Keywords: Autonomous Systems, Model-Based Design
Abstract: In this paper, we address the control of speed-sensorless induction motors subjected to load torque and distortion voltage, both of which are uncertain and unstructured. The problem is considered under the following assumptions: (i) the rotor angle and stator currents are measurable, and (ii) all motor parameters are known. A novel nonlinear disturbance observer (DOB)-based control scheme is developed to achieve complete cancellation of these external perturbations while enhancing transient performance. We first design the controller within a direct rotor field-oriented control (RFOC) framework, assuming the availability of rotor flux. This assumption is then relaxed by adopting an indirect RFOC approach, in which the proposed controller is combined with a flux estimator. Both designs are validated through numerical experiments, with stability guarantees provided.

Keywords: Autonomous Systems, Human-Machine Teaming, Human Machine Systems
Abstract: This paper presents an adaptive shared control path-following law that corrects deviations from a desired path when a craft is manually steered. A reference model generates ideal motion, which is aligned with the actual trajectory to compute real-time corrections of lateral deviations. The method preserves longitudinal disturbances to maintain operator awareness, while enhancing stability and intuitive control. The stability properties relies on the underlying motion control system, however it is locally asymptotically stable for a range of controllers and calculations of the path. We emphasize the benefits on an experimental case study of a remotely operated fully actuated ferry by demonstrating how the reference model may be chosen to achieve both course-stability and fixed heading support in different modes of operations. The method optimize operator intuitive steering with a seamless path corrections and support for station-keeping, evaluated from both expert and non-expert operators.

Keywords: Assistive Technology and Rehabilitation Engineering
Abstract: Series elastic actuators (SEAs), owing to their inherent compliance and safety, have been widely adopted in rehabilitation robots. However, friction and external disturbances can severely degrade trajectory tracking accuracy and even compromise the stability of human-robot interaction. To address this issue, this paper proposes an adaptive fuzzy control method with friction compensation. In the proposed scheme, a fuzzy logic system is employed to online approximate the effects of friction and disturbances, while a robust term is incorporated to suppress approximation errors. Within the Lyapunov framework, the stability of the closed-loop system is rigorously established. Simulation results demonstrate that the proposed controller enables rapid convergence of joint tracking errors and ensures smooth and stable control inputs under the influence of friction and disturbances, validating its effectiveness and robustness.

Keywords: Human-Computer Interaction, Assistive Technology and Rehabilitation Engineering, Virtual and Augmented Reality
Abstract: A compliant two-degree-of-freedom (2-DOF) upper limb rehabilitation robot was developed, offering safe and flexible motion control suitable for active patient participation in rehabilitation training. Building upon this hardware, a game-based interactive training system was implemented to guide patients through virtual tasks, while simultaneously collecting multimodal data including motion trajectories, electromyographic (EMG) signals, and functional near-infrared spectroscopy (fNIRS) measurements. Comprehensive analysis of kinematic parameters such as velocity and acceleration, muscle activation patterns, and cortical functional connectivity was performed alongside clinical assessments using the Fugl-Meyer scale. The intervention demonstrated significant improvements in motor performance, neuromuscular coordination, and cortical functional connectivity. Patients exhibited enhanced task execution, improved muscle control, and strengthened brain functional networks related to motor processing, accompanied by notable increases in Fugl-Meyer scores. These findings indicate that the system effectively facilitates upper-limb motor recovery, neuromuscular integration, and brain network plasticity. The platform shows great promise for safe and adaptive rehabilitation therapy, with potential for broad clinical application.

Keywords: Human-Machine Interfaces, Human Machine Systems, Machine Learning
Abstract: A major challenge for robust motor imagery electroencephalogram (MI-EEG) decoding is posed by nonstationarity. Current cross-subject/session algorithms are primarily focused on feature engineering rather than data preprocessing, while variations of user's mental and physical state is often undefined. An inter-state difference pattern subtraction framework was proposed to improve MI-EEG decoding under user state transitions. Three distinct user states were described and a pattern subtraction method based on common spatial pattern (CSP) was developed to alleviate state-related nonstationarity. The regularized CSP followed by support vector machine algorithm (RCSP-SVM) was then applied for feature extraction and classification. Experiments were conducted on four participants' MI-EEG data, and significant improvement in transfer learning performance was achieved by the proposed method.

Keywords: Human Performance;Assistive Technology and Rehabilitation Engineering;Model-Based Design
Abstract: This paper proposes an algorithm for the rehabilitation analysis of hemiplegic patients, leveraging musculoskeletal modeling to assess key aspects of motor function. By constructing a personalized musculoskeletal model, the algorithm analyzes joint torques, muscle activation patterns, and movement coordination, allowing for the identification of issues such as muscle atrophy, gait abnormalities, and impaired motor control. Experimental results show significant reductions in the range of motion (ROM), joint torque, and muscle activation in the affected limbs of the patients. These findings highlight the impact of central nervous system injury on motor function and demonstrate the need for individualized rehabilitation strategies. The proposed method offers a more precise approach to rehabilitation assessment, with potential to improve treatment efficacy and optimize recovery outcomes for hemiplegic patients.

Keywords: Adaptive Assistance;Design, Analysis, and Evaluation;Model-Based Design
Abstract: This paper investigates fixed-time leaderless consensus control for multi-agent systems (MAS) modeled by heat conduction partial differential equations (PDE). A distributed control strategy using adaptive neural networks integrates spatiotemporal dynamics to ensure fixed-time convergence of consensus errors. By constructing PDE-adapted Lyapunov functions and applying fixed-time stability theory, control laws and neural network weight update algorithms are derived. Theoretical analysis rigorously proves that system states achieve MAS consensus within a predefined time. Simulation results validate the scheme’s robustness against nonlinear dynamics and coupling uncertainties, enabling high-precision fixed-time consensus.

Keywords: Adaptive Aiding, Model-Based Design, Design, Analysis, and Evaluation
Abstract: This paper investigates containment control for distributed parameter multi-agent systems (DP-MASs) under system uncertainties, limited non-collocated measurements, and observer faults. An adaptive fuzzy point observer is designed to estimate unmeasurable states and compensate for faults. Meanwhile, a Takagi-Sugeno (T-S) fuzzy model approximates system uncertainties, based on which a piecewise fuzzy containment control strategy is constructed. Subsequently, Lyapunov-based analysis guarantees the stability and convergence of the closed-loop system. Ultimately, numerical results verify the effectiveness and robustness of the method.

Keywords: Human Machine Systems, Adaptive Aiding
Abstract: This article proposes an active fault-tolerant control strategy for a two-degree-of freedom (2-DOF) helicopter system subject to sensor gain faults. First, to address the issue of inaccurate state measurements caused by sensor faults, a state observer is designed to reconstruct the system states, and an adaptive parameter is introduced to estimate the fault in real time. Then, the uncertainties in the system are estimated and compensated by employing a radial basis function neural network. Subsequently, the stability of the closed-loop system is proven based on Lyapunov stability theory. Finally, MATLAB simulation results are presented to validate the effectiveness of the proposed method.

Keywords: Autonomous Systems, Decision Support Systems
Abstract: Allocating tasks reasonably is the precondition for a cluster of multiple Unmanned Surface Vessels (USVs) to cooperatively complete tasks. In task allocation for multi-USV, the constraint condition of turning and evaluation indicator are designed, and a Clustering and Optimization Algorithm based on Ant Colony Optimization (COA-ACO) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is proposed. Simulation experiments show that under the search angle constraint, turning angle of multiple USVs could be effectively reduced, and the collisions of their preset paths could be avoided simultaneously. Consequently, the proposed method provides a more effective allocation solution which makes the USV turning smoother in the meanwhile.

Keywords: Design, Analysis, and Evaluation, Assistive Technology and Rehabilitation Engineering
Abstract: In response to the treatment accessibility challenges faced by constipation patients and the current global shortage of intelligent abdominal massage devices, this study developed a control system for an abdominal massage robot. We developed a modular control system based on the series-parallel hybrid structure of the robot, implementing both software and hardware designs. An efficient mobile phone interface was also designed. Rooted in traditional Chinese medicine theory, the system incorporates massage trajectory tracking functionality, supports multiple massage techniques, and includes a safety module. Experimental testing demonstrates that the control system functions stably and effectively achieves the expected control outcomes, validating its practical feasibility and reliability.

Keywords: Assistive Technology and Rehabilitation Engineering, Human Machine Systems, Design, Analysis, and Evaluation
Abstract: Robotic exoskeletons are widely used in medical rehabilitation, industry, and daily activity assistance. However, current exoskeleton design pipelines are often tailored to specific application, which limit the generalizability and scalability of design knowledge, making it more difficult for newcomers to this field. To address this issue, this study proposes a systematic and generalizable design guideline for powered lower-limb exoskeletons based on biomechanical and energetic analyses of human gait. The analysis highlights the sagittal plane, terminal stance, and pre-swing phases as the key directions and timing for effective mechanical assistance. The ankle, particularly during plantarflexion, plays the most critical role in delivering torque during walking and thus should be prioritized when designing exoskeletons for energy-efficient assistance. Based on this principle, we present a series of previously developed soft, cable-driven ankle exoskeletons designed accordingly. The study outlines the design pipeline and discusses the benefits of using cable-based actuation. Actuator requirements were derived from gait data, and cable transmission and friction losses were analyzed to enhance efficiency. Overall, this work contributes a structured framework for exoskeleton joint selection and system design, offering practical guidance for developers and promoting broader innovation in wearable robotics.

Keywords: Autonomous Systems
Abstract: Multimodal sentiment analysis (MSA) is essential for natural human–machine interaction, enabling robots to perceive users’ emotions through text, audio, and visual cues. However, achieving precise cross-modal alignment while preserving modality-specific signals remains challenging. We propose a Text-Dominant Disentangled Fusion Network (TDFN), which establishes text as the dominant modality to guide cross-modal alignment, while disentangling shared and specific features with orthogonality constraints and auxiliary supervision. Experiments on MOSI, MOSEI, and CH-SIMS demonstrate superior performance, highlighting TDFN’s potential to enhance emotional intelligence in human–robot interaction.

