United States

Accurate recognition and tracking of human instances captured at a distance in various poses, particularly those in distress during chaotic events, is crucial for various applications, including disaster relief, damage assessment, and rapid distress-aware triage in battlefield scenarios. These applications have been increasingly empowered by the recent evolution of AI technologies integrated into unmanned platforms, such as UGVs and UAVs. Particularly, identification of injured or distressed individuals via pose analysis is a crucial component of these applications in diverse real-world environments. However, pose identification on video feeds from unmanned systems faces significant challenges due to adverse conditions, such as platform vibration-induced noise, low camera resolution, and camera pose uncertainty. Additionally, data scarcity in the human pose domain also becomes a major bottleneck to achieving successful pose identification. To tackle the above adverse conditions, the Army-DTRA collaborative project called Archangel currently aims to advance the science behind pose/casualty identification by leveraging the multimodal LLM (MLLM)-based scalable framework for automatically generating pose transition descriptions for Composed Pose Retrieval (CPR) benchmarks. Archangel is a customer-reimbursed program conducted by the DEVCOM Army Research Laboratory (ARL) and sponsored by the Defense Threat Reduction Agency (DTRA). Archangel aims to develop advanced AI models integrated into small platforms to recognize human instances in distress primarily for rapid distress-aware triage in chaotic events. Archangel started in early 2020 and is divided into two phases: Phase 1 and 2. Phase 1 ended in 2023 and the deliverables include enhanced machine learning models for detecting human instances in various poses captured by UAVs flying at relatively high altitudes and long ranges. Archangel is currently in Phase 2, focusing on identifying individuals in distress in the context of disastrous events or military conflicts to prioritize the search and rescue of victims or Warfighters in severe physiological conditions. In this workshop, we will introduce our pose identification framework called AutoComPose, the first MLLM-based framework for automatically generating pose transition descriptions that eliminates the need for costly human annotations and enhancing CPR accessibility. AutoComPose is dataset-agnostic, integrating seamlessly with any human image dataset. In this work, we constructed two benchmarks AIST-CPR and PoseFixCPR, as new testbeds for CPR evaluation. Beyond CPR, we will also demonstrate MLLMs’ potential for structured motion understanding, reducing annotation costs while improving scalability. Finally, we will highlight how our MLLM-based pose retrieval establishes the foundation for detecting casualties crucial for distress-aware triage.

NAML 2026

Human Pose Identification Using Multimodal LLMs for Rapid Distress-Aware Triage

casualty detection

human pose identification

multimodal llms

computer vision

### Welcome to the 8th annual workshop on Naval Applications of Machine Learning!

NAML is the premier annual event showcasing current machine learning research relevant to Naval applications. We welcome participation from the defense community, industry, and academia. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team.

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications

* To connect machine learning researchers with experts in military requirements

* To build and strengthen collaborations within the defense research community

Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.

**General Chairs:**

Katie Rainey, Naval Information Warfare Center Pacific

Nicole Isoda, Naval Information Warfare Center Pacific 

**[More about NAML 2024](https://sites.google.com/go.spawar.navy.mil/naml/)**

Please log in to explore this event. You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this workshop.

Registration for this workshop has now closed. We appreciate your interest and hope to welcome you at a future event. If you’d like to hear about similar upcoming events, please check our website or join our mailing list.

Registration for this workshop has now closed.

NAML 2024

The 8th annual workshop on Naval Applications of Machine Learning is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. The goals of NAML are to spread awareness of current machine learning research relevant to Naval applications, to connect machine learning researchers with experts in military requirements and to build and strengthen collaborations within the defense research community.


poster

### Welcome to the 10th annual workshop on Naval Applications of Machine Learning!

NAML is the premier annual event showcasing current machine learning research relevant to Naval applications. We welcome participation from the defense community, industry, and academia. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. 
Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.
#### On-Demand Access

* This is an on-demand site and all the recorded content from the workshop is available here. 
* Access to the recordings is exclusively available to registered participants.
* Each registrant has received an email notification, with instructions on how to access the content.

