AI-Driven Urban Mobility Solutions: Shaping Bucharest as a Smart City

Abstract

The metropolitan agglomeration in and around Bucharest, Romania’s capital and largest city, has experienced significant growth in recent decades, both economically and demographically. With over two million residents in its metropolitan area, Bucharest faces urban mobility challenges characterized by congested roads, overcrowded public transport routes, limited parking, and air pollution. This study evaluates the potential of AI-driven adaptive traffic signal control to address these challenges using an agent-based simulation approach. The authors focus on Bucharest’s north-western part, a critical congestion area. A detailed road network was derived from OpenStreetMap and calibrated with empirical traffic data from TomTom Junction Analytics and Route Monitoring (corridor-level speeds and junction-level turn ratios). Using the MATSim framework, the authors implemented and compared fixed-time and adaptive signal control scenarios. The adaptive approach uses a decentralized, demand-responsive algorithm to minimize delays and queue spillback in real time. Simulation results indicate that adaptive signal control significantly improves network-wide average speeds, reduces congestion peaks, and flattens the number of en-route agents throughout the day, compared to fixed-time plans. While simplifications remain in the model, such as generalized signal timings and the exclusion of pedestrian movements, these findings suggest that deploying adaptive traffic management systems could deliver substantial operational benefits in Bucharest’s urban context. This work demonstrates a scalable methodology combining open geospatial data, commercial traffic analytics, and agent-based simulation to rigorously evaluate AI-based traffic management strategies, offering evidence-based guidance for urban mobility planning and policy decisions.

1. Introduction

Urban populations continue to expand rapidly, resulting in increasing mobility demands and intensified pressure on transportation infrastructure in metropolitan areas. The challenges associated with urban mobility become more pronounced, along with the cities’ growth, significantly impacting traffic congestion, air pollution, citizens’ quality of life, and economic productivity. Therefore, there is an urgent need for efficient and sustainable urban mobility to ensure the vitality and resilience of metropolitan regions.

Artificial Intelligence (AI) technologies have become more accessible, and they have substantial potential to revolutionize urban mobility solutions. AI can effectively address pressing urban challenges by enabling adaptive traffic management systems, predictive analytics for public transportation, and efficient integration of micro-mobility solutions. AI-driven mobility solutions offer proactive rather than reactive management of transportation networks, reducing congestion, enhancing environmental sustainability, and improving overall urban life.

This study establishes foundational insights into integrating AI within urban mobility, using Bucharest as a critical case. The insights gained here provide a basis for future, broader explorations of AI applications in integrated logistics, specifically for last-mile and final delivery.

Bucharest, the capital and largest city of Romania, represents a pertinent case study of these urban mobility challenges. With a metropolitan area population exceeding two million residents, Bucharest faces severe urban mobility issues, including chronic traffic congestion, limited parking availability, overcrowded public transport services, and elevated air pollution levels. According to the TomTom Traffic Index, Bucharest is ranked the 4th most congested European city and 22nd globally [1]. Addressing this problem is necessary for improving daily life and also for enhancing the city’s economic attractiveness and sustainability.

1.1. Review of the Current State of the Research

A primary focus in current research is the intersection of smart mobility practices with urban development and tourism. Marchesani et al. [2] argue that the integration of smart mobility practices significantly influences tourism inflows by improving overall accessibility and connectivity within cities. This point is echoed by Wawer et al. [3], who suggest that the tech-savvy orientation of Generation Z brings new demands for sustainable urban mobility, thereby advocating for the adoption of Information and Communication Technology (ICT) solutions to engage this demographic effectively. The research indicates a clear relationship between innovative mobility solutions and urban attractiveness, which is a key factor for city rejuvenation and economic sustainability.

One study [4] has demonstrated the utility of geospatial technologies in analyzing how transportation hubs significantly influence business location decisions within urban agglomerations. Specifically, in Bucharest’s metropolitan area, it has been shown that passenger traffic from international airports and railway stations has a substantial impact on hotel clustering patterns, underscoring the importance of strategically aligning transportation infrastructure with hospitality businesses through integrated urban planning and advanced technological tools.

The popularity and accessibility of AI technologies has raised the need for a re-evaluation of traditional transportation management systems. Přibyl et al. [5] emphasize the need for interdisciplinary approaches to urban tunnel management within smart cities, suggesting that AI can enhance traffic management efficiency beyond conventional methods by utilizing interconnected data systems. This analysis supports the broader consensus that AI’s role is decisive in transforming urban mobility from reactive to proactive frameworks, enhancing real-time decision-making processes.

