IEEE/CAA Journal of Automatica Sinica

OpenClaw in the Wild: Security Analysis of Autonomous Agents

Wanlun Ma, Qing-Long Han, Xiaogang Zhu, Wei Zhou, Junwu Xiong, Zhihang Ren, Sheng Wen, Yang Xiang

2026, 13(6): 1257-1273. doi: 10.1109/JAS.2026.126209

Abstract(91) HTML (23) PDF(3)

Abstract:
Autonomous self-hosted AI agent platforms are rapidly evolving from prompt-response assistants into persistent systems that can maintain long-lived state, invoke tools, ingest external content, and execute environment-changing actions. While this transition enables practical automation, it also introduces lifecycle security risks that cannot be fully explained by prompt-level analysis alone. In this paper, a security analysis of OpenClaw is presented, with OpenClaw serving as a representative autonomous agent operating environment and a concrete case study for broader security challenges in emerging agent ecosystems. A trust-boundary-first perspective is adopted to examine how attacks propagate across five boundary classes: Channel-Access, Session-and-State, Tool-Execution, External-Content, and Extension Supply-Chain. The results presented in this paper show that threats such as indirect prompt injection, memory poisoning, unsafe tool invocation, data exfiltration, and malicious skill abuse are not isolated anomalies; rather, they are stage-specific manifestations of a common systems problem in which untrusted influence progressively crosses into higher-privilege contexts. Based on this analysis, the defense-in-depth implications for OpenClaw deployments are discussed, including boundary-aware isolation, capability-scoped tool mediation, memory integrity controls, extension governance, and evidence-oriented operational oversight. This study provides a practical framework for evaluating and hardening long-running, tool-capable, autonomous AI agents in realistic deployment settings.

Fixed-Time and Predefined-Time Synchronization of Inertial Memristive Neural Networks With Proportional Delay and Mismatched Switching Jumps via Stochastic Particle Swarm Optimization

Qiang Lai, Jun Wang, Leimin Wang, Minghong Qin

2026, 13(6): 1274-1287. doi: 10.1109/JAS.2025.126008

Abstract(16) HTML (8) PDF(4)

Abstract:
This paper studies the fixed-time synchronization (FxTS) and predefined-time synchronization (PTS) of a class of inertial memristive neural networks (IMNNs), which have unbounded proportional delay independent of time linearity and mismatched switching jump coefficients. Based on Filippov solution theory and Lyapunov methods, this work develops an enhanced FxTS criterion delivering a tighter upper bound on convergence time and thereby extending guaranteed fixed‐time behavior to a wider range of networks. A refined PTS condition that incorporates additional state-dependent terms accelerates error decay and reduces conservatism during the transient response. Numerical simulations show that the convergence rate of PTS under this strategy is significantly improved. Moreover, an optimization model with minimum control energy and dynamic error as objective functions is proposed to obtain more accurate controller parameters, and the stochastic inertia weight particle swarm optimization (SIWPSO) algorithm is introduced to solve the optimization model. Numerical studies not only validate the theoretical results for both FxTS and PTS but also demonstrate a secure communication application in which the chaos of the IMNNs acts as a masking carrier and enables perfect encryption and decryption of a complex test signal through SIWPSO-optimized FxTS and PTS.

LL-Refiner: Learning Adaptive Refinement for Ultra-High-Definition Low-Light Image Enhancement

Xunpeng Yi, Qinglong Yan, Yibing Zhang, Zongrong Li, Hao Zhang, Zhanchuan Cai, Jiayi Ma

2026, 13(6): 1288-1300. doi: 10.1109/JAS.2026.125939

Abstract(31) HTML (11) PDF(4)