[![](https://assets.underline.io/markdown_image/1/image/60cd12aa376c191d3467ea2ee60944ea.png)](https://naml2026.org/)

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications
* To connect machine learning researchers with experts in military requirements
* To build and strengthen collaborations within the defense research community

#### *"What we do as a Navy is a human business. Don’t pass up an opportunity to come together as a team. Don’t pass up leadership opportunities, and don’t outsource them to a PowerPoint. If you have an opportunity to come together as a team, to increase belonging, do it."*
Admiral John M. Richardson, 31st Chief of Naval Operations

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Please log in using the email address you registered with.** If you have any issues logging in, please contact us at <naml2026@underline.io>.

Let’s celebrate the past decade by investing in the next. 🚀

Please log in to explore this event.

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Explore the latest advancements in machine learning tailored to Naval applications 🔎**

With on-demand access to NAML 2026, you can watch all lectures & browse poster presentations at your convenience, while also engaging directly with presenters via the Underline platform - ask questions, exchange ideas, and stay connected.

**If you still haven’t registered, you can do so now.**
👇
**Take advantage of a special one-time offer: full access for just $99, with complimentary registration for students, government, and military employees.**

**Get access here: https://naml2026-ondemand.paperform.co/**

Please check your spam folder, or contact our support team at **<naml2026@underline.io>** if you registered, but still haven't received an access email.

Let’s celebrate the past decade by investing in the next. 🚀

Please register!

The 10th Annual Workshop on Naval Applications of Machine Learning (NAML 2026) was held in San Diego, California, March 2-5, 2026. NAML is a forum for interdisciplinary experts from government, industry, and academia to discuss results of ongoing research on theory and applications of machine learning to address some of the most challenging problems faced by our naval forces. 



Sound transmission in the shallow ocean is highly dependent on the characteristics of the sea floor. An effective seabed, suitable for long-range transmission loss modeling, can be predicted with machine learning using passive sonar. Specifically, spectrograms of passing ships of opportunity (SOO) have been used in a ResNet-18 algorithm to classify an effective seabed type. One challenge with this approach is defining a catalog of seabed classes that are acoustically distinct. Currently, porosity is hypothesized to be the most significant parameter of the seabed; for the current work a catalog is defined with 15 seabed classes based on increasing porosity. This catalog is used to generate synthetic SOO spectrograms to train ResNet-18 to predict the seabed class. The trained algorithms are tested on measured data from the seabed characterization experiment (SBCEX 2017 and 2022) conducted in the New England mud patch. Preliminary results show that categorizing seabeds by porosity is a simpler and more robust approach than the previous catalog based on acousto-elastic parameters. In addition to investigating the seabed catalog, this work highlights the impact of water sound speed variability on deep learning results. Because the sound speed profiles vary between the 2017 and 2022 SBCEX data, differences between the performance of models trained on these datasets show the significance of understanding the water sound speed in order to accurately predict seabed type.

Ship noise for deep learning seabed classification

Due to climate change, the amount of ice cover in polar regions has become more variable. Because ice cover impacts sound propagation, proper sonar performance depends on an understanding of the presence or absence of ice. This project investigates the possibility of using deep learning to make this distinction. In preparation for ocean testing, initial measurements were collected in a laboratory tank. The experiment used a linear chirp from 10-50 kHz recorded at variable source and receiver positions in the tank. Approximately 150 liters of pebble ice were added to the water tank, and sound was recorded before ice was added, while it was present, and after it melted. From each one-second signal, the spectrogram was calculated and labeled based on the presence of ice. This labeled data was used to train a convolutional neural network (CNN) for binary classification for the presence or absence of ice. The trained CNN was applied to data collected on different days with different ambient noise conditions. The most successful CNN reached 81% accuracy on the unseen data. These results indicate the potential for deep learning to identify ice coverage in naval applications.

