Implementation Considerations

Authors: Lieutenant Colonel Joseph West, Joshua C Keene, Lieutenant Colonel Kate Tollenaar and Priyam Dalmia

Abstract

In future warfare, operational advantage will favour those who adapt more quickly. As robotic and autonomous systems and artificial intelligence (RASAI) become integral to the battlespace, their capacity to evolve must extend beyond initial design, and adaptability must become a deliberate capability. This paper defines RASAI capabilities that can be modified and adapted between engagements to address a changing environment or to exploit or counter a rapid advancement in technology. It argues that adaptability is no longer solely a human trait but a required characteristic of intelligent machines operating in dynamic conditions. The paper distinguishes tactically adaptable RASAI (TAR) from related ideas like flexibility or robustness, outlines key enablers (modular hardware, updatable models, embedded technical roles) and analyses institutional barriers preventing adaptation at the tactical edge. It presents international case studies, identifies current Australian Defence Force (ADF) gaps, and proposes a set of structural, technical and doctrinal reforms, including raising a deployable TAR support element, to operationalise TAR as a core capability within Army’s human–machine force structure. While the findings are relevant to all domains, we focus primarily on the land domain.

Introduction

In war, operational advantage favours the side that adapts more quickly;^[1] this principle has held across conflicts, from early 20th century campaigns to modern multi-domain operations. In an environment of agile adversaries and rapid technological change, adaptability is not just desirable; it is essential. The Australian Army recognises the importance of adaptability. Land Warfare Doctrine (LWD) 1 describes a credible land force as ‘adaptive and relevant’.^[2] The ‘Adaptive Campaigning’ framework reinforces it, embedding continuous learning and adjustment as core principles,^[3] but adaptability is no longer just a human trait. As robotic and autonomous systems and artificial intelligence (RASAI) become embedded in Army formations,^[4] their capacity to be rapidly modified in response to recent operational experience must be treated as a core requirement. Tactical adaptability must be designed into RAS from the outset, creating tactically adaptable RASAI (TAR).

The faster the adaptation, the greater the tactical advantage and, while dynamic adaptation during contact may be an ambition, it is unrealistic with current technology. A more immediate and practical opportunity lies in enabling adaptation of RASAI between engagements by integrating new observations, reconfiguring payloads, or updating control parameters based on lessons from previous missions.^[5]

Traditionally, adaptability—the ability to improvise under fire and adapt to new situations—has been seen as a critical leadership trait. However, we argue that tactical adaptability now extends to machines and must become an institutional capability, enabled by doctrine, modular hardware, updatable software, and human–machine workflows. RASAI systems must support rapid adjustment in configuration, control logic, and tactical function between deployments. In a forthcoming Australian Army Research Centre Occasional Paper related to this study, we define TAR as ‘A RASAI capability that can be modified and adapted between engagements to address a changing environment^[6] or to exploit or counter a rapid advancement in technology’.^[7]

This article considers tactical adaptability as it applies to RASAI and distinguishes it from related concepts such as flexibility, robustness and resilience. It explores key enablers, including mission command architecture, software modularity, model retraining, and human-in-the-loop structures, and provides examples from Australian and allied forces, including intelligence, surveillance and reconnaissance detachments, RAS teams, and logistics elements. It also identifies institutional barriers that currently inhibit this capability across Defence.

Two options exist to obtain TAR: one is to require each RASAI capability to be tactically adaptable on its own. The other is to raise tactical adaptability as a capability for RASAI systems, which could be achieved through a TAR support element—a specialist function responsible for enabling force-wide RASAI adaptation. This article considers the latter. Its goal is to provide a doctrinally grounded and operationally useful roadmap for engineering tactical adaptability into the emerging human–machine force structure of the Australian Defence Force (ADF).

RASAI do not yet dominate the battlespace; however, the prevailing view is almost unanimous that they will.^{[8],[9],[10],[11],[12],[13],[14],[15],[16],[17],[18]} The urgency of the recommendations made in this article is predicated on the assessment of when that will occur. This being the case, the ADF must act to integrate tactical adaptability as a core capability in the design, procurement and operation of RASAI systems. The recommendations presented here aim to ensure that the ADF prepares not only to meet the challenges posed by these emerging technologies but also to exploit the full potential of RASAI systems in future operations. By embedding adaptability into the force structure, the ADF will be equipped to remain competitive, agile and capable in the face of rapidly evolving technological and operational environments.

Artificial Intelligence

‘Artificial Intelligence (AI) is the study of agents that receive percepts from the environment and perform actions.’^[19] As such, it is a key component in RASAI capabilities. Notably, AI systems typically make decisions by learning from previous data to identify, and act on, important trends and patterns. One of the predominant AI training methods is supervised learning, which uses exemplar data consisting of labelled input–output pairs and applies an optimisation algorithm to minimise the prediction error. This approach assumes the training data is both comprehensive and representative of the deployment environment. In real-world Defence contexts, that assumption rarely holds. Operational data may be sparse, degraded or deliberately adversarial, making model performance fragile and transient. To remain effective, AI-enabled systems must support regular updates, including model retraining and parameter adjustments, based on recent mission data. Defence must assume that current models and methods will be periodically outperformed and design its systems with the expectation of continual algorithmic improvement. This requirement is amplified by the rapid pace of AI advancement. In 2024, over 7,000 new AI research contributions were accepted in just the two pre-eminent AI conferences,^[20],[21] and these are typically innovations considered the most profound.

In addition to the requirement for continuous model updates, there is also a real risk that new AI methods will emerge unexpectedly, causing abrupt and significant leaps in performance, rendering previous approaches obsolete. These advances can arise suddenly, often from what appear to be incremental or minor methodological changes, but they can be transformational in outcomes.

For example, while ChatGPT appeared to emerge suddenly, the core techniques—transformer architectures, reinforcement learning from human feedback, and large-scale pretraining—were already well-established methods.^[22] It was the careful integration and optimisation of these components, combined with unprecedented volumes of training data and compute resources, that delivered a language capability that surpassed expectations and redefined industry benchmarks. This example highlights the nonlinear nature of AI progress, where a modest change can lead to disproportionate gains in performance. Importantly, such advancements may originate from a single researcher or small team and are often software-only improvements that can be deployed almost immediately. This underscores the importance not only of ensuring continuous and rapid model update cycles but also of building and retaining the technical capability and infrastructure to discover, adopt and exploit major breakthroughs before adversaries do—or shortly after. It also points to the need to standardise and modularise onboard compute across platforms. This allows AI developers to work within known resource constraints when designing or replacing models, supports modular scaling where additional compute is justified, and reduces uncertainty during field deployment of new methods.

These characteristics of AI mean that adaptability is no longer solely a human trait: it must now be treated as a design requirement for AI-enabled systems operating in the tactical environment. For RASAI platforms, this includes system-level changes like updating machine learning (ML) models, reconfiguring modular payloads, adjusting autonomy parameters, and modifying control logic between missions. Such adaptations must occur on tactical timelines and reflect direct lessons from recent operational experience; therefore TAR is a requirement for any system dependent on AI models.

Defining Tactical Adaptability

Tactical adaptability is the ability of a military unit to modify its tactics, techniques or procedures (TTPs) quickly in response to changing battlefield conditions.^[23] Tactical adaptability operates on timescales relevant to the tactical battle, typically short—either between missions or in phases of a single operation.^[24] At its core, it shortens the feedback loop from observation to action—recognising what has changed and implementing adjustments in the next engagement, identifying what is not working or what the enemy is doing differently, and rapidly adjusting to retain the advantage in the next encounter.

Tactical adaptability differs from several related concepts:^[25],[26]

• Flexibility—the ability to perform multiple predefined roles. A multi-role aircraft or a cross-trained soldier is flexible, capable of selecting from established options. Flexibility is static: it draws from a fixed menu. Adaptability involves creating or modifying new approaches, not just choosing among existing ones.
• Robustness—the ability to function despite a range of disturbances or unexpected stresses. An armoured vehicle that withstands repeated impacts is robust. While robustness supports adaptability by providing survivability, it does not entail tactical change.
• Resilience—the ability to recover after disruption, damage or failure. A resilient unit can regroup and return to the fight. But resilience alone may simply restore the status quo; unless combined with adaptation, it risks repeating failure.

Adaptability requires changes in response to the environment—not merely selecting from existing options. While tactical adaptability can be achieved through flexible systems, this is only the case when the change is driven by observed environmental or adversarial changes.

Traditionally, adaptability might involve fitting field-expedient armour, rerouting convoys based on fresh intelligence, or reorganising teams to address an unexpected gap. In the context of RASAI, this could mean updating a drone’s image recognition model overnight to identify a new enemy uniform by morning. This is not pre-planned flexibility; it is tactical adaptation.

Crucially, tactical adaptability is based on observations in timeframes meaningful to the tactical element. It allows forces, and the systems they employ, to evolve during a campaign, incorporating new information to improve their next encounter. Tactical overmatch is guaranteed not by static capability but instead by frequent and fast improvement between contacts. It is this ability to evolve under pressure that distinguishes a reactive force from an adaptive one.^[27],[28]

From Mindset to Capability

Adaptability has often been discussed as a human quality: the ingenuity of soldiers and commanders to make decisions under pressure. The Australian Army prides itself on a culture of improvisation and initiative, embodied in the doctrine of mission command. Mission command decentralises decision-making, giving subordinates the freedom to execute the commander’s intent in the way they see as best, given unfolding reality. This approach enables on-the-spot tactical adjustments. However, while mission command encourages adaptive behaviour, it does not authorise units to adapt their capabilities, such as modifying hardware or software, within the same timeframe. In other words, soldiers are empowered to be adaptive, but not to adapt their equipment.

To fully harness tactical adaptability, it must be considered a capability, not just a trait of individuals. In military terms, a capability can be developed through the Fundamental Inputs to Capability (FIC) as outlined in the One Defence Capability System:^[29] organisation, command and management, personnel, collective training, major systems, facilities and training areas, supplies, support, industry, and data. While some FIC support adaptability:

• mission command within organisation and command and management promotes decentralised decision-making, and Army’s Adaptive Warfare Branch institutionalises a human adaptation cycle
• collective training deliberately stresses forces under uncertainty
• facilities and training areas provide the environments for adaptation, though they are not themselves adaptive
• Defence’s growing emphasis on data as a strategic asset underpins adaptability in AI-enabled systems.

Other FIC elements, however, are less aligned:

• Major systems remain rigid rather than modular.
• Supplies and support are geared to efficiency rather than agility.
• Industry engagement often favours long-cycle acquisition rather than rapid adaptation.

Recognising these gaps is essential if tactical adaptability is to be treated as a fully developed capability.

Australian doctrine already points to the importance of adaptation. LWD 1: The Fundamentals of Land Power identifies adaptability as essential for land forces operating in dynamic, complex environments. The Defence Strategic Review 2023^[30] and Army’s Future Land Operating Concept^[31] similarly describe forces that can ‘adjust rapidly’ to change. Current doctrine articulates the need for adaptive action but offers no specific framework for tactical adaptation, particularly when it comes to adapting technology and materiel during operations.^[32]

Bridging this current gap will require tactical units having the ability and authority to implement deliberate system changes between engagements. This would elevate capability adaptation from ad hoc practices, which are born of necessity, to institutionalised procedures that are repeatable, supportable and scalable. A soldier adding body-armour plates to a vehicle to account for increased enemy heavy weapons is showing initiative, while a unit with access to modular kits and embedded technical specialists to harden vehicles in theatre based on observed enemy threat has formal adaptability. The latter reflects a repeatable and scalable force function, not just individual ingenuity.

A relevant example is software modification in combat systems. An adaptive approach treats software, particularly AI models in surveillance drones, targeting systems or navigation, as dynamic and updatable in the field. If an algorithm fails to detect a new threat signature, an adaptable force can refine or retrain it between missions via reach-back to a lab or through forward-deployed tech teams. The Army Robotic and Autonomous Systems (RAS) Strategy recognises this, noting that future autonomous capabilities must be ‘intelligent and adaptable’ and that modernisation must enable ‘rapid updates to enhance or replace capabilities in service’.^[33] The advantages offered by RASAI systems are not enduring; their continued utility increasingly depends on how quickly and precisely they can be modified to recognise changes. Electronic warfare air platforms already operate this way, with mission data and threat libraries updated in or near theatre to account for new radar signatures.^[34],[35] This precedent reinforces the feasibility of capability adaptation and provides a baseline methodology for extending similar practices to RASAI.

Case Studies in Tactical Adaptation

Below we briefly outline three tactical adaptability scenarios. These examples are drawn from allied experience and show how local changes delivered operational benefits in compressed timeframes.

AI Evasion Tactics:^[36],[37] US Marines successfully bypassed an AI-enabled surveillance system during a training exercise by performing unexpected movements such as somersaulting, hiding under cardboard boxes and walking in single file. The soldiers’ behaviour was not within the initial AI training dataset; therefore the computer vision system did not recognise their behaviour as human. While these actions are extreme, enemies’ adaptations could be much more subtle as they explore their own counter-AI adaptations. If this occurred during operations, the data related to these missed detections would need to be captured and subsequently used to retrain the model before the next engagement. While the Marines defeated the system on day one, they would not have had the same success on subsequent days if the system had been TAR, as the model would have been retrained using the new data. RASAI systems must not only be able to detect threats but also be capable of adaptation in response to changes in the environment or adversary behaviour.

Countering New Camouflage Tactics:^[38],[39] In mid-2022, Ukrainian reconnaissance units deployed small drones which were able to identify Russian forces on the ground. In response to this, Russian forces deployed new camouflage and decoy techniques that confused the drones’ object recognition systems. Ukrainian soldiers collaborated with remote volunteer tech teams to retrain the AI model using fresh imagery of the camouflage patterns. Within days, a software patch was issued to fielded drones, which restored their ability to detect enemy positions. This case demonstrates real-world tactical AI adaptation, where soldiers equipped with the right tools became ad hoc systems integrators, modifying fielded capabilities in near real time. While demonstrating the need for TAR, ad hoc adaptation in this manner introduces a number of risks. The lack of a technical regulatory control or a TAR equivalent to electrical and mechanical engineering instructions (EMEIs) could mean that unit-sourced modifications occur without the necessary quality assurance mechanisms. Additionally soldiers independently engaging local tech teams could result in inconsistent configurations between systems, undermining interoperability. A capability approach to TAR is needed to ensure modification of systems does not weaken the force, much as the Army’s Technical Regulatory Framework and EMEIs assure the safety and effectiveness of materiel modifications.

Rapid Reconfiguration to Defeat Jamming:^[40],[41] During a NATO electronic warfare exercise, a robotic vehicle unit encountered jamming that disabled its primary radio control links. Prior to the mission planned for the next day, technicians reconfigured the vehicles overnight by installing alternative communications modules operating on different frequencies and including an optical data link currently immune to jamming. These platforms were modular and could be rapidly swapped out to adapt to the adversary’s changing behaviour. This case illustrates how modular hardware, supported by technically capable personnel and empowered operational authority, enables tactical adaptability at the system level.

These cases illustrate the practical need for tactical adaptability. The adaptations applied in all of these examples, however, are local, without an organisational framework in place. Local changes without clear organisational guidelines introduce unmanaged technical regulatory framework risks. This suggests that a more controlled and deliberate process for TAR is required.

The next sections examine what enables such adaptations (enablers), what currently hinders them (gaps) and how tactical adaptability can be embedded as a formal warfighting capability.

Enablers of TAR

TAR depends on deliberate design choices, support structures and operational permissions that allow systems to be modified between engagements. This section identifies the key enablers that turn RASAI platforms from fixed-function assets into adaptive tactical systems.

Flexibility—Modular, Reconfigurable Hardware

Hardware modularity enables physical reconfiguration of RAS platforms in response to new mission requirements or operational insights. Platforms that support standardised payload interfaces, hot-swappable sensors, and plug-and-play effectors can be adapted in the field without depot-level intervention. For example, an unmanned ground vehicle (UGV) designed with open payload architecture can rapidly switch from a signals intercept role to a logistics mule role by replacing sensor and power modules. This capacity depends not just on mechanical compatibility but on software abstraction layers that allow new payloads to be recognised and integrated with minimal downtime. While this example is more aligned with flexibility as opposed to adaptability, advances in sensor and payload technology mean that a flexible, modular design provides the foundation for RASAI hardware adaptation between tactical encounters.

Defence RAS strategies^{[42],[43],[44]} highlight a shift towards open digital backbones and common standards, critical if future platforms are to support modular TAR. Without this design philosophy, tactical elements are locked into static configurations, limiting their ability to respond to newly observed threats. This places a responsibility on the industry FIC to deliver open architectures and modular standards, rather than proprietary or closed designs, if TAR is to be realised in practice.

Field-Updatable Software and Machine Learning Models

RASAI systems increasingly depend on software-defined functions, especially those involving computer vision, navigation, sensor fusion, and autonomous control. To be adaptable, these systems must support rapid update cycles, ideally within 24 to 48 hours of new observations to be tactically relevant. This includes both rule-based logic (e.g. updating mission parameters or control routines) and ML-based models (e.g. retraining a detection network to account for a new enemy uniform or drone signature).^[45],[46] Effective TAR requires mechanisms to inject updated models onto deployed devices, ideally through secure, low-bandwidth digital delivery. In-theatre retraining or re-parameterisation is feasible if forward-deployed units include technical specialists or if systems support modular inference pipelines where new models can be uploaded without full redeployment. Critical to achieving confidence in this process are the acquisition and certification frameworks, falling under the FIC of organisation and command and management, which allow authorised units to make bounded software changes without breaching compliance regimes. Additionally the collective training FIC must prepare units to work with systems that change between missions, and command and management must enable authorisation to apply these changes.

Intent-Driven Control and Mission Logic

Mission command for humans is well understood; however, for RASAI it translates to systems that accept intent-level commands rather than rigid scripts. Systems that support configurable task graphs, goal reweighting, or bounded autonomy parameters allow for tactical adaptation without requiring complete redesign. For example, a robot tasked with route clearance might be instructed to prioritise speed over stealth based on new threat assessments. Likewise, the sensitivity of detecting adversaries might increase during higher risk periods. If the system architecture supports the injection of these intent changes between missions, whether via a control interface, mission script or application programming interface (API), it becomes tactically adaptable.^[47],[48]

This requires that TAR platforms be designed with an expectation of operator parameter selectability. Systems that only accept pre-certified mission templates or require developer-side modification are inherently fragile. Consider a counter uncrewed aerial system (UAS) jammer that cannot have its jamming band or power settings changed by the operator in theatre. If the adversary shifts their drone to an off-band frequency or reduces radar cross-section (RCS) and the jammer is fixed to pre-certified settings, the jammer is rendered ineffective, and hostile UAS will penetrate defended airspace.

Tactical Data Feedback and Model Update Infrastructure

Data collected by RASAI systems during operations—video, sensor logs, mission outcomes—must be extractable, processable and usable to inform adaptation across the battlespace. This requires onboard data logging architectures, compression and transmission standards, and back-end tools or personnel that allow for rapid triage and incorporation into retraining cycles.^[49]

Equally important is the ability to push model updates or software patches back to the system rapidly. This cycle—data extraction, insight generation, model refinement, redeployment—must be shortened to days, not months. Systems that lack bandwidth-efficient update paths or store model parameters in ways that prevent local modification are not TAR compliant in practice, regardless of their stated autonomy.

Embedded Technical Personnel and RASAI Support Roles

Embedding roles with the technical competence to modify systems in theatre is critical to maintaining the quality of and trust in AI systems.

