Keywords: Aerial Robotics, Robot Vision, Embodied AI
Abstract: Real-time UAV multi-vessel tracking plays a pivotal role in maritime safety and surveillance, but is critically impeded by dynamic maritime fog. Under such adverse conditions, conventional tracking pipelines suffer from severe feature collapse and temporal flickering. Furthermore, existing enhancement methods entail prohibitive latency unsuitable for edge deployment, while standard domain adaptation relies on vulnerable visual cues, overlooking robust semantic knowledge. To reconcile state-of-the-art accuracy with strict edge constraints, we propose a language-guided progressive inheritance learning framework, termed FogTrack. We first construct a physics-aware fog synthetic engine to generate homologous stochastic fog streams. Next, to realize this progressive inheritance, a three-stage feature-language isomorphic distillation is designed to progressively restore the representations of the visual student network. Since high-level semantics are significantly more resilient to fog degradation than low-level pixel features, we leverage robust vision-language (VL) representations as offline semantic anchors to perform textaligned semantic grounding, semantic-driven clarity alignment, and enforce trajectory continuity via a novel instance-level temporal consistency loss. Crucially, this inheritance paradigm successfully transfers degradation-invariant knowledge from large foundation models to resource-constrained edge devices. By internalizing this VL knowledge into the visual latent space during training, FogTrack deploys a fully decoupled, purely vision-based lightweight pipeline without multimodal computational overhead. Extensive experiments demonstrate that FogTrack achieves state-of-the-art tracking performance and robustness against dynamic maritime fog, while ensuring real-time UAV edge inference. The source code and new dataset are located at https://github.com/vision4robotics/FogTrack.

Keywords: Cybernetics Automation and Control, Motion Control, Deep Reinforcement Learning
Abstract: The application of optimization-based approaches to address state constraints plays a critical role in ensuring the flight safety of unmanned aerial vehicles (UAVs). However, predictive optimization strategies often incur substantial computational cost due to the inherent nonlinearity and strong coupling of UAV dynamics. This paper proposes an innovative event-triggered reference governor (ET-RG) designed for UAV control to enforce constraints within the optimal control framework. To balance computational overhead with control performance, a reinforcement learning (RL)-based event-triggering mechanism is introduced, enabling adaptive activation of the RG only when necessary. The effectiveness of this approach is validated through simulations performing a hovering control task, demonstrating that the proposed method achieves constraint-compliant tracking performance while substantially lowering the computational burden compared to conventional RG designs.

Keywords: Cybernetics Automation and Control, Systems Modeling & Control
Abstract: The air supply control of proton exchange membrane fuel cells (PEMFCs) is often adversely affected by model uncertainties. To this end, an approximation-free control method with prescribed performance is developed in this study. First, to quantify the convergence overshoot and steady-state range of the oxygen excess ratio (OER) tracking error, a flexible performance function is constructed, in which an auxiliary shifting function is incorporated to ensure satisfaction of the initialization-value-based feasibility condition (IFC). Following this, a low-complexity controller based on an equivalent error transformation is developed, which effectively avoids the use of intelligent approximators, parameter estimation, or command filtering. Theoretical analysis shows that the proposed method maintains the tracking error within the prescribed bounds, and comparative tests further verify its superiority.

Keywords: Embodied AI, Human-Robot Interfaces, Social Robotics
Abstract: Physical intelligence (PI)—the ability of robots to sense, reason, and act skillfully in unstructured environments—has recently advanced from laboratory research toward deployable systems. Science Centre Singapore (SCS) is harnessing this momentum to enrich public STEM education across all age groups, from primary-school students to adult learners. This paper describes two interactive PI demonstrations developed for RoboFest 2026 (9–12 April, SCS): (1) a Unitree G1 humanoid robot trained using Action Chunking with Transformers (ACT), benchmarked against the π0 and π0-FAST vision–language–action foundation models, to autonomously serve popcorn to visitors; and (2) a hands-on teleoperation station where visitors operate a leader–follower robot arm pair and experience first-hand how demonstration data for Vision-Language-Action (VLA) models is produced. Over four days, more than 30,000 visitors attended RoboFest; more than 3,000 experienced the popcorn robot demonstration at close range and over 5,000 tried the robot arm. A pre/post comprehension survey of 125 visitors across four age groups revealed a consistent gain of +21.6 percentage points in PI pipeline comprehension, rising from a near-zero baseline of 2.4% to 24% post-visit. These results demonstrate that carefully designed hands-on PI exhibits can effectively communicate cutting-edge robotics concepts to broad public audiences.

Keywords: Robot Vision, Embodied AI, Robotics and Automation Applications
Abstract: Robotic pick-and-place in cluttered tabletop settings usually demands task-specific perception models, curated datasets, or retraining whenever new objects appear. We describe a modular perception-to-action pipeline that sidesteps these requirements by combining open-source components---SAM2 for instance segmentation, CLIP for few-shot recognition, and GPT-4o for instruction grounding---so that a new object category can be onboarded with roughly ten reference images and no model retraining. Given a top-down RGB image and a natural-language command (typed or spoken), the system segments every object, matches each instance to a few-shot class prototype, and uses an agentic VLM workflow to resolve the command into pick-and-place targets. A fixed manipulation routine (approach--grasp--lift--transport--place) then executes the motion on a real Fanuc arm. We evaluate the perception module on 43 cluttered real-world scenes (333 annotated object instances) and report F1 scores of 96.3% (marker) and 97.4% (injector), substantially outperforming a Grounded SAM2 baseline on classification reliability. A quantitative grounding evaluation over 303 auto-generated commands yields 94.1% action accuracy. End-to-end demonstrations confirm that the pipeline can ground spoken commands into successful pick-and-place executions without task-specific training.

Keywords: Robotics and Automation Applications, Systems Modeling & Control, Neural Networks
Abstract: Cable-driven hyper-redundant robots (CDHRRs) enable flexible motion in complex environments, and their accurate control depends critically on precise modeling of the driving cables. However, practical cables exhibit inherently nonlinear coupled behaviors. Classical linear models and classification-based models are often insufficient to capture configuration-dependent force variations, cable-length errors induced by cable-guide disk intervals, and friction-induced tension distortion. These unmodeled effects hinder reliable segment-wise tension estimation, degrade control precision, and impair motion stability. To address these challenges, this paper develops a coupled geometric-mechanical modeling approach together with a hybrid controller for CDHRRs. The geometric module integrates elastic-deformation compensation and cable-guide disk modeling to determine contact locations and effective cable length. The mechanical module combines a quasi-static force model with a deep learning (DL)-based tension estimator. It provides a configuration-dependent tension baseline while compensating for friction-induced distortion. The resulting unified framework enables accurate estimation of cable length and segment-wise tension. Based on the proposed model, a hybrid tension-configuration controller is developed and experimentally validated. It achieves a 44.7% reduction in angular tracking error, a 53.4% improvement in configuration zeroing precision, and a 16.0% reduction in settling time. Additionally, the proposed model achieves a 20.9% reduction in mean root mean square error (RMSE) for cable tension prediction compared with the baseline, along with improved prediction consistency. These results demonstrate improved control accuracy and motion stability.

Keywords: Systems Modeling & Control, Motion Control, Intelligent Transportation Systems
Abstract: The performance of the designed controller is highly dependent on the model selection adopted in the design process. As a semi-empirical model featuring nonlinear characteristics and transient working states, UniTire is widely applied in tire dynamics research and vehicle simulation analysis. Adopting the semi-empirical paradigm based on theoretical physics combined with experimental data fitting, it strikes a balance among model accuracy, generality and engineering practicality. This study establishes an intelligent vehicle trajectory tracking control scheme by adopting the UniTire modeling approach. The control method adopts the theoretical framework of nonlinear model predictive control (NMPC). For the convenience of controller derivation and implementation, this research appropriately simplifies the complete UniTire formulation while preserving its core unified exponential characterization of tire dynamic behavior, thereby obtaining a simplified UniTire model suitable for control-oriented design. A joint simulation platform combining Carsim and MATLAB is adopted to validate the feasibility and overall performance of the developed control system.

Keywords: Decision Support Systems, Large Language Models, Neural Networks
Abstract: Complex confined built environments such as metro stations typically rely on pre-defined evacuation plans to guide pedestrians to exits. However, this approach relies heavily on manual calibration and is insensitive to diverse and unforeseen conditions encountered during real-world emergencies. Simulation-based methods, which can provide precise evacuation plans, are constrained by time-consuming simulation processes and computational delays, making them difficult to provide timely decisions under critical conditions. To address these challenges, this paper proposes a novel simulation-in-the-loop emergency evacuation framework based on large language models to generate evacuation plans in a real-time manner. Various evacuation scenarios and the associated optimal evacuation plans are encoded into semantic descriptions and subsequently used to fine-tune a language model for generalised evacuation decision-making. When an emergency occurs, the language model can infer an evacuation strategy from the ongoing environmental conditions without running additional simulations. Experimental results show a 94.99% match rate between the simulation-based results and those generated by the language model, with a BLEU-4 score of 96.05 indicating high textual consistency. These results demonstrate the substantial potential of the proposed method to accelerate emergency response.

Keywords: Image Processing, Neural Networks, Computational Intelligence
Abstract: Synthetic aperture radar (SAR) platforms generate massive volumes of raw complex echoes, while downlink bandwidth and onboard storage remain limited. Existing onboard compression schemes are still dominated by block adaptive quantization, which is efficient but often degrades image quality and phase fidelity at low bit rates or in high-contrast scenes. Recent learned image compression models improve rate-distortion performance for optical imagery, yet they are not designed for complex-valued SAR RAW data and usually break the amplitude-phase coupling of the echo signal. This paper presents a phase-preservation-oriented learned compression framework for SAR RAW data based on complex-valued neural networks. Built on a hyperprior and slice-based entropy model, the proposed method introduces three key designs: complex generalized divisive normalization for phase-safe feature normalization, amplitude-guided complex channel attention for selective feature enhancement, and complex window attention for local dependency modeling in the latent space. Experiments on operational Sentinel-1 Level-0 RAW products and an AID-based simulated dataset show that the proposed model consistently outperforms strong learned baselines, including TCM and DCAE, in RAW-domain SQNR, focused-domain PSNR, and focused-domain phase MAE. On Sentinel-1, our method achieves BD-rate reductions of 14.10% in SQNR and 15.49% in PSNR relative to the ELIC anchor, while reducing phase MAE most substantially and keeping decoding time close to lightweight learned baselines. These results indicate that complex-domain modeling is a practical direction for high-fidelity SAR RAW compression.

Keywords: Optimization, Computational Intelligence, Discrete Event Systems
Abstract: Feature selection is essential for improving predictive performance and reducing redundancy in high-dimensional data. However, conventional filter methods ignore feature interactions, while swarm intelligence approaches often suffer from redundant selections and inefficient search. We propose a redundancy-aware graph-clustering-guided framework for swarm-intelligence-based feature selection. A feature graph is constructed to model pairwise dependencies, and community detection is employed to group highly correlated features into clusters. Building on this representation, a redundancy-aware selection mechanism is introduced to penalize inter-feature redundancy while promoting diversity during optimization. The framework is optimizer-agnostic and can be integrated into different swarm intelligence algorithms. We instantiate it using particle swarm optimization (PSO), differential evolution (DE), and grey wolf optimization (GWO), yielding three variants: GCPSO, GCDE, and GCGWO. Extensive experiments on benchmark datasets demonstrate that the proposed framework consistently improves classification accuracy while producing more compact feature subsets, outperforming both the original algorithms and classical methods, including mRMR, ReliefF, FScore, IG, and CIFE.

Keywords: Aerial Robotics, Planning and Control, Robot Vision
Abstract: Autonomous navigation of low-altitude unmanned aerial vehicles (UAVs) in cluttered environments is challenging due to partial observability, limited onboard perception, and inefficient exploration in end-to-end reinforcement learning. This paper proposes a teacher-guided asymmetric reinforcement learning framework for end-to-end visual navigation of low-altitude UAVs in the Isaac Sim 5.1 environment. A privileged teacher policy is first trained using obstacle-state information to acquire reliable navigation priors. A deployable student policy is then learned with an asymmetric actor-critic architecture, where the actor takes depth images and proprioceptive states as input, while the critic uses privileged information during training. To improve policy transfer, an annealed knowledge distillation strategy is adopted: the student is strongly guided by the teacher in the early stage, and the guidance is gradually removed to enable autonomous reinforcement refinement. Experimental results show that the proposed method achieves higher success rates, lower collision rates, and faster convergence than baseline methods.

Keywords: Deep Learning, Image Processing
Abstract: 半监督语义分割中的关键挑战 （SSSS）正在有效利用未标记的数据进行改进 模型以有限的注释训练。然而，存在 方法主要依赖视觉线索，限制了模型 对于数据的外观中心视图，当 忽略几何信息，导致次优结果 性能，特别是在物体边界上。 基于这一见解，我们提出了一个具几何感知能力的方案 半监督式知识蒸馏框架 转移的语义分割（{GeoKDS^4}） 几何先验：从几何教师模型到 以语义为导向的学生模型。该过程面临两个 挑战：（1） 它们之间的架构差异导致 对于错位表示和（2）来自 教师在复杂体中几何建模不足 区域可能会干扰蒸馏过程。应对 这些问题，我们引入两个组成部分：A 相关引导几何知识蒸馏（CGKD） 通过以下方式来减少表示差异的模块 显

Keywords: Deep Learning, Localization & Tracking, Image Processing
Abstract: Localizing the active speaker by jointly leveraging audio and visual information is important for applications such as human-computer interaction, intelligent meeting analysis, and visual assistance for people with hearing impairments. This paper focuses on a practical downstream setting in a pipeline, where an approximate sound-source azimuth is assumed to be available and image information is further used to localize the active speaker. Under this setting, we propose a lightweight explicitfusion method that encodes the sound-source azimuth as a spatial mask aligned with the RGB image and feeds it into the detector as a fourth input channel. In this way, the model can directly exploit the directional prior for speaker localization. Built upon YOLO26n, the proposed method achieves 0.9935 mAP@.5, 0.7949 mAP@.5:.95, and 0.9746 F1, while remaining suitable for deployment on edge devices. Further analysis shows that, although the input is extended to four channels, transferring pretrained weights from the original three RGB channels remains crucial or performance. Meanwhile, different guidance formulations and bandwidth settings also lead to clear performance differences. When the primary metric mAP@.5 is used for model selection, the final selected configuration is achieved with a Gaussian guidance mask of angular bandwidth 2σ = 2◦.

Keywords: Deep Learning, Robot Vision
Abstract: Transformer-based person re-identification (ReID) performs well on standard benchmarks but remains vulnerable in crowded fine-grained scenes, where visually similar identities are easily confused and occluded patch tokens become unreliable. This problem is particularly severe in basketball player ReID due to near-identical uniforms and frequent body overlap. We propose CAR-TransReID, a lightweight extension of TransReID with two plug-in components. Adaptive Confusion-Aware Pair Mining (ACAPM) builds a dynamic identity confusion bank and applies adaptive-margin hard-negative supervision, while Reliability-Guided Residual Token Aggregation (RGRTA) estimates token reliability from token-global consistency and enhances the global feature through residual fusion. Experiments on BallShow, Market-1501, and DukeMTMC-reID validate the proposed design. CAR-TransReID improves TransReID from 93.6/91.1 to 95.2/92.3 in Rank-1/mAP on BallShow, and achieves 95.7/89.5 on Market-1501 and 90.8/82.6 on DukeMTMC-reID. ACAPM introduces no inference overhead, and RGRTA adds only 0.89M parameters and 1.43% latency.

Keywords: Deep Reinforcement Learning, Deep Learning, Neural Networks
Abstract: Accurate delineation of nasopharyngeal carcinoma (NPC) lesions is crucial for radiotherapy planning, but remains challenging in clinical MRI due to irregular tumor morphology, indistinct boundaries, and complex surrounding anatomy. Existing CNN-based and CNN–Transformer segmentation models improve local and global feature representation, but decoderside reconstruction may still suffer from background interference and unstable lesion boundaries. In this paper, we propose EMATransUNet, a 3D TransUNet-based segmentation model enhanced with a decoder-side Enhanced Multi-Scale Attention (EMA) module at the third and fourth decoder stages. EMA projects grouped 3D features into a 2D in-plane representation and broadcasts the learned attention cues back to the 3D feature space, thereby strengthening lesion-related responses while suppressing redundant background interference. Experiments on a private clinical NPC MRI dataset show that EMA-TransUNet achieves a DSC of 74.64%, the highest Recall of 77.24%, and the lowest HD95 of 5.51 mm among the compared methods. Compared with 3D TransUNet, EMA-TransUNet reduces HD95 by 0.92 mm while introducing only 0.013M additional parameters and 0.294G additional FLOPs boundaries in challenging cases.

Keywords: Robot Vision, Embodied AI, Deep Learning
Abstract: Low-light garment grasping remains challenging for home service robots because strong foreground-background coupling degrades garment segmentation and grasp localization. Infrared-based multimodal perception can alleviate photometric degradation, but explicit cross-modal alignment often introduces additional computational overhead. In this paper, we propose DFBNet, a lightweight encoder--decoder segmentation network for low-light robotic garment grasping. In the encoder, the Cross-modality Transfer (CT) module selectively enhances inter-modal complementary features through discrepancy-aware gating. In the decoder, the Mask-Guided Attention (MGA) module decouples foreground and background feature streams using foreground/background probability masks. Based on the segmentation output, we further introduce a grasp-annotation-free multi-scale ridge detection strategy for grasp point localization from depth images. We also contribute DFBClothes, a low-light multimodal garment dataset with pixel-level annotations. Deployed on a Baxter robotic platform, DFBNet achieves 91.31% mIoU, an 82.50% grasp success rate, and 32.27 FPS on an RTX 4090 with only 18.61M parameters. Experiments show that our method achieves the best grasp success rate among the compared methods while providing a favorable tradeoff among segmentation accuracy, efficiency, and grasping performance.

Keywords: Robot Vision, Biologically-Inspired Robots and Systems, Embodied AI
Abstract: Event cameras can capture human motion with low latency and high dynamic range, providing an emerging sensing paradigm for action recognition. A large body of recent work converts event data into a sequence of dense frames and feeds them into a video classification model for prediction. Although the frame-based methods take advantage of pre-trained backbones from the image domain to achieve high accuracy, most existing works sacrifice the sparsity of events due to dense processing, thereby increasing model complexity and latency. To boost the efficiency of frame-based methods while fitting the sparse nature of events, we propose the Sparsity-Aware Vision Transformer Network (SVTNet) with two novel event-oriented designs. Instead of direct event-to-frame conversion, we introduce an event representation method named Progressive Cumulative Event Representation (PCER) that adaptively integrates spatiotemporal cues across multiple temporal slices to better preserve fine-grained information. To leverage the spatial sparsity of events to accelerate model inference, we present the Motion Prior-Guided Token Sparsification Module (MPTS) that prunes redundant visual tokens hierarchically based on the joint motion-semantic importance. Comprehensive experiments show that SVTNet achieves state-of-the-art accuracy on multiple benchmark datasets and maintains much lower model complexity than existing frame-based methods.

Keywords: Social Robotics, Modeling
Abstract: A Stackelberg-based dual-policy learning framework is proposed for social robot navigation in complex, dynamic environments. In such settings, robots must balance multiple, often conflicting objectives, including safety, efficiency, and social compliance. Conventional multi-objective reinforcement learning approaches typically rely on manually tuned reward weights, which are often insufficient to capture these trade-offs and may lead to unstable or suboptimal behaviors.

To address this challenge, the multi-objective decision-making problem is formulated as a hierarchical Stackelberg game, where different objectives are modeled as interacting components with explicit priority relationships. Within this framework, the policy is decomposed into a leader–follower structure: the leader policy enforces safety by regulating the action scale (i.e., velocity magnitude), while the follower policy determines the motion direction to balance goal-directed efficiency, human-aware interaction, and obstacle avoidance.

The framework is trained using the Soft Actor-Critic (SAC) algorithm, with both policies optimized in an end-to-end manner. The follower policy generates normalized directional weights through a softmax layer, enabling a structured combination of multiple behavioral objectives. This design decouples motion magnitude from direction selection, leading to improved learning stability and enhanced policy interpretability.

The proposed method is evaluated in a Webots-based simulation environment with dynamic pedestrians and complex obstacle configurations. Experimental results demonstrate superior performance over both classical and learning-based baselines, including ORCA, CADRL, LSTM-RL, SAC, and A2C, in terms of success and failure rates. In addition, the generated trajectories are smoother, the navigation behavior is more socially compliant, and the policy exhibits stronger robustness in highly dynamic multi-agent scenarios.

Overall, the framework provides an effective and principled solution for multi-objective decision-making in social robot navigation, and highlights the potential of integrating game-theoretic reasoning with reinforcement learning.

Keywords: Systems Modeling & Control, System of Systems
Abstract: 基于激光雷达的定位常因几何退化环境下因配准不良和可观测性丧失而失败。现有的被动缓解策略或基于视觉的规划方法对激光雷达来说要么不足，要么计算量过大。本文提出了一种退化感知路径规划框架，主动避免此类区域。我们首先通过对黑森矩阵条件数的谱分析推导可量化的退化因子，以表征局部可观测性。其次，我们引入了紧凑的二维风险图，将方向性风险剖面编码为64位整数，实现常时查询且无在线模拟开销。最后，通过惩罚混合A*成本函数内的简化区域，规划者主动引导机器人走向几何约束丰富的路径，显著增强了在复杂环境中的导航稳健性。

Keywords: Cooperative Systems and Control, Optimization, Cyber-physical Systems
Abstract: This abstract explores the potential to reduce the whole fuel consumption by jointly optimizing speed dynamic for a large population of connected and automated vehicles(CAVs). To avoid the computation burden of the large population of vehicles, a novel decentralized control scheme is designed by employing the mean field game (MFG) theory. Thanks to the communication by vehicle to everything (V2X) that the real-time velocity distribution of the vehicles is available, the receding horizon sense is introduced to MFG to avoid the open-loop prediction error of velocity distribution within the optimization horizon. The solution is derived from a novel numerical method for the general nonlinear MFG problem. A virtual connected traffic simulation platform is built in MATLAB/Simulink, and simulations are conducted to show the effectiveness of the proposed strategy.

Keywords: Robotics and Automation in Unstructured Environment, Deep Reinforcement Learning, Intelligent Transportation Systems
Abstract: Addressing the challenges of autonomous mobility and attitude maintenance for lunar rovers in low-gravity environments, this paper proposes an end-to-end reinforcement learning-based jumping control method for a parallel four- wheel-legged rover. The method eliminates dependency on terrain elevation maps, relying solely on proprioceptive sensing to execute jump commands. Considering the strong coupling of the five-bar parallel mechanism, an impedance-control-based action space is designed for safety. A phase-aware reward scheduling mechanism, utilizing contact force and acceleration cues, dynamically activates corresponding reward terms to resolve dynamics discrepancies across takeoff, flight, and landing phases. Specifically, a force-balanced leveling reward is proposed to minimize the variance of vertical contact forces, enabling implicit attitude self-leveling on uneven terrain without elevation maps. Simulation results demonstrate that the proposed strategy enables stable jump execution and rapid posture recovery on unknown rough terrain under lunar gravity.

Keywords: Motion Control
Abstract: To improve the trajectory tracking accuracy of autonomous agricultural vehicles during practical operations, this paper proposes an extended state observer (ESO) enhanced decoupled non-smooth integral (NI) control scheme. First, to address the coupling between longitudinal and lateral deviations that can degrade control precision, a trajectory deviation-based hierarchical model is established, enabling independent control of longitudinal and lateral motions without mutual influence. Second, the NI controller is designed to enhance the tracking accuracy of both longitudinal and lateral deviations under the hierarchical framework. Finally, an extended state observer is incorporated to effectively compensate for external disturbances. The stability of the proposed control scheme is rigorously demonstrated using Lyapunov analysis, and simulation results confirm the effectiveness and superior performance.

Keywords: Multi-Agent Systems, Optimization, Energy Efficiency
Abstract: This paper investigates a distributed aggregative optimization problem for multi-agent systems subject to communication and set constraints. A distributed optimization algorithm is developed, in which each agent updates its decision using locally available information and limited interactions with neighboring agents. To reduce communication burden, an event-triggered communication scheme is introduced such that information exchange occurs only when predefined triggering conditions are satisfied. As a result, unnecessary transmissions are avoided and communication efficiency is improved. In addition, a projection operator is incorporated to ensure that all decision variables remain within the feasible set throughout the optimization process. Furthermore, the convergence of the proposed algorithm is rigorously analyzed. Finally, a case study on a networked multi-electrolyzer system is presented to demonstrate the effectiveness of the proposed method for coordinated energy management.