Multi-Agent Mapping and Tracking-Based Electrical Vehicles with Unknown Environment Exploration

Abstract

This research presents an intelligent, environment-aware navigation framework for smart electric vehicles (EVs), focusing on multi-agent mapping, real-time obstacle recognition, and adaptive route optimization. Unlike traditional navigation systems that primarily minimize cost and distance, this research emphasizes how EVs perceive, map, and interact with their surroundings. Using a distributed mapping approach, multiple EVs collaboratively construct a topological representation of their environment, enhancing spatial awareness and adaptive path planning. Neural Radiance Fields (NeRFs) and machine learning models are employed to improve situational awareness, reduce positional tracking errors, and increase mapping accuracy by integrating real-time traffic conditions, battery levels, and environmental constraints. The system intelligently balances delivery speed and energy efficiency by dynamically adjusting routes based on urgency, congestion, and battery constraints. When rapid deliveries are required, the algorithm prioritizes faster routes, whereas, for flexible schedules, it optimizes energy conservation. This dynamic decision making ensures optimal fleet performance by minimizing energy waste and reducing emissions. The framework further enhances sustainability by integrating an adaptive optimization model that continuously refines EV paths in response to real-time changes in traffic flow and charging station availability. By seamlessly combining real-time route adaptation with energy-efficient decision making, the proposed system supports scalable and sustainable EV fleet operations. The ability to dynamically optimize travel paths ensures minimal energy consumption while maintaining high operational efficiency. Experimental validation confirms that this approach not only improves EV navigation and obstacle avoidance but also significantly contributes to reducing emissions and enhancing the long-term viability of smart EV fleets in rapidly changing environments.

1. Introduction

The field of smart EV/EVs perception is rapidly evolving, addressing various challenges in situational awareness and autonomous navigation. Current research explores diverse applications, from disaster relief operations to self-driving EVs navigating unknown environments. This study aims to develop a cooperative, multi-agent mapping and tracking system for EVs operating in unstructured and dynamic environments. Unlike traditional route optimization methods, which primarily focus on minimizing travel time and cost, this research emphasizes situational awareness. By enabling EVs to dynamically perceive, interpret, and navigate their surroundings through real-time data processing, the proposed system enhances autonomous decision making, improving safety, efficiency, and adaptability in complex environments.

A key requirement for smart electric vehicles (EVs) is situational awareness, which enables them to understand their environment and identify obstacles in relation to their position. Effective situational awareness is essential for autonomous navigation, ensuring that EVs can adapt to dynamic surroundings and make informed decisions in real time. In this study, the term “multi-robot system” refers to a coordinated network of smart EVs that collaboratively explore and map their environment. These EVs function as autonomous robotic agents, utilizing distributed decision making to enhance mapping accuracy and obstacle avoidance. To achieve this, they must employ efficient methods for representing obstacles, allowing them to process environmental data and navigate effectively in complex, unstructured environments.

This study focuses on the development of a mathematical model that optimizes multi-agent EV mapping and navigation. The proposed framework integrates theoretical formulation and technical implementation, ensuring its practical applicability in real-world autonomous vehicle systems. A major challenge in autonomous EV navigation is the lack of efficient perception and mapping systems, particularly in unstructured or unknown environments. Traditional navigation methods rely on predefined maps and road infrastructure, which limit adaptability. This research introduces a self-adaptive system where EVs collaboratively create and refine topological maps in real time, enabling dynamic and responsive navigation. To achieve accurate environmental representation, EVs must efficiently process spatial data. This study explores data structures that allow EVs to accumulate and integrate sensory information over time, ensuring accurate obstacle detection and route planning. For example, an EV mapping a room may use LiDAR sensors to measure distances by emitting laser pulses and analyzing their reflections. By overlaying these measurements onto a grid-based occupancy map, the EV can differentiate between occupied and free spaces, optimizing navigation decisions. Beyond navigation, this research extends its findings to autonomous delivery systems, considering multiple parcel drop-off methods such as direct handover to recipients, secure parcel lockers, and designated drop-off points. The choice of method depends on recipient availability, security constraints, and vehicle capabilities. Understanding these logistical factors is essential for integrating autonomous EVs into real-world delivery networks. By addressing both navigation and operational efficiency, this study enhances the potential of EVs for multi-agent coordination, adaptive mapping, and efficient service deployment, (see Figure 1).

Enhancing the clarity of smart electric vehicle (EV) representations is crucial for improving visualization and data interpretation. Adjusting the background color to white is recommended, as it enhances the visibility of structural details, making key elements such as sensor placements, object detection outlines, and trajectory paths more distinguishable. This adjustment is particularly beneficial for diagrams illustrating occupancy grids, wireframe mapping, and object tracking models, as it reduces visual noise and improves readability. While the primary focus of this research is on multi-agent EV mapping, we also explore the role of drones in supplementing this process. Aerial drones can provide additional environmental data, offering an elevated perspective that enhances mapping accuracy. By integrating drone-collected data with EV navigation systems, the framework improves situational awareness, obstacle detection, and real-time route optimization, ensuring more efficient and adaptive navigation in complex environments.

An occupancy grid is a widely used method for creating environmental maps, providing an intuitive representation of spatial structures. This study explores how aerial drones can complement ground-based EV/EVs by offering an elevated perspective of the environment. Drones can collect valuable data on terrain, obstacles, and spatial structures, which EVs can then integrate into their navigation framework to improve mapping accuracy. However, as the mapped area increases, the memory cost of maintaining an occupancy grid becomes a significant challenge. While a single EV may handle its mapping efficiently, multi-EV coordination requires an optimized approach to sharing and processing data. The challenge arises when multiple EVs attempt to exchange and integrate large amounts of mapping data, leading to potential computational and communication bottlenecks. This study addresses these challenges by developing a cooperative multi-agent mapping system that enables EVs to efficiently share and update environmental data in real time, ensuring scalable and efficient navigation in large and dynamic environments.

The proposed system enables smart EV/EVs to collaborate by exchanging real-time data on road conditions, traffic congestion, and energy-efficient routes. This cooperative approach allows EVs to dynamically adjust their paths based on shared intelligence, optimizing travel efficiency while reducing energy consumption and emissions. By leveraging multi-agent coordination, the framework ensures that EVs work together to enhance both operational performance and environmental sustainability. However, real-time data exchange presents challenges in high-memory consumption and processing complexity, especially when mapping large areas. When transitioning from 2D occupancy grids to 3D spatial representations, memory demands increase significantly. Instead of a flat grid, mapping in three dimensions requires volumetric representations, such as voxels (box cells) or point clouds. A point cloud is a collection of x, y, and z coordinate points in 3D space, commonly generated using depth cameras or LiDAR sensors. These sensors capture both color information and depth measurements, allowing EVs to construct a detailed spatial model of their surroundings. While point clouds provide high accuracy in mapping, they introduce computational challenges, particularly for real-time data processing and object recognition. Unlike human perception, which can intuitively interpret 3D structures, EVs must process and analyze these data using advanced algorithms. This study explores efficient methods for handling large-scale 3D mapping and data exchange, ensuring that EVs can operate collaboratively and effectively in complex environments.

Unlike traditional vehicle navigation, which relies on static maps and centralized traffic data, smart EV/EVs must operate in dynamic and uncertain environments. This requires continuous real-time perception, adaptive routing, and energy-aware decision making. The proposed system addresses these challenges by integrating collaborative mapping, efficient route optimization, and energy-efficient path planning, ensuring sustainable and precise EV operations. Traditional fuel-based vehicle navigation models primarily optimize speed and fuel consumption, which are not directly applicable to EVs. Instead, smart EVs must consider energy efficiency, battery constraints, and charging station availability. Unlike combustion-engine vehicles, EVs require real-time route adjustments to account for dynamic charging opportunities and terrain-based energy variations. The proposed framework integrates adaptive decision making and multi-agent coordination, optimizing navigation specifically for EV operations, ensuring both efficiency and sustainability. A major challenge in autonomous EV navigation is understanding and representing environmental structures. Point cloud representations, which map environments in 3D space using x, y, and z coordinates, provide a detailed spatial model but introduce computational challenges. Converting a point cloud into a 3D mesh allows for a more structured representation, where each point is connected to its neighbors, improving path planning and obstacle detection. However, updating and maintaining these representations iteratively remains a challenge, requiring advanced algorithms for efficient real-time processing. This research builds upon previous studies in autonomous vehicle routing, parcel delivery logistics, and mapping techniques. By analyzing existing methodologies, this study provides a comprehensive overview before introducing an innovative approach to multi-agent EV navigation and mapping.

The effectiveness of smart electric vehicle (EV) navigation depends on clear and efficient environmental representations. To enhance visibility and improve data interpretation, adjusting visualization models—such as occupancy grids, wireframe mapping, and object tracking models—is essential. A white background in these models enhances structural details, making sensor placements, object detection outlines, and trajectory paths more distinguishable. While this study primarily focuses on multi-agent EV mapping, it also explores how drones can supplement this process by providing additional environmental data. Aerial drones offer an elevated perspective, capturing crucial information on terrain, obstacles, and spatial structures, which enhance real-time navigation for EVs. This collaborative data-sharing approach allows EVs to integrate both aerial and ground-based data, improving overall mapping accuracy. However, traditional occupancy grid mapping presents significant memory limitations, especially when scaling up to larger areas or multi-EV coordination. While a single EV can efficiently process its own data, multiple EVs sharing a memory-intensive map can cause processing delays and communication bottlenecks. To address these challenges, alternative 3D spatial perception methods such as point clouds and 3D meshes are explored. Point clouds, which store x, y, and z coordinates, provide detailed depth perception but also introduce computational challenges for real-time decision making.