Acoustic Binary Classification of Ice Cover Conditions Using Deep Learning

General Problem: Future Distributed Maritime Operations (DMO) will rely on large-scale, heterogeneous swarms of unmanned systems (UxS). A critical barrier is achieving resilient autonomy in Disconnected, Intermittent, and Limited (DIL) environments. Standard distributed learning paradigms—such as Federated Learning—fail in these tactical scenarios for two reasons. First, they assume a fixed, reliable communication graph, which collapses under jamming. Second, they cannot handle "functional heterogeneity," where diverse agents (e.g., USVs vs. UAVs) possess incompatible state spaces, making simple parameter averaging mathematically invalid.

Proposed Framework: Sheaf Hypergraph Neural Networks. We propose a mathematically rigorous framework for decentralized representation learning based on Cellular Sheaves over Hypergraphs. Unlike standard Graph Neural Networks (GNNs) using scalar weights, we model the multi-agent system as a sheaf—a topological structure explicitly modeling algebraic constraints between local agent data.

Sheaf Diffusion for Resilient Consensus: We formulate coordination as a diffusion process governed by the Sheaf Laplacian. This operator drives the system toward a "harmonic" state (global consensus) via local exchanges. Because the diffusion is topological, it is inherently robust to node dropouts and dynamic topology changes common in contested spectrums.

Learning Geometric Constraints: We introduce a Sheaf Hypergraph Neural Network (SHNN) where the "communication protocol" is learned. The network learns Restriction Maps—linear transport operators that translate data between heterogeneous agents. During training, the model automatically discovers how to align feature spaces (e.g., sonar vs. optical) to create a unified common operating picture without a central fusion engine.

High-Order Hypergraph Modeling: We utilize hypergraphs to capture multi-way interactions (e.g., triangulation tasks), ensuring the learning process respects high-order mission dependencies rather than reducing them to pairwise links.

Naval Applications: This framework offers foundational technology for NIWC/NRL priorities in trusted autonomy:

Swarm Robotics and Heterogeneous Teaming: For Distributed Autonomous Systems (DAS), this replaces heuristic "flocking" rules with topological guarantees. It enables functionally diverse swarms to maintain formation and allocate tasks even when 50% of communication links are severed.

Collaborative Spectrum Sensing (EW): Distributed agents collaboratively classify hostile RF emitters. The sheaf framework aggregates partial signal features into a global result without transmitting raw IQ data, significantly reducing bandwidth.

Multi-Modal Sensor Fusion (ISR): A decentralized training framework for joint target recognition. Learned restriction maps mathematically ensure that acoustic, thermal, and visual features are fused consistently across the fleet.

Sheaf-Theoretic Hypergraphs for Resilient Distributed Autonomy and Learning in Contested Environments

Division 584 is developing several applied artificial intelligence (AI) capabilities to support cyber security engineering, including the Policy Related Assessment of Controls using Artificial Intelligence (PRAC-AI) and the Controlled Unclassified Information Categorizer (CUICAT); as well as the Technology Transition Classifier (TTC), the Technical Library Chatbot (TCLB) and AI@FlankSpeed. PRAC-AI, a Naval Innovation Science and Engineering (NISE) in NIWC PAC, ingests Risk Management Framework (RMF) test result exports (TREs), uses Python and the Pandas library to filter inherited controls, retrieves authoritative configuration information from policy documents, and generates draft test results using generative artificial intelligence (GenAI) and retrieval-augmented generation (RAG). The Information System Security Engineer (ISSE) maintains responsibility for approval, but PRAC-AI significantly reduces writing and formatting effort and improves consistency across assessments. CUI-CAT employ fine-tuned Bidirectional Encoder Representations from Transformers (BERT) models to detect and categorize all 126 categories of CUI, whereas the TTC categorizes weekly highlights into one of nine NAVWAR Technology Transition categories. The TLCB uses a lightweight retrieval-augmented approach to allow personnel to interact with Naval Information Warfare Center (NIWC) Pacific publications. In parallel, the division is collaborating with NearlyHumanAI through a Small Business Innovation Research (SBIR) effort on AI@FlankSpeed, an Azure-hosted, FlankSpeed-attached platform that builds custom generative AI workflows and exposes application programming interfaces (APIs) for integration with FlankSpeed applications and other cloud-hosted services.