Keywords: Assistive Technology and Rehabilitation Engineering, Adaptive Aiding
Abstract: Traditional exoskeleton research, primarily focused on reducing metabolic cost on level ground, often overlooks the critical role of localized muscle fatigue in incline walking tasks. This study proposes using muscle fatigue as a primary metric to guide the selection of an optimal assisted joint (knee or hip). In a preliminary case study, a participant underwent a 13-minute high-intensity incline walking protocol (12°, 1.2 m/s) designed to induce fatigue. We used surface electromyography (sEMG) to monitor eight key lower limb muscles, quantifying fatigue by analyzing the temporal changes in mean frequency (MF) and median frequency (MDF). The results clearly identified the quadriceps femoris as the first and most significantly fatigued muscle group, strongly suggesting the knee is the potential optimal target for assistance. Furthermore, we observed distinct muscle fatigue patterns—a "fluctuating decline" in the quadriceps versus a "continuous decline" in the hamstrings—which implies that future exoskeletons could benefit from dynamic, adaptive assistance strategies that recognize these different physiological states.

Keywords: Machine Learning
Abstract: Facial Expression Recognition (FER) is challenging due to the subtle and ambiguous nature of human expressions. To address this issue, we propose RA-FER, a retrieval-augmented framework that enhances representation learning by leveraging semantically similar images from a retrieval gallery. Specifically, RA-FER employs a dual cross-attention mechanism to integrate complementary information from both the query and the retrieved samples, while a router adaptively selects the most informative features for expression prediction. Extensive experiments on three real-world benchmarks demonstrate the robustness and superior performance of RA-FER.

Keywords: Assistive Technology and Rehabilitation Engineering, Adaptive Aiding, Human-Computer Interaction
Abstract: This paper presents a human gait symmetry estimation algorithm based on multi-sensor fusion, which is applied to the motion control scenario of assist-as-needed mobile walking assistive robots. It addresses the limitations of traditional single-sensor gait symmetry calculation, significantly improving the stability and accuracy of gait symmetry calculation. By fusing the foot movement data collected by the Inertial Measurement Unit (IMU) and the foot movement data obtained by the RealSense D435 depth camera based on YOLOv5, an ellipsoidal outer-bounding set-membership filtering is used to achieve precise estimation of foot movement states. On this basis, a human gait symmetry algorithm based on foot movement states is designed. Through a multiple data fusion mechanism, the stability and reliability of the collected foot movement data are further enhanced, thereby enabling the calculation of the Human Gait Symmetry Index (GSI). The experimental results show that compared with the single-sensor scheme, this algorithm has significantly optimized the stability and accuracy indicators, providing an innovative technical pathway for the motion control strategy of intelligent walking robots.

Keywords: Machine Learning, Autonomous Systems, Model-Based Design
Abstract: Visual Place Recognition (VPR) has evolved from handcrafted descriptors to deep learning approaches, yet significant challenges remain. Current approaches, including Vision Foundation Models (VFMs) and Multimodal Large Language Models (MLLMs), enhance semantic understanding but suffer from high computational overhead and limited cross-domain transferability when fine-tuned. To address these limitations, we propose a novel zero-shot framework employing Test-Time Scaling (TTS) that leverages MLLMs' vision-language alignment capabilities through Guidance-based methods for direct similarity scoring. Our approach eliminates two-stage processing by employing structured prompts that generate length-controllable JSON outputs. The TTS framework with Uncertainty-Aware Self-Consistency (UASC) enables real-time adaptation without additional training costs, achieving superior generalization across diverse environments. Experimental results demonstrate significant improvements in cross-domain VPR performance with up to 210× computational efficiency gains.

Keywords: Machine Learning, Human Machine Systems
Abstract: Object re-identification aims to match images or videos captured by different cameras. Its objective is to identify the same object when provided with a query image from a collection of images recorded by diverse surveillance devices. Object re-identification tasks still face huge challenges, such as low resolution of images taken in complex scenes, occlusion phenomena in images, changes in lighting, and changes in perspective and posture, especially for 3D reconstruction tasks during distribution network operations. The recognition accuracy of object re-identification in complex scenarios needs to be further improved. Previous work did not pay attention to the impact of multi-scale feature interaction. This article proposes a novel multi-scale feature interaction and perception with transformer method (MSFIP-TM) for object re-identification. It designs a multi-scale perception module to extract multi-scale features, which have the advantage of expanding the receptive field and facilitating the capture of small differences in different feature maps, thereby improving the robustness of object re-identification methods in complex scenes. Superior experimental performance is achieved on two large-scale Market-1501 and DukeMTMC-ReID datasets.

Keywords: Machine Learning, Supervisory Control
Abstract: The Transferability of machine learning algorithms is important for various security applications. An image contains richer information than plain text, which plays more and more important roles in visually based applications. However, it often requires large number of labeled images for reliable model training, and is less efficient. On the contrary, Zero-shot image classification tries to accurately predict the classes of unseen target images by learning from labeled source images. Traditional zero-shot image classification algorithms model images in relatively independent manner, and do not fully explore semantic correlations of images. To solve this problem, in this paper, we propose an zero-shot image classification (ZSIC) method by modeling implicit semantic correlation Transferability (ISCT) for. We first obtain implicit semantic representations of images by training classifiers with labeled source images. Instead of using implicit semantic representations for direct classification, we explore the implicit semantic correlations between images for prediction. This is achieved by dividing the line between two images and then obtaining the implicit semantic representation of each segment. The implicit semantic representations of segments are concatenated in a matrix form for each two images. The matrix representation can be seamlessly combined with various well designed networks for zero-shot image classification. We evaluate the performance of ISCT on four public datasets. Experimental results prove the effectiveness of the proposed method.

Keywords: Machine Learning
Abstract: Visible image stitching for power equipment monitoring is hindered by low-texture regions, cross-modal inconsistencies, and illumination variations. This paper proposes a cascaded framework using a truncated VGG16 network to extract robust features, combined with saliencyweighted, region-aware keypoint sampling. A hierarchical matching strategy (adaptive NNDR + RANSAC) refines correspondences. Experiments on a power equipment dataset show the method outperforms traditional techniques (SIFT, ORB) in keypoint count, matching accuracy, and stitching quality, effectively addressing multimodal challenges for panoramic monitoring

Keywords: Human-Computer Interaction, Human-Machine Interfaces, Machine Learning
Abstract: Multimodal large language models (MLLMs) in the biomedical domain have shown superior performance over general-domain models when processing complex and fine-grained biomedical images. However, existing biomedical MLLMs are typically limited to single-modal tasks, and their performance degrades significantly on multi-modal tasks due to modality discrepancies. To address the drop in performance caused by cross-modal gaps, we propose Granular Fusion LLaVA (GF-LLaVA), a novel framework that enhances visual comprehension across diverse modalities by integrating additional visual encoders and fusing multi-granularity visual features. Experimental results on Med-VQA datasets demonstrate that GF-LLaVA consistently outperforms previous state-of-the-art models across multiple metrics, highlighting its superior performance for multi-modal medical tasks.

Keywords: Usability Engineering, Machine Learning
Abstract: Nucleic acid testing (NAT) technology is the ‘gold standard’ for diagnosing infectious diseases. However, traditional NAT is limited in its application in scenarios requiring rapid testing due to its cumbersome process, high personnel and environmental requirements, and lengthy time consumption. This paper integrates microfluidics and robotics technology to develop a fully automated multiplex nucleic acid rapid testing system-iNAT. Using microfluidic technology, we designed a microfluidic chip to load nucleic acid testing reagents and samples, and complete biochemical reactions. Using robotic technology, a nucleic acid rapid testing robot was developed, including mechanical design, electrical systems, control modules, and intelligent algorithms. Through a closed-loop control strategy involving multiple modules, the system achieves fully automated testing from sample processing to result analysis, with the entire process taking only 40 minutes. Experimental results indicate that the system can accurately perform quantitative detection of samples, with a linear fitting goodness of 0.9999 for gradient sample detection results. Additionally, it can effectively detect multi-target samples, demonstrating strong potential for clinical testing. In summary, these results indicate that the iNAT system we have developed breaks through the limitations of traditional nucleic acid testing, offering rapid, precise, and fully automated advantages to provide efficient solutions for various scenarios. It is expected to drive the development of molecular diagnostics toward greater convenience and intelligence, showcasing broad clinical application prospects and market potential.

Keywords: Design, Analysis, and Evaluation, Model-Based Design
Abstract: Compared with cylindrical designs, rectangular soft robots provide larger workspace and more stable contact in confined environments. However, their practical application is limited by the inherent low stiffness and the complex influence of embedded cables on kinematic modeling. This paper presents a constant curvature kinematic model for cable-driven rectangular soft robots that accurately predicts the bending deformation. The effect of cable routing paths on robot stiffness is systematically analyzed, deriving analytical expressions between routing parameters and stiffness. Results show that parallel routing yields minimum stiffness, while non-parallel paths significantly enhance stiffness while maintaining necessary compliance. The proposed kinematic model is validated through MATLAB implementation, FEM simulation, and physical experiments, demonstrating consistency within a 45° bending range. Impact tests further confirm the stiffness enhancement of different cable routings, verifying the effectiveness of the analysis method.

Keywords: Model-Based Design, Human Machine Systems, Human-Machine Interfaces
Abstract: Understanding natural human behavior is essential for designing effective and well-perceived automation in physical human–robot interaction (pHRI). While model-based control strategies are increasingly applied in assistive systems, most current approaches assume humans behave deterministically, which is contradicting evidence from neuroscience that highlights the stochastic nature of human motor control. This paper presents a user study with 21 participants performing goal-directed locomotion while physically pushing a smart wheelchair. By analyzing unconstrained human-only trials, we focus on characterizing human inherent variability in the context of physical coupling. Our results reveal structured patterns of task-relevant and task-irrelevant variability across repetitions, suggesting that variability is not random but systematically shaped by the task. These findings offer important insights for future shared control systems that aim to accommodate, rather than disregard, human movement variability.

Keywords: Human-Machine Interfaces, Human Machine Systems, Machine Learning
Abstract: This paper investigates the prediction of critical diabetes-related events using brain–computer interface systems. It reviews recent advances, outlines measurement techniques, and examines correlations between glycemia and brain activity. Preliminary results indicate the feasibility of developing predictive models for early event detection. Furthermore, the study highlights the potential of non-invasive approaches for estimating blood glucose levels, offering a foundation for future clinical applications.

Keywords: Human-Machine Teaming, Machine Learning, Human-Computer Interaction
Abstract: Continuous prediction of joint angles of human is of significant value for rehabilitation training, nursing, and human-robot interaction (HRI), however, the prediction accuracy and real-time performances are far from satisfactory. This study proposes a continuous prediction method using Mamba and fuzzy entropy-based multi feature fusion of surface electromyography (sEMG) to improve prediction performances. The core advantage of Mamba lies in its powerful information integration capability and operational efficiency. Mamba selectively fuses all historical information through its selective state space model (SSSM) mechanism, effectively filtering redundancy and accurately updating predictions for the current time step. Secondly, Mamba abandons inefficient sequence recursion and adopts efficient matrix operations for computation, thereby breaking through the limitations of step-by-step training and achieving parallel processing across time steps. Fuzzy entropy (FE), a new indicator to measure the regularity of time series, can effectively capture the nonlinear characteristics of sEMG. The integration of FE with conventional features enhances the representational capacity of sEMG signals. By inputting FE features into the Mamba model for training, the model can better capture the changes in sEMG signals under different motion states, thereby significantly improving the ability to recognize human motion intentions in sequence data. Experimental results show that the proposed prediction method has significant advantages in prediction accuracy and computational efficiency over existing research results based on recurrent neural network (RNN), long short-term memory network (LSTM), and temporal convolutional network (TCN).

Keywords: Machine Learning, Human-Machine Interfaces, Assistive Technology and Rehabilitation Engineering
Abstract: An exoskeleton-Brain-Computer Interface (BCI) system integrates neuroimaging technology with wearable robotic exoskeletons to control assisted movement. Non-invasive neuroimaging technologies are used to measure the hemodynamic response and electric potential, translating neural signals into interpretable information for BCI systems. Neuroimaging signals acquired under varying environmental (domain) conditions, such as indoor and outdoor settings, exhibit different patterns that introduce variability. This variability in signal patterns presents significant challenges for BCI system design and control. In this study, the effectiveness of deep learning (DL) algorithms: Convolutional Neural Networks (CNN), Transformers, and Long Short-Term Memory networks (LSTM), in the same and cross-domain testing is performed. When DL classifiers are trained and tested in the same environment, all models yielded high performance accuracy and strong results across evaluation metrics, including precision, AUC, recall, and F1-scores. In contrast, when trained and tested in a cross-environment (domain), it results in reduced accuracy, showing the impacts of the domain shifting on model generalization. To resolve this challenge in BCI applications, transfer learning is used as a domain adaptation strategy. Results show that when classifiers are fine-tuned using data from both target and the source domains, significant improvements in accuracy and model robustness are observed. Transfer learning significantly improves DL models' performance to generalize performance across domains. Overall, this work underlines the importance of the domain adaptation (transfer learning) technique in practical BCIs control in diverse real-world conditions. In the future, focus will be on advanced domain adaptation techniques to further improve models' robustness and applicability.

Keywords: Assistive Technology and Rehabilitation Engineering
Abstract: Functional electrical stimulation has been widely applied in stroke rehabilitation, but existing systems often lack continuous adjustment based on patient-specific feedback. This study presents a contralaterally controlled stimulation system that introduces the wrist joint angle as an intermediate variable, establishing two mappings: from surface electromyography of the unaffected side to joint angles via a two-layer long short-term memory network, and from stimulation current to joint angles on the affected side using nonlinear fitting. This design enables real-time closed-loop modulation of stimulation parameters, allowing the affected limb to mirror unaffected-side movements, thereby enhancing patient engagement and potentially accelerating motor function recovery. The prediction accuracy (R^2 > 0.95) and bilateral trajectory correlation (r > 0.9) obtained from experiments with healthy subjects and stroke patients demonstrate the system’s feasibility and rehabilitation potential.

Keywords: Assistive Technology and Rehabilitation Engineering, Human Performance, Design, Analysis, and Evaluation
Abstract: Functional electrical stimulation (FES) has emerged as a promising technique for motor rehabilitation, yet quantifying its effects remains challenging. While current assessment methods utilizing clinical scales, kinematic parameters, and biomarkers offer comprehensive evaluations, they fail to delineate the specific motor function domains influenced by FES, and quantify both therapeutic outcomes and functional improvements. To address these limitations, we introduced multi-objective decision analysis method to propose a novel multidimensional evaluation framework for assessing FES-driven motor performance under varying stimulation parameters. For this, we collected surface electromyogram (EMG) and inertial signals under different parameters of FES-driven wrist movements. We found a relationship between the FES stimulus parameters and different movement level features. Our findings demonstrate that completion indicators generally increased with increasing stimulus amplitude, while movement stability indicators showed an inverse relationship with stimulus frequency. Notably, fatigue indicators exhibited dual trends, reflecting different aspects of motor unit performance. This work providing a crucial foundation for subsequent precise modulation of FES parameters and timely evaluation of rehabilitation progress.

