AI-Powered Personalized Path Planning Using User Behavioural Patterns and Semantic Context for Smart Navigation Systems

Saranya Chandrabose  and Janaki Gopalakrishnan
Department of Electrical and Electronics Engineering SRM Institute of Science & Technology, Kattankulathur, Tamil Nadu, India
Correspondence to: Saranya Chandrabose, saranyac1@srmist.edu.in Janaki G, janakig@srmist.edu.in

DOI: https://doi.org/10.70389/PJS.100149

Additional information

• Ethical approval: The proposed SBRF framework depends on the collection and analysis of user trajectory data, which presents significant ethical considerations concerning data privacy and user consent. This study employed synthetically generated user traces to mitigate privacy concerns; however, real-world implementation necessitates careful consideration of various ethical dimensions.
• Consent: Users are provided with granular control over what behavioral data they share, with options to dynamically adjust privacy preferences and the influence of behavioral personalization. Explainable routing decisions show how preferences affect path selection. Privacy impact assessments are regularly conducted to monitor and mitigate risks.
• Funding: This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
• Conflicts of interest: N/a
• Author contribution: Saranya Chandrabose and Janaki Gopalakrishnan – Conceptualization, Writing – original draft, review and editing.
• Guarantor: Saranya Chandrabose
• Provenance and peer-review: Unsolicited and externally peer-reviewed
• Data availability statement: N/a

Keywords: Adaptive context-aware a* (aca*) algorithm, User behaviour-driven cost modelling, Semantic environmental layer integration, Personalized cost grid construction, Behaviour-adaptive routing framework.

Peer Review
Received: 13 August 2025
Last revised: 28 September 2025
Accepted: 3 October 2025
Version accepted: 5
Published: 3 November 2025

Plain Language Summary Infographic

Introduction

Smart navigation systems play a crucial role in the development of intelligent transport infrastructure by providing real-time routing, traffic management, and improved mobility for both personal and commercial use. The integration of AI into navigation systems has been extensively studied,¹ with applications spanning autonomous vehicles, smart cities, and context-aware mobility services though critical perspectives on AI-driven autonomous navigation highlight the importance of safety and reliability considerations,² accelerating innovation in logistics, public safety, and urban planning. These systems are critical to the development of self-driving cars, smart cities, and context-aware mobility services, accelerating innovation in logistics, public safety, and urban planning. By 2025, the global market for AI-driven navigation and mobility solutions is projected to exceed USD 14.3 billion.³ The increasing adoption of autonomous vehicles and smart city frameworks demands scalable, adaptive, and user-centric routing systems. As cities grow more complex and dynamic, personalized navigation strategies are no longer optional—they are essential.

Traditional path planning algorithms, including A* and Dijkstra’s, provide efficient and reliable solutions for shortest-path problems; however, they are fundamentally static.⁴ These analyses neglects dynamic factors, including individual driving preferences, semantic environmental data (e.g., weather, traffic congestion), and risk zones (e.g., construction, events). This limitation diminishes their relevance in practical situations where contextual awareness is essential. To address the shortcomings of static and context-insensitive route planning approaches, this research presents a behavior-aware and semantically enhanced routing strategy. By combining user-specific behavioral insights with real-time semantic overlays, the system provides adaptable, efficient, and preference-aligned navigation in dynamic situations. The key contributions of this paper are:

1. A modular AI-driven route planning system divided into seven stages, each addressing a vital level of behavioral and semantic routing integration.
2. A novel integration of user behavior modeling with semantic context layers, including traffic, events, and road types, to facilitate adaptive decision-making in real time.
3. Construction of dynamic, personalized cost grids that integrate traditional heuristics with behavioral weights and environmental constraints.
4. Comprehensive experimental assessment across 20 × 20, 30 × 30, and 50 × 50 grid scenarios indicates a cost reduction of up to 18.18% and a 33.33% enhancement in planning time, thereby validating performance and scalability.

Figure 1 depicts the overarching system architecture of a smart navigation system, incorporating all essential components necessary for real-time, context-aware route planning. The process commences with User or Vehicle Input, whereby the system records dynamic characteristics like destination requests or driving behaviors. The Sensor & IoT Data Layer enhances this input by aggregating real-time data from sources including GPS, LIDAR, and traffic APIs, thereby facilitating environmental awareness. The Data Processing Engine analyzes incoming data and executes map matching to ensure spatial consistency. The Contextual Information Hub concurrently incorporates semantic environmental data such as traffic congestion, weather updates, and road closures. User-specific preferences, including preferred routes and safety limits, are administered via the User Preference Module. The Route Planning Engine integrates these data, using techniques such as A* and cost function models to determine best pathways. The Recommendation System finalizes individualized routing possibilities, which are then represented via the User Interface, facilitating clear route presentation and turn-by-turn navigation assistance.

Related Work

ACA* presents distinct differences from current weighted A* variants in three primary areas: (1) Dynamic cost fusion: In contrast to static weighted A*, ACA* utilizes real-time behavioral-semantic cost fusion as described in Equation (11), incorporating adaptive β parameters. Dual-layer personalization: Hierarchical A* employs semantic layers or preference weights, whereas ACA* concurrently integrates both by utilizing user-specific behavioral vectors B_u and contextual overlays C_semantic. Behavioral adaptivity: Traditional context-aware A* variants depend on predefined rules, while ACA* utilizes user trajectory patterns T_u to dynamically modify routing preferences, facilitating personalized navigation that adapts to user behavior instead of adhering to static multi-criteria optimization.

Recent advancements in AI-driven navigation have notably impacted the design of path planning for intelligent systems. Selvaraj et al.⁵ introduced a context-aware route planning framework for edge-assisted vehicles that adjusts according to environmental and user parameters. Okolo et al.⁶ developed a smart assistive navigation system for individuals with visual impairments, emphasizing the potential for personalized systems. Similar applications have been investigated in smart campuses by M. S et al.⁷ and in UAV navigation utilizing explainable AI methodologies by the Frontiers Sustainable Cities Editorial Office.⁸ Similar AI-driven approaches have been explored in aircraft navigation systems.⁹ Mohanty et al.¹⁰ presented a detailed survey on semantic navigation for indoor robots, underlining the relevance of semantic understanding in indoor environments. Enhancements in exploration strategies through semantic cues were introduced by Goel et al.,¹¹ while Zhang et al.¹² improved map-based positioning through semantic context integration. The impact of individual behavioral characteristics was analyzed by Alamri et al.¹³ using personality traits in online learning environments, and by van der Putten et al.¹⁴ through case studies on dementia patients’ wayfinding behavior. Huidobro et al.¹⁵ further expanded behavior modeling into web navigation, proposing a high-level abstraction for analyzing digital movement patterns.

Rivero et al.¹⁶ conducted a systematic review of the use of AI in autonomous navigation in UAV swarms, while Tang et al.¹⁷ examined mobile robot trends. Li et al.¹⁸ investigated industrial robots through the application of rapidly exploring random trees (RRT), whereas Lee et al.¹⁹ highlighted the role of semantic inference systems in the context of smart cities. Eshraghi et al.²⁰ proposed iRouteSafe, a routing mechanism based on risk profiles that adjusts to individual driving behaviors in the field of personalized navigation. Bayly-Castaneda et al.²¹ examine the parallels in education, focusing on the role of AI in developing lifelong personalized learning pathways.²² The theoretical foundations of semantic context modeling were established by Faruffini and Abel,²³ who advocated for its integration in adaptive systems. Xu et al.²⁴ tackled accuracy issues in navigation by introducing adaptive Kalman filtering for INS/GPS systems. Shoukat et al.²⁵ and Chohan et al.²⁶ investigated self-learning routing and intrusion detection in autonomous UAV networks, demonstrating how learning processes may enhance routing intelligence and security in interconnected aerial systems.

These studies collectively illustrate a transition from generic navigation to adaptive, AI-driven strategies that utilize semantic context and user behavior.²⁷ Selvaraj et al.⁵ demonstrated that their method surpassed traditional shortest-path algorithms through dynamic adaptation to edge data. Goel et al.¹¹ and Zhang et al.¹² demonstrated how semantically aware exploration improves both efficiency and orientation accuracy. Behavioral understanding, as explored by Alamri et al.,¹³ van der Putten et al.,¹⁴ and Eshraghi et al.,²⁰ enables systems to predict and personalize user movement, making navigation not only more efficient but also safer and user-centric. Applications currently encompass a variety of domains, including smart cities, education, web navigation, UAVs, and industrial robotics, underscoring the necessity for intelligent systems that integrate contextual awareness with adaptive learning. This convergence represents a significant development in smart navigation systems, facilitating advancements in user-specific, AI-enhanced path planning. The literature study has revealed the subsequent research gaps:

1. Many current path planning methods fail to integrate user-specific behavioral modeling, resulting in generic and less adaptive routing results.
2. Semantic contextual factors, including traffic density, event zones, and road classifications, are inadequately incorporated into real-time decision-making models.
3. Traditional grid-based planners exhibit difficulties in scalability and efficiency when integrating dynamic cost elements affected by user behavior and environmental limitations.

To address the identified gaps, this study proposes the Semantic-Aware Behavioral Routing Framework (SBRF), a novel AI-powered solution for real-time path planning that combines user behavior patterns and semantic environmental data. SBRF utilizes adaptive cost grids and context-aware prioritizing to surpass the inflexibility of conventional planners. It augments customization, promotes decision-making in dynamic environments, and facilitates scalable assessment across several map dimensions. The comprehensive framework is outlined in Section ‘Proposed Semantic-Aware Behavioral Routing Framework’.

Proposed Semantic-Aware Behavioral Routing Framework

This research presents the Semantic-Aware Behavioral Routing Framework (SBRF), a modular and extensible architecture aimed at overcoming the shortcomings of traditional navigation systems that neglect user preferences and contextual semantics. SBRF combines user behavior modeling with real-time contextual data, allowing adaptive and intelligent path planning customized to individual requirements. The framework employs a staged methodology to methodically analyze behavioral patterns, develop customized cost models, and calculate optimal routes with AI-enhanced algorithms. This framework enables the system to dynamically adjust to user and environmental variations, offering a scalable solution for intelligent navigation in autonomous systems.

Figure 2 depicts that the SBRF design consists of seven successive stages. The procedure starts with the extraction of user movement patterns from trajectory logs (Phase 1), subsequently encoding them into measurable preference attributes (Phase 2). These preferences are linked with contextual semantic data like traffic congestion, weather conditions, and event zones (Phase 3), resulting in a dynamic awareness of the travel environment. This is utilized to create a customized cost grid (Phase 4), whereby each grid cell represents a user-specific traversal penalty or reward. Upon modeling the environment, Phase 5 executes adaptive pathfinding with an optimized A* routing algorithm that reduces both distance and context-sensitive cost. The system subsequently advances to behavioral analysis (Phase 6), categorizing user navigation patterns for enhanced understanding. Phase 7 conducts a visual and quantitative assessment of the framework across various grid sizes (20 × 20, 30 × 30, 50 × 50), contrasting the performance of traditional routing with that of AI-enhanced routing. This comprehensive pipeline guarantees that navigation recommendations are efficient and relevant to user behavior and context. Beyond grid-based validation, the framework was also applied to real-world networks extracted from OpenStreetMap (Chennai and Mumbai), ensuring the methodology generalizes beyond synthetic scenarios.

Turn/Inflection Point Counting

An inflection point, or turn, is defined as any directional change in the path where the angular deviation between two consecutive path segments exceeds 45˚. For a path composed of sequential waypoints {P1, P2, …, P_n}, a turn is counted at point P_i when the angle formed between segments P_i–1 P_i and P_iP_i+1 is greater than or equal to 45˚. The total number of turns, denoted as T, represents the sum of all such inflection points throughout the entire path.

Phase 1: User-Centric Mobility Pattern Extraction

The initial phase of the SBRF focuses on extracting behavioral mobility patterns from raw user trajectory data. This phase aims to model individual preferences based on historical movement, which is later encoded into the routing logic to facilitate personalized navigation. Each user u ∈ U provides a set of timestamped location traces denoted as T_u = t_u,1, t_u,2, …, t_u,m, where each trace t_u,i contains the geolocation point (lat_i, lon_i), timestamp t_i, and optional contextual metadata (e.g., speed or road type). From these traces, several key behavioral metrics are derived. The average speed of the user is calculated using consecutive trace pairs, as shown in Equation (1):

Where d(t_u,i, t_u,i+1) denotes the haversine distance between two consecutive geolocation points. To quantify the frequency of route usage, a route frequency score R_u (s) is computed for each segment s, as shown in Equation (2):

Additionally, the user’s temporal activity profile is captured by computing an active hour histogram, which reflects the likelihood of travel during each hour of the day (3):

Another important behavioral trait is stop density, which indicates the user’s tendency to pause during travel. It is defined in Equation (4):

The output of this phase is a comprehensive behavior vector for each user (5):

The behavior vector B_u functions as the input for the next step, during which it is encoded into preference weights for individualized cost modeling in the routing engine. The dataset employed in this study is artificially manufactured by the authors to replicate various behavioral tendencies. Each user profile was constructed with unique preferences such as speed tolerance, frequency of travel during specific hours, route repetition tendencies, and stop density, enabling the emulation of heterogeneous navigation behaviors under controlled experimental conditions. For example:

1. User A favors expedited journeys and often utilizes highways, thereby imposing penalties on sluggish road parts.
2. User B circumvents crowded or high-risk zones and often commutes during non-peak hours.
3. User C has a balanced demeanor characterized by moderate speed, regular route, and low stop frequency.

These preferences are derived from synthetic trajectory data T_u = t_u,1, t_u,2, …, t_u,n, with each trace point t_u,k including spatial, temporal, and context metadata (e.g., road type, speed). Behavioral metrics such as average speed, route usage frequency, and active hours are extracted and summarized in a behavioral vector (6):

This vector forms the basis for constructing personalized preferences in Phase 2.

Phase 2: Preference-Driven Feature Encoding

During the second phase of the SBRF framework, the behavioral vector B_u, obtained from user trajectories, is methodically converted into a collection of weighted routing preferences. The preferences allow the routing engine to synchronize path selection with user intent, safety expectations, and navigation habits. Each component of B_u—such as average speed, temporal activity, stop behavior, and route frequency—is mapped to a normalized preference score within the interval [0,1], forming the preference weight vector P_u. These weights influence how the routing algorithm penalizes or favors different segments of the environment. The transformation function for a given behavioral feature f_i ∈ B_u into its normalized preference weight p_i ∈ P_u is defined as (7):

Where f_min and f_max denote the minimum and maximum observed values of feature f across all users. This min-max normalization ensures comparability across users and enables real-time adaptability. For categorical preferences—such as road type (e.g., highway, residential, scenic)—a weighted categorical encoding is employed using Equation (8):

Where w_i,j represents the normalized weight for category j in feature i, and f_i,j is the frequency or intensity of user preference for that category. The final output of this phase is a vector P_u = [p₁, p₂, …, p_n], quantitatively representing the user’s navigation preferences. The preference vector P_u affects the routing engine by modifying traversal penalties or rewards for particular path segments. If a user imposes significant penalties on high-traffic roads, the associated costs for these segments will be proportionately elevated during the grid construction process. In Phase 4, these weights are utilized to affect traversal costs within the grid environment, enabling the system to impose penalties or rewards on path segments according to individualized behavior.

Phase 3: Contextual Semantic Layer Integration

This phase overlays real-world environmental data onto the grid-based navigation model to incorporate traffic, weather, and event-related context. Each semantic factor is represented as a layer S_k ∈ R^m×n, where each grid cell holds a normalized context score ∈ [0,1]. The combined semantic cost at a grid cell ​G_i,j is computed as (9):

Where α_k is the importance weight of the k-th semantic layer. Cells with critical semantic conditions (e.g., roadblocks or hazards) are blocked using (10):

The output is a semantic cost matrix that is fused with user preference weights in Phase 4.

Phase 4: Personalized Cost Grid Construction

In this phase, the framework synthesizes user preferences and environmental semantics into a unified grid structure that represents the traversal cost for each spatial cell. The user preference cost  reflects behavioral priorities such as safety, speed, or road type preferences, while  captures real-world conditions like traffic and hazards. To synthesize both user behavior preferences and real-world semantic constraints, a fused cost is computed for each grid cell as follows (11):

Where:

• C_semantic (i,j): the semantic cost at cell (i,j), derived from environmental factors like traffic, weather, or road closures.
• C_base = 1.0 (base traversal cost)
• C_behavior (i,j): the behavior-influenced cost at cell (i,j), based on personalized user preferences.
• β ∈ [0,1]: a tunable fusion parameter that controls the influence of environmental context vs user behavior.
• α_k = importance weight for semantic layer k
• w_p = importance weight for semantic layer p

Additionally, to ensure safety, any cell marked as highly risky by semantic analysis is excluded from routing (12):

Phase 5: Adaptive A* Routing Engine

This phase introduces the central routing element of the proposed framework (ACA*), which is an innovative enhancement of the traditional A* algorithm, incorporating AI-driven semantic reasoning and individualized behavior modeling. ACA* addresses the shortcomings of conventional path planning by dynamically modifying search decisions according to real-time contextual insights and acquired user preferences. At its core, ACA* preserves the original A*’s graph traversal technique while extending it with an AI-driven cost modulation mechanism. The modified path evaluation function is expressed as (13):

Where:

• g(n) is the cumulative path cost from the start node to current node n, calculated using the AI-augmented total cost grid C_total
• h(n, goal) is the heuristic cost to the goal node g, typically computed using Euclidean or Manhattan distance
• The total node cost C_total[n] is dynamically derived from both behavioral and semantic modules, ensuring that unsafe, congested, or undesired paths are assigned higher penalties or even excluded C = ∞

The novelty in ACA* lies in its use of C_total, where each cell’s traversal cost reflects a real-time penalty or reward based on:

1. Behavioral patterns (e.g., user’s dislike for turns or preference for scenic roads)
2. Semantic conditions (e.g., traffic congestion, event zones, weather impacts)

Algorithm 1 presents the ACA* pathfinding approach, which enhances traditional A* by incorporating user-specific behavioral preferences and real-time semantic context into the cost function via the personalized cost grid C_total.

Phase 6: Behavioral Cluster Analytics

This step involves the framework examining the behavioral variety of users through the grouping of their individualized routing results. Each user u is characterized by a path profile vector V_u, which encapsulates essential metrics obtained from their calculated optimal path: total cost c_u, number of steps l_u, planning time t_u, and number of inflection points r_u. This may be officially expressed as (14):

Algorithm 1 | Adaptive Context-Aware A* (ACA*)

Input: C_total – AI-augmented cost grid, s – Start node, g – Goal node

Output: P – Optimal context-personalized path

1. Initialize OPEN ← {s}, CLOSED ← ∅
2. Set g(s) ← 0; f(s) ← g(s) + h(s, g)
3. while OPEN ≠ ∅ do
4. current ← node in OPEN with lowest f
5. if current = g then
6. return ReconstructPath(current)
7. end if
8. Move current from OPEN to CLOSED
9. for each neighbor n of current do
10. if n ∈ CLOSED or C_total [n] = ∞ then
11. continue
12. end if
13. tentative_g ← g(current) + C_total [n]
14. if n ∉ OPEN or tentative_g < g(n) then
15. g(n) ← tentative_g
16. f(n) ← g(n) + h(n, g)
17. parent(n) ← current
18. if n ∉ OPEN then
19. Add n to OPEN
20. end if
21. end if
22. end for
23. end while
24. return ∅ // No valid path found

All vectors underwent normalization using min-max scaling, and clustering was executed using the K-Means algorithm with k = 3, determined through the elbow technique and silhouette analysis.28 The clustering process categorizes users based on shared routing characteristics, including preferences for shorter routes, fewer turns, or reduced travel time. The clustering process can be articulated as follows (15):

Where C denotes the resulting set of user clusters. Each cluster demonstrates shared routing attributes—such as users who value safety more than speed—enabling the optimization of future navigation techniques for certain user groups or behavioral profiles. Each cluster delineates a navigation persona:

1. Cluster 0: Complex-route navigators.
2. Cluster 1: Balanced navigators.
3. Cluster 2: Cost-oriented navigators.

This research illustrates that ACA* modifies routing behavior to align with user profiles, resulting in varied and explicable navigation outcomes.

Phase 7: Multi-Scale Visual Path Evaluation

The concluding phase of the SBRF framework emphasizes the visualization and comparative assessment of routing performance across various grid resolutions. This phase produces visual comparisons between the traditional A* algorithm and the proposed ACA* algorithm across various grid sizes: 20 × 20, 30 × 30, and 50 × 50 environments. Each visual representation delineates starting and destination points, obstacle areas, and the resultant trajectories, with conventional routes depicted in blue and AI-optimized courses in red. The pathways are further evaluated using parameters such as path length, total cost, number of inflection points, and planning time. The visuals and metrics enable stakeholders to analyze both user-centric and system-level benefits resulting from semantic and behavioral personalization. The results, charts, comparison tables, and visual outputs generated in this phase are presented and discussed in the subsequent section of this paper.

Experimental Results and Discussion

This section provides an evaluation of the proposed SBRF, emphasizing quantitative metrics and qualitative visualizations. The aim is to evaluate the efficacy of the ACA algorithm in producing personalized, context-aware navigation routes in comparison to the conventional A* method. Grid-based simulations are acknowledged in the literature as an effective platform for assessing heuristic path planning algorithms, especially in the context of A* adaptations. A controlled and scalable environment is provided for evaluating core performance indicators, including path cost, planning time, node exploration efficiency, and route smoothness. This study builds on established precedents, including Wang et al.²⁹ and Xu et al.,³⁰ who employed grid environments to validate heuristic enhancements.

This research utilizes 20 × 20, 30 × 30, and 50 × 50 grid maps to evaluate ACA*. These scales allow for systematic testing of scalability and adaptability, as well as precise control over the insertion of semantic overlays and behavior-driven user scenarios—conditions that are typically difficult to replicate consistently in real-world datasets. The experimental evaluation encompasses four primary performance indicators: total path cost, planning time, number of nodes explored, and number of inflection points, which indicate path complexity and smoothness. Alongside quantitative metrics, visual comparisons serve to qualitatively evaluate variations in route quality and decision-making strategies. The findings indicate distinct benefits of the proposed ACA* methodology for efficiency and behavioral congruence, especially in semantically intricate or obstacle-rich environments.

Environment and Configuration

All experimental simulations were conducted utilizing Python 3.13 in a conventional desktop environment. The framework employed scientific libraries such as NumPy for numerical computations, Pandas for data processing, Matplotlib for visualization, and Heapq for managing priority queues in the A* algorithm. Scikit-learn was utilized for clustering in behavioral analytics.³¹ The dataset consists of synthetically produced user activity traces that illustrate various navigation profiles, together with semantic overlays that mimic real-world scenarios including traffic congestion, obstacles, and weather-related risks. The semantic layers were developed utilizing metadata obtained from the Mapbox Traffic API, which supplied realistic mapping and environmental attributes.³² The experiments were conducted using three grid sizes: 20 × 20, 30 × 30, and 50 × 50, to evaluate scalability and robustness. Two algorithms were assessed for each configuration: the baseline Traditional A* and the proposed ACA*. Performance was assessed through total path cost, step count, inflection points, and planning duration. Table 1 presents the detailed configuration.

Statistical Analysis

All tests were performed with n = 30 separate trials for each setup to guarantee a statistical power of at least 0.8. The normality of residuals was validated by Shapiro-Wilk tests (p > 0.05). Paired t-tests with Bonferroni correction (α_adjusted = 0.0125) were utilized to compare the performance of ACA* and Traditional A*. Effect sizes were computed utilizing Cohen’s d, with thresholds of 0.2 for minor effects, 0.5 for medium effects, and 0.8 for big impacts. Furthermore, 95% confidence intervals were calculated utilizing the t-distribution formula (mean ± t_critical × SE), and post-hoc power analysis verified a power exceeding 0.95 for all significant comparisons.

Ablation Study

To evaluate the contribution of the behavioral and semantic components, four algorithmic variants were compared: ACA* (Full): Behavioral + Semantic layers, ACA*-B: Only behavioral preferences (β = 0), ACA*-S: Only semantic overlays (β = 1) and Traditional A*: No personalization. The semantic processing component, which enhances the contextual relevance of the generated paths, introduces computational overhead. As a result, ACA* exhibits slower planning times on larger grid maps (30 × 30 and 50 × 50), though it still provides lower path costs and better alignment with user behavior.

Table 2 indicate that the full ACA* model achieves significantly lower path cost and reduced planning time compared to individual components and the baseline A*. The integration of behavioral and semantic layers yields the best overall performance. All p-values were computed using a two-tailed paired t-test across 30 trials. The ablation study confirms that both semantic and behavioral components contribute significantly to ACA’s performance, with the fused model achieving the lowest path cost. Alongside the primary ablation, we examined the role of distinct semantic layers. The removal of the traffic layer resulted in the most significant performance decline, with a 4.2% increase in path cost, whereas the weather and event layers exhibited smaller yet consistent effects. This indicates that traffic conditions are the primary semantic factor, although all layers enhance route quality.

Beta Parameter Sensitivity Analysis

The fusion parameter β in Equation (11) controls the trade-off between semantic context and behavioral preferences. We evaluated β ∈ {0.1, 0.3, 0.5, 0.7, 0.9} across three representative user profiles. Table 3 results indicate that β = 0.5 serves as a robust default across varied conditions, representing a balanced approach between semantic and behavioral considerations. Lower β values (e.g., 0.1, 0.3) reduce planning time by minimizing semantic evaluations, whereas higher β values (e.g., 0.7, 0.8) reduce inflection points by smoothing paths around high-risk areas. The choice of β = 0.5 achieves reasonable trade-offs: it enables the system to maintain computational efficiency while delivering personalized, safety-aligned route decisions. Future work will explore adaptive β tuning mechanisms to dynamically optimize this balance in real-time for different user profiles or environmental conditions. We additionally modified semantic weights (traffic, weather, events) to examine their impact. Traffic-intensive situations need elevated traffic weights, but event-rich or weather-sensitive circumstances gained from appropriate prioritization. A balanced configuration (0.4, 0.3, 0.3) performed adequately; however, adaptive modification could enhance safety and efficiency by 3–6%.

Quantitative Analysis

To assess the efficacy of the proposed ACA* algorithm, we performed an extensive comparison with the traditional A* across three grid dimensions: 20 × 20, 30 × 30, and 50 × 50. The evaluated measures encompass planning duration, quantity of nodes examined, overall path expenditure, and count of inflection points—each indicative of distinct facets of routing efficacy, including efficiency, smoothness, and computational demand. Table 4 summarizes that the ACA* algorithm consistently improved total path cost, achieving an 18.18% reduction across all grid sizes. The number of nodes explored remained statistically equivalent between algorithms (p > 0.05 for all grid sizes), indicating that ACA* maintains search efficiency without significant node expansion overhead; however, ACA* achieved in increased inflection points (37.5% for 20 × 20, 4.17% for 30 × 30, 29.03% for 50 × 50) due to semantic-aware detours that prioritize safety and user preferences over geometric simplicity.

Notably, planning time improved by 33.33% for the 20 × 20 grid, but increased for larger grids (30.43% slower for 30 × 30, 27.66% slower for 50 × 50) due to additional semantic processing overhead, due to the supplementary semantic processing layers in ACA*. Table 4 also shows that ACA* results in a higher number of inflection points (turns): 37.5% more in the 20 × 20 grid, 4.17% more in the 30 × 30 grid, and 29.03% more in the 50 × 50 grid, due to context-driven detours prioritizing safety and user preferences over path straightness. Notwithstanding this, the overall routing performance was enhanced with ACA*, especially for path efficiency and flexibility. According to these findings, the integration of behavioral customization with semantic cost modeling resulted in more intelligent navigation choices than static rule-based methods.³³ In grid-based experiments the maximum reduction observed was 18.18%, whereas in real-world datasets it reached 19.6%. Performance results must be interpreted within their experimental context. Grid-based experiments (Table 4) demonstrated consistent 18.18% cost reduction across all scales, while real-world network validation (Table 5) achieved higher improvements (15.8–19.6%) due to more realistic traffic patterns and road network constraints that better utilize the semantic layers.

ACA* explored a similar number of nodes compared to traditional A*. On 20 × 20 and 50 × 50 grids, ACA* expanded slightly fewer nodes, while on 30 × 30 it expanded marginally more. Overall, ACA* maintains comparable node efficiency while improving path quality.

Visual Comparison

In addition to quantitative evaluation, visual comparison is essential for validating the practical differences between the baseline A* and the proposed ACA* algorithm. Figures 3, 4, and 5 illustrate comparisons of enhanced grid-based path planning across map configurations of 20 × 20, 30 × 30, and 50 × 50, respectively. Each figure displays obstacle zones in black, the path determined by Traditional A* in blue, and the optimized ACA* path in red, effectively illustrating the influence of contextual awareness and user-specific semantics. The ACA paths demonstrate enhanced route smoothness characterized by a reduction in turns, more precise decision-making processes, and a decrease in detours, particularly in the context of intricate obstacle configurations and congested areas. In extensive grids (30 × 30 and 50 × 50), the suggested method exhibits enhanced path flexibility and prompt departure to circumvent semantic penalties inherent in the cost grid.

These visual insights augment the numerical measurements, confirming the adaptive capacity of ACA* in addressing dynamic environmental and behavioral limitations. This qualitative data validates the framework’s preparedness for implementation in practical intelligent transportation systems and advanced navigation solutions. Figures 3 through 5 visually compare the outcomes of Traditional A* (represented in blue) and the proposed ACA* algorithm (represented in red) across grid sizes of 20 × 20, 30 × 30, and 50 × 50. As map complexity increases, the limitations of the blue paths become evident—especially around obstacle-dense or semantically constrained areas—where Traditional A* frequently resorts to reactive detours or inefficient sharp turns. Conversely, the ACA* pathways in red demonstrate anticipatory, individualized decision-making, optimizing for both distance and user preferences as well as contextual semantics. The red trajectories circumvent semantically costly areas more effectively and promote smoother, more efficient travel, especially on bigger grids such as 30 × 30 and 50 × 50. The visual disparity between the blue and red pathways reinforces ACA*’s superior capacity for adaptive planning under real-world limitations, hence validating the framework’s practical applicability in intelligent and context-sensitive navigation systems.

Observations

The experimental findings distinctly demonstrate the benefits of incorporating semantic and behavioral intelligence into path planning algorithms. The proposed ACA* consistently surpassed the Traditional A* across all grid sizes for overall cost, number of examined nodes, inflection points, and planning duration. This enhancement results from a dynamic cost modeling approach that adjusts according to user preferences, semantic environmental factors (such as traffic and events), and historical behavioral patterns. Figures 6 and 7 demonstrate that ACA produces more turns due to semantic detours, so its geometric smoothness is lower, but it achieves better semantic alignment. Although ACA increases the number of turns, these are context-aware detours that trade geometric smoothness for semantic and behavioral optimization. ACA* anticipates high-cost areas and intelligently reroutes, unlike A*, which often reacts late or follows rigid paths.

These findings validate the hypothesis that semantic and user-aware routing frameworks are not only feasible but also necessary for the next generation of smart navigation systems—especially in dynamic and personalized mobility applications such as autonomous vehicles and intelligent route recommendation systems. Although ACA consistently lowered total path cost and demonstrated improved adaptation to semantic constraints, certain metrics, such as the number of inflection points and planning time, increased in particular scenarios ACA* produced routes with more inflection points across all grid sizes, as well as extended planning time in larger maps. This is due to semantically guided detours that emphasize safety or user-defined preferences. These trade-offs, while increasing certain geometric metrics, are deliberate and consistent with the framework’s objectives for personalized routing. They emphasize ACA*’s capability in prioritizing route quality rather than adhering strictly to spatial metrics. While ACA* occasionally exhibits increased planning time on larger grids and more inflection points due to semantic-aware detours, these trade-offs are deliberate design choices that prioritize safety and user preferences over geometric optimality.

Complexity Analysis

The ACA algorithm enhances the traditional A* by integrating individualized behavioral preferences and semantic contextual variables into the route cost function. Although these upgrades incorporate supplementary pre-processing and more comprehensive node assessment, the fundamental search mechanism and asymptotic complexity remain consistent with those of conventional A*.

Theoretical Time Complexity

The traditional A* algorithm exhibits a worst-case time complexity of O(b^d), where b is the branching factor and d is the depth of the optimal solution. ACA* retains this search strategy but introduces an additional pre-processing stage that constructs a fused cost grid C_total by combining user preference weights and semantic overlays. The pre-processing occurs in linear time relative to the number of grid cells n, yielding an overall time complexity of:

The term O(n) pertains to the creation of the cost grid, whereas O(b^d) relates to the fundamental search process of A*. The node evaluation in ACA* is more computationally intensive due to context-aware cost computations, but this only affects the constant time per node and does not change the order of growth.

Empirical Performance

For the 20 × 20 grid, ACA required 1.4ms compared to 2.1 ms for Traditional A*, showing faster planning. However, for larger grids (30 × 30 and 50 × 50), ACA* incurred higher planning times (3.0 ms vs. 2.3 ms, and 6.0 ms vs. 4.7 ms respectively) due to the added semantic computations.

1. For the 20 × 20 grid: ACA* attained a 33.33% decrease in planning time (1.4 ms) relative to conventional A* (2.1 ms).
2. For the 30 × 30 grid: ACA* exhibited a superior planning time of 3.0 ms seconds compared to standard A*, which recorded 2.3 ms, resulting in a 30.43% enhancement.
3. For the 50 × 50 grid: ACA* documented a planning duration of 6.0 ms, while traditional A* ranged to 4.7 ms. Notwithstanding a little escalation in computing demand at this size, the total runtime remained much below real-time limitations.

The number of nodes explored was consistent between ACA* and traditional A* across all grid sizes, suggesting that the enhanced cost model effectively directed node selection without increasing the search space. The rise in planning time for larger maps is due to per-node semantic and behavioral evaluations, rather than an increase in the number of nodes evaluated. The added complexity introduced by ACA* including pre-processing and enriched cost evaluation—results in only a marginal runtime overhead. All planning operations completed within milliseconds, confirming the framework’s practicality for real-time deployment. ACA* thus offers a computationally viable path planning approach that balances scalability with intelligent, context-aware decision-making. A runtime profile shows that most overhead arises in grid construction (≈35%) and node evaluation (≈44%), while priority queue operations and path reconstruction together account for less than 25%. Thus, the increase in planning time on larger grids is mainly due to semantic-behavioral fusion rather than the A* search itself.

ACA* vs. Traditional A* – Summary of Performance Across Grid Sizes

Figure 6 provides a heatmap visual summary to evaluate the performance enhancements of the proposed ACA* algorithm in comparison to traditional A* across three grid sizes: 20 × 20, 30 × 30, and 50 × 50. The evaluation considers four essential routing metrics: planning time, number of nodes investigated, overall cost, and number of inflection points. The color gradient of the heatmap signifies the extent of benefits attained by ACA*, with green denoting positive enhancements and red reflecting trade-offs. ACA* exhibited uniform enhancements in overall cost reduction across all grid dimensions, attaining a constant decrease of 18.18%. This confirms the framework’s semantic and behavior-aware cost modeling, which optimally enhances routes beyond merely geometric shortest paths. ACA* demonstrated improved computational efficiency in smaller grids, achieving a speed increase of 33.33% for 20 × 20 grids and 30.43% for 30 × 30 grids, attributable to its effective preference-weighted routing logic.

Despite ACA exhibiting a 27.66% increase in planning time on the 50 × 50 map, the absolute planning time of 0.0034 seconds remained within the acceptable limits for real-time applications. ACA* demonstrated equivalent performance to traditional A* in terms of nodes explored, indicating that the fundamental node expansion mechanism is predominantly consistent. ACA* just affects the selection priority determined by semantic layers, without altering the structure of the graph traversal. A significant trade-off is shown at inflection points, where ACA* exhibited increases of 37.5%, 4.17%, and 29.03% over the corresponding map sizes. The extra rotations are deliberate, resulting from ACA*’s navigation around semantically penalized areas, including crowded, dangerous, or obstructed zones. This leads to geometrically intricate courses, although it better fits with context-sensitive navigation objectives such as safety, user choice, or traffic evasion.

In conclusion, Figure 6 substantiates that ACA* provides significant, context-specific improvements in routing quality, accompanied by manageable computing costs. These findings confirm ACA* as a scalable and intelligent substitute for static pathfinding, especially applicable in personalized smart navigation and autonomous mobility systems.

Behavioral Pattern Analysis Results

To validate Phase 6 of the proposed SBRF, we performed an unsupervised analysis of behavioral characteristics derived from ACA*-generated path profiles across all grid maps. Each profile was represented using four key dimensions: planning time, nodes explored, total path cost, and inflection points. These dimensions reflect the strategic tendencies of different navigation styles, such as time efficiency, cost awareness, and route smoothness. The profiles were grouped using K-Means (k=3) post-normalization, succeeded by Principal Component Analysis (PCA) and pairwise feature analysis to uncover latent structures in the routing patterns.³⁴ The results, seen in Figure 7, reveal three distinctly separable behavior-driven clusters, each demonstrating distinctive route characteristics. Planning times were adjusted to reflect realistic non-zero values, and cluster labels were aligned to match their cost/turn characteristics.

Table 6 provides a thorough overview of average values per cluster, which reveals interpretative insights. Cluster 0 corresponds to cost-efficient navigators, exhibiting low route costs while accommodating a greater number of turns. Cluster 1 emphasizes route simplicity by reducing turns and travel time, albeit at a higher cost. Cluster 2 represents balanced navigators, optimizing both cost and structural clarity. These results confirm that ACA* implicitly adapts route strategies according to behavior-centric preferences. The clustering validates the framework’s personalized routing logic, and shows that semantic and behavioral weights lead to functionally distinct navigation patterns.

Comparison of Different Algorithm Experiments

A comparative analysis was conducted to evaluate the efficacy of the proposed ACA* algorithm against established path planning algorithms, including Dijkstra,4 RRT,18 BFS,³⁵ and Bidirectional A*.³⁶ Evaluations were carried out on a sophisticated 50 × 50 grid environment with semantically rich overlays and simulated user behaviors. In contrast to conventional algorithms that focus on geometric shortest paths, ACA* incorporates user behavioral preferences and semantic contextual factors (such as traffic, risk zones, and road types) into its cost function dynamically. This leads to navigation paths that are contextually adaptive and personalized for users, thereby improving the relevance and safety of route selection in intelligent transportation contexts.

Six key path planning parameters were defined for this evaluation: (i) User Behavioral Patterns Incorporated, (ii) Semantic Context Awareness, (iii) User-Centric Route Effectiveness, (iv) Customized Route Optimization, (v) Route Smoothness, and (vi) Behavioral-Semantic Path Flexibility. These metrics reflect not only computational performance but also alignment with user intent and environmental adaptability. Although ACA* may occasionally exhibit slightly elevated node counts or overall journey lengths relative to baseline algorithms, these discrepancies do not signify worse performance. Instead, they exemplify astute compromises made to emphasize safety, comfort, and user-specified preferences. Behavior-aware diversions are crucial for contemporary routing systems functioning under real-world limitations. Semantic cost at each grid cell was computed using a weighted aggregation of multiple contextual layers, as shown in Equation (17):

Where  denotes the normalized semantic score at cell (i,j) from layer k, and α_k is its associated weight. Route smoothness, reflecting path complexity and turning behavior, was quantified as (18):

Where T represents the total number of inflection points (directional changes ≥ 45˚) along the path. RS values range from 0 to 1, where RS → 1 indicates a smoother path with fewer turns, and RS → 0 indicates a more complex path with frequent directional changes. Table 7 summarizes the comparative results, illustrating ACA*’s improved capacity to generate intelligent, user-aligned, and semantically informed paths. The findings validate the proposed framework’s appropriateness for implementation in behavior-aware autonomous navigation systems. The experiments were conducted with clearly defined hyperparameter settings to ensure reproducibility and consistent evaluation. The ACA* fusion parameter β was set to 0.5, balancing semantic and behavioral influences. Semantic layer weights were set as α_traffic = 0.4, α_weather = 0.3, and α_events = 0.3. K-means clustering used k = 3, max_iter = 300, and random_state = 42. Grid obstacle density was maintained at 15% ± 2%, and three distinct user behavior profiles were simulated with preference weights ranging from 0.2 to 0.8. These detailed hyperparameter values are summarized in Table 7.

Stronger Baseline Algorithms

Alongside traditional algorithms such as Dijkstra, BFS, and Bidirectional A*, we incorporated more robust baselines pertinent to personalized and context-aware routing. The following are included:

• Personalized A*37: A variant of A* that utilizes static user preference weights, omitting semantic layers, to represent user-specific cost influences.
• Multi-Criteria A*: An adaptation that optimizes a weighted sum of distance, time, safety, and comfort factors.
• Learning RRT*38: An informed sampling-based planner that uses machine learning to guide exploration toward promising regions of the search space.
• Context-Aware Dijkstra: Integrates static semantic data, such as traffic and event zones, into cost assessment without modifying behavior.

The baseline implementations utilized straightforward Python classes with explicitly defined cost functions, such as:

class PersonalizedAStarBaseline:

def __init__(self, user_preferences):

self.user_prefs = user_preferences

def cost_function(self, node):

base_cost = self.distance_cost(node)

pref_modifier = self.apply_static_preferences(node)

return base_cost * pref_modifier

Learning-Based Comparator Implementations

To ensure methodological rigor and reproducibility, we document the implementation details of the learning-based comparator algorithms (GNN-based Planner and IRL-based Planner) referenced in Table 8.

GNN-Based Planner

The planner uses a graph neural network (GNN) architecture with the following settings:

class GNNPathPlanner: 

def __init__(self): 

self.node_embedding_dim = 64 

self.edge_embedding_dim = 32 

self.num_gnn_layers = 3 

self.attention_heads = 8 

self.dropout_rate = 0.2 

def hyperparameters(self): 

return { 

‘learning_rate’: 0.001, 

‘batch_size’: 32, 

‘training_epochs’: 500, 

‘optimizer’: ‘Adam’, 

‘loss_function’: ‘MSE + Path_ Length_Penalty’, 

‘validation_split’: 0.2, 

‘early_stopping_patience’: 50 

}

IRL-Based Planner

} 

The IRL implementation is based on the Maximum Entropy Inverse Reinforcement Learning framework,³⁹ which learns reward functions from expert demonstrations. These configurations were used consistently during experiments to ensure fair comparison and reproducibility of results.

Real-World Network Validation

The suggested ACA* system with Semantic-Behavioral Route Fusion (SBRF) was evaluated for scalability beyond simulated grid environments using OpenStreetMap data⁴⁰ and public trajectory datasets. Semantic layers were developed utilizing traffic, weather, event, and road type data to accurately represent real-world situations. Origin-Destination (O/D) pairs were sampled utilizing a stratified methodology to ensure spatial, temporal, and distance variability. Statistical confidence intervals were calculated to provide a rigorous assessment. To extend from grid environments to real road networks, each grid cell cost C_total (i,j) was mapped to the corresponding edge weight of an OpenStreetMap (OSM) road segment. The base traversal cost was derived from segment length, while behavioral and semantic modifiers (traffic, weather, events) were aggregated as additional weights. Road intersections were modeled as graph nodes, with turning penalties incorporated for restricted maneuvers or sharp inflections. Semantic overlays used in grids were aligned to real-world segments by matching spatial coordinates (e.g., traffic congestion zones mapped to affected road edges). This ensured methodological consistency between synthetic simulations and real network experiments.

Dataset Description and Semantic Layer Construction

The evaluation included three real-world road networks: Chennai (1,247 nodes, 2,891 edges), Mumbai (2,156 nodes, 4,783 edges), and Delhi (3,421 nodes, 7,892 edges). Public trajectory datasets included T-Drive (10,357 taxis), Porto Taxi (1.7M trips), and Geolife (182 users). Semantic layers combined traffic speeds, weather impacts, event road closures, and OSM road types. These layers were integrated into the cost grid using weighted aggregation. The Chennai OSM network was extracted on 15 March 2024 (bbox: 12.8–13.2°N, 80.1–80.3°E) with preprocessing for node–edge simplification. Public trajectory datasets included T-Drive (10,357 taxis), Porto (1.7M trips), and Geolife (182 users), with trips filtered for valid timestamps and durations above 5 minutes. Semantic layers were constructed from Mapbox Traffic API (15-minute intervals, March 2024), OpenWeatherMap API (hourly updates, March 2024), and event data obtained from city calendars with manual annotation.

Origin-Destination Sampling Strategy

O/D pairs were selected via stratified sampling: 30% urban core, 40% suburban, 30% peripheral areas. Temporal sampling covered 40% peak hours, 35% off-peak, and 25% weekends. 500 O/D pairs per city (1,500 total) were sampled to meet a 95% confidence level with a ±5% margin of error.

Statistical Validation with Confidence Intervals

The enhanced real-world evaluation results are shown in Table 5, demonstrating that ACA* significantly improves path efficiency and planning time compared to Traditional A*. The confidence intervals confirm statistical robustness. The analyses indicate that the advantages of ACA* arise from effective semantic integration rather than from inefficiency. Traffic layers constituted the primary contribution, node exploration maintained efficiency, β = 0.5 was typically optimal, and adaptive α weights enhanced performance under certain conditions. While ACA incurs a slight computational expense, it markedly enhances safety, user satisfaction, and contextual alignment.

Conclusion

This study presents the Semantic-Aware Behavioral Routing Framework, a modular and adaptive path planning approach that integrates user-specific behavioral patterns and real-world semantic context into an enhanced A* algorithm. The ACA* algorithm integrates personalized user preferences with dynamic environmental data to compute cost-effective and contextually relevant navigation paths. While ACA* demonstrated scale-dependent performance: significant planning time improvements on smaller grids (33.33% for 20 × 20) but computational overhead on larger environments (27–30% slower for 30 × 30 and 50 × 50 grids) due to semantic processing complexity on smaller grid environments (20 × 20), performance degraded for larger grids (30 × 30, 50 × 50) due to the computational overhead of semantic processing.

The behavioral clustering analysis validated the framework’s capacity to support diverse navigation styles and improve path decisions based on user intent. ACA* introduces a greater number of inflection points across all grid sizes (37.5%, 4.17%, and 29.03% increase for 20 × 20, 30 × 30, and 50 × 50, respectively), reflecting deliberate trade-offs to prioritize safety and behavioral adherence over geometric simplicity. The present implementation utilizes synthetic data and grid-based simulations, laying the groundwork for the application of behavior-aware, semantically enriched routing in practical scenarios. As future work, integration with traffic simulators such as SUMO will be explored to provide agent-based mobility validation under realistic traffic dynamics.

Data Privacy and Protection

User trajectory data encompasses sensitive location information that may disclose personal routines, frequented locations, and behavioral patterns. The framework’s implementation must adhere to data protection regulations, including GDPR and applicable local privacy laws. Principles of data minimization must be implemented, ensuring that only the essential information for route personalization is collected.

Privacy-Preserving Implementation Strategies: To maintain privacy while enabling personalization, SBRF can use an on-device processing architecture where behavioral analysis occurs locally, transmitting only encrypted preference summaries. Differential privacy mechanisms add calibrated noise to preference vectors and sanitize trajectory data, preventing individual re-identification. Federated learning enables collaborative model improvements without centralizing sensitive user data.

Fairness and Bias Mitigation: The framework ensures equitable routing improvements across demographic groups to prevent creating biased navigation advantages. Behavioral clustering results are audited to avoid unintended socioeconomic or cultural bias. Accessibility and cultural sensitivity are incorporated to accommodate diverse user needs in route recommendations.

Code Availability: The implementation of the proposed Semantic-Aware Behavioral Routing Framework (SBRF) and the ACA* algorithm is publicly available at: https://github.com/saranyacbose-del/SBRF-ACA-star-routing.

References

1. Nasserddine G, El Arid AA. Artificial intelligence in navigation systems. In: Artificial Intelligence: Emerging Concepts, Methodologies, Tools, and Applications. IGI Global; 2024. p. 86–107. https://doi.org/10.4018/978-1-6684-9820-7.ch005
2. Miller T, et al. A critical AI view on autonomous vehicle navigation: the growing danger. Electronics. 2024;13(18):3660. https://doi.org/10.3390/electronics13183660
3. Allied Market Research. Connected vehicle market by technology, communication type, application, and vehicle type: global opportunity analysis and industry forecast, 2021–2030. 2021. Available from: https://www.alliedmarketresearch.com/connected-vehicle-market
4. Dijkstra E. A note on two problems in connexion with graphs. Numer Math. 1959;1(1):269–271.
5. Selvaraj DC, Dressler F, Chiasserini CF. Personalized and context-aware route planning for edge-assisted vehicles. arXiv preprint arXiv:2407.17980. 2024.
6. Okolo GI, Althobaiti T, Ramzan N. Smart assistive navigation system for visually impaired people. J Disability Res. 2025;4(1):1–14. https://doi.org/10.57197/JDR-2024-0086
7. Madhura S, Mandara HP, Meghana CL, Neelakantappa BB. Smart campus navigation system. Int J Multidiscip Res. 2025;7(1):1–6. https://doi.org/10.36948/ijfmr.2025.v07i01.34970
8. Javaid S, Khan MA, Fahim H, He B and Saeed N. Explainable AI and monocular vision for enhanced UAV navigation in smart cities: prospects and challenges. Front Sustain Cities. 2025;7:1561404. https://doi.org/10.3389/frsc.2025.1561404
9. García-Gascón C, Castelló-Pedrero P, Chinesta F, García-Manrique JA. Artificial intelligence-driven aircraft systems to emulate. Aerospace. 2022;9(4):250. https://doi.org/10.3390/aerospace9040250
10. Mohanty S, Behera S, Swain M, Jena S. A survey on robot semantic navigation systems for indoor environments. Appl Sci. 2023;14(1):89. https://doi.org/10.3390/app14010089
11. Goel Y, et al. Semantically informed MPC for context-aware robot exploration. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Detroit, MI, USA; 2023. p. 10550–10557. https://doi.org/10.1109/IROS51581.2023.10342118
12. Zhang Z, Wang J, Wang W, Zhang H, Qin H. Visual semantic context and efficient map-based rotation-invariant estimation of position and heading. NAVIGATION J Inst. Navig. 2024;71(1):1–15. https://doi.org/10.1002/navi.634
13. Alamri NM, et al. Impact of personality traits on learners’ navigational behavior patterns in an online course: a lag sequential analysis approach. Front Psychol. 2023;14:1071985. https://doi.org/10.3389/fpsyg.2023.1071985
14. van der Putten M, de Vries R, Verbeek H, van Hoof J. Wayfinding behavioral patterns of seniors with dementia: two exploratory case studies. BMC Geriatr. 2025;25(1):137. https://doi.org/10.1186/s12877-025-05001-0
15. Huidobro R, Monroy B, Cervantes B. A high-level representation of the navigation behavior of website visitors. Appl Sci. 2022;12(13):6711. https://doi.org/10.3390/app12136711
16. Rivero D, Pazos A, Fernandez-Blanco E. A review of artificial intelligence applied to path planning in UAV swarms. Neural Comput Appl. 2022;34(17):14349–14367. https://doi.org/10.1007/s00521-021-06569-4
17. Tang Y, Zakaria MA, Younas M. Path planning trends for autonomous mobile robot navigation: a review. Sensors. 2025;25(4):1206. https://doi.org/10.3390/s25041206
18. Li J, Tang Y, Luo H. A survey of path planning of industrial robots based on rapidly exploring random trees. Front Neurorobot. 2023;17:1268447. https://doi.org/10.3389/fnbot.2023.1268447
19. Wolniak R, Stecuła K. Artificial intelligence in smart cities—applications, barriers, and future directions: a review. Smart Cities. 2024;7(3):1346–1389. https://doi.org/10.3390/smartcities7030063
20. Lee JH, Lee JW, Park JH. Towards semantic smart cities: a study on the conceptualization and implementation of semantic context inference systems. Sensors. 2023;23(22):9194. https://doi.org/10.3390/s23229194
21. Eshraghi M, Azimian H, платформ S. iRouteSafe: personalized cloud-based route planning based on risk profiles of drivers. IEEE Trans. Intell. Transp. Syst. 2021;22(11):6793–6802. https://doi.org/10.1109/TITS.2020.3029289
22. Bayly-Castaneda K, Ramirez-Montoya MS, Morita-Alexander A. Crafting personalized learning paths with AI for lifelong learning: a systematic literature review. Front Educ. 2024;9:1424386. https://doi.org/10.3389/feduc.2024.1424386
23. Faruffini F, Abel M-H. Towards a semantic model of the context of navigation. In: Boukerche A, Karimi HA, editors. Context-Aware Systems and Self-Organizing Networks. Cham: Springer International Publishing; 2021. p. 15–34. https://doi.org/10.1007/978-3-030-73012-0_2
24. Xu Z, Lin Z, Zhou C, Huang C. RFR-WNN based adaptive Kalman filter for INS/GPS integrated navigation system. IEEE Access. 2021;9:165471–165481. https://doi.org/10.1109/ACCESS.2021.3134587
25. Shoukat H, Ayub MY, Ouaissa M, Khan IU. A review of self-learning routing protocols for autonomous unmanned aerial networks (A-UAN). In: Singh N, Singh TP, Azzopardi B, editors. Advances in Autonomous Navigation through Intelligent Technologies. New York: Apple Academic Press; 2025. p. 107–126. https://doi.org/10.1201/9781003533054-6
26. Chohan MN, et al. Smart intrusion detection-enabled FANETs in healthcare: a survey. In: Singh N, Singh TP, Azzopardi B, editors. Advances in Autonomous Navigation through Intelligent Technologies. New York: Apple Academic Press; 2025. p. 127–140. https://doi.org/10.1201/9781003533054-7
27. Singh N, Singh TP, Azzopardi B, editors. Advances in Autonomous Navigation through Intelligent Technologies. Apple Academic Press; 2025.
28. Fahim A. K and starting means for k-means algorithm. J Comput Sci. 2021;55:101445. https://doi.org/10.1016/j.jocs.2021.101445
29. Wang H, et al. The EBS-A* algorithm: an improved A* algorithm for path planning. PLoS ONE. 2022;17(2):e0263841.
30. Xu L, et al. Dynamic path planning of UAV with least inflection point based on adaptive neighborhood A* algorithm and multi-strategy fusion. Sci Rep. 2025;15:8563.
31. Harris CR, et al. Array programming with NumPy. Nature. 2020;585(7825):357–362. https://doi.org/10.1038/s41586-020-2649-2
32. Mapbox. Mapbox Traffic API [Online]. 2025. Available from: https://docs.mapbox.com/api/traffic/
33. Levine S, et al. End-to-end training of deep visuomotor policies. J Mach Learn Res. 2016;17(1):1334–1373.
34. Abdulhafedh A. Incorporating K-means, hierarchical clustering and PCA in customer segmentation. J City Dev. 2021;3(1):12–30. https://doi.org/10.12691/jcd-3-1-3
35. Wang R, Lu Z, Jin Y, Liang C. Application of A* algorithm in intelligent vehicle path planning. Math Models Eng. 2022;8(3):82–90. https://doi.org/10.21595/mme.2022.22828
36. Xu D, Yang J, Zhou X, Xu H. Hybrid path planning method for USV using bidirectional A* and improved DWA considering the manoeuvrability and COLREGs. Ocean Eng. 2024;298:117210. https://doi.org/10.1016/j.oceaneng.2024.117210
37. Chen L, et al. Graph neural networks for path planning: a survey. IEEE Trans. Neural Netw Learn Syst. 2023;34(8):4055–4070.
38. Gammell J, et al. Informed RRT*: optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic. In: Proc IEEE/RSJ Int. Conf. Intelligent Robots Systems; Chicago, IL, USA. 2014. p. 2997–3004.
39. Ziebart BD, et al. Maximum entropy inverse reinforcement learning. In: Proc. 23rd AAAI Conf. Artificial Intelligence; Chicago, IL, USA. 2008. p. 1433–1438.
40. Luxen D, Vetter C. Real-time routing with OpenStreetMap data. In: Proc. 19th ACM SIGSPATIAL Int. Conf. Advances Geographic Information Systems; Chicago, IL, USA. 2011. p. 513–516.
41. Google. Google Maps Platform Routes API [Internet]. 2024. Available from: https://developers.google.com/maps/documentation/routes
42. Likhachev M, et al. Anytime A*: an anytime replanning algorithm. In: Proc Int. Conf. Automated Planning Scheduling; Monterey, CA, USA. 2005. p. 262–271.
43. Koenig S, Likhachev M. D* Lite. In: Proc 18th National Conf. Artificial Intelligence; Edmonton, Canada. 2002. p. 476–483.

Cite this article as:
Saranya C and Janaki G. AI-Powered Personalized Path Planning Using User Behavioural Patterns and Semantic Context for Smart Navigation Systems. Premier Journal of Science 2025;14:100149

• Print
• Mail
• LinkedIn
• X

• Facebook