Cardiovascular interventions often rely on X-ray fluoroscopy for real-time image guidance, which involves occupational health risks of working with ionizing radiation. Magnetic Resonance Imaging (MRI) presents a radiation-free alternative and offers superior soft tissue contrast. However, the use of MRI for procedural guidance has been limited by several technical challenges, including poor visibility and control of the endovascular device. This study aims to improve real-time MRI guidance by introducing an image-driven robotic approach. The visualization of the endovascular guidewire was automated by detecting passive susceptibility markers using a convolutional neural network (CNN). The guidewire was manipulated using a modified MR-safe CathBot platform to enable real-time control and image-based haptic feedback. MRI sequences with optimal contrast-to-noise ratio for AI-based marker detection were used, allowing relatively high temporal resolution. This integrated approach introduces a new concept for improving control and precision during MRI-guided endovascular interventions, with potential for improved occupational safety and clinical outcomes.
Natural Orifice Transluminal Endoscopic Surgery (NOTES) imposes exceptionally high requirements on the operational precision and tissue safety of flexible robots. To address the prevalent "force-position conflict" during complex interactions, this paper proposes a bi-level master-follower shared control framework based on a dynamic virtual fixture (VF) and adaptive model predictive control (MPC). In this framework, the upper level utilizes dynamic VF to guide operator intent and restrict motion in non-safe regions, while the lower level incorporates an adaptive MPC controller. The core innovation lies in an adaptive weight scheduling mechanism that dynamically adjusts the position and force tracking weights by perceiving the terminal contact force state in real time. Furthermore, the rigorous mathematical foundation of the proposed framework is established by analyzing closed-loop nominal stability, recursive feasibility under practical environmental bounds, and the passivity of the haptic interaction loop. Trajectory tracking and ex vivo surgical simulation experiments, benchmarked under a standardized Independent Peak-Performance Tuning Protocol, demonstrate that the proposed method strictly confines the contact force within the 0.5 N safety threshold with zero force overload, while simultaneously suppressing the trajectory RMSE to within 1.32 mm. This study effectively breaks through the bottleneck of force-position synergistic control in unstructured environments, significantly enhancing the accuracy of fine operations while absolutely guaranteeing the physical safety of soft tissues.
Peripheral Intravenous Catheterization (PIVC) is one of the most common invasive medical procedures, with over one billion procedures performed annually. However, its failure rate remains unacceptably high at 35% to 50% in patients with Difficult Intravenous Access (DIVA), posing significant clinical challenges. In this paper, we present Cathbot-Pro, a handheld robotic device designed to enhance the safety and accuracy of automated PIVC. The system integrates a novel real-time vein segmentation algorithm that leverages Near-Infrared (NIR) imaging alongside established image processing techniques, including Contrast-Limited Adaptive Histogram Equalization (CLAHE) for contrast enhancement and Alternate Sequential Morphological Filtering (ASF) for smoothing shapes. The device was designed, fabricated, and tested for robustness and usability on a phantom model with 14 expert users. The vein segmentation module was independently evaluated using a mock dummy device on a pool of 25 volunteers. The results demonstrate that the proposed system effectively achieves a 97% success rate on adult phantoms and 93% on pediatric phantoms, with 100% first-attempt success in user trials. Additionally, the integration of the NIR guidance module shows promise in reducing complications associated with failed insertions.
Preoperative planning in Robot-Assisted Minimally Invasive Surgery (RAMIS) is crucial for surgical safety, and efficiency. However, existing planning methods rely heavily on the surgeon’s experience and face risks such as collisions or limited visibility due to improper coordination between multiple robotic arms. This study proposes a preoperative planning algorithm based on deep reinforcement learning—TrocarOpt-DRL, aiming at optimizing RAMIS preoperative configurations to improve planning efficiency and surgical safety. It uses an actor-critic architecture to model the preoperative planning problem as a Markov decision process, and optimizes the strategy using a reward function based on surgical evaluation metrics. Taking paraesophageal hernia repair as an example, the simulation results showed that TrocarOpt-DRL outperformed traditional methods with a 6–11% improvement in overall planning score and showed an approximately 80 times higher computational efficiency compared with particle swarm optimization. In further comparative experiments conducted on a real robotic platform, the preoperative planning approach based on TrocarOpt-DRL demonstrated higher operational efficiency, more stable motion characteristics, a safer operating environment, and an improved user experience.
