The three-layer temporal decoupling architecture provides a practical template for safely integrating LLM-assisted planning with real-time multi-agent control in contested environments.
Abstract
Unmanned surface vehicle (USV) swarms operating in communication-denied maritime environments face degraded formation control when inter-agent state exchange is disrupted. This paper presents a three-layer control architecture integrating (1) LLM-assisted strategic mission planning with formal safety verification, (2) predictive tactical coordination combining physics-based motion extrapolation with online-learned neighbor behavior models, and (3) DMPC + ADMM execution for constrained formation control. The predictive coordination module in simulation-based evaluation across five representative scenarios reduces formation error by 76.0% (under simulation conditions) during 60 s communication outages compared to zero-hold prediction (Cohen d = 0.88, p < 0.001, DMPC + ADMM validated). The adaptive topology manager dynamically selects among star, mesh, and tree configurations via a utility function balancing communication quality, threat exposure, and overhead, reducing communication overhead by 71.7% (for the particular cases investigated) under the evaluated conditions while maintaining formation accuracy. Stability analysis using multiple Lyapunov functions and average dwell time theory guarantees global uniform exponential stability under topology switching, with explicit error bounds under communication denial derived via Gronwall-type arguments. The framework is validated through 1250 simulation trials across five scenarios with rigorous statistical analysis. The three-layer temporal decoupling architecture provides a practical template for safely integrating LLM-assisted planning with real-time multi-agent control in contested environments.
Cooperative unmanned-aerial-vehicle (UAV) swarm architectures considered in this study rely on inter-agent communication for distributed coordination. In the evaluated consensus baseline, coordination degrades sharply as packet loss approaches 40%, motivating a control architecture that remains operational when shared information becomes unreliable. We present ADMOS (Adaptive Dual-Mode Operational Swarm), an architecture comprising three coordinated elements: (i) distributed state estimation using Covariance Intersection (CI), which provides consistent fusion under unknown cross-correlation when its covariance assumptions hold; (ii) a hybrid supervisor that transitions each agent between cooperative and autonomous operation using hysteresis and a positive minimum dwell time to bound switching in the implemented controller; and (iii) a higher-order control-barrier-function (HOCBF) safety filter, solved as a quadratic program at 10 Hz, that provides the corresponding forward-invariance certificate when the tightened constraints are feasible and the stated uncertainty bound holds. Under the stated dynamics, bounded-disturbance, common-equilibrium, and dwell-time assumptions, we derive a conditional input-to-state-stability (ISS) bound for each admissible global mode assignment and switching sequence satisfying the stated Lyapunov conditions, and explicitly identify the conditions under which the certificate does not apply. In 80-trial simulations per scenario using a high-fidelity quadrotor simulator, ADMOS achieved asset-protection rates of 92.5% under healthy links and 84.0% under full jamming, compared with 78.4% and 42.0%, respectively, for the consensus baseline; no safety violations were observed in the evaluated trials under feasible operation and the stated sensing/uncertainty conditions. Ablation experiments quantify the contribution of the individual architectural elements. The observed resilience is associated with transitioning from cooperative estimation and control to bounded autonomous operation when shared information becomes unreliable; the claim is restricted to the evaluated scenarios and analytical assumptions.
Mohamed Hijazy Shazin Hassan, Dandu S. N. Venkata Pavan Raghavendra Varma, Jaswanth Kumar Bobbadi et al.· Discover Vehicles· 0 citations
This paper proposes AERIS, an offline policy improvement framework for multi-UAV ISAC that learns from fixed flight logs under centralized training and decentralized execution and designs STAR-CRDT, an offline multi-agent RL algorithm that performs support-aware local action rectification and distills only trusted improvements into the decentralized actor.
Ziyuan Wang, Yi-Fan Sui, Wei Wei et al.· 0 citations
Safe cooperative navigation of unmanned aerial vehicle (UAV) formations through three-dimensional environments with dense obstacles, dynamic threats, and unstructured terrain requires jointly addressing goal-directed navigation, formation keeping, hard safety constraints, and formation-topology consistency. We present FALCON-MASAC, a safety-integrated multi-agent reinforcement learning framework that decomposes this task into four complementary layers: (1) a hierarchical leader-follower paradigm that pairs a pre-trained virtual leader with followers learning a distributed cooperative policy; (2) a dual-scale entity-risk attention encoder (DSER-AE) that structures heterogeneous observations into entity-scale and risk-scale semantic tokens and fuses them through intra-scale self-attention and inter-scale full-sequence self-attention over the concatenated tokens; (3) a safety shield built on signed distance functions and high-order control barrier functions (SDF-HOCBF) that maintains the conditional safety certificate while the robust quadratic program remains feasible and the high-order admissibility conditions hold, and switches to a bounded best-effort fallback otherwise; and (4) a bypass-side commitment coordination layer that suppresses trajectory chattering and mitigates crossing conflicts among neighboring UAVs. The safety analysis explicitly quantifies the SDF linearization error and dynamic-obstacle prediction uncertainty and describes the bounded fallback used when certified execution is unavailable. Under the centralized-training-with-decentralized-execution (CTDE) paradigm, simulation experiments show that FALCON-MASAC substantially outperforms representative baselines: it attains a 97.3% success rate (21.7 percentage points above the strongest pure-MARL baseline and 13.0 percentage points above a post-hoc CBF-RL variant), a success-conditional steady-tracking formation error of 1.524 m, and a minimum clearance of 2.732 m. The framework also generalizes well along two out-of-distribution dimensions: obstacle density and dynamic-obstacle speed.
Yiming Shang, Changping Du, Rui Yang et al.· Unmanned Systems· 0 citations
The mission planning problem of using ground-fixed communication base stations to guide Unmanned Aerial Vehicle (UAV) swarms back under communication denial is addressed. The core challenge is to optimally match limited resources with massive UAV demands under constraints such as time windows, base station exclusivity, and relay continuity, which we formulate as an NP-hard combinatorial optimization problem. We first build a static model maximizing comprehensive benefits, incorporating base station heterogeneity and a super-linear congestion penalty for load balancing. We then extend it to a rolling horizon dynamic framework. Through task state partitioning and frozen resource inheritance, this extension decomposes the long-term optimization into sequential finite-horizon subproblems, enabling online decisions as UAV information is gradually revealed. To solve these models, we propose CMSA-MSWOA, which integrates elite opposition-based learning and Lévy flights to navigate the fragmented feasible solution space. Simulation results show 100% guidance coverage across scales from 100 to 500 UAVs in static scenarios, with the benefit advantage over the best benchmark growing from 10.0% to 65.5% as scale increases. In dynamic scenarios, the rolling framework satisfies all constraints and achieves full coverage. While our framework performs robustly in simulations, the current evaluation assumes idealized communication conditions; validation under more complex interference and external testing remains future work. Overall, our model and algorithm offer a useful simulation-based closed-loop framework for resource scheduling in denial environments, providing a foundation for further validation under more realistic field conditions.
High-efficiency area coverage using Unmanned Aerial Vehicle swarms is a fundamental capability for large-scale search-and-rescue missions. In disaster scenarios lacking prior knowledge, where agents rely on local sensing and decentralized decision-making, stochastic exploration strategies provide a practical solution. Asymmetric Lévy flight has recently emerged as a promising theoretical framework for scalable autonomous coverage in complex environments. However, existing studies primarily focus on kinematic optimization in particle-level simulations and rarely consider practicable strategies tailored for real-world swarm deployment. This limitation becomes particularly critical in dense swarms, where execution-level deadlocks frequently arise as agents mutually obstruct access to their intended waypoints, undermining coverage continuity and reliability. When deployed on physical platforms with standard obstacle-avoidance mechanisms, these deadlocks can severely degrade operational performance. To address this challenge, this paper proposes D-ALF, a stochastic swarm coverage planner augmented with a deadlock-aware exploration policy. By integrating an asymmetric collaborative consensus mechanism, the framework enables agents to actively resolve local congestion, maintain continuous and coordinated stochastic exploration, and translate theoretical coverage efficiency into physically realizable swarm operation. Extensive simulations and real-world experiments demonstrate that D-ALF actively accounts for execution-level deadlocks and consistently enhances coverage efficiency, operational reliability, and scalability compared with classical asymmetric Lévy flight strategies.
The integration of unmanned aerial vehicles (UAVs) into shared airspace requires on-board safety mechanisms for self-separation without cooperation between platforms or centralised air-traffic services. This paper proposes the UAV Safety Bubble (USB), which is a six-layer geometric model that defines an adaptive safety envelope around the UAV. Each layer corresponds to a distinct physical contribution: platform dimensions, positioning uncertainty, communication performance, wind disturbance, detection-processing latency, and avoidance manoeuvrability. The model was evaluated through 5300 closed-loop Monte Carlo simulation instances across six conflict scenarios: static-obstacle avoidance; head-to-head encounters; 90°, 30°, and 60° approaches between two dynamic UAVs; and a non-reacting-intruder case. Platform, sensing, communication, and environmental parameters were sampled from uniform distributions across operational ranges. Under these simulation assumptions, no instance produced an intrusion into the fifth layer (USB5), which was the model’s operational separation boundary. The smallest clearance between the USB5 outer edge and the obstacle’s centre of mass was 35.36 m, recorded in the 90° approach scenario, and the one-sided 95% upper bound on the composite intrusion probability over the 1000 independent parameter sets was 0.299%. The adaptive envelope achieved this at per-scenario median route-length overheads of 15–46% relative to nominal straight-line routes. The USB complements downstream planners by providing an adaptive no-go region. The avoidance manoeuvres evaluated are restricted to the horizontal plane; full 3D avoidance is left for future work.
Dominik Jerinić, T. Radišić, Karolina Krajček Nikolić et al.· Drones· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
