Low-altitude uncrewed aerial vehicle (UAV) swarms require reliable coordination under uncertain dynamics, limited wireless communication resources, and rapidly changing operating conditions. Continuous information exchange can introduce considerable communication overhead, especially when multiple UAVs share bandwidth-constrained networks. This article develops an AI-enabled decentralized networked control framework with dynamic event-triggered communication for low-altitude UAV swarms. A robust adaptive controller is first constructed to accommodate nonlinear coupling effects and unknown time-varying parameters. A lightweight neural predictor is then incorporated into the event-triggering mechanism to adjust communication thresholds according to the predicted evolution of the system error. Consequently, control information is transmitted only when necessary, reducing redundant network updates while maintaining closed-loop control performance. The proposed framework also provides a decentralized implementation in which each UAV relies primarily on locally available information, improving its suitability for scalable swarm systems. Compared with decentralized intermittent-feedback designs for nontriangular, time-varying systems, the framework also uses predicted state error in dynamic communication scheduling. Numerical experiments with interconnected UAV subsystems demonstrate that the proposed strategy maintains bounded system states and control inputs while significantly reducing communication transmissions. In particular, communication reductions of 45.5% and 95.1% are obtained for the two control channels. The results indicate that AI-assisted event-triggered control (ETC) can effectively coordinate control performance and communication efficiency in resource-constrained UAV swarm networks.
Control barrier function (CBF) provides an optimization-based framework for guaranteed obstacle avoidance in robotic systems. However, its performance can degrade under system perturbations. This article presents a safety-aware predictive control approach using input-to-state safety CBFs (ISSf-CBFs) for a four-wheel independent steering and four-wheel independent driving (4WIS–4WID) mobile robot. A discrete bounded error disturbance observer (DBEDOB) estimates external disturbances and provides bounds on the estimation errors. The ISSf-CBF constraints incorporate these disturbance estimates and error bounds, with tightened formulations to ensure robust obstacle avoidance and safety guarantees. Integration of these constraints into a model predictive control (MPC) framework enables safe robot operation. Simulation results demonstrate the proposed method’s effectiveness in maintaining CBF performance and robot safety under varying disturbances.
This article investigates localized data-driven self-triggered control (STC) for multiagent systems (MASs), where each follower computes its local control gain using only locally collected state and input data. Noisy data-driven protocols are developed to address the heterogeneity of control gains induced by locally sampled data. A system lifting method is employed to reformulate the model-based triggering condition as a tractable data-driven self-triggering mechanism (STM). Data-based stability conditions are established in the form of linear matrix inequalities (LMIs) for each follower, and their feasible solutions are used to construct the triggering matrix for determining the next transmission instant. The proposed data-driven STC conditions involve only low-dimensional LMIs. Simulation results on autonomous vehicle systems validate the effectiveness of the proposed method.
This article develops two funnel-based output feedback resilient control schemes for a class of uncertain nonlinear cyber–physical systems (CPSs) subject to false data injection attacks through sensor networks. An improved prescribed performance funnel is constructed by combining a shifting function with a finite funnel. Its reciprocal boundary starts from zero, thereby removing the conventional initial-feasibility restriction, and the output reaches a user-specified accuracy after a prescribed time. Because neither the system states nor the attack-free output is available, a state observer is driven only by the compromised output. A Lyapunov-based logic switching rule is then integrated into the controller to compensate for the unknown combined control and attack direction while ensuring finitely many switches and a positive dwell time. A Nussbaum-based controller provides an alternative continuous adaptation strategy. The proposed funnel controllers guarantee that the output signal always stays within the improved prescribed performance funnel and that all signals in the closed-loop system remain bounded. Simulations and controlled comparisons on a single-link robotic manipulator illustrate the prescribed performance and resilience properties of both schemes.
Transformer-based large language models (LLMs) have shown strong potential for natural-language-driven robotic control and code generation, while mixture-of-experts (MoE) architectures provide an efficient sparse activation paradigm for scaling model capacity. Deploying LLMs on uncrewed aerial vehicle (UAV) edge platforms can improve mission responsiveness, autonomy, and communication robustness. However, the large memory footprint, loading latency, and inference cost of existing models remain major obstacles for resource-constrained UAV systems. We propose DroneMoE, a lightweight MoE framework tailored for drone edge deployment. Built on DeepSeek-Coder-V2-Lite and fine-tuned via quantized low-rank adaptation (QLoRA), DroneMoE combines static expert pruning with dynamic expert activation. The former removes low-importance experts before deployment to reduce memory footprint, while the latter adaptively selects experts using quantile-based routing confidence thresholds to accelerate inference beyond fixed Top- $k$ routing. We curated a dataset mapping natural-language instructions to Python control code across multiple UAV software development kits (SDKs). Experimental results demonstrate that DroneMoE reduces memory usage, loading time, and inference cost while maintaining acceptable UAV control-code generation quality under both single-SDK and multi-SDK settings. Users can flexibly adjust thresholds to fit specific drone hardware constraints and SDK requirements. Code and datasets will be publicly available at https://github.com/Yizhan-FENG/DroneMoE