At a minimum, force elements employing RASAI systems should include:

• operators trained to recognise when system adaptation is required
• technicians able to replace or integrate modular hardware
• digital specialists capable of deploying new models or software builds
• liaison roles to coordinate with reach-back support elements.

This distribution of roles aligns closely with Scholtz’s (2003)^[50] framework for human–robot interaction, which highlights supervision, intervention, adjustment, repair and coordination as essential human functions in robotic operations. The current Army structure does not routinely embed this blend of roles at the tactical level. For TAR to be viable, these support capabilities must be normalised, whether as attached specialists or organically trained within combat teams.

Authorisation and Institutional Leeway

Tactical adaptability requires authority to modify systems at the point of need. Current governance models, based on centralised control of system modifications, are incompatible with TAR. A revised framework must enable bounded, auditable changes by authorised users in theatre. This may include pre-approved model update parameters, sandboxed execution modes for isolated testing, or the establishment of tactical adaptability detachments with modification authority. Without such mechanisms, RASAI platforms remain static, constrained by approval delays while adversaries continue to adapt.

Gaps in the ADF’s TAR Today

While the ADF recognises the need for agility, its current structures and processes remain misaligned with the requirements of TAR. RASAI systems are not static platforms; they are software driven and data dependent, and should be modular by design. However, five critical gaps prevent them from being adaptable in practice:

• No Formal TAR Teams or Structures: The ADF lacks any dedicated force element tasked with in-theatre modification of RAS. Existing trades are not structured or trained to perform daily software changes, reconfigure AI pipelines, or implement hardware substitutions on deployed autonomous systems. Updates must typically route through original equipment manufacturers (OEMs) or acquisition channels, delaying responsiveness. Without a defined unit or role empowered to adapt RASAI systems between missions, TAR remains conceptual rather than executable. The ability to change a system must be delivered with it. A useful precedent exists in the Joint Electronic Warfare Operational Support Unit, which routinely generates and distributes new mission data files to update deployed air platforms in response to emerging threats. This demonstrates the feasibility of embedding adaptation as an institutional function. By abstracting this model, the ADF could establish a Joint RASAI Operational Support Unit to deliver equivalent responsiveness for autonomous systems, providing modular payload integration, bounded software changes, and rapid model updates in theatre.
• Absence of Uniformed Technical Specialists in RASAI: There is no organic technical capability to support and create updates for fielded RASAI systems. Roles such as AI model handlers, ML engineers, digital integrators or data engineers do not exist in the ADF’s force structure. While Defence Science and Technology Group (DSTG) involvement in supporting ADF operations is well established, DSTG does not doctrinally provide real-time effects or operational support directly within the combat elements. Even where informal talent exists within units, there is no mechanism to authorise or support its use for TAR tasks. This creates a vacuum of personnel and skills around deployed RASAI systems, particularly as to who is creating the new AI models and methodologies. This gap is mirrored in the civilian sector, where no established trade or profession exists for deployed RASAI model handling, further limiting Defence’s ability to draw on external expertise and generate role-ready AI technical specialists.
• Inaccessible and Disconnected Data Streams: RASAI systems generate and rely on enormous volumes of data from sources such as onboard cameras, sensors, mission logs and system health monitors. But in most cases, this data is neither stored in accessible formats nor connected to a centralised analysis or retraining pipeline. Classification issues, bandwidth constraints, and stovepiped architectures mean tactical units cannot extract useful insights or enable AI retraining cycles. Without tactical-to-strategic data federation, TAR cannot occur: the system cannot learn what the fight teaches it.
• No Pathway for Rapid Field-Level Software Changes: Defence acquisition and compliance frameworks remain aligned to centrally tested, version-controlled software updates. Tactical teams that identify a model or logic issue cannot, typically, modify systems without triggering multi-stage certification. Emergency procurement pathways exist, but no standing process supports continuous integration or bounded in-theatre updates for RASAI systems. This has the potential to freeze needed adaptations behind bureaucratic processes which are disconnected from operational tempo. It is unclear if recommendations relating to improving capability timelines from the Defence Strategic Review and the National Defence Strategy have been meaningfully implemented; however, TAR could provide a mechanism to operationalise those reforms by institutionalising bounded, rapid update processes.
• Doctrinal Silence on RASAI Adaptation: Doctrine does not yet account for the adaptation of autonomous systems between engagements. A review of current ADF doctrine reveals no defined framework for tactical system modification or autonomy adaptation in theatre—only ambitions. Concepts like mission command, while enabling human initiative, have not been extended to RAS platforms. There is no doctrinal authority, process or expectation for tactical-level changes to autonomy settings, mission parameters or control logic. As a result, even units inclined to adapt their systems must navigate ambiguous policy terrain. There is also limited guidance on how AI-enabled systems should be managed, retrained, or tuned in the field, an absence that inhibits initiative.

While the requirements of TAR are more prevalent in high-intensity peer conflicts, some operational environments may impose limits on adaptation tempo, support availability or legal frameworks, particularly in coalition or grey zone operations. Therefore, TAR must be scalable and selectively applied based on mission type, risk tolerance and operational constraints. However, if TAR is permitted only in high-intensity conflicts, the ADF risks entering the conflict without a mature understanding how to effectively adapt its systems, thereby exposing a greater vulnerability when the stakes are highest.

