This PDF file contains the front matter associated with SPIE Proceedings Volume 13051, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Foundation models, including Chat-GPT and its many variants, have come into prominence in the natural language processing (NLP) community thanks the ubiquity of text data readily available on the internet and the design of modern transformer architectures that can effectively learn from such data. However, popular DoD domains, such as autonomy and logistics, are faced with additional challenges not present in NLP and vision. In this talk, we discuss some of those challenges and opportunities for addressing them.

Battlespace management requires rapid processing of large amounts of data to facilitate informed decision-making. Large Language Models (LLMs) have demonstrated near or above human performance on a wide range of cognitive tasks. Current LLMs are unreliable, have poor explainability, and are prone to bias and hallucinations. As such they are unsuitable for many defense applications, but their abilities can be studied with future, improved LLMs in mind. Specifically, LLM capabilities to synthesize defense-relevant data and make decisions in a combat environment have been largely unexplored. The battlefield information and tactics engine (BITE) uses LLMs as observers and decision-makers in a military environment. A multiplayer video game focusing on modern mechanized combat, Squad by Offworld Industries Ltd., is used as an operating environment due to its moderate realism levels and focus on audio communication between players. BITE is tasked with ingesting tactical data, providing summaries of the current situation, and giving order to a squad of human players. The present work aims to qualitatively assess the suitability of BITE, and LLMs in general, for use in battlespace management systems. Shortcomings are identified in the areas of spatial awareness, decision-making time, and reliability. However, BITE exhibits instances of competent leadership and demonstrates a generalized understanding of modern mechanized combat. While current LLMs are currently deeply unsuitable for combat environments, BITE and similar approaches show promise in wargaming and training applications.

Image geo-registration is an essential technology with a wide range of applications in the geospatial intelligence space. Prior to recent deep learning advancements, image geo-registration primarily used area-based image correlation algorithms. The “line-photogrammetry” [1] concept, which uses lines instead of points for georegistration and triangulation, has not been widely adapted due to the difficulties of reliably and accurately detecting lines from geospatial images. With deep learning, this is no longer a challenge. At BAE Systems GXP, we have taken advantage of the recent advancements in deep learning and have developed DeepTie, which uses both tie points and tie features, for image geo-registration and triangulation.

DeepTie uses two deep learning models: (1)TieFeature; and (2)TiePoint. In this paper, we focus on the TieFeature model, as tie features provide much richer information than tie points. Analogous to identifying a person, to use the TiePoint model is to only use a person’s nose to identify that person, while the TieFeature model uses the whole face of the person for identification. While the TiePoint model is much faster than the TieFeature model, tie features are much more reliable and accurate than tie points. We have developed tie feature matching algorithms to reliably and accurately match tie features for images which are taken from different years, different seasons, and different sensors.

Image-to-image correspondence is important in numerous remote sensing applications ranging from image mosaicking to 3D reconstruction. While many local features used for these methods aim for robustness to changes in viewpoint/illumination, recent studies have suggested that traditional feature extractors may lack stability in multi-temporal applications. We have discovered that this is especially true in multi-modal sensor contexts, such as corresponding high-resolution UAS images to broad area overhead imagery (e.g., satellite images). This paper explores the performance of various local feature extraction methods as they pertain to image-to-image correspondence in scenes captured at different times, with different sensors. Experiments here specifically evaluate co-registration between low-altitude, nadir UAV frames, and imagery collected from satellite sources. Due to challenges in the localization of imagery with significantly different resolutions, spatial extents, and spectral characteristics, two further studies are presented beyond baseline evaluation. First, images undergo histogram matching to better understand how the discussed algorithms’ performance changes as image characteristics become more or less similar. Secondly, experiments are performed where key point feature matches are refined with information taken from segmentation maps inferred by pre-trained segmentation models. These methods are evaluated in regions where satellite and UAV images have been collected at different times, with spatial correspondences being hand-labeled.

With the proliferation of the Internet of Things (IoT), military science and technology have been reshaped and will continue to evolve to address modern warfare challenges. The Internet of Battlefield Things (IoBT) has emerged as a pivotal research domain for military operations. IoBT utilizes machine intelligence and networked communication technologies, building on IoT foundations to meet modern military warfare demands. These IoBT devices are being relied upon more and more to assist soldiers and are increasingly becoming indispensable tools for supporting soldiers in maintaining clear perception in a complex battlefield environment. Particularly, camera-based IoBT devices need to provide a clear perception of the environment to facilitate agile and precise decisions. When encountered with an object, these devices need to recognize and label that target object (e.g., human, enemy bunker, tanks) and take appropriate actions. However, the restricted angle of view and the computationally intensive nature of image recognition algorithms further limit these resource-bounded devices for timely object labeling. To address this, we propose a distributed approach by utilizing a group of geo-distributed camera-based IoBT devices in the battlefield environment to rapidly and accurately identify and label objects. We propose a Collaborative Object Labeling Approach, called COLA, that accumulates the opinions of several IoBT devices to label each object. Our proposed distributed approach is based on committee formation and reaching accurate verdicts in real-time. We perform extensive experiments to analyze the performance of our proposed algorithm based on several metrics. These experiments demonstrate COLA’s potential to significantly enhance operational efficiency in dynamic and distributed battlefield environments.

Collections of autonomously behaving systems, or swarms, are predicted to be an important component of the US DoD strategy. Therefore, research into how to create swarms with suitable characteristics, behaviors, and function for these different purposes is in the interest of the US military. However, there are challenges in swarm research, including technical limitations of existing hardware, the need to address both individual drone level behavior as well as the complexities of the entire swarm behavior, and the parameter combinatorics that may be relevant to swarm performance in operations.

This presentation proposes methodologies for the computer simulation research and analyses for experimentation on swarm behavior. Swarm performance data from computer simulation experimentations using simulation software were analyzed through multiple steps to investigate how individual and entire swarm characteristics might affect how well the swarm performed a mission.

Within the landscape of military technology, the time has come for automated systems to assume missioncritical responsibilities demanding computational speeds that are beyond human capabilities. Swarm robotics exhibits a diversity of applications in many fields including surveillance, reconnaissance, and more. Regardless of the specific application, meticulous planning and optimization of traversal routes coupled with seamless communication among individual robots assumes a pivotal role. Optimized cooperative route planning enables heightened operational efficiency, yielding a reduction in operational costs while elevating overall productivity. The significance of optimization extends to each robotic ensemble and an understanding of the intricacies of route selection fosters enhanced collaborative synergy. Consequently, the swarm attains a heightened capability to undertake intricate and multifaceted missions, transcending the limits of individual capabilities. Traditional approaches have restricted the potential of swarm-based systems. There has been little focus on cooperative route planning for swarms, essentially eliminating the advantages that swarm systems provide. We propose to rectify this imbalance by considering and addressing frequently overlooked variables of dynamic obstacles and cooperative route planning. Our approach leverages the fusion of advanced machine learning techniques and the reinforcement of communication channels among the constituent agents within the swarm. The deployment of efficient machine learning models facilitates real-time adaptation to the ever-changing optimal path, effectively mitigating the negative impact of dynamic obstacles in the mission environment. Furthermore, by reinforcing the communication infrastructure within the swarm, our approach fosters a heightened sense of synergetic collaboration among swarm agents. This approach unleashes the swarm robotics’ full potential by amplifying its efficiency and expanding the scope of possible applications, ushering in a new era of versatile and high-performing swarm robotic solutions.

Reinforcement Learning (RL) has become a widely used approach for pursuit-evasion games. However, the behavior of such RL models is hard to analyze, often leading to a lack of trust. This paper describes a study in which we used machine learning (ML) approaches to develop metareasoning policies that control pursuers’ strategies. The proposed approach enables pursuer agents to capture a faster evader by choosing simple pursuit strategies collaboratively. The results show that some metareasoning policies perform better than any pursuer strategy combinations. Our approach provides an innovative way for the pursuer agents to reason about their opponents and adapt their strategy, which could have significant implications for the design of intelligent agents in real-world applications.

Virtual Reality (VR) has made significant strides, offering users a multitude of ways to interact with virtual environments. Each sensory modality in VR provides distinct inputs and interactions, enhancing the user's immersion and presence. However, the potential of additional sensory modalities, such as haptic feedback and 360° locomotion, to improve decision-making performance has not been thoroughly investigated. This study addresses this gap by evaluating the impact of a haptic feedback, 360° locomotion-integrated VR framework and longitudinal, heterogeneous training on decision-making performance in a complex search-and-shoot simulation. The study involved 32 participants from a defence simulation base in India, who were randomly divided into two groups: experimental (haptic feedback, 360° locomotion-integrated VR framework with longitudinal, heterogeneous training) and placebo control (longitudinal, heterogeneous VR training without extrasensory modalities). The experiment lasted 10 days. On Day 1, all subjects executed a search-and-shoot simulation closely replicating the elements/situations in the real world. From Day 2 to Day 9, the subjects underwent heterogeneous training, imparted by the design of various complexity levels in the simulation using changes in behavioral attributes/artificial intelligence of the enemies. On Day 10, they repeated the search-and-shoot simulation executed on Day 1. The results showed that the experimental group experienced a gradual increase in presence, immersion, and engagement compared to the placebo control group. However, there was no significant difference in decision-making performance between the two groups on day 10. We intend to use these findings to design multisensory VR training frameworks that enhance engagement levels and decision-making performance.

Robot manipulation has been on the rise in the past decade. This has yielded tremendous growth in low-cost commercial off-the-shelf robot arms. Employing such arms on the ground, aerial vehicles, and space robots is called mobile manipulation, and it is an active research topic in robotics. The U.S. Department of Defense (DoD) has been interested in using mobile manipulation for military applications like intelligence, surveillance, and reconnaissance (ISR) missions. The idea is to replace soldiers with mobile manipulating robots to complete essential tasks in hazardous areas like unexploded ordnance (UXO). One of the challenges in mobile manipulation is to complete the desired tasks beyond the operator’s line of sight. Immersive technologies like VR/AR could address this challenge. VR/AR provides immersive situational awareness to the operator in the remote site. It also maps human body motions to the robot arm. Integrating such technologies enables the operator to leverage his experience to complete manipulation tasks through mobile robots, called telemanipulation. This paper presents hardware/software design and experiment results of telemanipulation.

Digital augmentation of direct view optics like riflescopes allows viewing of the scene with near-zero latency while simultaneously offering increased situational awareness from fielded sensors and systems data. Existing solutions to digitally augment a direct view optic add a considerable amount of weight/size, require the digital projection system to be designed into each specific viewing optic, or interfere with operational function of the optic. Here we present an externally mounted, compact optical waveguide capable of displaying information such as ballistic data overlaid onto the field-of-view of a riflescope. Volume phase holograms form the input/output couplers and the pupil expander of the waveguide. The holographic waveguide satisfies the contrast ratio needed for daylight bright environments, allows for operability of the diopter adjustment, minimally impacts eye relief, and requires no refocusing to view the symbology.

The information presented by augmented reality (AR) systems may not appear exactly like information perceived in the “real” world. Shortfalls in AR technology give rise to anomalies like tracker errors, lack of opacity compared to the background, and reduced field of view (FOV) compared to the human visual system’s field of view. These anomalies can make users feel like the information presented in an AR training system is not believable. This lack of belief can lead to negative training, where trainees adjust how they train due to flaws in the training system and are therefore less prepared for actual battlefield situations. We have completed an experiment to investigate trust, reliance, and human task performance in an augmented reality three-dimensional experimental scenario. Specifically, we used a methodology in which simulated real (complex) entities were supplemented by abstract (basic) cues presented as overlays in an AR head-worn display (HWD) for a visual search and awareness task. We simulated shortfalls of different AR display systems to determine which of the properties most affect training efficacy. Results from our experiment will feed directly into the design of training systems that use AR displays and will help increase the efficacy of training.

MITRE Adversarial Threat Landscape for AI Systems (ATLAS™) is a globally accessible, living knowledge base of adversary tactics and techniques based on real-world attack observations and realistic demonstrations from artificial intelligence (AI) red teams and security groups. There are a growing number of vulnerabilities in AI-enabled systems as the incorporation of AI increases the attack surfaces of existing systems beyond those of traditional cyberattacks. We developed ATLAS to raise community awareness and readiness for these unique threats, vulnerabilities, and risks in the broader AI assurance landscape. Anna will discuss the latest ATLAS community efforts focused on capturing cross community data on real world AI incidents, growing understanding of vulnerabilities that can arise when using open-source models or data, building new open-source tools for threat emulation and AI red teaming, and developing mitigations to defend against AI security threats.

The recent push for fair, trustworthy, and responsible Artificial Intelligence (AI) and Machine Learning (ML) systems have pushed for more explainable systems that are capable of explaining their predictions/decisions and inner workings. This led to the field of Explainable AI (XAI) going through an exponential growth in the past few years. XAI has been crucial in making AI/ML systems more comprehensible. However, XAI is limited to the model that it is being applied to, for both post-hoc or transparent models. Even though XAI can explain the decisions being made by the ML systems, these decisions are based on correlation and not causation. For applications such as tumor classification in the medical field, this can have serious consequences as people’s lives are affected. A potential solution for this challenge is the application of causal learning, which goes beyond the limitations of correlation for ML systems. Causal learning can generate analysis based on cause-and-effect relations within the data. This study compares the results of explanations given by post-hoc XAI systems to the causal features derived from causal graphs via causal discover for image datasets. We investigate how well XAI explanations/interpretations are able to identify the pertinent features within images. Causal graphs are generated for image datasets to extract the causal features that have a direct cause- and-effect relation with the label. These features are then compared to the features highlighted by XAI via feature relevance. The addition of causal learning for image datasets can aide in achieving fairness, bias detection, and mitigation to provide a robust and trustworthy system. We highlight the limitation of XAI tools such as LIME to make predictions based on physical features from images, whereas causal discovery can go beyond the simple pixel based perturbations to identify causal relations from image attributes.

With the spread of powerful AI models, human-machine interaction has improved rapidly. In this situation, recognizing facial expressions of human emotion is essential for successful human-machine collaboration/ teaming. While Convolutional Neural Network (CNN) models were widely used for facial emotion classification, Transformer-based models, known for excelling in NLP tasks, have demonstrated superior performance in areas like image classification, semantic segmentation, and object detection. This study investigates the effectiveness of using a transformer based model, the Face-transformer model a powerful tool for facial identification, that will be fine-tuned for the Facial Emotion Recognition (FER) challenge. We aim to modify the face transformer architecture to recognize emotional states from facial photos using the extensive facial emotion recognition datasets, opening the door for more natural and responsive machine interactions. Our initial findings suggest that the Face-Transformer model holds promise for bridging the gap between machine interpretability and human emotions, potentially paving the way for more natural and responsive human-computer interactions.

Artificial reasoning systems via Artificial Intelligence (AI) and Machine Learning (ML) have made tremendous progress within the past decade. AI/ML systems have been able to reach unprecedented new levels of autonomy for a multitude of applications ranging from autonomous vehicles to biomedical imaging. This new level of intelligence and freedom for AI/ML systems requires them to have a degree of human-like intelligence in terms of causation beyond the correlation. This, however, has remained a major challenge for investigators when combining causality with AI/ML systems. AI/ML systems that are capable of generating cause and effect relationships are still in their infancy, as the literature highlights. The lack of investigations for causal reasoning systems that are capable of using datasets other than tabular data is well highlighted within literature. Causal learning for image, audio, video, radio-frequency, and other modalities still remain a major challenge. While there are open-source tools available for causal learning with tabular data, there is a lack of tools for other modalities. To this extent, this study proposes a causal learning method with image datasets by using existing tools and methodologies. Specifically, we propose to use existing causal discovery toolboxes for investigating causal relations within image datasets by converting image datasets into tabular form with feature extraction using tools such as auto-encoders and deep neural networks. The converted dataset can then be used to generate causal graphs by using tools such as the Causal Discovery Toolbox to highlight the specific cause and effect relations within the data. For AI/ML systems using causal learning for image datasets via existing tools and methodologies can provide an extra layer of robustness to ensure fairness and trustworthiness.

Artificial intelligence (AI) is quickly gaining relevance as a transformative technology. Its ability to rapidly fuse and synthesize data, accelerate processes, automate tasks, and augment decision-making has the potential to revolutionize multi-domain warfighting through data-centric operations and algorithmic warfare. As the military relies more on AI-enabled Decision Aids to increase the efficiency and effectiveness of decision-making, it highlights the need to effectively assess them before deployment. Modeling and simulation (M&S) environments are essential for assessing these rapidly evolving AI-enabled systems. Accepted analytical frameworks are needed to guide ways to represent and model AI sufficiently within M&S environments for accurate assessment. In this paper, we identify common characteristics within the main categories of AI and investigate how those characteristics can be best represented across the main categories of M&S. We provide two use cases to highlight an assessment of AI-enabled Decision Aids for cybersecurity and aeromedical evacuation problems. Our example use cases demonstrate how to leverage a framework for analytic assessment of AI within M&S environments.

In the face of increasing complexity and uncertainty in decision-making across various domains, there is a pressing need for frameworks that can adeptly navigate the multifaceted nature of such challenges. This paper introduces a novel, generalized framework designed to enhance multi-criteria decision-making under uncertain conditions. By integrating advanced data processing techniques with dynamic decision plan generation and adaptive mechanisms, the framework offers a robust solution to making informed decisions that are both flexible and responsive to changing information landscapes.

Central to the framework is its methodological rigor, which encompasses a comprehensive approach to data processing, including feature extraction and anomaly detection. This ensures decisions are grounded in accurate and up-to-date information. The framework employs a systematic algorithm for generating initial decision plans, which are then dynamically expanded to include co-requisite actions, ensuring comprehensive coverage of necessary interventions. A distinctive feature of the framework is its adaptive feedback mechanism, which refines decision plans based on the outcomes of implemented actions and new data, facilitating continuous improvement and relevance of decision strategies.

Furthermore, the paper details the application of advanced visualization techniques, such as Sankey diagrams, to elucidate the flow of information and the interdependencies between actions within the decision-making process. This not only enhances the transparency and interpretability of decisions but also fosters stakeholder engagement and consensus.

The versatility of the proposed framework is demonstrated through its applicability to a variety of domains, from emergency response planning to business strategy development, showcasing its potential to significantly improve decision-making practices. Through a meticulous examination of the framework's methodology and its application across diverse scenarios, this paper contributes to the advancement of decision science, oYering a scalable and adaptable solution to the challenges of multi-criteria decision-making under uncertainty.

The CDAO's Responsible AI team focuses on operationalizing the DoD AI Ethical Principles, sustaining the DoD's tactical edge through concrete actions, processes, and tools. This presentation provides a deep dive into a key piece of the DoD’s approach to Responsible AI: the Responsible AI Toolkit. The Toolkit is a voluntary process through which AI projects can identify, track, and mitigate RAI-related issues (and capitalize on RAI-related opportunities for innovation) via the use of tailorable and modular assessments, tools, and artifacts. The Toolkit rests on the twin pillars of the SHIELD Assessment and the Defense AI Guide on Risk (DAGR), which holistically address AI risk. The Toolkit enables risk management, traceability, and assurance of responsible AI practice, development, and use.

Reinforcement learning (RL) agents offer significant value for military applications by effectively navigating complex, dynamic environments typical of mission engineering and operational analysis. Once trained, these agents can be employed to inform mission planners on optimal strategies, tactics, or even innovative utilization of different military platforms within a given scenario. In recent years, RL has become a major research area for automation and solving complex sequential decision-making problems. However, a notable challenge lies in the inherent black-box nature of RL models and their inability to explain their decisions and actions. This limitation serves as a major adoption barrier, especially in Defense. This paper aims to study EXplainable RL (XRL) within an operational context. XRL is a distinct branch of Explainable Artificial Intelligence (XAI) techniques that provides the necessary transparency to make AI models more transparent to address this challenge. This research is an effort to gain insight into the behavior of RL agents in an operational environment and to discuss explainability and interpretability through the lens of different roles within the decision-making pipeline.

Large Language Models act as an all-inclusive ground for engagement. However, the output of these models is not always aligned with the correct human preferences or values. This can produce a large range of inappropriate and offensive content. In human culture/society, proverbs act as a way to transmit and shape human values and behavior. In this research, we experiment to determine whether or not proverbs can play the same role for LLMs in transmitting and aligning to human values. Additionally, we research to determine how we can prompt LLMs to generate their own set of rules for categories of human values and use that as a means to align future-generated content to said values.

The Modeling and Reasoning Artificial Intelligence Course of Action Generation Hub (MRAI Hub) seeks to advance the research and development of artificial intelligence (AI) technologies in support of combat planning by improving sharing and collaboration across the Department of Defense. The MRAI Hub will include a data and information repository, test and evaluation capability, and common operational challenges. This talk will provide a high-level overview of course of action (COA) generation AI, present the MRAI Hub vision and mission, and discuss current status and efforts, including ways to engage.

There is a recognized need for accelerated research and development and transition of AI supporting DoD command and control (C2) to include but not limited to enhanced planning, wargaming, and course of action development to maintain decision dominance on the battlefield. As such the Air Force Research Laboratory Information Directorate (AFRL/RI) has formed the AI Agent Community of Practice (CoP) to rapidly identify, evaluate, and engage the right community members with government stakeholders to address high priority operational problem sets and AI technology challenges. The AI Agent CoP focuses on sequential decisionmaking AI to include technologies such as Reinforcement Learning, Computational Game Theory, Evolutionary Algorithms, Search, Optimization, etc. A key component of the AI Agent CoP is to use digital AI wargaming competitions to spur research and evaluate AI technology for transitioning to operational partners.

In setting up the AI Agent CoP, AFRL/RI held a pilot competition, using a digital hex-based wargame. For the competition, CoP members created blue agents for a small operational scenario including naval, air, and long-range fires. The submitted agents were evaluated against a known red agent, an unknown red agent, and qualitative human evaluators. This paper reports on the pilot competition.

Remaining competitive in future conflicts with technologically-advanced competitors requires us to accelerate the development of robust artificial intelligence (AI) for wargaming. In this study, we contribute to the research of AI methods to develop intelligent agent behaviors in combat simulations by investigating the effectiveness of a multi-model approach as compared to a single-model approach. To accomplish this, we first develop a multi-model framework that leverages supervised learning, reinforcement learning, and scripted behavior models. We then design and conduct an experiment where we compare the performance of each one of eight individual behavior models to four multi-models composed of varying numbers of these individual behavior models. Using a combat simulation with a random scenario generator, we find that a multi-model approach improved the mean score by 62.6% over the mean game score of the best-performing single-model. Additionally, we find that a multi-model with more embedded behavior models outperforms a multi-model with fewer behavior models. The outcomes of this study contribute to the broader research endeavors aimed at expanding AI capabilities to handle the intricate and expansive state-spaces characteristic of combat modeling and simulation.

Legatus AI, an Air Force program supporting joint and coalition partners, is addressing the need for the creation of a common architecture of software and hardware for campaign level planning to empower the efficient development of AI and core AI capabilities. Building off lessons learned in previous command and control programs, Legatus AI seeks to close the capability gap by transforming AI development from non-reusable boutique solutions to an ecosystem of modular and reusable processes for configuring, training, evaluating, and deploying reinforcement learning and game theory agents for planning and wargaming environments. The prototype architecture added reinforcement learning capabilities to StreamlinedML to allow for communication between AI agents and games using a Docker-based solution. The architecture was tested on the Airlift Challenge simulation to ensure intended functionality and demonstrate the initial capability. Future work will incorporate additional agents and Unity-based wargames. The maturation of this architecture is a key cornerstone for future collaborative research and development of adaptable AI and plug-and-play wargaming platforms.

Machine Learning has played a major role in various applications including Visual Slam and themal image process. In this paper, we discussed the possibility of generating a thermal map using LWIR images and a deep learning-based visual slam network and the value that the thermal map can create. We summarized the advantages and applicability of various deep learning-based visual slams and confirmed the results of nice slam, which generates the most curious Dense map. In order to apply Visual SLAM technology, time series, scene repetition, and images from various angles for one scene are required. However, most LWIR data sets consist of one shot for each scene or are unidirectional driving data. To solve this, we created a scenario using the LWIR driving dataset and created a repetitive route through repetition. RGB-Depth SLAM Mapping was performed on the constructed data set, and the results were evaluated and the limitations of the current approach were discussed. Finally, we summarized future directions for creating stable 3D thermal maps in indoor and outdoor environments by resolving the limitations.

The evolution of robots from tools to teammates will require them to derive meaningful information about the world around them, translate knowledge and skill into effective planning and action based on stated goals, and communicate with human partners in a natural way. Recent advances in foundation models, large pre-trained models such as large language models and visual language models, will help enable these capabilities. We describe how we are using open-vocabulary 3D scene graphs based on foundation models to add scene understanding and natural language interaction to our human-robot teaming research. Open-vocabulary scene graphs enable a robot to build and reason about a semantic map of the environment, as well as answer complex queries about it. We are exploring how semantic scene information can be shared with human teammates and inform context-aware decision making and planning to improve task performance and increase autonomy. We highlight human-robot teaming scenarios involving robotic casualty evacuation and stealthy movement through an environment that could benefit from enhanced scene understanding, describe our approach to enabling this enhanced understanding, and present preliminary results using a one-armed quadruped robot interacting with simplified environments. It is anticipated that advanced perception and planning capabilities provided by foundation models will give robots the ability to better understand their environment, share that information with human teammates, and generate novel courses of action.

Military operations invariably involve devices at the edge (e.g. sensors, drones, handsets of soldiers, etc.) In edge environments, good network connectivity cannot be assumed due to Denied, Degraded, Intermittent, or Low-bandwidth (DDIL) conditions. A DDIL environment poses unique challenges for deploying AI applications at the edge, particularly in the execution of Machine Learning Operations (MLOps). In this paper, we present a framework to address these challenges by considering three important dimensions: (i)the ML model lifecycle activities, (ii) specific DDIL induced challenges at the edge and (iii) the application stack. We discuss three realistic use cases in detail to explain the use of this approach to identify the underlying design patterns. We believe that use of this framework can lead to a responsive and reliable AI deployment under varying operational conditions.

In the era of data-intensive edge computing, the orchestration of Data Distributed Inferencing (DDI) tasks poses a formidable challenge, demanding real-time adaptability to varying network conditions and compute resources. This study introduces an innovative approach to address this challenge, leveraging Gradient Boosting Regression (GBR) as the core predictive modeling technique. The primary objective is to estimate inferencing time based on crucial factors, including bandwidth, compute device type, and the number of compute nodes, allowing for dynamic task placement and optimization in a DDI environment. Our model employs an online learning framework, continuously updating itself as new data streams in, enabling it to swiftly adapt to changing conditions and consistently deliver accurate inferencing time predictions. This research marks a significant step forward in enhancing the efficiency and performance of DDI systems, with implications for real-world applications across various domains, including IoT, edge computing, and distributed machine learning.

In tactical edge networks, the volatility of computational and communication resources complicates the consistent processing of data. Previous work has developed a system that allows execution of inference task applications throughout a network using a variety of adaptations to machine learning models that offer accuracy and latency trade-offs as conditions change in order adaptively perform deterministic resource allocation at the tactical edge. This paper expands on previous work that proposed a system that allows execution of inference task applications throughout a network and then developed a resource allocation algorithm that optimally places intelligence, surveillance, and reconnaissance related machine learning tasks throughout the network. We propose utilizing stochastic optimization to analyze the computational time and performance for inference tasks. Instead of adhering to deterministic averages, we use sample average approximation as a solution technique to optimize and analyze the inherent uncertainties of tactical edge environments, optimizing over the distribution of inference data rather than average-case scenarios. This paper verifies Jensen’s inequality gap within deterministic optimization and proposes an improved resource allocation algorithm that optimally places tasks throughout a network under uncertainty. We present initial results on a military relevant tactical edge network scenario.

Machine learning (ML) at the edge typically involves pushing deep neural network (DNN) models ever closer to the sensor. In practice, a DNN deployed to a dynamic environment will quickly become obsolete if it cannot be updated to accommodate new or modified classes. Especially for low-Size, Weight, and Power (SWaP) hardware, the data, hardware, and time requirements for retraining a DNN remain cost prohibitive. Technical challenges include 1) catastrophic forgetting where retraining only on new data overwrites prior knowledge; 2) class imbalance where there exists only a handful of novel class samples compared to the thousands of training samples; and 3) high energy costs required to run the backpropagation retraining for hours on high-end GPUs.

In this work, we evaluated the Constrained Few-Shot Class Incremental Learning (C-FSCIL) framework for sequentially learning the CIFAR100 dataset. The C-FSCIL framework modularizes the layers of an arbitrary DNN into 1) a frozen, pre-trained feature extractor, 2) a retrainable fully connected layer, and 3) an explicit prototype vector memory matrix. We investigated the effects of different task-agnostic feature extractors trained via fully supervised, weakly supervised, and self-supervised training. Using a CLIP-trained ConvNeXt-L for the frozen feature extractor, our C-FSCIL implementation sequentially learned 40 additional classes over the base session of 60 classes, with a final accuracy of 79.9% over the 100 classes, sacrificing only 7.2% points of accuracy from the base session.

With the current growing availability of datasets coming from multiple sources and domains, systems onboard our military assets have an immediate need of being functional in multiple domains, handling large amounts of data, and implementing fast and appropriate analyses for these datasets. However, onboard systems often have very limited computational resources upon which to process these tasks. Generalized additive models (GAMs), which are statistical model that are better able to account for non-linear relationships between independent and dependent variables, have been shown in multiple application areas to yield more accurate estimates than models that only assume linear relationships. GAMs require fitting smooth functions of independent variables, and potentially interactions between two or more independent variables with tensor products, commonly using penalized basis splines (P-splines). However, when trying to make predictions for previously unseen feature vectors and without the benefit of significant computing power and necessary software packages, computing the necessary P-spline basis values in order to make predictions becomes computationally intractable. We show that approximating the smooth functions of a trained GAM with cubic splines yields very good estimates that result in minimal loss of accurate compared to the original trained GAM model while significantly decreasing computational cost. These two benefits have resulted in our method being adopted by a primary contractor handling high-fidelity analyses. We will show our implementation results on simulated and real datasets, discuss our findings, and conclude with remarks on future work.

The DARPA Robotic Autonomy in Complex Environments with Resiliency (RACER) program is focusing on developing highspeed offroad ground autonomy for military applications. The presentation will discuss the progress made to advance the state of the art in this domain and address the challenges that remain.

Out of distribution (OOD) detection has shown immense promise to enable Automatic Target Recognition models for defense applications. However, many defense applications have constraints that make current best practices for training OOD detection models challenging. These include: the need to perform fine-grained classification of identified targets, low amounts of labeled data to train models, limited availability of Subject Matter Experts to accurately label new data, and the potential need to incorporate new classes of targets as they are discovered. Given these constraints, we propose to build a fine-grained classifier with robustness against OOD data through an active learning approach - designed to further classify objects after detection through some coarse-grained object detection model. This paper will explore active learning methods for Automatic Target Recognition applications, with experiments conducted using the fine-grained overhead imagery dataset, ShipsRSImageNet, along with samples from the DOTA dataset as an exposure set. Our contributions will include recommendations to achieve fine-grained Automatic Target Recognition with robustness against OOD data with minimal labeling from Subject Matter Experts.

Deep learning-based approaches, such as Convolutional Neural Nets (CNNs), have shown high performance in classifying contents of images. CNNs, however, have the notable drawbacks of potentially high computing costs, poor explainability, and wide performance variance if the underlying imagery data deviates from the training baseline. As advanced image processing capabilities are matured, the on-board detection of objects in space-based imagery is increasingly proposed. On-board satellite processing applications, which may be resource-limited, can drive the need for simpler models that reduce the necessary computing burden for edge computing applications. This raises the question of how well classic computer vision techniques can compete with more modern approaches. This paper characterizes and compares the performance of multiple computer vision models for the application of distinguishing maritime vessels from typical clutter in commercial electro-optical (EO) satellite imagery. A Support Vector Machine (SVM) model using manually curated features is compared to multiple DL-based models spanning a range of model sizes, with the goal of determining whether classical approaches can compete favorably with DL when computational resources are taken into consideration. Differences in performance and processing resources are characterized between the approaches. Findings include that the SVM-based model may approach the accuracy of some CNN-based models for classifying images of clouds in satellite EO imagery for smaller DL-based models. However, even the smallest DL-based models, which take about the same computational resources as the SVM-based model, generally out-perform the SVM-based model. This finding may have implications for the operational use of on-board processing techniques for satellite payloads.

In recent years, computer vision research has witnessed transformative changes with the integration of generative artificial intelligence (AI) models. The generative models have been widely researched in the field of semantic segmentation. In this survey paper, we present a comprehensive review of the generative models, with a specific focus on Generative Adversarial Networks (GANs), Diffusion Models (DMs), and Variational Autoencoders (VAEs), in the realm of semantic segmentation. We incorporate these generative models for model training, image synthesis, semantic label synthesis, image-label pair synthesis, domain adaptation, feature learning, and boundary localization for semantic segmentation. We also perform a thorough comparative analysis highlighting the approach, task, datasets involved, strengths, and weaknesses of the GANs, DMs, and VAEs-based semantic segmentation models. Our comparative evaluation showed a wide range of research works carried out in the generative semantic segmentation domain. This survey consists of diverse generative methodologies, serving as a comprehensive resource for researchers and enthusiasts contributing to the field of generative semantic segmentation.

We present a method for maritime platform defense using constrained deep reinforcement learning (DRL), showing how competing desires to reliably defend a fleet and conserve inventory may be managed through a dual optimization strategy. Against persistent and variable raids of threats, our agents minimize inventory expenditure subject to a constraint on the average time before a threat impacts the fleet being defended. Critically, the additional inventory consideration is introduced only after the agent has learned to defend the fleet well enough to consistently satisfy the constraint. In evaluations against a realistic simulation environment and with variable multi-ship geometries, we find that our strategy may be tuned to either (1) enable the agent to make significant gains in efficiency while losing very little in terms of reliability or (2) closely track specified reliability constraints while reducing inventory expenditure even further. The result is an agent with considerably stronger long-term viability, since the conserved inventory may be used for future engagements. We speculate on the potential of this method to provide a tunable, trustworthy artificial assistant to human decision-makers tasked with defense scheduling.

Flight planning in support of multi-domain operations is a complex and manpower intensive process. Multiple factors such as changing environmental conditions, navigational data and constraints, air-traffic procedures and policies, maturation of routes and coordination with all stakeholders contribute to this complexity. These changing constraints result in multiple updates and deviations to the flight plans which eventually impact mission objectives and timing. This work applies state-of-the-art machine learning technologies to automatically determine the quality of generated flight plans to enable rapid verification and approval for plan filing. It also identifies preferred routes that are input into planner that results in higher flight plan acceptance rates. We leverage voluminous sources of real operational flight planning data and filed planning data to develop the models. Supervised learning is leveraged to predict whether a generated plan will be filed or rejected. The unsupervised learning component identifies flight plan preferences for feedback to the planner system. These models could be eventually deployed in a flight planning operational environment to reduce human effort, cost, and time to generate flyable plans. The results from this work could also be used to improve planner rules and search algorithms.

The responsible and trustworthy use of AI is critical to the safe and secure processing and understanding of the large volumes of data collected everyday by the Department of Homeland Security. For example, the Transportation Security Administration takes five million images a day and screens millions of air passengers, while U.S. Customs and Border Protection processes more than 1 million images per day and more than 1 billion packages per year in achieving their mission to protect the U.S. homeland. In April of 2023, the Department of Homeland (DHS) Secretary, Alejandro Mayorkas, established the department’s first ever artificial intelligence task force (AITF). The guiding mission of the AITF is to advance the use of AI to support critical homeland security missions. This presentation will focus on the findings and results of the DHS AITF including policies for the responsible use of AI within the department, the selection and initiation of AI pilot projects, and the development of a framework for the successful implementation of AI within DHS. Finally, multiple DHS inspired problems will be presented to introduce the breadth of the DHS missions that will be impacted through the use of AI.

Combined, joint, intra-governmental, inter-agency and multinational (CJIIM) operations require rapid data sharing without the bottlenecks of metadata curation and alignment. Curation and alignment is particularly infeasible for external open source information (OSINF), e.g., social media, which has become increasingly valuable in understanding unfolding situations. Large language models (transformers) facilitate semantic data and metadata alignment but are inefficient in CJIIM settings characterised as denied, degraded, intermittent and low bandwidth (DDIL). Vector symbolic architectures (VSA) support semantic information processing using highly compact binary vectors, typically 1-10k bits, suitable in a DDIL setting. We demonstrate a novel integration of transformer models with VSA, combining the power of the former for semantic matching with the compactness and representational structure of the latter. The approach is illustrated via a proof-of-concept OSINF data discovery portal that allows partners in a CJIIM operation to share data sources with minimal metadata curation and low communications bandwidth. This work was carried out as a bridge between previous low technology readiness level (TRL) research and future higher-TRL technology demonstration and deployment.

In order to achieve a comprehensive situational awareness suite, there is a need to standardize how we access and manipulate information streams. These information streams are often singular, reducing information to what is relevant for a targeted sensor context, which limits their ability to allow multiple autonomous platforms to perceive based on the total available information. To enable distributed situational awareness, here we present an approach that standardizes the management of information streams across contexts of varying sensor domains. These information streams enable individual platforms to dynamically query any relevant information where it will co-register the sensor streams and be ready for perception systems to rationalize the information. This system is enabled through the Robot Operating System 2 (ROS 2), which provides a distributed publisher-subscriber model for communication in addition to standardization for autonomous systems, such as positional referencing, sensor processing, and dynamic mapping. Based on ROS 2’s system components, we show a system that enables sensor information to retain its generalized form, which other platforms can manipulate for their targeted need without the unnecessary destruction of data to perceive within its system. The development of this cross-platform approach to sensor co-registration is a major step toward our goal of distributed situational awareness and will enable significant research efforts in the future.

This paper details the lessons learned from designing a specialized competition focused on network attack simulation, specifically emphasizing lateral movement in networks. Our experience in planning the competition reveals several critical design principles that could guide the development of future competitions. Among these principles are the need for a carefully curated simulation environment that mimics real-world network behaviors and vulnerabilities without neglecting the limitations of current capabilities, the necessity for defining a clear scoring system that accurately reflects the effectiveness of different strategies while allowing for interpretability and style, and the importance of ensuring that the competition has clear roles for participants with varied levels of expertise in both cybersecurity and reinforcement learning. This paper offers lessons learned for those interested in organizing similar competitions or developing benchmarks in the area of cyber reinforcement learning. We strongly advocate for the use of network attack simulation competitions as an effective means to accelerate research, foster community collaboration, and identify emergent strategies and techniques. By outlining our design process, the challenges faced, and the solutions implemented, we provide actionable insights to assist practitioners, researchers, and government sponsors.

We consider a scenario in multi-domain operations where the users in one domain need to perform searches on information hosted by a provider in another domain. It is very common in many scenarios that information cannot be shared openly across different domains, and users may want to obfuscate their searches and prevent the search provider from learning the intent of their searches. In scenarios where search privacy is important, the use of a large language model can help implement an obfuscation approach relying on generation of decoy queries to obfuscate the real query. In this paper, we consider different alternative approaches to use large language models for search privacy, compare their strengths and weaknesses, and discuss their effectiveness.

In this work, we demonstrate the potential of dynamic reinforcement learning (RL) methods to revolutionize cybersecurity. The RL framework we develop is shown to be capable of shutting down an aggressive botnet, which initially uses spear phishing to establish itself in a Department of Defense (DoD) network. To ensure a suitable real-time response, we employ CP, a transformer model trained for network anomaly detection, to factorize the state space accessible to our RL agent. As the fidelity of our cyber scenario is of the utmost importance for meaningful RL training, we leverage the CyberVAN emulation environment to model an appropriate DoD enterprise network to attack and defend. Our work represents an important step towards harnessing the power of RL to automate general and fully-realistic Defensive Cyber Operations (DCOs).

Deep learning (DL) has revolutionized machine learning tasks in various domains, but conventional DL methods often demand substantial amounts of labeled data. Semi-supervised learning (SSL) provides an effective solution by incorporating unlabeled data, offering significant advantages in terms of cost and data accessibility. While DL has shown promise with its integration as a component of modern network intrusion detection systems (NIDS), the majority of research in this field focuses on fully supervised learning. However, more recent SSL algorithms leveraging data augmentations do not perform optimally “out of the box” due to the absence of suitable augmentation schemes for packet-level network traffic data. Through the introduction of a novel data augmentation scheme tailored to packet-level network traffic datasets, this paper presents a comprehensive analysis of multiple SSL algorithms for multi-class network traffic detection in a few-shot learning scenario. We find that even relatively simple approaches like vanilla pseudo-labeling can achieve an F1-Score that is within 5% of fully supervised learning methods while utilizing less than 2% of the labeled data.

In recognizing the importance of network traffic monitoring for cybersecurity, it is essential to acknowledge that most traditional machine learning models integrated in network intrusion detection systems encounter difficulty in training because acquiring labeled data involves an expensive and time-consuming process. This triggers an in-depth analysis into zero-shot learning techniques specifically designed for raw network traffic detection. Our innovative approach uses clustering combined with the instance-based method for zero-shot learning, enabling classification of network traffic without explicit training on labeled attack data and produces pseudo-labels for unlabeled data. This approach enables the development of accurate models with minimal limited labeled data for making network security more adaptable. Extensive computational experimentation is performed to evaluate our zero-shot learning approach using a real-world network traffic detection dataset. Finally, we offer insights into state-of-art developments and guiding efforts to enhance network security against ever-evolving cyber threats.

In this article, the power of physics-informed neural networks is employed to address the issue of model identification for complex physical systems, focusing on the application of a magnetorheological (MR) damper setup. The research leverages the Bouc-Wen hysteresis model, a well-established representation of nonlinear behavior in MR dampers, to inform the training process of a series of cascaded neural networks. The core objective of this research is to develop a surrogate model capable of accurately predicting the dynamic behavior of MR dampers under various operational conditions. Traditionally, MR dampers pose significant modelling challenges due to their nonlinear and hysteresis-rich characteristics. The approach explored in this article combines physics-based insights with the capabilities of neural networks to resolve the complexity associated with the modelling process. The methodology involves the formulation of a physics-informed loss function, which embeds the Bouc-Wen hysteresis model’s governing equations into the training process of the neural networks. This fusion of physical principles and machine learning enables the networks to inherently capture the underlying physics, resulting in a more accurate and interpretable surrogate model. Through experimentation, the effectiveness of the physics-informed neural network approach in surrogate modeling for MR dampers is demonstrated. The model developed exhibits decent predictive performance across a range of input parameters and excitation conditions, offering a promising alternative to conventional black-box machine learning and physics-based methods. Furthermore, this research showcases the potential for physics-informed machine learning in modelling complex physical systems, offering a perspective on the utility of this approach in other engineering and scientific domains. The application of this methodology further facilitates improved control and optimization strategies in various engineering applications.

As a result of the increased use of machine learning (ML) applications in current society, and multi-domain operations in the foreseeable future, there is a surge in adversarial machine learning (AML). One of the domains where AML can prove to be a critical problem is cyber security; the use of AI in security software creates new attack vectors for adversaries. One example of such an attack vector is the use of adversarial domain generation algorithms (DGAs). These adversarial DGAs claim to generate malicious domains that successfully evade deep learning-based DGA detectors. We test two state-of-the-art DGA detectors that make use of deep learning: DGA Detective¹ and B-ResNet,² against four different adversarial DGAs. The tested DGAs all use different adversarial techniques and provide a fitting reflection of the types of DGAs present in literature. Additionally, they can be implemented by adversaries with basic programming and AI knowledge. We find that both DGA detectors reach near-perfect performance on real malware domains, but see a dramatic decline in performance on adversarially generated domains. To counteract the adversarial DGAs, we test two methods to improve adversarial robustness of the detectors: adversarial training and residual loss. Adversarial training results in a ~12%-20% average increase in accuracy on a data set comprised of benign domains, existing malware domains and adversarially generated domains, for both detectors. The residual loss improves average detection accuracy on the same data set with ~4% for DGA Detective, but makes for a ~5% decline in average accuracy for B-ResNet.

The automotive industry must implement some of the best cybersecurity practices to protect electronic systems, communication networks, control algorithms, software, users, and data from malicious attacks. An artificial intelligence (AI) & machine learning (ML) based automotive cybersecurity system could help identify potential vulnerabilities in electronic vehicles as they are part of the vehicular cyber physical systems (CPS). The primary reason that cybersecurity challenges in the automotive domain differ significantly from those in other sectors are threats to computer systems within vehicular CPS can cause direct harm to the drivers of the vehicles. There is a pressing need for better understanding of various attacks and defensive approaches for vehicular CPS to better protect them against any potential threats. This study investigates AI/ML attacks and defenses for vehicular CPS systems to extract a better comprehension of how cybersecurity affects computational components within the vehicular CPS, what the standards are & how they differ, the types of prominent attacks against these systems, and finally an overview of defensive approaches for these attacks. We provide a comprehensive overview of the attacks and defensive techniques/methodologies against vehicular cyberphysical systems.

Physics-informed neural networks (PINNs) are a recently-developed scientific machine learning (SciML) approach useful for predicting behavior of systems governed by differential equations (DEs). Compared to classical methods like finite element method (FEM), PINNs can be easily set up and trained on general DEs and geometries. In this work, we will discuss uses of PINNs in different scientific domains. Our focus will be on the use of pinn-jax, an open-source library we have designed to enable easy development and training of PINNs on varied problems, including forward prediction and inverse estimation. We have designed pinn-jax to be easily extensible while also featuring implementations of some common techniques for enhancing PINNs, and we will demonstrate these on different problems. Particular attention will be paid to evaluating PINNs’ performance on problems that vary in behavior across different temporal and spatial scales.

Machine learning is a pervasive technique in contemporary applications, representing a subfield of artificial intelligence dedicated to machines emulating human behaviors. Neural networks, a prominent class of machine learning models, excel in decision-making tasks. Nevertheless, the empirical nature of designing a neural network structure poses challenges, with practitioners often facing the dilemma of incorporating excessive neurons, leading to prolonged training times, or insufficient neurons, resulting in training failures. This paper presents a solution by introducing a method that recommends an appropriate range of neuron numbers for a neural network, leveraging clustering methods to enhance structural design efficiency.

Sensor data from multiple sources provides valuable insights for data-driven decision making. Sensor fusion is the process of combining data from multiple sensors to produce a more complete view than would be possible with a single sensor alone. To implement sensor fusion at the tactical edge, heterogeneous data sources are often combined. Various types of sensors, such as cameras, lidar, radar, magnetic, etc. are frequently blended for optimal decision making. Multidomain operation (MDO) brings a unique set of challenges to the subdomain of sensor fusion. MDO software architects must consider such challenges as the need to operate in a resource-constrained, fast-paced, and contested environment. Software architects must focus on such design features as high data velocity, high data volume, scalability, and real-time processing. This work offers a blueprint for benchmarking a multi-input video processing pipeline in search of a service degradation point. We throttle the number of input video feeds while quantitatively measuring the pipeline’s performance. We employ an established Nvidia DeepStream SDK framework used for building commercial computer vision and video analytics applications. We begin with a double input application performing object detection, semantic segmentation, and a tracker. We continue scaling up the number of video inputs meant to simulate pipeline load variability frequently present in the field. While objectively measuring the pipeline’s execution parameters, we simulate the point of channel saturation and demonstrate the degradation of the quality of service. We postulate that it may be logical to apply such a benchmarking template in order to establish the range of limits and predict the anticipated pipeline processing times. We conclude with some of the potential future developments of the proposed template.

In the U.S. Army the variables of the Operational Environment (OE) defined as Political, Military, Economic, Social, Information, Infrastructure, Physical Environment, and Time (PMESII-PT) are used for mission analysis and course of action development. In this research, we discuss how to modify existing simulators in order to add non-kinetic operational variables and investigate how they may shape and influence mission outcomes. We use the StarCraft II (SC2) Learning Environment (LE) which provides an interface for Artificial Intelligent agents to control game entities, gather observations, and adjust actions algorithmically. We develop a military-relevant scenario in SC2LE that will be perturbed by the emergence of non-kinetic challenges that affect the simulation outcome. Finally, we investigate reward functions that leverage the OE context to discover an optimal course of action. By integrating PMESII-PT variables into simulators, this research significantly enhances the realism of military simulations, facilitating the development of algorithms that improve operational planning and decision-making for real-world scenarios.

Robots moving in formations can serve as the core functionality of coordinated maneuvers, which have real-world applications for logistics-based tasks, mobile asset protection, environmental monitoring, or wide-area surveillance. In order for autonomous multi-robot systems to be useful in such applications, they must be able to seamlessly execute actions, strategies, and behaviors in a coordinated manner. An example of coordinated behavior in multi-robot systems is the movement of robots across terrain in various geometric formations. The general goal of this work is to leverage formations inspired by military mobility and maneuver to guide the behaviors of autonomous multi-robot systems. Generally, our approach aims to develop coordinated multi-robot behaviors by utilizing domain knowledge to guide multi-agent systems. In this work, we measure a global-planning based approach for control of a multi-robot system that is tasked with executing specific geometric formations in a Robot Operating System (ROS) based photo-realistic simulation environment (ARL Unity ROS Simulator (AURS)). This work highlights a quantitative assessment of global-planning based approaches for executing simple multi-robot coordinated behavior. Finally, the results of these experiments establish a baseline by providing metrics to evaluate a set of global navigation planners using a planning only approach to multi-agent coordinated maneuver.

This paper presents a thorough investigation into the convergence of Particle Swarm Optimization (PSO) and Multi-Disciplinary Design Optimization (MDO), two pivotal methodologies in the realm of computational optimization. By harnessing the strengths of PSO's heuristic search capabilities and MDO's integrative design approach, this study explores the synergistic potential of combining these methods to tackle complex optimization challenges. Through a systematic literature review and bibliometric analysis, we delve into the evolution, methodologies, applications, and outcomes of this interdisciplinary integration, drawing from a diverse array of scholarly works. Our analysis reveals a growing trend in the application of PSO within MDO frameworks, highlighting significant advancements, identifying gaps in the current literature, and suggesting fruitful directions for future research. The findings underscore the robustness and adaptability of PSO-MDO integration across various domains, offering insights into its potential to enhance optimization practices and contribute to the advancement of engineering and technology. This study not only charts the current landscape of PSO and MDO convergence but also sets the groundwork for future explorations in this promising research domain.

This study delves into a comprehensive comparison between Fuzzy Logic (FL) and Artificial Neural Networks (ANN) in the context of the Inverse Depletion problem. Both methodologies, recognized for their distinct capabilities in handling complex problems, are assessed for their efficacy, accuracy, and computational efficiency. Initial observations highlighted the inherent flexibility of FL in managing uncertainty and the adaptive nature of ANN in recognizing patterns from intricate datasets. A series of benchmark scenarios were established to gauge the performance of each model. Results indicate that while FL offers more interpretable solutions, ANNs often outpace in terms of prediction accuracy. However, the choice between the two largely hinges on the specific requirements of the problem at hand, including the available data quality and the desired output precision. This research underscores the importance of understanding the nuances of each method and provides insights to practitioners on selecting the optimal approach for tackling the Inverse Depletion problem in the field of nuclear forensics.

This Chaos theory has long been a fascinating realm of study, offering insights into systems characterized by sensitivity to initial conditions and complex, unpredictable behaviour. Among these intriguing systems, the "Capsule-Shaped Equilibrium Curve Chaotic System" stands out due to its distinctive and intricate dynamics. In this paper, we present a novel approach to understanding and predicting the behaviour of this complex chaotic system through the application of Recurrent Neural Networks (RNNs). Our investigation begins with a thorough examination of the capsule-shaped equilibrium curve chaotic system, revealing its underlying principles and revealing its chaotic nature. We employ the scalability of neural networks to propose an innovative approach for predicting the temporal progression of this system. A promising avenue for modeling the dynamic behavior of chaotic systems with a high degree of precision is offered by the RNN framework, which is capable of capturing temporal dependencies. We delve into the details of our prediction methodology, including data preprocessing, network architecture, and training strategies tailored to the unique characteristics of the capsule-shaped equilibrium curve chaotic system. We conduct extensive experiments and provide quantitative evaluations of prediction precision and dependability to evaluate the predictive capabilities of our neural network-based strategy. Furthermore, this investigation contributes to the comprehension of chaotic systems and opens the door to practical applications in diverse fields, such as physics, engineering, and finance, where precise predictions of chaotic dynamics are essential for making decisions and controlling systems

This article addresses the challenge of image restoration using generative adversarial networks (GANs) specifically tailored for images that include undesired objects. This task becomes crucial in scenarios where there is a need to eliminate incidental elements like random pedestrians or obstructive objects such as text, symbols, or drawings that hinder the main content's clarity. These objects can sometimes entirely obscure critical parts of the image, leading to a distortion of the intended information. In this study, we introduce an innovative approach to image reconstruction by leveraging a generative adversarial network (GAN) architecture enhanced with a two-path discriminator for distinct texture and color analysis. Our model, which integrates the stability advantages of Wasserstein GANs, effectively addresses common GAN challenges like mode collapse and training instability. The sophisticated design of our generator and discriminator results in superior image quality for reconstruction tasks, surpassing the performance of traditional single-path GANs in terms of accuracy and visual fidelity.

The growing sector of Extended Reality (XR), comprising Virtual Reality (VR), Augmented Reality (AR), and Mixed Reality (MR), requires successful approaches for localization and mapping. Simultaneous Localization and Mapping (SLAM) has gained more importance in XR uses. This research conducts an in-depth bibliometric examination of the utilization of SLAM in XR settings, analyzing publications from the last ten years. We investigate the developing trends, fields of knowledge, prominent contributors, and important publications that have influenced this intersection. Our study indicates a notable increase in research on incorporating SLAM in XR, especially in fields such as gaming, architectural visualization, and remote collaboration. A study of the region highlights North America and the Asia-Pacific as key players in this field of research. The research also discusses the difficulties involved in integrating SLAM into XR, such as handling the computational requirements, guaranteeing real-time processing, and accomplishing successful sensor fusion. This paper strives to provide a thorough overview of the current status of SLAM in XR, pinpointing gaps that already exist and suggesting directions for further research. These observations aim to help researchers, developers, and industry professionals who want to use SLAM's capabilities in the quickly changing XR field.