Indoor navigation has turned into a critical area for research on account of its applications across multiple fields, along with robotics, healthcare, as well as smart buildings. Compared with outdoor navigation, interior settings pose some special problems such as fewer GPS signals, detailed arrangements, and moving impediments. This paper provides a detailed look at multiple methods used for finding routes in indoor spaces, like graph-based methods, probabilistic methods, and machine learning methods. We evaluate these algorithms based on their own accuracy, computational efficiency, scalability, and robustness across multiple indoor scenarios. The paper discusses the strengths and limitations with each approach and provides understandings into future research directions within the field.
Nowadays navigation is an important task that is embedded in every aspect of our daily routines.
Navigation is generally defined as the process of directing the movements of a vehicle, nave, plane,
and people from one place to another [
Indoor navigation i.e., localization, is the process of determining the location and orientation of a
person or object inside a building and guiding them to a place of interest in that location. Applications
for this technology range from autonomous robots to assistance technologies for visually impaired
people to navigation in smart buildings [
With satellite-based devices currently available for precise position data, outside navigation is
far easier. However, indoor navigation presents a unique set of challenges that call for a different
approach. For instance, while GPS performs exceptionally well outdoors, it may not work well indoors
because to elements including signal attenuation, multipath, and dynamic spillway development [
Thus, indoor navigation has received much interest, especially in the scenarios of smart
buildings, healthcare facilities, and commercial spaces. Due to the failure of GPS-based localization
to work indoors, different solutions have been developed, using Wi-Fi fingerprinting, Bluetooth, and
sensors [
The subject of the study was to compare different algorithm types and summarize their pros and cons, as well as best practices in environments suitable for different solutions. To solve these 6th International Conference Recent Trends and Applications in Computer Science and Information Technology ∗ Corresponding author. † These authors contributed equally.
erisa.bekteshi@unisalento.it (E. Bekteshi); claudio.pascarelli@unisalento.it (C. Pascarelli) 0000-0002-0678-979X (E. Bekteshi); 0000-0002-9854-7703 (C. Pascarelli)
© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). problems, indoor navigation systems were evaluated using a systematic approach that incorporates advanced localization technologies and algorithmic frameworks, as outlined in the subsequent section.
This study adopts a narrative review approach to explore and synthesize the current state of research concerning traditional and machine learning algorithms for indoor navigation. A narrative review is a recognized methodology for providing a comprehensive and critical overview of a research topic without the rigid procedural constraints typical of systematic reviews or bibliometric analyses. It is particularly suited to fields where the literature is heterogeneous in terms of methodologies, outcomes, and technological focus, as is the case for indoor navigation.
The literature considered in this review was identified through searches conducted in major
academic databases, including Google Scholar, ScienceDirect, and Web of Science [
While care was taken to prioritize peer-reviewed journal articles and recent conference
proceedings [
Through this methodological lens, the study seeks to compare the traditional indoor navigation methods—such as SLAM-based approaches and classical path-planning algorithms—with more recent machine learning-based techniques, particularly deep reinforcement learning (DRL), highlighting the relative advantages, limitations, and future research directions.
To achieve indoor navigation that is as seamless as outdoor navigation, the integration of advanced
technologies is essential. The primary methods adopted are Wi-Fi-based positioning systems, LiDAR,
visual and deep learning enhanced SLAM (Simultaneous Localization and Mapping). At large, these
technologies increase the accuracy, robustness, and adaptability of systems in indoor environments,
improving overall user experience [
To enhance the user experience, these improvements focus on accuracy, robustness, and
adaptability in complex indoor environments. For instance, many 3D deep learning methods today
aim to leverage technical progress from robotics and autonomous driving to consume less energy
and perform in real-time by processing raw point cloud data [
Deep Learning neural networks (DNNs) have excelled at the understanding and extraction of
a high level of intelligence for outlandish datasets, such as point clouds, which can achieve tasks
such as object detection, semantic segmentation, and scene reconstruction [
These technologies serve as the foundation for the algorithmic approaches compared in the next section, ranging from traditional SLAM to data driven DRL methods.
