United States

Analysis of civil considerations is the topic that may receive the least amount of time in the
planning course at the Expeditionary Warfare School (EWS) but is the step that can
potentially take the most time during operational planning. The Marine Corps Planning
Process (MCPP) calls for planners to evaluate the civil terrain by analyzing six
characteristics of the Area of Operations (AO): Areas, Structures, Capabilities,
Organizations, People, and Events (ASCOPE). For each of these ASCOPE categories, six
operational variables are to be considered: People, Military, Economic, Social,
Informational, and Infrastructure (PMESII).

Crosswalking the ASCOPE categories with the PMESII variables produces thirty-six
categories of civil considerations that each must be analyzed individually. The data that supports this analysis is typically complex, includes a large amount of unstructured text,
and is unique to the Area of Operations (AO). It is typically not possible to re-use previous
analyses or even data sources when performing MCPP in a new region.

Under sponsorship of ONR Code 34’s Artificial Intelligence To Operations program, the
Civil-Considerations thrust of our COA Optimization, Analysis, and Comparison Harness
(COACH) project uses the ASCOPE/PMESII conceptual structure to automate the process
of identifying, retrieving, and organizing civil considerations information.

COACH’s Civil Considerations Service (CCS) uses a Large Language Model (LLM) to identify
information relevant to each of the thirty-six crosswalk categories in a corpus of
unstructured text. Depending on the specific LLM used, the CCS can process non-English
data and produce English-language translations of the identified information. The CCS
produces a draft of the PMESII/ASCOPE crosswalk that users can then review and modify
to produce the final MCPP planning product.

The CCS is currently being integrated with the Higher Echelon Mission Planner (HEMP) and
the Small Unit Mission Planner (SUMP) tools, both of which are also AI2OPS projects.

NAML 2026

Automation to Address Civil Considerations in the Marine Corps Planning Process

civil considerations

marine corps planning process

### 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
<br><br>

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. 



technical paper

Extreme heat and cold stress pose serious risks to warfighter health, performance, and mission success. Individual susceptibility to thermal strain varies widely, so real-time physiological monitoring with intelligent analysis is critical to detect early signs of heat or cold injury [1]. We propose a two-stage wearable sensor-AI framework that integrates both mission planning and mission execution. In the planning stage, expected weather, operational activity, and warfighter demographics will serve as inputs to a Bayesian network-based causal inference model. This enables commanders to forecast likely physiological strain (e.g., elevated heart strain, skin temperature, hydration risk), simulate alternative scenarios, and adjust mission design (e.g., workloads, schedules, gear distribution) proactively before deployment. In the execution stage, we will employ continuous multimodal sensing (wrist, chest, helmet sensors) to capture real-time physiological and environmental data. Probabilistic AI models (Bayesian networks with temporal predictors) will infer key strain markers such as Relative Cardiac Cost, Skin Temperature, and Electrodermal Activity, and issue short-horizon forecasts (e.g., minutes ahead) to provide early warning of unsafe conditions. In prior work, we successfully deployed Empatica Embrace Plus wristbands and REED environmental sensors on 100 high-activity workers, building a Bayesian network regression model to predict skin temperature, cardiac cost, and electrodermal sweat signals from activity and weather data. This field study showed that low-cost wearables can yield actionable estimates of physiological heat stress under real-world conditions. Building on these results, our proposed approach will adapt the framework to military operations: integrating soldier-grade sensors (wrist, chest, helmet), tracking markers such as Relative Cardiac Cost (RCC), Skin Conductance Level/Response (SCL/SCR), skin temperature, and enhancing the AI model for real-time prediction and early warning. We will validate the system in the relevant field, simulate training exercises, and demonstrate integration with mission planning tools. This research aligns with ONR (Office of Naval Research) and NRL (U.S. Naval Research Laboratory) priorities in human performance and force health protection [2]. By enabling individualized monitoring and AI-based decision support, the proposed framework aims to improve warfighter survivability, reduce heat/cold casualties, and sustain mission readiness in extreme environments.

References
1. Josh Foster, Simon G. Hodder, Alex B. Lloyd, and George Havenith. 2020. Individual Responses to Heat Stress: Implications for Hyperthermia and Physical Work Capacity. Frontiers in Physiology Volume 11-2020, (2020). DOI: https://doi.org/10.3389/fphys.2020.541483. 
2. Office of Naval Research. n.d. Code 34: Warfighter Performance Department. https://www.onr.navy.mil/organization/departments/code-34.

Wearable Sensor Data-enabled AI Reasoning Framework for Warfighter’s Thermo-Physiological Survivability and Mission Readiness in Extreme Temperature

The development of robust vision-based navigation systems for maritime environments is critically dependent on the availability of high-quality, accurately labeled training data. Acquiring such datasets presents formidable logistical and technical challenges. A key difficulty lies in generating synchronized truth data, which requires capturing precise, time-aligned pose information, encompassing position and attitude, for both the UAV and the ship simultaneously. Without this critical truth data, collected imagery lacks the fidelity needed to train the advanced models essential for autonomous navigation. The Vision POD (VisPOD) system was purposefully designed to address these challenges and enable the creation of robust datasets for maritime machine learning applications.

As an advanced data collection system, VisPOD integrates with unmanned aerial vehicles (UAVs) to capture high-quality datasets for the development of vision-based navigation models. VisPOD captures multi-modal datasets across varied positions and environmental conditions, providing the comprehensive data needed for robust algorithm development. Additionally, its precise pose tracking and adaptable imaging capabilities ensure datasets are optimized to meet operational objectives. 

With its scalable architecture and ability to adapt to mission needs, VisPOD represents a powerful tool for advancing maritime data collection and driving innovation in machine learning applications. VisPOD is the essential data collection component of the Vision Navigation (VisNAV) pipeline, delivering datasets that power the automated labeling process, ultimately driving algorithm training and model development. This presentation will explore VisPOD’s capabilities and its role in delivering high-quality datasets for next-generation vision-based navigation machine learning models.


VisPOD’s Vision: Transforming Data Collection for Smarter Machine Learning Models

The mass deployment of Unmanned Systems (UxS) to the fleet burdens operators with increasingly complex decision-making demands that threaten their ability to conduct effective Command and Control (C2) mission management. The decision space required for UxS operations will expand exponentially as future missions scale in size, employ heterogeneous multi-domain platforms, and operate within Denied and Degraded Environments (D2E). Addressing these challenges requires unified, integrated C2 tools that are context-aware, objective-driven, and capable of synthesizing diverse autonomy-enabling technologies. No single capability can meet these needs in isolation; closing these gaps demands a cohesive toolchain that weaves together multiple models into a unified C2-focused solution.

We introduce the Decision Engine for Situational Context, Autonomous Replanning, Task Execution & Scheduling (DESCARTES), a framework designed for rapid integration of advanced autonomy models supporting C2 mission management. DESCARTES incorporates optimized path planning, dynamic environment discretization, priority task queuing, agentic Artificial Intelligence (AI), large language models (LLM), and reinforcement learning (RL) models in a modular architecture.

DESCARTES consists of three core components: the event cloud, the datacore, and an advanced discretizer and path planner called Multi-resolution Adaptive Graphs for Intelligent Cellular Automata Route Planning (MAGICARP). The event cloud, inspired by the Air Force Research Laboratory’s Cognitively Enhanced Complex Event Processor (CECEP), uses MQTT to implement a trigger–response mechanism in which specialized topics activate event sequences defined by behavior models. The datacore functions as a behavior model that records event triggers into a local database, automatically generating new Postgres tables and entries when previously unseen event types are introduced.

MAGICARP is a multi-resolution environment discretization and path-planning framework that partitions the operational space according to local conditions while preserving global structure. Drawing on cellular automata principles, MAGICARP represents the environment as interacting hexagonal cells whose local updates shape higher-level topology. An entropy-guided refinement mechanism directs computational attention to regions of high informational change, enabling localized updates that support scalable, high-fidelity route planning in time-critical domains.

Together, these components enable DESCARTES to evaluate C2 mission plans against behavior models that provide contextual understanding, examine, feasibility, capability, and adherence to rules, doctrine, and commander intent. DESCARTES can integrate this data and replan missions, bring attention to constraint and restraint violations, and provide contextually aware decision support to C2 operators.

Decision Engine for Situational Context, Autonomous Replanning, Task Execution & Scheduling

Two central challenges in multi-agent reinforcement learning (MARL) are (i) effectively aggregating information from local observations through inter-agent communication, and (ii) improving communication efficiency by reducing unnecessary message exchanges. The solutions for these challenges are inherently conflicting to each other. While information aggregation is crucial for distributed agents to build global awareness and make effective decisions, indiscriminate aggregation often leads to excessive communication overhead. In this talk, we present a novel centralized training with de-centralized execution (CTDE) framework, BElief-based Predictive Auxiliary Learning (BEPAL) with communication reduction, to address both challenges. Compared to traditional reinforcement learning, agents in BEPAL maintain a richer and more accurate latent representations of environment dynamics with the help of the auxiliary learning task, which reconstructs a belief of the global environment from local observations and inter-agent communications. In addition, it uniquely incorporates an unlearning-based communication advantage baseline, enabling agents to learn when to suppress uninformative messages during training, thus improving communication efficiency. We evaluated BEPAL on four cooperative MARL benchmarks: Predator-Prey, Traffic Junction, Google Research Football, and Robotic Warehouse. In all environments, BEPAL consistently outperforms prior methods, achieving an average 31.33% improvement in performance and demonstrating more stable convergence. The communication unlearning stage effectively reduces communication by 22.25% with nearly no impact to the agent’s performance.

Communication Efficient Multi-Agent Systems using BEliefbased Predictive Auxiliary Learning

Cognitive overload in naval personnel is a critical operational risk, yet forecasting it remains challenging. While existing deep learning models can predict performance from physiological data, their reliance on purely statistical mappings without biological constraints makes them brittle to sensor changes and provides limited insight into the underlying neural mechanisms of fatigue or stress.

Our work introduces a new paradigm by embedding a classic model from computational neuroscience, the Wilson-Cowan equations, within a deep learning architecture. This structurally informed approach allows us to learn a personalized, low-dimensional dynamical system for each individual, governed by the physics of neural excitation and inhibition.

Our framework comprises three novel machine learning stages. First, a sensor encoder, pre-trained with an anchor-view contrastive objective, fuses heterogeneous sensor data into a modality-agnostic latent space. Second, a deviation-based variational autoencoder extracts a static trait vector by modeling each individual as a unique delta from a learned population-level baseline. Finally, a hypernetwork maps this trait to the specific parameters of a personalized Wilson–Cowan stochastic differential equation, which models the evolution of the subject’s cognitive state as a continuous trajectory.

Validation on complex synthetic systems demonstrates the model’s ability to accurately recover the system’s underlying governing dynamics from observational data. We further show its effectiveness by training on high-fidelity laboratory data and applying it to diverse, high-demand tasks analogous to operational contexts, such as simulated driving and surveillance camera monitoring. The system learns to distinguish between stable focus states and unstable overload states.

This framework provides a direct measure of neural stability. For naval applications, it enables robust fatigue monitoring using available sensors and lays the groundwork for adaptive interfaces in command and control systems that can reconfigure displays in response to an operator’s measured cognitive state, mitigating overload before performance degrades.

Neuro-Dynamic Digital Twins: Biologically-Constrained Cognitive State Modeling from Physiological Signals

SOAR (Smart Obstacle Avoidance through Reinforcement learning) presents a novel, adaptive framework for robust collision avoidance in multi-rotor aerial vehicles operating in dynamic, unpredictable environments. Traditional obstacle-avoidance systems for drones often rely on static rule sets, hand-tuned controllers, or assumptions about environmental regularity. These approaches struggle when confronted with rapidly evolving conditions such as sudden wind gusts, shifting obstacles, or the presence of other aircraft. SOAR addresses this limitation by training an end-to-end reinforcement learning (RL) agent capable of continuously adjusting its control strategies in response to real-time disturbances while maintaining adherence to a prescribed flight path.