Enabling Consumption of AI/ML Use Cases in the IL5 Environment

Naval tasking requires an intense focus that can burden cognitive resources and may reduce warfighter efficiency and performance accuracy. Understanding cognitive workload (CW) through brain computer interface (BCI) technology may provide solutions to optimize performance of operational procedures. This project reflects an ongoing effort [1] to measure and manipulate cognition through standard psychological tasks to develop a real-time CW classification approach. Here, we report on classifier performance of CW using electroencephalogram (EEG) cognitive task datasets for human in the loop optimization of warfighter performance. We created software (modified from a public source [2]) to reproduce a MATBII [3] cognitive workload experiment, where users perform a multi-tasking exercise with varying cognitive difficulty. We administered an experimental design to collect low and high CW data from our research team. We generated relevant plots to analyze our in-house data against highly processed public datasets. We structured and processed EEG data according to methods that can be conducted for real-time applications, including procedures for noisy EEG channel interpolation. In our search for an optimal real-time model, we tested multiple CNNs with and without pretraining on externally sourced CW datasets [4], using leave-one-out cross-validation for classification accuracy. Following our simulated real-time approach during MATBII task performance, we were able to achieve a mean cross-subject accuracy of 65% (range: 54%-70%) for binary classification of high vs low CW on our small dataset. Notably, we can obtain higher accuracies (mean: 73%; range: 66%-85%) by introducing additional subject-specific training and are currently researching methods to implement this reliably in real-time systems. Altogether, these results directly inform development of real-time BCIs for Navy use, and specifically for use cases that continuously demand and fatigue cognitive resources.
[1] Onners, J., Ronquillo, N., Trammel, T., Houst, D., Sengphanith, R., Wymbs, N. (2025, February 25). Real Time Cognitive Engagement Classification using CNNs. [Poster presentation]. NAML 2025, San Diego, CA, United States. https://doi.org/10.48448/tpkd-xg20
[2] Cegarra, J., Valéry, B., Avril, E., Calmettes, C., & Navarro, J. (2020) OpenMATB: A Multi-Attribute Task Battery promoting task customization, software extensibility and experiment replicability. Behavior Research Methods, 52, 1980–1990. https://doi.org/10.3758/s13428-020-01364-w
[3] Comstock Jr J. R., Arnegard R. J. (1992). The multi-attribute task battery for human operator workload and strategic behavior research. Available at: https://ntrs.nasa.gov/citations/19920007912.Lastedaccessthe02/02/2024
[4] Hinss, M.F., Jahanpour, E.S., Somon, B. et al. Open multi-session and multi-task EEG cognitive Dataset for passive brain-computer Interface Applications. Sci Data 10, 85 (2023). https://doi.org/10.1038/s41597-022-01898-y

Cognitive Workload Estimation of Noisy EEG using ML

Problem Context
Machine learning efforts often begin not with “big data,” but with small, imperfect data. Legacy systems store information in non-machine-readable formats, and institutional repositories suffer from poor maintenance and broken links. These challenges compound when integrating data sources with misaligned data definitions, increasing ML workflow complexity. This work demonstrates a practical framework for making small, messy datasets analytically usable through careful preprocessing, transformation, and model selection.

Technical Approach
This paper demonstrates techniques for integrating heterogeneous government datasets to address foundational issues: (1) geospatial crosswalk construction using representative-point projection rather than geometric centroids for correct tract-to-region assignment; (2) variance stabilization for bounded, right-skewed rate data using Box-Cox transformation; (3) multicollinearity reduction in compositional features using isometric and additive log-ratio transformations; (4) model evaluation via stratified 10-fold cross-validation to address sample-size constraints, comparing linear regression and classification approaches. Together, these transformations mitigate spatial misalignment, heteroscedasticity, and multicollinearity while retaining observations—avoiding the need to remove influential cases (e.g., via Cook’s distance) or drop collinear features.

Results
Linear regression models achieved poor explanatory power (R² = 0.47 ± 0.28) with high cross-validation variance, indicating that continuous rate prediction is unreliable under current data constraints. Reframing the problem as binary classification (high-risk vs. baseline regions using 75th-percentile thresholds) yielded 91 % ± 4 % accuracy with F1 = 0.74 ± 0.16, demonstrating stable performance despite class imbalance. SHAP analysis provided ranked feature importance, enabling interpretable model outputs suitable for government decision-making contexts.

Implications
A well-designed pipeline combining careful preprocessing with appropriate model selection can yield actionable insights despite challenging data conditions. Geospatial crosswalk construction, variance stabilization, and compositional feature handling enabled meaningful analysis where raw data integration would have failed. While regression could not produce reliable forecasts (R² = 0.47 ± 0.28), classification identified high-risk regions (91 % ± 4 % accuracy), supporting prioritization and resource-allocation decisions. These methods are relevant for domains facing similar data constraints. The dataset and crosswalk methodology are publicly available on Kaggle for reproducible validation and adaptation.

David Elgas
MS Data Science Candidate, Harvard University
BSAE, Aerospace Engineering, San Diego State University


Making Imperfect Data Work: A Practical ML Framework for Small-Sample, Multi-Source Data Integration

This project leveraging AI for Mission Assurance and Quality Management Systems in the Navy proposes the development of an advanced artificial intelligence (AI) framework to enhance mission assurance and quality management systems (QMS) within the Navy. By leveraging cutting-edge technologies such as natural language processing (NLP), large language models (LLMs), and specialized AI agents, the framework will address critical challenges in quality assurance, operational readiness, and controlled unclassified information (CUI) management. The solution emphasizes the secure handling of proprietary information, specifications, and requirements, ensuring compliance with authority to operate (ATO) standards.
Fine-tuned LLMs will be central to the framework, enabling domain-specific capabilities such as summarization, topic modeling, and document comparison (e.g., specification-to-functionality alignment). These models will be trained on curated datasets containing CUI, proprietary data, and technical documentation, ensuring relevance and accuracy. Vector databases will facilitate efficient inference and retrieval of embeddings, supporting rapid analysis and decision-making. Specialized agent assistance will automate routine tasks, such as data curation and wrangling, while providing actionable insights to Navy personnel.
To meet the Navy's stringent security and operational requirements, the AI solution will be deployed as a containerized application optimized for small footprint and hardware constraints. On-premises deployment will ensure data sovereignty and minimize exposure to external threats. Prompt engineering techniques will enhance user interactions, enabling tailored responses and improving human-machine collaboration.
The framework addresses key challenges in mission assurance and QMS, including managing large volumes of sensitive data, ensuring alignment between specifications and functionality, and maintaining operational readiness. By automating quality assurance processes, enforcing standardization, and delivering real-time insights, the project will transform the Navy's ability to ensure reliability, efficiency, and security in its operations. This innovative approach positions the Navy to leverage AI for strategic advantage in modern warfare.

Artificial Intelligence for Mission Assurance

The fusion of Artificial Intelligence (AI) with signal processing has highlighted signal labeling as an essential part of the workflow. This poster/presentation aims to explore the methodologies in signal labeling and the integration of deep learning techniques to automate and optimize the process, while focusing on an example of automated labeling of flight maneuvers. While some flight test data sets are already labeled, a large set of historical data may be missing or have less accurate labels. Using AI/ML methods to iteratively label these data sets allows the recovery of this historic data for use in analysis.

Enhancing AI Applications with Automated Signal Labeling – A Flight Test Data Application

Integrated photonics is an innovative technology that merges optical components with traditional electronic circuits, leveraging mature, cost-effective semiconductor manufacturing processes to enable high-speed data transmission and lower power consumption than conventional electrical interconnects. Currently, the market is experiencing robust growth, primarily in data centers and high-performance computing, where photonics-based transceivers are essential for meeting the escalating bandwidth demands of cloud services and AI workloads. Looking ahead, envisioned applications extend beyond telecommunications to include highly integrated solutions like co-packaged optics (CPO) placed directly alongside silicon chips and analog computation systems for lightspeed, energy-efficient matrix multiplication. However, scaling such systems does not come without challenges, such as modulation, loss, noise, and footprint. To address such concerns, we have employed advanced numerical optimization techniques, commonly called inverse design, to the engineering of photonic devices and circuits. Inverse design methods can be used to optimize a user-defined photonic metric with respect to thousands of design parameters using computationally efficient algorithms such as the adjoint state method, leading to, for example, photonic devices and circuits that have near-unity energy efficiency and minimal chip footprints. Furthermore, recent revelations in statistical learning methods such as artificial intelligence and scientific machine learning can be used to enhance inverse design algorithms. We will present our work on integrating the latest automatic differentiation tools into photonic inverse design workflows, constructing physics-informed neural operators as data-free surrogate partial differential equation (PDE) models, and ultra sample efficient optimization methods leveraging low-accuracy surrogates.

Engineering State-of-the-Art Photonic Devices and Circuits with AI-Enhanced Inverse Design

Major theoretical and experimental advances in designing controllable quantum systems have brought us closer to Richard Feynman's vision of using quantum mechanics to simulate nature. Today, we can access up to hundreds of qubits across a range of device types, including superconducting qubits, neutral atoms, trapped ions, photonics, and more. But to unlock the full potential of quantum computing (QC)—meaningful speedups for real-world problems like simulating chemistry and materials science—we need to scale these quantum systems to orders of magnitude more qubits and do so cost-effectively. This milestone, known as quantum utility, marks the point where quantum computers become truly useful. To achieve quantum utility, it will be necessary to tightly integrate quantum processors (QPUs) with classical, high-performance computing (HPC) systems. This is because, rather than acting as general-purpose processors, QPUs function more like accelerators or coprocessors, designed to tackle highly specialized tasks that are beyond the reach of classical computers. This hybrid HPC-QC model will be essential not only in the near term—the noisy, intermediate-scale quantum (NISQ) era—but also for future fault-tolerant quantum computation (FTQC), as quantum error-correction schemes will rely heavily on classical co-processors. To achieve so-called utility-scale quantum computing, it will be necessary to integrate with existing heterogeneous HPC infrastructures and develop a hybrid quantum-classical full computing stack. In other words, we will need to build a quantum supercomputer. 

We present a full quantum-classical stack with tight integration of HPC and QC in multiple layers. At the highest layer, we extend familiar HPC programming environments to include QC capabilities, making it easier for developers to run quantum tasks alongside classical, high-performance ones. This includes a quantum accelerator API, which allows seamless execution of quantum algorithms, and a smart scheduler that efficiently distributes quantum workloads across available resources. At a lower level, we provide an HPC solution for circuit compilation and real-time quantum error correction that uses known hardware profiles to synthesize optimized fault-tolerant circuits. Integration at this layer will require fast communication between the quantum and classical components, and can be enabled by ultra-low latency interconnects. Novel algorithms and heterogeneous coprocessing—including CPU/GPUs, QPUs, and probabilistic processing units (PPUs)—allow the system to partition a particular problem into subproblems and send them out to the right coprocessors, enabling even more efficient execution. To aid in designing and evaluating the hybrid stack, we provide resource estimates for utility-scale (FTQC) problems in materials science as well as demonstrate HPC-QC integration workflows for several near-term (NISQ) applications.

Premium content

Next from NAML 2026

Ship noise for deep learning seabed classification

Similar lecture

Archangel: UAV-based Human and Casualty Detection at The Edge

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES