This PDF file contains the front matter associated with SPIE Proceedings Volume 12544, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Modernization of line-of-sight communications, with beam-forming, mesh networking, and technologies making communications less detectable and susceptible to interference and jamming in contested environments help to secure lines of communications on and on the battlefield. We present a self-adjusting drone-hosted autonomous mesh network to deliver critical data in situations not historically supported by traditional configurations of radios. We develop autonomous collaboration behaviors to maintain a dynamic and adaptive tactical mesh network that ensures high quality communications in uncertain battlefield environments. While recent advances in tactical communications (e.g., 5G, optical, etc.) allow high bandwidth, directed line-of-sight (LOS) peer to peer information sharing between both manned and unmanned assets, current high-bandwidth communications infrastructures for ground and air lack the exibility to support highly dynamic and mobile operations, often requiring manned ground stations and established, planned-out infrastructure. By hosting a tactical mesh network on a swarm of mobile unmanned platforms, we extend the communications range of the network beyond LOS and around occlusions such as terrain and urban sprawl. We achieve this through a combination of rapidly exploring random graphs to identify candidate mesh configurations, non-convex genetic optimization to refine these configurations, and distributed multi-agent control algorithms to maintain the dynamic mesh in-situ. Our solution optimizes resiliency where feasible, often allowing the network to continue to function in the presence of a dead or otherwise compromised agent.

Multi agent hybrid dynamical systems are a natural model for collaborative missions in which several steps and behaviors are required to achieve the goal of the mission. Missions are tasks featuring interacting subtasks, such as the decision of where to search, how to search, and when to transition from a search behavior to a rescue behavior. Control in hybrid systems is poorly understood. Theoretical results on state reachability rely on restrictive assumptions which hinder formal verification and optimization of such systems. Further difficulties arise if there are no a priori ordering or termination conditions on the intermediate steps and behaviors. We present a flexible framework to enable decentralized multi agent hybrid control and demonstrate its efficacy in a class of multi-region search and rescue scenarios. We also demonstrate the importance of dynamic target modeling at both levels of the hybrid state, i.e. estimating which region targets are in, how search behavior affects this estimate, and how the targets move between and within regions.

We take the first step to demonstrate the feasibility of using Modular, Extensible, Interoperable Autonomy (MEIA) to support the Internet of Military Things (IoMT) by implementing MEIA for an Aurelia drone. We set up a pipeline for autonomy development (PAD) which includes tools and processes that facilitate autonomy development. We implement both the middleware and internal data management to run our MEIA solution which includes the autonomy algorithms capable of executing various missions. We provide the results of our final demonstrations which mark the completion of this initial step to demonstrate MEIA as a viable cross-domain, autonomy architecture for IoMT.

Autonomous underwater vehicles (AUVs) are gaining increasing attention due to their promising potential for underwater multi-agent tasks in both military and civil applications. Formation control for AUVs, as the basic problem in cooperation of multiple AUVs, is gaining increasing attention due to the unique difficulties compared with formation control for surface and aerial vehicles. In this paper, we propose a formation control algorithm for underactuated AUVs while tracking a given trajectory. The proposed strategy, which leverages a leader-follower approach, is used to reduce the complexity of the control algorithm. During the formation control process, each agent tracks the same trajectory formed by pre-set waypoints with constant surge velocity by using lower-level PID controllers. Simulations are conducted to validate the proposed formation control algorithm. The AUVs’ position coordinates are fed to a distributed beamforming system to demonstrate the ability to form AUV swarm coherent beams.

