Freezing of gait (FOG), a debilitating symptom of Parkinson's disease, can manifest in three sub-types: shuffling, trembling, and akinesia, with occurrence and frequency varying across patients. While deep learning (DL) models show promise in FOG detection, their robustness and generalization across subtypes are limited by data scarcity and imbalances between FOG/non-FOG classes and among subtypes. To address this, we propose a subtype-aware FOG augmentation technique enabling training of DL models to perform consistently across subtypes. Specifically, we introduce Hierarchical Coarse-to-Fine conditional Generative Adversary Network (Hi-CF cGAN), a two-stage model that generates subtype-conditioned FOG-like ankle accelerations that are realistic and diverse, as verified through visualization, UMAPs, and Maximum Mean Discrepancy comparison against real signals. We evaluate its effective-ness by training CNNs for FOG detection with both general (subtype-stratified) and personalized (subtype-variant, based on patient-specific subtype composition) augmentation via Hi-CF cGAN, benchmarking against classical augmentations and baseline (no augmentation). Compared to baseline, general augmentation with Hi-CF cGAN effectively improves average detection rates of FOG, trembling FOG, and especially the previously overlooked minor subtypes, shuffling FOG (from 66.8% to 81.6%) and akinesia FOG (from 58.7% to 77.9%). These improvements exceed those of classical augmentations, demonstrating superior real-ism, richness, and adaptability of Hi-CF cGAN-generated data in addressing FOG/non-FOG and subtype imbalances. Personalized augmentation further enhances accuracy on targeted subtype(s) compared to general augmentation, highlighting its potential for tailored model optimization.

This study presents and experimentally validates an adaptive control method for human-exoskeleton interaction through online adaptation of desired joint trajectories. Leveraging gait phase and human-exoskeleton interaction torque estimators, our approach enables seamless assistance adaptation to varying walking patterns and speeds. Specifically, a pretrained neural network approximates the exoskeleton's dynamics, enabling real-time interaction torque estimation from kinematic measurements and commanded motor torques alone. These estimates drive a gradient-descent update of the joint reference trajectories, minimizing a cost function that penalizes both interaction torques and trajectory modification, ensuring bounded convergence and stability without user-specific parameter tuning. We compared our adaptive controller with a fixed-trajectory gait-phase-based controller during overground and treadmill walking at three self-selected speeds ranging from 0.4 to 0.8 m/s. In 16 participants, the adaptive controller significantly reduced the hip and knee interaction torques by 51.2% +/- 11.1 and 63.9% +/- 29.7, respectively, during over-ground walking. Muscular effort significantly decreased in Bicep Femoris (21.0% +/- 34.5) and Rectus Femoris (28.1% +/- 34.6), while remaining unchanged in other muscles. Cadence and gait speed increased by 7.6% +/- 5.2 and 10.7% +/- 8.3, respectively, indicating that participants could walk faster with less effort due to trajectory adaptation. Postadaptation trajectories more closely resembled those of walking without the exoskeleton, and exoskeleton-torques aligned more closely with human biological torques. Our proposed adaptive controller, which requires only exoskeleton kinematics, also maintained performance during treadmill walking across speeds, demonstrating speed-invariant behavior compared to the nonadaptive controller.

Lower-limb exoskeletons offer high-dose gait rehabilitation, but their effectiveness depends on accurate coordination between the user and robotic controller. We propose a real-time method to estimate and represent the user’s intended motions from hip and knee joint angle feedback. A kinematic feature vector is obtained through a processing pipeline including gait phase estimation, standardization, and recursive fitting of a von Mises basis function model to the hip and knee flexion angles. We show that the basis functions coefficients form a rich feature vector for classifying and detecting changes in gait patterns, while also implicitly providing real-time maximum likelihood estimates of the joint kinematics over a gait cycle, independent of user’s current gait phase within the cycle. The algorithm was evaluated on 16 participants across 16 treadmill and overground walking conditions, demonstrating accurate reconstructions of joint kinematics. We show that gait cycle kinematics are accurately modeled by the von Mises basis model, with coefficients of determination above ${0}.{90} \pm {0}.{03}$ . The feature vector varies meaningfully across different walking conditions, enabling efficient real-time detection of changes, such as asymmetric limping.

The Velocity Flow Field (VFF) lower-limb exoskeleton controller is widely applicable for gait rehabilitation because it provides the user with considerable agency over their gait; however, previous studies reported the feeling of "walking through water", and resistance to the user's efforts. In this work, a mathematical explanation for the viscous damping behavior when users deviate from the reference trajectory is presented. The controller was corrected and an adaptation law is proposed that synchronizes the speed gain with the user's current walking speed by minimizing the average mechanical work transferred between the user and exoskeleton per step. Experiments comparing a fixed and adaptive controller with 12 participants walking at 0.4 +/- 0.1 body length/s on a treadmill showed that the adaptive controller tracks changes in walking speed, while reducing the energy absorbed by 0.589 +/- 0.126 J/step compared to the fixed controller at the fastest walking speed. Analysis of changes in muscle effort and interaction torques with a human-exoskeleton interaction portrait showed that for most participants, the adaptive controller at medium and fast speeds substantially reduced user-controller disagreement and increased user agency over the walking motion. These positive results suggest that optimizing the energy supplied per step can serve as an effective coordination mechanism, enabling personalized and real-time adjustments of walking speed between the user and the exoskeleton.

Recent advancements in machine learning and computer vision enable direct prediction of mechanical properties from microstructure images. The feasibility of this process hinges on the material structure-property relationship, richness of the dataset, and the choice of machine learning approach. This study investigates the application of a deep learning model to directly predict the yield strength (YS), ultimate tensile strength (UTS), and true stress-strain curve of the cast-forged AZ80 alloys from SEM microstructure images. We manufactured 27 cast-forged AZ80 magnesium alloy components using varied process parameters, creating a diverse dataset of AZ80 microstructures and mechanical properties through their characterization. In addition to predicting magnesium alloy properties, we address challenges related to data imbalance, brightness and contrast variability, and microstructure long-range heterogeneity. We demonstrate that synthetic data oversampling using a denoising diffusion probabilistic model effectively improves the model's prediction accuracy via balancing the minority classes. A rigorous analysis of the model's performance shows that the model accurately predicts the YS, UTS, and Ramberg-Osgood equation's parameters ( K and n ). In image-out validation, the model achieves average percentage errors of 2.10 % (YS), 2.15 % (UTS), 1.50 % ( K ), and 5.47 % ( n ). In class-out validation, the errors are 6.27 %, 9.58 %, 4.69 %, and 10.24 %, respectively. (c) 2025 Chongqing University. Publishing services provided by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/)

Background Functional electrical stimulation (FES) delivers transcutaneous electrical current to motor nerves to artificially evoke muscle contractions and produce joint torque. In neurorehabilitation, FES is commonly applied via a large surface electrode pair placed over a muscle group – an approach that we here refer to as single-electrode stimulation (SES). However, the torque-generating capacity of SES is limited. Targeting specific regions of high electrical excitability, or motor points, may enhance neuromuscular activation and increase joint torque output. Here we tested whether continuous multiple motor point stimulation (mMPS) increases tetanic knee torque compared with SES in the quadriceps and hamstrings. We also quantified quadriceps recruitment overlap as a secondary aim. We evaluated quadriceps and hamstring responses to mMPS and SES in neurologically intact participants, recording tetanic knee torque and assessing recruitment overlap relative to maximal voluntary contraction (MVC). Results Stimulating all quadriceps motor points produced approximately 51% greater knee extension torque than SES (33.3% MVC vs. 22.1% MVC). In contrast, stimulating all hamstring motor points did not significantly increase knee flexion torque. Within the quadriceps, proximal and intermediate vastus lateralis motor points contributed the most to knee extension torque, whereas the distal motor point had no significant effect. Discrepancies between observed and predicted MPS torques were primarily associated with rectus femoris stimulation, while vastus medialis contributed the least. Conclusions Stimulation of all quadriceps motor points generated 51% greater knee extension torque compared with conventional SES, highlighting the potential of mMPS to improve the effectiveness of FES interventions. Optimization of motor point selection should emphasize proximal and intermediate vastus lateralis sites, while also including vastus medialis. Rectus femoris stimulation should be considered only when channel allocation permits and when the torque benefit outweighs redundancy. In contrast, hamstring stimulation did not yield additional knee flexion torque underscoring muscle-specific limitations. These findings provide practical guidance for electrode placement and channel allocation, suggesting that optimal quadriceps stimulation can be achieved by prioritizing a limited subset of motor points.

Lower-limb exoskeletons have demonstrated great potential for gait rehabilitation in individuals with motor impairments; however, maintaining human-exoskeleton coordination remains a challenge. The coordination problem, referred to as any mismatch or asynchrony between the user's intended trajectories and exoskeleton desired trajectories, leads to sub-optimal gait performance, particularly for individuals with residual motor ability. Here, we investigate the virtual energy regulator (VER)'s ability to generate coordinated locomotion in lower limb exoskeleton. Contribution: (1) In this paper, we experimented VER on a group of nine healthy individuals at different speeds (0.6m/s - 0.85m/s) to study the resultant gait coordination and naturalness on a large group of users. (2) The resultant assisted gait is compared to the natural and passive (zero-torque exoskeleton) walking conditions in terms of muscle activities, kinematic, spatiotemporal and kinetic measures, and questionnaires. (3) Moreover, we presented the VER's convergence proof considering the user contribution to the gait and introduced a metric to measure the user's contribution to gait. (4) We also compared VER performance with the phase-based path controller in terms of muscle effort reduction and joint kinematics using three able-bodied individuals. Results: (1) The results from the VER demonstrate the emergence of natural, coordinated locomotion, resulting in an average muscle effort reduction ranging from 13.1% to 17.7% at different speeds compared to passive walking. (2) The results from VER revealed improvements in all indicators towards natural gait when compared to walking with a zero-torque exoskeleton, for instance, an enhancement in average knee extension ranging from 3.9 to 4.1 degrees. All indicators suggest that the VER preserves natural gait variability and user engagement in locomotion control. (3) Using VER also yields in 13.9%, 15.1%, and 7.0% average muscle effort reduction when compared to the phase-based path controller. (4) Finally, using our proposed metric, we demonstrated that the resultant locomotion limit cycle is a linear combination of human-intended limit cycle and the VER's limit cycle. These findings may have implications for understanding how the central nervous system controls our locomotion.

The velocity vector field (flow) controller is a well-established control strategy for lower limb exoskeletons. In this paper, we analyze this controller and propose modifications to improve its performance. We demonstrate that flow control acts as a variable proportional-derivative error regulator, where the parameter Γ represents the desired norm of the hip-knee joint velocity vector (path speed). Based on this, we introduce two modifications to Γ: (1) a constant Γ set to the mean desired path speed, and (2) a variable Γ that mimics natural path speed during unassisted walking. We compared the modified flow controllers with a slow-Γ version in experiments involving seven participants walking on a treadmill at 0.6 m/s , 0.8 m/s , and 1.0 m/s . Compared to the slow-Γ controller, the RMS tracking error decreased by 30.7 ± 11.3% and the range of motion of the knee increased by 48.2 ± 5.5% for the mean-Γ controller, while the variable-Γ controller had 32.4 ± 14.7% smaller RMS error and 50.5 ± 6.5% larger range of motion of the knee. Additionally, the slow-Γ controller consistently applied resistive power, whereas participants reported more comfortable and natural gait with the modified controllers. We also compared them with the original tuning of flow controller, with results indicating superior performance from the proposed modifications. These findings demonstrate effectiveness across different walking speeds and offer a tuning strategy for future flow controller use. ### Competing Interest Statement The authors have declared no competing interest. NSERC, Discovery - RGPIN 2018 Canada Foundation for Innovation, https://ror.org/000az4664 Ontario Research Foundation, https://ror.org/051639p85 New Frontiers in Research Fund (NFRF), NFRFE-2018-01698, NFRFE 2022-620

This work presents a real-time gait phase estimator using thigh- and shank-mounted inertial measurement units (IMUs). A multi-rate convolutional neural network (CNN) was trained to estimate gait phase for a dataset of 16 participants walking on an instrumented treadmill with speeds varying between 0.1 to 1.9 m/s, and conditions such as asymmetric walking, stop–start, and sudden speed changes. One-subject-out cross-validation was used to assess the robustness of the estimator to the gait patterns of new individuals. The proposed model had a spatial root mean square error of 5.00±1.65%, and a temporal mean absolute error of 2.78±0.97% evaluated at the heel strike. A second cross-validation was performed to show that leaving out any of the walking conditions from the training dataset did not result in significant performance degradation. A 2-sample Kolmogorov–Smirnov test showed that there was no significant increase in spatial or temporal error when testing on the abnormal walking conditions left out of the training set. The results of the two cross-validations demonstrate that the proposed model generalizes well across new participants, various walking speeds, and gait patterns, showcasing its potential for use in investigating patient populations with pathological gaits and facilitating robot-assisted walking.

This work investigates the optimal cost function composition for human gait at different walking speeds. Kinematic and kinetic data for walking at four walking speeds were collected from five individuals without any known disability. The data was then used to recover optimal cost functions in a predictive simulation environment with musculoskeletal models. Twenty inverse optimal control (IOC) problems were solved for cost function weight tuning using the previously developed and validated Adaptive Reference IOC (AR-IOC) algorithm. Given the walking speed range examined (0.6–1.5 m/s), the converged cost function weights suggest that the increase in walking speed attributes to a reduction of foot sliding penalty weight and weight increase for the center of mass (CoM) acceleration and stability as confirmed by several experiments. Furthermore, we did not observe any significant weight shift in effort reduction between the upper and the lower body with respect to walking speed. The obtained results from this study can be used in a toolbox for obtaining subject- and task-specific cost functions and assisting the development of personalized rehabilitation technologies.

This paper presents experimental validation of a recently proposed Assist-As-Needed controller which utilizes human-robot interaction force to adapt the exoskeleton's reference trajectory. Using the estimated dynamics of Indego lower-limb exoskeleton in single and double support regimes, we accurately estimate the human-exoskeleton interaction torques all over the gait cycle. This enables the exoskeleton to adapt its desired joint angle trajectories by minimizing its physical interaction with the user, thereby resolving the human-exoskeleton physical disagreement. The results of our experiments on two able-bodied participants show that the proposed reference trajectory adaptation method can boost performance metrics such as assistance efficiency and assistance coordination with user intention. Moreover, we observed reductions in muscular activity, with decreases ranging from 13% to 54%, across muscles acting on the hip or knee joints. While direct measures of gait stability were not assessed, the proposed trajectory adaptation is associated with improved foot clearance metrics, specifically minimum toe and maximum heel clearance. This improvement, coupled with a reduction in users' muscle activity, may suggest enhancements in perceived stability compared to using a fixed desired trajectory for controlling the exoskeleton.

Intermittent pneumatic compression (IPC) systems apply external pressure to the lower limbs and enhance peripheral blood flow. We previously introduced a cardiac-gated compression system that enhanced arterial blood velocity (BV) in the lower limb compared to fixed compression timing (CT) for seated and standing sub7 jects. However, these pilot studies found that the CT that maximized BV was not constant across individuals and could change over time. Current CT modelling methods for IPC are limited to predictions for a single day and one heartbeat ahead. However, IPC therapy for may span weeks or longer, the BV response to compression can vary with physiological state, and the best CT for eliciting the desired physiological outcome may change, even for the same individual. We propose that a deep reinforcement learning (DRL) algorithm can learn and adaptively modify CT to achieve a selected outcome using IPC. Herein, we target maximizing lower limb arterial BV as the desired out19 come and build participant-specific simulated lower limb environments for 6 participants. We show that DRL can adaptively learn the CT for IPC that maximized arterial BV. Compared to previous work, the DRL agent achieves 98% ± 2 of the resultant blood flow and is faster at maximizing BV; the DRL agent can learn an “optimal” policy in 15 minutes ± 2 on average and can adapt on the fly. Given a desired objective, we posit that the proposed DRL agent can be implemented in IPC systems to rapidly learn the (potentially time-varying) “optimal” CT with a human-in-the-loop.

Despite the increased use of lower limb exoskeletons as gait training and mobility assistive devices, their controllers often lack the ability to synchronize and adapt to meet individual users' needs. This letter investigates two control approaches for lower limb exoskeletons: a real-time kinematic state-dependent estimation of desired torques with an inverse dynamics model and a data driven component in the first approach, and a pre-defined torque control based on gait speed and phase in the second approach. These controllers are linearly combined to shift the controller behavior between pure kinematic state-dependent and pure gait phase-dependent control. These combinations were tested during overground and treadmill walking with nine able-bodied participants. The linearly combined controller with a greater emphasis on kinematic state-dependent control produced a more natural gait in terms of spatiotemporal metrics. This is reflected by 0.1 m/s increases in overground walking speed and 5% decrease in percent stance compared to walking with a passive exoskeleton. This controller also decreases the overall activity of lower limb muscles by up to 25% and thigh co-contractions by up to 40%. Participant feedback through a questionnaire, in terms of perceived effort, walking naturalness, and stability, also favored the aforementioned controller.

Human-robot physical interaction contains crucial information for optimizing user experience, enhancing robot performance, and objectively assessing user adaptation. This study introduces a new method to evaluate human-robot co-adaptation in lower limb exoskeletons by analyzing muscle activity and interaction torque as a two-dimensional random variable. We introduce the Interaction Portrait (IP), which visualizes this variable's distribution in polar coordinates. We applied this metric to compare a recent torque controller (HTC) based on kinematic state feedback and a novel feedforward controller (AMTC) with online learning, proposed herein, against a time-based controller (TBC) during treadmill walking at varying speeds. Compared to TBC, both HTC and AMTC significantly lower users' normalized oxygen uptake, suggesting enhanced user-exoskeleton coordination. IP analysis reveals this improvement stems from two distinct co-adaptation strategies, unidentifiable by traditional muscle activity or interaction torque analyses alone. HTC encourages users to yield control to the exoskeleton, decreasing muscular effort but increasing interaction torque, as the exoskeleton compensates for user dynamics. Conversely, AMTC promotes user engagement through increased muscular effort and reduced interaction torques, aligning it more closely with rehabilitation and gait training applications. IP phase evolution provides insight into each user's interaction strategy development, showcasing IP analysis's potential in comparing and designing novel controllers to optimize human-robot interaction in wearable robots.

Inverse dynamics is a common tool for determining human joint torques during walking. The traditional approaches rely on ground reaction force and kinematics measurements prior to analysis. A novel real-time hybrid method is proposed in this work by integrating a neural network and dynamic model that only requires kinematic data. An end-to-end neural network for direct joint torque estimation is also developed based on kinematic data. The neural networks are trained on a variety of walking conditions, including starting and stopping, sudden speed changes, and asymmetrical walking. The hybrid model is first tested in a detailed dynamic gait simulation (OpenSim) which results in root mean square errors less than 5 N.m and a correlation coefficient of greater than 0.95 for all the joints. Experiments demonstrate that the end-to-end model on average outperforms the hybrid model across the whole test when compared to the gold standard approach which requires both kinetic and kinematic information. The two torque estimators are also tested on one participant wearing a lower limb exoskeleton. In this case, the hybrid model (R > 0.84) has significantly better performance than the end-to-end neural network (R > 0.59). This indicates that the hybrid model is better applicable to scenarios which differ from the training data.

Intermittent pneumatic compression (IPC) systems are employed to treat vascular diseases. It has been shown that applying cardiac-gated compression effectively enhances femoral blood velocity (BV), but the optimal compression timing likely varies between individuals and may vary over time. While a previous work has shown the usability of one heartbeat ahead BV estimation to optimize the compression timing, that study was limited to a single treatment session and the BV estimator performance may deteriorate for the next sessions. Therefore, the goal of this study is to develop BV estimators and evaluate their accuracy over a longer time-scale. Six participants wore a custom IPC system and experienced random cardiac-gated compression timings for 1.5 hours per day for two days. A data- driven model was trained on electrocardiogram and applied pressure data to predict femoral BV one heartbeat ahead in a closed loop manner. The mean R 2 for this model across participants on the second session was 0.74 ± 0.09 and the mean absolute error was approximately 3%, which is a reduction of only 11% compared to the first sessions, for both metrics. This study is the first to show that BV across IPC sessions can be predicted using a pre-trained model. This work may lead to a significant improvement in IPC performance with only an initial model training session.

This paper investigates the effects of the cast-forged process on the hardness distribution of the AZ80 magnesium alloy. AZ80 material is cast at four different cooling rates and die-forged at three forging temperatures. The microstructural characteristics of & alpha;-Mg grain size, Mg17Al12 phase dissolution and precipitation, and dynamic recrystallization are critical determinants of the material hardness. The processing parameters influence and provide controllability over these microstructural characteristics. Artificial neural network models are developed to predict the hardness based on a given combination of cooling rate, forging temperature, and location. This data-driven model is then used to predict the hardness values for all the measurement points on an unseen cooling rate and forging temperature, producing the predicted hardness contour maps. The model accurately captures the relationship attributed to the effect of processing parameters on the evolution of several different microstructure features. The prediction of the hardness distribution results in the percentage of the average hardness errors of 1.44 % & PLUSMN; 0.82 and 2.05 % & PLUSMN; 1.25 for one-I-beam-out and one-condition-out scenarios, respectively. The model's average normalized root mean squared error values for one-I-beam-out and onecondition-out predictions are 0.0586 & PLUSMN; 0.014 and 0.0640 & PLUSMN; 0.016, respectively. The predicted contour maps accurately predict the hardness distribution of unseen cast-forged conditions and resemble the hardness variations in the actual cast-forged components.

Careful consideration of the preform shape is essential when designing a forging process. A high-quality forging process must promote process-related grain refinement and an unbroken grain flow devoid of cavities or folding, in addition to minimizing the amount of generated flash while achieving a complete die fill. The desired forged part properties can be obtained by optimizing the preform shape. However, three-dimensional shape optimization presents challenges in design generation and design evaluation due to the resource-intensive demands of each task. To address these challenges, we propose a multi-objective optimization framework consisting of a parametric computer-aided design (CAD) model for shape generation, data-driven models for shape evaluation, and a multi-objective evolutionary optimization algorithm to search the design space effectively. This computational framework is used to evolve an optimal preform shape, which was ultimately cast using permanent mold casting (PMC) and then hot forged under elevated temperature conditions. We compared the forging outcome of the optimal preform with a baseline cylindrical billet which was produced according to the same sequence of manufacturing steps. Comparative analysis of the laboratory-scale forging results revealed that the cast-preform and cast-billet produced about 6% and 12% flash material, respectively. Quasi-static tensile and stress-controlled cyclic tests were also conducted to evaluate mechanical properties. While comparable yield and ultimate tensile strengths were observed in both forgings, a significant increase in fracture strain was observed in the preform forging, suggesting improved toughness. In general, the forging outcome of the optimized preform proved to be superior to the billet forging.

Many daily activities involve bimanual coordination while simultaneous movement of the hands may result in interference with their movements. Here, we studied whether the central nervous system could use the relevant information in composite feedback, i.e., a weighted sum of positional information of nonvisible and visible hands, to improve the movement of the nonvisible hand. Our results suggest the ability to decode and associate task-relevant information from the composite feedback.

The rapid progress in artificial intelligence (AI) based image generation led to groundbreaking achievements, like OpenAI's DALL-E 2, showcasing state-of-the-art generative models in deep learning and computer vision. Recently, the Denoising Diffusion Probabilistic Model (DDPM) has emerged as a strong contender, excelling in generating high-resolution images with complex features similar to those found in real-world images. In this study, we investigate DDPM's potential as both generator and predictor of scanning electron microscope (SEM) images, encompassing both known and unseen microstructural conditions. To rigorously evaluate DDPM, we curated a comprehensive dataset comprising 27 distinct cast-forged AZ80 magnesium alloy components with varied process parameters and microstructure features. Some conditions were held back during training to test DDPM's predictive abilities for unseen scenarios. Our study demonstrates the model's remarkable capacity to capture the inherent physical relationships between process parameters and microstructure features. We scrutinize the synthesized images alongside real-world SEM counterparts, undertaking a comprehensive analysis of various morphological properties. Remarkably, the results show the model's performance, with an average error of 6.36 % +/- 0.42 for measured microstructural properties in seen conditions and an equally impressive 6.67 % +/- 0.85 for unseen conditions. This study envisions a transformative shift in materials science, as advanced AI predictive models offer new potential to streamline the laborious process of microstructure image generation.