United States

Cyber-physical systems (CPS) increasingly require real-time intelligence and autonomous operation in environments where centralized analytics cannot meet latency, bandwidth, and reliability constraints. Traditional architectures that rely on cloud, or server-based processing may introduce delays and vulnerabilities that compromise system responsiveness during critical events. We propose an on-edge intelligent architecture utilizing expert device-layer agents, each powered by a specialized Small Language Model (SLM) fine-tuned with physics and behavior informed data. Each agent will be deployed directly to a compact embedded platform, e.g. Raspberry Pi unit, interfaced with individual field devices. 

Our approach integrates two adapter types: one trained on device-specific operational physics and another trained on historical behavioral data. This dual-adapter design enables increased contextual awareness for anomaly detection, data diagnostics and natural-language explanations. 

The experimental framework compares the deployed agents with a centralized baseline solution running on a remote server, creating controlled comparisons between conventional systems and distributed devices. This architecture allows for direct measurement of latency reduction from local inference, and an increase in accuracy performance from physics and behavior aware expert agents. 

Validation is performed using hardware-in-the-loop, power-grid simulators with physical SEL hardware to generate representative data streams. Two sets of three distinct SEL hardware devices are deployed. One set will operate with the proposed decentralized method, and the other set will operate with the typical centralized solution. This work introduces a pathway towards highly interpretable, low-latency, and resilient on-edge intelligence for mission-critical CPS applications.

NAML 2026

Expert Device-Layer Agents for Enhanced On-Edge Intelligence in Cyber Physical Systems

small language models (slm)

on-edge intelligence

cyber physical systems

ai agents

natural language reasoning

anomaly detection

The aim of ESSCIRC and ESSDERC is to provide an annual European forum for the presentation and discussion of recent advances in solid-state devices and circuits.
The level of integration for system-on-chip design is rapidly increasing.

This is made available by advances in semiconductor technology. Therefore, more than ever before, a deeper interaction among technologists, device experts, IC designers, and system designers is necessary. ESSCIRC and ESSDERC are governed by a common Steering Committee and share Plenary Keynote Presentations and Joint Sessions bridging both device and circuit communities.

Highlights:
* 3 joint keynote presentations
* 3 keynote presentations ESSCIRC
* 3 keynote presentations ESSDERC
* Invited papers with overall coverage of all aspects of advanced devices and circuits
* Presentation of IEEE and ESSCIRC/ESSDERC Awards
* Educational Events

JOINT PLENARY talks live broadcast: September 13, 2021 
ESSCIRC and ESSDERC KEYNOTE talks live broadcast: September 15, 2021 
LIVE Q&A INTERACTIVE sessions for all Conference Technical sessions: September 14 – 16 – 17, 2021 
LIVE Q&A INTERACTIVE sessions for all Educational Events: September 20-21-22, 2021 
(on demand videos available from September 6 until November 30, 2021)

**Please see here for the [ESSCIRC Guide](https://drive.google.com/file/d/10dghzuv4aQBty4QfTIY9Q8hXyqsue8Le/view?usp=sharing), [ESSDERC Guide](https://drive.google.com/file/d/10dx9y2tm9cHPDa4upNUYjfXCf44aPry5/view?usp=sharing), and the [Program at a Glance](https://drive.google.com/file/d/10eU0OLTXEpiNzI5hkqv4mSr43T8cedPP/view?usp=sharing)**

Please register for educational events at the following link: https://www.registerme.eu/event/esscirc-essderc-2021/

Please register for the ESSCIRC/ESSDERC Educational Events to join the event.

ESSCIRC ESSDERC Educational Events

The Educational Events of ESSCIRC ESSDERC 2021. 

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. 



Naval operators must execute missions characterized by heterogeneous platforms, contested environments, uncertain sensing, and tightly coupled policy and doctrinal constraints. In contrast, contemporary machine learning ecosystems often presume expert users, well-structured inputs, and narrowly scoped tasks, assumptions that often break down in operational naval contexts. 

This talk introduces LEXICON, an adaptive multi-agent orchestration framework that enables operators to express tasking in natural language. LEXICON leverages a constellation of specialized large language model (LLM) agents to convert these high-level intents into deterministic, verifiable, and auditable workflows that integrate seamlessly with existing planning tools, simulation backends, and retrieval-augmented knowledge sources. The result is a science-driven bridge between unstructured human intent and rigorous algorithmic execution, enabling mission planning pipelines that are transparent, reproducible, and compatible with the Navy’s emerging autonomy and decision-support architectures. 