Unlike traditional vehicle navigation, which relies on static maps and centralized traffic data, smart EVs must operate in dynamic, uncertain environments. This requires continuous real-time perception, adaptive routing, and energy-aware decision making. The proposed system addresses these challenges by integrating collaborative mapping, allowing EVs to share and update environmental data dynamically, efficient route optimization, which adjusts paths based on real-time traffic and energy efficiency metrics, and energy-aware navigation, which considers battery constraints, terrain variations, and charging station availability. Traditional fuel-based vehicle navigation primarily focuses on speed and fuel consumption. However, EVs require real-time route adjustments to optimize energy usage and charging opportunities. Unlike combustion-engine vehicles, smart EVs must consider battery levels, regenerative braking, and terrain-based energy fluctuations. The proposed framework integrates adaptive decision making and multi-agent coordination, ensuring optimal efficiency and sustainability in EV navigation. This research builds upon previous studies in autonomous vehicle routing, parcel delivery logistics, and mapping techniques. By analyzing existing methodologies, this study establishes a theoretical foundation before introducing an innovative approach to multi-agent EV navigation and mapping.

2. Description and Methodologies

Representations used in autonomous vehicle mapping are often overlooked, yet they play a crucial role in navigation and environmental perception. This section reviews existing research on autonomous vehicle routes, environmental mapping, and multi-agent systems, summarizing key findings that serve as a foundation for the proposed approach. Rather than relying on assumptions specific to this study, this discussion references prior research to provide insights into various mapping strategies, their applications in autonomous navigation, and potential optimizations that inform this work. In autonomous navigation, data structures are essential for EVs to accumulate, process, and integrate environmental information over time. One common approach is the occupancy grid, which represents an environment by mapping distance measurements collected using sensors such as LiDAR. For instance, an EV mapping a room with a LiDAR sensor would emit laser pulses at different angles, measure the time taken for the signals to bounce back, and overlay the gathered data onto a grid. This process allows the EV to classify areas as occupied or free, creating an intuitive and structured spatial representation. While occupancy grids are effective, they present challenges when scaling up to larger environments. As the mapped area increases, memory requirements grow significantly, making it difficult for multiple EVs to efficiently share and integrate data. To address these challenges, alternative multi-agent mapping approaches are explored, allowing collaborative data processing and real-time updates while maintaining computational efficiency.

As smart EV/EVs generate and exchange environmental maps, the increasing memory demands can slow down data processing, especially when transitioning from 2D to 3D representations. Unlike traditional occupancy grids, which represent space using a flat grid, three-dimensional mapping requires volumetric models such as voxels or point clouds. A point cloud is a collection of x, y, and z coordinates representing the 3D structure of an environment. These are commonly generated using depth cameras—which function as a hybrid between traditional cameras and LiDAR sensors—capturing both color and depth information. This representation enables iterative data accumulation, allowing EVs to overlay and refine their perception of surroundings over time. However, while point clouds are effective in detailed spatial mapping, they introduce challenges in data processing, object recognition, and navigation efficiency. This study approaches EV navigation and mapping from an engineering perspective, focusing on how EVs autonomously construct, share, and refine their maps. While logistical concerns such as delivery efficiency and route planning are relevant, they are not the primary focus. Instead, the goal is to enhance the EVs’ ability to perceive, navigate, and adapt in unknown or dynamic environments. To improve spatial awareness, point clouds can be further processed into 3D mesh representations, where each point is connected to its neighboring points, forming a structured surface. This approach solves some of the ambiguities present in point clouds, making it easier for EVs to interpret obstacles and traversable areas. However, 3D meshes introduce additional computational challenges, as they require iterative updates and efficient processing algorithms to remain useful in real-time applications. This study explores various mapping strategies, analyzing their strengths and limitations to develop optimized solutions for multi-agent EV coordination and real-time environmental adaptation (see Figure 2).

This study explores various mapping and tracking strategies for smart electric vehicles (EVs), ranging from high-level environmental representations to photorealistic object modeling. By examining these approaches, we aim to identify how their unique characteristics can be leveraged to enhance algorithmic performance and improve EV navigation. At the core of this research is the development of advanced mapping techniques, focusing on how EVs perceive, interpret, and track their surroundings. This study primarily investigates multi-agent EV mapping, where vehicles work collaboratively to construct and refine real-time spatial representations. By integrating efficient mapping and tracking algorithms [1], this research seeks to optimize EV perception, navigation, and adaptability in dynamic environments.

Smart EV/EVs rely on environmental mapping and obstacle tracking to navigate efficiently. Large-scale mapping enables EVs to construct a static representation of their surroundings, while local tracking focuses on identifying dynamic obstacles in real time. These two approaches are complementary—while mapping aids in path planning, real-time tracking prevents collisions and helps EVs adapt to environmental changes [2]. For example, an autonomous EV vacuum cleaner builds a map of its cleaning area, identifying cleaned and uncleaned spaces. It also detects obstacles, such as pets or furniture, and adjusts its path accordingly [3]. Both mapping and tracking rely on a continuous data processing cycle—an EV first forms a belief about its environment, takes sensor measurements, compares its observations to expectations, and updates its internal representation to improve accuracy. This study focuses on multi-EV collaboration, where multiple autonomous EVs operate as a coordinated system to enhance mapping efficiency and spatial awareness [4]. Rather than functioning as independent units, these EVs share data and optimize navigation collectively. In the Multi-EV Systems Lab at Taif University, research demonstrates that deploying multiple EVs for large-area mapping significantly improves efficiency, accuracy, and adaptability. By mapping expansive regions simultaneously, a multi-robot system enhances real-time decision making, navigation performance, and environmental perception compared to a single-EV system.

Multi-EV systems offer significant advantages in efficiency and robustness. With multiple EVs operating collaboratively, tasks can be completed faster, and system resilience is enhanced. For instance, if one EV malfunctions, the remaining EVs can continue the mapping process, ensuring task completion without interruption. The optimization of EV routing and navigation has been extensively studied, forming the foundation for smart EV mapping and tracking. Researchers introduced the electric vehicle routing problem [5] (EVRP), incorporating time windows and recharging stations to optimize EV movement. Others expanded on this by applying a metaheuristic approach [6], integrating fast-charging stations to improve efficiency. A team refined further EV routing by considering nonlinear charging functions [7], ensuring accurate range estimation and effective battery usage. Two researchers contributed by addressing time-dependent waiting times at recharging stations [8], a critical factor in dynamic EV routing. Finally, Çatay and Keskin [9] explored the role of quick-charging stations, refining strategies for minimizing energy consumption and optimizing EV navigation. Collectively, these studies provide a strong theoretical and methodological framework for real-world applications in autonomous EV navigation and multi-EV coordination. The concept of distributed multi-EV systems plays a crucial role in autonomous EV routing. Unlike centralized approaches, each EV operates independently, making real-time decisions based on its own energy constraints and environmental data. This decentralized approach improves navigation efficiency, particularly in complex urban environments, where charging demands and energy consumption significantly impact route planning. By integrating real-time energy considerations, distributed EV systems enhance overall fleet performance, ensuring optimized energy usage and adaptive navigation in dynamic and unpredictable conditions.

A centralized system relies on a single EV or computer to control all operations [10]. While this setup simplifies coordination, it also introduces a single point of failure and limits scalability. As more EVs are added, the central system may become overwhelmed, restricting the number of EVs that it can efficiently manage. In contrast, distributed systems eliminate this bottleneck by allowing scalability to any number of EVs, enabling more efficient and resilient operations. This paper is structured into four sections, each addressing a key aspect of multi-EV navigation and mapping. The first section introduces a framework for EV collaboration, emphasizing energy-efficient route optimization [11]. This framework integrates energy consumption into the electric vehicle routing problem (EVRP), ensuring that EVs balance battery limitations with real-world traffic conditions and charging station availability while mapping in a distributed manner. The second section explores topological mapping, a memory-efficient map representation. Unlike traditional maps that measure distances between objects and walls, topological maps abstractly represent connectivity between different areas. For example, in a multi-room building, a topological map captures which rooms are accessible from others, rather than their exact spatial distances. This mapping approach is particularly well suited for distributed multi-EV systems, as it allows an unlimited number of EVs to cooperatively analyze the structural relationships in an environment. The third section shifts focus from mapping to object tracking, maintaining the lightweight processing approach. While previous sections explored efficient mapping representations, this section emphasizes developing computationally efficient tracking algorithms that can run directly on EV hardware without requiring extensive external processing [12]. By integrating scalable distributed mapping, energy-aware navigation, and real-time tracking, this study aims to enhance multi-EV coordination, ensuring optimal route planning and obstacle avoidance in dynamic environments.

To ensure real-time performance, the algorithm must operate with high computational efficiency. This is achieved, in part, by representing the electric vehicle (EV) using a minimal set of data points—specifically, in three dimensions: length, width, and height. This simplified model forms a 3D bounding box, significantly reducing the computational load while maintaining the necessary accuracy for effective navigation. Additionally, this study incorporates NeRF-Based Visual Odometry for Autonomous Vehicles, a technique that enhances both path estimation and object localization relative to the EV as it navigates through its environment [13]. This approach enables the EV to maintain continuous situational awareness, accurately tracking its position while identifying surrounding objects in real time. By integrating this method, the system improves trajectory estimation and ensures robust, adaptive navigation in dynamic environments.

To optimize the navigation and efficiency of smart EV/EVs, this research introduces a method that simultaneously calculates both the position of objects and the trajectory of EVs. This is achieved using Neural Radiance Fields (NeRFs) [14], a type of neural network [15] that has recently gained attention for its ability to encode the geometry and color of objects with remarkable precision. The NeRF enables the generation of photorealistic images, allowing EVs to perceive their surroundings in high detail, which is crucial for accurate path planning and obstacle detection. The proposed system focuses on jointly optimizing the positions of objects and intelligent EVs by employing an advanced routing framework powered by real-time data and artificial intelligence. This system dynamically adjusts EV routes based on factors such as traffic density, battery status, and environmental conditions, enabling vehicles to reach their destinations more efficiently while reducing both energy consumption and harmful emissions. This approach significantly enhances the sustainability and overall performance of smart EV operations. The methodology unfolds in four stages. The first stage involves the development of a distributed mapping framework to model the operational structure of multiple EVs working collaboratively. In the second stage, a subset of smart EVs was utilized to analyze the target environment, with the surroundings represented through a topological map. Unlike traditional maps that measure physical distances, topological maps focus on the relationships and connections between various locations, providing a more abstract yet efficient representation. The third stage centers on the implementation of the algorithm directly on EV hardware, designed for fast, real-time processing to support quick decision making during navigation. Finally, in the fourth stage, the system measured optical distances using the NeRF to enhance the EV’s capability to understand object connectivity within its environment. This process improves the accuracy of path estimation and strengthens the vehicle’s ability to adapt to complex, dynamic settings.

3. Wireframe Mapping and Environment Representation

It is both a map representation and the corresponding framework that allows a group of EVs to construct this map of the environment [16]. This includes how they can update the map, how can they explore the area, and how and when they come into communication range of one another. They share what they have learned. Experimentally, we have an EV moving through an environment, building one of these wireframe maps (see Figure 3):

The purpose of the wire mapping process in this research work is to illustrate the connection of EV/EVs, devices, and equipment to each other in an organized manner. The research results helped us understand and document the network structure and facilitate maintenance and network management. The purpose of using a wireframe to represent networked environments is to simplify the mapping process. The research results enabled the network to be presented in an organized and logical manner that makes it easy to understand how EV/EVs and devices are connected and interact between them. It also helps in accurately determining the locations of electrical devices and vehicles, facilitating exploitation, maintenance, and improvement in the network. We notice, as presented in Figure 4, that the environment is represented by a collection of straight lines. So, this is one of the assumptions we are making to achieve this extreme memory efficiency. Now, we might wonder what happens if we have an environment that is not well represented. In that case, we could perhaps use a different sort of geometric primitive like a spline, but that is beyond the scope of this work. We assume that we are mapping out something like a man-made structure, which is generally well represented by these rectilinear maps. The wireframe itself is a labeled embedded directed graph, and we are going to explain now what that means, with the Wireframe-Env. Representation.

A graph, in mathematical terms, is defined as a set of vertices (nodes) and edges (connections) that link these vertices. The following mathematical formulation represents the model proposed in this study. This model is an extension of previous research, building upon the foundational work of studies such as [X], [Y], and [Z]. While these prior approaches provide the basis for understanding EV routing and coordination, our model introduces a novel optimization framework specifically designed for autonomous EV networks. This framework incorporates dynamic EV coordination and real-time environmental adaptation, addressing the limitations of earlier models and enhancing the efficiency and flexibility of smart EV systems. The corresponding equation reflects this advanced approach, highlighting the integration of dynamic variables that allow EVs to respond to changing environmental conditions while maintaining optimal route planning and energy efficiency:

$$\mathrm{W}=\left\{\mathrm{V},\mathrm{\xi },\mathrm{\Phi }\left(\mathrm{\upsilon }\right),\mathrm{\lambda }\left(\mathrm{\upsilon }\right)\right\},$$

(1)

$$\varphi :\nu \to {\mathbb{R}}^2,$$

(2)

$$\lambda :\nu \to \left\{\mathrm{N},\mathrm{O},\mathrm{F}\right\}$$

(3)

So, in the wireframe, these vertices are going to correspond to corners in the environment and the edges. In this case, directed edges represent walls between the corners. It is an embedded graph because each one of these corners is embedded in ${\mathbb{R}}^2$, meaning that there is an actual physical location. So, an x y coordinate associated with each, in addition to the embedding each label or each vertex, has a label.

To enhance the visual clarity of the wireframe simulation, it is recommended to modify the image background to white. This adjustment significantly improves the contrast between wireframe elements and the surrounding space, making key structural details, such as edges and vertices, more distinguishable. This improvement is particularly important when visualizing the overlaid measured wireframe map against the black ground truth, as it enhances overall readability. Such clarity ensures that smart EV/EVs can effectively utilize the map for navigation and obstacle avoidance. As illustrated in Figure 5, the map includes elements that can be classified as either nominal occlusions or frontier points. A nominal point refers to any point from which two edges emerge, and, based on this definition, most points within the map are considered nominal. However, as the map evolves, it becomes useful to label specific points differently to reflect their roles in navigation. For instance, when an obstacle blocks the smart EV’s view of a rear edge, the endpoint of the occluding edge is identified as an occlusion point. The projection of this point onto the hidden edge is called a frontier point, representing the boundary between the observed environment and unexplored areas. This concept is central to the exploration algorithm, as it helps guide the EV in discovering new sections of the environment. The smart EV updates its understanding of the environment using a computational framework known as a particle filter, as depicted in Figure 6. While designing this model, certain points were labeled differently to capture their specific roles within the mapping process. Some points act as boundary points because they limit the EV’s line of sight, revealing only partial views of certain edges. These boundary points are essential in shaping the exploration algorithm, as they help identify areas that require further investigation. The particle filter operates through a continuous cycle of prediction, measurement, and updating, allowing the smart EV to refine its environmental model over time. The process begins with initialization, where an initial set of sample particles represents possible states of the EV within its environment. As the EV moves, these particles are updated to reflect potential changes in position. The system then takes sensor measurements and compares the observed data with predicted values, adjusting the particle distribution based on the accuracy of these predictions. To improve precision, a resampling process is carried out, selecting particles that most likely represent the EV’s true state. This cycle of prediction, measurement, and updating continues until the EV successfully reaches its goal. Throughout this process, the direction of the edges in Figure 5 symbolizes the evolving boundaries of the EV’s perception of its environment. The dynamic interplay between forecasting, real-time data collection, and continuous model refinement ensures that the smart EV can adapt effectively to new and changing environmental conditions.

The particle filter framework serves as the foundation for how a smart electric vehicle (EV) dynamically updates its internal map. If unfamiliar with Bayesian filtering, the underlying concept is relatively straightforward: the EV makes a prediction about its environment, compares this prediction with real-time sensor measurements, and then updates its belief based on the differences between the two. This continuous loop of prediction, measurement, and belief adjustment allows the EV to maintain an accurate understanding of its surroundings. At its core, the particle filter is designed to estimate probability distributions that represent the EV’s environment. This is achieved using a sample set of particles, with each particle representing a possible state of the EV or its surroundings. These particles are propagated through the system’s dynamic model, simulating how the EV and its environment might evolve over time. As new sensor data become available, the algorithm adjusts the likelihood of each particle based on how closely it matches the observed measurements. Particles that closely align with the real-world data are assigned higher probabilities, while less accurate particles are filtered out through a process called resampling. This ensures that only the most relevant particles are retained, enhancing the accuracy of the EV’s internal map. In the context of wireframe mapping, the particle filter is represented visually by blue circles (as shown in Figure 7). The Wireframe Particle Filter operates similarly to traditional particle filters but is tailored for mapping structural environments. The key difference lies in how the particles are used—not just to track the EV’s position but also to map the geometry of the surrounding environment. This approach allows the EV to predict spatial structures, compare them with sensor inputs, and refine the wireframe map accordingly. This process forms the basis of the exploration algorithm, enabling the EV to continuously improve its understanding of the environment. By integrating prediction, measurement, and resampling, the Wireframe Particle Filter ensures real-time updates to both the EV’s location and its environmental map, providing a robust framework for autonomous navigation and obstacle detection.

Each particle is an entire map. We can think of these particles almost as a map hypothesis. One particle might have an edge that does not exist in another, or the location of these corners could be different as the smart EV, then continues to take future measurements. The likelihood of each map or particles is going to be adjusted and, in this manner, we are able to both keep the most likely map at any given time, while still considering other options, and also remove individual edges. We can see the map being constructed and the transparency of the smart EV. Now, one thing that we want to point out is, when the EV is going to turn this corner, we will notice that the information in the top left disappears. While conducting experiments and tests, we calculated unexpected peaks and edges in unexplored areas within the survey potential. Wireframe maps were drawn, and the green line symbolizes the known directed wall, while the red line symbolizes the expected wall and conflicting directions. This indicates that the smart EV had a mismatch between what they would expected to see versus what it saw. An image of the Wireframe-Directed Edges (see Figure 8) is depicted below.

