EdgeAgentX-DT: Resilient Edge Intelligence

• EdgeAgentX-DT is an advanced extension of EdgeAgentX that combines digital twin simulation with generative AI to improve resilience and diversity in training tactical network policies.
• It employs a multi-layer architecture integrating on-device intelligence, digital twin synchronization, and scenario generation to address rare, extreme operational conditions.
• Empirical evaluations demonstrate improvements in latency, throughput, and learning convergence under adversarial stressors compared to baseline models.

Searching arXiv for the cited EdgeAgentX-DT and base EdgeAgentX papers to ground the article. EdgeAgentX-DT is an extension of EdgeAgentX that integrates digital twins and generative AI for resilient edge intelligence in tactical networks. In the formulation reported for military communication networks, it augments a federated learning and centralized-training/decentralized-execution multi-agent reinforcement learning stack with a live network digital twin and a generative scenario engine, so that policies are trained and validated against a broader distribution of conditions than those encountered in ordinary field-derived training alone (Ray, 28 Jul 2025). The system is positioned for contested environments characterized by RF jamming/spoofing, intermittent or degraded connectivity, dynamic topology from mobility, harsh compute and energy constraints, unexpected node failures, traffic surges, and strict latency and throughput objectives (Ray, 28 Jul 2025). Relative to the original EdgeAgentX framework, which integrated federated learning, MADDPG-based MARL, and adversarial defenses for military communication networks (Ray, 24 May 2025), EdgeAgentX-DT adds a synchronized virtual environment and a generative training layer intended to improve convergence speed, steady-state network performance, and resilience under compound stressors (Ray, 28 Jul 2025).

1. Origins and conceptual scope

EdgeAgentX-DT emerged as an advanced extension of EdgeAgentX, a three-layer edge AI framework for military communication networks integrating federated learning, multi-agent reinforcement learning via MADDPG, and adversarial defense (Ray, 24 May 2025). In the original framework, distributed intelligent agents ran on soldier radios, autonomous UAVs, and tactical IoT sensors; each node observed local state such as wireless channel quality, queue length, battery, and nearby allies, and acted by choosing channels, adjusting transmit power, selecting next hops, or deciding whether to transmit or hold data (Ray, 24 May 2025). Training was centralized through MADDPG critics that observed joint state-action tuples, while execution was decentralized through local actors, and federated learning synchronized model parameters without raw data sharing (Ray, 24 May 2025).

The original EdgeAgentX paper did not define a Digital Twin component and did not present an “EdgeAgentX-DT” variant (Ray, 24 May 2025). EdgeAgentX-DT was subsequently defined as a system that integrates digital twin simulations and generative AI-driven scenario training into the EdgeAgentX architecture (Ray, 28 Jul 2025). In that formulation, the key limitation being addressed was that the original EdgeAgentX still relied on whatever environments it encountered during training, which was described as insufficient coverage of rare or extreme conditions (Ray, 28 Jul 2025). EdgeAgentX-DT therefore adds a live network digital twin that mirrors field conditions and a generative scenario engine using diffusion models and transformers to create diverse and adversarial scenarios for training within the twin (Ray, 28 Jul 2025).

This suggests a conceptual shift from learning only from operational trajectories and federated aggregation to a closed-loop “calibration-then-train” regime. A plausible implication is that EdgeAgentX-DT should be understood not merely as a deployment framework, but as a training-and-validation architecture in which real telemetry, simulation fidelity, and scenario synthesis interact continuously (Ray, 28 Jul 2025).

2. Multi-layer architecture

EdgeAgentX-DT is described as a multi-layer architecture with three principal layers: on-device edge intelligence, digital twin synchronization, and generative scenario training (Ray, 28 Jul 2025). At the physical layer, distributed edge nodes such as radios, UAV relays, and sensors execute local policies under decentralized execution. These agents observe local state variables including channel quality, neighbor status, and queue length, and take actions over routing, power, channel selection, and mobility (Ray, 28 Jul 2025).

The middle layer is a digital twin environment, denoted as $\mathcal{M}_T$, which serves as a software replica of nodes, links, traffic, propagation, and mobility and is maintained so as to mirror the real environment $\mathcal{M}_E$ (Ray, 28 Jul 2025). The twin maintains an estimated network state $\hat{s}(t)$, while the real network has state $s(t)$. Synchronization occurs periodically or on events, with the twin updated from field telemetry according to

$$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$

The twin supports state setting, scenario injection, time stepping, pausing, fast-forward, instrumentation, detailed KPI export, and policy validation before deployment (Ray, 28 Jul 2025).

The upper layer is a generative scenario training layer. It contains a Scenario Generator that injects conditions into the twin using diffusion models for continuous or spatial state snapshots and transformers for discrete event sequences over time (Ray, 28 Jul 2025). Diffusion models are used for artifacts such as channel or interference maps and link-matrix states, while transformers generate event sequences such as node failures, jamming windows, and traffic surges (Ray, 28 Jul 2025).

A central coordinator spans all three layers. It orchestrates secure federated aggregation with anomaly detection, performs centralized MARL updates on twin-generated experience, disseminates updated policies to devices, and can validate candidate policies in the twin before field deployment (Ray, 28 Jul 2025). This coordinator is structurally consistent with the original EdgeAgentX coordination layer, which already combined global model aggregation, hierarchical federated learning, and centralized training with decentralized execution (Ray, 24 May 2025).

The architectural decomposition is summarized below.

Layer	Main function	Representative elements
1. On-device edge intelligence	Real-world decentralized execution	Radios, UAV relays, sensors, local agents
2. Digital twin synchronization	Virtual replica of the tactical network	$\mathcal{M}_T$, $\hat{s}(t)$, telemetry-driven updates
3. Generative scenario training	Synthetic and adversarial scenario injection	Diffusion models, transformers, scenario generator

A plausible implication is that the middle layer changes the status of simulation from an offline engineering tool to an active runtime component in the learning loop, while the upper layer changes scenario generation from manual scripting to data-driven conditional synthesis (Ray, 28 Jul 2025).

3. Learning stack and control formulation

EdgeAgentX-DT inherits the learning substrate of EdgeAgentX: federated learning on real edge data, centralized training with decentralized execution for MARL, and adversarial defenses against poisoning and jamming (Ray, 24 May 2025). In EdgeAgentX, federated aggregation is described conceptually through hierarchical FedAvg-like weight averaging, with the next global parameter set given by

$\theta_{\text{global}^{t+1} = \frac{1}{N}\sum_{i=1}^{N}\theta_i^{t}$

using local parameters from $N$ agents (Ray, 24 May 2025). The system uses hierarchical aggregation, robust server-side anomaly detection on model updates, and secure communication via authentication and encryption (Ray, 24 May 2025).

The MARL component retains a MADDPG-like structure in which each agent has an actor $\pi_i$ and critic $\mathcal{M}_E$0, critics are trained centrally on joint state-action tuples, and actors execute locally under partial observability (Ray, 24 May 2025). The reward in the original EdgeAgentX system is defined as

$\mathcal{M}_E$1

with additional small penalties when adversarial behavior is detected, although the exact penalty term is not given (Ray, 24 May 2025).

EdgeAgentX-DT introduces what it terms an enhanced two-loop learning process (Ray, 28 Jul 2025). The outer loop keeps the twin calibrated to field data, while the inner loop uses simulated experience, including generative scenarios, to improve policies (Ray, 28 Jul 2025). Operationally, the training procedure is described as alternating among real interaction, federated aggregation, twin synchronization, scenario generation, twin episodes, centralized MARL update, and policy deployment (Ray, 28 Jul 2025).

The digital twin provides observations and rewards analogous to the physical environment, so that policies can be trained in a consistent MDP-like structure (Ray, 28 Jul 2025). The coordinator can run multiple twin episodes per real episode, often at 5–10× the rate of real interactions, resource dependent (Ray, 28 Jul 2025). This suggests that the twin is used not only for safety validation but also for accelerating policy iteration by increasing the volume and diversity of experience available between field updates.

A distinct feature of the EdgeAgentX-DT formulation is targeted scenario sampling. The Scenario Generator supports targeted sampling of failure modes that recently caused low reward or high latency, described as a curriculum inspired by difficult scenarios discovered during learning, though not implemented through a dedicated adversary agent (Ray, 28 Jul 2025). A plausible implication is that the training distribution becomes adaptive: rather than sampling broadly from historical distributions alone, the system preferentially allocates compute to scenarios that expose current policy weaknesses.

4. Digital twin synchronization and generative scenario synthesis

The digital twin in EdgeAgentX-DT models node positions and mobility, radio propagation, fading distributions, interference, link capacities, routing queues, traffic flows, and topology (Ray, 28 Jul 2025). Synchronization is performed at intervals $\mathcal{M}_E$2 or on events such as topology changes or jamming detections (Ray, 28 Jul 2025). Field telemetry includes measured link KPIs such as latency, bandwidth, and loss, as well as mobility traces, node health, and events; these measurements are used to overwrite or patch state variables and recalibrate propagation and traffic parameters (Ray, 28 Jul 2025).

The synchronization mechanism is deliberately simple in the reported formulation. No explicit Kalman or particle filtering equations are specified; instead, synchronization is implemented via direct state patching and recalibration using $\mathcal{M}_E$3 (Ray, 28 Jul 2025). This distinguishes EdgeAgentX-DT from generalized digital twin proposals in which state-estimation updates might be formalized differently. A plausible implication is that fidelity management is treated as a systems-engineering problem rather than as a separate filtering-theoretic contribution.

The generative layer uses two different model classes. Diffusion models generate continuous or spatial state variables, with scenario states denoted by $\mathcal{M}_E$4 and sampled as

$\mathcal{M}_E$5

from a model trained on historical and simulated data, including 1,000 layouts and interference patterns with varied jammers (Ray, 28 Jul 2025). Transformer-based generators produce event sequences

$\mathcal{M}_E$6

which may encode node failures, jammer activations on specific channels and regions for given durations, and traffic spikes in designated areas (Ray, 28 Jul 2025).

The conditioning variables for the diffusion model include traffic load level, number and location of jammers, and environmental factors, enabling targeted sampling of rare or adversarial conditions (Ray, 28 Jul 2025). The transformer is trained on historical exercises and composed scenarios so as to capture temporal correlations across events (Ray, 28 Jul 2025). Constrained generation is used to preserve feasibility and avoid impossible scenarios by enforcing domain rules on topology and physics (Ray, 28 Jul 2025).

The full significance of this layer is methodological. In the original EdgeAgentX, adversarial training exposed agents to jamming and perturbations in simulation, but there was no twin-based generative scenario engine (Ray, 24 May 2025). EdgeAgentX-DT therefore shifts from a fixed or manually curated set of hostile conditions to a learned scenario manifold with controllable parameters (Ray, 28 Jul 2025). This suggests a tighter coupling between training diversity and operational telemetry, particularly if generators are periodically retrained offline as new data accumulate.

5. Network setting, evaluation protocol, and empirical results

The evaluation setting for EdgeAgentX-DT is a tactical multi-hop mesh over a $\mathcal{M}_E$7 area with $\mathcal{M}_E$8 mobile nodes plus one base station or coordinator, and traffic directed to a gateway (Ray, 28 Jul 2025). The digital twin is implemented in an OMNeT++-based simulator with custom wireless channel models calibrated from telemetry (Ray, 28 Jul 2025). Training uses 1,000 episodes with 100 time steps per episode and five random seeds (Ray, 28 Jul 2025). Adversarial conditions include jamming and poisoning, with up to 20% of nodes compromised for malicious model updates and robust aggregation enabled (Ray, 28 Jul 2025).

The baseline set includes the original EdgeAgentX, Independent RL, Federated RL without MARL, a Centralized RL oracle, EdgeAgentX-DT without GenAI, and EdgeAgentX without Defense (Ray, 28 Jul 2025). This baseline structure extends the one used in the original EdgeAgentX study, which also compared against Independent RL, Centralized RL, Federated RL without MARL, and EdgeAgentX without Defense (Ray, 24 May 2025).

Under moderate load, EdgeAgentX-DT is reported at approximately $\mathcal{M}_E$9 ms average latency, compared with approximately $\hat{s}(t)$0 ms for EdgeAgentX, approximately $\hat{s}(t)$1 ms for Federated RL without MARL, approximately $\hat{s}(t)$2 ms for Independent RL, and approximately $\hat{s}(t)$3–$\hat{s}(t)$4 ms for the Centralized RL oracle (Ray, 28 Jul 2025). At an offered load of $\hat{s}(t)$5 Mbps, throughput is reported as approximately $\hat{s}(t)$6 Mbps for EdgeAgentX-DT, approximately $\hat{s}(t)$7 Mbps for EdgeAgentX, and approximately $\hat{s}(t)$8 Mbps for Federated RL without MARL, with the DT-based system yielding approximately $\hat{s}(t)$9 higher throughput than the next-best baseline (Ray, 28 Jul 2025).

Learning convergence also improves. EdgeAgentX-DT reaches a plateau after approximately $s(t)$0–$s(t)$1 episodes, compared with approximately $s(t)$2 for EdgeAgentX, approximately $s(t)$3 for Federated RL without MARL, and slower convergence to a lower asymptote for Independent RL (Ray, 28 Jul 2025). The reported summary is approximately $s(t)$4–$s(t)$5 fewer episodes than EdgeAgentX and reduced variance across seeds (Ray, 28 Jul 2025).

Under jamming, EdgeAgentX exhibits throughput degradation of approximately $s(t)$6 and latency increase of approximately $s(t)$7 relative to the no-jamming baseline, whereas EdgeAgentX-DT exhibits throughput degradation of approximately $s(t)$8 and latency increase of approximately $s(t)$9 (Ray, 28 Jul 2025). Under poisoning with $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$0 malicious updates, robust aggregation and twin validation are reported to prevent degradation, with DT-based performance remaining stable while vanilla FedAvg fails under attack (Ray, 28 Jul 2025).

The ablation EdgeAgentX-DT without GenAI converges in approximately $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$1–$$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$2 episodes and yields throughput approximately $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$3 better than original EdgeAgentX, but with weaker resilience to composite jamming-plus-failure events (Ray, 28 Jul 2025). This isolates the contribution of generative scenario diversity relative to the twin alone.

6. Compound-stressor behavior, security posture, and limitations

A central case study in EdgeAgentX-DT involves simultaneous stressors: at $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$4 s a jammer activates within a $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$5 km radius producing approximately $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$6 packet loss locally; at $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$7–$$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$8 s network-wide load doubles; and at $$\hat{s}(t) \leftarrow f(\hat{s}(t), s_{\text{real}}(t)).$$9 s a key relay node fails (Ray, 28 Jul 2025). In that scenario, conventional routing suffers more than $\mathcal{M}_T$0 throughput drop and latency spikes greater than $\mathcal{M}_T$1 s, with fragmentation and heavy loss (Ray, 28 Jul 2025). EdgeAgentX experiences an initial drop to approximately $\mathcal{M}_T$2 throughput, recovers to approximately $\mathcal{M}_T$3 by around $\mathcal{M}_T$4 s, and sees latency approximately double during adaptation (Ray, 28 Jul 2025). EdgeAgentX-DT shifts channels and paths proactively, exhibits relay-like behavior around the jammed zone, balances redundancy and load during the surge, dips to approximately $\mathcal{M}_T$5 throughput and recovers to approximately $\mathcal{M}_T$6 by around $\mathcal{M}_T$7 s, with latency increasing only approximately $\mathcal{M}_T$8 briefly (Ray, 28 Jul 2025). Over the $\mathcal{M}_T$9 s stress window, it delivers approximately $\hat{s}(t)$0 more data than EdgeAgentX and approximately $\hat{s}(t)$1 more than conventional routing (Ray, 28 Jul 2025).

Security and robustness are inherited partly from EdgeAgentX and partly from DT-specific controls. From the original framework, adversarial defense includes anomaly detection on model updates, adversarial training against jamming and perturbations, and secure communication through authentication and encryption (Ray, 24 May 2025). EdgeAgentX-DT adds twin isolation, encrypted telemetry, access control, and policy rollback if twin validation fails (Ray, 28 Jul 2025). The twin serves as a safe sandbox for adversarial training and red-team “wargaming” without risking mission assets (Ray, 28 Jul 2025).

Several limitations are explicit. Evaluation remains simulation-based with $\hat{s}(t)$2 nodes, and broader-scale field validation is pending (Ray, 28 Jul 2025). Twin fidelity and sync latency are assumed sufficient for effective transfer; events faster than the synchronization cadence $\hat{s}(t)$3 may be missed without predictive stepping (Ray, 28 Jul 2025). Scenario coverage depends on generator training data, so black-swan events remain difficult (Ray, 28 Jul 2025). Compute requirements at the coordinator or twin host are nontrivial, and the system still relies on connectivity to disseminate updates (Ray, 28 Jul 2025).

There is also a broader systems implication from adjacent edge-agent security work. Studies of edge-local and hybrid agent deployments on IoT hardware emphasize that deployment architecture itself shapes attack surface, particularly with respect to provenance enforcement, failover windows, and silent boundary crossings under fallback (Zhan et al., 26 Feb 2026). Although EdgeAgentX-DT is not evaluated in that home-automation setting, a plausible implication is that its twin synchronization, fallback, and policy-distribution channels require architectural security analysis in addition to model-level robustness claims.

Future directions identified for EdgeAgentX-DT include field trials on real radios with emulation-backed twins, quantification of minimum fidelity for effective transfer, adaptive twin fidelity and predictive simulation, online generative-model updates with telemetry, hierarchical or federated twins for larger-scale networks, more sophisticated adversaries, and human-AI teaming interfaces for scenario steering and policy transparency (Ray, 28 Jul 2025). These directions indicate that the framework is best understood as an initial integrated blueprint for twin-enabled resilient edge intelligence rather than as a finalized operational doctrine.