Graph neural networks (GNNs) are becoming increasingly popular for multi-agent autonomy applications. For mission planning, when numerous autonomous agents are involved, even simple behaviors can overwhelm the human commander. This is because the number of possible interactions between agents and possible outcomes for coordinating effective teaming grows exponentially. Artificial intelligence can be used to learn effective agent behaviors to create what we define as a Virtual Commander (VC). The VC can be used to learn friendly and/or adversary behaviors in a complex mission planning scenario. A VC may be comprised of multiple machine learning components, such as behavior classification and behavior prediction. Based on classifying a particular behavior mode at any instant in time, future behaviors can be learned and therefore predicted to aid a commander in effective mission planning. In this work we compare 2 methods of behavior mode classification for a virtual commander using graphs. The first method uses an image-based digit (IBD) classifier for mode classification, while the second method uses adjacency matrix-based (AMB) classifier to perform mode classification. Though both classifiers predict the VC mode classification correctly <90% of the time, we show that the AMB performs better than IBD, while also reducing the data generation complexity for training and testing by multiple orders of magnitude.

In this paper, we explore the advantages and disadvantages of the traditional guidance law, proportional navigation (ProNav), in comparison to a reinforcement learning algorithm called proximal policy optimization (PPO) for the control of an autonomous agent flying to a target. Through experiments with perfect state estimation, we find that the two strategies under control constraints have their own unique benefits and tradeoffs in terms of accuracy and the resulting bounds on the reachable set of acquiring targets. Interestingly, we discover that it is the combination of the two strategies that results in the best overall performance. Lastly, we show how this policy can be extended to guide multiple agents.

When operating autonomous systems across multiple operational domains (air, surface, or undersea), the ability to develop and test existing and novel algorithms requires the support of unique tactics and communication methods. This work presents an update of technical advancements in support of unmanned, distributed, multidomain intelligence, surveillance, and reconnaissance (ISR) mission sets for increased communication and battle space awareness. Specifically, we present an ecosystem of technologies designed to execute a specific multidomain operations (MDO) concept, elements of which are prototyped in simulation while exploring development in hardware. The former (simulation) leverages an open source simulation tool, SCRIMMAGE, to prototype the range of capabilities envisioned while the latter (hardware) investigates the multifaceted aspects of what is tangibly required to implement those capabilities in a realistic environment. The hardware team incorporates multiple open source software tools and low-cost surrogates developed in-house to accomplish its goals through rapid experimentation at local testing sites.

We consider a variant of the target defense problem where a single defender is tasked to guard a target region from a sequence of incoming intruders. Each intruder’s objective is to breach the target boundary without being captured and the defender’s objective is to capture as many intruders as possible. The intruders appear sequentially on a fixed circle surrounding the target, resulting in a sequence of 1-vs-1 games between the defender and the intruders. Each 1-vs-1 game is terminated when the target is breached or the intruder is captured. The defender has to start the next game as soon as the current game ends. Each intruder knows the entry point of the last intruder and this information is used to find an optimal entry point. Each game is analyzed by dividing it into two phases: partial information and full information phase. We utilize the notions of engagement surface and capture circle to analyze the strategies for the defender as well as the intruders. Furthermore, we analytically compute the capture percentage for both finite and infinite sequences of intruder arrivals. Finally, the theoretical results are verified through numerical examples using Monte-Carlo type random trials of experiments.

Autonomous robots are ideal to explore partially known locations and minimizing the number of lives put at risk for humanitarian assistance after disastrous events. We present an alternate solution to autonomously identify a safe region in a partially-known environment where a robot will navigate to and transmit data. Our approach involves sending a navigation goal for the robot to travel to autonomously. As the robot moves, it collects important information including radio data communications (RSS), and its current location. Once complete, it identifies the highest RSS data received and travels back to starting location to transmit a 5 MB file.

In this work, we propose a method to perceive a terrain’s geometric and other surface properties for efficient motion planning in outdoor environments. Our method incorporates two perception branches to identify the terrain’s elevation and roughness separately. The first branch uses an elevation map created using LiDAR point clouds to compute a cost map that represents critical elevation changes. The second branch uses a vision-based cost map trained using RGB images, IMU, and robot odometry. Then, least-cost waypoints are calculated on a combined cost map and are followed using the Dynamic Window Approach (DWA). Our planner navigates along the least-cost waypoints while adaptively varying the velocities to reduce vibration. We evaluate our method’s performance on a Husky robot in real-world environments. We observe that our method leads to higher success rates, and lower vibrations compared to state-of-the-art methods.

The Distributed and Collaborative Intelligent Systems and Technology (DCIST) CRA aims to study the underlying science questions and develop new methods and concepts that will enable these abilities through behaviors that are not brittle and preprogrammed but rather adaptive, resilient, and learned. Thrust 1 – Multi-Robot Collaborative Autonomy handles a large team of autonomous robots acquiring information from an unknown and dynamic environment while each agent performs individual and team complex missions. In this paper, we present an overview of this thrust and experimental results about robust navigation, distributed semantic mapping and localization, multi-agent communication, decision-making and maximization of knowledge about adversarial dynamic targets.

The Tactical Behaviors for Autonomous Maneuver Collaborative Research Program (TBAM-CRP) is focused on developing and experimentally evaluating coordinated and individual behaviors for small groups of autonomous agents to learn doctrinal as well as novel tactics for maneuvering in military relevant environments. The TBAMCRP will find novel solutions to these maneuver problems in analogical simulations representing complex realistic terrain. The first two-year sprint topic is “coordinated and adversarial tactical maneuver in complex terrains”, with an operational scenario entitled “Movement to Contact”. In this scenario, contact with adversarial positions is a constant concern – in some situations this contact should be avoided through use of terrain features and cover; in other missions the adversary positions should be met with a posture of tactical overmatch through coordinated maneuver - the synchronized actions of a distributed system.

We discuss real-world challenges for multi-agent autonomy for defending high value targets using vision-based Unmanned Aerial Vehicles for detecting and intercepting adversarial intruders. Specifically, we address the defense of a hemispherical dome encapsulating high value targets. We discuss vision-based detection of intruders and the design of a control policy for pursuit and interception. We evaluate the performance of the algorithms in both the simulated environment and the real world. We extend the framework to multiple defenders and intruders using graph neural networks (GNNs). We show how GNNs can be trained on small graphs and deployed on large teams of defenders.

Patrolling is generally classified as either regular or adversarial patrolling. Regular patrolling aims to periodically visit important locations so that the duration between visits to locations is minimized. Regular patrolling, however, is typically deterministic. In the presence of an adversary which is able to observe the patrollers’ behavior before deciding on a plan for intrusion, deterministic patrolling allows an intruder to invade a location when it knows that patrollers will be elsewhere. Adversarial patrolling addresses this problem by using stochastic strategies that resist the intruder’s ability to learn and predict the patrollers’ actions.

The use of unmanned aerial vehicles (UAVs) for intelligence, surveillance, and reconnaissance (ISR) is one of the most well-known applications of the technology. UAVs performing ISR can carry out missions that would be too dangerous or otherwise too complicated for manned systems to complete. Today, explicit multi-agent information theoretic optimizations present challenges, requiring approximations and computational shortcuts to determine the expected information gain of an action. The approximation enables distributed optimization at the cost of efficiency in terms of under-utilization of certain sensors during tracking. While some techniques introduce multi-step planning to address this issue, these come at the cost of an additional computational burden. By using information theoretic metrics as reward functions in a Multi-Agent Reinforcement Learning (MARL) framework, we train RL agents to select actions which maximize expected information gain without having to estimate this quantity at runtime. This approach has potential to out-perform existing techniques, which rely on truncated estimates of expected information due to computational limitations. Our current prototype focuses on developing a cooperative learning strategy by modeling UAVs tracking down adversarial ground targets in an area of interest. Missions are expected to be highly dynamic where suboptimal conditions are likely to occur. To replicate this in our research, RL agents exist in a partial-knowledge environment where they learn to leverage various sensors and information available to complete the mission together. Our next step is to add a policy to optimize the number of UAVs needed to scan an area of interest while still tracking targets.

Adversaries or internal threats wishing to exfiltrate sensitive data may choose a method that obfuscates the fact that a transfer is even taking place by camouflaging it inside of allowable data. This technique is known as steganography, which is the practice of concealing data (such as hidden communications or sensitive data of any kind) in an existing data transfer medium, such as images or video. Video streams may be targeted more readily than other media because of the bandwidth they offer compared to single images. Current methods for mitigating the risk associated with data exfiltration through steganography sanitization through unconditional alteration video frames. However, these approaches are limited as they do not detect if a steganography attack is occurring. Deep Neural Networks (DNNs) excel at pattern recognition, presenting an opportunity to leverage them for detection of steganography. Having the capability to detect steganography could prove extremely useful, with uses such as prompting security administrators to investigate further or to immediately halt a video stream. We investigate Least Significant Bit Steganography with both encrypted and cleartext payloads and show that machine learning can detect both.

Aerial video recognition is challenging due to various factors. Prior work on action recognition imposes constraints in terms of unavailability of object detection bounding box ground-truth inhibiting the application of localization models and computational constraints preventing the usage of expensive space-time self-attention. Optical flow and pretrained models for detecting human actor performing action do not work too well due to domain gap issues. Our contributions1, 2 are as follows: 1. We present a frequency-domain space-time attention method that encapsulates long-range space-time dependencies by emulating the weighted outer product in the frequency domain. 2. We present a frequency-based object background disentanglement method to inherently separate out the moving human actor from the background. 3. We present a mathematical model for static salient regions and an identity loss function to learn disentangled features in a differentiable manner.

We introduce a new method of building digital reference architectures (RA) using modular, reusable elements within the RA itself. Modular RAs using this method define the domain of interest via reusable elements that are used as building blocks for downstream system developments. This method eliminates the reliance on static documentation and enables the building of tailored systems aligned to current RA objectives and content. An advantage of this method is that users interact with the RA without the need for software licenses or architecture expertise. This paper describes the underlying modular RA design and tools developed for user interaction.

The development of modular open system architecture (MOSA) approaches for Positioning, Navigation and Timing (PNT) systems is a critical need. In this paper we present pntOS, an open architecture for building PNT estimators that integrates seamlessly with other open architectures targeting related problems. The pntOS MOSA is designed to allow isolated development of plugins for navigation filter algorithms, sensor integration strategies, integrity approaches, and network buses. Critically, a developer who is a domain expert in a particular field (e.g. radar signal processing) can write a plugin incorporating their field of knowledge into pntOS in isolation, without needing to understand any other part of pntOS. Once a set of plugins has been developed, any particular plugin can be easily swapped with another plugin, allowing for rapid trade-space analysis and reconfiguration of systems at minimal cost. The pntOS MOSA is designed to support integration into embedded real-time systems, post-processed analysis, and Monte-Carlo simulations.

As dependence on GPS-based PNT services increases, there is an increasing need to enable PNT systems that complement GPS with other sensors to endure in conditions where GPS is degraded or unavailable. With the plethora of complementary sensors and their varying trade-offs, a modular and adaptable system structure is the most sustainable approach. A key enabler to allow this type of development is a PNT information data standard. ASPN is a data standard that describes what PNT data should be exchanged for consistent usage and interoperability of PNT estimators across different systems, sources, and users. Under the current version, ASPN 2023 defines the message content of 53 PNT sensors to enable more agile development and usage of PNT systems that use complementary sources. Furthermore, ASPN 2023 is communitydeveloped with 162 participants across over 50 organizations across the US government, industry, and academia. ASPN 2023 hosts an ecosystem for sharing the data model, creating experimental versions of the data model, as well as contributing community-developed tools and solutions.

This paper presents a video surveillance-based face image security system that utilizes post-quantum cryptography algorithm for enhanced security. The system accurately extracts face images from the surveillance video using the multi-task cascaded convolutional neural network (MTCNN) and subsequently encrypts them using the ring learning with errors (ring-LWE) algorithm. The encrypted facial images are then transmitted to a remote server for secure storage. The proposed system is implemented on a central processing unit (CPU) platform and its operations are accelerated using a graphics processing unit (GPU). The evaluation results on an NVIDIA GeForce RTX 3090 Ti GPU for videos from the standard reference database of National Institute of Science and Technology (NIST) show that the proposed system can perform face image encryption and decryption as fast as 1.33 ms and 0.51 ms on GPU platform for parameter set n = 256 and q = 7681, respectively. Moreover, the analysis of security parameters such as histogram, correlation coefficients, and entropy proves that the proposed system outperforms its predecessors in terms of confidentiality. The proposed system is also assessed using various security parameter sets to showcase its compatibility with diverse computational resources.

