⚡ electrical engineering

ADMOS: An Adaptive Dual-Mode Operational Swarm for Resilient Asset Defense under Communication Jamming

This paper introduces ADMOS, an adaptive dual-mode UAV swarm architecture that combines distributed state estimation, a hysteresis-based hybrid supervisor, and higher-order control barrier functions to maintain resilient asset defense and zero safety violations even under severe communication jamming where traditional consensus-based systems fail.

Original authors: Mohamed Hijazy Shazin Hassan, Dandu S. N. Venkata Pavan Raghavendra Varma, Jaswanth Kumar Bobbadi, Neeraj Sree Mailee Dudaboyina, Narendra Sirisipalli, Sri Hari Sirisipalli, Giri Rajasekhar Gunnu

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mohamed Hijazy Shazin Hassan, Dandu S. N. Venkata Pavan Raghavendra Varma, Jaswanth Kumar Bobbadi, Neeraj Sree Mailee Dudaboyina, Narendra Sirisipalli, Sri Hari Sirisipalli, Giri Rajasekhar Gunnu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a team of tiny, buzzing drones tasked with guarding a precious treasure chest. Their superpower isn't just flying; it's talking to each other. They share a "hive mind," constantly updating a shared map of where everyone is and where the bad guys are. This works beautifully—until a sneaky enemy hits them with a radio jammer.

In the old way of doing things (the "consensus baseline"), if the jammer cuts off the chatter, the drones get confused. They keep trying to agree on a plan using stale, broken information. It's like a group of friends trying to dance in the dark while someone keeps changing the music; eventually, they trip over each other. The paper shows that once the radio noise gets too loud (about 40% of messages lost), this old system completely collapses, and the treasure is lost.

Enter ADMOS, a new, smarter way to run the swarm. Think of ADMOS not as a single brain, but as a team of independent agents with a very clever "Plan B."

The Three Magic Ingredients

ADMOS works by combining three specific tricks, like a superhero suit with three distinct power-ups:

The "Don't Guess" Map (Covariance Intersection):
When the drones talk, they sometimes repeat the same story. Old systems get overconfident, thinking they know the truth because they heard it twice. ADMOS uses a special math trick called Covariance Intersection. Imagine if you and a friend both tell you a story, but you aren't sure if you're hearing the same event twice. Instead of doubling your confidence, ADMOS says, "Okay, I'll be a little more cautious." It keeps the map honest, even when the radio is full of static. This prevents the drones from becoming dangerously overconfident about where they are.
The "Cool-Down" Switch (Hybrid Supervisor):
This is the decision-maker. It watches how "noisy" the connection is. If the noise gets too high, it flips a switch: "Okay, team chat is dead. Everyone, go solo!" But here's the clever part: it doesn't flip-flop. It uses a "hysteresis" rule (like a thermostat that doesn't turn the heat on and off every second) and a "dwell time" (a mandatory waiting period). This stops the drones from panicking and switching modes back and forth like a flickering lightbulb. Once they go solo, they stick with it for a while to make a solid decision.
The "Safety Net" (Control Barrier Functions):
Even when flying solo, the drones have a strict safety rule. A mathematical "force field" (a quadratic program) constantly checks their path. If a collision is about to happen, this safety net overrides their aggressive attack plan and says, "Brake!" It's like having a guardian angel that won't let you crash, even if you're trying to be a hero.

The Big Test: Simulated Chaos

The authors didn't just guess; they ran 80 different simulations for each scenario, pitting ADMOS against the old "consensus" team. They tested three situations:

Perfect Radio: The old team got the treasure 78.4% of the time. ADMOS got it 92.5% of the time.
Total Jamming (Radio Dead): The old team's performance crashed to 42%. They gave up. ADMOS, however, kept protecting the treasure 84% of the time.
Outnumbered: When the bad guys outnumbered the drones 3-to-1, the old team failed miserably (45%), while ADMOS held the line at 78%.

The most important discovery wasn't just that ADMOS won more often; it was how it lost. As the jamming got worse, the old team's performance dropped like a stone off a cliff. ADMOS, however, slid down a gentle slope. It degraded "gracefully." Even when the radio was completely dead, the drones didn't panic; they just switched to their solo mode and kept doing their job.

What This Isn't (And What's Next)

It's important to know what this paper doesn't say. This is a simulation, not a real-world flight test yet. The authors are very clear that they haven't flown these drones outside in the wind or tested them against real, physical jammers yet. They also ruled out "Byzantine" attacks (where an enemy pretends to be a friendly drone) and message forgery; their system assumes the messages are just lost, not faked.

The paper proves that in these high-fidelity computer simulations, this specific architecture works. It shows that letting a drone act "intelligently alone" when the group chat fails is a better strategy than trying to force the group to agree when they can't hear each other.

The authors suggest that if this works in the real world, it could protect power plants, ports, or event perimeters. But for now, the proof is in the code: a team that knows when to talk and when to go solo is much harder to jam than a team that just keeps shouting into the void.

Technical Summary: ADMOS – An Adaptive Dual-Mode Operational Swarm for Resilient Asset Defense under Communication Jamming

Problem Statement
Cooperative Unmanned Aerial Vehicle (UAV) swarms designed to defend fixed assets typically rely on continuous inter-agent communication to maintain a shared situational picture and execute consensus-based coordination. This architecture possesses a critical vulnerability: electronic attacks (radio jamming) that increase packet-loss rates degrade the shared state estimate. In standard consensus designs, estimators assuming independent measurements (e.g., Kalman filters) become overconfident when messages are reused across network cycles, leading to underestimated error covariances. Empirical evidence presented in the paper indicates that once packet loss exceeds approximately 40%, these systems suffer an abrupt, catastrophic collapse in coordination, rendering the swarm ineffective. The paper addresses the challenge of maintaining asset protection in the "communication-denied" regime without relying on stronger agreement mechanisms.

Methodology: The ADMOS Architecture
The authors propose ADMOS (Adaptive Dual-Mode Operational Swarm), an architecture composed of three distinct, composable elements designed to ensure resilience through intelligent autonomy rather than forced consensus.

Distributed State Estimation via Covariance Intersection (CI):
To address the issue of unknown cross-correlations induced by message reuse in lossy networks, ADMOS replaces standard independence-assuming fusion with Covariance Intersection. This method fuses local estimates ( $\hat{x}_i, P_i$ ) with neighbor estimates ( $\hat{x}_j, P_j$ ) using a weight $\omega$ optimized to minimize the trace of the resulting covariance.
- Mechanism: $P_i^+ = (\omega P_i^{-1} + (1-\omega)P_j^{-1})^{-1}$ .
- Guarantee: The fused estimate remains consistent (the reported covariance never underestimates the true error) regardless of the unknown correlation between inputs, preventing the "overconfidence" that drives consensus systems to diverge.
Hybrid Mode Supervisor with Hysteresis and Dwell Time:
Each agent operates in one of two discrete modes: Cooperative (C) or Autonomous (A).
- Cooperative Mode: Uses fused neighbor estimates for distributed guarding laws.
- Autonomous Mode: Relies solely on local sensing (within range $r_s$ ) for local interception laws.
- Arbitration: The supervisor monitors a link-health signal, defined as the trace of the error covariance ( $\eta_i = \text{tr}(P_i)$ ). Switching from C to A occurs when $\eta_i$ exceeds a high threshold ( $\eta_{hi}$ ), and reversion to C occurs when it drops below a low threshold ( $\eta_{lo}$ ).
- Stability: A minimum dwell time ( $\tau_D$ ) and hysteresis prevent "chattering" (rapid, unstable switching) and Zeno behavior, ensuring the system remains stable even under intermittent jamming.
Higher-Order Control Barrier Function (HOCBF) Safety Filter:
A safety filter sits atop the nominal control commands (from either mode) to guarantee forward invariance of the safe set (collision avoidance).
- Implementation: The filter solves a Quadratic Program (QP) at 10 Hz. Constraints are tightened by a robust margin proportional to the estimator uncertainty ( $\sqrt{\text{tr}(P_i)}$ ), ensuring that agents with higher uncertainty maintain larger separation distances.
- Fallback: If the QP becomes infeasible (e.g., in dense packing scenarios), the agent executes a braking command to halt, preserving safety at the cost of mission progress.

Stability and Theoretical Guarantees
The paper proves that the closed-loop system is Input-to-State Stable (ISS) with respect to estimation errors and disturbances for any assignment of modes across the swarm.

Theorem 1: Under assumptions of Lipschitz dynamics, shared equilibrium, and bounded disturbances, the hybrid system satisfies an ISS bound. The guarding error remains bounded by a function of the disturbance magnitude, ensuring that performance degrades gracefully rather than collapsing as packet loss increases.
Graceful Degradation: The architecture is designed such that as communication degrades ( $p \to 1$ ), agents transition to autonomous mode, and the ISS bound ensures the swarm remains stable and safe, albeit with reduced coordination efficiency.

Key Contributions
The paper makes four primary contributions, validated through high-fidelity simulation:

Consistent Estimation: Replacing independence-assuming fusion with Covariance Intersection to maintain estimator consistency under unknown correlations, trading peak sharpness for robustness.
Non-Zeno Supervision: A hybrid supervisor using hysteresis and dwell time that guarantees finite switching rates and prevents instability caused by mode chatter.
Feasibility-Aware Safety: A HOCBF-based safety filter that couples constraint tightening with estimator uncertainty, providing a formal safety guarantee that degrades gracefully with estimation quality.
ISS Certificate: A formal proof of Input-to-State Stability for the dual-mode system, establishing that the swarm remains bounded under any mode sequence, including total communication denial.

Results
The system was evaluated in a high-fidelity quadrotor simulator across 80 Monte Carlo trials per scenario.

Asset Protection Rate (APR):
- Healthy Links (S1): ADMOS achieved 92.5% APR vs. 78.4% for the consensus baseline.
- Full Jamming (S2): ADMOS maintained 84.0% APR, whereas the baseline collapsed to 42.0%.
- Numerical Disadvantage (S3, 3:1 intruder ratio): ADMOS achieved 78.0% vs. 45.0% for the baseline.
Degradation Profile: The baseline exhibited a sharp "cliff" in performance near 40% packet loss. ADMOS demonstrated a monotonic, gentle decline from 92.5% to 84% across the full range of packet loss (0% to 100%).
Safety: Zero safety violations (collisions) were recorded for ADMOS across all trials. The baseline incurred safety violations when jamming was present.
Ablation Study: Removing any single component (CI, dwell time, or HOCBF) resulted in significant performance drops or safety failures, confirming that each element is necessary.
Computational Cost: The full pipeline executes in under 1 ms per agent at a 10 Hz control rate, well within real-time budgets.

Significance and Claims
The paper claims that resilience in defensive swarms is achieved not by forcing agents to agree harder under adverse conditions, but by allowing them to act autonomously and intelligently when agreement is impossible.

Architectural Shift: The contribution is architectural and certificate-level rather than tactical. The specific guarding or interception laws can be swapped without invalidating the stability proof, provided the estimation, supervision, and safety layers remain.
Graceful vs. Catastrophic: The primary significance is the shift from "catastrophic collapse" (observed in consensus baselines) to "graceful degradation." An adversary can degrade the performance of an ADMOS swarm but cannot induce a sudden, total loss of coordination by crossing a critical jamming threshold.
Scope: The authors explicitly limit their claims to the defined threat model (radio jamming causing packet loss, no message forgery or Byzantine agents). They acknowledge that the results are currently based on simulation and that the double-integrator abstraction introduces a bounded disturbance gap relative to real quadrotor dynamics.

In conclusion, ADMOS demonstrates that a defensive swarm can maintain a credible level of asset protection even under total communication denial by decoupling estimation consistency, mode arbitration, and safety constraints under a unified stability framework.

The Three Magic Ingredients

The Big Test: Simulated Chaos

What This Isn't (And What's Next)

Technical Summary: ADMOS – An Adaptive Dual-Mode Operational Swarm for Resilient Asset Defense under Communication Jamming

More like this