Soft wearable assistive devices offer inherent compliance and enhanced comfort for human-robot interaction. However, the lack of intrinsic sensing in most soft actuators complicates reliable force estimation and system integration. Effective exosuit assistance further requires user-specific personalization, typically achieved using time-consuming, costly, and laboratory-bound metabolic instrumentation based calibration. We present a soft bilateral knee-flexion exosuit powered by a custom sensorized pneumatic artificial muscle (sen-pSTAM) actuator. By unifying actuation and sensing, the sen-pSTAM enables real-time transmitted force estimation with an average RMSE of 3.98N across walking speeds of 3.5-5.5 km/h, eliminating requirements of external force sensors. The actuator geometry is synthesized using a generative neural-network based framework through a multi-constrained, gait-kinematics-informed optimization. The exosuit integrates textile-based anchoring, a ROS-enabled embedded sensing-control-actuation architecture with on-board edge-inference, and an untethered pneumatic supply. Rapid bilateral personalization of exosuit control is achieved through an innovative human-in-the-loop Bayesian optimization framework driven by transmitted force feedback, completing single-speed calibration in 2.93 ± 0.87 minutes and delivering peak assistive forces of 42.26 N. Wireless surface electromyography (sEMG) evaluation across subjects demonstrates effective assistance, with a maximum muscle activation reduction of 32.17% in the gastrocnemius medialis, thus paving the framework for real-world deployment.
Upper limb amputations severely impair motor and sensory functions, affecting independence in daily life. Myoelectric prostheses can restore motor function by translating muscle signals into movement, but they often lack adequate sensory feedback, which is essential for intuitive control. The literature presents many approaches to provide grasping force feedback, but they focus only on the normal force component. However, human mechanoreceptors are also sensitive to shear forces, which are crucial for effective and safe object manipulation. This study proposes a novel encoding scheme using multiple vibration motors to convey both normal and shear components of grasping force, measured by custom tactile sensors placed on prosthetic fingers. The system was preliminarily tested in 17 non-disabled participants while doing two different experiments. The first evaluated whether users could identify object weight and force direction from the feedback. The second assessed whether the added shear feedback improved slippage detection compared to normal force feedback alone. Participants achieved high success rates in decoding weight and direction (median: 80–88%), demonstrating that the feedback was intuitive even without prior systematic training. Moreover, slippage detection improved significantly with the inclusion of shear force feedback: participants recognized slippage in 82% of trials with combined feedback, compared to 64% with only normal force. Overall, this preliminary study highlights the potential of integrating multidirectional force feedback in hand prosthetics to improve safety and usability, especially by detecting and reacting to abrupt force changes during object manipulation.
Robot-assisted vitreoretinal surgery requires micron-scale precision, stable depth perception, and continuous awareness of instrument-tissue interaction. This paper presents a comprehensive evaluation of the GEYEDANCE platform, a multimodal shared-control architecture based on Optical Coherence Tomography, haptic shared control, acoustic feedback, and augmented decision-making. Two validation campaigns were conducted: a formative evaluation on synthetic eye phantoms using the research platform, and a pre-clinical in vivo study on porcine eyes using a clinically approved robotic system. On the research platform, the assessment followed the nominal seven-step epiretinal membrane workflow, while the detailed formative outcome reporting followed six evaluated task blocks, with flap creation and grasping assessed jointly as flap initiation/grasp. On the clinical platform, the evaluation focused on representative subtasks: grasping, vitrectomy, and endolaser photocoagulation. These findings demonstrate that multimodal assistance can effectively enhance safety and decision-making in robotic-assisted eye surgery when aligned with surgical practice.
High-quality, comprehensively annotated datasets are essential for advancing artificial intelligence (AI) and machine learning (ML) in gastrointestinal disease diagnosis and monitoring, such as Celiac disease (CeD). Current clinical video capsule endoscopy (VCE) recordings lack synchronized image-pose information, constraining both algorithm training and interpretability. This study introduces a robotic wire-driven VCE platform that enables controlled and repeatable capsule motion within benchtop intestinal simulators to generate synchronized, multimodal data. The system integrates a linear-drive translation module and a two-degree-of-freedom tendon-actuated orientation unit, reproducing physiologically realistic trajectories extracted from in-vivo motion datasets. Experimental results demonstrate sub-millimeter translational precision (0.59 mm ± 0.3 mm) and angular errors below 2°, with overall spatial RMSE of 3.26 mm ± 1.54 mm during intestinal navigation emulation. By providing synchronized image–pose pairs under programmable conditions, the platform bridges a critical gap in AI dataset quality, enabling the development, benchmarking, and generalization of computer-vision models for gastrointestinal disease assessment. Furthermore, its repeatable actuation and real-time control interface establish a realistic training environment for clinicians, enhancing procedural understanding and endoscopic skill acquisition. This versatile framework extends beyond CeD to broader applications in robotic endoscopy, medical imaging, and data-driven healthcare innovation.
Cable-driven robots are becoming popular mechanisms for various gait rehabilitation tasks. However, gait-induced force disturbances in these systems result in increased force-tracking errors under pathological gait conditions. The present study aims to develop an enhanced feedforward force controller to compensate for gait-cycle disturbances. The proposed approach combines offline disturbance modeling with online disturbance compensation. In the offline stage, participant-specific disturbances are identified via spectral analysis of cable force sensor signals. In the online stage, a disturbance feedforward command is synchronized with the gait phase estimated from thigh-mounted IMU data. The disturbance profile is iteratively updated to compensate for modeling errors and gait variability. To evaluate the performance of the enhanced feedforward controller, nine healthy participants were recruited. Participants walked in three different gait scenarios: baseline, simulated hemiplegic gait, and simulated diplegic gait. Across all gait patterns, the enhanced feedforward-based controller reduced root mean squared error of force by 50% compared to the nominal feedforward controller. Improved force-tracking accuracy may enable more consistent delivery of designed force intervention, thereby supporting more reproducible and effective gait training for patients.
Wearable robots require firm attachment to transmit assistive forces, but excessive or sustained strapping pressure increases discomfort and soft-tissue injury risk. Most studies rely only on force measurements at the robot’s attachment point, missing limb–cuff interactions such as micro-slippage and tissue deformation. We present a spatiotemporal strapping-pressure framework using three actively controlled inflatable straps and a sensorized cuff with twelve miniature 3D force sensors that measure in-cuff interaction forces, especially shear, alongside comfort ratings. Twenty-seven static configurations, grouped by total bladder pressure, and one dynamic configuration that redistributed pressure over time at constant total pressure, were tested under controlled cuff-to-arm relative motion using a collaborative robot. Configurations sharing the same total pressure showed statistically equivalent end-effector stiffness and damping, whereas different totals produced distinct mechanical responses. Raw cuff-sensor quasistiffness decreased with pressure due to reduced transmitted shear; after linear scaling to the load-cell data (R² = 66.9%) to compensate for incomplete surface coverage, the corrected quasi-stiffness recovered the end-effector trend. End-effector energy dissipation changed little with pressure, while in-cuff measurements revealed up to 64% lower interface losses at higher pressures. The dynamic configuration preserved mechanical coupling relative to a matched static medium-pressure condition while achieving significantly higher comfort (24% improvement, p = 0.0005), a strategy relevant to upper-limb exoskeletons for stroke rehabilitation where prolonged wear is required.
Magnetic tracking is a convenient and cost-effective technique for localizing medical instruments with embedded permanent magnets. However, for medical continuum robots (CRs), models trained on datasets of unconstrained magnets are not optimal because the distal-tip position and orientation are physically coupled. In this study, we combine the constant curvature model, Cosserat rod theory and the magnetic dipole model to efficiently generate simulation data for training a CR-specific tracking model. We also propose a physically-informed loss function that not only minimizes the estimation error but also ensures that the estimated position is coupled with the estimated orientation. We have built a tendon-driven CR with an embedded permanent magnet at the distal tip and a magnetic sensing array with 16 sensors. We evaluate the proposed loss function in comparison with the traditional loss function on simulation and collected datasets. Experimental results show that the proposed loss function performs better than the traditional one on both datasets, achieving a tracking accuracy of 0.7167 ± 0.2866 mm in position and 2.1682 ± 1.4113° in orientation on static unseen poses. In a trajectory-tracking experiment, the accuracy was 1.5845 ± 0.7463 mm and 2.5689 ± 1.8164° while maintaining an overall tracking frequency of approximately 120 Hz. Even under the most challenging tested loading condition, the position error remained 1.0916 ± 0.8168 mm, demonstrating robustness under contact-induced deformation.
Muscle fatigue, particularly in the quadriceps, can adversely affect squat mechanics and increase lumbar injury risk. Although robotic ankle-foot orthoses (AFOs) have been studied primarily for walking, their potential to assist squatting remains underexplored. This study developed and evaluated a Human-in-the-Loop optimization framework to personalize robotic AFO assistance during squatting using electromyography (EMG). Seven unimpaired male participants completed a two-day protocol including acclimation, HIL optimization, and validation. Bayesian optimization used rectus femoris EMG on the assisted leg as the cost function to tune descent and ascent assistance parameters, implemented through a stiffness-based control law. Personalized assistance reduced rectus femoris activation by 15.48% - 20.28% relative to the unpowered, no-device, and generic-assistance conditions. Vastus medialis and vastus lateralis also decreased significantly by 7.1% - 12.76% relative to all baseline conditions, while trunk and contralateral muscle activities showed no significant compensatory increases. Metabolic costs were significantly reduced by more than 11% relative to the unpowered and no-device conditions. These findings suggest that EMG-based personalization reduces knee extensor demand and lowers steady-state metabolic cost relative to the unpowered and no-device conditions during squatting.
Lung diseases cause one of the world’s biggest health concerns. Although varied types of approaches and devices have been developed to achieve diagnosis and treatment goals, it remains a big challenge for examining the lung end bronchial diseases due to the complex physiological structure of this region. Herein, we report a hybrid magnetic actuation strategy for accomplishing simultaneous examination and treatment functions for the lung end bronchus with a submillimeter multifunctional fiberscopic robot. Firstly, the principle of the proposed strategy and the development of the submillimeter multifunctional fiberscopic robot are introduced. Further, the model of the proposed hybrid magnetic actuation strategy with a fiberscopic robot is established. Next, both the high-precision performance of endoscopic imaging (error of about 2 μm) and magnetic actuation (maximum standard error of about 13.5 μm) have been investigated. After that, with the established submillimeter fiberscopic robotic system, an interventional demonstration inside a printed bronchial tree phantom is carried out and multiple functions, including scanning imaging, liquid delivery, and laser ablation, have been performed at the end bronchial region. This research paves an effective solution for the interventional control of miniature continnum robots, shining new light on early diagnosis and treatment of more hard-to-reach diseases in the human body.
Fault detection and diagnosis (FDI) in multi-degree-of-freedom (multi-DOF) robotic systems is essential for ensuring operational integrity in life-critical applications, such as robotic-assisted surgery and advanced bionics. Traditional methods often struggle with limited data sources and the masking effects of complex motion dynamics on fault localization. The theoretical innovation of this work lies in a novel, hierarchical FDI architecture that synergistically integrates frequency-domain signature modeling with bidirectional temporal learning to decouple motion-induced power fluctuations from subtle fault signals. We utilize the Bode Equation Vector Fitting (BEVF) method to precisely model non-stationary dynamic fault signatures, providing a high-fidelity reference baseline. A two-stage classifier is then employed: a Bidirectional Long Short-Term Memory (BiLSTM) network first localizes faults to a specific joint with 94.4% accuracy by exploiting bidirectional temporal dependencies in the power residuals. Subsequently, a Support Vector Machine (SVM) diagnoses the fault type (mechanical or electrical) with an overall accuracy of 76.3%. This approach successfully identifies high-impact electrical faults while capturing subtle mechanical deviations often masked by the robot's internal compensatory control loop. Our framework demonstrates a robust and non-invasive solution for FDI, significantly improving diagnostic granularity and providing actionable insights for high-reliability robotic systems.
In contrast to electromyography (EMG), which quantifies electrical activity, Forcemyography (FMG) captures mechanical deformations during contraction, making it less vulnerable to noise, electromagnetic interference, and perspiration. However, it still depends on muscle volumetric changes, sensor contact variability, and fatigue, limiting prosthetic control reliability. This paper introduces a novel human-machine interface (HMI) for transradial prosthetic devices, utilizing force-sensitive resistors (FSRs) embedded in the prosthetic socket rather than affixed to muscles, ensuring independence from muscle-related variability. Since the system relies on residual-limb movements, it ensures full independence from muscle activity. Experimental validation involving 30 healthy individuals and one subject with upper-limb impairment revealed enhanced signal stability and binary classification accuracy of 94.7% versus 82.9% for a single-channel threshold-based EMG baseline. The proposed FSR-based system demonstrated superior robustness, sustaining high classification accuracy under both perspiration and fatigue conditions, while EMG exhibited substantial performance degradation. Key contributions include: (i) a novel sensing system for binary prosthetic control based on stump rotation; (ii) a constant current-excitation circuit that linearly converts resistance changes; (iii) a microcontroller-implemented transformation algorithm optimizing FSR output; and (iv) a proof-of-concept prototype with two FSR sensors. Results validate enhanced performance and robustness of the FSR-based HMI for binary control tasks, providing a promising proof-of-concept for robust and user-friendly prosthetic control.
We propose a sensorless remote catheter control system that delivers real-time haptic feedback. Bilateral control combined with a reaction force observer estimates contact forces and transmits them to the operator in real time, while frequency-shaped feedback improves the perceived sensitivity and reduces fatigue. Using three catheter types and three operators, we evaluated position–force synchronization, loop delay, and vessel-wall load relative to manual manipulation. The average position error was approximately 2.14 × 10-5 m, and the force error was 0.493 N. Cross-correlation (≤ 1 Hz) showed no measurable lag in the low-frequency synchronization analysis within the 1 ms resolution. For force equivalence, the prespecified primary analysis employed a two one-sided tests (TOST) procedure based on estimates obtained from a linear mixed-effects (LME) model, with the smallest effect size set to ±10% of the manual mean; equivalence was supported (b = -0.029, 90% confidence interval [-0.119, 0.061]), whereas a stricter ±5% sensitivity analysis was not supported. All operators completed the tasks after minimal system familiarization. These findings indicate that the system replicates manual precision in a benchtop phantom setting while enabling real-time haptic feedback and maintaining vessel-wall loading comparable to that of manual manipulation. Our sensorless haptic technology represents a promising step toward future clinical translation of remote catheter systems, with applications in safer, standardized, and more intuitive endovascular procedures.
This study presents a skeleton-based approach for gait analysis using millimeter-Wave radar sensors. mmWave radar is non-intrusive, privacy-preserving, unaffected by lighting conditions, and both cost-effective and energy-efficient. Current radar-based gait analysis methods typically use micro-Doppler signatures to identify walking phases and extract features. In contrast, our approach leverages a pose-estimation model to reconstruct the human skeleton from radar point cloud data, enabling comprehensive full-body analysis. This enables a more intuitive and detailed gait analysis while also allowing for direct signal-level comparisons with wearable sensor-based approaches. In our study, we recruited 78 participants and conducted gait tests across four distinct environments to evaluate the effectiveness of the proposed method. The results showed that gait features, including stride time, stride length, and stride velocity, exhibited good to excellent Intraclass Correlation Coefficients (ICCs) when compared with wearable sensors. Additionally, we are the first to analyze sub-phase features such as swing, stance, and double support using a radar system. Our findings show that mmWave radar sensors can accurately capture stride-level gait features, while their performance is less effective for detailed sub-phase analysis. This study underscores the significant potential of mmWave radar for gait analysis in older adults, providing a low-cost, non-contact, and privacy-preserving solution.
The most commonly prescribed ankle/foot prostheses are energy storage and return (ESR) devices, which have good performance in level ground walking, but have limited functionality in tasks requiring ankle mobility, such as stair ascent or descent. Powered prostheses aim to address this limitation by actively controlling the joint with actuators, but they tend to be heavier and taller and become unusable when the battery is depleted, hindering their clinical adoption. In this paper, we present a semi-active ankle prosthesis that integrates the energy storage and return characteristics of a passive prosthesis with the active joint articulation of a powered prosthesis. The proposed prosthesis combines a novel brake mechanism with a series elastic actuator to enable Passive and Active operation modes within a compact build height (141 mm) and lightweight (1.4 kg without batteries) form factor. Experimental validation with one lower-limb amputee demonstrates that the proposed prosthesis reproduces the gait dynamics of a passive prosthesis, independent of battery charge, in Passive mode, and matches the range of motion and actuation capabilities of a powered device during stair ascent and sit-to-stand transitions in Active mode. Automatic transitions between modes within a gait cycle allow for incline adaptation. Moreover, automatic transitions between gait cycles allow for seamless transitions between passive walking and active stair ambulation. This combination of functionalities has never been demonstrated in an ankle prosthesis and has the potential to improve mobility among the amputee population.
The population of individuals with mobility impairments has been significantly increasing due to aging, diseases, or accidents. Gait analysis plays a crucial role in investigating musculoskeletal functions and gait parameters, enabling the assessment of mobility diseases, and providing valuable data for optimizing physical therapy treatments. Smart walkers can not only provide physical support to patients but also monitor the user’s walking gait. In this research, force sensors and built-in encoders are used to extract various gait parameters. Machine learning techniques, which have the ability to adapt to users with different gait patterns, are used to analyze users’ gait through interaction force. Three machine learning models, Support Vector Machine, Random Forest, and Neural Network are selected to perform gait analysis. Experiments with 5 subjects are conducted to evaluate the performance of the machine learning models in gait parameters estimation. Results from this pilot study show that the accuracies of the three machine learning models are all above 90% and the performance is much better than the traditional method and Random Forest model has slightly better accuracy among the small sample-size populations. In addition, SVM has the fastest response with 1.2 ms prediction time, which is suitable for online evaluation.
