and challenges.
Smart Cities 2021,4, 429–475. [CrossRef]
30.
Nikitas, A.; Michalakopoulou, K.; Njoya, E.T.; Karampatzakis, D. Artificial intelligence, transport and the smart city: Definitions
and dimensions of a new mobility era. Sustainability 2020,12, 2789. [CrossRef]
31.
Cugurullo, F. Urban artificial intelligence: From automation to autonomy in the smart city. Front. Sustain. Cities 2020,2, 38.
[CrossRef]
32.
Dashkevych, O.; Portnov, B.A. How can generative AI help in different parts of research? An experiment study on smart cities’
definitions and characteristics. Technol. Soc. 2024,77, 102555. [CrossRef]
33.
Xu, H.; Omitaomu, F.; Sabri, S.; Li, X.; Song, Y. Leveraging Generative AI for Smart City Digital Twins: A Survey on the
Autonomous Generation of Data, Scenarios, 3D City Models, and Urban Designs. arXiv 2024, arXiv:2405.19464.
34.
Wolniak, R.; Gajdzik, B.; Grebski, M.; Danel, R.; Grebski, W.W. Business models used in smart cities—Theoretical approach with
examples of smart cities. Smart Cities 2024,7, 1626–1669. [CrossRef]
35.
Shulajkovska, M.; Smerkol, M.; Noveski, G.; Gams, M. Enhancing Urban Sustainability: Developing an Open-Source AI
Framework for Smart Cities. Smart Cities 2024,7, 2670–2701. [CrossRef]
36.
Fera, F.T.; Spandonidis, C. An Artificial Intelligence and Industrial Internet of Things-Based Framework for Sustainable
Hydropower Plant Operations. Smart Cities 2024,7, 496–517. [CrossRef]
37.
Alharbi, E.; Cherif, A.; Nadeem, F. Adaptive Smart eHealth Framework for Personalized Asthma Attack Prediction and Safe
Route Recommendation. Smart Cities 2023,6, 2910–2931. [CrossRef]
38.
Hayajneh, A.M.; Aldalahmeh, S.A.; Alasali, F.; Al-Obiedollah, H.; Zaidi, S.A.; McLernon, D. Tiny machine learning on the edge:
A framework for transfer learning empowered unmanned aerial vehicle assisted smart farming. IET Smart Cities 2024,6, 10–26.
[CrossRef]
Smart Cities 2025,8, 34 14 of 14
39.
Ismagilova, E.; Hughes, L.; Rana, N.P.; Dwivedi, Y.K. Security, privacy and risks within smart cities: Literature review and
development of a smart city interaction framework. Inf. Syst. Front. 2022,24, 1–22. [CrossRef]
40.
Westraadt, L.; Calitz, A. A modelling framework for integrated smart city planning and management. Sustain. Cities Soc. 2020,
63, 102444. [CrossRef]
41.
Herath, H.; Mittal, M. Adoption of artificial intelligence in smart cities: A comprehensive review. Int. J. Inf. Manag. Data Insights
2022,2, 100076. [CrossRef]
42.
Ullah, Z.; Al-Turjman, F.; Mostarda, L.; Gagliardi, R. Applications of artificial intelligence and machine learning in smart cities.
Comput. Commun. 2020,154, 313–323. [CrossRef]
43.
Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. Training
language models to follow instructions with human feedback. Adv. Neural Inf. Process. Syst. 2022,35, 27730–27744.
44.
Mills, N.; Rathnayaka, P.; Moraliyage, H.; De Silva, D.; Jennings, A. Cloud edge architecture leveraging artificial intelligence
and analytics for microgrid energy optimisation and net zero carbon emissions. In Proceedings of the 2022 15th International
Conference on Human System Interaction (HSI), Melbourne, Australia, 28–31 July 2022; IEEE: Piscataway, NJ, USA, 2022; pp. 1–7.
45.
La Trobe University. La Trobe University—Facts and Figures. 2024. Available online: https://www.latrobe.edu.au/about/at-a-
glance/facts-figures (accessed on 11 November 2024).
46.
Moraliyage, H.; Mills, N.; Rathnayake, P.; De Silva, D.; Jennings, A. Unicon: An open dataset of electricity, gas and water
consumption in a large multi-campus university setting. In Proceedings of the 2022 15th International Conference on Human
System Interaction (HSI), Melbourne, Australia, 28–31 July 2022; IEEE: Piscataway, NJ, USA, 2022; pp. 1–8.
47.
Reiss, A.; Hendeby, G.; Stricker, D. Confidence-based multiclass AdaBoost for physical activity monitoring. In Proceedings of the
2013 International Symposium on Wearable Computers, Zurich, Switzerland, 8–12 September 2013; pp. 13–20.
48.
Reiss, A.; Stricker, D. Introducing a new benchmarked dataset for activity monitoring. In Proceedings of the 2012 16th International
Symposium on Wearable Computers, Newcastle, UK, 18–22 June 2012; IEEE: Piscataway, NJ, USA, 2012; pp. 108–109.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
Available via license: CC BY
Content may be subject to copyright.
