Keywords: Robotics, Motion Control, Modeling and Control of Complex Systems
Abstract: Developing agile flapping robots like insects is still an important task to achieve successful potential applications of the flapping robots. Authors have been working on developing flapping robots using a new type of piezoelectric material, that is, piezoelectric fiber composites, while mimicking the structure and the control of insects. To enhance the performance of the flapping robots, this paper describes an introduction of click mechanism to a flapping robot using piezoelectric fiber composites inspired from insects. The simulation-based design is performed to achieve flapping motions with large flapping amplitude and high speed by utilizing the click mechanism. A prototype of flapping robot using the designed click mechanism has been fabricated and large flapping amplitude and high speed due to the click mechanism have been confirmed by experimental results.

Keywords: Robotics, Motion Control, Modeling and Control of Complex Systems
Abstract: Mobile systems with multiple robots are widely used in warehouse and production logistics, where work efficiency is determined not only by the quality of local navigation of an individual robot, but also by the consistency of movement of the entire fleet. With the increasing number of robots in a limited environment, characteristic problems arise: conflicts at intersections, mutual delays, queues in narrow corridors and mutual blockages. Fully decentralized approaches provide good local adaptation, but do not guarantee a global order of movement. Fully centralized approaches allow you to create conflict-free plans, but they are sensitive to computational complexity and the mismatch between a discrete model and continuous execution. The paper proposes a hybrid fleet management framework for mobile robots that combines a centralized level of coordination based on multi-agent pathfinding and decentralized execution using the Nav2 navigation stack in the ROS2/Gazebo environment. Conflict-based search is used as a centralized planning method that forms a globally consistent discrete plan. For practical implementation, the Fleet Manager intermediate coordination module has been developed, which converts a discrete plan into a sequence of goals and waypoint commands for Nav2, tracks the progress of robots and initiates replanning in case of deviations. Experimental validation was conducted at two levels: a discrete Python model and continuous simulation in Gazebo. Three modes are compared: centralized coordination in a discrete model as a reference, decentralized execution using Nav2 alone, and hybrid management of MAPF + Nav2. The results show that in simple scenarios, Nav2-only mode can provide acceptable behavior, but in conditions of corridors and conflicts, the likelihood of blockages increases and waiting times worsen. The hybrid approach enhances stability and predictability of fleet motion by integrating global coordination with local adaptability.

Keywords: Robotics, Motion Control, Optimal Control
Abstract: Autonomous drone racing requires time-optimal trajectories that are both dynamically feasible under the full nonlinear quadrotor model and safe with respect to reliable gate traversal. This paper presents a tractable trajectory-optimization framework based on successive convexification augmented with systematic convexity exploitation. We incorporate geometric gate decomposition by representing gate-passage requirements as convex constraints (intersections of half-spaces and second-order cones), thereby certifying complete traversal without integer variables or intricate nonconvex constraint mechanisms. Beyond accelerating the main optimization, we show that the exploited convex structure also enables strong initialization: a single convex program can provide a high-quality initial guess and, in many cases, a competitive solution. Extensive comparative evaluations demonstrate significant computational improvements from successive convexification and convexity exploitation, while real-world flight experiments validate the physical feasibility and repeatability of the resulting trajectories.

Keywords: Robotics, Real-time Systems
Abstract: Reliable long-term localization is essential for autonomous navigation, yet significant drift often arises when transient environmental changes are incorrectly matched to the pre-built static map. We propose S2Loc, a hierarchical framework that replaces passive error tolerance with active constraint verification. By explicitly decoupling feature reliability from feature existence, the system ensures that only spatiotemporally consistent geometric features are utilized for localization. Specifically, the Stable Structure Perception (SSP) module functions as a strict validation layer to extract dominant structural planes while actively intercepting structural inconsistencies. This allows the Structure-Guided Matching (SGM) registration strategy to leverage exclusively these validated constraints to optimize the pose estimation. Extensive experiments on the NCLT dataset and in industrial environments subject to frequent layout changes demonstrate that S2Loc reduces translational errors by an average of 46.6%, maintaining robustness even under environmental variations (with a Global Change Rate reaching up to 41.8%).

Keywords: Robotics, Sensor/Data Fusion
Abstract: Autonomous exploration with Unmanned Ground Vehicles (UGVs) demands a balance between exploration efficiency and navigation safety. However, existing methods typically focus on known regions or frontiers, neglecting the informative potential of unknown regions,resulting in redundant backtracking or suboptimal efficiency. To address this issue, we introduce HSURE, a hierarchical safety-aware autonomous exploration framework guided by unknown regions. We explicitly decompose unknown regions by topological connectivity and construct an observability-driven safety assessment using dynamic sparse graphs in known space. Leveraging these hybrid representations, we devise a coarse-to-fine safety-aware exploration planner. The global planner computes an optimal coverage path and the local planner refines this into a safe and executable trajectory. Extensive simulation and real-world experiments demonstrate the proposed method outperforms state-of-the-art baselines, significantly reducing exploration time and trajectory redundancy while maintaining navigational safety.

Keywords: Robotics, Sensor/Data Fusion
Abstract: Compact passive single-axis prosthetic knees are widely prescribed because of their low mass and short build height, yet their swing motion is largely unregulated. At low cadences, this can cause (i) insufficient minimum toe clearance (MTC) due to premature knee extension after peak flexion, and (ii) knee-flexed initial contact caused by terminal impact and prolonged late swing, both of which increase fall risk. This paper presents a compact uniaxial prosthetic knee that electronically actuates a conventional load-brake through a toggle mechanism, enabling swing-phase knee-angle holding at arbitrary timings with reduced actuator torque. The knee applies braking at two key instants: it holds the knee at peak flexion until the predicted MTC instant and locks the knee at full extension in late swing until initial contact. A prototype (build length 205 mm, mass 918 g) was tested in level walking with one unilateral transfemoral amputee at cadences of 60–90 rpm, compared with two conventional passive knees. At 75 rpm, the proposed knee increased MTC from 0.5 mm to 23.7 mm and eliminated toe scuffing, while at 60 rpm, fall prevention was achieved by maintaining the fully extended position.

Keywords: Robotics, Sensor/Data Fusion, Modeling and Control of Complex Systems
Abstract: In the pursuit of general robotic manipulation, dexterous hands require reliable and scalable proprioception, which remains a key bottleneck due to limitations in sensing integration and generality. This paper presents HandCept, a visual-inertial proprioception framework for accurate joint angle estimation without relying on embedded joint sensors. HandCept integrates a wrist-mounted RGB-D camera and distributed 9-axis IMUs to estimate link poses through complementary sensing modalities. A zero-shot visual pipeline trained on synthetic data provides global 6D pose observations, while inertial measurements deliver high-frequency orientation updates. These asynchronous signals are fused via a latency-compensated Extended Kalman Filter, enabling real-time, drift-free estimation under dynamic conditions. Experimental results show that HandCept achieves joint angle errors within 2^{circ} to 4^{circ} without observable drift, outperforming visual-only and inertial-only baselines. The proposed framework further demonstrates strong IMU stability and cross-device uniformity, allowing a shared reference frame and simplified calibration. HandCept provides a generalizable and hardware-efficient solution for dexterous hand proprioception, supporting robust closed-loop manipulation in real-world environments.

Keywords: Adaptive Control, Nonlinear Systems and Control, Networked Control
Abstract: This paper proposes an adaptive trajectory tracking scheme for underactuated AUVs subject to time-varying ocean currents and uncertainties. A fast-convergent RBFNN observer with fractional-order terms is designed to accurately estimate disturbances at the kinematic level. By integrating Prescribed Performance Control (PPC) into the backstepping framework, tracking errors are strictly confined within a predefined performance funnel. Simulation results verify that the proposed method ensures superior transient stability and robust disturbance rejection compared to standard backstepping controllers.

Keywords: Adaptive Control, Fuzzy and Neural Systems, Multi-agent Systems
Abstract: This article addresses the asymptotic formation control issue for networked unmanned surface vehicles (USVs) with output and input constraints. Specifically, a leader-follower scheme is developed to make each USV follow its reference trajectory without collisions among vehicles. Moreover, a global prescribed performance control is proposed to ensure tracking errors converge to optimal constrained boundaries regardless of initial conditions. Additionally, an adaptive formation controller is designed with the auxiliary variable to guarantee asymptotic convergence of the tracking system towards the origin with a small residual set, where the first-order filtering is introduced to avoid repeated differentiation and the fuzzy logic system is employed to attenuate effects of uncertainties and disturbances. Based on Lyapunov stability theorem, the rigorous closed-loop stability is achieved in terms of convergence and boundedness. Finally, numerical simulations show the effectiveness of the proposed method.

Keywords: Learning-based Control, Nonlinear Systems and Control, Optimal Control
Abstract: 重型卡车的高速机动因非线性动力学而带来重大控制挑战，而高重心使其容易发生灾难性翻车。为此，我们提出了一个风险厌恶型非线性模型预测控制（NMPC）框架。 首先，构建结构化残差物理知情神经网络（rPINN），以补偿模型与实际非线性车辆动力学之间的不匹配，从而提升多步预测的准确性，同时确保物理一致性。 其次，通过基于网格的可达性分析离线构建安全操作包络（SOE）。 在不同速度下识别出稳定边界，并用凸多胞体近似实现实时优化。 此外，稳定包络映射到可微风险势场，并集成到NMPC中作为软约束。 该机制主动降低风险，防止车辆接近稳定边界。模拟表

Keywords: Learning-based Control, Nonlinear Systems and Control, Robotics
Abstract: Reinforcement learning (RL) has shown promise for quadrotor control, but sim-to-real transfer remains highly challenging: policies that work well in simulation often fail or even crash in reality. The core difficulty lies in unmodeled dynamics and disturbances, including actuator-level thrust dynamics affected by nonlinear aerodynamic drag, body-relative airflow, and multimodal environmental noise, that are usually omitted or over-simplified in simulation. To address these issues, we propose an RL-based controller with three key design choices: (i) careful selection of temporal parameters such as motor inertia and control delay to stabilize reward-action mapping, (ii) an input space design that uses body-rates history and rotation-matrix attitude representation to balance efficiency, robustness, and accuracy, and (iii) diffusion-based noise modeling that captures complex real-world disturbances beyond simple parametric approximations. We validate our approach on Crazyflie quadrotors across diverse and challenging trajectories unseen during training, where our policy significantly outperforms classical controllers in tracking accuracy. These results show that principled design in modeling and training can enable reliable zero-shot sim-to-real transfer of RL policies for quadrotor flight.

Keywords: Nonlinear Systems and Control, Learning-based Control
Abstract: Aiming at the challenge of achieving precise control performance when the flow angle of unmanned aerial vehicles (UAVs) cannot be measured directly, a predefined-time disturbance-rejection control (PTDRC) method without flow angle measurements is proposed. Based on the six-degree-of-freedom (6-DOF) nonlinear model of the UAV, an affine nonlinear form containing position, path angle, flow angle, angular rate, and velocity loop is derived; an flow angle estimation network (FAEN) is developed using a deep learning–based approach, in which the network structure contains three stacked gated recurrent unit (GRU) layers followed by a fully-connected layer; and a PTDRC method for the UAV without flow angle measurements is designed based on the estimated flow angle. Simulation results demonstrate that the FAEN exhibits stronger robustness against coefficient perturbations compared with the extended Kalman filter (EKF). Under different aerodynamic coefficient perturbations and turbulence intensities, the proposed PTDRC method achieves higher control accuracy and stronger disturbance rejection performance than the linear active disturbance rejection control (LADRC).

Keywords: Nonlinear Systems and Control, Learning-based Control, Motion Control
Abstract: Deep Reinforcement Learning (DRL) shows great potential in achieving high-performance autonomous navigation but suffers from high sample complexity and exploration risks. This paper proposes a safety-aware Model-Based Reinforcement Learning (MBRL) framework to address these issues. First, we integrate a High-Order Control Barrier Functions (HOCBFs) module to ensure safety. This module strictly constrains policy actions within safe regions, significantly minimizing training violations. Second, we apply a modular Stochastic Value Gradient (SVG) scheme that aligns policy updates with safety-filtered execution. Finally, to enhance sample efficiency, we employ a Dyna-style architecture augmented by a Sparse Gaussian Process (SGP) dynamics model. We initialize the model via offline pre-training and fine-tune it online. This design reduces model mismatch and keeps the computational cost manageable. Experimental results show improved learning efficiency and safety performance compared to the baselines.

Keywords: Learning Systems
Abstract: To address the artifacts arising from low-quality geometric initialization in 3D Gaussian Splatting under sparse views, this paper proposes a general, plug-and-play self-supervised point cloud optimization framework. Seamlessly integrated into the InstantSplat pipeline, our method leverages monocular depth estimation as a geometric constraint. Specifically, it constructs a gradient-based spatial confidence mask to suppress edge noise in the depth prior, dynamically calibrates depth scales via confidence-weighted statistical moment matching, and optimizes the point cloud using a combination of Charbonnier loss and an adaptive anchor regularization mechanism. Experimental results demonstrate that our approach effectively eliminates geometric artifacts, significantly enhancing the visual quality and robustness of novel view synthesis compared to the baseline.

Keywords: Learning-based Control, Robust and H infinity Control, Linear Systems
Abstract: In the present study, a Q-learning model predictive control with adaptive adjustment of learning period is proposed for unknown linear discrete-time systems to optimize the mixed H_2/H_infty performance index. To improve the learning efficiency of the control gain, the influence function is used to construct the quantitative relationship between the measured data and learning performance. Meanwhile, to achieve better control performance in the case of system disturbance mutation, the receding horizon optimization of model predictive control combined with Q-learning is presented. Finally, effectiveness of the proposed algorithm is numerically evaluated.

Keywords: Modeling and Control of Complex Systems, Sensor Networks
Abstract: Maximal covering location problems (MCLP) are fundamental in spatial optimization with widespread applications in sensor networks, smart cities, and robotics. The growing need for fine-grained decision-making in modern systems results in both large-scale candidate facility location and demand point sets, rendering this NP-hard combinatorial optimization beyond the practical computational limits of existing solvers. Simple aggregation methods improve scalability, but often incur a significant loss in solution quality, motivating the need for intelligent aggregation mechanisms. To address this, the paper proposes an adaptive aggregation framework that effectively reduces problem dimensionality while maintaining high-quality coverage. The proposed method establishes an iterative framework of aggregation-optimization-refinement, driven by novel criteria: a coverage inconsistency evaluation that detects discrepancies between approximate and actual coverage, and a local gain potential analysis that identifies regions with high potential for solution improvement. Using these indicators to selectively refine spatial granularity, the framework adaptively preserves high-quality coverage while maintaining computational efficiency. Simulations with up to 100,000 points demonstrate that the proposed algorithm achieves high-quality coverage comparable to leading solvers with substantial runtime reductions, confirming its scalability for solving large-scale MCLP.

Keywords: Energy Efficiency, Control Applications, Learning-based Control
Abstract: ，同时减少极化 4182; 延长电池寿命ᦁ 2; 精准的控制至关 7325; 要。 本文将深度结合 带脉冲充电的强 1270; 学习提出一个 新型充电控制方 7861;。 关注贫困群体 传统方法的泛化 3021; 力，这 方法动态调整初 2987; 条件 训练，使模型能 2815; 适应多样化的ࠠ 5; 电 环境并迅速响应 6032; 的作情况 场景。 它有效地增强了 7169; 型的 泛化能力，同时 6174; 著抑制 充电时的偏振效 4212;。 模拟结果 证明通过该方法 5292; 极化电压保持ߎ 1; 变 约0.2V，温度在 充电过程。

Keywords: Intelligent and AI Based Control, Learning Systems, Learning-based Control
Abstract: 视觉语言导航的泛化能力 （VLN）模型受到以下的稀缺性极大限制。 高质量、多样化的训练数据。传统数据 采集方法依赖于昂贵的机器人平台， 人工注解劳动密集型，导致有限的注释 数据规模和显著的场景偏差。为了克服这一点 瓶颈，本文提出了一种新颖的端到端解决方案 自动化VLN训练数据综合框架。其核心 创新在于高效地大规模生产 训练数据——具有精确几何结构和丰富的语义， 以及自然语言指令——仅使用多视图RGB 目标场景的图像序列作为输入。具体来说，我们 首先，采用三维高斯喷溅技术 重建高保真、可渲染的神经场景 从图像中表示。随后，我们设计了一个 数据驱

Keywords: Optimal Control
Abstract: Multi-UAV cooperative task allocation is a typical multi-objective constrained optimization problem with discrete combinatorial structure, strong constraint coupling, and conflicting objectives. To address these issues, this paper proposes a State-aware(S) Decomposition-guided(D) Multi-Objective Marine Predators Algorithm (MOMPA), abbreviated as SD-MOMPA. A multi-objective model is first established by considering total system energy consumption, maximum completion time, and load variance under task uniqueness, energy, capacity, and time-window constraints. Then, the decomposition idea from MOEA/D is introduced into MOMPA to enhance search directionality along different Pareto preferences; state-aware phase scheduling is designed to adaptively coordinate exploration and exploitation using population diversity, improvement rate, and feasibility ratio; and local replanning is applied to infeasible solutions to preserve favorable task-allocation structures. Comparative experiments show that SD-MOMPA achieves competitive convergence performance and obtains a well-distributed Pareto solution set, demonstrating its effectiveness for constrained multi-UAV task allocation.

Keywords: Multi-agent Systems
Abstract: This paper studies the consensus problem for stochastic multiagent systems (SMASs) under compound noises (additive and multiplicative noises) over stochastic networks via a Large Language Model (LLM).Existing challenges primarily stem from the difficulty in obtaining appropriate control gains to rapidly mitigate noises under stochastic network conditions.Furthermore, the simultaneous presence of these networks and compound noises complicates the transformation into unified error dynamics.To address these issues, this paper applies LLM inference to the adaptive selection of control gains alpha(k).By utilizing the LLM's powerful reasoning capabilities and ability to process complex contextual information, this approach dynamically selects appropriate control gains to cope within complex environments.Specifically, a stochastic approximation (SA) protocol is employed where the LLM-designed gains, subject to specific convergence constraints, effectively reduce the influence of noises for SMASs.In addition to ensuring state boundedness through Lyapunov theory, a semi-decomposition technique to establish consensus among agents is employed.The effectiveness of our LLM-based control scheme is ultimately demonstrated by a representative numerical example.

Keywords: Nonlinear Systems and Control, Multi-agent Systems
Abstract: This paper addresses the problem of distributed Nash equilibrium seeking for multi-agent systems composed of heterogeneous linear agents. Unlike most existing works that assume static cost functions or single-integrator dynamics, we consider a dynamic noncooperative game where the cost functions evolve over time and the agents are subject to high-order physical constraints. We conduct a hierarchical design consisting of a distributed algorithm for the Nash equilibrium seeking at the upper level and a reference-tracking controller at the lower level. Rigorous theoretical analysis of the system performance reveals that the algorithm performance is fundamentally limited by the variation rate of the realtime NE. Numerical simulations are presented to validate the effectiveness of the proposed algorithm.

Keywords: Learning-based Control, Robotics
Abstract: Dexterous manipulation of articulated objects necessitates navigating a high-dimensional configuration space riddled with discontinuous contact dynamics. Standard end-to-end Reinforcement Learning often struggles in this regime, frequently converging to jittery, unstable policies. To bridge this gap, we propose a hierarchical control framework that synergizes the extensive exploration capabilities of Population-Based Training with the structural priors of Geometric Fabrics embedded directly into the learning loop. We construct a decoupled attractor landscape that guides the arm and fingertips toward task-relevant poses, strictly confined by anisotropic repulsive metrics for self-collision avoidance and singular barrier potentials for joint limits. This formulation not only leverages the diverse experience collection of a population of agents but also effectively projects the RL problem into a physically consistent manifold, transforming the policy's role from raw motor command generation to high-level energy shaping. Extensive validation in the massively parallelized IsaacLab simulation environment demonstrates that our GF-Delta strategy achieves a 97.47% success rate in drawer-opening tasks. Compared to standard joint-space baselines, our method exhibits superior convergence efficiency and generates smooth, feasible trajectories without requiring complex reward engineering.

Keywords: Multi-agent Systems, Intelligent and AI Based Control, Motion Control
Abstract: To address the problem of multi-UAV cooperative pursuit of a high-speed evader in obstacle-rich environments, this paper proposes a hybrid cooperative pursuit algorithm trained under the centralized training with decentralized execution paradigm. The proposed algorithm employs an attention encoder to produce compact representations of dynamic neighbor observations and adopts the Twin Delayed Deep Deterministic Policy Gradient framework to output continuous potential field parameters, achieving cooperative pursuit and obstacle avoidance control through attention-weighted potential fields. The Dual-Channel Gradient Decoupling architecture is designed to segregate the multi-network optimization pathways, thereby resolving the Actor--Critic gradient conflict arising from the shared attention encoder. In addition, the Critic-Anchored KL Regularization is introduced to suppress the target-value non-stationarity caused by fluctuations in the attention distribution. Simulation experiments demonstrate that the algorithm attains an average capture success rate of 88% with a peak of 92%, outperforms the baseline in both convergence speed and final performance.

Keywords: Learning-based Control, Robotics, Control Applications
Abstract: This paper proposes a deep Koopman operator-based output regulation framework to address the pursuit-evasion (PE) game problem of quadrotors with strongly nonlinear dynamics. First, for the nonlinear quadrotor dynamics, a deep neural network is trained to lift the physical states into a high-dimensional latent linear space. Then, an output matrix is introduced to map the lifted states to task-relevant outputs. Next, a discrete-time linear quadratic regulator (LQR) with an output-weighted quadratic cost is formulated to yield a Riccati-based controller for efficient regulation. In contrast to existing Koopman-based control schemes, the proposed approach improves the modeling fidelity for nonlinear flight dynamics. Finally, simulation demonstrates that the proposed controller outperforms the baseline controller and also provides a principled and computationally efficient alternative to purely reinforcement learning (RL)-based policies.

Keywords: Learning Systems, Modeling and Control of Complex Systems
Abstract: This work studies imperfect-information dynamic Defender-Attacker Blotto games under graph constraints. In this game, two players make multi-stage decisions on how to allocate and move limited resources to control critical nodes, with each player observing only part of the opponent’s resources at each stage, highlighting the challenges of imperfect information. Existing reinforcement learning methods suffer from low sample efficiency and lack theoretical guarantees, while classic counterfactual regret minimization (CFR) algorithms struggle with low computational efficiency in large action spaces. To address this, we propose a Progressive-Expanding Counterfactual Regret Minimization (PE-CFR) algorithm. This method uses hybrid strategy pruning to construct a high-quality initial action subset and applies progressive expansion mechanism to explore the full action space. Experiments on networks with 3 to 6 nodes show that PE-CFR significantly outperforms the baseline in both convergence speed and computational efficiency. The equilibrium generated reveals that resource allocation strategy should consider network topology and the opponent's mobility, rather than solely consider values of nodes.

Keywords: Automated Guided Vehicles
Abstract: ，交通标志检测在 智能交通系统（ITS）与高级 驾驶辅助系统（ADAS）通过提供关键 支持驾驶安全和事故预防。致 解决小靶遗漏、遮挡等问题， 以及复杂环境中的低探测精度，即 论文提出了一种新的交通标志检测算法 针对基于YOLOv8n框架的小型对象进行了优化。 首先，通过积分开发C2f-CGLU模块 通道门控线性单元（CGLU）进入瓶颈 主干中C2f模块的结构，因此 加强模型对局部细节的感知， 部分遮挡符号的通道更新功能。 此外，设计轻量级功能精炼融合 引入（FRF）模以重建 YOLOv8n 的交叉卷积颈部结构;该模块， 称为C2f-FRF，促进了更优越的特征细节融合 同时减少总参数数。 此外，原始的损失函数ඪ

Keywords: Control of Distributed Generation Systems, Energy Efficiency, Control of Smart Power Delivery Systems
Abstract: Distributed Energy Storage Systems (DESSs) are critical for grid stability, yet their effective scheduling relies on accurate knowledge of the global total capacity. However, the capacity data of individual storage units is commercially sensitive. Traditional distributed consensus algorithms allow for decentralized estimation but often expose the initial states of nodes to neighbors, leading to privacy leakage. To address this challenge, this paper proposes a privacy-preserving distributed estimation strategy based on textit{Dynamic Differential Privacy}. We design a time-varying noise injection mechanism where Laplace noise is added to the consensus process. The noise scale is initialized at a high level to mask raw data and follows a decay function, reducing to zero at a pre-set time step T. Theoretical analysis proves that this method ensures the privacy of individual units during the transient phase and achieves exact convergence to the true global capacity at step T, effectively balancing the trade-off between privacy protection and estimation accuracy.

Keywords: Control of Distributed Generation Systems, Nonlinear Systems and Control, Multi-agent Systems
Abstract: Frequency stability in islanded microgrids is often compromised by the low inertia of inverter-based interfaces and the prevalence of measurement noise in weak communication links. Traditional secondary control strategies, such as asymptotic or finite-time methods, typically suffer from slow convergence rates and are sensitive to noise, leading to steady-state fluctuations or prolonged recovery times. To address these challenges, this paper proposes a noise-resilient distributed secondary control strategy tailored for the emergency frequency recovery of weak microgrids. By employing a prescribed-time control framework with time-varying gains, the proposed method ensures that frequency synchronization is achieved exactly within a user-defined time window, regardless of initial system states. Theoretical analysis reveals that the inherent growing gain of the controller effectively suppresses bounded measurement noise as the deadline approaches, ensuring deterministic recovery precision. Numerical simulations on a modified IEEE 33-bus system validate that the proposed strategy can restore frequency to the nominal value within 5 seconds, exhibiting superior robustness compared to conventional fixed-gain approaches.

Keywords: Energy Efficiency, Multi-agent Systems
Abstract: With the increasing penetration of electric vehicles (EVs), vehicle-to-grid (V2G) technology enables them to act as flexible resources within the power grid. Electric vehicle aggregators (EVAs) play a crucial role in coordinating large-scale EV charging and discharging, yet they face the dual challenge of maximizing economic benefits of EVAs while maintaining user satisfaction. To address this issue, this paper develops a solution method based on Stackelberg game theory. As leaders, EVAs determine charging prices and V2G compensation to optimize profits and achieve peak shaving and valley filling, while users, as followers, respond to regulatory signals and adjust their charging behaviors accordingly. An embedded cross multi-agent actor–critic (EC-MAAC) algorithm is further proposed to address the scalability issues arising from large-scale EV participation in V2G systems, while alleviating the reliance of conventional game-theoretic solvers on convexity and differentiability assumptions for tractable equilibrium computation. Simulation results demonstrate that the proposed EC-MAAC approach achieves a 15.1% reduction in EVA operating costs, a 44.6% increase in user rewards, and an improvement in SoC satisfaction from 91.2% to 96.8%, effectively balancing economic efficiency and user satisfaction.

Keywords: Multi-agent Systems, Networked Control, Adaptive Control
Abstract: This paper investigates the distributed optimal coordination problem for multiple uncertain Euler-Lagrange systems. We first propose a novel measurement-based distributed optimal coordinator to generate the local reference trajectory without relying on communication networks. Notably, the coordinator gains are designed as two time-varying decreasing functions with distinct decay rates, which avoids the dependence on global information and, meanwhile, mitigates the error terms stemming from the replacement of communicated variables with physical outputs. Then, a norm-estimation-based adaptive controller is designed to achieve accurate reference tracking, where only one updated parameter needs to be introduced, regardless of the dimensions of disturbances and uncertainties. By constructing an appropriate composite Lyapunov function and utilizing the property of a perturbed time-varying differential inequality, it is proved that agents’ outputs can asymptotically converge to the global optimum. The proposed scheme is fully distributed and inherently immune to cyber-attacks.