Keywords: Assistive Technology and Rehabilitation Engineering, Human-Computer Interaction, Cognitive System Engineering
Abstract: Stroke severely impairs hand function, particularly in patients with low Fugl-Meyer hand subscores, where rehabilitation is challenged by both motor and cognitive deficits. This study aimed to compare neural engagement under different motor imagery (MI) command modalities in order to optimize brain–computer interface (BCI) rehabilitation strategies. Six post-stroke patients performed hand-focused MI tasks within a five-class paradigm under two conditions: direct command (explicit cues) and active calculation (self-derived MI). Electroencephalography (EEG) was used to assess neural activity through Movement-Related Cortical Potentials (MRCPs) and Event-Related Desynchronization (ERD) in the alpha and beta bands. Direct commands elicited earlier MRCP onset, stronger negative amplitudes, and more sustained ERD compared to the delayed and attenuated responses under active calculation. These findings indicate that structured cues facilitate more robust motor cortex activation, while cognitively demanding tasks may hinder consistent engagement. Direct commands appear more effective in eliciting neural activation for stroke patients with severe impairments, suggesting that reducing cognitive load enhances rehabilitation potential. Clinically, adapting BCI command strategies to patients’ cognitive capacities may improve outcomes, and future work should explore hybrid approaches that balance accessibility with cognitive engagement.

Keywords: Assistive Technology and Rehabilitation Engineering, Machine Learning, Human Performance
Abstract: The objective diagnosis of Parkinson's disease (PD) is challenged by the difficulty in quantifying corticomuscular circuit dysfunction. To address this, we propose a neurophysiologically-inspired Symmetric Bidirectional Cross-modal Network (SBC-Net). The framework first introduces our proposed Rest-AdaptNorm method to calibrate inter-subject variability and features a core bidirectional cross-attention mechanism to dynamically model the directed information flow between cortical motor commands (EEG) and subsequent muscle activation (EMG). Evaluated on a challenging dataset with seven diverse motor tasks under a strict leave-one-subject-out (LOSO) cross-validation, SBC-Net achieves state-of-the-art performance with an F1-score of 89.69%. Crucially, its interpretability analysis reveals a disruption of the directional, time-delayed corticomuscular communication patterns in PD patients, which highly aligns with known pathophysiology. SBC-Net provides a robust and interpretable tool for decoding corticomuscular dynamics, demonstrating significant potential for clinical translation.

Keywords: Assistive Technology and Rehabilitation Engineering, Usability Engineering
Abstract: Parkinson's disease (PD) is the second most common neurodegenerative disorder worldwide, with primary clinical manifestations including resting tremor, bradykinesia, rigidity, and postural gait disturbances, among which limb tremor is the most prominent feature. This paper focuses on visual-spatial motion detection methods, addressing current research challenges such as difficulties in extracting finegrained features from videos and inadequate integration of temporal and spatial dimensions. We propose an innovative Temporal-Spatial Dual-Dimensional Edge Graph Convolutional Attention Model (T-GAU). The model introduces two-dimensional feature streams: static spatial edges and dynamic temporal edges. Static spatial features are learned through an iterative Graph Convolutional Network (GCN) combined with a graph attention mechanism (GAT), while dynamic temporal features are captured using a variant of Gated Recurrent Unit (GRU) blocks integrated with a self-attention mechanism. Evaluated on the Tim-Tremor dataset, the model achieves an accuracy of 99.74% in the tremor binary classification task and 94.48% in predicting disease severity. This approach provides a novel, accurate, and reliable diagnostic solution for assessing upper limb tremors in PD patients.

Keywords: Human-Computer Interaction, Machine Learning, Assistive Technology and Rehabilitation Engineering
Abstract: The trend of global population aging continues to intensify, promoting the growth of demand for active rehabilitation training based on human-computer interaction. Surface electromyography (sEMG) signals can capture human movement intentions and have garnered significant attention in the field of human-computer interaction. To support rehabilitation training for older adults, this paper has established a community-oriented dataset of forearm 8-channel sEMG data for older adults gestures, which includes 40 subjects aged between 60 and 80 years. On this basis, a novel 8-channel sEMG gesture recognition method for older adults, named SyCoT-Res , is proposed. This method effectively addresses the issues of insufficient feature extraction and gradient vanishing in traditional methods when processing high-dimensional time series data by embedding channel feature extraction modules (CFE) and temporal feature extraction modules (TFE) within the network, combined with deep residual learning. Five-fold cross-validation results indicate that the proposed method achieves a recognition accuracy of 98.31% ± 0.72% on the self-built dataset, demonstrating sound recognition performance and stability.

Keywords: Human Machine Systems
Abstract: This paper addresses the parameter identification problem for fractional-order impedance models, which are increasingly used to describe the complex dynamics of systems such as robotic manipulators in interactive tasks. The main challenge lies in accurately and simultaneously estimating both the model coefficients and the fractional differential orders from noisy input-output data. To tackle this, an identification scheme based on a hybrid function basis is proposed. The method combines Block-Pulse functions and Legendre polynomials to construct operational matrices for fractional integration, effectively transforming the dynamic identification problem into an algebraic optimization problem. Numerical simulations demonstrate the effectiveness of the proposed method, showing high estimation accuracy with errors below 4% for all parameters, even under significant measurement noise. The results confirm that the hybrid basis approach offers superior performance compared to traditional block-pulse basis representations, providing a robust and efficient framework for fractional-order system identification.

Keywords: Human Machine Systems, Assistive Technology and Rehabilitation Engineering
Abstract: Lower-limb assistive exoskeletons hold broad application prospects, but achieving precise assistance remains a challenge. In this work, we employ LSTM to accurately recognize the user's gait. Based on adaptive oscillators capable of delay-free tracking of periodic/quasi-periodic signals, we enhance them with phase oscillator compensation and an initialization procedure to achieve precise assistance for the user's diverse gait modes. The effectiveness of this method is validated through simulations and experiments.

Keywords: Human-Computer Interaction, Adaptive Aiding, Human Machine Systems
Abstract: Climbing is a physically demanding and limb coordination requirement task which is often required in special activities. Although many pieces of equipment have been developed to help people during climbing, most of them are passive without the ability to provide active assistance or lack flexibility resulting in restricted movement for people. This paper introduces an active upper-limb assistive suit aimed for climbing movement. In order to simultaneously maintain movement flexibility and provide effective assistance, the actuation system is developed based on a cable driven approach. The actuator’s output is connected to a customized glove through a Bowden cable. This configuration takes the wearer’s hand serving as an anchor point, thereby enhancing the suit’s flexibility and simplifying control. A cascade control framework is implemented to fulfill the assistance control strategy. The upper-level is an assistive force profile generator that employs an impedance control algorithm. This algorithm uses position displacement as its input and generates the desired assistive force as output. The lowerlevel is a force profile tracker, which combines admittance control with a PI velocity control algorithm. The feasibility and effectiveness of our developed device are verified through pullup movement. Experimental results show that muscle activations for Trapezius, Latissimus Dorsi, Infraspinatus, Biceps Brachii and Pectoralis Major reduce by 17.1%,13.4%,18.5%,7.3% and 11.6% respectively.

Keywords: Human-Machine Interfaces, Human-Computer Interaction, Human Machine Systems
Abstract: The pursuit of high-accuracy gaze estimation in unconstrained environments has long driven the development of large, computationally intensive deep learning models. This trend poses a significant barrier to practical deployment on resource-constrained platforms, creating a persistent trade-off between performance and efficiency. This work challenges the conventional notion that model scale is the only path to accuracy, proposing instead a novel training paradigm designed to unlock the full potential of lightweight architectures. We construct an efficient and accurate framework based on the state-of-the-art hybrid vision transformer, FastViTHD. The cornerstone of our methodology is an Auxiliary Attention Supervision (AAS) mechanism, implemented as a parallel heatmap generation task. This auxiliary objective provides an explicit, pixel-wise supervisory signal that forces the model's shared backbone to learn a robust and spatially precise representation of the eye regions. By guiding the model to first learn where to look, we significantly enhance the quality of features available for the primary task of regressing gaze direction. Crucially, this multi-task learning approach substantially improves performance without adding any computational overhead at inference time. We conduct a rigorous evaluation on the challenging MPIIFaceGaze dataset using the standard Leave-One-Subject-Out (LOSO) cross-validation protocol. Our method achieves a highly competitive mean angular error of 3.81°, demonstrating performance comparable to state-of-the-art methods that utilize much heavier backbones. This work validates that intelligent supervision design is a powerful and effective alternative to model scaling for achieving high-performance gaze estimation.

Keywords: Human-Machine Teaming, Autonomous Systems
Abstract: In transportation systems, humans, as operator or passenger, are always somewhat in-the-loop. Human-in-the-loop is an evocative metaphor originating from aviation and taking a central place in automated driving safety. Driving automation shifts the human role from operator to passenger but new types of crashes occur. This paper proposes a collaborative steering system based on admittance control, which provides interactivity together with high trajectory tracking performance, and adjustment of the computed trajectory to manual input. Collaborative steering provides a technical means to connect the driver with the automation so as to form a driving team with shared authority. Improved user acceptance and intuitive operation are anticipated advantages of the proposed practical control.

Keywords: Human-Machine Interfaces, Cognitive System Engineering, Machine Learning
Abstract: Human–machine interfaces are commonly used in various types of systems where control of the system and actuators is required. The purpose of the interface lies in the interaction between humans and computers – from microcontrollers to single-board computers or personal computers. This study proposes novel approaches to controlling systems (prosthetics) using biosignal processing and machine learning algorithms. The work also presents a methodology for acquiring synchronized data using the Lab Streaming Layer (LSL), ensuring reliable and time-aligned input for subsequent analysis. This work highlights the integration of biomedical sensing and machine learning for innovative healthcare solutions, and demonstrates the potential for more intuitive, adaptive, and efficient control of assistive devices, paving the way for improved patient comfort and usability.

Keywords: Machine Learning
Abstract: Pose estimation algorithms aim to detect and precisely localize human keypoints in human keypoints from human instances. Current approaches primarily focus on modifications to the network architecture, with loss functions emphasize the accuracy of individual keypoint predictions, overlooking the intrinsic relationships between keypoints. In this work, we introduce two types of loss functions: Keypoint Center Loss and Link Relationship Loss, which supervise the centers of keypoints and the biological connections between keypoints, respectively. During the same training epochs, our methods achieve improved convergence performance.

Keywords: Machine Learning
Abstract: Visual–tactile fusion enables robots to integrate global geometry from vision with local contact cues from touch, providing complementary insights for reliable manipulation. However, existing approaches often model spatial, temporal, and semantic information in isolation, which weakens cross-modal coherence and makes them brittle under noise and occlusion. To address this, we propose an Adaptive Multi-Scale Attention (AMSA) framework that achieves unified and noise-resilient visuo–tactile fusion through four synergistic modules: (i) MGSA, which preserves geometry by learnable multi-scale Gaussian alignment; (ii) ATG, which enhances temporal consistency via gated sliding-window attention; (iii) LSG, which injects compact semantic priors using prototype-guided attention; and (iv) UAFC, which adaptively re-weights fused representations based on uncertainty cues. Under robustness-aware optimization, AMSA achieves superior cross-modal alignment and stability across diverse manipulation scenarios while maintaining a lightweight 2.1M-parameter footprint, outperforming recent vision–tactile baselines by a large margin.

Keywords: Machine Learning
Abstract: Pear flowers play an important role in agricultural production. The manual detection of target flowers has the problems of low efficiency, high cost, and dense growth in natural environments. Therefore, this article focuses on the monitoring problem of target pear flowers. Based on the YOLOv11 model, two improved modules, FEM and SimAM, are introduced to verify the detection rate of the improved model and the original model through experiments. After adding the FEM module, the trained mAP50 reached 0.934, with an accuracy of 0.981. The F1-score reached its maximum value of 0.88 at a confidence level of 0.548. After training with the SimAM module, the mAP50 reached 0.937, with an accuracy of 0.977. The F1-score reached its maximum value of 0.88 at a confidence level of 0.546. Two methods for monitoring Pear flowers alleviate the pressure of manual detection and provide a new platform for the development of agricultural mechanization and intelligence.

Keywords: Model-Based Design, Human-Computer Interaction, Machine Learning
Abstract: To address inadequate capture of non-stationary signals, MultiScaleResidualCNN is proposed, using multi-scale kernels to parallelly capture temporal features and residual connections to enhance weak signal transmission, boosting electromyographic (EMG) features capture. Experiments on 6 subjects recognizing 10 gestures show individual accuracies over 97.44% (optimal 99.87%) and 96.97% overall,enhancing model accuracy. For edge deployment under wearable constraints, two lightweight strategies are tested: Int8 quantization reduces size by 3.91% with ≤0.04% performance loss; Depthwise separable convolution (DSC) cuts size by over 80%. These provide precision-efficiency trade-offs, advancing practical EMG deployment.

Keywords: Model-Based Design, Human-Computer Interaction, Machine Learning
Abstract: Electroencephalography (EEG) is directly linked to the brain’s emotional processes and, unlike behavioral signals such as facial expressions, is not easily susceptible to artificial manipulation—granting it significant advantages in emotion recognition. However, its performance is prone to interference from noise, individual differences, and acquisition conditions, resulting in compromised robustness. While multimodal approaches that combine EEG with peripheral physiological signals (PPS) have emerged, many of these methods independently model feature extraction for each modality. This can lead to inconsistent feature representations, complicate the fusion process, and fail to fully leverage the complementary information across different modalities. To tackle these challenges, we propose a multimodal emotion recognition framework with two key components: a universal CNN-based feature extraction module, which can automatically learn spatial and temporal representations from diverse physiological signals, and a Transformer-based fusion mechanism, which captures global inter-modal dependencies through self-attention. Experiments on the DEAP dataset show that our framework delivers robust performance, highlighting its versatility and robust generalization capability in emotion recognition.

Keywords: Model-Based Design, Usability Engineering, Machine Learning
Abstract: A robust billet character recognition approach tailored for industrial robotic vision systems operating in extreme environments is proposed. To address the challenges of partial occlusion, uneven illumination, and character degradation commonly encountered in real-world robotic applications, a geometrically constrained Spatial Distribution Module(SDM) is introduced to mitigate detection omissions across multiple text regions. Furthermore, a Cross-Region Feature Fusion Module(CRFFM) and a Missing Character Supplement Module(MCSM) are developed to promote inter-region feature synergy and recover missing character information, ensuring consistent and complete recognition. Experimental results on real-world industrial datasets demonstrate that the proposed method achieves a character recognition accuracy of 99.9%, significantly surpassing conventional techniques and showcasing strong potential for deployment in industrial robotic vision tasks.