When a radio jammer floods the airspace around a protected asset, a cooperative drone swarm can lose the very thing that makes it a swarm: the constant exchange of neighbour states that keeps its members coordinated. A new simulation study reports an architecture designed to survive exactly that loss, by letting each drone decide when to stop trusting shared information and fall back on its own sensors. In the reported trials, the approach held asset protection at 84.0 percent under complete communication denial, while a consensus-based baseline collapsed to 42.0 percent.
The study, published as open-access research in Discover Vehicles, introduces ADMOS, short for Adaptive Dual-Mode Operational Swarm. The work was carried out by Mohamed Hijazy, Shazin Hassan, and colleagues, with corresponding author Giri Rajasekhar Gunnu. The paper frames the problem precisely: cooperative unmanned-aerial-vehicle architectures rely on inter-agent communication for distributed coordination, and in the evaluated consensus baseline coordination degrades sharply as packet loss approaches 40 percent. A jammer that raises packet loss on the shared channel can therefore cripple a defensive swarm without ever touching an individual airframe.
ADMOS combines three coordinated elements. The first is distributed state estimation using Covariance Intersection, a fusion method that remains consistent even when the correlation between two incoming estimates is unknown. This matters because an estimator that assumes independence can become overconfident when previously shared state information is rebroadcast and fused again, since the resulting estimates are correlated. Covariance Intersection avoids that failure mode at the price of conservatism: when correlations are weak, its covariance can exceed what known correlations would allow.
The second element is a hybrid supervisor that switches each agent between cooperative and autonomous operation. The switch is governed by a normalized estimator-health signal, computed as the trace of the position-error covariance divided by a fixed reference scale, which makes the quantity dimensionless. When the signal crosses an upper threshold of 1.6, the agent transitions to autonomous mode; it returns to cooperative mode only when the signal falls below a lower threshold of 0.8. The gap between the two thresholds provides hysteresis, and a minimum dwell time of 0.25 seconds limits how often the controller may switch. Together these rules bound the switching rate in the implemented sampled controller, with every accepted transition separated by at least the enforced dwell interval.
The third element is a safety filter built on higher-order control barrier functions, solved as a quadratic program at 10 hertz. Because the barrier function has relative degree two under the double-integrator dynamics used for analysis, acceleration enters only at the second derivative, requiring the higher-order construction. The filter computes the minimum-norm modification of the nominal control command that keeps the tightened safety constraints satisfied, and it tightens those constraints using the estimator’s own uncertainty: the certified safe region contracts as uncertainty grows. When the quadratic program becomes infeasible, the system invokes a braking controller whose formal safety guarantee holds only inside a defined recoverable set; outside that set, the authors make no unconditional safety claim.
On the analytical side, the paper derives a conditional input-to-state-stability bound for each admissible global mode assignment and switching sequence, under stated Lyapunov, common-equilibrium, bounded-disturbance, and dwell-time assumptions. The certificate is explicitly conditional. If the cooperative and autonomous modes do not share a common equilibrium, if the dwell-time condition is violated, or if the disturbance bounds fail, the corresponding guarantee does not apply, and the paper identifies these failure boundaries rather than burying them.
The evaluation used a custom high-fidelity six-degree-of-freedom quadrotor simulator with actuator saturation and an inner-loop attitude controller, run on a Raspberry Pi 4 Model B companion-computer configuration. Each scenario comprised 80 Monte Carlo trials with randomized initial conditions and intruder strategies, and each engagement lasted 60 seconds. Defenders started in a 24-to-26-metre radial annulus around the asset, intruders in a 45-to-60-metre annulus, with intruder speed capped at 4.0 metres per second. The packet-erasure channel was swept from 0 to 100 percent loss. The measured end-to-end control cycle averaged 0.58 milliseconds, with a 95th-percentile value of 0.95 milliseconds, comfortably inside the 100-millisecond control budget.
The headline results show a consistent advantage. Under healthy links, ADMOS achieved an Asset Protection Rate of 92.5 percent against the baseline’s 78.4 percent. Under complete jamming, ADMOS reached 84.0 percent while the baseline fell to 42.0 percent, meaning ADMOS lost 8.5 percentage points across the sweep while the baseline lost 36.4. Under a three-to-one intruder numerical advantage, ADMOS attained 78.0 percent against 45.0 percent. The authors are careful about statistics: the 14.1-point healthy-link difference falls below the conservative distribution-free Hoeffding screening bound of 30.36 percentage points for 80 trials, so no separation is claimed there, while the 42.0- and 33.0-point differences under jamming and numerical disadvantage do exceed it. No safety violations were observed in the reported feasible-operation trials, and minimum inter-agent distance stayed at or above the required minimum throughout.
Ablation experiments in the full-jamming scenario quantified each component’s contribution. Removing Covariance Intersection dropped protection to 58.0 percent as the independence-assuming estimator became overconfident. Removing the dwell-time mechanism dropped it to 71.0 percent through excessive mode switching. Removing the safety filter left protection nearly intact at 82.5 percent but produced nonzero safety violations, making that variant unsafe under the tested conditions. A packet-loss sweep further showed the baseline degrading sharply near 40 percent loss while ADMOS declined gradually, though the authors note this supports the graceful-degradation objective only for the tested configurations and does not establish a universal critical jamming threshold.
The limitations are stated plainly. All results come from simulation; the gap between the double-integrator analytical model and the quadrotor realization is represented by a bounded disturbance envelope of 0.50 metres per second squared, which is not hardware-flight calibration. The threat model covers only radio jamming on an independent packet-erasure channel: spoofing, message forgery, Byzantine behaviour, and direct sensor compromise are excluded, and an adversary that forges messages would defeat the consistency premise behind Covariance Intersection. The dwell-time latency is potentially exploitable by a jammer that toggles faster than the dwell interval, and the safety guarantee is conditional on quadratic-program feasibility. The authors list hardware-in-the-loop and outdoor flight validation as the most immediate next step, followed by extension to integrity-protected links and adversaries beyond the current threat model. Within those boundaries, the study supports a specific and useful claim: a swarm that knows when to stop sharing can keep defending when the spectrum is taken away.
Subject of Research: A dual-mode UAV swarm control architecture that switches from cooperative to autonomous operation under communication jamming, evaluated in simulation.
Article Title: Adaptive dual mode operational swarm control for resilient asset defense under communication jamming
Article References: Hassan, M. H. S., Varma, D. S. N. V. P. R., Bobbadi, J. K., Dudaboyina, N. S. M., Sirisipalli, N., Sirisipalli, S. H., & Gunnu, G. R. (2026). Adaptive dual mode operational swarm control for resilient asset defense under communication jamming. Discover Vehicles, 2(1), Article 34. https://doi.org/10.1007/s44465-026-00041-0
Image Credits: AI Generated
DOI: 10.1007/s44465-026-00041-0
Keywords: UAV swarm, communication jamming, Covariance Intersection, control barrier functions, hybrid control, asset defense, packet loss, input-to-state stability, quadrotor simulation, resilient autonomy, mode switching, safety filter
News Source: Denise Maddox. (October 29, 2026). Drone Swarm Switches to Autonomy When Jamming Cuts Its Communications. Scienmag.