The Ground Troop Formation Identification research demonstrates how Machine Learning (ML) can be employed to classify ground force formations from the mass of individual observations made by local sensors and tactical information received from the connected battlespace at 93.75% accuracy at inference. This research examined suitable Machine Learning options, resulting in the development of a Random Forest (RF) algorithm software solution that was then integrated with a representative airborne mission system environment consisting of a Data Link Processor (DLP) and a Tactical HMI. This allowed a more realistic testing of how it could perform in a real world environment. This research displayed the results within a platforms tactical HMI for clear presentation. This system would aid an already burdened operator by automatically performing the complicated task of quick and effective Tactical Situational Awareness (SA) analysis, securing operational advantage through improved speed, accuracy and quality of decision making.

This Weapon Package Selection Innovation Initiatives Internal (I3) Project demonstrated how a chosen Artificial Intelligence (AI) technology could assist a human operator in the decision making process of selecting weapons carried by a complex mix of manned and unmanned platforms to targets, in highly dynamic scenarios. This was an innovative application of AI technology to complex decision making in a challenging new environment, where more traditional approaches may severely impact the operational effectiveness of new weapon delivery capabilities. The project developed two Machine Learning (ML) technologies, a subset of AI, for initial comparison - a Neural Network and a Random Forest. It was concluded that both AI technologies could be used to assist a human operator in the decision making process of weapon selection. The project has also provided a springboard for future research opportunities within General Dynamics Mission Systems UK (GDMS-UK).

During the summer of 2022, the United States Military Academy hosted a robotics apprenticeship program during which interns programmed maritime robots to move autonomously. By the end of the apprenticeship, the robots were able to compete against each other in a force-on-force competition based on capture-the-flag game rules. This game mimics tactics performed in military operations and is used for studying new military tactics and operations involving humans and robots working with each other and in human-robot teams. The live tests were planned using Monterey Phoenix,1 a Navy-developed language, approach and tool for behavior modeling. Monterey Phoenix was used to generate a set of possible scenarios that could occur during the robot competition based on the game rules and expected conditions. Scenario variants containing search and rescue (SAR) operations were included to help with planning in case of a man overboard or robot malfunction. The analysis in Monterey Phoenix led to the exposure of some new SAR scenario variants that were not previously considered, including: the weather being unsafe but shoreside permitting continued play; two men overboard events happening simultaneously on both teams; SAR is deployed despite not being signaled to; and lastly one team has a man overboard event but the other team is unaware and continues playing the game. Having a library of these and other possible scenario variants helped the team consider possible causes for their occurrence and avoid or mitigate unwanted outcomes that could arise should they occur during live competition.

The Aquaticus capture-the-flag (CTF) game1, 2 was created several years ago at the Massachusetts Institute of Technology (MIT) as a testbed for human-robot teaming experiments utilizing maritime robots controlled by the MOOS-IvP autonomy stack3, 4 and humans in motorized kayaks. Monterey Phoenix (MP)5 is an open-source program for scope-complete behavior modeling created at the Naval Postgraduate School. Our objective in the work presented in this paper is to create an interface between MP logical (discrete-event) behavior models, which are text script files, and the MOOS-IvP autonomy stack used for investigating continuous-time maritime robot autonomy, including complex behaviors such as automatic collision avoidance. These physical robot behaviors are achieved by the MOOS-IvP application by generating an appropriate stream of Helm behaviors at a repetition rate of four commands a second (speed and azimuth values) by approximate solution of nonlinear differential equations. We developed a tool that extracts relevant data from the Monterey Phoenix model in JSON format and then generates an IvP Helm behavior tree, allowing for simulations of robot behavior in the MOOS-IvP environment. The results of this research can be expanded upon in the future to allow interfacing between Monterey Phoenix and other simulation environments as well as to be used in reinforcement learning environments to generate AI/ML-enabled agents to either play the CTF game in a virtual environment or to be downloaded into robots to play the CTF game live at the multi-domain operations (MDO) Human-Robot Teaming Sandbox (MDO-HuRT-S) at the United States Military Academy.

As warfare looks to the future and the need for the internet of military things (IoMT) grows, we discuss how autonomy fits into this paradigm. We define common terms relating to autonomy to promote common understanding between autonomy developers, and we analyze a variety of autonomy architectures, examining what they do correctly to support IoMT and where they fall short. We discuss our general philosophy concerning autonomy – that it must be multi-layered to be effective – and provide an overview for our Modular, Extensible, Interoperable Autonomy architecture that supports IoMT and the future of warfare.

Experiences from recent conflicts show the strong need for providing timely a precise situation picture in order to improve the situational awareness of commanders. Situational Awareness requires comprehensive information delivered at the right time and at the right place, which must be adapted to the user, his role and current task; it must be conforming to the communication and visualisation devices currently in use. Therefore, a lot of sensor data and corresponding information have to be considered, exploited and combined in a flexible way. Smart sensor suites comprising different multi-spectral imaging sensors as core elements as well as additional non-imaging sensors may contribute decisively to the needed complete situation picture. The smart sensor suites should be part of a smart sensor network – a network of sensors, databases, evaluation stations and user terminals. Its goal is to optimize the use of various information sources for military operations like situation assessment, intelligence, reconnaissance, target recognition and tracking. Such a smart sensor network will enable commanders to achieve higher levels of situational awareness. Such a smart sensor network will enable commanders to achieve higher levels of situational awareness based on increased flexibility in using combined smart sensors. This paper presents a prototype of an Open System Architecture based on a system-of-systems approach. The open system architecture enables combining different sensors in multiple physical configurations, like distributed sensors, co-located sensors combined in a single package, sensors mounted on a tower, sensors integrated in a mobile platform, and use of trigger sensors. The mode of operation is adaptable to a series of scenarios with respect to relevant objects of interest, activities to be observed, available transmission bandwidth, etc.

The INTERACT Project, funded by the DG-DEFIS of the European Commission and managed by the European Defence Agency (EDA), aims at enhancing the capabilities of European armed forces to safely, effectively and flexibly operate unmanned and manned systems in joint or combined operations. The challenge to achieve this lies in creating overarching interoperability concepts for defence systems in general and unmanned systems in particular. INTERACT proposes to use selected NATO STANAGs to engender compatibility for military systems. But the lack of a promulgated STANAG for UxS (Unmanned Systems) control in an all-domain context is identified as a major gap regarding this endeavour. As a response the INTERACT project is elaborating a set of interoperability concepts and standardisation proposals, which will enable the coordinated deployment of multiple and potential heterogeneous platforms by a single, standardised control station as well as the controlled hand-over of platforms between INTERACT compliant control nodes. The INTERACT solution creates a holistic approach and includes the proposal for concepts and design of a set of interoperable standardized interfaces between subsystems and payloads within an unmanned system (intra-system interoperability) to ease the upgrade and adoption of novel payloads and maintaining and upgrading equipment and components to the state-of-the-art, as well as the proposal for inter-system interface standardization in order to pave the way for future operational concepts where autonomous assets will flexibly operate together in organized heterogeneous UxS teams. Beneath the system interoperability INTERACT will also address the human-machine interaction by proposing a common design solution for standardisable user interfaces. The INTERACT consortium consists of 4 major European RTOs as a core team supported by a strong alliance of 15 representative European defence industries, SMEs and RTOs from 11 different nations.