Indoor navigation presents several challenges; nonetheless, certain algorithms can assist in overcoming these obstacles. These algorithms are categorized into traditional and deep reinforcement learning (DRL)-based algorithms.
Traditional methods include SLAM, global planning, and local planning. SLAM (Simultaneous
Localization and Mapping) is one of the most common approaches employed in robotic indoor
applications, where algorithms create maps of the environment and localize a robot at the same
time, using the position information gathered by LiDAR, cameras, Wi-Fi, or Ultra-Wideband (UWB)
[
Visual SLAM, using camera systems, is more affordable and widely used in aerial vehicles and
mobile robots, but can suffer from reduced performance in low light or with reflective surfaces
[
Traditional autonomous navigation often employs Simultaneous Localization and Mapping
(SLAM) to build a map of the environment while simultaneously estimating the robot’s pose within
that map. Algorithms like Karto-SLAM, which is based on graph optimization, are used in this process.
Global planning then uses this map to find an optimal route from a starting point to a goal, often
using algorithms like Dijkstra’s algorithm as implemented in the Navfn planner. Local planning,
such as with the Dynamic Window Approach (DWA) or Timed-Elastic-Band (TEB), then adjusts this
global plan in real-time to avoid obstacles and account for dynamic changes in the environment [
Traditional path – planning algorithms, such as Dynamic Window Approach (DWA) and Times
Elastic Band (TEB), offer several strengths in indoor navigation. These strengths excel in path planning
and efficiency, with DWA providing high temporal efficiency and shorter routes, particularly in
environments where line-of-sight (LOS) conditions are sufficient and lastly can provide a quick path
calculation in simpler environments [
Another strength in traditional algorithms is in the safety features where TEB can generate a
route with the least number of collision while maintaining safe distances from obstacles and making
it highly effective in static or well-mapped environments [
From an implementation point of view, these algorithms are easier to implement than AI-based
algorithms, as they do not require extensive training data or high computational resources, ensuring
predictable behavior in structured scenarios [
They have also been used effectively with pre – existing maps, resulting in reliable performance
when precise environmental data are available [
These strengths make traditional methods ideal for controlled environments, but their strict dependence on the static maps and limited adaptability in dynamic environments highlight the need for complementary approaches, such as machine learning in more complex scenarios.
Despite their strength, traditional algorithms like DWA and TEB offers limitations in real
world environments. One of the most important limitations is the environmental adaptability: these
algorithms perform poorly in dynamic environments, struggling with unpredictable obstacles (e.g.
sudden pedestrian movement), and cannot generalize well to new situations [
Sensor Dependencies further deepens these algorithms to not be more reliable because they are
heavily reliant on sensor precision and accuracy, wheel odometer accumulates errors due to slipping,
LiDAR suffers in featureless corridors (“corridor effect”), necessitating redundant sensor arrays to
maintain line-of-sight (LOS) conditions [
Notably, performance degrades in sufficiently complex scenarios: for example, while performing DWA, collision rates increase by ~40% in cluttered environments; and while performing TEB, computational latency increases exponentially with the number of obstacles in a situation.
These algorithms do not learn; they remain inflexible to changes in their environment without
being manually recalibrated. Compounding these problems is an infrastructure burden, where
specific pre-mapping and regular upkeep increase deployment costs by an order of magnitude of
2–3× versus data-driven alternatives. These constraints highlight the reason why modern systems
are moving toward hybrid architectures, merging the interpretability of traditional methods with the
flexibility that AI affords to create a balance between stability and flexibility [
DRL-based approaches employ agents to acquire optimal navigation policies via interaction
with the environment, providing adaptability to dynamic changes and intricate circumstances. In
mobile robotics, a Deep Reinforcement Learning (DRL) agent can acquire navigation skills within a
warehouse setting by getting feedback, either incentives or penalties, contingent upon its actions, such
as advancing, turning, or halting [
This adaptability is especially beneficial in intricate situations where conventional rule-based
navigation systems may struggle, such as congested surroundings or regions with erratic human
behavior [
This is distinct from classic approaches that require extensive manual tuning and reprogramming whenever the environment changes. Furthermore, DRL algorithms can utilize transfer learning approaches to expedite the learning process in novel situations. A robot taught to navigate one warehouse can swiftly adjust to a different layout by refining its existing policies instead of starting from the beginning.
Deep reinforcement learning algorithms offers several strengths in indoor navigation systems,
such as environmental adaptability by learning optimal navigation policies through continuous
interaction with both static and dynamic environments without relying on pre – existing maps or
precision sensors [
A distinguish feature of this algorithm is its ability to effectively integrate and execute complex
decision-making task through their trial – and – error, capabilities, utilizing policies, reward, and
value function to maximize performance with approaches like Soft Actor-Critic (SAC) showing
particularly efficient sample usage and lower collision rates compared to traditional methods [
Another feature that deep reinforcement learning algorithms has that the traditional algorithms
does not have is the benefit of the performance such as SAC demonstrating efficient sample usage
and lower collision rate, has better computational efficiency, achieves higher rewards in testing
scenarios, can function effectively in maples environments. [
The configurability of DRL architectures allows for both model – free and model – based
configurations along with the capability to handle continuous or discrete action spaces and hence
allows developers a great deal of flexibility when devising their architectures, depending on their
specific application requirements. These qualities allow DRL to perform well in complex, real-world
navigation tasks, especially when environmental unpredictability is a major factor [
However, while being powerful and adaptable to the environment, DRL algorithms face major
obstacles such as exceedingly virtual trial and error paths, unavoidable collision happening in the
training process, along with high requirement of calculating resources, resulting in difficulty to
implement it in real world problems [
Moreover, their significant reliance on the large, heterogeneous datasets and the need to collect
location-related sensitive information raises privacy and security issues receiving the attention
of scholars and practitioners [
The flexibility of DRL architectures supports both model – free and model – based designs, with
the ability to manage continuous or discrete action spaces, offering developers a broad spectrum of
configurations based on specific application needs. These attributes make DRL particularly suitable
for complex, real-world navigation tasks, especially where environmental unpredictability is a
challenge [
Despite their strength and adaptability of the environment, DRL algorithms face significant
challenges, including extensive virtual training requirements, inevitable collisions during the
training phase, and substantial computational resource demands, which complicate real-world
implementation [
Additionally, their performance depends heavily on large, heterogeneous datasets and raises
privacy and security concerns due to the collection of sensitive location information [
In a summarized way comparison of indoor navigation algorithms discussed are presented in table 1.
Indoor navigation remains a complex and evolving field, demanding solutions that can navigate challenges such as dynamic environments, occluded signals, and diverse spatial layouts. This study has compared traditional algorithms—such as SLAM, DWA, and TEB—with machine learningbased approaches, particularly Deep Reinforcement Learning (DRL), to highlight their respective strengths and limitations. Traditional methods excel in controlled, static environments with high predictability, offering low-cost, easy-to-implement solutions with high interpretability. However, their dependency on precise mapping and limited adaptability restricts their effectiveness in realworld, dynamic scenarios.
In contrast, DRL-based algorithms demonstrate significant advantages in adaptability and autonomous decision-making by learning from real-time interactions with the environment. Their ability to operate without pre-mapped data and adapt policies through trial-and-error makes them highly suitable for complex and unpredictable settings. Nevertheless, DRL faces considerable implementation challenges, including high training costs, computational demands, and privacy concerns related to data acquisition.
environments training dataset
The findings emphasize the growing need for hybrid navigation architectures that blend the robustness and interpretability of traditional algorithms with the flexibility and learning capabilities of AI-based methods. As indoor navigation continues to expand across domains such as smart infrastructure, robotics, and assistive technologies, future research should focus on optimizing hybrid solutions, improving training efficiency, and addressing data privacy through federated learning and secure data management practices. Such advancements will be critical for enabling reliable, scalable, and context-aware indoor navigation systems in the years to come.
During the preparation of this work, the author(s) used X-GPT-4 and Gramby in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.