Adaptive Command in Modern C2

• Adaptive command is a decision architecture that continuously revises commands based on real-time feedback from dynamic environments and human inputs.
• It applies techniques like prompt-driven policy adjustments, belief updates, and online optimization across domains such as air defense, robotics, and game AI.
• The approach balances full autonomy with human–AI collaboration, ensuring rapid, resilient decision-making under communication and operational constraints.

Searching arXiv for papers on adaptive command and related command-and-control frameworks. I’m unable to access the arXiv search tool in this interface, so I’m grounding the article strictly in the supplied arXiv records. Adaptive command denotes a class of decision architectures in which commands, plans, or policies are revised online as observations, human instructions, environmental conditions, or measured performance change. In future Command and Control (C2), it is defined as a capability in which human commanders and AI-enabled decision-support tools jointly generate, assess, and revise courses of action (COAs) in real time under highly dynamic, uncertain, and communication-constrained conditions (Madison et al., 2024). Related work instantiates the same general idea in autonomous air defense, real-time strategy (RTS) games, tabletop command interpretation, swarm robotics, off-road navigation, adaptive predictive control, and terminal guidance, where adaptation occurs through belief updates, fuzzy inference, prompt-driven policy adjustment, decentralized replanning, uncertainty-aware command vetoing, or extremum-seeking set-point generation (Das, 2019, Ma et al., 5 Aug 2025, Prabhakar et al., 2020, Delgado et al., 6 May 2026).

1. Conceptual scope and recurring structure

Across the literature, adaptive command is not a single algorithm but a recurring systems pattern. A command layer receives high-level intent, sensor updates, or a measurable objective; an execution layer maps that input into concrete actions; and an adaptation mechanism reconditions the mapping when the operating context changes. In C2 formulations, this appears as continuous feedback loops between humans and AI, dynamic reallocation of cognitive and computational resources, and preservation of shared expertise through interactive machine learning (Madison et al., 2024). In autonomous control settings, the same pattern appears as online identification, recurrent policy adaptation, or replanning against updated target distributions or uncertainty estimates (Gaudet et al., 2021, Samson et al., 19 Jun 2025).

Domain	Adaptive object	Representative papers
Future C2	COAs, role allocation, shared knowledge	(Madison et al., 2024, McDowell et al., 2024)
Air defense	BDI plans, threat scores, weapon allocation	(Das, 2019)
RTS games	High-level policy, behavior modulators, tactical adapters	(Ma et al., 5 Aug 2025, Ma et al., 21 Jul 2025)
Tabletop HRI	Object/action/target grounding from multimodal commands	(Gajewski et al., 2024)
Swarm robotics	Spatial target distribution and reference trajectories	(Prabhakar et al., 2020)
Off-road UGVs	Velocity commands under slip uncertainty	(Samson et al., 19 Jun 2025)
Adaptive control and guidance	Set-point commands and attitude-rate commands	(Delgado et al., 6 May 2026, Gaudet et al., 2021)

A common misconception is that adaptive command necessarily implies end-to-end online learning. The literature is more heterogeneous. Some systems freeze the base policy or LLM and adapt only through prompting or lightweight modules; others update statistical models, fit new distributions, or adapt command signals directly; still others treat explainable modularity as the basis for adaptation rather than gradient-based learning (Ma et al., 5 Aug 2025, Ma et al., 21 Jul 2025, Gajewski et al., 2024).

2. Human–AI adaptive command in contested C2

In the future-C2 literature, adaptive command is motivated by Denied, Degraded, Intermittent, and Limited communications, heterogeneous multi-domain data streams, and accelerating operational tempo. These conditions are held to invalidate slow, linear staff processes derived from the industrial era and to require real-time human–AI partnerships that can maintain coherence under jamming, sparse bandwidth, and shortened decision windows (McDowell et al., 2024).

Two complementary formalizations recur. One maps tasks along a spectrum of integration according to complexity, decision timescale, and information certainty via a function $$\phi(C,\tau,\sigma)\to\{\text{Human},\text{AI},\text{Shared}\}$$. The other defines a human-machine partnership as $P=(H,A,R)$, where $H$ is the set of human roles, preferences, and biases, $A$ is the set of AI algorithms and models, and $R$ is a real-time, bidirectional role-assignment policy. An adaptive command system is then defined by the requirement that under environmental change $\delta E$, there exists an adaptation rate $\alpha>0$ such that $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$ (McDowell et al., 2024).

Operationally, this formulation compresses the Military Decision-Making Process into a fast iterative loop. The envisioned iCOAs suite continuously generates COAs, scores them on multiple objectives, and exposes them to humans for rapid steering. Decision latency reduction is expressed as $L=T_0-T_{\text{new}}$, with $T_{\text{new}}\approx T_0/\beta$ and $P=(H,A,R)$0. Unity of effort is formalized as

$P=(H,A,R)$1

where $P=(H,A,R)$2 is the feature vector representing unit $P=(H,A,R)$3’s plan at time $P=(H,A,R)$4. Adaptive collective knowledge is represented by

$P=(H,A,R)$5

so that new data and human feedback update the shared knowledge base at a tunable rate (McDowell et al., 2024).

The more explicitly machine-learning-centered SIML agenda decomposes adaptive command into three research areas: human–AI interaction algorithms for planning in complex, dynamic situations; methods for fostering resilient human–AI teams through roles, configurations, and trust; and techniques for scaling algorithms and teams across diverse contexts. It frames COA generation under uncertainty using multi-agent POMDPs and hierarchical reinforcement learning, and treats demonstrations, corrections, preference feedback, compressed communications, trust monitoring, and dynamic role assignment as central components of adaptive command rather than peripheral interface choices (Madison et al., 2024).

Illustrative scenarios make the operational claim concrete. In mission planning, iCOAs can spawn 200 candidate COAs in parallel, refine all 200 in $P=(H,A,R)$6 after staff feedback, and reduce latency from $P=(H,A,R)$7 to $P=(H,A,R)$8. During execution under fire, a local iCOAs-S instance can maintain $P=(H,A,R)$9 within $H$0 despite full voice silence by sending compressed intent vectors rather than full radio traffic. In after-action review, setting $H$1 for an unanticipated capability causes future COA generation to immediately absorb the new lesson into the shared knowledge base (McDowell et al., 2024).

3. Autonomous air-defense command as BDI adaptation

A concrete autonomous C2 realization appears in the air-defense system developed by Das, where agents are implemented in JACK using a BDI-style architecture with explicit BeliefBase, Goals, and Plans, and with Meta-Plan Reasoning governing plan selection (Das, 2019). The system is entirely agent-based, runs without human intervention in an air combat simulated environment, and performs electronic counter-countermeasures, threat assessment, and weapon allocation.

The architecture centers on two agent families. The Surveillance Radar Agent monitors jamming by processing sequential radar returns $H$2 and storing the time-stamped relative difference

$H$3

If $H$4 exceeds a learned threshold, the plan “RadarIsJammed” is posted; otherwise “RadarIsNotJammed.” The Air Defense Direction Center Agent fuses tracks from central and local radars to produce a target set $H$5 and computes a fuzzy threat score $H$6 for each candidate target. Plans are then prioritized for threat clustering, cluster-priority assignment, “RedAircraft–Interceptor” pairing, and optimal weapon-target allocation (Das, 2019).

Threat assessment is modeled as a Mamdani-style fuzzy inference system over six attributes: range $H$7, velocity $H$8, altitude $H$9, angle of attack $A$0, target type $A$1, and intent class $A$2. Rule firing strengths are products of trapezoidal or semi-trapezoidal membership functions, and the inferred threat score is defuzzified by weighted average:

$A$3

with output fuzzy sets $A$4. Weapon allocation is then formulated as an integer linear program maximizing target-value-destroyed,

$A$5

subject to inventory and target-count constraints. For large instances, a genetic-algorithm “greedy eugenics” solver is used within a fixed $A$6 planning window (Das, 2019).