The proposed system is developed and evaluated using the NVIDIA Isaac Sim environment, which enables high-fidelity simulation of quadcopter dynamics, sensor noise, environmental chaos, and multi-agent interactions. This platform allows the training pipeline to expose the agent to a broad distribution of challenging scenarios that would be impractical or unsafe to reproduce in real-world testing. Environmental perturbations include wind fields, transient forces, unexpected moving obstacles, and simulated aircraft exhibiting independent motion models. By integrating these diverse disturbances into training, SOAR emphasizes resilience, generalization, and situational awareness.

A key innovation of SOAR lies in its unified reward structure that jointly penalizes collisions, path deviation, and unstable control actions while positively reinforcing precise trajectory following under chaotic conditions. Rather than merely avoiding obstacles, the agent must intelligently compensate for external forces and maintain stable flight, transforming the problem from simple reactive avoidance to proactive, context-aware navigation. This encourages the emergence of control behaviors that dynamically balance safety, efficiency, and path fidelity.

The resulting system demonstrates the potential for RL-driven flight controllers to outperform conventional approaches in scenarios where environmental conditions cannot be fully predicted or modeled. SOAR contributes a scalable methodology for training robust aerial autonomy and highlights the value of simulation-rich RL as a design paradigm for next-generation drone navigation systems.

SOAR: Smart Obstacle Avoidance through Reinforcement learning

This abstract outlines a framework for developing intelligent, interpretable control systems by distilling policies from Reinforcement Learning (RL) agents into instruction-tuned Large Language Models (LLMs). We leverage navy-relevant, Gymnasium-compatible simulators, such as gym-sailing, as an initial training testbed. While this specific environment models foundational tasks such as upwind navigation and tacking, the underlying methodology is designed to be scalable and directly applicable to a wide range of more complex, high-stakes naval control systems, including autonomous underwater vehicle navigation and dynamic multi-vessel coordination. We first train an RL agent to act as an "oracle," generating an optimal dataset of state-action trajectories. This data is then used for policy distillation via supervised fine-tuning (SFT) of advanced, instruction-based LLMs, such as Llama 3.1 and Qwen3 of different sizes and quantization, transforming them into specialized reasoning agents capable of real-time, closed-loop control.

The advantage of the resulting LLM-based controller is its inherent interpretability. Unlike the opaque RL agent, the LLM can articulate a "chain-of-thought" rationale for its control decisions, such as explaining why a specific rudder angle was chosen in relation to wind direction, boat speed, and the target location. These reasoning agents offer high-level insights into the system's functioning, provide actionable strategic recommendations, and excel at generalization. They can adeptly handle dynamic priority switching in multi-objective settings—such as re-tasking from "maximum speed" to "maximum comfort" or "avoid new obstacle"—by processing these new objectives as natural language instructions, a feat unachievable by traditional RL agents without costly retraining.

This capability has direct and powerful applications for naval personnel training. By integrating these reasoning agents into simulators, trainees can engage in an interactive learning loop, querying the AI "skipper" or "tactical officer" for its strategy. This transforms the simulator from a simple practice tool into a dynamic, "explain-by-example" instructional platform, significantly accelerating skill acquisition & deep strategic understanding.

This RL-to-LLM distillation framework is not merely theoretical; it has been successfully validated in other high-stakes, complex domains. As a key proof point, we applied this method to distill policies from a sophisticated multi-agent RL controller responsible for the end-to-end liquid cooling management of HPC systems, such as the Frontier supercomputer. In that case, the distilled LLM controller consistently matched the performance of the original RL policy across key metrics. This approach can also be proven effective in a host of classic OpenAI Gym control environments, demonstrating a robust and generalizable pathway for creating high-performing, interpretable agents from optimized, black-box policies.

