United States

We present a novel approach to developing autonomous agents for the MCTF (Maritime Capture the Flag) competition by implementing both the environment and agents entirely in JAX. Our primary motivation is to leverage JAX hardware acceleration to significantly speed up reinforcement learning and machine learning training pipelines, enabling more extensive experimentation and faster iteration cycles. 

Our JAX-based implementation demonstrates decent performance in controlled settings, achieving 100\% win rates against random and purely offensive strategies red agents. However, the agent struggles to break ties against purely defensive opponents, a challenge consistent with results observed in previous competition years. These results validate that our JAX environment reproduces known strategic dynamics while offering substantially improved training efficiency. 

Furthermore, our work is important in linking the faster training back to the original environment, demonstrating sim-to-real issues often seen in reinforcement learning applications in the real world. A critical finding concerns the sim-to-real transfer gap: when deployed in the original Pyquaticus, agent performance degrades. Against random agents, we achieve 3 wins, 0 losses, and 7 ties. Notably, agents exhibit out-of-bounds navigation errors more frequently than in the JAX environment, highlighting fundamental differences between our accelerated training environment and the competition platform. 

This transfer problem represents the primary challenge for future work. We outline several promising directions to address these limitations: physics-based analysis to identify and minimize environmental discrepancies, followed by either fixing the training environment and retraining, or applying domain adaptation techniques during training or inference to compensate for environmental differences. Other methods include recent sim-to-real research such as progressive neural networks to mitigate catastrophic forgetting during cross-environment transfer, and multi-environment training strategies. These approaches collectively aim to preserve training efficiency while achieving robust competition performance.

NAML 2026

Accelerated Multi-Agent Reinforcement Learning for Maritime Autonomous Agents: Towards Bridging the Sim-to-Real Gap

reinforcement learning acceleration

sim-to-real transfer

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

technical paper

### 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. 🚀

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. 



Small-UAV autonomy often couples perception to action without provable policy guarantees, and cloud inference is unsuitable in contested communications. We present OSIRIS, an onboard framework that converts edge-vision detections into fail-closed, deadline-aware mission interlocks on commodity hardware—Raspberry Pi 5 with Hailo-8/8L and VisionFive 2 (RISC-V) with Hailo-8L—with CPU (NCNN) fallbacks. Detector outputs are exposed as boolean gates (e.g., pad_present, person_in_AOI, hazard_nearby) that must satisfy both confidence and time budgets; under low confidence, missed deadlines, or module faults, gates abstain/deny rather than allow. A compact policy DSL binds gates to actions (e.g., “land only if pad_present ∧ not person_in_AOI within r meters”) and compiles a machine-checkable safety report (evidence, thresholds, timing) for after-action review. OSIRIS does not plan routes; it authorizes or blocks mission segments at execution time and can be used alongside text-to-mission planners only as related work. We will present a recorded, offline ArduPilot software-in-the-loop (SITL) with Gazebo demonstration for three gates—(1) landing-pad confirmation for final descent, using INT8 YOLOv8n with a deterministic ArUco baseline; (2) people/vehicle exclusion inside the area of interest; and (3) a hazard sentinel for speed/AGL derating—implemented with INT8 YOLOv8n on Hailo and NanoDet/NCNN as CPU fallback. We report frame-to-decision latency (P50/P95), deadline-miss rate, gate FN/FP at the chosen operating point, throughput (FPS), and energy per decision (J/decision), and provide a controlled cross-silicon comparison (Arm vs. RISC-V) under identical clips and thresholds; fault-injection tests (dropped frames, delayed inferences) verify abstention behavior. Distribution A artifacts—gate definitions, policy-DSL templates, Hailo/NCNN pipelines, run scripts, and example video→gate→policy traces—enable partial reproduction. Assumptions: VLOS operations, preloaded AOI polygons, recorded demo (no live flight). Benefit to the Navy: proof-carrying, cloud-independent interlocks that reduce operator burden and risk while increasing policy compliance (safe landing, human/vehicle exclusion) on low-SWaP edge platforms, advancing auditable autonomy and human–machine teaming for C2 at the tactical edge.


OSIRIS - Onboard Sensor Inference for Rules-Integrated
Safety: Deadline-Aware Policy Gates for UAV Missions at the
Edge

Sea Strike 2043 sketched a “command center of the future” but left the core question open: what brain actually drives it? Atombeam’s Persistent Cognitive Machine (PCM) is a working prototype of that brain, implemented as Kennedy PCM for maritime command and control. Instead of treating sensor and intelligence feeds as logs or text tokens, PCM encodes them as trajectories on a continuously evolving geometric manifold, providing true persistent cognition at theater scale.

We will demonstrate Kennedy PCM on a notional island-defense vignette inspired by Sea Strike: a Red amphibious invasion attempting to force a strait against Blue forces. The system ingests multi-domain, multi-timescale data (radar, AIS, UxS swarms, ISR, cyber events, operator inputs) and (1) builds a layered manifold of the battlespace, (2) discovers and retains operational patterns over months of simulated conflict, and (3) searches this learned geometry for courses of action under explicit constraints (ROE, risk, asset limits). In the demo, Kennedy uses a year of prior engagements to predict Red’s feint and strait-transit timing, identifies a geometry-driven COA that exploits an emergent 8-second vulnerability window on the landing force, and explains its recommendations via “geometric certificates” that can be replayed and audited by watch teams.

We explicitly contrast this with today’s LLM/agent stacks. LLMs must repeatedly “start from zero,” re-discover context every shift, and rely on brittle toolchains and cloud-scale compute; they do not naturally remember years of operations or share that memory across units. PCM instead learns continuously, storing experience directly in geometry, with a footprint small enough to operate within a single system-on-chip. This makes it practical to embed Kennedy-class cognition in thousands of UAS and other edge platforms, allowing them to operate coherently and complete missions even under EMCON or total loss of reach-back communications. Multiple PCMs federate by exchanging compact geometric patterns rather than raw data or prompts, enabling learning across strike groups while respecting bandwidth limits and classification boundaries.

We will conclude with implications for NIWC Pacific’s cognitive C2 vision, JADC2 integration, human–machine teaming experiments, and scaling PCM to live Navy data and future CCA/NGAD-style, UAS-heavy kill webs. Preferred format is a 20-minute short talk with live demo; the material can be adapted to a lightning talk if required.



Kennedy PCM: Persistent Geometric Cognition for the Naval
Command Center of the Future

As the number of new use cases on the Jupiter data platform has grown annually, so has the number of user questions that are sent to Jupiter’s technical support emails and dedicated office hours. To address this influx, “Jupiter Assist” was developed as the platform’s in-house large language model (LLM) chatbot. It is able to take advantage of years of documentation and technical support conversations using a retrieval-augmented generation (RAG) framework to provide accurate, domain-specific answers in real-time. Additionally, we cover the integration of generative models within Tableau tusing TabPy and Databricks MLflow, which provide the technical architecture for Jupiter Assist.

An LLM-Powered Technical Support Tool: Jupiter Assist

Reinforcement learning (RL) has demonstrated remarkable success in solving complex decision-making problems, yet its adoption in critical domains is hindered by the lack of interpretability in its decision-making processes. Existing explainable AI (xAI) approaches often fail to provide meaningful explanations for RL agents, particularly because they overlook the contrastive nature of human reasoning—answering” why this action instead of that one?” To address this gap, we propose a novel framework of contrastive learning to explain RL selected actions, named VisionMask. VisionMask is trained to generate explanations by explicitly contrasting the agent’s chosen action with alternative actions in a given state using a self-supervised manner. We demonstrate the efficacy of our method through experiments across diverse RL environments, evaluating it in terms of faithfulness, robustness and complexity. Our results show that VisionMask significantly improves human understanding of agent behavior while maintaining accuracy and fidelity. Furthermore, we present examples illustrating how VisionMask can be used for counterfactual analysis. This work bridges the gap between RL and xAI, paving the way for safer and more interpretable RL systems

Why the agent made that decision: contrastive explanation learning for reinforcement learning

AI systems deployed in military contexts present unique risk profiles not addressed by traditional cybersecurity models. This work introduces a hybrid threat modeling and risk validation methodology that aligns AI-specific attack vectors from MITRE ATLAS with two risk control frameworks: the OWASP Risk Rating Methodology (RRM) and the US Navy's Operational Risk Management (ORM) model. Our approach begins with threat enumeration using MITRE ATLAS to identify attack paths relevant to military AI systems, including prompt injection, data poisoning, and misuse of model outputs. We then apply OWASP RRM to score technical likelihood and impact. These results are translated into ORM terms of Severity, Probability, and a Risk Assessment Code (RAC) to support risk decisions by command elements. We validate this method through structured analysis of Retrieval-Augmented Generation (RAG) systems used for mission planning and decision support. Failure modes such as hallucinated intelligence summaries and source contamination are modeled and scored, producing risk matrices in both OWASP and ORM formats. This enables both cybersecurity teams and operational leadership to act on a shared, mission-aligned view of AI risk. The resulting framework provides a practical, adaptable methodology for AI risk assessment in coalition environments.

A Hybrid Threat Modeling and Risk Validation Framework for Mission-Critical AI Systems

If two experts disagree on a test, we may conclude both cannot be 100% correct. But if they completely agree, no possible evaluation can be excluded. This asymmetry in the utility of agreements versus disagreements is explored here by formalizing a logic of unsupervised evaluation for classifiers. Its core problem is computing the set of group evaluations that are logically consistent with how we observe them agreeing and disagreeing in their decisions. Statistical summaries of their aligned decisions are inputs into a Linear Programming problem in the integer space of possible correct or incorrect responses given true labels. Obvious logical constraints, such as, “The number of correct responses cannot exceed the number of observed responses,” are inequalities. But in addition, there are “axioms”—universally applicable linear equalities that apply to all finite tests. The practical and immediate utility of this approach to unsupervised evaluation using only logical consistency is demonstrated by building no-knowledge alarms that can detect when one or more LLMs-as-Judges are violating a minimum grading threshold specified by the user.

Logical Consistency Between Disagreeing Experts and Its Role In AI Safety

This presentation explains how an AI-enabled adjudication framework—combining machine learning, large language models, and interpretable simulation—has recently accelerated and improved naval wargaming and simulation. Developed through collaboration between the Naval Postgraduate School, AlphaPlay, New York University, and the University of California, Santa Barbara, the system was recently applied in an operational wargame and demonstrates how integrating AI directly into the adjudication workflow significantly increases the speed, transparency, and analytical fidelity of multi-domain wargaming.
Traditional adjudication often relies on human judgment and legacy simulations that can be slow, opaque, or difficult to adapt. By contrast, our AI-assisted architecture computes engagement outcomes—including dogfights, missile exchanges, defensive penetrations, and platform-level kill assessments—in seconds. These gains result from a hybrid approach that combines machine-learned survival models with rule-based timing and firing logic. The entire pipeline runs without network connectivity, enabling use in classified environments.
A core technical contribution is the use of explainable AI throughout the system. Each adjudicated event produces a machine-generated reasoning trace, integrating deterministic modeling with LLM-based natural-language explanations to describe why an outcome occurred. This produces a transparent battle log that documents assumptions, intermediate calculations, uncertainty sources, and causal factors. The explainability layer enables adjudicators to validate results, accelerates player feedback cycles, and supports systematic refinement of the underlying ML components.
AI also enables more rigorous and repeatable scenario experimentation. The system can rapidly evaluate alternative laydowns, variations in force structure, base/favorable/unfavorable conditions, and multi-team (red/blue/green) interactions while maintaining consistent adjudication logic. Fog-of-war conditions are generated through AI-structured information partitioning, creating asymmetric data access without bespoke scenario scripting. This results in a scalable environment for exploring operational contingencies with greater temporal and analytical resolution.
From an organizational perspective, integrating AI into wargaming workflows strengthens human–machine teaming. Players receive faster and more coherent feedback; adjudicators gain transparent, reproducible AI-generated reasoning; and analysts can compare outcomes across large scenario libraries supported by consistent ML-based adjudication. This enhances the utility of wargames for force-design analysis, operational planning, and data-driven decision support.
Collectively, this research illustrates how integrating modern AI methods into adjudication can transform next-generation wargaming by improving speed, transparency, analytical rigor, and trustworthiness while maintaining essential human oversight.




AI-Assisted Engagement Adjudication in Naval Wargaming

Fuel distribution and logistics is a complicated but necessary field. It requires a greater understanding of the environment than just getting from A to B, it needs everything in between. That is precisely what Logistics Bulk Fuel Expeditionary Analysis and Simulation Tool (LOGBEAST) brings to the table for the United States Marine Corps (USMC). 

Currently, Marines must engage in laborious manual planning using decentralized data in spreadsheets—a process that is time-consuming, tedious, and prone to human error. Additionally, it is not accessible by the wider Marine Corps or Department of War’s (DoW) Logistics Enterprise. Using classic optimization modeling techniques along with data acquired from throughout the Indo-Pacific theater, LOGBEAST takes in a detailed configuration from the user and returns a comprehensive fuel delivery plan with estimated delivery times that take into account: physical transfer rates of fuel, queuing of trucks, slow-downs due to congestion in population centers, and other factors impacting ground travel with large vehicles.