Adaptation is distributed across several mechanisms. New events invoke JACK’s Meta-Plan Reasoning library; “relevant(e)” prunes plans, “context()” enforces belief consistency, and “getInstanceInfo().def” exposes a fuzzy-computed rank used to select the highest-ranked plan. Fuzzy interpolation handles previously unobserved threat signatures without enumerating all 1,215 rule combinations. The Surveillance Radar Agent also logs $A$7 measurements and refits a Generalized Extreme Value model using the Kolmogorov–Smirnov test, declaring jamming when the KS statistic exceeds the critical threshold (Das, 2019).

Evaluation combines logical consistency, distribution fitting, and end-to-end simulation. Over 500 simulated runs, the Surveillance Radar Agent never pursued conflicting goals simultaneously. The radar agent’s $A$8 sequence was best fit by a GEV law with $A$9 at $R$0, and the ADDCA’s minimum and maximum threat order statistics were best fit by Inverse-Gaussian laws. In a $R$1 defended-airfield scenario involving 10 fighters, 10 bombers, 10 cruise missiles, 50 smart bombs, 1 AWACS, 1 UAV, and random jamming, $R$2 occurred in 231 of 500 runs, prompting frequency hops in 41% of cases; 91 messages were sent to ADDCA for interceptor tasking; no false-positive jamming declarations were observed; jamming-detection latency was $R$3; frequency-hop success rate was 100%; SRA planning cost was $R$4 per $R$5 check; and the ADDCA achieved $R$6 of maximal possible TVD for up to 90 threats, degrading gracefully beyond that resource horizon (Das, 2019).

This air-defense line of work is important because it shows adaptive command in a strong autonomy sense: continuous belief revision, dynamic plan re-ranking, fuzzy threat interpolation, and online statistical refitting all occur without a human remaining in the loop. A plausible implication is that later human–AI C2 visions can be read not as replacements for autonomy, but as attempts to reinsert controllable partnership structure into domains where autonomous adaptation is already technically feasible.

4. Natural-language-conditioned command adaptation

A second major strand treats adaptive command as language-conditioned policy adjustment. In the StarCraft II system explicitly named Adaptive Command, the architecture combines a Game State Processor using a Chain of Summarization, an LLM-based Strategic Advisor, a Behavior Tree Executor, and a natural-language interface with whisper-small and edge-tts. The frozen LLM computes an initial policy $R$7 and later adjusts policy via $R$8, while the behavior tree selects atomic actions through $R$9. No online gradient updates are performed; policy adjustment is purely prompt-driven re-conditioning of behavior modulators (Ma et al., 5 Aug 2025).

This design yields measurable benefits for low- and mid-skill play. Against Blizzard AI (Protoss, “Very Hard” Level 6), novice players improved from 0% win rate in the control condition to 42% with Adaptive Command, intermediate players from 58% to 75%, and expert players remained at 100% in both conditions. Instruction-following scores were 4.1 for novices, 4.4 for intermediates, and 2.8 for experts; helpfulness scores were 4.7, 4.2, and 3.1 respectively. The only reported baseline is standard play, and no formal ablation is reported, although a rough internal comparison stated that removing Chain of Summarization doubled LLM latency to about 1200 ms from about 600 ms without win-rate gain (Ma et al., 5 Aug 2025).

TacticCraft implements a different language-conditioning strategy. It freezes a pre-trained DI-Star policy and attaches lightweight adapter modules to each action head, conditioned on a tactical tensor $\delta E$0 encoding nine high-level tactical archetypes. A natural-language command is mapped to $\delta E$1 by a fine-tuned LLM or lightweight classifier, softmaxed over the nine categories, and injected additively into the penultimate activation of each decision head. Training uses a PPO-style objective regularized by a KL penalty to keep the adapted policy close to the pre-trained one:

$\delta E$2

The extra inference cost is modest: base DI-Star is about 60 ms per decision, while the adapter-augmented system is about 63 ms with about 0.5 M extra parameters and +2% memory overhead. In an emergent case study, Configuration C at Epoch 5 K with $\delta E$3 emphasizing (+1 Zergling) and (Lurker) discovered a hybrid +1-Zergling $\delta E$4 Lurker timing attack with 73% win rate versus Level 10 (Ma et al., 21 Jul 2025).

A third formulation emphasizes interpretable multimodal grounding rather than policy modulation. In tabletop scenarios, a short RGB video clip containing one spoken command is processed by Whisper for time-stamped transcription, Phi-3 mini for a JSON command parse, Mediapipe for hand landmarks and pointing vectors, GroundingDINO for open-vocabulary object detection, spatial filtering by proximity to the pointing line, and Segment Anything for final masks. The output is a minimal JSON symbol structure for object, action, and target, together with pixel masks, and can be passed directly into symbolic planners such as CRAM, KnowRob, or graph-task. The evaluation is qualitative only: no quantitative metrics, precision/recall, IoU, or ROC curves are reported, and failure cases are attributed chiefly to model errors such as MediaPipe blur sensitivity or Whisper mistranscriptions (Gajewski et al., 2024).

Taken together, these systems clarify that natural-language adaptive command need not mean the same thing operationally. It can mean prompt-driven strategic advice with fixed model weights, adapter-conditioned tactical modulation around a frozen policy, or modular command interpretation that yields inspectable symbolic outputs. This suggests that the decisive technical choice is not whether language is present, but where language enters the control stack: as policy prior, as head-level bias, or as semantic grounding for downstream planners.

5. Embodied robotics, swarm control, and command uncertainty

In swarm robotics, adaptive command is instantiated as dynamic reshaping of a spatial target distribution for decentralized ergodic control. User commands arrive through a tactile tablet interface or Unity VR interface, are converted into a distribution $\delta E$5 over a domain $\delta E$6, and are combined with environment-driven distributions to form

$\delta E$7

Agents then minimize the ergodic metric

$\delta E$8

with consensus over shared Fourier coefficients and low-level tracking via RT-RRT*. The system responds not only to user shading commands but also to detected easter-eggs or disabling devices, which add attractive or repulsive terms to $\delta E$9 and trigger immediate replanning (Prabhakar et al., 2020).

This formulation is explicitly adaptive in both control and communications architecture. Commands and environment events are broadcast through a Java interface over LTE, each agent runs a decentralized ergodic planner as a ROS node on Jetson TX2 hardware, and only Fourier coefficients rather than full trajectories need to be shared. In virtual reality simulation with 6 quadcopters, the time-averaged Fourier reconstruction tracked the changing $\alpha>0$0 within one planning horizon of about 5 s, and coverage error $\alpha>0$1 dropped below 0.05 within 8 s after each update. In DARPA OFFSET FX3 field trials with 4 ground robots, user commands elicited a collective response in under 10 s despite loss of up to 2 robots, an easter-egg caused the swarm to form a perimeter within 15–20 s, and coverage error remained below 0.1 even under intermittent LTE dropout (Prabhakar et al., 2020).

For off-road autonomous navigation, the emphasis shifts from command generation to command admissibility under uncertain terrain–vehicle interaction. The DRIVE protocol standardizes data collection for skid-steer UGV slip by sampling commands across the feasible domain, executing each for 6 s, measuring wheel and body-frame motion, and estimating slip transfer functions by Gaussian smoothing. In the wheel frame, slip is $\alpha>0$2; in the body frame, $\alpha>0$3. The resulting local slip map $\alpha>0$4 is estimated by weighted averaging over sampled commands (Samson et al., 19 Jun 2025).

The key adaptive-command contribution in this setting is the unpredictability metric

$\alpha>0$5

which compares commanded and realized kinetic energies after orientation penalties. When $\alpha>0$6, $\alpha>0$7; when $\alpha>0$8, $\alpha>0$9. The metric is proposed as an online estimate of command uncertainty for risk-aware deployment. One may set a threshold $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$0, for example 0.2–0.3, above which commands are vetoed or replaced with lower-speed or lower-curvature alternatives, or add $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$1 directly to a planner’s cost. Empirically, ice exhibited a median $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$2 about 1.6 times higher than grass, asphalt, and gravel combined, and realized body-frame speeds on ice were limited to about $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$3 versus a $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$4 command ceiling (Samson et al., 19 Jun 2025).

These robotic formulations broaden adaptive command beyond plan revision. In the swarm case, the command itself is a time-varying objective distribution that agents collectively realize. In the DRIVE setting, adaptation may consist in rejecting commands whose uncertainty is too high. The common element is not the specific control law, but the online remapping from nominal operator intent to executable motion under current environmental conditions.

6. Adaptive command generation, guidance, and control-theoretic realizations

In adaptive control, command generation can be separated from low-level stabilization and optimized directly against a measured but unknown performance function. The ECG/PCAC framework does this by combining an extremum-seeking-based command generator with Predictive Cost Adaptive Control. At each sample $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$5, a measured cost $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$6 is available, though its functional form is unknown. ECG updates a command $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$7 through a discrete-time dither–demodulate–filter–integrate loop and feeds that command into PCAC, which solves a constrained receding-horizon quadratic program using an online-identified model updated by recursive least squares with variable-rate forgetting. The architecture therefore combines command generation, closed-loop system identification, and online optimization in a single loop (Delgado et al., 6 May 2026).

The ECG dynamics are given explicitly by high-pass and low-pass filtering of the measured cost, integration of a gradient proxy, and addition of a dither term:

$\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$8

with the mean-value dynamics following $\frac{dP}{dt}\ge \alpha\cdot\|\delta E\|$9 under standard time-scale separation assumptions. PCAC then solves a quadratic program over predicted outputs, tracking references, incremental inputs, and soft output constraints. The numerical studies cover an undamped oscillator, a disturbed double integrator, and an exponentially unstable plant, showing stabilization, command following, disturbance rejection, and asymptotic movement of $L=T_0-T_{\text{new}}$0 toward the extremum of the unknown cost (Delgado et al., 6 May 2026).

A reinforcement-learning realization appears in terminal adaptive guidance for hypersonic strike weapons. Here the policy observes an 8-dimensional vector derived directly from radar-seeker outputs,

$L=T_0-T_{\text{new}}$1

corrupted by small scale-factor error, and outputs normalized commanded rates for bank angle, angle of attack, and sideslip. PPO with a recurrent architecture is trained over an ensemble of perturbations in aerodynamic coefficients, atmospheric density, sensor scale factors, actuator lag, actuator failure, and maneuvering targets. The reward combines a shaping term that penalizes line-of-sight rotation, a small quadratic control penalty, and a terminal bonus for successful impact under path and speed constraints (Gaudet et al., 2021).

The reported performance is strong under off-nominal conditions. Over 5000 randomized episodes, the nominal optimized ensemble achieved mean miss distance $L=T_0-T_{\text{new}}$2 with 100.0% success for miss $L=T_0-T_{\text{new}}$3 and miss $L=T_0-T_{\text{new}}$4, and 0.0% constraint violations. Under larger parameter variation, actuator failure, divert maneuvers, and evasion scenarios, success remained above 98% in the 3-DOF setting, with only rare heating, dynamic-pressure, or load violations. A preliminary integrated 6-DOF version achieved mean miss $L=T_0-T_{\text{new}}$5 and 99.7% success in the no-divert case, and mean miss $L=T_0-T_{\text{new}}$6 with 65.9% success in the full ensemble (Gaudet et al., 2021).

These control-theoretic and RL-guidance formulations expose a core technical divide within adaptive command. One line adapts commands by explicit online optimization around a measurable objective, with stability, constraints, and identification occupying center stage. The other line learns a recurrent mapping from compact observations to commands that can instantly re-adapt through hidden state without in-flight retraining. This suggests that “adaptive command” names a systems property rather than a preferred learning paradigm: the ability to revise actionable control outputs fast enough to remain effective when the plant, environment, task, or adversary changes.

Across the literature, the main unresolved issues are also recurrent. Some systems emphasize human–AI trust, role assignment, and explainability; some achieve full autonomy without human intervention; some rely on prompt-driven reconditioning rather than parameter updates; and evaluation rigor ranges from detailed quantitative operational metrics to purely qualitative demonstrations. The concept is therefore best understood as an umbrella for real-time command revision under uncertainty, not as a settled canonical architecture.