Interpretable Naval Control: Reinforcement Learning Policy Distillation for Reasoning-based LLM Agents

Compass

Current United States Marine Corps (USMC) marksmanship and tactical assessments rely on disparate technologies and methods that often provide incomplete human performance data and/or delayed results. Furthermore, assessments often require a one-to-one instructor-to-shooter ratio, which limits throughput for basic marksmanship assessments and impedes measurement of tactical/force-on-force training at scale. Our current program is investigating novel solutions to these issues through the pioneering application of artificial intelligence/machine learning (AI/ML) capabilities embedded within tactical computer fabrics. Sensors (acoustic and accelerometry) and microcontroller units (MCUs) are strategically placed in a malleable fiber and woven into the fabric, creating an imperceptible wearable system with fully encapsulated data collection and performance analysis and prediction. The core of our methodology lies in the application of AI/ML models, including neural networks (NNs), distributed across the fiber’s MCUs. The AI/ML models are designed to analyze the complex factors that contribute to marksmanship performance, including shot timing, precision, and shooter movement. A pilot study focused on determining shot detection, using labeled acoustic data, found that individual tactical computer fibers on both left and right arms achieved high accuracy in shot detection of blind test data. The NN models were first trained on ~800 shots across eight shooters with different postures. The NN was composed of separable convolution and dense layers with ~4000 trainable parameters, small enough to be deployed within the MCU itself. A method of combining the global efficiency of shot detection by the NN and precise peak finding through Shannon energy algorithms was also demonstrated. Live-fire testing included USMC zeroing procedures, the Infantry Marksmanship Assessment course of fire, and complex live-fire and maneuver events. Across these events, the shot timing captured by the tactical computer fibers was within >95% of the data captured by the Joint Marksmanship Assessment Package, the current gold standard for marksmanship assessment in the USMC, and the AI/ML models achieved >93% accuracy in predicting shot detection, with a low rate of false positives. The models also achieved >91% accuracy in predicting shot precision based on pre-shot shooter movement data captured via accelerometry and validated via grouping measurement (mean radius), providing deeper insight into individual shooter lethality. Future research efforts will focus on several key areas to further enhance the capabilities of the tactical computer fabric system. We plan to expand the dataset used to train the AI/ML models, incorporating data from a diverse range of shooters with varying skill levels and experience. We will also explore incorporating additional physiological data measured by the fibers to further improve the predictive capabilities of the AI/ML models.

Artificial Intelligence/Machine Learning (AI/ML)-Enabled Tactical Computer Fibers for Measurement and Evaluation of Live-Fire Shot Timing and Precision

Rapid advancements in Large Language Models (LLMs) and Transformer-based capabilities are driving increased adoption of Artificial Intelligence (AI). These capabilities have the potential to offer transformative impacts on operator effectiveness and capability development, but their increased usage presents major challenges related to risk assessment, risk mitigation, and risk tracking. These adoption difficulties are further amplified in highly-regulated and compliance-driven industries, such as defense.
At the core of these difficulties are: (i) demand for and implementation of AI capabilities outpacing regulatory guidance and contractual requirements, (ii) poor compatibility between AI-specific risks and existing non-AI risk management and tracking methodologies, and (iii) the lack of quantitative measures for risk severity and composite risk weighting.
In response, we introduce a practical AI risk framework consisting of: (i) a standardized taxonomy of ten AI risk types spanning three top-level categories (Responsibility, Security, and Accountability), providing comprehensive coverage that can be systematically mapped to existing standards and governance artifacts, and (ii) an organizational risk calibration method that leverages semi-automated, scenario-based questionnaires and structured qualitative-to-quantitative mappings to derive scoring criteria and category weights for composite risk scoring. This approach is designed to integrate with existing risk management processes in highly regulated environments and to support consistent, repeatable evaluation of AI-enabled capabilities.

Premium content

Downloads

Next from NAML 2026

Wearable Sensor Data-enabled AI Reasoning Framework for Warfighter’s Thermo-Physiological Survivability and Mission Readiness in Extreme Temperature

Similar lecture

Course of Action Optimization, Analysis, and Comparison Harness (COACH)