Key elements include: 
-Multi-agent orchestration with clearly scoped roles wired in a GraphFlow that enforces ordering and termination. 
-A strict data contract and metadata-driven Execution Coordinator that turns REQUIRED_TOOLS lists into deterministic tool runs, with pre- and post-execution gates checking inputs, outputs, and coverage against tool metadata. 
-A retrieval-augmented context layer: document ingestion and embedding build a tagged index; a deterministic DocRetriever agent uses it to enrich context and close gaps, with “doc overrides” merged under explicit rules and checked by the same context-integrity and gating mechanisms as user-provided context. 
-Guardrails including a tool allowlist with an inline JSON to function-call shim that converts agent-emitted call blocks into actual tool invocations, strict versus adaptive execution modes, and a context-integrity policy that prevents agents or RAG from fabricating or silently dropping critical information. 
-Telemetry, tickets, and bounded loopbacks, where detected defects (for example missing geometry or context drift) generate targeted repair requests back to the responsible agent rather than silently continuing. 

We illustrate the approach with Navy-relevant proxy scenarios, including hotspot search and patrol routing with no-fly zones and scan policies, using run logs and state traces to show both successful missions and common failure modes. The talk is organized in two parts: a system overview covering the multi-agent plus RAG architecture and guardrail mechanisms, followed by results and lessons learned from these scenarios—focusing on how to wrap LLM-based agents around legacy tools without violating policy and safety constraints, how much structure data contracts and guardrails need to be useful in naval settings, and what open issues remain for deploying similar orchestration layers in operational environments.

From Intent to Action: Multi-Agent LLMs Coordinating Naval Planning Tools

A transfer learning approach was adopted, using a VGG16 neural network model to classify earthquakes from non-earthquake events in North-East China close to the North Korean test site. Magnitudes below 4 are considered for classification of tectonic events (earthquakes) from explosions. Because of the scarcity of observed explosion data, synthetic seismograms were generated using SW4 to augment the observed data so that overfitting doesn’t happen during validation. Our preliminary experiment shows promising results which may help us identify any potential future small event.

Machine-Learning Event Classification in North-East China using augmented data

The Small Weapons Automated Scoring (SWAS) system is a cost-effective, computer vision-based solution designed to modernize and enhance marksmanship training. SWAS automates the traditionally manual, time-consuming, and subjective process of scoring bullet impacts on targets. By leveraging a standard IP camera and a sophisticated machine learning model, the system provides shooters with immediate, objective, and data-rich feedback on their performance. The application processes this video feed in real-time, feeding each frame into a purpose-built computer vision model trained to detect and locate new bullet impacts. Key outcomes include improved training efficiency, detailed performance analytics for individual and programmatic review, and a significant reduction in the personnel and resources required for range operations.

The system's primary objective is to deploy a low-cost, lane-based automated scoring system that provides instantaneous feedback and robust data logging. SWAS empowers instructors and shooters with precise, actionable data, accelerating skill development and creating a comprehensive digital record of performance over time.

Small Weapons Automated Scoring

In support of developing a credible digital twin for prognostic health management of Navy-relevant equipment, one must first assess the quality of captured data before processing for either digital twin development or health assessment. This paper focuses on assessment of data quality in time-series vibration sensor data. Accelerometers are common in many machinery and equipment health management systems. Sources of poor quality, or corrupted vibration sensor data include internal sensor faults, wiring and connector problems, and incorrect or loose sensor mounting. Other common types of sensors used for health management, such as current or pressure sensors may have similar failure modes. The signal path from measurand to health management processing algorithms are themselves systems that require monitoring and assessment.

Numerous computational techniques exist to varying levels of maturity for monitoring the health of a system from which one can acquire data, including variants of statistical process control and scientific Machine Learning. Those techniques are ubiquitous for assessing equipment health. Whatever the chosen technique to monitor a process, one must curate the data appropriately. In collaboration with the Naval Surface Warfare Center Carderock Division, our team has developed a classifier using machine learning to distinguish between “good” and “bad” data from piezoelectric accelerometers. Algorithms extract salient features from vibration data and predict the class of data quality.

This paper describes the preprocessing, feature extraction and training of the classifier for data quality. We demonstrate our methodology on data sets acquired from machinery on a representative ship platform, from a shaker in a laboratory, and a combination thereof. A secondary goal of the research was to explore the use of simulated sensor data (in this case from a shaker) to supplement the relatively small amounts of actual faulty sensor data. The methodology also includes the use of real modeled sensor data to generate representative laboratory data for model development. This is a common problem in the development of health monitoring systems, for sensor as well as machinery faults. The results show both the ability to detect sensor quality issues and the successful use of supplemental data in model training. 


Feature Engineering and Classifier Training Method for Vibration Data Quality Assessment

My team at the University of Arizona’s SpaceTREx Laboratory has been working extensively with planetary scientists on advancing spacecraft technologies for high-priority reconnaissance missions. Among them are plans for fast flyby of small bodies, moons, and interstellar objects. For these application areas, we have been proposing spacecraft swarms that fly by a body or inspect it, providing rapid 3D mapping and imagery. The organization and coordination of these swarms are achieved through the use of Machine Learning algorithms that evolve the craft design, swarm organization, and trajectories of multiple crafts. Building on these flyby mission concepts, we are beginning to explore alternative techniques, including rendezvous, impactor/ramming, and landing/docking. Many of these architectural and operational techniques may be applicable for robotic UUVs. Using a swarm or fractionated modular craft, we may split groups depending on needs to take on specialized tasks, including maintaining communications links, inspecting a craft, or performing analysis from a complex debris field. Given a target, maybe an uncooperative craft, one or a few may get close, but others in the swarm would maintain a safe distance. If the craft is uncooperative and is on a collision course with vital assets, one consideration may be to use a ramming spacecraft that impacts and conjoins with the uncooperative craft and moves it off onto a safe trajectory. In other examples, we have used a spacecraft swarm to perform multipoint observation of a complex 3D event without any craft being a single point of failure for the evolving mission. These projects are early-stage and currently supported by NASA and Brinson Foundation funding. We have made promising progress in simulation and are planning towards an early-stage hardware demonstration in the laboratory through external support in preparation for field demonstrations.

Advancing Machine Learning Directed Design and Organization of Spacecraft Swarms Concepts Towards Terrestrial and Underwater Applications

Maritime vessel controls and operational technology (OT) systems are getting more complex and interconnected. With industry trends aiming to reduce crew, automate tasks, and improve efficiency, these networks are expanding in scale, intricacy, and criticality for vessel operation and maintenance. The standard controller area network (CAN) bus for maritime vessel networks, developed by the National Marine Electronics Association (NMEA), known as NMEA2000. NMEA2000 is an application layer network protocol built on the ISO11783 standard and compatible with automotive SAEJ1939, it uses unique message identifiers known as Parameter Group Number, to define the data within each communication frame.

Despite its widespread use, NMEA2000 remains a relatively unexplored domain, particularly in understanding normal versus abnormal network behavior, due to the unavailability of open-source datasets. To address this gap, we constructed a NMEA2000 system consisting of five nodes: GPS/Radar, Wind Speed/Direction sensor, and Multifunction Display. Using this setup, we collected datasets to analyze system behavior and developed an intelligent fingerprinting techniques for each sensor, establishing a baseline of the normal operating system. We subject the system to controlled attacks to evaluate the accuracy and effectiveness of the fingerprints. This work represents a foundational step towards enhancing security and reliability in maritime OT systems.

Intelligent NMEA2000 Network Fingerprinting

The Citizen Broadband Radio Service (CBRS) utilizes the 3.55-3.7 GHz band for dynamic spectrum sharing, enabling commercial cellular network operations without causing harmful interference to high-priority Navy incumbent systems. This protection is facilitated by a dedicated Environmental Sensing Capability (ESC) that continuously monitors for incumbent activity and triggers network responses. As spectrum sharing extends to additional frequency bands beyond CBRS, augmenting dedicated sensing with complementary sensing modalities may be necessary for enhancing detection performance and response times. 
This presentation proposes a novel approach for sensing that leverages commercial cellular base stations as distributed sensors to detect incumbent spectral utilization. The approach employs the You Only Look Once (YOLO) real-time object detection architecture applied to a time–frequency representation (spectrogram) of received signals. YOLO models are trained on a combination of synthetic and measured radar datasets that encompass diverse radar types, pulse structures, and signal-to-noise ratios (SNRs). The framework simultaneously detects radar emissions and classifies their type, enabling efficient and reliable identification of incumbent activity.
In contrast to signal processing and classical machine learning methods that rely on handcrafted features and fixed detection thresholds, the proposed deep learning pipeline adapts to complex, nonstationary environments and maintains high detection performance across varying SNRs and interference conditions. The sensing capability is validated using a laboratory testbed comprising an NVIDIA Aerial Platform-based 5G base station with a GPU-accelerated physical layer integrated with the Open-Air Interface (OAI) software stack. Radar waveforms specified by the National Telecommunications and Information Administration (NTIA) for the ESC certification were utilized for performance testing. Executing advanced machine learning (ML) inference on a GPU-powered base station yields end-to-end latencies that are significantly better than current CBRS deployments. For this study, we target a chirp radar waveform as specified by NTIA for ESC certification. Our YOLO-based radar detector achieves approximately 98% precision and 95% recall at moderate-to-high SNRs, with sub-10 ms ML model inference per spectrogram. The improved sensing performance enables commercial networks to initiate timely mitigation actions that can protect incumbent systems while maintaining the cellular communication quality of service.
The proposed YOLO-based distributed sensing framework demonstrates that commercial 5G base stations can serve as efficient, real-time spectrum sensors, achieving improved detection accuracy and low latency. The experimental results establish deep learning–based sensing and AI-driven spectrum agility as a practical foundation for next-generation dynamic spectrum sharing systems. 


AI-Enabled Real-Time Sensing Using 5G Base Stations for CBRS Spectrum Sharing

Emerging GenAI software developer tools (e.g. Claude Code, Codex, Antigravity, etc.) in industry are transforming what it means to be a software engineer. However, current enterprise GenAI providers (NIPRGPT, CapraGPT, DoD HPC, etc.) across the DoD have multiple gaps, including limited application programming interface (API) access to large language models (LLMs) and the supporting infrastructure required to share agentic workflows across the enterprise. 

In this talk, the Naval Information Warfare Center (NIWC) Atlantic's Generative AI for Software Development (GenAI4SW) team investigates how to affordably scale AI tool deployments from the individual to the team level. As part of our efforts, we examine the infrastructure and workforce development strategies required to share LLM endpoints, build agentic workflows, and debug LLM-based applications. Our experiences highlight how GenAI tools are leveraged at a Navy Warfare Center and highlight grassroots efforts surrounding AI-enabled software engineering best-practice. 

As part of this effort, the team deployed multiple 49-billion parameter reasoning models across the DoD High Performance Computing Modernization Program (HPCMP) and shared OpenAI-compliant APIs that could be integrated into software development environments across the team. Our contributions are extended to included: (1) an architecture for sharing Model Context Protocol (MCP) servers to enable LLMs to call external tools, (2) observability frameworks to enable application debugging with LLMs, and (3) human-centered interfaces for building (e.g., n8n) an interacting with (e.g., Open WebUI) LLM-based applications. 

Our efforts are designed to enable and advertise the use AI tools across the software development lifecycle, influence enterprise adoption of these workflows, and demonstrate capabilities of existing computational infrastructure for GenAI within the DoD.

Generative AI Infrastructure for Software Development

Understanding the state of a given workforce is a prevalent issue across commands and organizations. One specific difficulty is predicting and planning for attrition. Knowing who and when an employee will attrit is paramount to maintaining a healthy workforce and ensuring ample time for transition periods between new and more senior employees. Predicting attrition, both on the individual and command level, is often a challenging problem to solve due to high class imbalance and data scarcity. Additionally, organization managers may be resource constrained, limiting the time available to parse through large candidate pools and find ideal candidates to backfill attrition-based losses. The proposed solution is two-fold: a machine learning (ML) based attrition prediction model and a novel candidate recommendation system. For attrition prediction, multiple custom models were developed, trained, and tested to compare and validate prediction performance including Deep Neural Networks (DNNs) and Convolution Neural Networks (CNNs) for individual predictions as well as Long Short-Term Memory (LSTM) models for command-level predictions. A suite of oversampling and feature engineering tools were employed in conjunction with the ML models. An explainability study was conducted for individual attrition predictions using SHapley Additive exPlanations (SHAP) to provide traceability between individual features and model outputs. For the candidate recommendation system, an ML-based resume intake pipeline was developed for reviewing resumes and evaluating job applicants for positions in specific command knowledge areas (KAs). Given the KA of an open position, the system recommends the top candidate matches using the skills extracted from their resume via a locally-hosted Large Language Model. Top-matching applicants are determined by leveraging vector embedding to compute a match score for each (KA, applicant) pair, where the match score quantifies the weighted cosine similarity between an applicant's skills and the KA's description and job-specific skills. The system concisely returns each applicant's most position-relevant skills for reference alongside the match score. The system output allows hiring managers to spend less time scanning through applicant resumes, and more time reviewing those applicants with relevant skills, resulting in better quality matches between candidates and job requirements. Both the attrition prediction models and the candidate recommendation system provide Naval Surface Warfare Center – Philadelphia Division with enhanced tools for business analytics and organizational needs.

Artificial Intelligence and Machine Learning for Business Analytics

Technology Goddesses™️was founded in 2002 as a volunteer-led 501.c.3 to improve the representation of diverse people in STEM degrees and occupations. This poster shares some of the metrics that need to change and how our approach to STEM education is different than many other programs.

Premium content

Next from NAML 2026

From Intent to Action: Multi-Agent LLMs Coordinating Naval Planning Tools

Similar lecture

Neuro-inspired Edge-AI for Autonomous Systems