This is also supported by recent research, highlighting the role of AI in dynamic optimization for complex, time-varying environments. In logistics, integrated reinforcement learning has been applied to dynamic path planning for automated guided vehicles, enabling systems to adapt routes in real time to minimize congestion and improve throughput in smart operations [6]. Similar principles are emerging in networked infrastructure management, where AI-driven optimization supports resource allocation, traffic management, and dynamic network slicing in 5G systems, demonstrating that adaptive, data-driven control can enhance throughput, reduce latency, and maintain service quality under fluctuating demand [7]. These studies underline the potential of AI techniques (particularly reinforcement learning and related adaptive control frameworks) to improve system efficiency by continuously responding to changing conditions. Our approach applies these concepts to the urban traffic domain, evaluating how an adaptive signal control strategy can mitigate peak-period congestion in a large-scale simulation of Bucharest.

While many studies report substantial performance gains from AI-based adaptive traffic control, field deployments often reveal mixed results due to practical constraints. In Toronto, a multi-agent reinforcement learning system (MARLIN-ATSC) deployed in the downtown network delivered corridor-level improvements, but gains were inconsistent due to detector errors, noisy data, and coordination difficulties across high-demand corridors [8]. In Pittsburgh, the decentralized SURTRAC system achieved 20–40% reductions in travel times and over 20% emission cuts during initial evaluations, yet follow-up observations found diminished performance during severe congestion or hardware failures [9]. The Sydney Coordinated Adaptive Traffic System (SCATS), widely implemented in Australia and abroad, has shown significant environmental benefits, such as more than 20% reductions in fuel use and emissions, but performance improvements vary by location, time of day, and prevailing traffic patterns [10,11,12].

Similar patterns emerge in other contexts. In Tehran, adaptive controllers maintained strong performance under stable flows but degraded markedly when faced with pedestrian phases, irregular incidents, or sensor noise [13]. In Las Vegas, field evaluations of SCATS on selected arterials reported measurable travel time reductions during certain periods, but performance gains were less pronounced or absent in others, illustrating the influence of local demand patterns and corridor characteristics on adaptive control outcomes [14]. These examples highlight that the effectiveness of adaptive traffic control depends on algorithm design, sensor accuracy, data quality, operational robustness, and the physical and institutional constraints of each network.

Consequently, the present study interprets its simulation results as indicative of upper-bound performance under idealized conditions, rather than direct forecasts of real-world outcomes. This framing reflects the acknowledged simplifications in the modelling process and is consistent with the variability observed in previous deployments.

Moreover, the concept of multimodal transport systems is gaining traction as cities strive to create cohesive mobility networks. Hasan and Fadli [15] argue that implementing multimodal transportation integrates various transport options such as walking, cycling, and public transport, thereby improving the quality of life in urban settings. This integration also enhances sustainability outcomes by reducing reliance on single-occupancy vehicles, a position reflected in the analysis by Mello and Faxina [16] that highlights the need for collaborative schemes among transport modes to make cities more efficient and environmentally friendly. Such studies highlight a consensus on the necessity of integrated systems that effectively utilize AI to manage complexities in urban transport dynamics.

In a global context, cities like Barcelona have been spotlighted for their innovative approaches to smart mobility. Soriano-Gonzalez et al. [17] address the key performance indicators for mobility logistics in Barcelona, advocating for a framework that aligns urban mobility with sustainability principles. This research illustrates how smart technologies are transforming from just tools into foundational components of urban policy and planning frameworks that prioritize operational efficiency and environmental responsibility.

Research into the application of autonomous technologies has demonstrated significant logistical efficiencies and sustainability improvements, especially within maritime and last-mile logistics contexts [18]. Extending these findings, the current literature suggests AI’s transformative potential in urban mobility, optimizing multimodal transportation systems, reducing congestion, and enhancing urban sustainability. These insights underline the scalability and adaptability of AI-driven solutions across various transportation segments, further emphasizing the necessity for comprehensive integration into smart city frameworks.

The implementation of these technologies comes with associated challenges. Mentsiev et al. [19] highlight both the potential and risks associated with integrating intelligent transport systems in urban environments. Issues such as cybersecurity and data privacy are primary concerns that must be strategically addressed to ensure public trust and the technology’s continued adoption. A disputed area within smart cities discourse is represented by the need for comprehensive regulatory frameworks to govern these advancements.

