MXene/graphene composite flexible impedance glove with wide linear range for cross-subject gesture recognition

Abstract

With the continuous evolution of human-machine interaction (HMI) technologies, natural and intuitive gesture recognition has emerged as a pivotal interaction modality. However, constrained by inter-individual differences in hand dimensions and wearing habits, existing gesture recognition techniques still suffer from limited generalization performance in cross-subject recognition. This paper proposes a flexible impedance glove based on MXene/graphene composites with wide linear response range, which enables lightweight and zero-shot cross-subject gesture recognition. The sensing layer of the glove is fabricated using MXene/graphene/spandex composites, delivering a wide linear tensile range of 0-90% (R² = 0.995), and exceptional dual-modal sensing capability, effectively suppressing baseline fluctuations induced by individual differences. To enable high-precision gesture signal acquisition and intelligent classification, a self-designed miniaturized, low-cost impedance measurement module and the Random Forest (RF) algorithm were integrated. The customized hardware module achieves a high signal-to-noise ratio (SNR) of 61.6 dB and an average measurement accuracy of 99.79%. By employing a computationally lightweight Random Forest (RF) algorithm, the integrated system delivers an intra-subject recognition accuracy of 99.43% in classifying 12 complex gestures. Remarkably, it achieves a zero-shot cross-subject recognition accuracy of 91.04%, which further improved to 97.85% with the introduction of merely 5% individual calibration data. Furthermore, the successfully developed applications in intelligent sign language translation and low-latency game control demonstrate the promising application potential of the proposed system in barrier-free communication, HMI and other related fields.

Introduction

Human-machine interaction (HMI) is evolving toward more intuitive and natural paradigms with the advances of technology. Traditional interfaces such as keyboards, mice, and touchpads are no longer sufficient to meet contemporary interactive demand¹. Owing to its innate efficiency and alignment with daily human communication habits, gesture recognition has emerged as a pivotal technology for realizing seamless HMI. By directly interpreting the postures and kinematics of the human hand, gesture recognition serves as a fundamental interaction module in numerous fields, including intelligent robotic interaction², virtual reality³, and sign language translation⁴.

Current mainstream methodologies for gesture recognition primarily encompass computer vision⁵, electromyography (EMG)⁶, data gloves⁷, and radio frequency sensing⁸. Although these methods can convert physical hand motions into digital signals, they all suffer from inherent technical limitations in practical applications. Computer vision techniques are highly susceptible to external environmental factors such as lighting variations, background clutter, and object occlusion, which inherently compromise recognition accuracy and reliability. EMG signals are intrinsically weak and vulnerable to interference from adjacent bioelectrical activities or external noise. Conventional data gloves integrate numerous sensors, leading to high costs and performance limitations imposed by sensor precision. In addition, radio frequency-based techniques are vulnerable to interference due to their inherent signal frequency characteristics.

Flexible strain sensing technology offers a highly promising upgrade to circumvent the limitations of traditional rigid interactive interfaces, owing to its exceptional deformation sensitivity, cost-effective fabrication, and superior conformal wearability. Driven by this trend, significant advancements have been achieved in flexible pressure and strain sensors based on two-dimensional nanomaterials, which exhibit exceptional electromechanical coupling capabilities and broad application prospects in wearable MEMS/NEMS devices and dual-mode sensing systems^9–11. Building upon these advanced materials, researchers have extensively explored various flexible sensing architectures, such as graphene-coated silk fabrics, graphene aerogels, printed carbon nanotube composite networks, as well as sensors integrated with microfluidic and bionic hierarchical structures. Furthermore, to fully decipher the complex physiological signals captured by these hardware interfaces, integrating flexible sensing arrays with advanced machine learning algorithms has emerged as a highly effective paradigm. This data-driven strategy has shown great success in broad human activity monitoring, such as intelligent sitting posture recognition and HMI^12,13. In the specific domain of gesture classification, synergizing these sensor architectures with machine learning models, such as Artificial Neural Networks (ANN) and backpropagation networks, these systems generally achieve outstanding recognition accuracies ranging from 84% to 100% in intra-subject gesture classification tasks^14–19. Despite the remarkable advances in intra-subject gesture recognition accuracy achieved in the aforementioned studies, insufficient cross-subject generalization ability remains a core bottleneck that restricts the practical application of flexible sensing-based gesture recognition systems²⁰. Current flexible sensors generally have a narrow linear sensing range, which fails to accommodate differences in initial pre-strain arising from variations in hand sizes among individual users, thereby triggering signal baseline drift and nonlinear response. Furthermore, most systems rely on sophisticated deep learning models and massive labeled datasets for training. New users are required to undergo tedious and time-consuming calibration procedures, which severely impair user experience and system practicability.