Currently deployed in DoW’s ADVANA platform, LOGBEAST provides a centralized industrial grade optimization model built for speed and efficiency using software that determines the optimal schedule of fuel deliveries from the hub to downstream squadrons positioned at spokes, then generates the shortest paths along real roads, routing around physical restrictions caused by civil infrastructure and/or user identified restricted zones.

LOGBEAST accounts for maximum realism such as the visualization of historical weather trends layered on high resolution satellite imagery, and user selection of route optimization by either time or distance and/or obstacle avoidance, plus provides an analysis of potential locations for fuel equipment laydown placement at certain locations.

Currently, LOGBEAST leverages a Retrieval-Augmented Generation (RAG) Large Language Model (LLM) enriched with USMC fuel planning doctrine enabling context-aware and data grounded responses to quickly and precisely answer questions for users. Future AI plans include the modeling fuel distribution via sea and air. Given LOGBEAST’s focus on the Indo-Pacific Theater, incorporation of the naval layer for island-to-island fuel transfers and aerial transport is critical to model the full scope of the problem USMC faces in the Pacific. To advance efforts, an agentic AI approach may be pursued that would allow the LLM to select not only the best models/data needed for the mission but to proceed with utilizing said tools and provide the user with the generated outputs. 

Ultimately, LOGBEAST fills a void felt by the Marine Corps that would otherwise render the fuel planning process vulnerable to slowdowns, human errors, broad generalizations, and a lack of centralized data sharing amongst all Services.


Logistics Bulk Fuel Expeditionary Analysis and Simulation Tool (LOGBEAST)

Presenting PEO C4I as an acquisition command, how S&T is done at PEO C4I and how industry/academia/government can help with AL/ML.

PEO C4I Science and Technology (S&T) at NAML

The motion response of a surface vessel in following or stern quartering seas may include a sudden uncontrollable turn, known as broaching. The physical reason for broaching is an instability in yaw caused by a combination of the hull forces, thus the rudder action may not have an effect at all. Broaching-to is not a new problem; it has been known since the Age of Sail [1], so experienced operators are aware of it.

A commonly known precursor to the broaching phenomenon is surf-riding. During a surf-ride, a ship accelerates or decelerates to match the wave celerity such that the ship travels with the incident wave profile. In the realistic sea, these periods of matching the speed of the wave profiles appear and disappear with acceleration, making analysis complex. Given the common frequencies inherent to the sea and typical ship speeds, surf-rides are typically linked with periods of “high-run” [2]. It follows that the instantaneous speed of the ship as well as the speed of the wave profile at the ship’s center of gravity are important parameters that inform surf-riding and potentially broaching.

The proposed method introduces a machine learning framework that links time series realizations of ship speed and wave celerity to the instantaneous yaw of the ship. A Long-Short Term Memory (LSTM) [3] neural network is trained with various combinations of tri-chromatic wave time series and surge velocity time series simulated with the reduced-order SimpleCode [3] as input and with yaw time series generated by SimpleCode as targets. The results of the trained neural network are evaluated against SimpleCode in ability to estimate large yaw events.

References

[1] Spyrou, K. J. “Perceptions of Broaching-To: Discovering the Past”, Chapter 22 of Contemporary Ideas on Ship Stability, Neves, M.A.S., Belenky, V., de Kat, J. O., Spyrou, K. and N. Umeda, eds., Springer, ISBN 978-94-007-1481-6, pp. 399-414, 2011.

[2] Themelis, N., Spyrou, K., and Belenky, V. “Surf-riding in Multi-chromatic Seas: “High-runs” and the Role of Instantaneous Celerity,” in Ocean Engineering, Vol. 120, pp 230-237, 2016.

[3] Hochreiter, S., and Schmidhuber, J., “Long Short-Term Memory,” Neural Computation, Vol. 9, No. 11, pp. 1735–1780, 1997.

[4] Weems, K. and Belenky, V. “Extended Fast Ship Motion Simulations for Stability Failures in Irregular Seas,” Proceedings of the 13th International Conference on Stability of Ships and Ocean Vehicles, Kobe, Japan, 2018.

Premium content

Downloads

Next from NAML 2026

OSIRIS - Onboard Sensor Inference for Rules-Integrated Safety: Deadline-Aware Policy Gates for UAV Missions at the Edge

Similar lecture

Natural Language Can Facilitate Sim2Real Transfer