Jing Xu

Deploying high accuracy models on resource-limited edge is essential for industrial applications, yet challenging due to the diverse and dynamic real-world environment. Cloud–edge collaboration helps by training high-capacity models in the cloud and deploying compressed models to edge nodes. However, traditional pipelines assume a fixed setting and lack the flexibility to adapt compression strategies to the heterogeneous and evolving environment. To overcome this, we propose CE-MAC, a cloud–edge collaborative large language model (LLM)-based multiagent system (LLM-MAS) for adaptive edge model compression. We adopt model pruning for its support of customizable and progressive compression, enabling flexible adaptation in dynamic edge environments. The LLM-MAS acts as the decision hub, generating device-specific pruning policies based on real-time feedback from both cloud and edge, enabling better cloud–edge co-evolution. We evaluate CE-MAC on image classification, keyword spotting, bearing fault diagnosis, and electricity theft detection tasks, achieving state-of-the-art performance across heterogeneous data and varying compression levels.

With the rapid advancement of artificial intelligence, multi-agent systems (MASs) are evolving from classical paradigms toward architectures built upon large foundation models (LFMs). This survey provides a systematic review and comparative analysis of classical MASs (CMASs) and LFM-based MASs (LMASs). First, within a closed-loop coordination framework, CMASs are reviewed across four fundamental dimensions: perception, communication, decision-making, and control. Beyond this framework, LMASs integrate LFMs to lift collaboration from low-level state exchanges to semantic-level reasoning, enabling more flexible coordination and improved adaptability across diverse scenarios. Then, a comparative analysis is conducted to contrast CMASs and LMASs across architecture, operating mechanism, adaptability, and application. Finally, future perspectives on MASs are presented, summarizing open challenges and potential research opportunities.

On-device deployment often relies on lightweight models that require fine-tuning when deployment environments differ from development conditions or when model performance changes over time. Although recent LLM-enabled agents are capable of reasoning about system states and forming adaptive strategies, they often incur substantial computational and budgetary overhead, making them impractical for resource-constrained edge settings. This paper presents Budgeted Agentic AI for Adaptive Lightweight Model Fine-Tuning (BA2), a framework for budget-constrained lightweight model adaptation that combines bounded optimization steps with auxiliary evaluation, rollback, and compression actions. BA2 enables agents to dynamically adjust adaptation strategies according to real-time budget conditions while maximizing system performance. BA2 explicitly accounts for constrained resources, including limited fine-tuning steps, tool invocation quotas, and token consumption. An Engineer–Manager architecture is introduced, where an LLM-based Engineer generates parameterized candidate operations from system logs, and a lightweight contextual-bandit Manager selects actions conditioned on both environmental states and remaining budgets. Experimental results demonstrate that BA2 achieves superior cost-performance trade-offs compared with static heuristics and inference-time search agents under tight budget constraints, while remaining competitive with strong baselines in higher-budget regimes.

This article addresses critical challenges in multiagent containment control (MACC) under adversarial disturbances and hard constraints by integrating zero-sum differential games, collision avoidance, energy efficiency, and formation preservation into a unified framework. First, a terminal-cloud collaborative control framework is built. For the cloud agent, a barrier-function-augmented reinforcement learning (RL) mechanism ensures collision-free path planning, and an integral RL algorithm via policy iteration solves the two-player zero-sum differential game for optimal control against adversarial disturbances. For the terminal agents, the event-driven distributed RL controllers optimize energy consumption while maintaining formation tracking under topology switching. Furthermore, the hierarchical architecture enables cross-layer algorithmic decoupling and event-driven adaptive RL. Rigorous theoretical analysis establishes the asymptotic stability of containment errors and uniform ultimate boundedness of actor–critic weight estimation errors, even under resource limitations and dynamic obstacle environments. Simulations of satellite swarms validate the framework’s ability to navigate environments with dense obstacles, tolerate node failures, and adapt to intermittent connectivity.

This article investigates reach–avoid games involving defenders equipped with capture radii, where both defenders and attackers have different speeds. The main challenge lies in using geometric methods to analyze different speed ratios, construct barriers, or defensive advantage angles, and divide the state space into defensive and offensive advantage regions. This article proposes optimal analytical strategies for players based on the corresponding payoff functions, depending on the attacker’s position within different winning regions under various speed ratios. In addition, we demonstrate the existence of a unique optimal target point within the offensive advantage region. Unlike numerical methods, which are limited by computational complexity and real-time application capabilities, the proposed method allows for the precise calculation of barriers and real-time updates in nonpoint capture scenarios. Finally, simulation results validate the effectiveness of the constructed barriers in multiplayer reach–avoid games.

Vision-based reinforcement learning (RL) methods enable efficient policy learning and adaptive decision-making for quadrotor uncrewed aerial vehicles (UAVs) navigation in complex, high-dimensional flight environments. Although end-to-end vision-based RL approaches are effective, they often function as closed-box models, lacking interpretability. We develop an explainable vision-based hierarchical RL algorithm for QUAV navigation, integrating perception, obstacle avoidance, and motion control into a unified framework. Due to the high-dimensional state space and complex dynamics of QUAV tasks, traditional RL methods often suffer from sparse and difficult-to-obtain rewards. To address this, we introduce the echoic hindsight experience replay mechanism, which accelerates convergence by transforming failed episodes into successful ones. Building on this, we propose an RL-based proportional-integral-derivative-retarded control method that leverages multirate measurements to enhance low-level control performance, improving maneuverability and precision in QUAV operations. Both simulated and real-world experiments demonstrate the effectiveness of our proposed method for UAV navigation in complex environments.

As embodied intelligence (EI), large language models (LLMs), and cloud computing continue to advance, Industry 5.0 facilitates the development of industrial artificial intelligence (IndAI) through cyber-physical-social systems (CPSSs) with a human-centric focus. These technologies are organized by the system-wide approach of Industry 5.0, in order to empower the manufacturing industry to achieve broader societal goals of job creation, economic growth, and green production. This survey first provides a general framework of smart manufacturing in the context of Industry 5.0. Wherein, the embodied agents, like robots, sensors, and actuators, are the carriers for IndAI, facilitating the development of the self-learning intelligence in individual entities, the collaborative intelligence in production lines and factories (smart systems), and the swarm intelligence within industrial clusters (systems of smart systems). Through the framework of CPSSs, the key technologies and their possible applications for supporting the single-agent, multi-agent and swarm-agent embodied IndAI have been reviewed, such as the embodied perception, interaction, scheduling, multi-mode large language models, and collaborative training. Finally, to stimulate future research in this area, the open challenges and opportunities of applying Industry 5.0 to smart manufacturing are identified and discussed. The perspective of Industry 5.0-driven manufacturing industry aims to enhance operational productivity and efficiency by seamlessly integrating the virtual and physical worlds in a human-centered manner, thereby fostering an intelligent, sustainable, and resilient industrial landscape.

In this work, the multistage semiimplicit Euler digital implementation is proposed for a switched antiwindup super-twisting control (AWSTC) to achieve the precise and robust attitude tracking. The algorithm of AWSTC is designed based on the singular perturbation redesign, where the high gain feedbacks are automatically activated for the global stability of tracking errors in presence of controller saturation. As a result, the attitude tracking performance is significantly improved, even when the initial conditions are far away from the origin. Then, the semiimplicit Euler method of AWSTC is proposed to mix explicit and implicit discretizations for attenuating the undesirable numerical chattering of the explicit Euler methods. In this situation, the selection of a relatively large sampling period is permitted, and thus the overestimation of controller gains for inducing the numerical chattering is also avoided. Both the simulation and experimental results on the quadrotor platform confirm the effectiveness and merits of the proposed algorithm.