The development of the proposed mapping and navigation system for smart electric vehicles (EVs) highlights the critical role of directed edges in improving environmental understanding [17]. Before incorporating directed edges, the system faced challenges in correctly interpreting spatial relationships. For example, imagine standing in a room, looking at the wallpaper, walking through a door, and then turning around. You would not expect to see the backside of the wallpaper, but that is exactly the kind of mistake that early EV models made. Without directional cues, the system incorrectly predicted the visibility of surfaces from unrealistic angles. By incorporating edge directionality, the system identifies mismatches between predicted and actual sensor data, allowing it to adjust without penalizing edges incorrectly. This adjustment improves environmental mapping accuracy and helps avoid errors in spatial interpretation [18]. A similar challenge arises when merging maps are created by multiple EVs. Initially, the merged maps may appear accurate, but issues such as collapsed vertical edges can occur when relying on bi-directional edges. By using directed edges, the system easily detects and rejects faulty map merges, ensuring more reliable environmental representations [19]. These directed edges are not just beneficial for map merging but also form the foundation of the Wireframe Exploration algorithm, guiding EVs as they navigate unfamiliar areas. The exploration algorithm relies on frontier points, which act as dynamic goals for EV movement. The EV continuously seeks the nearest frontier point—the edge of its current sensing range. As the vehicle moves closer, the frontier shifts, revealing more of the environment. This approach encourages EVs to explore systematically, expanding their mapped territory efficiently. However, during mapping, unexpected vertices and edges occasionally emerged due to process errors in network design and device connections. These issues were resolved by reviewing the original network schematics and correcting the configurations [20]. In real-world experiments, map elements were rearranged to eliminate inaccuracies and ensure precise environmental representation. As the EV moves, the evolving map reflects its current understanding of the environment. For instance, when the EV turns a corner, previously visible areas might disappear from the map if the system’s predictions do not match new sensor inputs. This discrepancy highlights the importance of directed edges, as they help prevent such mismatches by maintaining consistency in environmental models [21]. To evaluate the system’s effectiveness, a series of simulated and real-world experiments were conducted. The experiments focused on mapping accuracy, navigation efficiency, and scalability. One key experiment tested how multiple EVs collaboratively built environmental maps using LiDAR and occupancy grids. The results showed that the system significantly reduced computational load while maintaining high spatial awareness [22]. In another test, varying numbers of EVs exchanged mapping data, demonstrating that increased collaboration reduced the mapping time and improved consistency. However, efficiency gains plateaued beyond a certain number of EVs, indicating diminishing returns with large fleets [23].

For real-time tracking, the MSL RAPTOR algorithm was used to detect and follow moving objects. The system outperformed traditional tracking methods, showing high adaptability to sudden environmental changes [24]. Further validation involved comparing the proposed approach with G-Mapping and point cloud-based SLAM. The new system achieved a 40% reduction in memory usage while maintaining high tracking accuracy, making it ideal for large-scale navigation tasks [25]. Most experiments were conducted in simulated environments using MATLAB and Gazebo, providing controlled conditions to assess mapping efficiency and tracking precision. Additionally, real-world trials with EV prototypes equipped with LiDAR and onboard computing systems revealed challenges such as sensor noise and network delays, which were less evident in simulations. Addressing these real-world issues will be a key focus of future research [26].

The system enables EVs to collaborate by exchanging real-time data on road conditions, traffic congestion, and energy-efficient routes. This cooperative approach allows vehicles to dynamically adjust paths based on shared intelligence, optimizing travel efficiency while minimizing energy consumption and emissions [27]. Multi-agent coordination ensures that EVs work together to achieve both operational efficiency and environmental sustainability. The exploration algorithm can be thought of as a “carrot on a stick”, motivating EVs to move toward unexplored frontier points. Once all frontiers are mapped, the system switches to a different exploration mode to maximize efficiency [28]. To enhance perception, drones are integrated as auxiliary mapping tools, providing aerial views that improve navigation in complex environments such as urban delivery zones or disaster response areas [29]. The real-time mapping process allows EVs to dynamically scan their surroundings using LiDAR sensors, neural network-based perception, and occupancy grids. Unlike static preloaded maps, this dynamic approach enables EVs to identify obstacles, adjust routes in real time, and collaborate with other vehicles to refine spatial awareness in challenging environments [30]. This research builds on existing studies, offering new insights into mapping strategies, their applications in autonomous navigation, and potential optimizations that advance the field of smart EV technology.

Lightweight Object Tracking and Representation

This section examines lightweight methodologies for object tracking and environmental representation, specifically designed for smart EV/EVs. The primary goal is to develop efficient algorithms capable of running directly on EV hardware without imposing significant computational demands. To achieve this, the approach simplifies object representations by reducing them to their core 3D dimensions—length, width, and height—which streamlines data processing and enhances system performance. By integrating neural network-based techniques, such as bounding box-based tracking, the system maintains high accuracy in detecting and monitoring objects relative to the EV’s position. This method not only improves real-time tracking but also supports dynamic object recognition in complex driving environments. Furthermore, the proposed methodologies are benchmarked against state-of-the-art algorithms to evaluate their performance. The results demonstrate notable advantages in terms of processing speed, adaptability, and energy efficiency, making these approaches highly suitable for real-world smart EV applications where fast, reliable, and resource-efficient object tracking is essential.

4. Multi-Agent Wireframe Mapping

As illustrated in Figure 9, the proposed system operates as a distributed system, meaning that it relies on minimal assumptions regarding the information shared among smart EV/EVs. Specifically, the system does not assume that the EVs share a common coordinate frame or origin, which allows for greater flexibility and scalability in dynamic environments [31]. When smart EVs come into communication range with one another, they must align their map information before merging their data effectively. This alignment process ensures that the environmental maps created by different EVs can be integrated accurately, enhancing the system’s overall spatial awareness. However, this is where the system encounters one of the limitations associated with using a sparse map representation. While sparse mapping is memory-efficient, it lacks the rich detail found in more data-intensive representations. These detailed maps can provide valuable context when determining whether two maps are correctly aligned. Without this additional information, the system may struggle to verify the accuracy of map alignments, potentially affecting navigation performance in complex environments [32,33]. Despite these challenges, the distributed nature of the system offers significant advantages in terms of scalability and robustness. By minimizing dependencies on shared coordinate systems, the framework allows EVs to operate independently while still being able to collaborate effectively when necessary. This approach not only reduces the risk of single points of failure but also improves the system’s ability to adapt to dynamic real-world conditions [34].

To address the challenges associated with map alignment in distributed smart electric vehicle (EV) networks, the proposed system draws inspiration from the computer vision field. Specifically, a feature vector is defined for each vertex in the map, based on the graphical structure. This vector incorporates relative distances and angles between neighboring vertices, which helps identify structural patterns within the map. By applying these features within the RANSAC (Random Sample Consensus) framework, the system proposes potential map alignments. RANSAC operates as a probabilistically rigorous trial-and-error method, effectively identifying the most likely alignment. However, it is important to note that the best alignment identified is not always the correct alignment, especially when dealing with misaligned maps, which can introduce numerous inaccurate edges that require correction later [35]. To mitigate this issue, the system incorporates a verification stage before accepting merged map information. When two smart EVs come within communication range, they exchange data, including both their environmental maps and their locations within those maps. The receiving EV uses the proposed alignment to navigate toward the other EV’s location, continuously comparing what it expects to see with what it observes. Based on these comparisons, the system can either reject the alignment entirely or approve the merger if the data are consistent [16]. This verification process significantly reduces the risk of propagating errors through the network, saving substantial effort in downstream corrections. The system also balances delivery time and energy efficiency through adaptive route recalculations based on real-time conditions. Experimental results show that increasing EV collaboration reduces the number of mapping iterations, improving overall efficiency without compromising speed. Moreover, the framework optimizes travel paths, minimizing energy waste while maintaining high levels of delivery reliability. Comparative studies demonstrate that this approach significantly outperforms traditional vehicle routing algorithms, particularly in EV-specific scenarios, by integrating real-time multi-agent mapping, wireframe-based navigation, and energy-aware route optimization [36]. For evaluation, wireframe mapping simulations were conducted using MATLAB, followed by more detailed experiments in Gazebo, which provided the improved modeling of LiDAR sensor errors. The system’s performance was benchmarked against G-Mapping, a widely used 2D particle filter method for creating occupancy grids. The results confirmed that the proposed method offers superior mapping accuracy, faster computational performance, and reduced memory usage compared to conventional approaches (see Figure 10).

The wireframe map, from a mathematical perspective, is described as a labeled embedded directed graph. In this structure, the vertices represent specific physical corners or key points within the environment, while the directed edges denote the connections or boundaries, such as walls or obstacles, that exist between these corners. The term “embedded” indicates that each vertex is associated with a specific physical location in space, typically defined by coordinates (e.g., x and y positions), allowing for accurate spatial representation. Additionally, the “directed” nature of the edges implies that these connections have a specific orientation or direction, which helps in defining pathways or movement constraints within the environment. This approach is like the graph-based methods commonly used in robotics, where environments are abstracted into networks of nodes and edges to facilitate tasks like path planning, navigation, and obstacle avoidance. The labels assigned to vertices and edges provide additional context, such as object types, distances, or connectivity properties, enhancing the map’s utility for autonomous systems like smart EVs.

The wireframe used in the proposed system is best described as a labeled, embedded, directed graph, and each of these terms plays a crucial role in defining its structure and functionality. Mathematically, a graph is simply a collection of vertices (or nodes) connected by edges. In the context of wireframe mapping, the vertices represent corners or key points in the environment, while the edges, specifically directed edges, represent walls or barriers between these corners, indicating the direction of connectivity. It is referred to as an embedded graph because each vertex is tied to a specific physical location in the real world, defined by x and y coordinates. This spatial embedding allows for the precise geometric representation of the environment, making it easier for smart EVs to interpret and navigate. To improve the accuracy of the wireframe model, experimental tests introduced the concept of delocalizing the feature vector for each vertex. Inspired by techniques from the computer vision field, this approach reduces dependency on fixed positions, making the system more flexible in dynamic environments [37]. A verification phase was also incorporated before merging maps from multiple EVs. During this phase, when two smart EVs transmit data, the system continuously scans and compares the expected environmental features with the actual sensor readings. This process ensures that only accurate mapping data are accepted, significantly reducing errors. Following initial tests, the system was further evaluated using Gazebo, a simulation platform that provided more accurate environmental models and improved error handling. The results demonstrated a noticeable reduction in mapping errors, with the average error rate in wireframe mapping dropping to 1.5%, even when using the same memory resources [38]. This improvement highlights the system’s efficiency in maintaining high mapping accuracy while optimizing computational performance.

The problem is to determine how a wireframe map compares to this other method. We had to make some decisions in order to compete on equal footing. The first decision is with an occupancy grid. We needed to choose the size of the squares because that affects the resolution, its able to capture. We chose a grid size that was equal to the threshold at which we would merge two wireframe corners into a single line, and, then, we chose a map size that the overall memory for the print of this grid matches that of the wireframe. It is a small environment, but, this way, they are on equal footing at this comparison. We are able to significantly outperform the other method in terms of accuracy, but this is not the kind of the situation in which we would use wireframe results. The wireframe that we want to test it on is a much bigger environment. This shows a floor plan of the National History Museum for one of the levels and is about 200 by 200 m. The red shows the measured wireframe map overlaid on the black ground truth, and we can see that it is accurate, visually speaking, but there are some corners that are cut. Despite that, this would be a perfectly usable map if we wanted to give it to a smart EV to acknowledge. Now, we can plan a path from one section to another and then use our onboard obstacle avoidance to avoid running into things another way. We evaluated this map, which is important if we ever have a multi-EV system that does actually work better with multiple EVs. So, we ran our experiment with multiple numbers of EVs at different communication ranges, each one multiple times, and this graph here shows the average number of iterations that it took to complete the map. It ran as expected, and we hoped for this decrease, although, for a given map size, it eventually saturates because of the addition of a new EV (see Figure 11 and Figure 12). The wireframe map is depicted in three colors: red, blue, and green. We started the experiment using different numbers of smart EV/EVs (numbers) 1, 2, 3, 5 and 8. Figure 11 summarizes the average scan and map completion iterations, identifying the used EVs and the distance covered to achieve this. We mention, for example, that, when using the number of three EV/EVs, we have the average repetitions for scanning and completing the map, as seen below. The NeRF would be a natural fit into this sort of framework. But these will be a natural fit together, and that is an area of active research in our laboratory.

The wireframe map is drawn in blue, and the distance covered to achieve this was 212 m. For the wireframe map in red, it was 143 m, and, in green, the distance covered to clear and complete the map was 80 m. We experimentally evaluated this map, and it is important to have a multi-EV system that works best with multiple EVs. We recorded satisfaction.

In conclusion, wireframe mapping proves to be a memory-efficient map representation that seamlessly integrates with the proposed framework, enabling a distributed network of smart EV/EVs to effectively explore and map their environment. Given its efficiency and compatibility with the framework, we have chosen to focus on wireframe mapping, skipping over topological mapping in this discussion. The wireframe approach not only optimizes memory usage but also supports the demands of distributed EV systems, ensuring smooth and synchronized operations even when multiple EVs operate in proximity, (see Figure 12). The implementation of lightweight tracking methods in smart EVs offers significant benefits, particularly in terms of energy efficiency. Traditional object tracking algorithms often require high computational power, which leads to increased energy consumption. In contrast, lightweight tracking techniques, such as simplified 3D bounding box representations and reduced processing overhead, enable smart EVs to track objects efficiently while minimizing energy usage. A notable example is the Monocular Sequential Lightweight Rotation and Position Tracking (MSL RAPTOR) framework. This system relies on minimal data processing while maintaining high tracking accuracy. By utilizing monocular vision and optimized feature extraction, MSL RAPTOR ensures efficient real-time tracking without overloading the EV’s onboard processors. As a result, smart EVs can allocate energy resources more effectively, leading to prolonged battery life and improved system performance. Furthermore, the integration of lightweight tracking with multi-agent wireframe mapping enhances the scalability of smart EV networks. Distributed tracking reduces redundant computations across multiple vehicles, promoting more synchronized navigation while maintaining lower power consumption. Ultimately, these methods contribute to the development of sustainable EV operations, optimizing both mapping accuracy and energy efficiency. Research results confirm that a distributed group of smart EVs can successfully explore and map environments, demonstrating the effectiveness of this approach in real-world scenarios.

5. Monocular Sequential Lightweight Rotation and Position Tracking on EVs

In this study, we focus on the Monocular Sequential Lightweight Rotation and Position Tracking (MSL RAPTOR) framework, which is designed for tracking objects relative to cameras. While MSL RAPTOR’s core functionality lies in object tracking, our primary goal is to run the algorithm directly on smart EV/EVs [39]. This aligns with our emphasis on lightweight computational approaches, which are critical for ensuring energy efficiency and real-time processing in autonomous systems. Unlike traditional tracking algorithms that demand substantial computational resources, MSL RAPTOR employs a lightweight representation, minimizing the need for extensive processing power while maintaining high tracking accuracy. This efficiency is achieved through optimized data handling and streamlined algorithmic design, which enables the system to operate effectively even on resource-constrained hardware. In practical applications, the framework can be observed in scenarios where a smart EV, such as a drone, tracks another object—like another drone—within a shared environment (see Figure 13). This capability highlights MSL RAPTOR’s adaptability in dynamic environments, offering robust tracking performance with minimal computational overhead. By integrating this framework, smart EVs can maintain precise positional awareness, enhancing both navigation efficiency and energy management.

Wireframe mapping is a memory-efficient map representation and goes along with this framework to enable a distributed group of smart EVs to explore and map an environment. At this point, we are going to skip over the other mapping method topological mapping and start talking about MSL RAPTOR. This, again, is a framework for tracking an object relative to the camera. But, again, our main goal here was to run the algorithm on the EV. But it is performed in a different way. The representation is lightweight and enabled to have required minimal processing power. MSL RAPTOR is an acronym. It stands for Monocular Sequential Lightweight Rotation and Position Tracking on EVs. We apologize for the slightly tortured acronym, but we wanted MSL or the lab title in the name plus the imagery evoked by the term “Raptor”. This is a little bit of a bird and prey heal for one smart Flight EV tracking another (see Figure 14). We have the results of an experiment that encodes a single sequential rotation over multiple time periods. The three images on the upper right side represent the movements recorded against location tracking on smart EV/EVs that were successfully completed.

The MSL RAPTOR (Monocular Sequential Lightweight Rotation and Position Tracking) framework operates by processing a sequence of monocular images or a video feed captured in real time. Using this 2D visual data, the algorithm estimates the 3D pose of objects within the environment, enabling effective object detection and tracking. In the experimental setup at Taif University Lab, one of the smart Flight EVs was utilized, equipped with an onboard camera for continuous image capture and an NVIDIA TX2 GPU for accelerated neural networking. The inclusion of the TX2 chip is critical as its integrated GPU significantly enhances the speed and efficiency of neural network computations, which form an integral part of the MSL RAPTOR framework. The object representation within MSL RAPTOR is designed to be lightweight. In its simplest form, tracking can be achieved using just the basic dimensions of the object—length, width, and height. This minimalistic approach reduces computational load while maintaining tracking accuracy. However, the framework is also flexible. When more detailed information about an object is available, it can be incorporated to improve performance. For example, tracking an open laptop is challenging with a simple 3D bounding box due to its irregular structure and empty spaces. By adding a few more key points, the system can more accurately capture the object’s shape, enhancing tracking precision. While it is possible to extend this representation to a dense point cloud for extreme accuracy, doing so would undermine the lightweight nature of the algorithm, increasing computational demands and defeating the framework’s purpose. The goal is to strike a balance between accuracy and efficiency, making it suitable for real-time applications where energy conservation and speed are critical. To improve the visual clarity of the MSL RAPTOR framework’s demonstration (see Figure 15), it is recommended to replace the gray background with a white one. This change enhances the contrast of key visual elements, making it easier to differentiate between various stages of the tracking process. A white background also improves the clarity of object representations, allowing for a more intuitive understanding of the step-by-step operations within the MSL RAPTOR framework [40]. This adjustment will help readers and researchers more effectively follow the framework’s workflow and tracking mechanics.

At a high level, the proposed system processes monocular or binocular images to estimate the 3D poses of objects in the environment. The framework is divided into two core components: the front-end and the back-end, both of which work together to enable real-time navigation and object tracking in smart EV/EVs. The front-end is responsible for processing incoming images and extracting angled 2D bounding boxes (BBs) that outline objects of interest. This information is then passed to the back-end, which uses an advanced algorithm based on the Unscented Kalman Filter (UKF) to update the 3D pose estimates for each tracked object. The UKF is particularly effective because it handles partial information efficiently, allowing the system to iteratively update the object’s state even when single measurements are insufficient. To further enhance navigation capabilities, the framework integrates Neural Radiance Fields (NeRFs), which combine real-time visual data from onboard cameras with depth data from LiDAR sensors. Cameras capture sequential images to estimate motion and refine positional accuracy, while LiDAR provides precise depth measurements for obstacle detection and terrain analysis. By fusing these inputs, the NeRF constructs an adaptive 3D spatial model, enabling smart EVs to navigate both indoor and outdoor environments with improved accuracy and reliability. An important feature of this system is its ability to quantify uncertainty in pose estimations. The UKF provides a rigorous measurement of uncertainty, which is then fed back to the front-end. This feedback loop allows the image processing pipeline to adjust dynamically, improving the accuracy of object detection and pose estimation over time. The smart EVs used in this study are equipped with onboard LiDAR sensors, depth cameras, and real-time processing units. These vehicles autonomously scan their surroundings, detect obstacles, and refine spatial data, enabling accurate indoor mapping and outdoor navigation. The distributed architecture of the system allows each EV to process and share environmental data, promoting scalability and real-time adaptability in dynamic environments. The front-end consists of two interconnected neural networks that collaborate to optimize data processing and tracking performance (see Figure 16). Together, these networks handle the transition from 2D image analysis to 3D pose estimation, ensuring robust and reliable navigation for smart EVs.

The core motivation behind the design of the proposed system is to achieve high-speed, real-time performance, which is critical for smart EV/EVs operating in dynamic environments. The framework integrates two key components: an object detector and a tracker network, each offering complementary capabilities to optimize both accuracy and efficiency. The object detector is a widely used type of neural network designed to analyze an entire image, identify objects within it, and draw bounding boxes around these objects while classifying them into specific categories (e.g., detecting a dog, bicycle, or truck in a scene). While this approach is highly effective for initial detection, it requires significant computational resources because it must process the entire image, which can result in slower performance compared to specialized tracking networks. To address this, the system incorporates a tracker network, often referred to simply as a tracker, which operates differently. Instead of scanning the whole image, the tracker focuses on the bounding box identified by the object detector and tracks its movement across subsequent frames. This enables the system to monitor how the object shifts over time, achieving much faster processing speeds than the detector. The tracker excels in scenarios where real-time responsiveness is crucial, as it only processes local information around the bounding box rather than re-evaluating the entire image. However, the tracker has a notable limitation: since it relies solely on localized data, it may lose track of an object if there are sudden movements, occlusions, or drastic changes in the environment. This issue was observed in experimental tests, where the tracker’s output was represented by green bounding boxes. Occasionally, the bounding box turned red, indicating that the system had reset the tracker using the object detector. This reset mechanism is triggered when the uncertainty metric, passed from the back end, spikes—signaling that the tracker’s predictions no longer align with actual measurements. For example, when a quadrotor drone briefly flew in front of a door, the tracker lost the object momentarily, but the system quickly recovered by reinitializing the tracker with the object detector. This dynamic interaction between the detector and tracker creates a robust front-end system capable of maintaining accurate tracking even in challenging conditions. The object detector initializes the tracking process, while the tracker handles continuous monitoring at high speeds, with the detector stepping in when needed to correct tracking errors.

Experimental results demonstrate that this hybrid system significantly outperforms traditional vehicle routing algorithms in EV-specific scenarios. By integrating real-time multi-agent mapping, wireframe-based navigation, and energy-aware route optimization, the framework achieves superior accuracy, faster mapping, and reduced computational overhead. Comparative analyses with conventional routing models confirm that smart EVs benefit from dynamic, data-driven path adjustments, resulting in efficient travel and minimal energy consumption. In practical applications (see Figure 17), the back-end plays a crucial role by providing uncertainty metrics that help regulate the front-end’s performance. The system’s ability to detect tracking failures and recover in real-time ensures robust, reliable navigation, making it highly effective for smart EVs operating in complex, fast-changing environments.

We are going to focus on how the back-end works and get familiar with the core idea here, to see this kind of predictable update cycle in action here. But our measurements are out at the front-end, so this is going to be a 2D bounding box. We need a way to predict what 2D bounding box we are going to see. So, we give it a 3D pose of the object, and, because of the way we are representing it, which is, again, with some subset points, we are able to project those points on for the image plane and draw the 2D bounding box around that. Then, we can compare that against the actual measured bounding box and use that to update this 3D pose. However, it is important to realize that there are multiple 3D poses that can give that same 3D information. So, why it is important to have a sequence of images from a single measurement? We are not able to uniquely define what the pose of the object is. Something we did that did not fully solve the problem but did help is the fact that we used angle bounding boxes as opposed to the more common axis-aligned bounding boxes. This tightly fits the object shown in the image and produces the space of poses that would give us the same 2D measurement. We wondered how the back-end works, and the core idea here is familiar. We are going to see this kind of predictive measure update cycle in action. Because our measurements or output of the front-end is going to be a 2D bounding box, we need a way to predict which 2D boundary box will appear. We will see, given the 3D pose of the object and the way that we are representing it, which is again some subset of points. We are able to project those points onto the image plane and draw the 2D bounding box around that. Then, we can compare that against the actual measured bounding box and use that to update this 3D pose. However, it is important to realize that there are multiple 3D poses that can give us the same 2D information. So, this is why it is important that we have a sequence of images, because, from a single measurement, we are not able to uniquely define what the pose of the object is. A high level looks at what the back-end is doing. We successfully completed displaying points on the image plane, as well as drawing a 2D bounding box. A comparison was made to the square, and the actual circumference was measured (see Figure 18).

We mentioned how these angled 2D bounding boxes are sent in, and, then, we used Bayesian filtering, in this case, an Unscented Kalman Filter (UKF), to estimate the pose and uncertainty of the object. If we have multiple objects, we can spin up a UKF to track each one. The framework can be handled, and the only thing we would need to add is some way to solve the correspondence problem when we run our object detectors so that we know which instance of this class bounding box corresponds to which object. We could accomplish that with something like a Hungarian algorithm. One thing that we want to mention about this UKF though, as part of its prediction step, is that it uses a dynamics model. And, if we do not know anything about what we are tracking, we must assume that it is going to be still, and any difference between any motion it has is going to be compensated for by the natural error correction of the UKF. However, because we obtained a class label as part of our object detector, we can use a class-based dynamics model. So, if we see that it is a car versus a Versa drone, we can use an accurate model, and that can help the tracker improve its performance [20]. Additionally, if we are tracking an object that has inherent symmetry in it, we can mention that based on the class well. Then, we can use that to reduce the state that we are estimating with the UKF at. For example, if we are estimating a soda can that is symmetric about the vertical axis, there is no real reason to estimate the yaw angle, and, in general, a smaller state space is easier for these filters to track. For the first way, we evaluated our results vs. RGB-D Methods (see Figure 19). Our experiments use the Hungarian algorithm. By using a dynamic category-based model, we were able to successfully use an accurate model. We evaluated our results on NOCS-REAL275, (see

Table 1. Results on NOCS REAL275.

Method	NOCS	ICP	Key Point Net	6-PACK	Ours	Method
Modality	RGB.D	RGB.D	RGB.D	RGB.D	Mono.	Modality
Bottle	Rerr	25.6	48.0	28.5	15.6	18.7
	Terr	1.2	4.7	8.2	1.7	9.8
Bow1	Rerr	4.7	19.0	9.8	5.2	16.2
	Terr	3.1	12.2	8.5	5.6	4.0
Camera	Rerr	33.8	80.5	45.2	35.7	23.4
	Terr	14.4	15.7	9.5	4.0	8.5
Can	Rerr	16.9	47.1	28.8	13.9	17.6
	Terr	4.0	9.4	13.1	4.8	9.0
Laptop	Rerr	8.6	37.7	6.5	4.7	17.4
	Terr	2.4	9.2	4.4	2.5	11.6
Mug	Rerr	31.5	56.3	61.2	21.3	35.3
	Terr	4.0	9.2	6.7	2.3	7.7
Overall	Rerr	20.2	48.1	30.0	16.0	21.8
	Terr	4.9	10.5	8.4	3.5	8.2

). In practice, rotation errors are reported in degrees and translation errors in centimeters. Figure 19 summarizes the details of the results obtained. We also have the percentage estimates and the estimated error value.

The MSL RAPTOR (Monocular Sequential Lightweight Rotation and Position Tracking) framework was tested using the Knox dataset, which is widely used for evaluating pose estimation algorithms. In comparison with state-of-the-art methods that rely heavily on RGB-D data (combining RGB images with depth information), MSL RAPTOR demonstrated impressive performance despite not using depth data. Although initial results showed that many competing methods performed well due to their depth-based inputs, MSL RAPTOR was able to achieve comparable or superior performance, significantly boosting our confidence in its capabilities. This success highlighted the framework’s robustness in handling complex scenarios with limited sensory input. The real challenge, however, was to determine whether MSL RAPTOR could operate effectively onboard a smart Flight EV in real-world conditions, beyond controlled datasets. The system’s performance was compared to the Single Shot Pose (SSP) method, which was the closest competitor capable of delivering similar results on hardware with limited computational power. The SSP uses a neural network that predicts 2D projections of an object’s 3D bounding box, followed by the Perspective-n-Point (PnP) algorithm to estimate the object’s pose. While the SSP was designed for speed, it managed to process data at only 3 Hz, whereas MSL RAPTOR achieved speeds of up to 9 Hz, making it nearly three times faster while maintaining higher tracking accuracy. Performance evaluations were further validated through cumulative error distribution plots, which illustrated the framework’s accuracy. In these plots, the x-axis represents the error threshold, while the y-axis shows the percentage of measurements within that threshold. Although all models eventually approach 100% as the error threshold increases, MSL RAPTOR reached this level significantly faster, indicating superior tracking performance with fewer errors. Real-world tests involved tracking objects at distances up to 8 m, even when they were moving rapidly, such as drones and other smart EVs. This was particularly challenging due to the high-speed motion and dynamic nature of the environment. To ensure accurate localization, the system utilized onboard odometry, allowing the smart EV to track the relative position of moving objects effectively, even when both the EV and the target were in motion. A key factor in MSL RAPTOR’s success was the integration of deep learning techniques with Bayesian filtering, which efficiently manages uncertainties in motion estimation. The framework optimizes Bayesian filtering using an approach inspired by the Cloth Simulation Filtering Method, commonly employed in LiDAR data processing to differentiate between ground and non-ground points. This method involves several steps: inverting the LiDAR point cloud, applying a cloth simulation to model terrain features, analyzing interactions between simulated cloth nodes, and ultimately extracting ground points to create accurate digital terrain models. By incorporating this technique, MSL RAPTOR reduced the amount of data required to track new objects, enhancing efficiency without compromising accuracy. Additionally, the system benefited from the use of class-specific labels, which unexpectedly improved tracking performance. Initially unplanned, this enhancement emerged during experimental testing and proved highly effective, particularly in refining object detection during complex navigation tasks. The ability to adapt and improve through such iterative discoveries further highlights the framework’s flexibility. While the initial focus was on wireframe mapping—a method for creating large, sparse maps—MSL RAPTOR was integrated to enable the real-time tracking of nearby objects using smart EVs. Experiments demonstrated the framework’s ability to maintain tracking accuracy across various distances and speeds, confirming its robustness under different operational conditions. To improve sensor reliability, especially in challenging environments, the framework incorporated interference mitigation techniques, such as bandpass filtering and LCD-based adaptive apertures. These methods enhanced the system’s ability to maintain accurate perception even in low-visibility conditions, where traditional sensors often struggle. Practical tests confirmed that dynamic noise filtering and adaptive exposure adjustments significantly improved object tracking reliability, ensuring consistent performance in both structured and unstructured environments.

In conclusion, the MSL RAPTOR framework has proven to be a highly efficient, accurate, and reliable solution for real-time tracking and navigation in smart EVs. Its ability to process pose estimations three times faster than comparable models, while reducing cumulative tracking errors, makes it ideal for applications in autonomous navigation, fleet coordination, and obstacle avoidance. The combination of deep learning with Bayesian filtering, alongside lightweight processing techniques, ensures that MSL RAPTOR is well suited for real-world deployments, offering robust performance in dynamic, fast-changing environments.

Addressing Ambient Interference and System Integration

This section explores strategies to mitigate interference from ambient light and improve system integration, particularly in challenging environments such as underwater or low-visibility conditions. To enhance signal clarity and communication reliability, the section introduces techniques like bandpass filtering, which isolates specific frequency ranges to minimize background noise, and LCD-based adaptive apertures, which dynamically adjust light exposure to optimize performance under varying lighting conditions. A key focus is the application of pressure-neutral resin casting, a cost-effective housing solution designed for small-scale robotic systems. This technology ensures structural integrity and protection from environmental factors like water pressure and corrosion, making it ideal for both underwater robotics and harsh terrestrial conditions. The discussion also highlights hybrid applications, where systems are designed to support simultaneous illumination and data communication. This dual-functionality approach improves the efficiency of robotic operations, enabling devices to transmit information while performing active sensing or lighting tasks. Finally, the section reviews commercially available Underwater Optical Wireless Communication (UOWC) modems and examines notable projects in the field of swarm robotics, emphasizing innovations that integrate robust communication systems with autonomous navigation capabilities.

6. NeRF Visual Odometry

Wireframe mapping has been explored as a method for smart EV/EVs to create large, sparse maps of their surroundings. Additionally, the MSL RAPTOR framework was introduced to enable the real-time tracking of nearby objects using lab-based EV prototypes. Now, the focus shifts to NeRF-based visual odometry, which not only helps localize objects in the environment but also estimates the EV’s own trajectory as it navigates through the space. NeRFs, or Neural Radiance Fields, are a type of neural network designed to reconstruct 3D representations of an environment based on 2D image inputs (see Figure 20). In the experiments conducted, the primary objective was to measure visual distance to accurately determine the locations of objects within the environment. Using a NeRF, the system was able to successfully capture the object color, intensity, position, and optional display direction, providing a comprehensive spatial representation. To improve the clarity and visibility of the Neural Radiance Field (NeRF) Exploration figure, it is recommended to modify the background color to white. A white background enhances contrast, ensuring that structural details, including object geometry, color encodings, and ray tracing elements, remain clearly visible. This adjustment helps in better visualizing how a NeRF contributes to EV navigation and object recognition, ultimately supporting autonomous movement and decision making within complex environments.

To begin, we are going to define what a NeRF is. A NeRF is a Neural Radiance Field. And it is a type of neural network. This network can encode the geometry and color of an object. We can query it at different points and obtain a density. A color density is technically related to the probability of a ray of light being stopped. But, in practice, if it is not a transparent object, we can use that as a stand-in for a kind of volume. This video on the right looks like it was just taken with a camera. It is impressive, but it is in fact an output of one of these networks. We can generate truly photorealistic images. The way that we will generate an image is assuming that we have a scene, in this case represented by a NeRF, and a position in the direction from which we want to generate. We can trace an array from the camera’s origin through a pixel and at different points along the ray. We can query this radian basically with the density and the color by combining all the values along the ray in a way known as alpha blending. We can obtain the value for that pixel, and repeating that process for each pixel in our image will give us that full image. There is one last benefit of this network. It is differentiable, and that means how a shift in the input and the shift in the output are related. We can back-propagate a change in output or a difference in the output through the network to update the input. This is going to be important for how we are able to update our EV, or our smart EV’s location, and object positions. Before we discuss our method, we determine the general high Visual Odometry level problem of visual odometry (see Figure 21).

Here, we have a classic vehicle problem, which usually deals with the given two images. We want to know how the camera, which has moved between them, is based on the kind of commonalities. Most typically, we will scan the image for the kinds of unique feature points, and we will try and find which points correspond to the same points in the other image, based on how these have moved. Regarding geometry, we are able to estimate how the camera has moved, and, if we conduct this experiment again, we can estimate how the camera that is on the EV has moved through the environment. There are other things we can do, such as a kind of a batch optimization over the previous frames to refine it, but at a high level (see Figure 22).

Vision-Only EV Navigation in a Neural Radiance World is the previous work that served as our inspiration. The goal was for the robot to plan and execute a trajectory through a world that was represented by a single nerve. There is one kind of subcomponent of this that we want to focus on. The prior work, Camera Pose Estimates, specifically shows how, in this method, with an environment, as a nerve, the EV was able to predict how we can correct our assumption about where we are in the world. In practice, it is the true position of the smart Flight EV and above. It is the belief where the EV thinks it is on the right in this box. There is an overlay of a transparent, real image that it receives, which is based on where it is overlaid with what it predicts, which is based on the incorrect pose. Now, based on the mismatch of these images, the EV can update its position, and we will see as it approaches the true position. The overlaid images appear to come into focus as the prediction starts to match the actual measurement. This is a little different than the previous two waves, but it is still similar to that same prediction measure. This is the update cycle that has been present throughout this entire paper. So, to summarize this previous work, it used a single NeRF of the environment and estimated a single camera pose at each time. What we will do in our method is estimate multiple cameras poses at multiple EVs earlier at one time along with each individual object pose, instead of having a single NeRF of the environment. We have a NeRF for each object, and, in Figure 22, there is a plant, a table and a chair NeRF VO-Overview. We assume that our smart EV has a camera on it and is going to be moving through the environment. It has some manners of performing odometry. Perhaps it is integrating an imu to give a noisy estimate of how it is moved. Each of these given EVs is the same EV, just at different time steps. So, it starts looking at the scene from behind the chair and circles it to look from behind, like MSL RAPTOR. The EV is going to be using an object detector to know what it looks at in each view. It has some manner of initializing that pose. However, the biggest assumption that we are making is the fact that we have a pre-trained nerve model for each of the objects in the scene. This is a big assumption, but we will introduce how future work can soften this NeRF VO-Sliding Window. We mentioned before how we are going to be optimizing multiple views at a given time, along with the object poses. But, if we always conducted all the EV views well, that trajectory, which is going to continue to grow as EV moves, and, eventually, the number of things that we are trying to estimate would be too big. Instead, we use what is called a sliding window approach, where we fix a set number of views and optimize those together, along with object poses. So, first, that includes views one, two, and three; then, two, three and four; then four, five, and six, and so on. And we can see on the side that these are what the EV is seeing from each of these corresponding views—the NeRF VO-Pose Graph. Now, a very common thing is a very natural expectation if we are estimating multiple objects in a single camera image. We will achieve multiple relative poses, one for each object from that view. But, if we have multiple views and we perform this, plus if we have connections between each camera view, then all those different connections that we mentioned are going to look like a tangled mess because there are cycles in this object, which is called the pose graph (see Figure 23).

These connections are noisy. We need a second level of optimization, known as post-graph optimization. However, we can avoid all of this and estimate fewer variables by always using a minimal pose graph for each of the windows in the sliding window framework. So, by using that strategy, all our poses are always relative to the first camera view, which is depicted in the NERF VO-Qualitative Results. These images show the first six of the camera views, and we see both the measurements. It is transparent, and the EV’s prediction of what is gone can be seen in solid. We can see initially that there is a mismatch as the sliding window progresses. We can see the first three frames be optimized. They will turn white, and a mismatch between the prediction and the measure will correct this window. Then, it will increment, continue to increment, and, once the screen (the view) turns gray again, that means it is out of range of the sliding window.

So, there is no longer optimization being performed on that camera poses. But, if the object appears in the new views, that position will still be optimized as well. So, as we can see at the end, all of what was previously mismatched between the initial estimate and the measurement are alive quantitatively. The error is depicted in Figure 24,

Table 2. NeRF VO experimental data results.

Mean Error	Translation (mm) Total Object Keyframe			Rotation (Degrees) Total Object Keyframe
Initial	122.7	98.3	54.1	2.8	5.6	1.7
Final	41.6	15.	51.4	0.56	0.63	0.56
Reduction	66	84	5	79	89	67

.

As expected, both the translation and rotation error decreased; otherwise, we would not see those views converge. But one very interesting thing that we did not expect originally was the fact that, if we break out like the error by the object and camera mean keyframe and break that down into rotation and translation, we will notice that the camera translation is corrected much less than these other values. At first, we were confused by that until we remembered the way in which we are optimizing this system, which is based on the photometric loss or the difference between what we render versus what we expect to see. If a camera is far away from an object and we rotate that camera, there is going to be a significant difference in those between the predicted and measured photos. However, if we translate the EV a couple millimeters or the camera a couple millimeters, there is not going to be a significant difference, which explains why this happens. But, overall, the error does decrease a fair amount, and, again, one of the big goals of our method is to understand the scene. Both translation and rotation decrease significantly.

7. Results and Discussion

In a distributed smart electric vehicle (EV) environment, each EV operates with its own set of localized data and unique orientation, often starting from different positions and facing different directions. This diversity creates challenges in terms of map alignment, object detection, and determining the true spatial structure of the environment. While each EV gathers its own data, the question arises: How do individual EVs determine a consistent understanding of their surroundings, especially when their initial conditions vary so significantly? When processing environmental data, the system typically starts with capturing an image, which is then analyzed by a neural network that identifies objects within the scene and generates 2D bounding boxes. These bounding boxes provide spatial references, but they do not offer information about the free space or empty areas within them. While more advanced neural networks, like semantic segmentation models, could provide pixel-by-pixel labeling for detailed object recognition, the current system prioritizes speed and efficiency, limiting the use of such computationally intensive methods. The results of this study indicate that the proposed mapping framework significantly contributes to energy efficiency in smart EVs. By employing a distributed wireframe mapping approach, EVs can reduce redundant computations and unnecessary movements, optimizing energy consumption across the fleet. Experimental data demonstrated that, as the number of EVs increased, the time required for mapping decreased notably. This improvement prevented excessive battery drain from prolonged scanning and navigation tasks, showcasing the effectiveness of collaborative mapping in distributed systems. Furthermore, integrating lightweight tracking methods such as the Monocular Sequential Lightweight Rotation and Position Tracking (MSL RAPTOR) framework enhances energy efficiency even more. This approach minimizes the need for extensive sensor processing, enabling EVs to function with lower power consumption while maintaining high tracking accuracy. The optimized tracking framework reduces the computational load on onboard processors, leading to extended battery life and improved operational range. Additionally, this study highlights the benefits of real-time adaptive navigation, where EVs dynamically adjust their routes based on changing environmental conditions. By selecting the most energy-efficient paths and avoiding high-consumption maneuvers, the system promotes smarter energy utilization. These findings confirm that the proposed system not only enhances mapping precision but also plays a critical role in reducing energy consumption, making it a practical solution for sustainable EV fleet management. Lightweight tracking techniques further contribute to this by reducing the computational burden, allowing EVs to operate efficiently while conserving battery life. The tracking framework minimizes redundant data processing, ensuring that only essential computations are performed.

As a result, smart EVs can allocate energy more effectively (see Figure 25), leading to longer operational times and improved navigation accuracy. While advanced object detection systems like Detectron2, developed by Meta (formerly Facebook), could be integrated into the framework, the current implementation relies on versions of YOLO (You Only Look Once) for real-time object detection, which aligns better with the system’s real-time performance requirements. Both YOLO and Detectron2 are flexible frameworks, and the choice depends on the specific application. In this context, the system primarily needs to detect objects and generate bounding boxes quickly and efficiently. When EVs operate in environments previously scanned by other drones, they can utilize pre-existing models to improve object recognition and predictive accuracy. This approach is closely related to meta-learning, a subfield of deep learning where models are trained to generalize across different tasks and environments. Unlike traditional neural networks, which are often task-specific (e.g., recognizing a specific type of object, like a cat), meta-learning networks can adapt to recognize a wide range of objects without requiring extensive retraining. For example, a standard neural network might only recognize office chairs of a specific design, while a meta-learning framework could identify various types of office chairs, even those with subtle differences in shape or color. This adaptability allows the system to render accurate object representations in real time, enhancing both object recognition and navigation capabilities. The system’s ability to generalize from prior data—whether from past experiments or other EVs—improves its performance in new, unfamiliar environments. Moreover, techniques such as interference mitigation using bandpass filtering and adaptive sensor calibration ensure consistent sensor accuracy even in challenging conditions like low-visibility environments. Practical experiments confirmed that these methods improve object tracking reliability, enabling smart EVs to maintain accurate perception in both structured and unstructured environments. This consistency is crucial for tasks like obstacle avoidance, autonomous navigation, and dynamic object tracking.

In conclusion, the integration of distributed mapping, lightweight tracking, meta-learning-based object recognition, and adaptive navigation strategies creates a robust, energy-efficient system for smart EVs. The framework not only improves mapping precision and tracking accuracy but also optimizes energy consumption, making it an effective solution for sustainable autonomous vehicle operations. The ability to leverage data from multiple EVs, adapt to new environments, and efficiently process sensor inputs ensures the system’s scalability and reliability in real-world applications.

The results of this study highlight the challenges and advancements in object detection, smart EV positioning, and environmental perception. A key finding is that smart EVs perceive their surroundings differently due to their varying initial orientations and positions. This variability complicates the process of constructing a cohesive environmental model. To address this, this study explored the use of angled bounding boxes for estimating available free space. However, the current approach generates only 2D representations of battery enclosures without assessing the interior volume. Advanced neural networks capable of pixel-by-pixel segmentation could provide a more detailed spatial analysis, though real-time processing remains a constraint. Object detection plays a crucial role in refining environmental perception, and this study compared the performance of different detection frameworks, such as YOLO and Facebook’s Detector 2. These models demonstrated flexibility in adapting to various object recognition tasks, provided that they output meaningful bounding boxes. The integration of drone-collected data further improved prediction accuracy, reinforcing the importance of leveraging external datasets. A notable advancement in this research is the exploration of “Meta-NeRF”, which extends beyond traditional neural networks that are typically designed for single-object recognition. Meta-learning capabilities enable the system to generalize across multiple object categories, allowing NeRF models to store and retrieve detailed shape and color information. A significant challenge, however, is ensuring accurate classification within the same object category. For example, variations in office chair designs must be correctly recognized to enhance rendering precision. These findings contribute to a broader understanding of smart EV navigation by linking object recognition with wireframe mapping and object tracking methodologies. Wireframe mapping provides a lightweight yet effective method for generating sparse environmental representations, while the Monocular Sequential Lightweight Rotation and Position Tracking (MSL RAPTOR) framework enhances real-time object tracking capabilities. Furthermore, interference mitigation techniques, particularly those addressing lighting variations, demonstrate the potential for improving detection robustness under changing illumination conditions. Future research should focus on optimizing tracking algorithms and exploring multi-agent collaborative frameworks, which could allow multiple EVs to share perception models, thereby improving real-time obstacle avoidance and path planning. By advancing neural network adaptability and integrating shared environmental data, this research lays the foundation for more efficient and intelligent smart EV systems.

8. Conclusions

This research introduces an innovative framework for optimizing object detection and smart EV positioning by leveraging differentiable neural networks. A key challenge identified is the reliance on training separate NeRF models for each object. To enhance scalability, future research should focus on category-level NeRF models, allowing generalization across diverse object types. Instead of creating new networks for each instance, dynamic adaptation techniques could refine neural weights in real time based on observed characteristics. Additionally, addressing environmental lighting variations through adaptive NeRFs would improve tracking robustness under different illumination conditions. Furthermore, this study explores the potential of differentiable simulation to refine mass and inertial property estimates, paving the way for real-time physics-based learning within smart EV systems. Beyond its technical advancements, this research holds significant real-world implications, particularly in autonomous navigation and robotic applications. Integrating adaptive NeRFs with wireframe mapping could improve decision making for autonomous vehicles, enabling them to navigate complex environments with greater efficiency. Moreover, enhanced object tracking methods have potential applications in logistics, surveillance, and dynamic robotic systems that require continuous environmental adaptation. Future studies should extend the framework to multi-agent systems, fostering collaborative mapping and decision making among EV fleets. By implementing shared perception models, autonomous systems could achieve more effective obstacle avoidance and path planning. Additionally, refining meta-learning strategies would allow smart EVs to generalize across diverse terrains, minimizing the need for extensive retraining on new datasets. In summary, this research presents a holistic approach to perception and navigation in smart EVs, beginning with a lightweight, distributed wireframe mapping technique for scalable environmental representation. The introduction of the Monocular Sequential Lightweight Rotation and Position Tracking (MSL RAPTOR) framework further enhances object tracking capabilities, striking a balance between computational efficiency and environmental awareness. By integrating these methodologies, this study contributes to the development of intelligent, adaptable EV systems, setting the stage for future advancements in autonomous mobility and collaborative AI-driven decision making.