Biomechanical energy harvesting to improve post-stroke mobility

Abstract

Foot drop is a common post-stroke gait impairment that reduces toe clearance and compromises walking safety. Existing assistive interventions are often limited by passive compensation, external power requirements, and manual tuning. Here, we introduce a wearable, gait-timing-adaptive, self‑powered neuroprosthesis (SPN) that augments ankle dorsiflexion and corrects gait patterns, avoiding labor-intensive parameter tuning and eliminating the charging and reliability burdens of external batteries. The SPN leverages gait’s out‑of‑phase mechanics: it harvests mechanical energy during stance on the non‑paretic limb and immediately converts it into electrical stimulation of the tibialis anterior on the paretic limb during swing, increasing dorsiflexion when foot clearance is required. In ambulatory stroke survivors with preserved walking capacity, foot drop, and lower-limb impairment (Fugl-Meyer LE Score: 13-27/34, Modified Ashworth Scale-Ankle Plantarflexors: 0-2), the SPN assistance produced immediate, clinically meaningful gains during outdoor walking: dorsiflexion angle improved from −1.1° to 6.8°, walking distance increased by 67.3%, speed by 43.5%, and gait variability decreased by 71.9%. These findings demonstrate the SPN’s potential to restore dorsiflexion, enhance mobility, and improve overall quality of life for individuals recovering from stroke, particularly those at later stages of life.

Foot drop after stroke reduces toe clearance and walking safety. Here, Pan et al. report a gait-timing-adaptive, self-powered neuroprosthesis that harvests gait energy to augment ankle dorsiflexion, improving outdoor walking distance and speed, and reducing gait variability in stroke survivors.

Introduction

1 in 4 people will experience a stroke in their lifetime¹, making stroke a leading cause of long-term disability worldwide² with profound personal and societal burdens³. Beyond acute neurological injury, many survivors face chronic motor impairments that limit independent mobility. Among these, foot drop—an inability to dorsiflex the ankle during swing due to impaired corticospinal drive and weakened dorsiflexors—is common and disabling. It reduces gait efficiency and stability⁴, increases fall risk⁵, and diminishes quality of life⁶, underscoring the need for practical and restorative interventions for post-stroke foot drop⁷.

Conventional ankle–foot orthoses (AFOs) are commonly prescribed aids for foot‑drop correction after stroke⁸. However, their compensatory effect is largely achieved by mechanically constraining ankle joint motion rather than restoring physiological dorsiflexion and toe clearance^9,10. This restriction not only limits natural gait kinematics but, with prolonged use, may also contribute to muscle disuse and secondary atrophy¹¹. Although robotic exoskeletons have been proposed as more active alternatives^12,13, their dependence on powered actuators, complex controllers, and expert calibration likewise limits portability and day‑to‑day applicability—especially for users with cognitive or motor deficits. Moreover, long‑term reliance on externally actuated devices can further reduce voluntary muscle engagement, potentially counteracting the goals of rehabilitative training.

Beyond mechanical assistive devices, other strategies, including pharmacological^14,15, surgical¹⁶, and behavioral¹⁷ interventions, have generally provided only modest and often transient improvements. A key limitation is that these approaches typically do not provide dynamic, gait-phase-specific assistance during walking, even though effective correction of foot drop requires timely restoration of ankle dorsiflexion during swing. Functional electrical stimulation (FES) has therefore emerged as a more physiological method to facilitate ankle dorsiflexion by activating the peroneal nerve during gait. Traditional FES systems, however, commonly rely on batteries or routine charging, as well as sensor-based gait-phase detection and algorithmic control. These requirements complicate setup, increase system complexity, and limit usability in daily-life settings¹⁸.

To overcome these limitations, we propose a self‑powered neuroprosthesis (SPN) that facilitates dorsiflexion through phase‑synchronized peroneal nerve stimulation powered entirely by the wearer’s own gait. Unlike conventional FES systems, the SPN does not require routine battery charging or separate gait sensors and control algorithms. Instead, it couples energy harvesting and stimulation delivery through reciprocal gait mechanics. This architecture enables phase-appropriate assistance during walking while reducing operational burden and maintenance demands, thereby improving the practicality of FES-based foot-drop assistance for stroke survivors. Specifically, biomechanical energy harvested during stance on the non-paretic limb is buffered and released near paretic toe-off to stimulate the tibialis anterior during swing, thereby facilitating ankle dorsiflexion and toe clearance (Fig. 1a–c). The system integrates three components: a heel-embedded energy harvester and a miniature power-management module on the non-paretic shoe, and a lightweight below-knee neurostimulator secured around the paretic calf with Velcro straps equipped with electrodes (Fig. 1b, d). By leveraging the natural out-of-phase coordination of bilateral gait, the SPN provides battery-free, gait-synchronized assistance that supports, rather than restricts, natural limb movement (Fig. 1e, f).