Keywords: Multi-agent Systems, Networked Control, Optimal Control
Abstract: This paper considers differentially private distributed Nash equilibrium seeking for aggregative games under an event-triggered mechanism. By incorporating an event-triggered scheme with a privacy protected aggregate estimation mechanism, a distributed Nash equilibrium seeking strategy is proposed. In the proposed strategy, the privacy protection is achieved by using Laplace noises to mask the information exchanged among the players, in which the transmission instants are determined by an event-triggered mechanism. By gradually weakening inter-player interactions, the proposed strategy ensures that players' actions can be driven exactly to the Nash equilibrium while ensuring rigorous epsilon-differential privacy with reduced communication costs. The effectiveness of the proposed strategy is verified using a numerical example.

Keywords: Robust and H infinity Control
Abstract: This work explores the observer-based controller problem for nonlinear networked dynamics where quantization non-idealities, exogenous disturbances, and asynchronous premise constraints are explicitly accounted for. A novel switching mechanism, leveraging the bounds of the MFs within an interval Type-2 Takagi–Sugeno (T–S) fuzzy framework, is proposed to control the approximated nonlinear model effectively. The proposed strategy obviates the need for precise MF information, accommodates premise asynchrony, and maintains low conservatism. The errors induced by quantization are formulated as bounded disturbances, which are suppressed under a prescribed H∞ performance index. Within the Lyapunov stability framework, conditions are derived to ensure asymptotic stability for the integrated system. The efficacy of the proposed approach, along with its advantages over existing methods, is subsequently confirmed via illustrative simulations.

Keywords: Intelligent and AI Based Control, Estimation and Identification, Sensor/Data Fusion
Abstract: Adapting frozen vision-language models (VLMs) to downstream tasks via prompt tuning has emerged as a parameter-efficient alternative to full fine-tuning. However, existing prompt tuning methods still lack explicit control over how strongly prompts should adapt to different samples, despite the significant variation in sample difficulty and prediction reliability---a critical limitation for trustworthy deployment in complex real-world scenarios such as autonomous perception and intelligent control systems. We propose UGPT (Uncertainty-Guided Dynamic Prompt Tuning), a framework that explicitly leverages visual prediction uncertainty as a sample-level control signal for prompt modulation. UGPT computes normalized entropy from a frozen zero-shot CLIP branch, maps it through a lightweight Uncertainty Injection Network (UIN) into prompt-space perturbation vectors, and applies broadcast injection to generate sample-adaptive dynamic prompts. Under the 16-shot few-shot setting, UGPT uses only 41K trainable parameters in total, achieves 81.58% Top-1 accuracy on ImageNet-1K, and obtains consistent improvements across four out-of-distribution benchmarks (Avg. OOD: 62.0%). Ablation studies and interpretability analyses confirm that the perturbation magnitude increases monotonically with uncertainty, validating the principled "harder samples receive stronger modulation" mechanism.

Keywords: Robotics, Man-machine Interactions, Control Applications
Abstract: Sparse-reward goal-conditioned reinforcement learning remains difficult in dexterous robotic control, where informative successful interactions are rare during early exploration. Although Hindsight Experience Replay (HER) improves sample efficiency by relabeling failed trajectories with achieved goals, standard HER usually adopts fixed hindsight-goal sampling heuristics and does not explicitly account for the agent's current competence. Meanwhile, existing curriculum strategies often rely on a single difficulty dimension, which may be insufficient for high-dimensional dexterous tasks. In this paper, we propose Competence-Aware Multi-Dimensional Curriculum HER (CAMD-HER), an extension of HER for sparse-reward goal-conditioned learning. CAMD-HER introduces a competence-aware hindsight relabeling mechanism that prioritizes candidate goals according to competence matching, novelty, and learning progress, and a multi-dimensional curriculum strategy that progressively adjusts task difficulty through success threshold, goal sampling range, reset perturbation, and observation noise. The proposed method is integrated into an off-policy HER-based training framework and implemented in a Stable-Baselines3-compatible manner. Experiments on the HandReach benchmark show that CAMD-HER improves training effectiveness over standard DDPG+HER by accelerating learning and achieving better performance under sparse rewards. The results indicate that jointly adapting replay selection and environment difficulty is an effective way to improve sample efficiency in goal-conditioned dexterous reinforcement learning.

Keywords: Man-machine Interactions
Abstract: 基于神经网络激增的人类动作识别 （SNN）因其 低功率优势。然而，现有基于SNN的方法 受限于训练困难，这些困难由 尖峰的不可微性，阻止了 准确性远超人工神经网络 （人工焦虑论）。为解决这一瓶颈，本研究提出了一个 集成两者结合的混合人工神经网络-卫星网络（混合AS） 两流架构中的网络范式。澳大利亚国立网络 布兰奇确保高 识别精度，而SNN分支则保持能量 效率。为了实现有效的跨分支 互补性，我们引入显著性评分 计算和多粒度采样机制 深度功能整合。HMDB-51上的广泛实验， UCF-101和Kinetics-400证明了混合AS 在保持 SNN的低功耗特性，提升效率 动作识别。

Keywords: Man-machine Interactions, Adaptive Control
Abstract: Spiking Neural Networks (SNNs) represent a class of biologically-inspired neural models that employ discrete spike-based communication, contrasting with the continuous-valued activations of traditional Artificial Neural Networks (ANNs). This event-driven paradigm offers significant advantages in power efficiency and temporal processing capabilities, positioning SNNs as promising alternatives for energy-constrained applications.However, two critical challenges impede their widespread adoption: the persistent performance gap compared to ANNs and the substantial computational overhead associated with training. The non-differentiable nature of spike generation prevents direct application of standard backpropagation, while existing training methodologies present fundamental limitations. ANN- to-SNN conversion techniques typically require extended temporal sequences and sacrifice inherent spatiotemporal dynamics,whereas surrogate gradient methods demand complete temporal unrolling, resulting in computational burdens comparable to conventional ANNs. To address these challenges, this work introduces the Integer Leaky Integrate-and-Fire (I-LIF) neuron model, which substantially reduces training complexity while maintaining adaptive capability.Additionally, we develop two specialized SNN architectural components: a convolution-based processing block and a Transformer-based attention block,specifically designed to overcome performance degradation in object detection applications. Extensive experimental validation on COCO,Gen1 and PASCAL VOC benchmarks demonstrates substantial performance improvements, confirming the potential of SNNs for complex visual perception tasks and advancing their practical deployment in real-world systems.

Keywords: Intelligent and AI Based Control, Fault Detection and Diagnostics, Control Applications
Abstract: Real-time defect detection in industrial and infrastructural environments often faces a dilemma: dense convolutions are computationally expensive, while lightweight operators provide limited inductive bias for high-frequency, anisotropic patterns such as cracks. This paper proposes MaskedStar, a masked depthwise convolution that restricts learnable kernel locations to two complementary star supports: a horizontal–vertical cross and a diagonal “X”. Building on MaskedStar, two plug-and-play backbone blocks: MaskedStar-S, which fuses dual MaskedStar branches with different dilations via lightweight ChannelSE attention, and MaskedStar-D, which additionally blends an optional full-context branch for mid-level pyramid features, are designed. Replacing the P3/P4 C3k2 blocks in a YOLO11n baseline yields consistent improvements on three benchmarks: NEU-DET (75.8/44.1 → 76.9/44.4), RDD-MotorBike (92.8/59.6 → 93.4/59.8), and RDD-Drone (59.7/33.6 → 62.0/35.8) for mAP50/mAP50:95. Overall, the observed improvements suggest that explicit orientation priors are a practical option for lightweight detectors targeting elongated patterns. Index Terms— Object detection, orientation-aware convolution, surface defect detection, road damage detection

Keywords: Robotics, Learning-based Control, Intelligent and AI Based Control
Abstract: Training robust vision-based robotic manipulation policies requires large-scale data, yet real-world collection is costly and unsafe. We present a high-fidelity simulation framework for data generation and policy learning, built on Isaac Lab with a closed-loop teleoperation pipeline for a UR5 manipulator and Robotiq 2F-85 gripper. The framework unifies demonstration collection, domain randomization, and policy training in a single pipeline. A central component is a semantic-aware parameter identifier that helps overcome a limitation of conventional DR by preventing unconstrained sampling from decoupling physical parameters from visual appearance, which otherwise yields semantically inconsistent scenarios that introduce spurious visual–physical correlations. To resolve this, we leverage a Vision-Language Model to infer material categories from RGB observations and assign physically plausible nominal values with confidence-aware bounds for friction and density. We validate the framework by training a Vision-Language-Action policy on the generated data. The results show that the parameter identifier attains high hit rates on material property prediction, and that the proposed DR meaningfully improves average task success over a position-only baseline.

Keywords: Learning-based Control, Robotics, Intelligent and AI Based Control
Abstract: Recent advances in vision--language--action (VLA) models enable end-to-end robotic manipulation by learning from demonstrations. However, existing approaches have predominantly focused on scaling model architectures or expanding dataset size, often overlooking the interaction between training data distribution and inference strategies. To address this gap, we jointly study data distribution and inference strategy for robotic grasping, emphasizing coordinated distribution design and adaptive inference. Our approach integrates a compositional data collection strategy that decomposes the task state space into orientation and spatial factors, efficiently expanding coverage with a limited number of demonstrations. Additionally, we propose an adaptive inference mechanism that dynamically adjusts the execution horizon during critical task phases, thereby enhancing task performance. Embedding-space analysis suggests that performance saturates once the relevant state-space dimensions are sufficiently covered. Real-robot experiments validate our approach, demonstrating a 98% grasp success rate with only 81 demonstrations. Furthermore, applying adaptive inference to a moderately covered dataset improves the success rate from 90% to 97%, suggesting that inference-level refinement can complement data coverage, though it cannot fully compensate for insufficient distribution design.

Keywords: Automated Guided Vehicles, Control Applications, Adaptive Control
Abstract: In complex environments such as industrial logistics, mobile robots relying on global planning are susceptible to communication latency, computational fluctuations, and non-convex environmental constraints. These factors often render reference paths stale or cause global guidance failures, resulting in stagnation or collisions at the control level. To address these navigation challenges, this paper proposes a candidate-set driven local resilient decision framework. By constructing a hybrid candidate set (model-based and policy-based) and incorporating receding horizon online evaluation, the proposed method reduces dependency on global paths while achieving an adaptive trade-off between local navigation efficiency and safety margins. Furthermore, an embedded fail-safe mechanism employs a conservative baseline strategy to guarantee robust operation within tightly constrained environments. Experimental results demonstrate that the proposed approach effectively avoids local deadlocks, reduces collision risks, and enhances navigation efficiency under conditions where global reference paths are significantly delayed or unreliable, while ensuring real-time performance.

Keywords: Automated Guided Vehicles, Control Applications, Robotics
Abstract: This paper considers the problem of straight-line tracking for unicycle mobile robots when the reference angular velocity is trivial and the persistence of excitation condition is a lack. To overcome this limitation, we present a controller to ensure global exponential tracking with input constraints. Simulation results confirm that the method achieves fast convergence.

Keywords: Automated Guided Vehicles, Energy Efficiency, Control Applications
Abstract: Four-wheel-independent-drive electric vehicles (FWID-EVs) provide high control flexibility and actuator redundancy, making them promising for improving both yaw stability and energy efficiency. However, existing studies mainly focus on stability enhancement, while the energy cost associated with direct yaw moment control is often neglected. In addition, the control emphasis between stability and efficiency should vary with the instantaneous vehicle stability state. To this end, this paper proposes a stability-aware yaw control framework for FWID-EVs. At the upper layer, a yaw stability controller is designed using prescribed performance control (PPC) and SMC to regulate the vehicle's yaw response and ensure lateral stability. At the lower layer, a phase-plane-based stability index is constructed, and a driving torque allocation strategy that considers tire utilization and motor efficiency is proposed. Simulation results demonstrate that the proposed approach can effectively maintain yaw stability while reducing energy consumption, thereby achieving a better balance between safety and energy efficiency for FWID-EVs.

Keywords: Automated Guided Vehicles, Motion Control
Abstract: In this paper, we propose a motion coordination strategy for multi-AGV systems operating on predefined closed paths, in which path preprocessing and dangerous circle detection serve as the two key mechanisms. Through path preprocessing, we explicitly analyze the network topology to identify conflict-prone segments and encode potential interaction relationships among AGVs, thereby transforming implicit collision and deadlock risks into a structured representation available for online decision-making. Building upon this representation, we employ a dangerous circle detection method to identify cyclic blocking configurations that are the fundamental cause of deadlocks under bidirectional motion. This enables us to shift deadlock avoidance from reactive resolution to proactive prevention by evaluating one-step-ahead motion states and suppressing the formation of dangerous circles via selective stopping and resuming actions, without resorting to global path replanning.

Keywords: Automated Guided Vehicles, Optimal Control, Motion Control
Abstract: Conventional Lane Centering Control (LCC) algorithms typically rely on the passive geometric tracking of the lane centerline during curve navigation. Consequently, they fail to replicate the active pose-adjustment mechanisms exhibited by human drivers throughout the entry, mid-curve, and exit phases, thereby degrading the human-like quality and safety margins of the system. To address these limitations, this study proposes a human-like trajectory generation approach tailored for full-process curve driving. First, an explicit parameter-driven trajectory generation model is established. By introducing five critical physical parameters, including a pre-positioning lateral offset and a time factor, this model quantitatively translates implicit human driving expertise into a controllable "pre-positioning mechanism." Furthermore, a hierarchical trajectory generation method is developed. The initial stage constructs a baseline trajectory incorporating pre-positioning features using the explicit parameters. Subsequently, a Quadratic Programming (QP) model integrating a jerk penalty is formulated to eliminate curvature discontinuities while faithfully preserving the pre-positioning geometric anchors, ensuring high-order dynamic smoothness. Finally, comprehensive closed-loop validation is conducted utilizing a Carla-Speedgoat Hardware-in-the-Loop (HIL) platform integrated with a curvature-feedforward Stanley controller. Experimental results demonstrate that the proposed approach, while fully compatible with the Conventional Centerline Tracking (CCT) approach, consistently generates dynamically executable trajectories with prominent human-like characteristics.

Keywords: Automated Guided Vehicles, Sensor/Data Fusion, Intelligent and AI Based Control
Abstract: Autonomous driving systems require reliable 3D perception for tasks such as detection and segmentation. In recent years, multimodal integration of LiDAR’s spatial information with camera’s contextual insights has achieved significant progress in this domain using supervised learning techniques. However, supervised models depend on extensive labeled data, which limits their widespread deployment. Selfsupervised learning (SSL) addresses this challenge by pretraining on unlabeled data to develop transferable features, yet prior SSL fusion techniques struggle with bridging modality differences, achieving precise alignment, and extracting meaningful abstractions. Inspired by the Joint Embedding Predictive Architecture, we propose a two-stage SSL pretraining method that fuses LiDAR and camera signals, improving semantic coherence and facilitating efficient adaptation to downstream tasks such as 3D object detection. The evaluations demonstrate significant performance gains in data-scarce scenarios, outperforming traditional SSL methods and competing supervised methods.

Keywords: Multi-agent Systems, Learning-based Control, Sensor Networks
Abstract: This paper investigates the distributed coverage control problem for multi-agent systems operating in complex environments characterized by unknown sensory density functions. To address this challenge, we propose a novel cooperative control framework that synergizes Graph Neural Networks (GNN) and Multi-Layer Perceptrons (MLP). Specifically, GNNs are employed to aggregate dynamic neighbor information, while MLPs approximate the local sensory density. This architecture enables agents to adaptively optimize their deployment based on estimated fields without requiring global prior knowledge. Lyapunov stability theory is employed to guarantee the convergence of the agents' positions, regarding the estimation errors of the density function. Numerical simulation shows that the proposed method outperforms the traditional Lloyd's algorithm by achieving performance comparable to benchmarks with access to the ground-truth density function. Furthermore, ablation studies validate the efficacy of the GNN component in enhancing both coverage performance and estimation accuracy.

Keywords: Robotics, Modeling and Control of Complex Systems, Nonlinear Systems and Control
Abstract: To overcome the limitations of existing moxibustion robots, such as reliance on predefined paths, poor adaptability to body surface, and lack of temperature regulation, this paper proposes a revolving moxibustion path planning method based on improved Ant Colony Optimization (ACO). First, a revolving moxibustion path planning model is established by integrating the body surface normals and heat transfer model, which enhances surface adaptability and enables temperature prediction during treatment. Second, the ACO algorithm is improved by designing a path transition probability function that incorporates directional guidance and turning angle constraints, along with an adaptive pheromone enhancement strategy based on the Sigmoid function. These enhancements accelerate the convergence of the algorithm while balancing global exploration and local exploitation. Experimental results illustrate that the proposed method performs well in surface adaptability, moxibustion efficacy, and operational efficiency, providing a valuable reference for further research and applications of moxibustion robots.

Keywords: Sensor/Data Fusion, Estimation and Identification, Robotics
Abstract: Ultra-wideband (UWB) ranging is a widely used solution for indoor positioning, but its accuracy degrades under non-line-of-sight (NLOS) conditions where range errors become biased, non-Gaussian, and heavy-tailed. In loosely coupled UWB--IMU fusion, these corrupted ranges first affect the nonlinear least-squares (LS) trilateration step, and the resulting erroneous position pseudo-measurements can destabilize a downstream extended Kalman filter (EKF). This paper proposes a robust trilateration front-end based on the Huber loss, solved via iteratively reweighted least squares (IRLS), and integrates it into a 15-state error-state inertial navigation EKF. The Huber-IRLS formulation bounds the influence of large residuals, improving robustness to NLOS-contaminated anchors without explicit NLOS detection or anchor rejection. Experiments in a controlled 3D indoor simulation with geometry-based NLOS modeling demonstrate that the robust method preserves nominal LOS accuracy while significantly improving NLOS performance. In the baseline NLOS scenario, the EKF position RMSE is reduced from 0.205 m (LS) to 0.131 m (Huber-IRLS), with the 95th-percentile error reduced from 0.376 m to 0.250 m. Moreover, the proposed Huber-IRLS solver achieves substantially lower runtime than lsqnonlin-based LS, reducing the total per-update computation from approximately 2.23 ms to 0.20 ms, supporting real-time deployment.

Keywords: Robotics
Abstract: This paper introduces Bidirectional Tight Informed Trees (BTIT*), an asymptotically optimal kinodynamic sampling-based motion planning algorithm that integrates an anytime bidirectional heuristic search (Bi-HS) and ensures the meet-in-the-middle property (MMP) and optimality (MM-optimality). BTIT* is the first anytime MEET-style algorithm to utilize termination conditions that are efficient to evaluate and enable early termination on-the-fly in batch-wise sampling-based motion planning. Experiments show that BTIT* achieves strongly faster time-to-first-solution and improved convergence than representative non-lazy informed batch planners on two kinodynamic benchmarks: a 4D double-integrator model and a 10D linearized Quadrotor. The source code is available href{https://github.com/yi213-robotic/Bidirectional-Tight-Informed-Trees}{textcolor{blue}{here}}.

Keywords: Learning-based Control, Optimal Control
Abstract: Model Predictive Control (MPC) is an established technique for controlling constrained dynamical systems, but its performance is highly sensitive to the tuning of its cost function weighting matrices and prediction horizon. Manual tuning is often time-consuming and yields sub-optimal performance for systems operating under varying conditions. This paper presents a novel framework that utilizes Reinforcement Learning (RL) for the online, simultaneous adaptation of these key parameters. By interacting with the closed-loop system, the RL agent learns a dynamic tuning policy aimed at maximizing a cumulative reward associated with tracking performance and control efficiency. To establish baseline safety, the framework first constructs a pre-certified stability-guaranteed set of MPC parameters offline. However, a critical challenge remains during online operation: dynamic parameter updates render the closed-loop system a switched system, where arbitrary switching can induce instability. To address this, we introduce a supervisory stability filter based on a Common Lyapunov Function (CLF). This mechanism rigorously enforces a monotonic decay in a fixed energy metric, functioning as a real-time safety gate that permits only stability-preserving updates from the RL agent. Finally, the framework's effectiveness is demonstrated through a comprehensive numerical evaluation comparing Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Trust Region Policy Optimization (TRPO), highlighting the superior tracking accuracy and stability of the proposed approach.

Keywords: Learning-based Control, Optimal Control, Multi-agent Systems
Abstract: This paper develops a distributed learning-based model predictive control (DMPC) framework with an explicit quadratic programming (QP) formulation for platooning control of heterogeneous autonomous surface vehicles (ASVs) under disturbances and modeling uncertainties. A composite control structure is adopted, in which the control signal consists of a nominal control term, a neuro-adaptive term, and a feedforward term. The nominal control term is obtained by solving a tube-based robust MPC problem, while a neuro-adaptive term estimates matched uncertainties online. The feedforward term is included to form a linear prediction model which facilitates the QP formulation. By integrating neural network (NN) based disturbance estimation into the tube-based MPC design, the proposed method preserves recursive feasibility, constraint satisfaction, and closed-loop stability. Simulation results demonstrate excellent control performance and improved disturbance rejection capability.

Keywords: Optimal Control, Automated Guided Vehicles, Robotics
Abstract: The data-driven model predictive control (MPC) approach is investigated for path following control applications of wheeled mobile robots (WMRs). By updating Hankel matrices continuously using input-output data, the robot motion dynamics can be encoded into an implicit linearized model, which facilitates the MPC controller design. A novel two-step path following control algorithm is proposed so that part of the control problem can be formulated into a standard form quadratic program (QP), which can be solved efficiently by off-the-shelf optimization software. The detailed path following control algorithm design is discussed for a linear equivalent model and the nonlinear unicycle model of WMRs. The linearization errors are estimated throughout the control sequence and explicitly included in a robust MPC problem formulation. Simulation results demonstrate excellent path following performance with the proposed data-driven method.

Keywords: Nonlinear Systems and Control, Robotics, Estimation and Identification
Abstract: This paper proposes a neural-network-based integral sliding mode control (NN-ISMC) method to address abrupt variations during the midair reconfiguration of modular reconfigurable flight arrays (MRFAs). First, an integral sliding function is constructed to eliminate the reaching phase and enhance robustness at the moment of separation. Then, an online neural network is employed to estimate the equivalent disturbance from the sliding variable and system states for feedforward compensation. A Lyapunov-based adaptive law is derived to guarantee convergence of all closed-loop signals. Finally, comparative simulations are provided for several representative scenarios to demonstrate effectiveness of the proposed method.

Keywords: Multi-agent Systems, Networked Control
Abstract: In this paper, we study attitude synchronization on SO(3) under switching communication topologies within a hybrid framework. A hybrid model is constructed to capture continuous attitude dynamics and discrete topology switching, and a distributed controller based on relative attitude information is adopted. Due to switching, the Lyapunov function may be discontinuous at switching instants. To address this issue, a trajectory-based bounding technique is developed. Under a minimum dwell-time condition, semiglobal exponential synchronization is established. Numerical simulations validate the theoretical results.

Keywords: Learning-based Control, Modeling and Control of Complex Systems, Nonlinear Systems and Control
Abstract: Fixed-wing Unmanned Aerial Vehicles (UAVs) are essential for high-dynamic aerial missions, yet fully exploiting their mechanical capabilities remains a challenge. Unlocking their potential for aggressive maneuvers via Deep Reinforcement Learning (DRL) is hindered by control jitter and safety concerns. In this paper, a Risk-Aware Smooth Reinforcement Learning framework (RAS-RL) is proposed to achieve aggressive flight without compromising safety or stability. Specifically, a learnable spectral filter is designed to denoise the observation stream, and Jacobian regularization is incorporated to suppress actuator jitter in closed-loop execution. Furthermore, a model-predictive risk-aware reward shaping term is constructed by combining a nominal dynamics model with a supervised residual dynamics approximator. Lastly, a curriculum learning schedule is employed that progressively increases command aggressiveness, observation noise, and the risk weight to stabilize training under the shaped objective. In high-fidelity JSBSim simulations, RAS-RL achieves a 96% success rate on randomized command tests with 0.8% safety violations, while maintaining low action fluctuation comparable to smoothness-oriented baselines.

Keywords: Control Applications, Optimal Control, Robust and H infinity Control
Abstract: This paper addresses the robust orbit-attitude control problem for satellites subject to directional jamming. Unlike conventional approaches that model interference as additive Gaussian noise, we formulate the anti-jamming problem as a zero-sum differential game based on the coupled nonlinear orbit-attitude dynamics, thereby transforming the problem from passive disturbance rejection to active geometry-aware anti-jamming control. A novel interference metric is introduced to characterize the intensity of directional jamming, which is explicitly determined by the beam-to-beam geometry between the target and the adversary. To solve the resulting Hamilton-Jacobi-Isaacs equation without requiring precise knowledge of the system dynamics, an off-policy integral reinforcement learning algorithm is developed. The proposed data-driven scheme iteratively approximates the Nash equilibrium solution using neural networks. Rigorous theoretical analysis guarantees the Uniformly Ultimate Boundedness of the weight approximation errors and the finite L_2-gain stability of the closed-loop system, thereby ensuring H_infty robustness. Numerical simulations validate the efficacy of the proposed method.

Keywords: Robotics, Learning Systems
Abstract: Autonomous quadrotor target tracking in cluttered environments remains a critical challenge, requiring persistent tracking of dynamic targets using exclusively onboard perception. Existing approaches frequently struggle with agile maneuvers, sim-to-real gaps, and prolonged target occlusions. To address these limitations, we propose a robust reinforcement learning framework. For high-bandwidth maneuverability and a minimized sim-to-real gap, our policy directly outputs collective thrust and body rates, bypassing the delays inherent in standard position controllers. Furthermore, we design a specialized observation space encoding target kinematics and introduce a novel geometry-agnostic visibility reward to proactively anticipate occlusions and maintain line-of-sight (LOS). To facilitate stable convergence and overcome early local optima, we implement a progressive curriculum learning strategy that incrementally scales obstacle density. Comprehensive simulations demonstrate that our framework maintains robust tracking capabilities and achieves exceptionally high success rates, even under severe occlusion conditions.

Keywords: Learning-based Control, Motion Control
Abstract: LiDAR-based and vision-based end-to-end path planning algorithms suffer from lighting variations and texture deficiency, leading to sudden trajectory planning failure and subsequent collision accidents. To address this problem, this paper proposes LV-Fusion Planner, a LiDAR-vision fusion end-to-end path planning algorithm based on the YOPO algorithm. The planner fuses LiDAR, vision and UAV self-state multimodal data. It combines dynamic motion primitive-based trajectory parameterization with end-to-end mapping from multimodal data to trajectory. It enables low-latency and fully autonomous navigation, and overcomes single-modal perception limitations in path planning.

Keywords: Multi-agent Systems, Control Applications, Robotics
Abstract: In this paper, we present a distributed framework for coordinated motion among multiple robotic swarms. The proposed approach is built upon the Smoothed Particle Hydrodynamics (SPH) paradigm, in which each robot is modeled as a particle and computes its control input based on the relative positions of neighboring robots within a finite communication range. Inspired by the immiscibility phenomenon observed in water–oil mixtures, we incorporate a differential interaction mechanism into the SPH framework to encode heterogeneous inter-group behaviors. As a result, in scenarios involving multiple robot groups with distinct affiliations, the swarm can spontaneously evolve from random initial configurations into multiple well-segregated and uniformly organized formations. Simulation results validate the effectiveness of the proposed method.

Keywords: Multi-agent Systems, Robotics, Modeling and Control of Complex Systems
Abstract: Multi-robot herding is of great importance for autonomous agricultural and livestock management systems. However, traditional methods often struggle to converge when dealing with evader swarms that are faster and have a strong tendency to disperse. This paper proposes a decentralized cooperative herding algorithm that combines a greedy strategy with dynamic convex hull envelopment. The algorithm uses a real-time dynamic convex hull to regulate the robot formation, while employing a greedy strategy that allows each robot to autonomously select the nearest target within its field of view, complemented by a rotational search mechanism to ensure target visibility. Based on a second-order dynamics model, a coverage angle constraint considering the speed ratio and a force-balancing control law are designed to ensure the dynamic stability of the encirclement. In a scenario with 6 herders managing 30 high-speed evaders, the proposed algorithm significantly reduces the average completion time and improves stability compared to a non-cooperative greedy strategy. The algorithm demonstrates strong robustness against obstacles, scalability to larger groups, and varying initial conditions, providing a feasible solution for practical ranch applications.

Keywords: Multi-agent Systems, Optimal Control, Intelligent and AI Based Control
Abstract: Multi-UAV systems have demonstrated significant potential in urban low-altitude transportation, particularly in disaster response and emergency supply delivery, owing to their high maneuverability and contactless delivery capabilities. However, under high-density arrival traffic and limited vertiport resources, conventional landing scheduling methods struggle to balance delivery timeliness for supplies with varying urgency against UAV battery safety. To address these challenges, this paper formulates a multi-UAV landing scheduling optimization problem for emergency logistics that jointly considers delivery efficiency with priority differentiation and low-battery risk, where the decision variables exhibit strong coupling. To solve this problem, we propose a priority-aware multi-UAV landing scheduling framework. The landing sequence is first determined by a genetic algorithm, followed by a Sequence-Based Approach with Yield strategy (SBAY) for approach control. The SBAY model introduces outer yield holding points to create expedited descent corridors for earlier-ranked UAVs in the landing sequence, thereby mitigating the conflict between strongly coupled decision variables and reducing low-battery risks. Simulation results demonstrate that the proposed method improves delivery efficiency for high-priority emergency supplies while ensuring battery safety, and effectively enhances overall landing scheduling performance.

Keywords: Robotics, Estimation and Identification, Learning Systems
Abstract: This paper addresses semantic boundary blurring and the conflict between geometric and semantic optimization in existing semantic visual SLAM systems based on Gaussian Splatting. Current semantic 3DGS methods typically employ simple feature regression strategies, which lead to averaged semantic features at object boundaries and often destabilize geometric tracking due to forced semantic constraints. To this end, we propose a novel 3DGS-SLAM system. First, we introduce Semantic Center Loss and Entropy Minimization Loss to replace conventional regression losses. Drawing upon metric learning principles, these losses consolidate intra-class features in the latent space and sharpen semantic boundaries. Second, we design a gradient decoupling mechanism that separates geometric updates from semantic optimization, preventing semantic gradients from interfering with camera pose estimation. Experiments on the Replica dataset demonstrate that our method achieves significantly improved semantic segmentation accuracy (mIoU) while maintaining high-precision trajectory estimation and superior rendering quality, outperforming current state-of-the-art approaches.

Keywords: Robotics, Automated Guided Vehicles, Motion Control
Abstract: Gas source localization is critical for industrial safety monitoring, disaster rescue, and environmental protection. This paper presents an adaptive source-seeking planner that explicitly balances exploration and exploitation under uncertainty. At each step, the robot samples candidate goal points and scores them by combining (i) an exploitation term derived from a Gaussian-like dispersion model and (ii) an exploration term computed as frontier-based information gain. To avoid search stagnation, a time-dependent penalization is introduced to reduce the attractiveness of early sampled goals, discouraging long-distance revisits. Moreover, the exploration--exploitation weight is adapted online using the spatial variance of a high-probability candidate set: dispersed candidates trigger stronger exploration, while concentrated candidates promote rapid exploitation toward the source. Simulation and real-robot experiments demonstrate that the proposed algorithm improves search efficiency in complex environments.

Keywords: Networked Control, Nonlinear Systems and Control, Multi-agent Systems
Abstract: This paper investigates leaderless bearing-constrained formation control for higher-order integrator systems utilizing bearing-only measurements. Unlike existing results, which are typically limited to first- or second-order dynamics, this work addresses systems of arbitrary order. We first establish a baseline for bearing-constrained formation control using state feedback and bearing measurements. By designing a bearing-feedback sliding surface and employing the backstepping technique, a distributed control law is synthesized to guarantee global convergence of the bearing constraint errors to zero. Subsequently, by leveraging the separation principle and the certainty equivalence principle, a distributed output-feedback control law is synthesized to achieve the target formation. Finally, the performance of the developed control schemes is demonstrated through numerical simulations.

Keywords: Sensor/Data Fusion, Signal Processing
Abstract: Multi-object tracking (MOT) in Unmanned Aerial Vehicle (UAV) swarm scenarios presents unique challenges due to dense formations, appearance similarity, frequent occlusion, and limited resolution. We propose a lightweight tracking framework based on Observation-Centric SORT (OC-SORT), specifically designed for appearance-free aerial tracking. We further introduce a Velocity Direction Consistency (VDC) module, which leverages temporally distant motion cues to construct a stable and discriminative descriptor. This helps suppress angular inconsistencies, improves association in dense formations, and significantly reduces ID switches when UAV trajectories overlap or targets undergo occlusion. Additionally, a trajectory memory module is employed to store recent observations and recover identities lost due to short-term occlusion without relying on visual features. Experimental results on the UAVSwarm Dataset demonstrate that our approach improves ID consistency and overall tracking accuracy compared to existing methods, while maintaining real-time performance.

Keywords: Signal Processing, Intelligent and AI Based Control, Learning Systems
Abstract: Digital Twin (DT)-driven smart inspection and diagnosis in marine environments rely on reliable underwater visual perception for state estimation, anomaly analysis, and virtual-physical consistency. However, underwater images are severely degraded by wavelength-dependent absorption and scattering, which hinders accurate inspection of underwater assets, ecological targets, and robotic operations. To address this issue, this paper proposes ACSC-UIR, an unsupervised underwater image restoration framework with adaptive color stretch and compensation (ACSC). The ACSC module supplements missing red-channel information while suppressing overcompensation in severely degraded scenes, and a multi-branch fully convolutional network is further designed to learn multi-scale restoration features without paired supervision. Experiments on UIEB and LSUI show that ACSC-UIR achieves state-of-the-art unsupervised performance in peak signal-to-noise ratio and structural similarity, while maintaining a favorable trade-off between restoration quality and model complexity. These results indicate that the proposed method can serve as an effective visual front-end for DT-driven smart inspection in underwater scenarios.

Keywords: Multi-agent Systems, Automated Guided Vehicles, Robotics
Abstract: Searching for small targets in large-scale natural environments is challenging due to limited flight resources and the sparse spatial distribution of targets. Traditional coverage strategies often rely on uniform sweeping patterns, which can lead to redundant observations and inefficient exploration. In this paper, we propose an environment-aware multi-UAV search framework for efficient large-scale small target detection. The proposed approach leverages environmental sensing and prior information to construct a probabilistic spatial belief map that estimates the likelihood of target presence. Instead of exhaustively covering the entire search area, high-probability regions are extracted and clustered into representative anchor points, significantly reducing planning complexity. A cooperative multi-UAV searching strategy is then developed by solving a rewarddriven team orienteering problem under flight distance and safety constraints. Extensive simulations demonstrate that the proposed method improves search efficiency and information collection compared with conventional coverage strategies. Field experiments with multiple UAVs further validate the feasibility and effectiveness of the framework in real-world environments.

Keywords: Smart Buildings
Abstract: The preservation of traditional Minnan-style architecture, characterized by complex topologies such as upturned eaves and elaborate facades, poses significant challenges for manual structural inspections. To overcome the labor-intensive and subjective nature of traditional methods, this paper proposes an autonomous Unmanned Aerial Vehicle (UAV) inspection framework tailored for safe and highly efficient defect detection in high-traffic campus environments. The proposed system integrates a hybrid path planning strategy, combining Boustrophedon and A-star algorithms, to ensure collision-free navigation and high-fidelity image acquisition across both expansive exteriors and spatially constrained interiors. Field deployments at Xiamen University validate the superiority of this framework, which cuts flight duration by more than half while increasing surface coverage to over 96 percent compared to manual piloting. Furthermore, we developed XMU AI-DETECT, a closed-loop visual decision-support platform coupled with a comprehensive deep learning-based object detection framework for robust anomaly identification against intricate brickwork backgrounds. This framework effectively transforms raw aerial data into actionable maintenance strategies, establishing a highly scalable paradigm for the preventative conservation of complex cultural heritage sites.

Keywords: Smart Buildings, Energy Efficiency
Abstract: Simulation-based building thermal and indoor air quality analysis requires coupling building energy simulation (BES) with computational fluid dynamics (CFD), yet the associated multi-software configuration workflow remains fragmented, expertise-demanding, and error-prone. This paper presents a multi-agent co-simulation framework that transforms natural-language building-analysis requests into controlled BES-CFD workflows within a digital twin environment. The framework organizes an analysis chain comprising objective-profile extraction, case-bundle planning, shared research memory, domain-specific artifact generation with context-isolated sub-agents, guarded runtime execution with policy-constrained overrides, and evidence reporting. Retrieval-augmented generation supports each sub-agent, while a human-in-the-loop approval gate and a two-layer artifact management scheme separate candidate outputs from the validated runtime baseline to improve traceability, reproducibility, and execution safety. A sports hall occupancy-driven thermal and air quality analysis task is used to evaluate the framework, with a cross-backbone assessment across six frontier LLMs. All backbones complete the workflow and produce identical physical outputs under the guarded execution policy, while exhibiting substantial planning-level diversity in artifact expressiveness and scenario design.

Keywords: Learning Systems, Fuzzy and Neural Systems, Process Automation
Abstract: Semantic segmentation of kelp holds critical significance for intelligent aquaculture and automated phenotypic analysis. However, conventional models suffer from representation collapse due to the extreme pixel-level long-tailed distribution between the dominant kelp body and sparse holdfast structures. To address this, we propose a Frequency-Calibrated UNet (FCUNet) featuring an Adaptive Frequency-Calibrated Classifier (AFCC). By integrating Adaptive Feature Norm (AFN) and Channel Group Scaling (CGS), the AFCC recalibrates logit distributions to prevent the marginalization of kelp holdfast category. Furthermore, we introduce an optimized compound loss that synergistically couples pixel-wise hard-sample mining with global structural consistency. Evaluated on our self-collected land-based Macroalgae Phenotyping Image Dataset (MPID), the framework demonstrates superior robustness across five mainstream backbones. Notably, the proposed method achieves significant performance gains for the holdfast category, improving the IoU and F1 Score by 0.1897 and 0.1933 respectively in the UNet architecture, effectively ensuring structural integrity for macroalgae phenotyping.

Keywords: Robotics, Optimal Control, Learning-based Control
Abstract: To address the challenge of balancing initial solution efficiency and asymptotic convergence in complex environments, this paper proposes CG-RRT*, a real-time path planning method guided by a continuous cost field. A lightweight neural network is employed to construct a global continuous cost field offline, effectively eliminating the high computational overhead of online inference. By adaptively adjusting the sampling radius based on the cost field and employing a two-stage strategy---integrating frontier guidance with ellipsoidal constraints---the algorithm achieves rapid exploration and stable convergence. Experimental results demonstrate that CG-RRT* achieves a 100% success rate under a 2% optimality threshold across various complex environments. With an average planning time of only 84~ms, the proposed method is approximately 7 times faster than the state-of-the-art NIRRT* while maintaining comparable path quality. Its robust generalization capability is further validated in unseen complex scenarios and real-world ancient architecture inspection tasks.

Keywords: Estimation and Identification, Robotics
Abstract: This paper studies motion estimation of moving energy emitters using passive sensors. The emitters may be light, acoustic, or radio sources. While the bearing vector pointing from the sensor to the emitter can be easily obtained, existing approaches mainly rely on the bearing-only motion estimation method. However, this method suffers from a fundamental limitation that the sensor must have lateral motion to ensure observability. Unfortunately, this lateral motion requirement often conflicts with the sensor's desired motion in many tasks. In this paper, we point out that the received signal strength, which can also be obtained easily in many ways, can greatly enhance motion estimation. Surprisingly, this strength information has not been well explored so far. Here, we propose a new bearing-strength method to fully exploit both the bearing and strength measurements. Our theoretical analysis shows that the system observability is significantly enhanced in the sense that the lateral motion condition is not required anymore. Real-world experimental results verify the proposed method and the theoretical analysis. It is notable that the benefit of the proposed method comes with no additional cost since it simply utilizes the received strength information that has not been fully exploited in the past.

Keywords: Nonlinear Systems and Control, Control Applications, Optimal Control
Abstract: Input-to-state safety is a crucial property for evaluating the robustness of a safe set against unknown disturbances. To guarantee the input-to-state safety of the safe set, input-to-state safe control barrier functions (ISSf-CBFs) have been proposed and further extended to a high-order version, termed as input-to-state safe high-order control barrier functions (ISSf-HOCBFs) to handle safety constraints with arbitrary relative degree. However, existing results on ISSfHOCBFs are limited to matched disturbances, which cannot affect the high-order control barrier function design until the control input is applied. In this paper, we extend the input-tostate safety guarantee from CBFs to HOCBFs under unmacthed disturbances, and propose a new type of ISSf-HOCBFs to handle safety constraints with arbitrary relative degree. A case study of an adaptive cruise control system is provided to demonstrate the effectiveness of the proposed methods

Keywords: Multi-agent Systems, Robotics, Man-machine Interactions
Abstract: To address the area coverage challenge in heterogeneous UAVs-UGVs systems, this article proposes a fixed-time distributed robust coverage control method based on human-machine interaction via large language model (LLM). This approach is applicable to scenarios involving system uncertainties and external disturbances. First, the multi-dimensional dynamic models of UAVs and UGVs are unified and the Voronoi diagram partitioning method is used to obtain the area and centroid for each region. A fixed-time observer is developed to estimate system uncertainties and external disturbances. A distributed fixed-time controller based on dynamic fast terminal sliding mode (FTSMC) is developed to achieve rapid error convergence within the fixed time. For scenarios such as issuing commands during emergencies, a method is developed where LLM parses human input dialogues and subsequently issues coverage tasks to the system. The Lyapunov function is employed to prove system stability under the designed control framework. Simulation results show that the coverage control error converges to near-zero within the fixed time. The proposed coverage control method enables the heterogeneous UAVs-UGVs systems to achieve area coverage within the specified timeframe.

Keywords: Learning Systems, Learning-based Control, Multi-agent Systems
Abstract: With increasingly complex mission requirements, current multi-robot systems often struggle to achieve stable and effective cooperative task planning in open-ended, partially observable, and communication-constrained real-world scenarios while ensuring the executability of joint plan. To address these challenges, this paper proposes a learning-based general hierarchical planning framework. At the top layer, a large language model is introduced for task understanding and decomposition of complex missions. The middle layer constructs a modular cooperative capability library, in which policies are uniformly encapsulated as waypoint generation modules, thereby constraining high-level reasoning and enhancing cross-task transferability. The bottom layer develops a waypoint-guided deep reinforcement learning motion planner that performs online collision avoidance and stable navigation using only local sensing. Experimental results demonstrate that, under diverse open-ended instructions, the top-layer planner can reliably generate executable task decomposition and assignment schemes; across different environment structures, the proposed motion planner consistently completes waypoint navigation, validating the generality and deployability of the proposed framework in open-task and communication-constrained scenarios.

Keywords: Learning Systems, Real-time Systems, Sensor/Data Fusion
Abstract: Accurate open-vocabulary 3D scene understanding requires semantic representations that are both language-aligned and spatially precise at the pixel level, while remaining scalable when lifted to 3D space. However, existing representations struggle to jointly satisfy these requirements, and densely propagating pixel-wise semantics to 3D often results in substantial redundancy, leading to inefficient storage and querying in large-scale scenes. To address these challenges, we present emph{PLAF}, a Pixel-wise Language-Aligned Feature extraction framework that enables dense and accurate semantic alignment in 2D without sacrificing open-vocabulary expressiveness. Building upon this representation, we further design an efficient semantic storage and querying scheme that significantly reduces redundancy across both 2D and 3D domains. Experimental results show that emph{PLAF} provides a strong semantic foundation for accurate and efficient open-vocabulary 3D scene understanding. The codes are publicly available at https://github.com/RockWenJJ/PLAF.

Keywords: Intelligent and AI Based Control, Robotics, Multi-agent Systems
Abstract: This paper presents HiveNav, a distributed semantic exploration framework for multi-UAV target search in structured environments. Inspired by the hierarchical decomposition and memory-driven reasoning style of CityNavAgent~cite{zhang2025citynavagent}, the full stack is redesigned for swarm robotics with three core upgrades. First, simulated RGB-D sensing is replaced with a real RGB-LiDAR reconstruction pipeline based on FAST-LIVO2~cite{zheng2025fastlivo2}, and a colorized semantic point cloud map is constructed for planning. Second, centralized planning is replaced with distributed hierarchical semantic planning (HSP) agents that run asynchronously on each UAV, enabling parallel decision making under heterogeneous observations. Third, a swarm-memory mechanism is introduced to share real-time state, future intent, and reservation information, reducing duplicated exploration and inter-agent conflicts. On top of this semantic layer, an SPH-based swarm navigation planner~cite{li2025sphswarm} is integrated as the motion-level planner for decentralized collision-free trajectory generation. Simulated experiments in multiple indoor scenes show consistent gains over geometric and centralized baselines, including higher task success, faster first-target discovery, and lower exploration overlap, while maintaining practical planner latency in the LLM-driven decision loop. These results demonstrate that colorized semantic mapping, distributed agents, and swarm memory are complementary and jointly critical for robust multi-UAV exploration.

Keywords: Control Applications, Nonlinear Systems and Control, Modeling and Control of Complex Systems
Abstract: To reduce the structural complexity and maintenance cost associated with conventional swashplate mechanisms, a swashplateless rotor with a teetering hinge is studied. The rotor achieves cyclic pitch control through periodic acceleration and deceleration of the motor combined with a tilted lag-pitch hinge. However, the blade motion and cyclic pitch generation mechanism of the rotor are complex, and its dynamic characteristics still lack systematic modeling and analysis. Therefore, the dynamic model of the blade is established based on Lagrangian mechanics and the blade element method. In addition, a closed-loop motor speed control scheme based on quasi-proportional resonant (QPR) control and active disturbance rejection control (ADRC) is designed. Simulation and bench experimental results demonstrate that the rotor can achieve effective cyclic pitch variation and adjustable thrust vector control, and that the theoretical analysis agrees well with the experimental results, providing a feasible solution for the simplification and lightweight design of rotor systems for micro aerial vehicles (MAVs).

Keywords: Nonlinear Systems and Control, Control Applications, Automated Guided Vehicles
Abstract: This paper presents a nonlinear model predictive control (NMPC) strategy for altitude regulation of a high- altitude balloon equipped with a vent valve and a ballast release mechanism. The control objective is to track commanded altitude references under time-varying atmospheric conditions while respecting actuator limits and the discrete nature of valve operation. To better reflect practical hardware, a duty-cycle actuation model is adopted: within each control interval, the controller allocates fractions of time for venting and ballast release, and the plant implements these commands through first- order valve dynamics. The proposed NMPC is implemented and evaluated in simulation with interpolated atmospheric data. Results on step changes in altitude reference demonstrate stable tracking and physically plausible on/off valve operation while avoiding the computational burden of mixed-integer predictive control.

Keywords: Nonlinear Systems and Control, Intelligent and AI Based Control, Modeling and Control of Complex Systems
Abstract: From the origin of life to the formulation of thermodynamic laws, understanding whether physical systems can manifest intelligence has remained a central scientific challenge. The ability to quantify intelligence is essential for determining whether a system exhibits intelligent behavior, and prior theoretical work has highlighted deep connections between intelligence and the fundamental principles of living systems. Dynamical systems offer a foundational framework for modeling the evolution of physical processes; however, their potential to exhibit intelligence has not been systematically characterized. In this paper, we introduce a mathematical modeling framework that defines dynamical systems composed of two interacting subsystems as Meta-Coupled Systems (MCS) to analyze the emergence of intelligence. Leveraging the concept of state entropy, we establish the Intelligence Emerging Principle (IEP) and derive its necessary and sufficient conditions. We claim that an isolated system that satisfies the energy conservation and an increment of positive total entropy, along with an entropy-declining subsystem, is able to generate intelligent behaviors. Furthermore, we specialize IEP under linear dynamical systems with different energy forms, and find that linear systems can also exhibit intelligence in some extent. Through both linear and nonlinear illustrative simulations, we verify the effectiveness of this mathematical framework as a principled method of investigating intelligence in dynamical systems.

Keywords: Nonlinear Systems and Control, Linear Systems
Abstract: For nonlinear systems, controllability is in general difficult to prove, let alone to compute the control inputs that achieve state transitions. In this paper, we consider a class of discrete-time bilinear systems which own not only algebraically verifiable controllability criteria but also algorithms of computing the required control inputs to achieve state transitions based on the root locus approach. However, since such control inputs are approximated rather than exact, there exist computation errors between the real terminal state and the given terminal state. How to minimize the errors becomes an important issue for the bilinear systems. We first analyze the errors and find through simulations that the selection of the root locus gain significantly influences the errors. As a result, under the constraint that all the control inputs have to be real numbers, the larger the root locus gains are, the smaller the errors will be. We then provide an efficient method to minimize the errors by adding one more control input, so that the locations of the open-loop transfer function’s zeros can be adjusted and a larger gain can be obtained. Simulations demonstrate our provided method and the previous algorithms to achieve transitions between arbitrarily given pairs of states can thus be improved.

Keywords: Nonlinear Systems and Control, Optimal Control, Control Applications
Abstract: This paper addresses a data-driven predictive control problem for discrete-time nonlinear systems subject to process disturbances and measurement errors. In the absence of an accurate system model, a lifting-based representation is employed to transform the nonlinear dynamics into a linearly parameterized form, enabling tractable controller design from data. To account for bounded uncertainties, a worst-case performance criterion is incorporated, leading to an optimization problem that can be reformulated as a semidefinite program (SDP). The resulting control strategy enforces constraint satisfaction while maintaining closed-loop stability. Numerical studies demonstrate that the proposed approach achieves reliable regulation under noisy measurements and outperforms existing data-driven control schemes in terms of performance.

Keywords: Nonlinear Systems and Control
Abstract: Accurate forecasting of freight flows along international transport corridors has become increasingly vital amid global supply chain reconfiguration. This study focuses on the Trans-Caspian International Transport Route (TITR), a key Eurasian corridor linking China, Kazakhstan, the Caspian region, and Europe. The research aims to enhance the precision of freight forecasting by applying a multi-factor data preparation and scenario assessment framework grounded in artificial intelligence and data-driven analytics. The methodology integrates heterogeneous datasets from 2010–2024, sourced from national and international institutions such as Kazakh Invest, Kazakhstan Temir Zholy, the World Bank, and OECD. Five principal factor groups were identified—infrastructure investment, technological advancement and digitalization, market liberalization, environmental policy, and cargo structure transformation—each represented by quantitative indicators and derived variables. Data preprocessing included normalization, outlier treatment, interpolation of missing values, dimensionality reduction, and feature engineering to ensure consistency and analytical validity. The proposed multi-factor approach enables scenario-based evaluation of freight flow dynamics under varying development conditions. By combining theoretical principles of transport economics with systematic data preparation and factor integration, the study establishes a robust methodological basis for forecasting freight volumes and supporting strategic decisions within the TITR corridor.

Keywords: Nonlinear Systems and Control, Automated Guided Vehicles, Estimation and Identification
Abstract: This paper investigates predictive station--keeping of a passive vehicle navigating in environmental flow fields governed by the Navier--Stokes equations. A computational framework combining Proper Orthogonal Decomposition (POD) reduced--order models, moving--sensor flow estimation, and nonlinear model predictive control (MPC) is developed. The vehicle is assumed to possess limited actuation capability and can only modify its vertical velocity while horizontal motion is determined by the surrounding flow. The goal is to exploit altitude--dependent flow structures in order to regulate the vehicle toward a desired spatial region. The work should therefore be interpreted as a computational investigation of the integration of reduced--order flow estimation and predictive control rather than the development of a new MPC theory. Simulation studies illustrate how predictive control can exploit environmental flow structures for station--keeping tasks.

Keywords: Micro and Nano Systems, Motion Control, Robotics
Abstract: Remotely controllable untethered microrobots with multiple functions and performances have attracted increasing research interests in recent years. Benefiting from the small sizes, these miniaturized agents have exhibited promising application potential in biomedicine, especially for therapy tasks in confined, tortuous, and vulnerable physiological environments. Targeted navigation is crucial for practical applications, where effective path planning methods and motion control algorithms are necessary for obstacle avoidance. Moreover, at in vivo scenarios, the peristalsis of biological tissue will lead to deviations between real-time environments and pre-registered environments. Thus, a planning method with online adjustment mechanism would worthy exploring. In this work, we proposed a rapidly-exploring random tree*(RRT*)-based hybrid planning scheme. In comparison to conventional RRT* methods, a potential field-based module is employed to modify the planned paths when possible collision may occur. Then, we designed a super twisting sliding mode controller (STSMC) to govern the motion control. A disturbance observer (DOB) was presented to estimate system states and lumped disturbances. Simulations and experiments were performed to validate our proposed schemes via a vortex microswarm. The results indicate that the hybrid scheme could effectively update the motion directions for the microswarm when new obstacles are added, and the motion controller is adequate to achieve precise trajectory tracking.

Keywords: Learning-based Control, Robotics, Motion Control
Abstract: In traditional material laboratory operations, a special filament with double-helical structures is often manufactured via manual work. This manual manufacturing process uses two needles to pull viscous melt materials along curved paths and at the same time rotate the needles. Therefore, the quality of produced filaments is always affected by human operation factors. Such factors as motion consistency, rotation coordination, and process stability all impact the quality of the filament. Hence, in order to improve repeatability and reduce differences caused by different operators, this paper researches a demonstration-guided robot viscoplastic filament drawing system. In this system, a robot manipulator first learns drawing skills from human experts, then independently completes the drawing process under specific endpoint constraint conditions. Firstly, this filament drawing task is formulated as a goal-directed sequential control problem. It requires the end-effector to carry out translational and rotational movements in a coordinated way. Secondly, for understanding the multi-mode characteristic of possible expert movement paths, a strategy based on flow-matching is proposed. This strategy is applied to learn a conditional continuous-time action generation mechanism. Thirdly, the expert demonstration data is collected through simulation methods. Subsequently, the strategy undergoes offline training to approximate the velocity field of action paths in the task space. Finally, the effectiveness of the proposed method is verified by both simulation and real-world examples.