One emerging area of research includes the role of artificial intelligence in decision-making processes related to urban mobility. Arora et al. [20] discuss how multi-agent systems driven by AI can enhance efficiency in urban settings by enabling real-time traffic management and smart resource allocation. This creates a more responsive urban mobility landscape, but the extent to which such systems can replace human-led management remains a subject of debate among scholars.

Controversies also arise regarding the balance between technological integration and the implications for social equity and access. Li et al. [21] suggest that while AI-driven solutions promise efficiency, there is a danger that they may exacerbate existing inequalities in urban transport access if not implemented thoughtfully. This caution is further supported by the necessity for participatory approaches in the design and deployment of smart city technologies, as highlighted by various researchers advocating for community involvement in shaping mobility solutions.

Furthermore, transitioning from traditional transport systems to smart solutions requires significant investment and strategic foresight. In emerging markets, cities are leveraging ICT and big data analytics as part of their strategic urban development plans, as noted by E-Elahi et al. [22]. However, different levels of technological and infrastructural readiness across regions create varying path dependencies affecting how smart mobility solutions are adopted and scaled.

The literature also emphasizes the importance of sustainable practices in implementing smart mobility. A study by Cepeliauskaite et al. [23] highlights the necessity for cities to adapt their transport strategies in alignment with global climate goals, stating that sustainable mobility solutions are the main approach for mitigating urban transport’s carbon footprint. The call for transformative actions is echoed in Tahmasseby’s [24] work. The study stresses that the integration of ICT with intelligent transportation systems is essential for improving mobility and also environmental outcomes in urban areas.

The integration of innovative technologies into mobility services determines a rethinking of urban planning processes. Kovačević [25] highlights promising practices from cities globally that successfully balance technology adoption with citizen engagement. He suggests that a model of smart transport must include diverse modes of mobility, including active and shared transport. This observation points to a fundamental shift in how urban planners must conceptualize transportation as part of a broader urban ecosystem.

Simulation methodologies, notably Multi-Agent Transport Simulation (MATSim), have been widely utilized to validate the potential benefits and practical impacts of new technologies in transportation planning and management. In urban contexts specifically, MATSim enables detailed modeling of individual transport behaviors and interactions, providing insights into how mobility solutions can optimize traffic flow, enhance operational efficiency, and improve overall sustainability. These methodologies help researchers and urban planners to accurately assess the real-world implications of deploying adaptive traffic management systems, predictive public transport fleet deployment, and micro-mobility solutions by simulating complex scenarios and offering granularity in the simulations’ outcomes.

International research highlights the benefits of AI-driven traffic management, but there remains a clear geographic gap in the literature for Bucharest and Romania, where few studies have rigorously evaluated such systems. Moreover, although adaptive signal control has been successfully demonstrated in other cities, the literature still lacks well-defined, scalable methodologies for implementing these adaptive systems in varied urban contexts.

1.2. Study Objectives

This study aims to explore how Bucharest can integrate AI technologies into its urban mobility ecosystem, setting the city on a path toward becoming a smart, efficient, and sustainable metropolitan area. Specifically, the paper investigates the potential impacts of adaptive traffic management, predictive fleet deployment in public transport, and micro-mobility solutions powered by AI. The goal is to demonstrate how AI-based interventions can significantly reduce congestion and environmental impacts.

The findings indicate that adaptive traffic control systems and predictive analytics for fleet management substantially reduce traffic congestion and contribute to sustainability objectives. Furthermore, the analysis highlights the essential roles of robust governance, transparent public–private collaboration, and adaptable regulatory environments in successfully implementing AI-driven solutions. Ultimately, this paper provides practical insights, scalable strategies, and actionable recommendations for urban planners, policymakers, and stakeholders interested in leveraging AI technologies to enhance urban mobility and sustainability.

By examining adaptive traffic control systems, predictive analytics, and micro-mobility solutions, this article sets the stage for a more comprehensive future investigation of AI-driven integration across multiple transport sectors. The remainder of this article is structured in four sections, as follows: The Materials and Methods section describes the study area, data sources, and simulation methodology, detailing the network preparation, demand modeling, and scenario design. The Results section presents the results of the simulation experiments, comparing fixed-time and adaptive signal control performance. The Discussion section discusses the implications of these findings for urban mobility planning and policy, while the final section concludes by summarizing key insights and suggesting directions for future research.

2. Materials and Methods

2.1. Study Area

This study focuses on Bucharest, Romania’s capital and largest urban agglomeration. The simulation model concentrates on the north-western area, centered on Piața Victoriei, [Victory Square] a critical congestion hotspot, and the surrounding arterial corridors. This area, with high congestion levels, was selected due to its representativeness for Bucharest’s systemic mobility challenges and its strategic importance for adaptive traffic management interventions.

The Research Methodology

The methodology follows a seven-step workflow, presented in the numbered outline below. Figure 1 summarizes this sequence, showing the progression from data acquisition through network preparation, scenario configuration, demand modeling, and simulation, followed by performance measurement and results interpretation. This structure clarifies the order of tasks and the flow of information used in the study.

• Data acquisition: The authors collected network geometry from OpenStreetMap (OSM) and traffic flow data from TomTom Move. The TomTom datasets included August 2024 Route Monitoring data at the corridor level and April 2025 Junction Analytics data for Piața Victoriei, enabling the estimation of relative demand and traffic distribution patterns.
• Network preparation: The OSM data was extracted and processed using QGIS 3.22. Lane counts, capacities, and free-flow speeds were standardized for urban arterial conditions. Minor service roads with negligible traffic impact were removed to ensure computational efficiency.
• Scenario configuration: For this simulation were prepared two scenarios, fixed time and adaptive control. Fixed-Time Control scenario applied generalized signal timing plans with static cycle lengths and green splits, representative of the current operational system. Adaptive Control scenario implemented the Laemmer self-organizing algorithm from the MATSim signals contrib module, configured with instantaneous detection of all approaches and dynamic green time allocation.
• Demand modeling: Daily vehicle demand for Piața Victoriei was derived from TomTom corridor and junction datasets. The demand profile was time-disaggregated to reproduce observed diurnal variations.
• Simulation execution: Both scenarios were executed in MATSim (release 2025.0) under identical boundary conditions.
• Performance measurement: Outputs were analyzed for average network speed, delay, average speed per hour and number of agents traveling simultaneously.
• Results interpretation: Quantitative results were compared between scenarios, with emphasis on relative improvements. Limitations arising from model assumptions, such as perfect detection and exclusion of pedestrian phases, were acknowledged.

This structured workflow ensures transparency and reproducibility and isolates the effect of adaptive signal control from other potential network or demand variations.

2.2. Data Sources

2.2.1. Road Network Data

The study area focuses on Piața Victoriei, one of the busiest and most complex traffic nodes in Bucharest. This junction serves as a convergence point for multiple arterial corridors, including Șoseaua Kiseleff, Bulevardul Lascăr Catargiu, and Bulevardul Ion Mihalache, and carries high volumes of both private and public transport vehicles throughout the day. The location is known for recurring congestion, particularly during morning and evening peak periods, with observed vehicle queues often extending into adjacent intersections. The observed existing fixed-time signal plans, coordinated in limited phases, provide only partial responsiveness to fluctuating demand and do not include dynamic adjustments based on real-time conditions.

The road network was derived from OSM data using QGIS to extract and preprocess the geometry of streets and intersections. Footways and non-motorized paths were excluded to reflect vehicle-based urban traffic. The OSM geometry preserves the actual lane configuration, intersection layouts, and connectivity for the study area, ensuring that the spatial representation reflects real-world conditions. However, the authors made targeted simplifications to facilitate the simulation process, including the removal of minor service roads with negligible impact on the modeled traffic flows, and the standardization of lane capacities and free-flow speeds to match urban arterial norms. These adjustments were necessary to ensure network connectivity and compatibility with the signal control module, while avoiding excessive computational overhead. The final network used in MATSim comprises 3364 nodes and 6884 links, capturing both major corridors and adjacent streets to ensure realistic routing and congestion propagation. Figure 2 presents the layout of the network. The nodes correspond to intersections and links represent individual road segments connecting these nodes.

2.2.2. Traffic Demand

Travel demand was generated to represent typical weekday patterns in Bucharest. The population file includes 23,050 agents making 46,100 trips over a 24-h period, based on disaggregated origin-destination pairs and time-of-day departure profiles calibrated to observed mobility trends. This scale reflects approximately 3% of Bucharest’s motorized trips, applying a downscaling factor to maintain computational tractability while preserving demand structure. This proportion was obtained by comparing the total number of motorized trips recorded by TomTom Route Monitoring across all monitored corridors in Bucharest on a representative weekday with the total number of vehicles recorded by TomTom Junction Analytics for Piața Victoriei during the same period.