An autonomous robot’s onboard computational resources are limited due to size, weight, and power limits. Thus, optimizing the use of those resources is essential so that the robot can complete its mission reliably and efficiently. Metareasoning, a branch of artificial intelligence, enables a robot to monitor and control its perception, mapping, planning, and other reasoning processes in response to changes in the robot and its environment. Metareasoning is implemented in a meta-level that is logically separate from the object level (which performs the reasoning processes). Previous work has developed metareasoning approaches for specific robotic systems, but these are not easy to generalize. This paper describes our implementation of a metareasoning approach in the Army Research Laboratory (ARL) ground autonomy stack, which is deployable on a variety of robotic platforms. This paper describes the general approach and our implementation of a metareasoning node that can switch the global and local path planners when a planning failure occurs. The results of simulated experiments show that adding metareasoning increases the likelihood of mission success in some cases. More research is needed to optimize the metareasoning approach. Ultimately comprehensive metareasoning that can control the most important aspects of object level reasoning will enable an autonomous robot to deploy its limited computational resources more effectively and complete its mission more reliably.

This work demonstrates that a novice user of the Monterey Phoenix (MP) tool can encode their expertise of robot mission tasks in the form of a finite state machine. State machine diagrams are among several approaches available for describing behavior. They are appropriate for capturing and understanding behavior rules of systems that can be described in terms of possible states and state transitions. Tools that offer state machine automation enable modelers to test the activation of states and trace the transitions between states, providing assistance with verification and validation of the behavior logic. The United States Military Academy (USMA) at West Point’s Robotics Research Center (RRC) in collaboration with the Naval Postgraduate School (NPS), developed a minimal executable state machine model to test its implementation of behavior logic for a single robot agent in Project Aquaticus, a human-robot teaming capture the flag competition. Monterey Phoenix, a formal behavior modeling language, approach, and tool was used to express the behavior logic of the robots participating in this game as a finite state machine model. MP was used to generate every possible trace through the state machine behavior logic up to the specified scope limit. This model informed the team’s understanding of the desired states for the single robot competitor in Aquaticus, assisted in identifying behavior improvements, and identifying dependencies among behaviors and where the process could possibly go wrong. In essence, this work demonstrates that encoding the finite state machine of the single robot tasks in MP allows experts to verify and validate expected robot behaviors and adjust as needed. Future work will look to expand from the single agent modeling to multi-agent modeling in MP to generate and verify possible sequences of events for the Aquaticus competition between competing teams. By using this modeling method to understand the state of each actor and their response to inputs from other actors, the team expects to improve the communication within the human-robot team and ultimately achieve a better understanding of command intent in complex, contested environments. Today’s mission-critical systems are actually, “systems of systems” with complexity that will surpass the designer’s cognitive ability to anticipate all interactions. Without tools, languages, and methodologies to analyze system-level behavior early in the design, mission systems can be at risk of emergent behaviors that may negatively impact safety and security. Our team believes that the MP environment can assist commanders in anticipating potentially unsafe or insecure emergent behaviors of human-robot teams in complex, contested environments and adjust human intent/command intent accordingly for allowable combat crew drill behaviors required to achieve assigned tasks and missions.

The performance of decision-making algorithmic approaches depends on a vast number of factors, including hyperparameters, which make some solutions difficult to find. In our previous work (MARDOC paradigm),1 we have found that even in a simple environment2 (i.e., a small map with few obstacles), merely changing the initial conditions or doctrinal guided policy (MARDOC) showed a significant impact on the converged behavior. Further, we found that not all policies were useful or desirable for military applications. In this paper, we focus on a complex environment (i.e., a larger map with a greater number of heterogeneous assets and a stronger adversarial force) to analyze the impact of different doctrinal control parameters on the performance and behavior of fixed doctrinal policies. Especially we prioritize the Red force assets for targeted maneuvers and attacks. We hypothesize that asset type and their corresponding coordination will have a significant impact on the performance of the Blue force. Our preliminary experiments in this complex environment showed that the performance varies tremendously depending on asset capability and coordination between teams.