Current gaps mean TAR is not yet viable at scale. Where adaptation occurs, it is typically improvised and driven by individual initiative rather than institutional support. Without formal structures, processes and authorities, the quality of the changes cannot be guaranteed. RASAI systems will only continue to deliver tactical advantage if they can evolve as quickly as the threat. Achieving this requires more than good design; it demands a force structure, workforce and doctrine that make machine adaptation a routine element of operations.

Recommendations

The following recommendations define the structural, technical and doctrinal changes needed to achieve TAR as a capability.

System and Procurement Foundations

Recommendation 1. Adopt a Formal RASAI Interface Framework

For TAR to be viable, RASAI systems must be modularly designed to support safe, field-level modification through open, standardised interfaces. This requires changes to both system architecture expectations and commercial arrangements:

• Develop and adopt a Defence-wide RASAI Interface Framework based on allied or NATO open mission system standards, to define mandatory interface layers for mobility systems, payload systems and decision/compute systems.
• Standardise RASAI compute hardware, allowing software and model updates to be deployed across different RAS platforms without unique dependencies.
• Retain control of standard development kits, interface definition and update tools within Defence, enabling secure in-house and third-party modification without OEM reliance.
• Prohibit closed, proprietary architectures that prevent field-level adaptation.

Recommendation 2. Embed Adaptability into RASAI Acquisition and Certification

All future RASAI procurements must include explicit TAR enablers. Systems must:

• mandate compliance with the modular framework in all RASAI acquisition contracts, ensuring vendors deliver systems with modularity, documented APIs, and containerised components that can be swapped or updated without proprietary engineering
• support modular payloads, open architectures and exposed control interfaces
• include abstraction layers that allow software and model updates without full re-certification
• contractually obligate vendors to provide tools for in-field configuration and patching, not just depot-level support.

Platforms that require OEM intervention or weeks-long validation cycles to change mission behaviour are unfit for TAR. Adaptability must become an acquisition gate for RASAI, not a retrofit.

Tactical Execution and Adaptation Infrastructure

Recommendation 3. Operationalise Tactical Data Feedback and Learning

For RASAI systems to adapt between missions, data must be captured, analysed and fed back rapidly. This demands:

• standardised logging architectures on all RASAI platforms, capturing sensor data, mission outcomes, and system anomalies
• on-site or forward-deployable analysts and triage tools to extract model-relevant data
• a live feedback loop that pushes refined models or logic updates back to deployed systems within days.

TAR requires more than raw data; it requires structured feedback infrastructure that turns recent operational experience into executable system improvements.

Recommendation 4. Establish a Model Update Pipeline

TAR requires a capability to generate, validate and deploy system updates within 24 to 72 hours. This must include:

• a field-ready pipeline for distributing software patches and ML model updates to deployed RASAI assets
• pre-approved adaptation boundaries, such as allowed model types, input limits and mission parameter envelopes, to expedite field-level approval
• local test harnesses or digital twin sandboxes for validating updates in theatre before deployment to live systems.

This pipeline should be treated as a standard support mechanism for RASAI operations, not an experimental workaround.

Recommendation 5. Raise a Deployable TAR Support Element

Create a dedicated force element responsible for the adaptation of fielded RASAI systems. This element, similar to electronic warfare or other specialised capability, would be embedded at the appropriate level and:

• include AI technicians, robotics engineers, software integrators and digital support specialists
• be capable of reconfiguring modular payloads, deploying newly updated AI models, and share new data across tactical forces. Updates could be applied daily, following a cadence similar to cryptographic key rotations
• maintain a pipeline to centralised analysis, model repositories, simulation environments and allied research facilities.

The proposed TAR support element will have appropriate technical expertise to conduct hardware and software changes within authorised parameters.^[51] Without it, modularity and AI-enabled platforms will remain static assets, regardless of threat evolution.

Workforce and Innovation Ecosystem

Recommendation 6. Develop a Uniformed RASAI Technical Workforce

RASAI adaptability depends on having technical expertise at the point of need and AI competent operators who understand the importance of correct labelling and the characteristics of AI. The ADF should therefore develop deeper expertise in artificial intelligence, robotics and software integration within its force structure. To make informed decisions during operations, combat leaders will require sufficient understanding of autonomy parameters, AI behaviour and the operational risk envelope associated with system updates. Crucially, these skills must be embedded within tactical units rather than concentrated in headquarters or specialist support organisations. Without uniformed digital specialists co-located with RASAI platforms, field-level adaptation will remain largely theoretical.

Recommendation 7: Integrate Academic and Industry Research Nodes for Rapid Method Adaptation and Innovation

In addition to regular model updates, new AI methodological innovation needs to be maintained. Maintaining this capability requires deliberate integration of Defence-approved academic and research partners into the RASAI adaptation cycle. TAR support elements should therefore be formally linked with accredited university research teams, DSTG, and national AI and robotics laboratories that can be mobilised to provide technical reach-back when required.

Operational RASAI systems should also be designed to support controlled method-level substitution. This includes the ability to replace object detection algorithms, autonomy control strategies, or data processing approaches using pre-authorised templates or sandbox validation environments. Rapid research and development pathways should exist to allow new approaches to be prototyped and evaluated within days or weeks when adversary tactics change.

To enable this interaction between operational units and external research partners, technical liaison officers should be embedded within TAR teams to translate operational problems into research tasks and integrate validated solutions back into deployed systems. A supporting ‘validated method repository’ should also be established, providing deployed units with access to modular models and technical approaches that have already been tested and approved for operational use.

While these recommendations provide specific mechanisms to support TAR, the development of tactically adaptable RASAI will ultimately require coordinated consideration across all FIC.

Conclusion

As intelligent systems become more prevalent in the battlespace, they must be delivered with adaptability in mind. Deployed AI-enabled platforms must be capable of having their behaviour modified between engagements, not over weeks or months but within the tempo of tactical operations.

RASAI systems will not deliver sustained warfighting advantage if they remain static and locked into their deployment configuration while adversaries evolve. TAR is the next capability step: the ability to update autonomy parameters, reconfigure platforms, adjust models, and redeploy systems in response to changing conditions.

The case studies in this paper from the US and NATO forces demonstrate how quickly tactical AI systems can be comprised and how quickly systems can be trained in response. The speed of adaptation necessitates a change in how we design, employ and evolve RASAI.

This paper proposes a pre-emptive framework for future RASAI integration. While RASAI platforms have not yet reached full saturation in the battlespace, the sentiment is clear that RASAI systems will become prolific. Defence needs to act now and consider the recommendations in this paper to avoid reliance on improvised adaptation once conflict begins.

Tactically adaptable robotic autonomous systems and artificial intelligence as a capability requires alignment across all FIC elements, which will deliver systems designed for field-level modification, embedded technical specialists, and streamlined update pipelines. Until this is achieved, even the most intelligent battlefield systems will be vulnerable to the simplest changes in a complex operational environment.

Endnotes

^[1] David W Barno and Nora Bensahel, Adaptation under Fire: How Militaries Change in Wartime (Oxford University Press, 2023).

^[2] Australian Army, Land Warfare Doctrine 1: The Fundamentals of Land Power (Australian Army, 2017).

^[3] Australian Army, Army’s Future Land Operating Concept (Australian Army, 2009).

^[4] Australian Army, Robotic & Autonomous Systems (RAS) Strategy v2.0 (Commonwealth of Australia, 2022).

^[5] Nand Mulchandani and John NT Shanaham, Software-Defined Warfare: Architecting the DOD’s Transition to the Digital Age (Center for Strategic & International Studies, 2022).

^[6] The changing environment includes changes in enemy conduct and capability, changing geography or topography, or changes in friendly conditions.

^[7] INFORMATION TO BE PROVIDED UPON PUBLICATION

^[8] Robotic & Autonomous Systems (RAS) Strategy v2.0.

^[9] Thomas Simpson et al., ‘Agile, Antifragile, AI-Enabled Command and Control (A3IC2)’, arXiv (2021), at: https://arxiv.org/abs/2109.06874.

^[10] Center for a New American Security, AI and Autonomy in Future Warfare (CNAS, 2020), at: www.cnas.org/research/defense/the-future-of-warfare.

^[11] Zachary Burdette et al., An AI Revolution in Military Affairs? How Artificial Intelligence Could Change How Militaries Fight and Win Wars (RAND Corporation, 2025).

^[12] Missy Cummings, Artificial Intelligence and the Future of Warfare (Chatham House, 2017).

^[13] United Nations Institute for Disarmament Research, Artificial Intelligence in the Military Domain and Its Implications for International Peace and Security: An Evidence-Based Road Map for Future Policy Action (UNIDIR, 2025).

^[14] Jacquelyn Schneider and Julia Macdonald, ‘Looking Back to Look Forward: Autonomous Systems, Military Revolutions, and the Importance of Cost’, Journal of Strategic Studies 47, no. 2 (2024): 162–184.

^[15] Kaleb McDowell et al., ‘Re-Envisioning Command and Control’, in International Conference on Military Communication and Information Systems, ICMCIS 2024, Koblenz, Germany, April 23–24, 2024 (IEEE, 2024), pp. 1–7.

^[16] James Black et al., Strategic Competition in the Age of AI: Emerging Risks and Opportunities from Military Use of Artificial Intelligence, RRA3295-1 (RAND Corporation, 2025).

^[17] Michael C Horowitz et al., ‘The Future of Military Applications of Artificial Intelligence: A Role for Confidence-Building Measures?’, Orbis 64, no. 4 (2020): 528–543.

^[18] Christina Balis and Paul O’Neill, Trust in AI: Rethinking Future Command (Royal United Services Institute, June 2022).

^[19] Stuart J Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th edition (Pearson, 2021).

^[20] 38th Annual Conference of Neural Information Processing Systems (NeurIPS) 10–15 December 2024 (Neural Information Processing Systems, 2024), at: https://media.neurips.cc/Conferences/NeurIPS2024/NeurIPS2024-Fact_Sheet.pdf.

^[21] Leah Hovig, ‘CVPR 2024 Breaks Paper and Attendance Records; AI Art Award and Demo Winners Announced’, IEEE Computer Society (website), 26 June 2024.

^[22] Long Ouyang et al., ‘Training Language Models to Follow Instructions with Human Feedback’, arXiv (2022), at: https://doi.org/10.48550/arXiv.2203.02155.

^[23] Nicholas Barber, ‘Adapt and Overcome: Promoting Tactical Adaptation in the Post-Afghanistan Army’, Australian Army Journal 11, no. 2 (2014).

^[24] Barno and Bensahel, Adaptation under Fire.

^[25] Daniel Bilusich et al., ‘Visualising Adaptive Campaigning: Influence Diagrams in Support of Concept Development’, Australian Army Journal 9, no. 1 (2012).

^[26] David Kilcullen, ‘Mobilisation and Australia’s National Resilience’, Australian Army Journal 20, no. 3 (2024).

^[27] Trent Scott, ‘Adapt or Die: Operational Design and Adaptation’, Australian Army Journal 6, no. 3 (2009).

^[28] Marion Schroeder, ‘Adapt or Die—How Does the Army Need to Adapt to the Changing Character of War?’, The Cove, 28 February 2025.

^[29] Department of Defence, The One Defence Capability System Manual V 2.0 (Commonwealth of Australia, 2025).

^[30] Department of Defence, National Defence: Defence Strategic Review 2023 (Commonwealth of Australia, 2023).

^[31] Army’s Future Land Operating Concept.

^[32] Scott, ‘Adapt or Die’.

^[33] Robotic & Autonomous Systems (RAS) Strategy v2.0.

^[34] United States Air Force, Air Force Doctrine Annex 3-51: Electromagnetic Warfare and Electromagnetic Spectrum Operations (U.S. Air Force, 2019).

^[35] U.S. Government Accountability Office, Electronic Warfare: DOD Actions Needed to Strengthen Management and Oversight, GAO-12-479 (U.S. GAO, 2012).

^[36] Max Hauptman, ‘Marines Outwitted an AI Security Camera by Hiding in a Cardboard Box’, Task & Purpose, 24 January 2023, at: https://taskandpurpose.com/news/marines-ai-paul-scharre.

^[37] Steve Orrin, ‘How the Military Is Preparing for AI at the Edge’, C4ISRNET, 27 June 2024.

^[38] Max Hunder, ‘Ukraine Collects Vast War Data Trove to Train AI Models’, RNZ, 20 December 2024, at: www.rnz.co.nz/news/world/537299/ukraine-collects-vast-war-data-trove-to-train-ai-models.

^[39] Kateryna Bondar, Ukraine’s Future Vision and Current Capabilities for Waging AI-Enabled Autonomous Warfare (Center for Strategic & International Studies, 2025).

^[40] ‘Czech Armed Forces Complete Training in Simulated Electronic Warfare Environment’, Defence Industry Europe, 21 May 2024.

^[41] Ameya Paleja, ‘US and NATO Allies Test AI-Powered Anti-Jamming Tech to Fortify Defenses’, Interesting Engineering, 21 November 2023.

^[42] Robotic & Autonomous Systems (RAS) Strategy v2.0.

^[43] Royal Australian Navy, RAS-AI Strategy 2040 (Navy, 2023), at: www.navy.gov.au/media-room/publications/ras-ai-strategy-2040.

^[44] Australian Defence Force, Concept for Robotic and Autonomous Systems (Commonwealth of Australia, 2020).

^[45] Aymen Rayane Khouas et al., ‘Training Machine Learning Models at the Edge: A Survey’, arXiv (2024), at: https://doi.org/10.48550/arXiv.2403.02619.

^[46] Lewis Shepherd, ‘Situational Awareness at the Tactical Edge: How Private AI Is Driving Defense AI Innovation’, Broadcom, 17 March 2025.

^[47] Ian Reynolds and Yasir Atalan, ‘Calibrating NATO’s Vision of AI-Enabled Decision Support’, CSIS (website), 7 August 2024, at: www.csis.org/analysis/calibrating-natos-vision-ai-enabled-decision-support.

^[48] ‘Summary of NATO’s Revised Artificial Intelligence (AI) Strategy’, NATO (website), 10 July 2024.

^[49] Dominic Perez, ‘CJADC2 Interoperability: AI-/ML-Based Sensor Fusion at the Edge’, Military Embedded Systems, 25 November 2024.

^[50] Jean Scholtz, ‘Theory and Evaluation of Human Robot Interactions’, Proceedings of the 36th Annual Hawaii International Conference on System Sciences (IEEE, 2003).

^[51] Australian Defence Force, ADF Concept for Command and Control of the Future Force (Commonwealth of Australia, 2019).