2.2.3. Calibration and Validation

Empirical traffic statistics were sourced from TomTom Junction Analytics and Route Monitoring [26]. This data provided corridor-level speed and congestion profiles (e.g., Titulescu–Moșilor route) and intersection-level movement and turn ratios for Piața Victoriei.

These datasets informed the network definition and simulation calibration, supporting the adjustment of capacity parameters and departure time distributions. Observed average speeds and congestion peaks were used to benchmark simulation outputs.

2.3. Simulation

The simulation was implemented in MATSim (Multi-Agent Transport Simulation). MATSim is an agent-based modeling framework that supports large-scale transport simulations with individual travelers pursuing activity schedules and optimizing routes over iterations. Its modular structure allows integration of adaptive traffic signal control logic.

Key modules used included signals contribution (modeling of traffic signals, including fixed-time and adaptive controllers) and analysis contribution (automated extraction of leg histograms and travel time statistics). Signalized intersections were identified based on the points shapefile extracted from OSM into QGIS, with node IDs used to define signal systems (see Figure 3). The authors created a traffic signal plan with baseline fixed-time. The signals were scheduled with a reasonable cycle, based on the observed pattern.

Figure 3 shows a close-up of an intersection modeled with signal groups in the simulation environment. The grey background represents the road layout, with beige nodes marking the locations where movements connect. The red squares and connecting lines depict the defined signal groups at the start of the simulation, all initialized in the red phase. Two scenarios were modeled in the simulation: fixed-time signal control and adaptive signal control. In the fixed-time control scenario, all signalized intersections operated with static cycle lengths and green splits derived from average observed timings, providing a baseline that reflects typical existing signal plans; in contrast, the adaptive signal control scenario allowed intersections to adjust green light time dynamically in response to real-time demand using adaptive logic based on observed queue lengths and approach volumes, enabling more responsive and efficient management of traffic flows under varying conditions.

The adaptive traffic signal control method implemented in this study is based on the decentralized, self-controlled algorithm originally proposed by Lämmer and Helbing [27] and implemented in MATSim’s signal module by Kühnel et al. [28]. It uses local sensor data to minimize waiting times and queue lengths at intersections by dynamically selecting which approach to serve next, guided by a priority index that considers predicted queues and outflow rates. While this algorithm is not newly developed for the present study, it applies artificial intelligence concepts by continuously analyzing real-time simulated traffic conditions and adjusting green time allocation dynamically to minimize queue lengths and balance approach delays. The Laemmer controller operates in a decentralized fashion, where each intersection makes decisions locally based on upstream and downstream traffic states, rather than following a pre-defined global schedule. A stabilizing regime ensures each approach is served within minimal intervals, preventing spillback and ensuring network stability.

Each intersection operates independently using local queue and flow data to dynamically allocate green times. The core of the algorithm is the priority index ${\mathsf{\pi }}_i$ for each approach i, calculated as:

$${\mathsf{\pi }}_i={\displaystyle \frac{q_i+{\widehat{q}}_i}{s_i}}$$

(1)

where $q_i$ is the current queue length, ${\widehat{q}}_i$ is the predicted arriving demand in the short-term horizon, and $s_i$ is the saturation flow rate. At every decision point, the approach with the highest ${\mathsf{\pi }}_i$ is given green, ensuring that the system minimizes total waiting time and prevents queue spillbacks. A stabilizing regime ensures fairness by guaranteeing minimum service intervals for all approaches, preventing starvation. This formulation enables fully decentralized, real-time adaptation to changing traffic conditions and has been demonstrated to reduce delays and improve network stability in agent-based simulations of urban environments.

Unlike fixed-time or simple actuated controls, this adaptive method can combine non-conflicting directions into stages, respect minimum green times, and react in real-time to changing traffic conditions (even during short overload periods), resulting in reduced delays and more stable queues in large-scale, agent-based simulations.

In the present simulation, the adaptive controller was configured to receive perfect detection inputs from all approaches, enabling instantaneous estimation of demand. This allowed the controller to adjust green phases in real time, extending or terminating phases as necessary to optimize throughput. The configuration parameters, such as critical gap times and maximum green limits, were set according to MATSim defaults for urban intersections but calibrated to the lane geometry of Piața Victoriei.

This approach is distinct from previous MATSim-based studies in two respects: first, it integrates the adaptive controller with TomTom-derived demand patterns at both citywide and junction levels, ensuring that the simulated control logic is responding to realistic traffic flows; second, it applies the controller to one of the most complex signalized junctions in Bucharest, characterized by multi-lane approaches, tram crossings, and heavy turning movements. While the adaptive algorithm itself is an established method, its application in this context, with integration of high-resolution floating-car data, represents a novel combination that has not previously been reported in the literature for Eastern European urban traffic.

The simulation results were compared to TomTom Move empirical data to assess the degree of realism in peak congestion replication and to evaluate the relative performance of adaptive control.

The simulation incorporates several simplifying assumptions to allow a controlled comparison between fixed-time and adaptive signal control. Signal plans in the fixed-time scenario use generalized green splits and cycle lengths derived from typical coordination settings, without site-specific fine-tuning. In the adaptive scenario, the signal controller operates with perfect and instantaneous detection of all approaching vehicles and is able to reallocate green time dynamically with no processing or switching delays. Pedestrian demands, crossing phases, amber intervals, and non-motorized modes are not explicitly modeled. All drivers are assumed to comply fully with traffic regulations, and vehicle performance characteristics are uniform. These assumptions remove many sources of real-world variability and constraint, meaning that the results for the adaptive scenario should be interpreted as an upper bound on achievable performance under the modeled conditions.

3. Results

This section presents the outcomes of the MATSim simulations comparing adaptive signal control and fixed-time signal control in Bucharest’s metropolitan area, with a particular focus on the Piața Victoriei zone and its connected corridors. The analysis includes network-wide performance indicators (average speed by departure hour and congestion levels) and compares these findings with observed traffic patterns derived from TomTom Move data.

3.1. Simulation Outcomes

The MATSim simulation was executed using a synthetic demand of 23,050 agents (in MATSim, agents are individual travelers within the simulated transport system. Each agent has a daily plan with specified activities, such as home or work, and the trips connecting them. During the simulation, these agents interact dynamically with congestion, traffic signals, and other agents, enabling the model to capture traffic patterns and the effects of control strategies like adaptive signaling on travel behavior). The agents completed 46,100 trips (an average of two trips per agent), with trip departure times and routes calibrated to reproduce observed traffic volumes and temporal patterns from the collected data. The road network was derived from cleaned OpenStreetMap data and calibrated using corridor-level and junction-level TomTom data where available (empirical data was collected in August 2024 and April 2025).

Figure 4 displays the results for the fixed-time scenario. Here, average speeds show significant hourly variability, with severe drops to 9–10 km/h (and even 6 km/h) during peak periods and only modest improvements during off-peak hours. This pattern reflects classic congestion dynamics under rigid signal plans lacking real-time responsiveness.

In contrast, Figure 5 shows the average simulated speed per departure hour for the adaptive scenario. Results indicate consistent performance across all hours, with average speeds maintained between 37 km/h and 39 km/h, suggesting that adaptive signal management can effectively smooth congestion peaks and support reliable travel times even under high demand conditions.

To validate the simulation patterns, TomTom Move data was analyzed for the Titulescu–Moșilor corridor in and for junction-level behavior at Piața Victoriei. Figure 6 summarizes TomTom’s hourly congestion levels and live speed estimates over a typical week (the data was collected in April 2025, where live congestion level is represented with a red line, usual congestion level with a dotted line and live speed with a blue bar).

Observed speeds fall to 10–20 km/h during peak periods and rise to 30–35 km/h off-peak, demonstrating marked rush-hour congestion consistent with urban traffic theory [1,29,30]. While the simulated adaptive scenario slightly overestimates average speeds compared to real-world data (by 5–10 km/h), the pattern of flattening peak-hour slowdowns aligns with the expected impact of well-coordinated, demand-responsive signal timing.

Conversely, the fixed-time scenario’s drastic speed declines during peak hours match the shape of observed congestion curves, underscoring the validity of the demand model and network structure in reproducing realistic bottleneck effects when advanced signal strategies are absent.

3.2. Implications, Opportunities and Recommendations

The results suggest that Bucharest’s metropolitan authorities could achieve meaningful congestion reductions through investment in AI-driven traffic signal control systems. While the magnitude of the simulated improvements should be interpreted cautiously, as they are reflecting simplifications in the network and demand model, the relative gains over fixed-time control are both substantial and consistent with real-world pilot outcomes in comparable cities.

Figure 7 illustrates the number of en-route agents over time for fixed-time versus adaptive signal control scenarios, highlighting the substantial operational advantages of adaptive approaches. The fixed-time control scenario shows pronounced peaks and sustained high numbers of en-route agents during morning and evening periods, indicating significant congestion and reduced network efficiency. By contrast, the adaptive control scenario maintains consistently lower levels of en-route agents throughout the day, with visibly flattened peaks, demonstrating its ability to respond dynamically to fluctuating demand and manage queues more effectively.

The simulation exercise integrated 19 major corridors (from TomTom route reports), and did not incorporate real-world calibrated signal cycle data or full turn-movement at intersections.

4. Discussion

The simulation results highlight the potential of AI-driven adaptive traffic management systems to address pressing urban mobility challenges in Bucharest’s metropolitan area. By integrating OpenStreetMap-derived networks, TomTom Move traffic data, and agent-based demand modeling in MATSim, this study demonstrates a practical approach for evaluating traffic management strategies in a real-world urban context.

A key finding is the stark contrast between the fixed-time and adaptive signal control scenarios. Fixed-time control, which mirrors many of Bucharest’s current intersections, results in significant congestion during peak hours, with average speeds dropping below 10 km/h. This closely resembles observed congestion patterns captured by TomTom Move, supporting the demand and network model’s overall plausibility.

In contrast, the adaptive control scenario maintains average speeds consistently around 37–39 km/h, demonstrating the theoretical benefits of responsive signal timing in smoothing peak-hour traffic and improving overall efficiency. While the adaptive scenario somewhat overestimates absolute speeds compared to observed data, it successfully captures the expected qualitative effect of demand-responsive control in flattening congestion peaks.

This pattern is consistent with urban mobility theory and with pilot deployments in other cities, suggesting that adaptive traffic signal control can meaningfully reduce congestion, enhance travel-time reliability, and support sustainability goals by reducing idling and stop-start driving. The results align with the broader literature on AI-driven urban mobility solutions, which emphasize the transition from reactive to proactive traffic management strategies enabled by real-time data and predictive analytics.

Comparable findings have been reported in other cities and modeling frameworks. In a realistic simulation of Dresden’s city center (13 coordinated intersections with buses, trams and 68 pedestrian crossings), Lämmer and Helbing’s decentralized, self-stabilizing control reduced average travel times and their variance relative to an optimized state-of-the-art controller and also reduced the temporal variability of speeds across the day, even under high demand. They also show shorter, more naturally distributed pedestrian red times, underscoring the reliability gains observed [31].

On a real Melbourne suburban network (Kew), de Gier, Garoni and Rojas compared fixed-cycle vs. adaptive signals and found that adaptive control improved the means of network observables and produced significantly smaller fluctuations, with the best performance when controllers used both upstream and downstream information, especially when pedestrian and transit constraints were not modeled in detail [32]. These findings suggest that the high stability observed in the MATSim simulation results is a known behavior in idealized adaptive-signal simulations, and that absolute speed values should be read as an upper bound on achievable performance, while relative improvements over the fixed-time baseline remain a robust indicator of benefit.

The study was conducted under some assumptions. The simulation uses generalized signal timing assumptions and does not fully capture real-world phasing, turn movements, or pedestrian crossing demands at Piața Victoriei and other key intersections. Demand data, while large-scale and consistent with overall trip volumes, lacks calibration against observed matrices or household travel surveys. These assumptions mean that while the relative difference between fixed-time and adaptive control is meaningful, the absolute speed levels should be interpreted with caution.

In addition, TomTom Move data for route-level and junction-level analysis, while highly valuable, was only partially available at the time of this study. Expanding coverage and integrating more junction monitoring over time would allow for more granular calibration of adaptive signal control strategies and richer validation of simulation outputs.

The magnitude of improvement observed in the adaptive scenario of this study reflects the idealized conditions of the simulation, including perfect detection, absence of pedestrian demand, and standardized lane capacities. Real-world deployments, such as those in Toronto, Sydney, and Pittsburgh, show that while adaptive control can deliver meaningful benefits, these are often moderated by operational constraints, variable detection accuracy, and external disruptions. Therefore, the adaptive results presented here should be viewed as an upper bound on potential gains, consistent with findings from prior field evaluations.

Based on the simulation results, the authors propose a system architecture for adaptive traffic signal control in Bucharest (Figure 8). This architecture connects local, decentralized adaptive control at intersections with centralized monitoring and optimization.

At the edge level, Intersection Nodes (IN) include sensors (e.g., cameras, loops, radar) feeding real-time data to AI-based controllers that optimize signal phases locally. The cloud-based Monitoring and Data Center (MDC) aggregates processed data from all nodes, enabling system-wide traffic management, performance monitoring, and operator interfaces. This design ensures responsiveness at the local level while maintaining strategic oversight and coordination city-wide.

The methodology demonstrated in this study is adaptable to a range of urban and regional contexts. In well-instrumented metropolitan areas, integration with existing detector networks, connected vehicle data streams, and central traffic management systems would allow near real-time operation of AI-driven adaptive control. In less digitally advanced cities or suburban and rural transition zones, the approach can still be implemented using reduced data inputs and lower-cost detection methods, such as aggregated counts from pneumatic tubes, manual surveys, or open-access datasets. While these alternatives would reduce the temporal and spatial resolution of the control system, they could still enable periodic re-optimization of signal timings based on observed patterns, yielding measurable benefits without full-scale infrastructure upgrades. The use of open-source platforms like MATSim also supports gradual scaling: initial models can be constructed from openly available map and traffic data, and progressively refined as additional sensor coverage and communication capabilities become available.

AI-driven traffic management systems raise broader considerations beyond operational efficiency. The use of high-resolution traffic data, whether from connected vehicles, mobile devices, or infrastructure-based detectors, introduces potential privacy concerns if adequate anonymization and data governance practices are not maintained. Furthermore, system benefits may not be distributed evenly: improvements in vehicle travel times could disproportionately favor car users over pedestrians, cyclists, or public transport riders unless multimodal priorities are explicitly incorporated into the control logic. In cities with unequal access to digital infrastructure, adaptive systems relying on advanced sensing and communication may also reinforce existing disparities, as neighborhoods with limited coverage or lower investment might not experience the same quality of service. Addressing these issues requires integrating privacy safeguards, equity-focused performance metrics, and multimodal optimization into future system designs.

5. Conclusions

This study demonstrates the feasibility and value of integrating AI-based adaptive traffic management into Bucharest’s urban mobility strategy. By leveraging open geospatial data, commercial traffic monitoring services like TomTom Move, and advanced agent-based modeling with MATSim, city planners can rigorously evaluate policy options and investment priorities.

The findings suggest that adaptive traffic control has the potential to substantially improve traffic flow, reduce congestion, and enhance travel-time reliability compared to fixed-time signal plans that fail to account for dynamic demand patterns. While results indicate somewhat idealized average speeds under adaptive control, the relative improvement over fixed-time scenarios aligns with empirical observations from cities that have implemented coordinated, responsive traffic signal strategies.

The study demonstrates a practical and scalable methodology for combining OSM-derived networks, commercial mobility data sources, and agent-based simulation tools like MATSim. This approach supports AI-informed traffic management strategies in large urban areas and can be adapted to improve data-driven decision-making for congestion reduction and sustainable mobility planning in cities. It demonstrates the potential of AI-driven adaptive traffic signal control to substantially improve traffic performance in a large-scale simulation of Bucharest’s Piața Victoriei area. By integrating high-resolution TomTom traffic data with MATSim’s signals module, we evaluated the effects of replacing generalized fixed-time plans with a fully demand-responsive controller. The results show marked improvements in average travel speeds and reductions in peak-hour congestion, with performance stability across the simulated day. While the absolute values reflect idealized modeling assumptions, the relative gains align with patterns reported in prior simulation-based adaptive control studies, reinforcing the robustness of the observed improvements.

The originality of this work lies in its integration of commercial-scale floating-car data with an open-source large-scale mobility simulation to evaluate adaptive control in a complex European urban junction. This approach is rarely applied in Romanian or Eastern European contexts. The study also advances the literature by combining methodological transparency with a reproducible simulation framework, enabling future researchers and practitioners to adapt the workflow to other urban networks. The findings provide an upper bound on achievable benefits, offering a valuable reference point for transport authorities considering the deployment of adaptive signal systems.

Limitations and Further Research

This study is limited by data availability and the scale of the simulation. The simulation did not incorporate real-world calibrated signal cycle data or full turn-movement counts at all intersections, relying instead on partial TomTom junction monitoring and generalized assumptions. Since only 19 major corridors were integrated from TomTom route reports, future work should include additional link-level monitoring and multi-modal interactions.

Despite limitations, the methodology proposed for integrating OSM-derived networks, commercial mobility data, and agent-based simulation tools like MATSim is scalable and able to evaluate AI-based urban traffic management strategies in a real urban/metropolitan agglomeration.