Keywords: Nonlinear Systems and Control, Robotics, Control Applications
Abstract: Output singularities and slow error convergence are two critical factors that cause sliding mode control failure. This paper proposes LogCosh Super-Twisting Control (LCSTC), a super-twisting sliding mode controller constructed with a nonsingular terminal sliding manifold that incorporates the LogCosh function. We also apply LCSTC to the path following task. LCSTC avoids singular behavior, achieves faster, finite-time convergence, and improves path-following accuracy. Based on this controller, we develop two planners. One is based on the upgraded RRT*(URRT*) algorithm for static scenes, and the other is based on the bi-robot APF(BiRAPF) algorithm for dynamic scenes. Both planners generate short paths while maintaining a safe distance, and they couple with LCSTC to form integrated plan-and-follow schemes. Simulations with LCSTC in representative two-dimensional scenarios show generally faster convergence and lower following RMSE than classical sliding mode control (SMC) and super-twisting control (STC). Experiments conducted on a magnetic microrobot further strengthen the dependable effects of the LCSTC controller in practical scenarios. The efficacy of two plan-and-follow schemes is also demonstrated by simulations.

Keywords: Intelligent and AI Based Control, Learning-based Control, Robotics
Abstract: This paper addresses the problem of isoline tracking in dynamic spatiotemporal scalar fields. This task is challenging because the field is unknown, time-varying, and can only be sensed through local pointwise measurements. Most existing methods are designed for static scalar fields. We present a physics-informed (PI) predictive control scheme for a nonholonomic vehicle. A local predictor is proposed to estimate the local field strength near a potential arrival point from short trajectory history and query-point information. The predictor employs physical residual regularization and multi-step recursive error to better capture the dynamics of spatiotemporal field evolution. A PI Predictive controller is designed with incorporation of the predictor to generate a set of feasible finite-horizon tracking trajectories and to select the optimal one. Only the first step of the optimal trajectory is implemented. The effectiveness and advantages of our scheme are validated via isoline tracking simulation experiments conducted on dynamic scalar fields.

Keywords: Learning-based Control, Intelligent and AI Based Control, Robotics
Abstract: Learning-based surgical manipulation often relies on accurate pose estimation and large annotated datasets, both of which are hard to obtain in dynamic surgical environments. We propose an end-to-end framework for learning mask-conditioned surgical grasping policies from teleoperated demonstrations. An online SAM+XMem pipeline generates target masks with minimal user input, while synchronized RGB images, masks, robot states, and actions are collected to train an Action Chunking Transformer (ACT) policy. Using RGB images and masks as dual visual streams helps the policy separate appearance from task-relevant spatial cues, improving robustness to occlusion. We evaluate the effectiveness of our method in grasp-centric surgical scenarios, specifically focusing on the Needle Pick task in both the SurRoL simulator and real-world settings. Results show that mask-conditioned input consistently improves task success, highlighting the value of explicit spatial guidance for reliable policy learning in visually challenging surgical scenes.

Keywords: Learning-based Control, Multi-agent Systems, Robotics
Abstract: This paper aims to address the distributed control problem within cooperative learning systems and achieve high-precision performance from the beginning to the end via an iterative learning process. Specifically, a PID-type distributed iterative learning control (DILC) approach is proposed for multi-agent networks to achieve learning-based output consensus. By formulating an augmented vector to incorporate all terms in the PID-type DILC, a rigorous necessary and sufficient condition is derived for consensus error convergence via two-dimensional system analysis. Notably, the presented theoretical results can be naturally generalized to conventional D-type or PD-type DILC approaches. Furthermore, simulations are conducted on wheeled mobile robots to demonstrate the effectiveness and superior performance of the proposed PID-type DILC.

Keywords: Intelligent and AI Based Control, Estimation and Identification
Abstract: Skeleton-based action recognition shows great potential in human-computer interaction and healthcare. However, mainstream high-precision models (e.g., 2S-AGCN) have a large parameter count (≈3.0M) and high computational complexity. This paper proposes a lightweight two-stream adaptive graph convolutional network (Lite 2S-AGCN), which simplifies the graph convolution structure and optimizes network depth, reducing parameters from 3.0M to 40K and model size from 11.7MB to 156KB. Experimental results show that compared with 2S-AGCN, Lite 2S-AGCN achieves 91.29% accuracy (vs. 58.37%) and 1.74ms average inference time (vs. 9.97ms) on CZU-MHAD (cross-person test), and 94.42% accuracy (vs. 35.58%) and 1.81ms average inference time (vs. 7.43ms) on UTD-MHAD. It outperforms the baseline significantly in inference speed with excellent recognition performance, making it suitable for resource-constrained edge and mobile devices.

Keywords: Sensor/Data Fusion, Robotics, Signal Processing
Abstract: For a long time, many point visibility estimation methods based on traditional geometric analysis or surface reconstruction have been studied. However, they struggle to handle scenes with varying density and noise. In this work, a real-time voxel-based point visibility estimation network, namely DeepPVE, is proposed to address the above problems. The scene points are voxelized into a grid and input into the encoder-decoder structure with sparse 3D convolution to extract local structural information and spatial relationships from the point cloud in an efficient manner. Compared to point-based approaches, the proposed DeepPVE can extract features that are easier to distinguish, leading to higher accuracy in visibility estimation. Moreover, it is robust against variations in point cloud density and noise. To overcome the lack of datasets, we propose an automatic point visibility data generator based on ray tracing which enables the self-supervised training of the proposed network, and a universal benchmark to evaluate the proposed estimation methods, enabling a comprehensive comparison. The visibility estimation tests, involving 600 objects and 300 pre-sampled viewpoints, confirm the generalization ability, adaptation to density variations, and robustness against noise of the DeepPVE in comparison with recent well-known methods. Furthermore, the relevant code will be released to promote the development of point visibility estimation.

Keywords: Sensor/Data Fusion
Abstract: Gesture recognition plays a crucial role in human-computer interaction, robotics, and assistive technologies. In this study, we propose a novel framework, MSCN-LSTM, for gesture recognition using EMG and IMU signals. The model combines Multi-Scale Convolutional Neural Networks (Multi-Scale CNN) with Bidirectional Long Short-Term Memory (BiLSTM) networks to integrate spatial features from EMG signals with temporal dependencies from IMU signals. Extensive experiments demonstrate that MSCN-LSTM outperforms traditional models like GRU, LSTM, and other multi-modal models such as DSC-GRU and Multimodal CNN across accuracy, precision, recall, and F1 score. Achieving a recognition accuracy of 97.41%, MSCN-LSTM excels in both robustness and precision. Additionally, the model demonstrates real-time applicability, with an inference time of 52.97 ms per sample, making it suitable for dynamic, real-time applications.

Keywords: Robotics
Abstract: Microrobots hold significant promise for targeted drug delivery, yet their precise navigation through vascular networks remains challenging due to the lack of accurate, high-resolution in vivo vascular maps. Existing imaging modalities are largely limited to ex vivo or phantom-based mapping, lacking the spatial resolution and penetration depth required for in situ guidance. Here, we present a methodology for constructing high-resolution 3D vascular network models in vivo using a reflection-mode photoacoustic microscopy system. Our approach achieves imaging of microvascular structures down to capillary levels in mouse ears and subcutaneous tumors. Through a dedicated image processing pipeline, raw volumetric data are transformed into quantifiable, navigable vascular maps that preserve topological fidelity. This work addresses the critical bottleneck of in situ map construction, providing essential support for microrobot navigation and targeted therapeutic delivery.

Keywords: Sensor/Data Fusion
Abstract: Gallium-based liquid metals (LMs) that remain fluid at room temperature combine fluidic mobility with high metallic electrical conductivity, emerging as highly attractive transformative materials in flexible electronics. However, constrained by their extremely high surface tension, liquid metals struggle to achieve reliable wetting and adhesion on flexible substrates, which severely hinders their widespread application in flexible electronics. Here, we report a magnetically controlled superwetting strategy that enables the construction of patterned liquid metal circuits on flexible substrates. Our approach is based on an iron-liquid metal composite (FeLMC), which effectively overcomes the high surface tension limits of liquid metals, enabling the LM to undergo ultrafast spreading on flexible substrates such as Ecoflex, thereby exhibiting remarkable superwetting performance. Even if prolonged exposure leads to a degradation in wetting performance, the coating's superwetting properties can be regenerated through simple mechanical stirring. We designed and fabricated a liquid metal based flexible strain sensor by using FeLMC as the wetting layer to monitor bending angle of a human finger. It exhibits a highly linear response in its relative resistance change with respect to the finger bending angle. The proposed method offers a highly promising way for advancing the low-cost and scalable manufacturing of liquid metal based flexible electronic devices and wearable systems.

Keywords: Robotics, Motion Control
Abstract: The Remote Center of Motion (RCM) constraint plays a critical role in robot-assisted minimally invasive surgery, ensuring that surgical instruments move around a fixed point and thereby reducing tissue damage near the surgical incision. RCM constraint methods can be categorized into mechanical and algorithmic approaches. Mechanical methods provide high accuracy but lack flexibility, while algorithmic methods offer greater adaptability; however, existing algorithmic approaches typically maintain RCM constraints in real time at the velocity level, which can fail under task conflicts or in complex environments. To address this issue, this paper proposes a novel robot path planning method based on two-stage sampling, which enforces RCM constraints through geometric construction rather than real-time correction, explicitly considering the global feasibility of the path. First, the robot performs a collision-free first-stage sampling in the task space to generate an initial path of the working points. Then, based on geometric relationships, all initial path points are mapped through the RCM point to compute the corresponding robot joint-end poses. Finally, a second-stage sampling is performed on the obtained joint-end poses to generate the final path in the robot’s joint or Cartesian space. To address sudden pose changes and increased RCM errors caused by initial path points being close to the RCM point, a distance-checking mechanism is introduced, and helical interpolation is applied to poses exceeding the threshold, ensuring path continuity and satisfaction of the RCM constraint. Experimental results demonstrate that the proposed method produces feasible paths that meet the precision requirements of medical robots under RCM constraints.

Keywords: Multi-agent Systems, Linear Systems
Abstract: Cooperative output regulation of singular multi-agent systems has wide applications in unmanned systems, smart grids, and other fields. Prior study provided the pioneering work on the cooperative output regulation with deterministic models by introducing the novel distributed observers. However, when the mathematical models of follower agents are unknown, traditional model-based design methods will no longer be applicable. This study extends the existing results to more general linear singular discrete-time multi-agent systems with unknown system matrices. To deal with the external disturbances, a coordinate transformation is employed to transfer the original tracking problem to a simplified stabilization problem. By using input and state data satisfying the full rank requirement, the feedback gain for each agent can be designed by solving data-based linear matrix inequalities. Then, a data-driven distributed control scheme is proposed to solve the problem under some mild conditions. Finally, an illustrative example is provided to verify the correctness of the proposed scheme.

Keywords: Linear Systems
Abstract: Controlling a reaction-diffusion PDE–ODE system presents a formidable challenge due to the inherent instability of both the open-loop PDE and open-loop ODE components. Consequently, a single controller must compensate for the instability of both the PDE and ODE states simultaneously. In this study, we address this challenging control problem inherent in reaction-diffusion PDE–ODE systems. Moreover, we extend our investigation to encompass a scenario where the connection between the PDE and ODE is subject to distributed delays, further complicating the control task. To tackle this formidable challenge, we employ a combination of predictor feedback and infinite-dimensional backstepping techniques to design an effective controller. We rigorously analyze the stability properties of the resulting closed-loop system, considering both L2 and L∞ norms, and establish proofs based on Lyapunov construction. Finally, we illustrate the effectiveness of our proposed controller through a simulation example.

Keywords: Control Applications, Nonlinear Systems and Control, Multi-agent Systems
Abstract: 本文探讨了协同形成控制 多伞翼系统的问题，涉及大量 滑翔翼和非凸翼的非线性飞行动力学 耦合约束包括碰撞避免和 队形维护。为了解决这些困难，一 基于交替方向的分布式控制器 提出了乘法法（ADMM），其中子问题 通过使用微分动力学高效求解 规划（DDP）、封闭形式投影和半正定投影 松弛（SDR）。受益于 的并行结构 ADMM 与约束的有效处理， 所提方法适用于实时控制和 具有良好的可扩展性，且对 副翼。多个副翼的模拟结果 阵型切换验证 拟议方法。

Keywords: Multi-agent Systems, Motion Control, Control Applications
Abstract: This paper investigates a position synchronization problem of multiple permanent magnet synchronous motor (PMSM) systems under parametric uncertainties and directed communication network. Based on the distributed internal model approach, a distributed position synchronization controller is proposed, which can simultaneously achieve position tracking, disturbance rejection, and robustness. Moreover, it can lead to superior transient performance. The simulation results demonstrate the effectiveness of our design.

Keywords: Linear Systems, Multi-agent Systems
Abstract: This paper studies a scalable distributed algorithm for solving linear equations over double-layered multi-agent networks. The network is divided into clusters, each consisting of an aggregator and a group of agents. Agents exchange local information within clusters, while aggregators coordinate information across clusters. Unlike existing results that require the inter-cluster communication graph to remain connected at all times, this paper considers a switching inter-cluster communication graph that may be disconnected at any time instant, as long as it satisfies a jointly connected condition. By modeling the resulting error dynamics as a linear switched system and exploiting its uniform complete observability (UCO) property, we establish exponential convergence of the distributed algorithm to the unique solution of a linear equation.

Keywords: Intelligent and AI Based Control, Learning-based Control, Signal Processing
Abstract: This paper proposes an integrated frequency identification and reinforcement learning based approach for active vibration control system to especially address the challenge from the frequency-switching excitation. Technically, we first propose a sparse autoencoder based method to identify frequencies, capable of directly extracting feature and achieving fast inference speeds, and then design an optimal frequency-specific control policy with the aid of reinforcement learning technique. In particular, the Beta distribution is adopted in the construction of FxLMS-based reinforcement learning environment and used to model the action in Markov decision process due to its bounded property, which not only results in a stable training process, but also accelerates the convergence of policy optimization. Our integrated method requires no manual parameter tuning when the frequency of the excitation dramatically changes, and effectively minimizes the total error energy under different frequencies of the exciter, demonstrated in the experiment of 8-channel AVC system.