Abstract:
Low-light image enhancement aims to address various degradations in low-light conditions, such as low illumination, noise pollution, color distortion, and missing scene content. With advances in digital imaging technology, the resolution of captured images has seen substantial improvements. This poses new challenges in achieving good enhancement performance for multi-scale details in ultra-high-definition images, as well as in managing the overhead for supporting the use of consumer-grade GPUs. In this paper, we proposed learning the adaptive refinement framework for ultra-high-definition image enhancement, termed LL-Refiner. It integrates the advantages of hierarchical adaptive refinement guided by coarse enhancement results to enhance the ultra-high-definition images. In detail, firstly, we conduct the coarse enhancement on the low resolution image by employing a Transformer-based coarse enhancement network. Secondly, the coarse enhancement output is fed into the adapted refinement injection. It assigns resolution-aware inputs as guidance to the corresponding adaptive aggregation module, which interacts with the backbone features of the adaptive refinement network. Ultimately, the adaptive refinement network incorporates a combination of hierarchical dense residual connection modules and lightweight convolutional modules at different resolution stages. Also, it integrates a multi-scale enhanced perceptual loss to progressively achieve ultra-high-definition image enhancement. Extensive experiments on ultra-high-definition image enhancement validate the effectiveness and superiority of the proposed method. Our code is publicly available at https://github.com/XunpengYi/LL-Refiner.

Fast Anomaly Detection and Joint State Estimation for Perturbed Nonlinear Systems With Prolonged Output Anomalies

Guangdeng Chen, Xiao-Jie Peng, Yan Lei, Chao Huang, Hongyi Li

2026, 13(6): 1301-1313. doi: 10.1109/JAS.2025.125900

Abstract(23) HTML (12) PDF(0)

Abstract:
State estimation under anomalies such as disturbances and faults remains a fundamental challenge in nonlinear systems, with its difficulty further exacerbated by potential network attacks. This study investigates fast anomaly detection and state estimation for perturbed nonlinear systems where actual outputs may be anomalous over a prolonged period. First, a fixed-time observer is constructed. By leveraging integral-type composite Lyapunov functions and homogeneity theory, the error bounds are proven under varying scenarios involving model disturbances, measurement noise, and nonlinearity. Based on these bounds, a fast anomaly detection mechanism is designed. Next, a cascade predictor is developed based on the fixed-time observer, which uses historical outputs from a previous time window to predict the current system state. Simultaneously, an algorithm is proposed to determine the reference historical output based on anomaly detection results, improving long-term prediction accuracy and mitigating the impact of anomaly detection delays. Finally, the secure state estimation is derived by fusing states from the fixed-time observer and the cascade predictor, depending on the anomaly detection results. The effectiveness of the proposed method is demonstrated through simulations on autonomous vehicles.

H_∞ Optimal Output Regulation of Unknown Linear Systems via an Adaptive Dynamic Programming and Internal Model

Yong-Sheng Ma, Jian Sun, Yong Xu

2026, 13(6): 1314-1324. doi: 10.1109/JAS.2026.125777

Abstract(25) HTML (8) PDF(2)

Abstract:
This paper delves into the $ H_\infty $ optimal output regulation problem for continuous-time linear systems with an unknown system model. By integrating the internal model principle with optimal control, we derive an optimal control policy and a worst-case disturbance policy through the formulation and solution of a zero-sum game problem. Subsequently, leveraging adaptive dynamic programming, we propose a policy iteration learning algorithm capable of learning both the optimal control policy and the worst-case disturbance policy directly from system data. The existing algorithms necessitate an initial stabilizing policy, a full-rank condition, and the storage of historical data to guarantee algorithm convergence. In contrast, we design a dual policy iteration algorithm equipped with an online learning mechanism, thereby eliminating these additional prerequisites. Simulation results with an autonomous ground vehicle underscore the effectiveness of our proposed algorithm, and its superiority is further demonstrated through comparisons with existing methodologies.

Quality or Quantity? Error-Informed Selective Online Learning With Gaussian Processes in Multi-Agent Systems

Zewen Yang, Xiaobing Dai, Jiajun Cheng, Yulong Huang, Peng Shi

2026, 13(6): 1325-1338. doi: 10.1109/JAS.2025.125993

Abstract(22) HTML (7) PDF(1)

Abstract:
Effective cooperation is pivotal in distributed learning for multi-agent systems, where the interplay between the quantity and quality of the machine learning models is crucial. This paper reveals the irrationality of indiscriminate inclusion of all models on agents for joint prediction, highlighting the imperative to prioritize quality over quantity in cooperative learning. Specifically, we present the first selective online learning framework for distributed Gaussian process (GP) regression, namely distributed error-informed GP (EIGP), that enables each agent to assess its neighboring collaborators, using the proposed selection function to choose the higher quality GP models with less prediction errors. Moreover, algorithmic enhancements are embedded within the EIGP, including a greedy algorithm (gEIGP) for accelerating prediction and an adaptive algorithm (aEIGP) for improving prediction accuracy. In addition, approaches for fast prediction and model update are introduced in conjunction with the error-informed quantification term iteration and a data deletion strategy to achieve real-time learning operations. Numerical simulations are performed to demonstrate the effectiveness of the developed methodology, showcasing its superiority over the state-of-the-art distributed GP methods with different benchmarks.

Optimization of Flying Ad Hoc Network Topology and Collaborative Path Planning for Multiple UAVs

Ming He, Peizhao Wang, Haihua Chen, Bin Sun, Hongpeng Wang

2026, 13(6): 1339-1352. doi: 10.1109/JAS.2025.125846

Abstract(20) HTML (10) PDF(0)

Abstract:
Multiple unmanned aerial vehicles (UAVs) play a vital role in monitoring and data collection in wide area environments with harsh conditions. In most scenarios, issues such as real-time data retrieval and real-time UAV positioning are often disregarded, essentially neglecting the communication constraints. In this paper, we comprehensively address both the coverage of the target area and the data transmission capabilities of the flying ad hoc network (FANET). The data throughput of the network is therefore maximized by optimizing the network topology and UAV trajectories. The resultant optimization problem is effectively solved by the proposed reinforcement learning-based trajectory planning (RL-TP) algorithm and the convex-based topology optimization (C-TOP) algorithm sequentially. The RL-TP optimizes the UAV paths while considering the constraints of FANET. The C-TOP maximizes the data throughput of the network while simultaneously constraining the neighbors and transmit powers of the UAVs, which is shown to be a convex problem that can be efficiently solved in polynomial time. Simulations and field experimental results show that the proposed optimization strategy can effectively plan the UAV trajectories and significantly improve the data throughput of the FANET over the adaptive local minimum spanning tree (A-LMST) and cyclic pruning-assisted power optimization (CPAPO) methods.

Bumpless Transfer Control for State-Dependent Switched Linear Systems With Dwell-Time Constraints

Feiyue Wu, Jie Lian, Dong Wang

2026, 13(6): 1353-1361. doi: 10.1109/JAS.2025.125816

Abstract(14) HTML (6) PDF(0)

Abstract:
This paper investigates the robust bumpless transfer (BT) control problem for a class of state-dependent switched linear systems. To reduce control bumps in switched systems, a state-dependent switching law with a specified dwell-time constraint is proposed. Combined with BT performance requirements, BT control is designed to incorporate both transition-dependent control and stabilizing control. In contrast to previous interpolation-based BT control design schemes, the proposed structure decouples the controller evaluation period from the dwell-time constraints, thereby enhancing flexibility in BT control design. Through the construction of transition-dependent Lyapunov functions, sufficient conditions are established to guarantee the exponential stability and BT performance for the switched system. The proposed method is extended to disturbed switched systems with a guaranteed $ {\cal{L}}_2$-gain upper bound. To demonstrate the effectiveness of the proposed BT control strategy, an example featuring an aero-engine model is presented.

An Efficient Evolutionary Algorithm for Few-for-Many Optimization

Ke Shang, Hisao Ishibuchi, Zexuan Zhu, Qingfu Zhang

2026, 13(6): 1362-1377. doi: 10.1109/JAS.2026.125852

Abstract(72) HTML (9) PDF(2)

Abstract:
Few-for-many (F4M) optimization, recently introduced as a novel paradigm in multi-objective optimization, aims to find a small set of solutions that effectively handle a large number of conflicting objectives. Unlike traditional many-objective optimization methods, which typically attempt comprehensive coverage of the Pareto front, F4M optimization emphasizes finding a small representative solution set to efficiently address high-dimensional objective spaces. Motivated by the computational complexity and practical relevance of F4M optimization, this paper proposes a new evolutionary algorithm explicitly tailored for efficiently solving F4M optimization problems. Inspired by S-metric selection evolutionary multi-objective optimization algorithm (SMS-EMOA), our proposed approach employs a $ (\mu+1) $-evolution strategy guided by the objective of F4M optimization. Furthermore, to facilitate rigorous performance assessment, we propose a novel benchmark test suite specifically designed for F4M optimization by leveraging the similarity between the R2 indicator and F4M formulations. Our test suite is highly flexible, allowing any existing multi-objective optimization problem to be transformed into a corresponding F4M instance via scalarization using the weighted Tchebycheff function. Comprehensive experimental evaluations on benchmarks demonstrate the superior performance of our algorithm compared to existing state-of-the-art algorithms, especially on instances involving a large number of objectives. The source code of the proposed algorithm will be released publicly. Source code is available at https://github.com/MOL-SZU/SoM-EMOA.

High-Order Interaction and Low-Order Parallelization of Features Fusion With Novel Mamba-UNet Architecture for Medical Image Segmentation

Qianhang Du, Zhenyu Lei, Jiujun Cheng, Masaaki Omura, Hideyuki Hasegawa, Shangce Gao

2026, 13(6): 1378-1391. doi: 10.1109/JAS.2026.125720

Abstract(33) HTML (8) PDF(1)

Abstract:
Medical image segmentation is an essential method for computer-aided diagnosis. Although image segmentation models based on convolutional neural networks (CNNs) and vision transformers (ViTs) have achieved significant advancements, CNNs struggle to effectively capture long-range dependencies, while ViTs face limitations in local information extraction and are hindered by quadratic computational complexity. Recently, their inherent issues have been successfully addressed by the state-space models in Mamba and 2D-selective-scan in Vision Mamba. However, the presence of noise and excessive redundant information in medical images limits the practicality of these methods. To address these challenges, we propose a highly effective and accurate high-low-order feature fusion visual state space module, named HL-VSS. This module primarily consists of two core components: multi-scale spatial convolution and high-low-order feature fusion (HLFF). The former component preliminarily suppresses noise and captures multi-scale feature information from medical images, accurately extracting edge and detail features for the fusion component. The latter processes these features, further reducing redundant information through high-order interaction with 2D-selective-scan, and fuses the local features obtained by low-order parallel Mamba, ultimately extracting deeper medical image features. We incorporate HL-VSS into a U-shaped architecture, named high-low-order feature fusion visual Mamba UNet (V-UNet). Comparison experiments and ablation studies are conducted on four publicly available medical image datasets to validate the strong competitiveness of V-UNet in medical image segmentation tasks. The code is available at https://github.com/ai-dqh0106/V-UNet_Code.

Bumpless Transfer Filtering for Continuous-Time Switched Systems With Admissible Dual-Modal Persistent Dwell-Time

Jian Zhang, Yanzheng Zhu, Rongni Yang, Xinkai Chen, Donghua Zhou

2026, 13(6): 1392-1408. doi: 10.1109/JAS.2025.125948

Abstract(27) HTML (8) PDF(2)

Abstract:
This paper investigates the issue of bumpless transfer H_∞ filtering for a class of continuous-time switched linear systems. A new definition on admissible dual-modal persistent dwell-time (ADM-PDT) switching is presented, which is more general than the existing mode-dependent or admissible edge-dependent persistent dwell-time switching. To match the ADM-PDT switching, the dual-mode-dependent filter utilizing previous mode information is designed, which can enhance the performance of mode-dependent filter widely used in the literature. To smooth the filtering output bumps caused by switching filter, a novel dual-mode-dependent bumpless transfer filter is constructed by introducing a time-interpolation-function-dependent transition filter. Then, the criteria for stability and H_∞ performance analyses are derived based on a transition-dependent dual-modal Lyapunov function, ensuring that the filtering error systems driven by the bumpless transfer filter and ADM-PDT switching are globally uniformly asymptotically stable with a non-weighted H_∞ noise attenuation performance. Finally, the validity and superiority of the developed theoretical results are illustrated by using a numerical example and an unmanned marine vehicle system.

Dynamic Neural Networks for Manipulability Optimization of Omnidirectional Mobile Redundant Manipulator Under Anti-Input Disturbance

Shijun Tang, Zhongbo Sun, Yunfeng Hu, Hong Chen

2026, 13(6): 1409-1427. doi: 10.1109/JAS.2025.125963

Abstract(21) HTML (9) PDF(0)

Abstract:
Manipulability optimization plays a crucial role in the motion control of omnidirectional mobile redundant manipulators (OMRM), effectively reducing the risk of singularity. However, existing methods often overlook obstacle avoidance or simplify obstacles as single points, limiting their practical applicability. To address these issues, this paper proposes a convex manipulability optimization scheme with physical and face-avoidance constraints (C-MOPOFC), where position tracking and matrix inversion are formulated as equality constraints, while physical limitations and obstacle avoidance are incorporated as inequality constraints. To enable real-time optimization, a resistant input disturbance recursive neural network (RID-RNN) is proposed, which solves the C-MOPOFC problem in an inverse-free manner, ensuring both real-time performance and robustness against disturbances. Additionally, it overcomes the limitations of traditional time-varying optimization solvers, which suffer from high computational complexity and weak disturbance suppression. Theoretical analysis proves that RID-RNN globally converges to the optimal solution of C-MOPOFC, even in the presence of noise. Finally, numerical simulations and physical experiments validate the proposed method, demonstrating its effectiveness in enhancing manipulability while ensuring safe operation in dynamic environments.

Large Language Model-Driven Evolutionary Optimization With a Hallucination-Resilient Mechanism

Weixiong Huang, Rui Wang, Tao Zhang, Sheng Qi, Shichao Fan, Ling Wang

2026, 13(6): 1428-1445. doi: 10.1109/JAS.2025.125876

Abstract(20) HTML (7) PDF(1)

Abstract:
Large language models (LLMs) have demonstrated significant potential as black-box optimizers due to their strong reasoning capabilities. However, challenges such as the hallucination phenomenon introduce instability and uncertainty, limiting their effectiveness. This paper proposes an LLM-driven evolutionary optimization framework, referred to as LLM-driven hybrid evolutionary optimization framework (LHO), that integrates LLMs with traditional evolutionary operators. LLMs accelerate the optimization process by generating high-quality solutions, while evolutionary operators ensure stability and provide performance guarantees. To further enhance robustness, we introduce a hallucination-resilient mechanism to mitigate the risks associated with LLM hallucinations. Experimental results on various benchmark tests, encompassing single-objective, multiobjective, and complex constrained multiobjective problems, confirm the effectiveness and practicality of the proposed framework, offering valuable insights and future directions for LLM as evolutionary optimizers.

Safety and Stability of Nonlinear Systems With Multiple High-Relative-Degree Time-Varying State Constraints: Theory and Its Application to UAV Tracking

Fan Yang, Jiangping Hu, Qingrui Zhou, Xiaoming Hu

2026, 13(6): 1446-1458. doi: 10.1109/JAS.2025.125990

Abstract(15) HTML (11) PDF(2)

Abstract:
It remains a challenge to stabilize a nonlinear system while satisfying time-varying state constraints for safety. In this study, we developed a time-varying nonlinear control barrier function (TVNCBF) approach to stabilize a nonlinear system with high-relative-degree state constraints. This approach mitigated potential conflicts among multiple control barrier function constraints, providing a theoretical foundation for safety control of affine nonlinear systems. We then proposed a safety control strategy for nonlinear systems and comprehensively analyzed its safety and feasibility. At the same time, considering a scenario where an unmanned aerial vehicle (UAV) avoids several dynamic obstacles during target tracking, we designed a direct safety control for the UAV system to ensure safe target tracking. Simulation results demonstrated the safety and stability of the UAV system under the proposed control strategy.

Data-Driven Algorithms for Finite-Horizon and Infinite-Horizon Indefinite Linear Quadratic Stochastic Optimal Control Problems

Guangchen Wang, Heng Zhang

2026, 13(6): 1459-1469. doi: 10.1109/JAS.2026.125747

Abstract(16) HTML (11) PDF(0)

Abstract:
This paper is devoted to devising data-driven algorithms for finite-horizon and infinite-horizon linear quadratic stochastic optimal control (LQSOC) problems. In our study, the diffusion terms of system dynamics are permitted to hinge upon both control and state variables, and the weighting matrices of cost functionals are allowed to be indefinite. It is acknowledged that the optimal controls of finite-horizon and infinite-horizon indefinite LQSOC problems are correlated with a generalized differential Riccati equation (GDRE) and a generalized algebraic Riccati equation (GARE). Herein, we propose two data-driven algorithms to approximate the solutions of these Riccati equations, and thereby determine optimal controls, without leveraging the information of all system parameters. Additionally, we prove the convergence of these algorithms and examine the impact of computational errors. Finally, we validate the performance of these data-driven algorithms via three simulation examples.

Bimodal Predicate Refinement With Decoupled Entity-Predicate Representations for Scene Graph Generation

Guoqing Zhang, Shichao Kan, Wanru Xu, Yi Jin, Yidong Li, Yigang Cen

2026, 13(6): 1470-1491. doi: 10.1109/JAS.2026.125786

Abstract(21) HTML (10) PDF(0)

Abstract:
Scene graph generation is crucial for visual understanding, which can provide a structural foundation for image captioning, image comprehension, and visual question answering. However, datasets with significant long-tail distributions adversely impact model performance. Existing approaches have proposed various fusion and training strategies to address the long-tail distribution issue. Unfortunately, many strategies often lead to the “entanglement” of predicate and entity representations. This means that the representation space of predicates becomes confused with the representations of subjects and objects, resulting in suboptimal model performance. To tackle this challenge, we introduce a novel approach called bimodal predicate refinement with decoupled entity-predicate representations (BiRef). Specifically, we initialize predicate representations using Gaussian distribution and then progressively refine effective predicate representations in both visual and semantic spaces. Then, the refined predicate information is adaptively weighted and fused from the visual and semantic branches to obtain the predicate representation for the triplet. Extensive experiments conducted on the visual genome, GQA, and open images datasets demonstrated that our method achieved state-of-the-art scene graph generation performance and effectively mitigates the prediction bias problem associated with long-tail distributions. Our code has been made open source on https://github.com/gavin-gqzhang/BiRef.

Construction of Conflict-Free and Efficient Cross-Organization Emergency Response Processes: A Petri Net-Based Approach

Qi Mo, Shichao Wei, Yuhang Zuo, Chengting Jiang, Fei Dai, Cong Liu

2026, 13(6): 1492-1511. doi: 10.1109/JAS.2026.125873

Abstract(344) HTML (15) PDF(29)

Abstract:
Usually, the disposal of the emergency is organized as a cross-organization emergency response process (CERP), where various resources are involved. The lack of these resources may cause resource conflicts that can delay or even suspend the CERP, thereby increasing the risk imposed on life, property, and the environment. In this paper, we propose a novel approach to construct conflict-free and efficient CERPs. This approach first presents a branching place-based method to decompose a CERP into a set of execution paths. In essence, an execution path refers to a process fragment without choice structures corresponding to some kind of process instance in the CERP. In practice, each execution of the CERP can only follow such an execution path. Next, it determines whether each execution path contains resource conflicts. If not, then the execution path is considered conflict-free; otherwise, it will be resolved using a delay-based strategy. Lastly, it introduces an execution path-oriented strategy to merge all originally conflict-free and resolved execution paths to form a resolved CERP, in which each execution of it is conflict-free and efficient. The proposed approach is implemented in the tool RCTool, and a group of experiments conducted on actual CERPs demonstrates that it is more effective in constructing conflict-free and efficient CERPs compared to existing proposals, and its computation overhead is also acceptable in practice.