To address the aforementioned challenges, this work proposes a flexible impedance-sensing glove based on MXene/graphene with a broad linear range for lightweight and zero-shot cross-subject gesture recognition. Through the synergistic optimization of material properties, hardware circuit architecture and algorithm lightweight design, the cross-subject generalization issue of flexible sensing-based gesture recognition is systematically resolved. In terms of materials, a composite conductive network of MXene and graphene is constructed, which delivers a wide linear tensile range and dual-mode sensing capability. This design guarantees signal uniformity for users with different hand sizes. At the hardware level, a compact impedance measurement module with a collaborative MCU-CPLD architecture is developed to enable high-fidelity signal acquisition in a miniaturized, low-cost form factor. On the algorithm side, a computationally lightweight Random Forest (RF) algorithm is employed for the classification of complex gestures. This design choice is intended to facilitate robust zero-shot cross-subject recognition and rapid user adaptation with minimal personal calibration data, without relying on complex deep learning models. Furthermore, two practical applications, namely intelligent sign language translation and low-latency game control, are developed based on the proposed system. The results validate its practical value in barrier-free communication, HMI and other related fields.

Results and discussion

The gesture recognition system based on MXene/graphene glove

This paper develops a gesture recognition system based on an MXene/graphene sensing glove. As illustrated in Fig. 1a, the system is capable of recognizing sign language information represented by complex hand gestures. Furthermore, the system supports gesture-driven program control for HMI, as demonstrated in Fig. 1b. The core component of the system is a sensing glove featuring an MXene/graphene composite conductive network (Fig. 1c–i). Benefiting from the superior strain-sensing properties of the composite material, the glove can accurately and stably capture resistance responses induced by hand motions. The system is further integrated with a self-developed miniaturized, low-cost impedance measurement module, as shown in Fig. 1c–iii. Equipped with a cooperative MCU-CPLD architecture and an adaptive gain mechanism, this module enables high-SNR, high-precision multi-channel signal acquisition within a compact physical footprint, endowing the system with excellent dynamic response performance and robust scalability for diverse application scenarios.

Fig. 1. Flexible impedance-based gesture recognition system based on MXene/graphene glove.

a Hand gesture translation schematic diagram. b HMI schematic diagram. c (i) Detailed structure of the sensing glove. (ii) Overall structure of the MXene/graphene glove. (iii) Impedance collection module. d Schematic diagram of signal flow in the system

The overall signal flow of the developed system is visualized in Fig. 1d. The MXene/graphene glove captures resistance variations induced by gesture movements. After signal preprocessing by the impedance measurement module, the acquired data are wirelessly transmitted to the host computer via TCP. The host computer implements data training and gesture recognition using machine learning algorithms, and the recognized results are subsequently delivered to the application terminal through a Flask server to realize multifunctional interactive applications.

The sensing mechanism and characteristics of MXene/graphene

To clarify the microstructural evolution of the conductive network under tension, SEM observations were carried out on composite films fixed at 0%, 10%, 40%, and 70% preset tensile strains, with gradient morphologies presented in Fig. 2a. At 0% strain, the composite film forms a dense and complete conductive network. Most MXene nanoparticles are wrapped inside stacked graphene sheets. Under 10% low strain, tensile force creates directional slip bands without through-thickness cracks, relative sliding between nanosheets separates overlapping graphene layers and exposes more embedded MXene along slip tracks. At 40% strain, modulus mismatch between substrate and coating generates nascent microcracks, and delaminated graphene sheets become visible, with MXene particles exposed and gathered at crack edges. When stretched to 70%, microcracks expand with enlarged interlayer gaps. Further separation of nanosheets releases more buried MXene, forming dense clusters at sheet boundaries, while large graphene lamellae bridge the crack gaps.

Fig. 2. Sensing mechanism and characteristics of the MXene/graphene composite conductive network.

a SEM surface morphology of the MXene/graphene composite conductive network under 0%, 10%, 40% and 70% tensile strain. b Schematic illustration of the testing setup. c Response of the composite conductive networks with varying MXene/graphene mixing ratios (1:3, 1:2, 1:1, 2:1, 3:1) under (i) tension and (ii) pressure. d Response time of the composite conductive network to (i) tension and (ii) pressure. e Stability test under 5000 cycles of stretching/release and pressing/releasing. f Voltage response of the composite conductive network to continuous (i) stretching/release and (ii) pressing/releasing. g Resistance response of the MXene/graphene glove at finger bending angles of 0°, 45°, 90° and 135°

To measure the electrical properties of the MXene/graphene composite conductive network, a DC power supply (DC-5030D) and an oscilloscope (UPO2204) were used for combined testing. As presented in Fig. 2b, the regulated power source supplied a constant direct current as the excitation signal, while precisely controlled mechanical deformation was applied to the sensing material. Electrical signals were eventually acquired by the oscilloscope. To determine the optimal material formulation, composite conductive networks with MXene/graphene mass ratios were firstly fabricated and systematically compared. As shown in Fig. 2c, a clear trade-off between sensitivity and linear working range was identified. Samples with higher MXene content deliver enhanced sensitivity but suffer from narrowed linear ranges due to increased film brittleness, while graphene-rich samples exhibit improved stretchability with compromised sensitivity as deformation is accommodated primarily by interlayer sliding of graphene sheets. The 1:1 composite strikes the optimal balance between these performance metrics. As shown in Fig. 2c–i, the relative resistance change (ΔR/R₀) of the 1:1 composite exhibits favorable linearity over a wide tensile range of 0-90% (R² = 0.995) with the sensitivity of 2.53E^-2/%. This linear performance is primarily ascribed to the synergistic percolation-microcrack mechanism, which is well supported by the gradual structural changes observed in SEM images. At low strain, nanosheet sliding widens interlayer spacing moderately, leading to slow resistance growth dominated by electron tunneling within the intact percolation network. At medium and high strains, microcracks tend to induce steep resistance increases, whereas cross-crack graphene bridges and boundary MXene clusters maintain secondary tunneling pathways. Such synergistic structural effects avoid sharp resistance jumps and sustain linear resistance evolution across a broad strain window^21–23. Under compressive stress (Fig. 2c–ii), the 1:1 composite similarly exhibits a highly reliable linear response of 0-3.65 kPa (R² = 0.968) with the sensitivity of –0.19 kPa^-1. As applied pressure increases, the loosely structured MXene/graphene conductive network undergoes densification inside the composite, which sharply multiplies conductive pathways until contact saturation occurs. Furthermore, the precise upper and lower limits of detection under both tensile strain and pressure were systematically determined via stepwise loading tests combined with the 3σ criterion (Fig. S1 in Supporting Information).

A comparison with state-of-the-art sensing structures from recent studies is presented in Table S1 in Supporting Information. Most reported sensors are restricted to single-mode operation and feature narrow linear ranges. In contrast, the MXene/graphene composite conductive network enables dual-mode sensing and retains excellent continuous linearity over a wide tensile range of 0-90%. The outstanding linearity within such a broad strain range substantially reduces the complexity of subsequent signal demodulation. More importantly, it delivers strong tolerance to variations in wearing size. This feature ensures stable performance within the optimal linear region, even when different users introduce varying levels of initial pre-strain due to dissimilar hand dimensions. A reliable material basis is established for extracting high-fidelity gesture features and achieving cross-subject recognition.

Rapid dynamic response is an essential property for the conductive network to meet the requirements of real-time monitoring. Figure 2d presents the transient response of the conductive network under mechanical loading and unloading. For tensile deformation (Fig. 2d–i), the response time and recovery time of the conductive network reached as low as 0.65 s and 0.60 s, respectively. Notably, an even faster dynamic response was observed under compressive loading, with the corresponding response time of 0.31 s and recovery time of 0.28 s (Fig. 2d–ii). The smooth, symmetric waveforms without obvious hysteresis demonstrated the superior structural resilience and rapid reconstruction capability of the conductive network within the composite. Long-term stability was evaluated via 5000 consecutive tensile-release and compression-relaxation cycles (Fig. 2e). Initially, a slight signal drift occurred due to the localized sliding of nanomaterials and the formation of preliminary microcracks within the matrix. Subsequently, the internal architecture reached a new dynamic equilibrium. In this steady state, the highly reversible opening and closing of microcracks enabled the MXene/graphene composite to effectively dissipate stress and prevent continuous structural damage, leading to a stabilized response. Consequently, the conductive network retained 83.98% of tensile response and 81.45% of pressure response after 5000 cycles, demonstrating reliable long-term durability.

Furthermore, signal reversibility and hysteresis characteristics under dynamic deformation were investigated through tests under stepwise cyclic stress. Figure 2f–i records the relative electrical output during stepwise cyclic tension ranging from 0-80% strain. Highly stable flat signal platforms were formed at each minor strain step. In addition, nearly overlapping response curves at identical strain levels during loading and unloading manifested low hysteresis. A highly symmetric stepwise response free of baseline drift and noticeable hysteresis was also achieved under progressive cyclic compression up to 6 kPa (Fig. 2f–ii). The low hysteresis verified that the MXene/graphene composite conductive network possesses excellent structural resilience and rapid reversible reconstruction of conductive pathways after mechanical deformation. This feature eliminates signal interference caused by previous movements and guarantees high-precision signal acquisition for continuous motion capture in gesture recognition systems. In addition to excellent mechanical cycling stability, the conductive network also maintains favorable temperature stability, detailed temperature drift test results are provided in Fig. S2 of Supporting Information.

Finally, sensing gloves integrated with the MXene/graphene composite conductive network were fabricated. