In recent years, deep learning techniques for motor imagery (MI) electroencephalography (EEG) decoding have shown great potential for advancing brain-computer interfaces (BCI). However, non-linearities and subject dif ferences in EEG signals negatively impact the robustness, applicability, and generalization of MI-EEG decoding, hindering BCI development. This paper proposes a novel network based on Bridge structures and Attention mech anisms (BANet). Inspired by transformer architectures, BANet incorporates multiple innovative Bridge blocks to extract temporal features of EEG signals from both local and global perspectives. These blocks model complex temporal relationships, enabling BANet to leverage information from diverse angles and enhance decoding perfor mance. Additionally, attention mechanisms and inception architecture are employed to focus on and extract EEG features across multiple dimensions and scales, further optimizing the utilization of temporal-spatial features. In various validation experiments, BANet demonstrates superior performance compared with recent representative deep learning methods, confirming its effectiveness in extracting local and global temporal and spatial features and its potential for developing robust BCI applications.

Soft pneumatic actuators provide compliant and comfortable hand rehabilitation, but conventional systems suffer from limited self-sensing capabilities and insufficient human-robot interaction (HRI). We present a segmented soft rehabilitation glove with a serpentine-embedded fiber Bragg grating (FBG) sensing structure, designed to alleviate bending-induced stress concentration and provide accurate, long-term joint-angle tracking. We further propose a Lightweight Temporal Multimodal Fusion Network (LTMFNet) that integrates FBG and surface electromyography (sEMG) signals. The network employs an Adaptive Time-Frequency Channel-wise Self-Attention (ATFCA) to capture discriminative temporal, spectral, and channel patterns in sEMG, while a convolutional minimal Gated Recurrent Unit (minGRU) efficiently models long-range spatiotemporal dependencies in FBG sequences. Experimental results demonstrate that the serpentine FBG configuration reduces stress concentration by 8.58% and yields higher tracking accuracy compared to linear layouts. On a 20-class gesture recognition task, the model achieves an accuracy of 97.89% (±0.36%), enabling real-time, intention-driven rehabilitation.

Pulmonary rehabilitation (PR) aims to improve lung function in patients with chronic respiratory disease (CRD). In recent years, significant advancements have been made in pulmonary rehabilitation technologies, demonstrating their potential for enhancing lung function in patients with respiratory diseases. The purpose of this study is to outline recent developments in the field of pulmonary rehabilitation guided by pulmonary rehabilitation robots, which has not been previously addressed in earlier reviews. To fill this gap, this paper first provides a systematic summary of the monitoring and actuation technologies of pulmonary rehabilitation robot systems and evaluates these technologies from multiple dimensions, including portability, wearability potential, invasiveness, and clinical applications, analyzing the potential for integrating various technologies into pulmonary rehabilitation robot systems. Furthermore, three technical directions are proposed: real-time precise monitoring, suitable structure and actuation strategies, and the intelligence of pulmonary rehabilitation robot systems. On the basis of these directions, this paper presents a comprehensive technical outlook for a soft wearable pulmonary rehabilitation robot system, providing reference and guidance for future research. To our knowledge, this is the first review of pulmonary rehabilitation robot systems and their key technologies. Additionally, the review section on respiratory assistive technologies simultaneously covers key technologies such as mechanical ventilation (MV), exoskeleton robots, and functional electrical stimulation (FES) for the first time. It also summarizes the respiratory assistive technology paradigm from the innovative perspectives of respiratory assistive modalities, targeted body sites, and types of ventilation for the first time. This study offers a broader perspective and a deeper understanding of pulmonary re-habilitation robots, with a technical outlook encompassing multimodal data fusion perception, respiratory event detection and intention recognition, full-phase assistance strategies, modeling, decoupling, and quantification of multiple-input multiple-output (MIMO) systems, as well as model-based interactive control strategies.

Stroke often results in impaired hand motor function, making effective hand rehabilitation essential for restoring activities of daily living (ADLs). Motor rehabilitation and neurorehabilitation are two major pathways to functional recovery. Rehabilitation gloves have proven to be effective tools for motor rehabilitation, and among them, soft robotic gloves (SRGs) have emerged as a research focus due to their lightweight design and inherent safety. Functional electrical stimulation (FES), which applies electrical currents to muscles and nerves, shows promise in promoting motor neural reorganization and restoring muscle strength in the hands of stroke survivors. The technologies applied to hand rehabilitation must possess the characteristics of safety, comfort, and practicality, while overcoming critical challenges such as portability, user-friendliness, and wearability. Motivated by the rehabilitation needs of post-stroke patients, this paper reviews recent advances in SRGs, FES, and hybrid hand rehabilitation systems (HHRSs) for hand rehabilitation, systematically examining progress in actuation strategies, intention sensing, and control algorithms across these three technologies. Furthermore, the limitations and technical challenges of current HHRSs are analyzed and four key future research directions are identified to pave the way for further development in this field.

Despite advancements in robotic surgery, recent developments in surgical robots have rarely focused on ensuring the safety of surgeons and other devices within the surgical environment. This paper introduces a novel real-time modeling and perception method tailored for surgical Dynamic Unstructured Environments, leveraging multisensory data. We propose the use of a uniform Cartesian grid (UCG) over the traditional radial non-uniform grid (RNUG) to establish a more accurate offline model of the surgical environment. Furthermore, our approach incorporates multisensory data fusion and UCGs for rapid online updates of environmental models. This dual strategy of offline modeling and online updating addresses both modeling accuracy and efficiency, making UCG modeling more suitable for multisensory fusion compared to RNUG. The proposed methodology enhances the overall safety and operational reliability of surgical robots.

Functional electrical stimulation (FES) has shown promise in restoring motor functions for patients with spinal cord injury and stroke. However, its clinical application is limited by insufficient accuracy in modeling muscle dynamics and the lack of robust control strategies under complex disturbances. To address these challenges, this study proposes a closed-loop framework that integrates high-precision modeling with strong robustness. A Hammerstein model enhanced by Kolmogorov-Arnold Networks (KAN) is constructed, where the explicit mathematical representation of KAN significantly improves the nonlinear dynamic modeling of muscle behavior. Additionally, a forgetting factor recursive least squares (FFRLS) algorithm is employed for online identification of time-varying parameters, achieving improved performance over traditional approaches. Further, a sliding-mode tube model predictive control (SMC-Tube MPC) strategy driven by surface electromyography (sEMG) feedback is developed. By combining the disturbance rejection capability of sliding mode control with the state constraint handling features of Tube-MPC, the proposed controller enables stable torque tracking under complex perturbations. The framework is validated on an experimental platform integrating a dynamometer, sEMG acquisition device, and electrical stimulator. Experiments with healthy subjects demonstrate high accuracy and strong robustness of the proposed system.

Caregiving and eldercare manipulation demands high motion smoothness and safety, as robots often operate in cluttered, contact-rich environments and in close proximity to humans. In our real-robot deployment, Action Chunking with Transformers (ACT) achieves strong task-level performance but suffers from execution-time instabilities—most notably high-frequency action jitter and inter-chunk discontinuities—that can substantially reduce manipulation reliability. We propose Action-Chunk Filtering (ACF), a lightweight, plug-and-play post-processing module that leverages ACT’s predicted action chunks to stabilize control commands without retraining the policy. ACF combines (i) intra-chunk, shape-preserving smoothing via Savitzky–Golay filtering, (ii) inter-chunk adaptive fusion with a speed-aware update rule to attenuate chunk-boundary artifacts, and (iii) rate/acceleration constraints to enhance robustness in safety-critical deployments. Across multiple caregiving-inspired manipulation tasks on Cobot Magic at 30 Hz, ACF raises the average success rate to over 85%. These results indicate that action-chunk-aware filtering offers an effective, low-cost pathway to stabilizing transformer-based imitation policies for real-world assistive manipulation.

Long-horizon robotic manipulation requires coordinating multiple motor primitives under uncertainty, especially in contact-rich and changing environments. Existing end-to-end visuomotor policies often lack explicit temporal structure, causing brittle execution and limited recovery after disturbances. Inspired by biological motor control, where reusable primitives are organized through phase decomposition and feedback-dependent transitions, we propose a bio-inspired phase-aware framework that represents execution as structured transitions over perception-grounded motor phases. The framework integrates three modules: a Multimodal Phase-and-Primitive Detector that extracts semantically and physically consistent phases from visual, proprioceptive, and force–torque signals; a Multimodal Perception Skill Graph (MPSG) that encodes feasible phase transitions and supports skipping, rollback, and recovery; and Promptable Phase Control, which converts language instructions into graph-level ordering constraints for task reordering without retraining low-level policies. Experiments on four multi-stage tasks show improved robustness, increasing the average disturbed-condition success rate from 25.0% to 73.8% relative to the monolithic Action Chunking with Transformers (ACT) baseline.

Motor imagery (MI) classification is an important research topic in brain-computer interfaces (BCI). Due to the nonlinear, nonsmooth, and complex topological structure of electroencephalography (EEG) signals, their distribution in high-dimensional space exhibits nonconvex manifold geometric features. Traditional Euclidean-based algorithms encounter difficulties in effectively processing such complex structures. In contrast, manifold learning methods have been widely investigated due to their capability of capturing the intrinsic geometric structure of data more effectively. However, most existing manifold learning approaches overlook the temporal coupling across different frequency bands in neural activity and suffer from structural information degradation during multistage manifold feature mapping. To address these challenges, this paper proposes a multiscale spatiotemporal convolutional fusion U-shaped Riemannian manifold learning network for MI-EEG decoding. Specifically, the proposed approach employs multiscale spatiotemporal convolution to effectively extract local temporal dynamic coupling features across time domains, partially mitigating the limited training data for EEG. Furthermore, a U-shaped manifold learning network with an encoder-decoder structure is utilized to capture global spatial dynamic information while alleviating the degradation of structural information during manifold mapping. Additionally, a Riemannian barycenter-based skip connection operation is introduced to normalize the SPD matrix distribution, reduce intra-class distance, and increase the inter-class distance of the features. Finally, the proposed method is evaluated on three widely used datasets. The experimental results demonstrate the superior performance and robustness of the proposed network in MI-EEG decoding.

Urban IoT devices face increasing challenges due to urbanization and complex traffic conditions. Traditional monitoring methods are inadequate for tracking urban traffic and ensuring public safety. Collaborative monitoring models involving multiple Uncrewed Aerial Vehicles (UAVs) and Uncrewed Ground Vehicles (UGVs) are emerging. However, sudden emergencies such as traffic accidents and geological disasters can disrupt monitoring tasks across urban traffic networks, leading to traffic paralysis and personal injury. This study develops a mathematical model for multi-UAV-UGV path planning, specifically tailored for dual mission applications, namely continuous monitoring and emergency response. A foresight-based task allocation and path optimization algorithm is proposed, which accounts for future situation prediction. This algorithm takes into account various constraints, such as the road network restriction and information timeliness, with the aim of optimizing system efficiency and response speed. Various factors are comprehensively considered to select response vehicles, and the potential loss caused by emergency situations of the entire road network is reduced in the case of rapid emergency treatment. Simulation results demonstrate that this method is capable of providing an efficient and reliable path for dual mission operation. Additionally, the algorithm exhibits good stability and strong adaptability to random emergency events.

Artificial muscles hold considerable promise in robotics owing to their lightweight and highly compliant characteristics. Nevertheless, existing bio-inspired actuation still lack human-like functionality and struggle to balance precision with compliance. Therefore, this study designs a Sarcomere-Inspired Recruitment Array Actuator (SRAA) motivated by the fundamental motor units of human muscles. The designed actuator employs a magnetic actuation unit array to realize progressive recruitment and cooperative force generation. A hierarchical structure combined with an integrated position-force control strategy enables dynamic allocation of output according to task requirements, thereby supporting fine-grained position regulation and tunable compliance. This study develops a humanoid robotic arm platform to validate the proposed design. Experimental evaluations demonstrate that the proposed system achieves millimeter-level positioning accuracy and stable mechanical response, while also exhibiting fast dynamic performance and high task adaptability.The robotic arm achieved an MAE of 0.04 mm in the dynamic water-filling task and maintained sinusoidal tracking MAEs below 0.25 mm for actuator displacement and 0.69° for forearm angle. The design provides a new pathway for compliant, human-like actuation in humanoids and rehabilitation robots.

The natural interaction and control performance of lower limb rehabilitation robots are closely linked to biomechanical information from various human locomotion activities. Multidimensional human motion data significantly deepen the understanding of the complex mechanisms governing neuromuscular alterations, thereby facilitating the development and application of rehabilitation robots in multifaceted real-world environments. However, existing lower limb datasets are inadequate for supplying the essential multimodal data and large-scale gait samples necessary for the development of effective data-driven approaches, and the significant effects of acquisition interference in real applications are neglected. To fill this gap, we present the K2MUSE dataset, which includes a comprehensive collection of multimodal data, comprising kinematic, kinetic, amplitude mode ultrasound (AUS), and surface electromyography (sEMG) measurements. The proposed dataset includes lower limb multimodal data collected from two cohorts, including 30 able-bodied young adults and 12 older adults, across different inclines (0 degrees, +/- 5 degrees, and +/- 10 degrees), speeds (0.5 m/s, 1.0 m/s, and 1.5 m/s), and representative non-ideal acquisition conditions (muscle fatigue, electrode shifts, and interday differences). The kinematic and ground reaction force data were collected with a Vicon motion capture system and an instrumented treadmill with embedded force plates, whereas the sEMG and AUS data of 13 muscles on the bilateral lower limbs were synchronously recorded. To validate the quality of the data, we quantified repeatability across locomotion modes, speeds, and inclines, examined physiological signatures under non-ideal acquisition conditions, and observed high agreement with existing public datasets. We also report baseline motion-intention recognition results, including joint angle estimation and gait phase classification, with a multimodal transformer model demonstrating accurate and stable performance. In addition, we present a control-oriented application in which an end-to-end model trained on the K2MUSE dataset provides hip assistance via a soft exoskeleton and yields consistent reductions in metabolic cost across multiple terrains. K2MUSE is released with the corresponding structured documentation, preprocessing pipelines, and example code, thereby providing a comprehensive resource for rehabilitation robot development, biomechanical analysis, and wearable sensing research. The dataset is available at https://k2muse.github.io/.

In medical image segmentation, the Segment Anything Model (SAM) is constrained by its reliance on manual prompts and limited generalization to downstream medical tasks. To address these challenges, we propose Dual-Interactive Tuning SAM (DIT-SAM), an enhanced SAM variant equipped with a dual-interactive tuning mechanism. By simultaneously optimizing information propagation and parameter adaptation, DIT-SAM enables prompt-free segmentation and improves generalization to downstream medical tasks. In the forward information flow, we introduce the Forward Feature Guidance Module (FFGM), which dynamically transmits multi-level semantic features from the encoder to guide the decoder. Conversely, in the backward information flow, the Backward Feedback Regulation Module (BFRM) regulates feedback from the decoder to the encoder through learnable tokens and cross-layer projections. Together, these two modules form a closed-loop optimization process of "Feature Guidance-Parameter Feedback," enabling deeper interaction between encoder and decoder. DIT-SAM adopts a parameter efficient fine-tuning (PEFT) strategy, updating only 1.52% of SAM's parameters. This preserves the foundational capabilities of SAM while enhancing its adaptation to domain-specific medical tasks. Extensive experiments on public medical segmentation datasets show that DIT-SAM enables prompt-free, end-to-end segmentation and outperforms current state-of-the-art methods in full-shot scenarios, while also achieving strong results in few-shot scenarios with minimal parameter updates. These results demonstrate its combined strengths in segmentation accuracy and parameter efficiency, offering a practical and scalable solution for real-world medical image analysis. The code will be available at https://github.com/yingyuhan/DIT-SAM.

Perception technology underpins robotic intelligence, functioning as the data foundation for large models and embodied intelligence. It facilitates data - driven autonomy, adaptive learning, and seamless human-robot interaction, with widespread broad applications in robotics, medical rehabilitation, and environmental monitoring. Although there have been advancements in sensor diversity, current research mainly focuses on the principles, materials, and other aspects of individual sensors. However, extant sensing technologies pose constraints on the progress of intelligent robotics and embodied intelligence. The advent of large-model-driven embodied intelligence imposes unprecedented demands on sensing capabilities. Nevertheless, fundamental limitations, particularly those in the integration with the human body, impede further breakthroughs. This paper systematically reviews recent progress in humanoid perception from three perspectives: proprioception, exteroception, and beyond-human perception. It comprehensively analyzes sensor principles, advantages, and limitations while identifying future directions such as multimodal fusion, deep integration of artificial intelligence, and emerging applications. These advancements reshape perception and sensing paradigms, laying the foundation for next - generation embodied intelligence. This study meticulously analyzes the current state and future trends, exploring the sensing infrastructure essential for the robotic era and providing a data backbone for future general AI - driven robots.

Exoskeletons have been shown to effectively assist humans during steady locomotion. However, their effects on non-steady locomotion, characterized by nonlinear phase progression within a gait cycle, remain insufficiently explored, particularly across diverse activities. This work presents a shank angle-based control system that enables the exoskeleton to maintain real-time coordination with human gait, even under phase perturbations, while dynamically shaping assistance profiles to match the biological ankle moment patterns across walking, running, stair negotiation tasks. The control system consists of an assistance profile online generation method and a model-based feedforward control method. The assistance profile is formulated as a dual-Gaussian model with the shank angle as the independent variable. Leveraging only IMU measurements, the model parameters are updated online each stride to adapt to inter- and intra-individual biomechanical variability. The profile tracking control employs a human-exoskeleton kinematics and stiffness model as a feedforward component, reducing reliance on historical control data due to the lack of clear and consistent periodicity in non-steady locomotion. Three experiments were conducted using a lightweight soft exoskeleton with multiple subjects. The results validated the effectiveness of each individual method, demonstrated the robustness of the control system against gait perturbations across various activities, and revealed positive biomechanical and physiological responses of human users to the exoskeleton's mechanical assistance.

Precise kinematic modeling and control of soft continuum robots are challenged by their inherent deformability and complex interactions with the environment. This paper proposes an adaptive kinematic modeling framework based on a Deep Belief Network with Event-Driven Incremental Learning, designed for artificial intelligence-enabled control of soft robotic systems. The deep belief network is first pre-trained using simulated data to establish an initial kinematic model, requiring only approximately 600 real-world samples for deployment. An event-driven incremental learning mechanism is then introduced to adapt the model online. This mechanism is guided by a spike intensity metric, which evaluates prediction errors and selectively triggers either fine-tuning of network parameters or the integration of new radial basis function nodes to compensate for unmodeled kinematic effects and environmental disturbances. The adaptive kinematic model is embedded within a closed-loop controller to achieve accurate trajectory tracking. Experimental validation is conducted on a tendon-driven soft origami manipulator, covering trajectory tracking under varying payloads and deformation constraints, disturbance rejection, and teleoperation tasks. The proposed framework achieves sub-millimeter average trajectory tracking accuracy under confined conditions and outperforms baseline deep belief network models and piecewise constant curvature approaches. The results demonstrate that the proposed artificial intelligence-based adaptive kinematic modeling method provides a data-efficient and robust solution for precise control of soft continuum robots in applications such as medical robotics and flexible manufacturing.

Underwater robots require sophisticated tactile sensing systems to execute high-precision operations in deep-sea environments. However, extreme hydrostatic pressure in deep-sea environments can cause sensor signal baseline drift and structural fatigue, or even device failure. Meanwhile, operational targets range from fragile organisms to rigid cultural relics, demanding that sensors possess both high sensitivity and a wide dynamic range to meet diverse tactile perception needs. To address these challenges, we present a bioinspired magnetic tactile sensor with pressure self-balancing capability, drawing inspiration from the hydrostatic skeletal mechanisms of marine organisms and the multi-level tactile sensing properties of human skin. The sensor incorporates a butterfly-shaped porous structure that enables internal and external pressure self-balancing under hydrostatic pressure, effectively suppressing signal drift to less than 1% even at 120 MPa. Through the synergistic design of biomimetic cilia and the porous structure, the device achieves a wide force sensing range from 0.0063 N to 718 N and a high sensitivity of 43.72 µT/N. During sea trials, a soft robot system integrated with the sensor demonstrated excellent performance in multiple dive missions, operating at depths of up to 2,284.7 meters. Across different depth environments, the adaptive system successfully accomplished force-closed-loop precision grasping of various deep-sea organisms, contour recognition, and cooperative sampling operations based on tactile interaction, fully validating its practical utility under real deep-sea conditions. This study provides a reliable tactile sensing solution for intelligent underwater robots, showcasing broad application potential in deep-sea resource exploration and ecological environment research.

Recent advancements in intelligent robotics have substantially heightened the demand for multidimensional haptic sensors. Traditional haptic sensors are constrained by their sensitivity, anti-interference capabilities, and measurement range, thereby limiting their application in complex environments. Inspired by the structure of human skin, a 3-D arrayed magnetic sensor (MCE) was developed, exhibiting high sensitivity (57.625N(-1)), a wide force measurement range, low hysteresis (1.04%), and high stability. The sensor comprises a three-layer design: the top layer includes a flexible square magnet that deforms under external forces to detect normal and tangential forces, the middle layer incorporates a flexible elastomer (Ecoflex 00-30) combined with a spring to achieve a broad range of force measurements, and the bottom layer consists of a flexible circuit board flexible printed circuit board (FPCB) embedded with an array of Hall sensors to convert magnetic field variations into electrical signals. The normal force measurement range of the MCE sensor spans from 0 to 16 N, with a minimum resolution of 10 mN, enabling the accurate detection of minor force variations. The tangential force is measured within a range of-7 to 7 N. Experimental findings indicate that the MCE sensor enables real-time slip detection and, when integrated with deep learning algorithms, attains a remarkable object recognition accuracy of 96.36%. This technological progression significantly augments tactile sensing functionalities, thereby facilitating advancements in intelligent robotics, such as precision grasping, and enhancing human-machine interaction, particularly in the realm of medical rehabilitation devices.

This letter proposes a modular variable configuration rehabilitation robot (MVCRR), aiming to meet the needs of multi-functional, full-cycle rehabilitation. MVCRR integrates a lower-limb training module and a sit-to-stand module, offering 16 actuated degrees of freedom and supporting four rehabilitation postures: supine, sitting, standing, and sit-to-stand transition, while integrating the functions of typical rehabilitation devices. Meanwhile, a split-reassembly mechanism is designed to enable flexible bedside deployment, while compact scissor mechanisms and optimized actuation design improve adaptability and efficiency. In addition, a distributed bilateral coordination control architecture is established, and a unified human-robot interaction control framework is developed based on a multi-posture dynamic model library, enabling compliant, assist-as-needed control. Eight active and passive training modes are designed, covering gait reconstruction, muscle strengthening, posture transitions, and bilateral coordination. Experimental results demonstrate high-accuracy joint tracking (RMSE <= 0.0083 rad) and torque regulation (RMSE <= 1.6848 Nm), and good natural-gait reproduction (joint-angle RMSE <= 0.094 rad, r >= 0.897), validating the effectiveness of rehabilitation training across multiple postures. Overall, MVCRR establishes a systematic and intelligent integrated solution that supports multi-posture training and bedside deployment, demonstrating strong potential for clinical application.