Fig. 1. Overview of the SPN.

a Schematic of SPN operation. b SPN configuration for post-stroke foot drop. Energy harvested on the non‑paretic side (N) powers stimulation of the tibialis anterior on the paretic side (P) to prevent foot drop. c Stimulation timeline of the SPN. Electrical stimulation is delivered near toe-off (dashed line). The gait cycle is time-normalized with touchdown at 0%. Note that the x-axis in Fig. 1c is time-normalized to the gait cycle (0-100%); the stimulation temporal position relative to the gait phases is further illustrated in Fig. 2a with phase-specific (stance/swing) labelling. d Fabricated SPN prototype. e Out-of-phase workflow. During late stance on the non‑paretic side limb, electromagnetic generation produces electrical energy that is conditioned by the power management circuit (PMC) and pulse-width modulation (PWM) modules into a biphasic waveform with the required frequency and amplitude. This drives tibialis anterior contraction on the paretic side, assisting ankle dorsiflexion and improving foot clearance. f Rehabilitation cycle of the SPN for post-stroke individuals.

Results

Overview of the SPN

The wearable SPN is a battery-free, lightweight, gait-adaptive FES neuroprosthesis (Fig. S1) designed to assist ankle dorsiflexion after stroke. Rather than relying on external sensors or battery-powered control, the device operates automatically during walking through mechanically coupled energy harvesting and stimulation delivery. This design enables gait-synchronized assistance in a compact, hard-wired format suitable for overground use.

Operating mechanism

A gravity-driven, gear-based switch-rail mechanism synchronizes mechanical energy harvesting on the non-paretic limb with bilateral gait dynamics (Fig. S2 and Movie S1). During early-mid stance on the non-paretic limb, heel loading depresses the harvester cover and disengages the generator from the switch rail (Fig. S2b). As the heel lifts during late stance, a spring raises the cover to re-engage the generator (Fig. S2c), thereby producing electrical power for paretic-side stimulation during the swing phase (Movie S2). The energy harvester weighs 56 g and, after heel-material removal, adds < 20 g net to the shoe. Its minimal distal mass and automatic operation allow post-stroke individuals to don the device within 2 min, ensuring seamless gait integration.

Energy harvesting and phase-adaptive stimulation

To individualize FES timing without external sensors or batteries, we exploited the out-of-phase nature of gait by embedding an energy harvester in the heel of the non-paretic shoe (Fig. 2a). As walking speed varied, the net power delivered to the SPN ranged from 0.16 to 1.2 W (Fig. 2b), with the corresponding rectified voltage profiles shown in Fig. S3. Across different gait conditions, the peak net power delivered to the SPN was 0.35 ± 0.08 W during uphill walking, 0.27 ± 0.13 W during downhill walking, 0.49 ± 0.23 W during stair ascent, and 0.41 ± 0.25 W during stair descent (Fig. 2c), with the corresponding rectified voltage waveforms shown in Fig. S4. Under these various conditions, the net power supplied to the FES module ranged from approximately 0.16 to 1.2 W (Fig. 2b, c), depending on walking speed and gait pattern (Movie S3). This power range indicates sufficient electrical availability to sustain pulsed neuromuscular stimulation (e.g., 35 Hz, 300 µs) across walking conditions, although the exact demand depends on stimulation parameters, electrode-tissue impedance, waveform, and circuit efficiency^19,20. Energy captured during late stance on the non-paretic limb powers stimulation of the paretic tibialis anterior during its subsequent swing phase (Fig. 2d), assisting dorsiflexion for improved toe clearance and stable foot placement. Each stimulation event draws down the capacitor voltage (Fig. 2d, e); the following non-paretic step recharges it to operating level, yielding a cyclical “one step, one stimulation” assist profile (Fig. 2d).

Fig. 2. Energy harvesting and electrical stimulation coupling.

a Representative gait pattern of the footdrop in stroke patients and the power generation phase for the electrical stimulation with the out-of-phase nature. b, c N = 6. Bars indicate the mean, and error bars represent the standard deviation (SD). Data are presented as mean ± SD. b Net power delivered to the SPN at different walking speeds. c Net power delivered to the SPN under various gait conditions. d Regulated voltage output of the energy harvester under the “one step, one stimulation” mode. Each non‑paretic side step charges the storage module to deliver a stimulation pulse to the paretic side; after stimulation, the stored voltage drops, then recharges on the next step, repeating cyclically. e The charging curve of a 22 mF capacitor as the wearer begins walking from a stationary position. At approximately 46 s, the voltage suddenly drops due to a transient gait pause. f Medium intensity stimulation voltage profile of the neuroprosthesis during 60 s of continuous walking at a speed of 0.8 m/s. g Power density of the proposed energy harvester compared with state-of-the-art footwear harvesters based on piezoelectric^59–65, triboelectric^66–71, electromagnetic^72–79, and hybrid^80–85 mechanisms.

Stimulation parameters—frequency, amplitude, and pulse width—were tuned to control motor-unit recruitment and adjust dorsiflexion torque (Fig. 2f). For comfort and safety, stimulation voltage was capped at 240 V²¹. The signal-generation electronics include a voltage regulator, booster, pulse-width-modulation (PWM) generator, and biphasic output stage (Figure S5a–c), with biphasic waveform dynamics shown in Figure S5d–f. Detailed energy-harvesting performance and the electrical output characteristics at each stage are presented in Figs. S3, S4, Figs. S6, S7 and Note S1.

A multi-stage frequency-up-conversion gearbox, coupled with the gravity-driven switch-rail mechanism (Fig. S2b), amplifies the rotor’s rotational speed under low-frequency heel-strike excitation. This configuration increases generator output and achieves power density approximately one order of magnitude higher than prior wearable harvesters (Fig. 2g and Table S1). The resulting energy suffices to support robust “one step, one stimulation” operation even during asymmetric, irregular, low-frequency stroke gait, ensuring consistent dorsiflexion assistance without external power sources. The SPN therefore unites energy harvesting and neuromuscular stimulation in a compact, battery-free device that autonomously synchronizes with gait. By channeling mechanical energy from the non-paretic heel (late stance) into real-time paretic dorsiflexor activation, it achieves precise, phase-appropriate functional assistance in post-stroke walking.

Experimental validation

To evaluate the SPN performance, we conducted both laboratory-controlled tests and real-world community walking trials with stroke participants.

Laboratory-controlled tests

To elucidate the underlying biomechanical mechanisms of the SPN, we assessed gait biomechanics in 8 stroke participants (Movie S4) during treadmill walking. Participants walked at self-selected constant speeds (Table S2) while instrumented with a 3D motion-capture system and wearable ultrasound to quantify joint kinematics and tibialis anterior muscle dynamics (Fig. 3a and Movie S4). With SPN assistance, the tibialis anterior stimulation timing profile more closely resembled that of healthy gait, with a 147.8% increase in tibialis anterior activity during terminal swing (Fig. 3b, c). Ultrasound revealed a small increase in dorsiflexor muscle thickness (+0.29%) relative to the assist-off condition, consistent with increased tibialis anterior engagement during SPN-assisted walking (Fig. 3d–f). Ankle kinematics improved accordingly before touchdown: the peak dorsiflexion angle increased by 85.9% relative to the assist off condition (Figs. 3g and S8b), whereas knee and hip flexion angles did not differ significantly between conditions (Fig. S8c–f). These dorsiflexion gains translated into enhanced foot clearance (Figs. S9 and S10), with the ensemble-averaged peak toe vertical displacement increasing by 26.6% (Fig. 3h) and participant-level peak displacement increasing by 14.8% under the assist-on condition (Fig. 3i). Muscle architecture and activation changes were consistent with these kinematic improvements: assistance increased tibialis anterior pennation angle relative to touchdown (Fig. S11a) and elevated swing-phase activation (Fig. 3b), supporting improved dorsiflexion and foot clearance further. Detailed tibialis anterior mechanics, morphological changes, and measurement methods are presented in Figs. S11 and S12. To provide an initial benchmark against a conventional clinical aid, we performed supplementary 6-meter overground walking tests under SPN, AFO, and assist-off conditions (Fig. S13a). Relative to assist off, SPN assistance increased paretic-side peak toe clearance by approximately 17.5% (~2 cm) during swing, an improvement comparable to that achieved with the AFO (Fig. S13b, c). However, the mechanism is fundamentally different: rather than providing passive mechanical compensation, the SPN uses self-powered, gait-coupled stimulation of the paretic-side tibialis anterior to recruit the wearer’s own dorsiflexor musculature and generate active ankle dorsiflexion, without requiring an external power source.

Fig. 3. Quantitative evaluation of SPN effectiveness during treadmill walking.

N = 7. (c, f, i): Bars represent the mean, error bars indicate the standard deviation (SD), and data are presented as mean ± SD. Two-sided paired t-test, and asterisks denote statistically significant differences, *P < 0.05, **P < 0.01. a Experimental setup for assessing tibialis anterior dynamics and lower limb kinematics. b tibialis anterior activity across the gait cycle in the assist-on and assist-off conditions. c Mean tibialis anterior activity during terminal swing in the assist-on and assist-off conditions (P = 0.044, two-sided paired-t test). d A wearable ultrasound probe for measuring tibialis anterior pennation angle (PNA) and thickness, with representative ultrasound data during concentric tibialis anterior contractions. e The Tibialis anterior thickness change relative to touchdown across the gait cycle in the assist-on and assist-off conditions. f Mean tibialis anterior thickness change relative to touchdown in the assist-on and assist-off conditions. g Ankle joint angle across the gait cycle in the assist-on and assist-off conditions. h Vertical toe displacement across the gait cycle in the assist-on and assist-off conditions. i Mean peak vertical toe displacement in the assist-on and assist-off conditions (P = 0.0038, two-sided paired t-test).

Trials in the community

Building on laboratory experiments demonstrating the prevention of foot drop, we extended our research to evaluate its practicality in outdoor, real-world settings. Nine stroke participants (Table S3) completed a 6-min walking trial in which they were instructed to cover as much distance as possible while walking back and forth on a straight, level 200-meter sidewalk in a shared community square (Fig. 4a and Movie S5). Compared with the assist off condition, participants walked substantially farther with SPN, with the maximum walking distance increasing by 67.3% from 226 m to 378 m (Fig. 4b), corresponding to an absolute gain of 152 m. We interpret this change as evidence of an immediate assistive effect under the tested community-walking condition. Future studies using matched single-session SPN-versus-FES comparisons and longitudinal head-to-head designs will be required to determine comparative clinical efficacy relative to existing FES-based interventions. Gait metrics also improved with SPN, including longer stride length (0.78 to 1.13 m; Figs. 4c and S14a), faster walking speed (0.62 to 0.89 m/s; +43.5%; Fig. 4d), greater foot-to-floor angle at touchdown (−1.1°–6.8°; Fig. 4e), increased ankle range of motion (29.8°–43.2°; Fig. 4f), and higher cadence (69.4–85.3 steps/min; Fig. 4g). In particular, the coefficient of variation of stride length decreased by 71.9% relative to assist off (Fig. S14b). Stair negotiation also improved with assistance (Fig. 4h and Movie S6): the foot-to-floor angle increased (3.3° vs. 18.6°; Fig. 4i), the toe‑stub event rate decreased (55.3% vs. 27.3%; Fig. 4j), and the total number of ascents within the test interval increased from 42 to 75 (Fig. 4k). Although foot drop and toe-stub events were not entirely eliminated during outdoor walking, the complexity of the outdoor walking task with a long duration, along with potential stimulating factors for foot drop in uncontrolled outdoor environments, was considered. Despite these factors, our observations indicate that the proposed wearable significantly mitigated foot drop. Overall, these results demonstrate the SPN’s multifunctionality and clinical potential, supporting its feasibility and effectiveness for community use.

Fig. 4. SPN improves mobility in individuals with foot drop.

N = 9, b-d: Bars represent the mean, with error bars indicating the standard deviation (SD). Data are shown as mean ± SD. f, g, i–k All metrics are displayed as box plots, where the center line represents the median (50th percentile), box edges correspond to the upper (75th percentile) and lower (25th percentile) quartiles, and whiskers extend to the minimum (Q1 − 1.5 × IQR) and maximum (Q3 + 1.5 × IQR) values within 1.5× the interquartile range (IQR). Individual points indicate outliers beyond the whiskers. Asterisks denote statistically significant differences (two-sided paired t-test with the mean difference as the test statistic, ***P < 0.001). a An experimental setup used to evaluate SPN performance in outdoor, community-based environments. b Average walking distance under assist-on and assist-off conditions. c Stride length of participants under assist-on and assist-off conditions (P = 1.49 × 10⁻⁵). d Average walking speed under assist-on and assist-off conditions. e Average foot-to-floor angle over a gait cycle under assist-on and assist-off conditions. f Ankle range of motion under assist-on and assist-off conditions (P = 6.4 × 10⁻⁶). g Cadence under assist-on and assist-off conditions (P = 1.96 × 10⁻⁴). h A foot drop participant with testing equipment navigating an outdoor stair environment. i Foot-to-floor angle under assist on and assist off conditions (P = 5.94 × 10⁻⁵). j Toe-stub event rate under assist-on and assist-off conditions (P = 4.08 × 10⁻⁴). k Number of steps ascended under assist-on and assist-off conditions (P = 8.05 × 10⁻⁴).

Discussion

The growing burden of aging-related neurodegenerative and cerebrovascular disease has increased the need for practical, user-acceptable mobility aids for people living with stroke and other gait-limiting conditions²². Although wearable robots and soft exosuits have advanced rapidly, many remain difficult to translate beyond controlled laboratory settings because of bulk, complexity, and limited usability in daily life. Here, we addressed this translational gap with a discreet, daily-life–oriented self-powered neuroprosthesis designed to assist ankle dorsiflexion and thereby reduce foot drop during walking.

Across both controlled treadmill testing and outdoor community walking, the SPN produced functionally meaningful improvements in ankle-foot behavior. In treadmill walking, the device increased dorsiflexion during swing and improved ankle positioning at initial contact and early stance (Fig. 3g), with temporally aligned electromyography (EMG) activity supporting phase-appropriate neuromuscular activation (Fig. 3b). Consistent effects were also observed outdoors, as reflected by IMU-derived foot-to-floor angle changes during overground walking (Fig. 4e). Together with the observed gains in toe clearance, stride length, and stride-length consistency, these findings indicate that the SPN can provide effective assistance not only in laboratory testing but also under more realistic walking conditions.

A key feature of the SPN is that it assists gait by engaging, rather than mechanically replacing, the user’s neuromuscular function. Unlike motor-driven ankle exoskeletons that deliver externally powered torque^23,24, the SPN harvests biomechanical energy from walking and uses that energy to stimulate the common fibular nerve, thereby eliciting active contraction of the ankle dorsiflexor muscles. By leveraging the natural out-of-phase coordination between the limbs during gait, the system delivers phase-appropriate stimulation at paretic-side swing initiation without requiring complex sensing, training datasets, or adaptive learning algorithms^25,26. This architecture reduces setup burden and system complexity while improving practicality for real-world deployment.

From a mechanistic perspective, the strongest evidence for SPN efficacy comes from its direct functional effects on gait. Dorsiflexion is biomechanically critical because it supports foot clearance during swing and promotes stable foot placement at ground contact. After a stroke, dorsiflexor insufficiency commonly leads to toe drag, compensatory gait strategies, elevated fall risk, and reduced community ambulation. Although simplified pendular models suggest relatively modest mechanical work requirements for dorsiflexion^27,28, empirical evidence underscores the importance of precisely timed dorsiflexor activation during the swing-to-stance transition^4,29. In this context, the immediate improvements in dorsiflexion angle, toe clearance, and walking performance observed here support the functional relevance of the SPN as a targeted intervention for post-stroke foot drop.

Ultrasound-based muscle morphology measures provided an additional, but more exploratory, perspective on the intervention response. Tibialis anterior pennation angle (Fig. S11a) and ultrasound-derived muscle thickness (Fig. 3e, f) both showed directionally positive shifts over the gait cycle, particularly during swing. However, the mean increase in dorsiflexor thickness was small (+0.29%), and its physiological significance should therefore be interpreted cautiously. In the context of post-stroke foot drop, where paretic dorsiflexor recruitment during swing is often limited, the observed direction of change is nevertheless consistent with increased tibialis anterior engagement during SPN-assisted walking. This interpretation is further supported by the concomitant improvements in dorsiflexion angle (Figs. 3g and S8a, b), toe clearance (Figs. 3h, S9 and S13b, c), stride length (Fig. 14a), and coefficient of variance in stride length (Fig. 14b), which provide more direct functional evidence than the small morphology shift alone. Prior work has shown that peroneal or superficial peroneal nerve stimulation can induce phase-specific improvements in stroke gait, including short-term carryover and neuromotor set-point modulation^18,30,31. The parallel changes in pennation and thickness are also physiologically plausible, because pennation can increase with muscle activation, muscle belly expansion, and tendon-aponeurosis strain^32,33, while transient viscoelastic effects and strain redistribution within the muscle-tendon unit may extend these changes beyond the stimulation window^34,35. In addition, stroke survivors often exhibit elevated tonic co-contraction around the ankle for stability^36,37, which may interact with FES-assisted dorsiflexion and contribute to morphology changes outside the immediate stimulation phase. Nevertheless, because dynamic ultrasound remains sensitive to probe placement and task-related variability, these morphology findings are best regarded as supportive, rather than primary, evidence of SPN-induced neuromechanical response^38–41.

The timing of peroneal nerve stimulation is another important consideration for both functional benefit and user comfort. In the present system, stimulation was initiated shortly after toe-off and remained active for more than 60% of the swing phase, rather than being confined only to terminal swing. This timing strategy was designed to facilitate ankle dorsiflexion during swing while improving foot positioning before heel strike. Such timing is clinically relevant because inadequate dorsiflexion commonly persists into late swing in individuals with foot drop, compromising foot placement before initial contact and increasing tripping or fall risk when assistance is absent^42,43. At the same time, stimulation beginning earlier in swing may be advantageous for some patients with more severe impairments in hip and knee flexion, whereas extending stimulation across the entire swing phase could increase tibialis anterior activation duration and potentially contribute to muscle fatigue⁴⁴. In our cohort, which spanned a range of lower-limb motor impairment and ankle spasticity severity (Tables S2–S4; Fugl-Meyer LE score: 13–27/34), the present timing strategy still produced significant improvements in ankle kinematics and foot clearance (Figs. S8a, b, 9 and 13b–c), supporting its functional relevance in the tested population. Nevertheless, this timing should not be interpreted as universally optimal. Patients with more severe hemiparesis may require stimulation that begins earlier in the swing phase, particularly when insufficient hip and knee flexion during the initial swing is a major contributor to impaired toe clearance. Future versions of the system should therefore incorporate more flexible timing control to broaden clinical applicability.

From a clinical and translational perspective, the SPN should be interpreted relative to existing assistive options. Compared with the custom-prescribed AFO, the SPN achieved comparable peak toe-clearance, while knee and hip kinematics remained broadly comparable among the SPN, AFO, and assist off conditions and did not exhibit a consistent condition-dependent pattern across participants (Figs. 8c–f and S13e–f). All AFOs used in this comparison were custom-prescribed thermoplastic solid AFOs (Table S3 and Fig. S13a), which reduced heterogeneity at the level of orthosis category. However, because these orthoses were individually fabricated for each participant, we did not perform a sensitivity analysis to determine whether participant-specific differences in orthosis fit or mechanical characteristics influenced the comparison; this should therefore be considered when interpreting the AFO benchmark. Overall, these findings suggest that the principal short-term effect of the SPN is targeted improvement in distal ankle-foot function and walking performance rather than broad normalization of whole-limb gait mechanics. This interpretation is consistent with the intended role of the device, which is to assist swing-phase dorsiflexion rather than directly modulate proximal joint motion. At the same time, established AFOs and modern lightweight FES systems each offer important practical and clinical advantages, including simplicity, maturity, and demonstrated efficacy in stroke rehabilitation. Although the present study included a direct comparison with a conventional AFO, direct head-to-head testing across the broader range of existing assistive technologies—particularly commercial FES devices and exoskeleton-based systems—was beyond its scope. We therefore do not position the SPN as a replacement for all existing approaches, but rather introduce it as a distinct assistive strategy that combines battery-free operation, gait-synchronized neuromuscular activation, and reduced system complexity. Consequently, comparisons with existing FES and exoskeleton-based approaches remain primarily qualitative and should be interpreted cautiously until direct head-to-head trials are performed. A broader comparison of the SPN with conventional AFOs, FES systems, and exoskeleton-based approaches is provided in Table S5.

More broadly, the ability to extend dorsiflexion-targeted neuromotor assistance beyond harnessed treadmill settings into community ambulation may have implications that reach beyond immediate gait mechanics^30,45. Improved mobility in stroke survivors has been associated with favorable cardiometabolic outcomes, reduced respiratory risk and all-cause mortality⁴⁶, and better mental health and quality of life⁴⁷. Although the present study evaluated only short-term responses, these findings raise the possibility that sustained use of a practical, self-powered gait-assist device could contribute to broader gains in activity, participation, and health.

Limitations

After completing community testing, two stroke volunteers requested a set of laboratory prototypes for daily use. While this positive feedback (Fig. S15 and Movie S7) is encouraging, our prototypes still require further optimization to transition into a fully mature product. First, the electrical energy generated by the energy harvester is stored in a supercapacitor (Fig. 2d, initial voltage: 0 V). The electrical stimulation module can only be activated once the capacitor voltage reaches 2.9 V. Consequently, wearers need to take a few steps in advance (Fig. 2e) to charge the capacitor to its operating voltage before receiving assistance from the SPN. In particular, in individuals with extremely slow or fragmented gait, the energy harvested from each step may be insufficient to sustain continuous stimulation, resulting in intermittent assistance. In our extended testing, this behavior was observed at walking speeds around or below 0.2 m/s and during prolonged gait pauses exceeding approximately 2 s (Fig. 2e). Second, stimulation frequency (35 Hz) and pulse width (300 µs) were fixed across participants, and although stimulation amplitude was adjusted individually, the available three-level amplitude setting was relatively coarse (Fig. S16). As a result, the selected setting may not have fully matched the functional and comfort requirements of every participant^21,48. Future work will focus on more refined and potentially closed-loop control of stimulation intensity to better accommodate inter-individual variability under different walking conditions.

Summary

We present a wearable, self-powered neuroprosthesis that coordinates dorsiflexor stimulation to avert foot drop during walking after stroke. By leveraging the out-of-phase nature of gait, the device harvests energy from the non‑paretic limb and delivers timed electrical stimulation to the paretic tibialis anterior, facilitating ankle dorsiflexion during swing. In controlled treadmill and real-world overground tests, the SPN increased foot clearance, stride length, and walking speed—metrics associated with gait safety and efficiency. These improvements were immediate, repeatable, and observed across multiple testing conditions, indicating robust on-demand assistance without complex setup or user tuning. The system operates without external batteries, integrates into standard footwear, and requires minimal donning time, supporting pragmatic use beyond the laboratory. Early findings suggest clinically meaningful enhancements in gait quality, reduced fall risk potential via improved foot clearance, and a pathway toward restorative neuromotor engagement rather than motion restriction. Given the lack of established, scalable engineering solutions to avert foot drop, this work represents a significant step toward practical, engineering-driven rehabilitation for stroke-related gait dysfunction.

Methods

Experiments involving patients

We enrolled twenty-one individuals post-stroke who exhibited foot drop. Baseline impairment was characterized using the Fugl-Meyer Assessment of the Lower Extremity (FMA-LE) and the Modified Ashworth Scale (MAS) for the ankle plantarflexors. Across participants, FMA-LE scores ranged from 13 to 27 out of 34, and MAS scores ranged from 0 to 2, indicating a range of lower-limb motor impairment levels and variable ankle spasticity (Tables S2–S4). All experiments and research protocols were conducted following the principles of the Declaration of Helsinki and received approval from the Institutional Review Board of the Hong Kong University of Science and Technology (HREP-2024-0258). Each participant gave informed written consent before testing, which included permission for the use of pictures/videos of their faces for publication purposes. The recruited participants were first-time users of the SPN, and they were given sufficient time to familiarize themselves with it before data collection, thereby reducing potential biases related to unfamiliarity. The order of the experimental conditions was randomized (Tables S6-S8) among participants to eliminate order and fatigue effects⁴⁹. Here, “randomized” refers to the order of condition administration within participants rather than to a parallel-group randomized controlled trial design. Due to a hardware malfunction, the test result of one participant in the treadmill experiment was excluded from the analysis.

Hardware

The total weight of the SPN is 132 grams (energy harvester 56 g, neuroprosthesis and cable: 76 g), with the energy harvesting module embedded in the shoe heel on the non‑paretic side, while the neuroprosthesis is routed around the waist to the surface of the tibialis anterior nerve on the paretic side. The energy harvester produces three-phase electricity, which is routed through a small hole drilled in the heel using a female Dupont connector (Fig. S1b), allowing for a plug-and-play interface with the male Dupont connector on the calf. To minimize the weight of the device, lightweight resin materials were selected for fabricating the energy harvester. To facilitate donning and doffing, the electrodes are integrated into the inner side of the Velcro strap on the paretic leg’s calf. The PCB circuit fabrication was provided by Jialichuang (Shenzhen JLC. Co., Ltd, China). To improve the durability of the hardware system, the power management circuit is installed in the heel and fixed with sealant. To provide sufficient installation space for the energy harvester, we used a hollow-heeled boot (Nanlong Boots) to install the energy harvester. The generator was supplied by Sanxing Motor (Sanxing Motor Co., Ltd.), and its gearbox was provided by Shuowei Manufacturing (Shantou, China).

Energy harvesting characterization

To quantify the net power available for stimulation, the mechanical-to-electrical energy conversion pathway was considered as three sequential stages: mechanical transmission (gearbox), electrical rectification, and charge storage. Net power delivered to the SPN was determined from electrical measurements acquired at the output terminal of the charge-storage module (Fig. S5a, b), thereby reflecting the cumulative effect of transmission, rectification, and charge-buffering losses.

Stimulation protocol and calibration

Bipolar surface stimulation was delivered to the superficial peroneal nerve using pre-gelled electrodes positioned around the fibular neck. Stimulation frequency (35 Hz) and pulse width (300 µs) were fixed across participants, following established FES parameters for post-stroke foot-drop assistance^21,48. Stimulation amplitude, by contrast, was adjusted individually during an initial fitting session. For each participant, the amplitude was gradually increased until visible ankle dorsiflexion was elicited comfortably, and was then refined using participant feedback to achieve an appropriate balance between functional assistance and comfort. In the present prototype, three discrete amplitude levels (low, medium, and high, Supplementary Fig. S16) were available, and the final level used for testing was selected for each participant based on both dorsiflexion effectiveness and comfort.

Textiles

We designed a textile system integrated with the SPN. The SPN is connected to the lower limb through a compliant textile strap. The textile interface consists of an adjustable waist belt (comprising a buckle and elastic band for ease of wear), one fabric harness for the shank, and two fabric harnesses for the thighs (Fig. 1b). On the non-paretic side, one Velcro strap is placed on the thigh and one on the calf to secure the wires. A buckle-style elastic band connects the wires from the non‑paretic side to those on the paretic side. The Velcro strap on the thigh of the paretic side is used to secure the connection between the wires and the neuroprosthesis, which is integrated with a Velcro strap on the calf of the paretic side. To facilitate cleaning and enhance wearing comfort, the electrodes are made from a sealed leather pouch. To improve the conductivity of the leather pouch electrodes, a small amount of water is sprayed on the surface before each use of the simulator. The entire wearable system is assembled using adjustable Velcro and buckle mechanisms, allowing the donning process to take no more than 2 min.

Indoor walking tests

The indoor walking tests consisted of two parts: treadmill walking tests and overground walking tests. To quantitatively assess the power output of the energy harvester during walking, as well as the effects of wearing the SPN on lower limb kinematics and dorsiflexor muscle activation, we conducted walking tests on a treadmill (Zebris Rehawalk, Germany). During an initial familiarization phase, each participant selected a self‑preferred comfortable walking speed according to their mobility level and stroke severity¹². The chosen speeds ranged from 0.2 to 0.8 m/s (Table S2) and were kept constant between the assist-on and assist-off conditions to eliminate the potential confounding effect of speed variation. Each participant completed four treadmill trials in total, consisting of two trials under the assist off condition and two trials under the assist on condition. Each trial lasted 2 min. If motion‑capture markers or EMG electrodes became detached during a trial, the trial was repeated to ensure complete and reliable data collection. A 15‑min seated rest interval separated the assist‑off and assist‑on conditions to mitigate fatigue effects¹². The indoor overground walking tests were conducted on a 6‑meter pathway marked by cone markers at both ends (Fig. S13a and Movie S8). During each task, participants were instructed to walk back and forth for 12 cycles under three randomized conditions (Table S7): SPN assist on, AFO, and assist off. The AFO used by each participant had been custom-prescribed and provided by the hospital or rehabilitation center where they received clinical care. Lower limb kinematic data were recorded using an eight-camera 3D motion capture system (Nexus 2.11, Vicon Motion Systems Ltd., Yarnton, UK) at a sampling rate of 250 Hz. The voltage output from the energy harvester was measured with an oscilloscope (RTE 1024, Rohde & Schwarz, USA). Surface EMG signals were recorded from the tibialis anterior using a wireless system (Delsys Inc., Natick, MA, USA) at a sampling rate of 2000 Hz. Electrode placement on the shank muscles followed established guidelines for surface EMG acquisition⁵⁰. To study tibialis anterior muscle architecture (e.g., thickness, pennation) during walking, we used a wearable Ultrasound-imaging-based Visual Feedback (UVF) system previously developed by our team⁵¹ to evaluate architectural changes under two experimental conditions: assist off and assist on. Data was collected using a wireless method to eliminate any interference from electrical cables.

Outdoor experiments

To evaluate the performance of the SPN under real-world walking conditions, nine stroke participants with clinically diagnosed foot drop were recruited for outdoor testing. Each participant performed two separate six-minute walk tests (6 MWTs): (i) level-ground walking and (ii) stair ascent and descent. Each task was conducted under two conditions: assist off and assist on. A 15-min seated rest period was provided between conditions within each task to minimize fatigue effects. For the level-ground 6 MWT, participants were instructed to walk back and forth along a 200-meter straight public pathway, covering as much distance as possible within six minutes. Cone markers were placed at each end of the path to guide turning points (Movie S6). For the stair-based 6 MWT, participants were instructed to ascend a public staircase (riser height: 16.5 cm; tread depth: 30.5 cm) repeatedly for six minutes, again aiming to cover the greatest possible vertical distance. All tests were conducted near a residential community in the Longhua District of Shenzhen. During both tasks, participants were not in proximity to any walls or handrails, thereby limiting the possibility of using external supports for balance assistance. The total distance walked by participants was measured using a measuring tape, while real-time ankle joint angles (Note S2) were recorded using Inertial Measurement Units (IMUs; BMI088, Crazyfile 2.1, USA), processed with Madhwick’s IMU and AHRS algorithms⁵². Touchdown (TD) and toe-off (TO) were identified from kinematic thresholds and confirmed by IMU-derived features. Steps with TD-TO inconsistency> 50 ms were excluded. Toe-stub event rate (IMU + visual): During the 6 MWT, foot- and shank-mounted IMUs flagged toe‑stub candidates when, in late swing (≤150 ms pre‑contact), the foot IMU showed a transient impact signature with low model‑based foot clearance. An experimenter simultaneously annotated visible toe‑to‑stair contacts. Events were counted only if IMU and visual annotations agreed (with missed visual cases added on review). Event rate = events/total steps × 100%. Stride length was calculated as the horizontal distance between heel markers upon each foot contact. The foot-to-floor angle was defined as the sagittal-plane angle between the foot (heel-to-toe orientation from a dorsum-mounted IMU) and the horizontal. Foot orientation was estimated from the dorsum IMU via accelerometer–gyroscope fusion, with initial gravity alignment during quiet standing and drift mitigation using stance-phase calibration (Fig. S17b).

Data processing

Joint kinematics were processed using a fourth-order Butterworth low-pass filter with a 10 Hz cutoff frequency. For each participant, we first inspected the EMG traces visually and removed segments contaminated by artifacts arising from suboptimal electrode placement, electrode motion, or baseline drift⁵³. The EMG data were zeroed to the mean values obtained from the entire gait cycle, full-wave rectified, and low-pass filtered using a 4th-order and bi-directional Butterworth filter with a cut-off frequency of 6 Hz⁵³. For envelope extraction, we identified runs of 15 consecutive strides that met the inclusion criteria within the analysis window, averaged the muscle activity across those strides, and normalized the resulting profile to its peak amplitude⁵⁴. Condition-level muscle activity was summarized by computing the time integral of the activity over a stride and dividing by the stride duration, yielding a time-averaged activation measure.

The dataset comprised 4849 B-mode ultrasound frames, corresponding to approximately 10 min of continuous recording. Muscle fascicle thickness and pennation angle were obtained by labeling individual fascicles in ultrasound data acquired with our previously described wearable sensors⁵⁵ during treadmill walking. To mitigate spatial drift associated with affine optical flow over extended sequences, a combination of manual correction and semi-automated tracking was employed, following previously established methods⁵⁶. The ultrasound, Vicon motion‑capture, and surface EMG systems were hardware‑synchronized. Touchdown events were identified from Vicon kinematics, and the corresponding frame indices were used to extract the aligned ultrasound image sequences for each gait cycle (Fig. S12). The power output of the energy harvester was computed as $V^2/R$, where $V$ is the load voltage and $R$ is the load resistance.

Statistics

Data processing, analysis, and visualization were performed in MATLAB (R2018b) and Origin 2023 (OriginLab Corporation, Northampton, MA, USA). For each participant and condition, kinematic variables (e.g., stride length, joint angles, and EMG signals) were averaged over 15 consecutive steady-state gait cycles selected from the middle portion of the treadmill recordings. This procedure was used to minimize transient effects associated with gait initiation and termination and to provide a stable representation of steady-state gait²⁸. Ultrasound-derived parameters were averaged over five consecutive gait cycles because of the lower frame-acquisition rate of the UVF system. Subject-level averages were then summarized across participants as mean ± standard deviation (s.d.). Statistical comparisons between conditions (“assist-off” vs “assist-on”) were conducted at the subject level using paired tests. The normality of within-subject differences (including averaged EMG features) was assessed with the Jarque–Bera test²⁸. When normality was satisfied, paired t-tests were applied; otherwise, Wilcoxon signed-rank tests were used⁵⁴. Two-sided permutation tests were additionally performed as a robustness check⁵⁷. When multiple biomechanical outcomes were tested, P-values were adjusted using the Benjamini–Hochberg false discovery rate⁵⁸. For each comparison, we report the effect size (mean difference or Hodges–Lehmann median difference), 95% confidence interval, test statistic, and two-sided P-value, with significance defined at α = 0.05 after correction where applicable.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Q.P. and Z.Y. conceived the project and designed the studies. Q.P., Y.L., and H.L. performed the biomechanics experiments and analyzed the experimental data. Y.Z., Y.W., and S.B. performed the investigation and designed the kinematics testing system for the lower limb. Z.L. worked on circuit design and optimization. J.Z. and M. P. analyzed the ultrasound and kinematics data. R.L., W.L., F.D., Z.L., X.Y., B.W., Y.H., and C.M. provided experimental assistance. Q.P., Y.L., B.H., C.M., and Z.Y. composed the manuscript. All authors provided critical feedback on the manuscript.

Peer review

Peer review information

Nature Communications thanks Jeongmin Kim and Toshiyuki Fujiwara for their contribution to the peer review of this work. A peer review file is available.

Funding

The research was supported by the Hong Kong Research Grants Council (GRF Project No. 16214025 and 16219826), and National Natural Science Foundation of China (Project No. T2522031).

Data availability

All data supporting the findings of this study are available within the article and its supplementary files. Any additional requests for information can be directed to, and will be fulfilled by, the corresponding author(s) or mention the public repository. Source data are provided with this paper.

Competing interests

United States Non-Provisional Patent Priority No. IP.PA.12708.US.PRV. Title: Biomechanical Energy Harvesting to Improve Post-Stroke Mobility. Assignee: The Hong Kong University of Science and Technology. Inventors: 1) Zhengbao YANG; 2) Qiqi PAN; Filed on 02 Sep 2025. The other authors declare that they have no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77394-3.