Keywords: Intelligent and AI Based Control, Multi-agent Systems, Robotics
Abstract: Stochastic compositional optimization (SCO) arises in a wide range of applications, including risk management and reinforcement learning. However, most existing methods for SCO rely on the smoothness assumption and require large batch sizes. To overcome these limitations, we propose a novel stochastic gradient-free algorithm, called SRMGF, designed for general nonsmooth nonconvex SCO. Unlike prior large-batch methods, SRMGF uses an increasing batch size for inner function evaluations, while employing only a single outer sample per iteration for gradient estimations. We further incorporate a momentum-based variance-reduction scheme to stabilize estimations. Theoretically, we prove that SRMGF converges to a (delta, epsilon)-Goldstein stationary point with a rate of mathcal{O}(delta^{-1/2}T^{-1/4}), achieving a total function query complexity of mathcal{O}(delta^{-4} epsilon^{-6}). Notably, SRMGF is the first algorithm to guarantee convergence without the need for large outer batches in nonsmooth SCO. Finally, numerical experiments on real-world tasks demonstrate the efficiency of SRMGF.

Keywords: Optimal Control, Learning-based Control
Abstract: The method of multipliers (MM) is a standard approach for equality-constrained convex optimization, but it often suffers from high computational cost in the primal update, and its dual update amounts to a gradient ascent step that can result in slow convergence. The bundle method of multipliers (BMM) mitigates these issues by incorporating a proximal bundle model into both the primal and dual updates, leading to a cheaper primal update and a more effective dual update. Nevertheless, the primal subproblem in BMM may still be expensive to solve. To improve computational efficiency, we propose a weighted BMM, which introduces a weighted proximal term to diagonalize the quadratic component of the primal subproblem. With this modification, the subproblem becomes significantly more tractable when solved using gradient-based dual approaches. We establish convergence of the proposed method under standard assumptions and demonstrate its superior time efficiency through numerical experiments.

Keywords: Intelligent and AI Based Control, Multi-agent Systems, Robotics
Abstract: Beyond-visual-range (BVR) air combat is a complex sequential decision-making problem. Reinforcement learning (RL) agents often struggle to discover multi-phase tactics in such high-dimensional spaces. While large language models (LLMs) offer strong strategic reasoning, integrating them directly into RL training remains an open challenge. To address this, a novel framework named KL-Prior Regularized PPO is proposed in this paper. This training-time-only framework incorporates a data-driven LLM tactical prior into the Proximal Policy Optimization (PPO) objective via a decaying KL-divergence penalty. This prior is constructed by analyzing 299 winning engagements from a baseline RL agent, successfully encoding the ``fire-and-extend" BVR doctrine into soft action distributions. Unlike inference-time fusion methods, actions in this approach are sampled exclusively from the PPO policy. This design preserves the stationary Markov Decision Process (MDP) assumption, which is critical for stable value function learning. The proposed method is evaluated in a high-fidelity 3D BVR simulator featuring F-16 aircraft, realistic sensors, and AIM-120C missile dynamics. Over 50,000 training episodes, KL-Prior PPO achieves a 28.8% accumulated win rate, outperforming the pure PPO baseline (25.0%). Furthermore, three failed inference-time fusion variants are analyzed. Entropy collapse and value function poisoning are identified as their fundamental failure mechanisms. These insights provide practical design guidance for robust LLM-RL integration in complex control domains.

Keywords: Optimal Control, Real-time Systems, Estimation and Identification
Abstract: This paper proposes an analytical Koopman-based linear model predictive control (MPC) method for real-time quadrotor trajectory tracking. While linear MPC offers computational efficiency, it sacrifices modeling fidelity; nonlinear MPC solved via sequential quadratic programming achieves high accuracy but requires multiple iterations at each control step. We develop a systematic procedure to derive Koopman observables that lift the dynamics into a quasi-linear model with state-dependent control matrix. An assumed state trajectory converts this to a linear time-varying system at each control period, enabling quadratic program formulation with guaranteed real-time solvability. An incremental nonlinear dynamic inversion (INDI)-based robust control allocation scheme is proposed, which requires no precise control effectiveness model. Simulation results demonstrate tracking performance comparable to nonlinear MPC with deterministic computation times. The proposed method requires no training data collection, making it straightforward to implement.

Keywords: Learning Systems, Multi-agent Systems, Networked Control
Abstract: Generalization analysis in personalized federated learning (PFL) is crucial for the design of learning algorithms. Existing theoretical studies primarily focus on characterizing the generalization performance of the global averaged model, while the performance of individual agents’ local models remains largely unexplored. To address this gap, we first establish agent-specific generalization error bounds for decentralized personalized federated learning (D-PFL) based on the notion of algorithmic stability. By reformulating the iteration process of D-PFL as a linear dynamical system and leveraging matrix decomposition techniques, we explicitly characterize how data perturbations propagate across the shared and personalized layers. Our analysis reveals the connection between network topology and agent-level generalization: agents with higher topological centrality benefit from more effective diffusion of perturbations through the communication network and consequently achieve improved generalization performance. In contrast, peripheral agents experience weaker perturbation dissipation and are therefore more susceptible to overfitting. Numerical experiments are provided to validate the theoretical findings.

Keywords: Optimal Control, Control Applications, Motion Control
Abstract: Fixed-wing UAV navigation in cluttered environments requires robust spatial-temporal planning and high-fidelity tracking under non-holonomic constraints. To address these challenges, this paper presents a hierarchical framework tightly coupling a kinematic-aware front-end, a MINCO-based trajectory optimizer, and an incremental Nonlinear Model Predictive Control (NMPC) scheme. To strike an optimal balance between computational tractability and geometric tightness, an adaptive Safe Flight Corridor (SFC) generation strategy is developed based on a 3D Dubins search. Furthermore, a novel Kinematics-Aware Terminal Regularization (KATR) mechanism is introduced. It effectively resolves kinematic boundary conflicts caused by naive zero-acceleration assumptions, preventing forced zero-curvature terminal states while unlocking tangential acceleration degrees of freedom. For robust execution, the incremental NMPC utilizes a physics-informed Zero-Control Invariant Set Extrapolation to eradicate terminal predictive conflicts, ensuring exact terminal convergence. ROS and PX4 SITL simulations demonstrate that the framework achieves millisecond-level online planning and high-fidelity 3D maneuver tracking with sub-meter root-mean-square error (RMSE).

Keywords: Estimation and Identification, Man-machine Interactions
Abstract: Real-time and accurate joint angle prediction is essential for human-machine interaction systems, yet it remains difficult due to substantial inter-subject variability and diverse locomotion tasks. While advances in neural networks have boosted prediction accuracy, data-driven approaches often struggle to achieve low-latency estimation in real-world online settings. To address this gap, we propose a dual-stage personalization Kalman filter (DP-KF) framework that couples long-term personalization with short-term dynamic correction. In the first stage, human physiological characteristics are incorporated as individualized prior conditions, and a conditional variational autoencoder (CVAE) is employed to generate a long-timescale baseline trajectory of joint angles. Meanwhile, a plantar-pressure matrix is fused to enrich gait, which improves the identifiability and stability of the learned latent variables. In the second stage, a Kalman filter (KF) operates under a system dynamics model and applies recursive estimation. It performs online short-timescale prediction of the residual between the long-timescale baseline trajectory generated in the first stage and the real-time observed joint angles. Finally, comparative experiments and ablation studies demonstrate that the proposed method achieves superior accuracy, while maintaining real-time performance.

Keywords: Adaptive Control, Optimal Control, Learning-based Control
Abstract: This paper proposes a safety-critical approximate optimal control framework, which synthesizes dual control for safety and optimality using Nesterov accelerated gradients with fixed-time stability (DCSO-NAG-FxT). For both safety and optimality objectives, a dual-objective function is constructed, and the Nesterov acceleration-based fixed-time learning law is applied to update the learning weights. Rigorous theoretical analysis establishes practical fixed-time stability for the closed-loop system, ensuring that the settling time of weights is uniformly bounded. By deriving explicit conditions for the bi-homogeneous weight updates, this work bridges the gap between accelerated gradient methods and fixed-time stability in adaptive dynamic programming. Furthermore, a safety-aware synthesis mechanism is integrated to guarantee the forward invariance of safe sets during the learning transient. Simulations on a nonlinear benchmark demonstrate faster convergence, lower cost, and zero safety violations.

Keywords: Control Applications, Motion Control, Optimal Control
Abstract: With the rapid advancement of information and automation technologies, autonomous ground vehicles (AGVs) have been widely deployed across various industrial and domestic applications. Trajectory tracking of AGVs remains a challenging task in practice due to the presence of external disturbances and physical constraint. To address this challenge, this extended abstract presents a tube-based model predictive control (MPC) framework. The key idea is to reformulate the original tracking problem as a regulation problem by analyzing the dynamics of the tracking error. A robust positively invariant set is then computed to capture the influence of external disturbances on the evolution of the tracking error. Furthermore, a gain-scheduling mechanism is developed to reduce the conservatism arising from the invariant set. The resulting tube-based MPC scheme ensures constraint satisfaction and tracking accuracy while effectively handling external disturbances.

Keywords: Robotics, Intelligent and AI Based Control, Control Applications
Abstract: Linear-actuator driven full-size humanoid robots (LF-HRs) provide high payload capacity and low energy consumption during static standing. Nevertheless, the coupled closed-chain kinematics and linear actuator dynamics introduce a deployment gap, making reinforcement learning (RL) locomotion policies trained on a 12-DoF opened-chain model difficult to transfer and deploy. To address this gap, we propose a robust two-stage closed-chain Sim2Sim gait transfer controller, which geometrically correlates rotational joint angles with linear actuator displacements through a high-fidelity virtual-physical joint space mapping. Specifically, sagittal-plane projection is employed to model the hip and knee linkages, and a Jacobian-based iterative solver is developed to accurately handle the parallel ankle mechanism, thereby guaranteeing coordinated and kinematically feasible motion under closed-chain constraints. Building on this mapping, we combine potential-based reward shaping and domain randomization to transfer the learned policy from the opened-chain training environment to a high-fidelity closed-chain model, followed by two-stage validation on unstructured terrains, including height field and rough ground. The proposed controller provides an effective pathway for the high-performance deployment of LF-HRs.

Keywords: Robotics, Modeling and Control of Complex Systems, Motion Control
Abstract: This paper explores a robotic system that synergizes the adhesion capability of wall-climbing robots with the agility of aerial vehicles for building maintenance applications. While wall-climbing robots can adhere to vertical surfaces, they suffer from limited obstacle-crossing abilities; conversely, unmanned aerial vehicles offer flexible maneuverability around walls but face challenges including limited proximity to surfaces, short endurance, and operational risks. To address these limitations, we propose a tether-connected tilt-rotor quadrotor robot that leverages the tensile force from the tether to enhance endurance and flexibility. The integration of a tilt-rotor mechanism with the tether enables decoupled position and attitude control, allowing the robot to approach walls at arbitrary orientations and substantially improving adaptability to complex environments. The dynamic model of this tethered tilt-rotor quadrotor is analyzed, with system identification employed to determine key parameters. Controller design and parameter tuning are subsequently performed based on the established model and identified parameters. Experimental results validate the effectiveness of the controller in angle command tracking and demonstrate the system's potential for wall inspection applications.

Keywords: Multi-agent Systems, Intelligent and AI Based Control, Learning-based Control
Abstract: This paper studies a coordination-layer problem for self-organized airship encirclement and hotspot surveillance. A team of airships moves on a prescribed standoff ring around a protected location, while its angular distribution is adjusted according to a hotspot-induced directional surveillance demand. To characterize this task, a density-based equal-mass criterion is introduced on the angular domain, where the hotspot increases the task value of the facing sector without removing the coverage demand on the remaining ring. Based on this formulation, Graph-MAPPO is developed by embedding a lightweight ring message-passing encoder into the MAPPO actor--critic architecture. The proposed method preserves the standard PPO optimization backbone, but injects a graph inductive bias matched to the fixed neighbor-only communication topology, so that each agent can aggregate multi-hop neighborhood information for local mass equalization. The model is intended for the tangential redistribution layer, with radial regulation and low-level stabilization assumed to be handled by an inner-loop controller. Comparative simulations, including multi-seed training, message-passing-depth ablation, random-initial-formation tests, and spacing-dynamics evaluation, demonstrate the advantages of Graph-MAPPO over an MLP-based MAPPO baseline and a handcrafted local controller in convergence and hotspot-oriented redistribution.

Keywords: Multi-agent Systems, Learning-based Control, Intelligent and AI Based Control
Abstract: Multi-agent adversarial games with pre-protection points are challenging due to partial observability, sparse rewards, delayed feedback, and dynamically evolving opponent strategies. In such settings, pure centralized training with decentralized execution (CTDE) reinforcement learning often suffers from low sample efficiency and fragile coordination, while standard behavior cloning struggles to represent multimodal expert policies and may lead to an unstable transition from imitation to autonomous exploration. To address these issues, we propose Generative-Initialized Gated-Adaptation (GIGA), a two-stage framework that bootstraps multi-agent reinforcement learning with generative expert priors. In Stage I, a conditional variational autoencoder is trained to model the multimodal distribution of expert behaviors, yielding a structured latent prior for policy initialization. In Stage II, a bias-initialized gating mechanism adaptively combines the frozen expert prior with a learnable residual policy under the CTDE paradigm, enabling a smooth transition from expert-guided behavior to maximum-entropy reinforcement learning. Together with twin centralized critics, GIGA improves convergence speed, final performance, and robustness across multiple training regimes, outperforming strong multi-agent reinforcement learning baselines and the expert policy used for initialization.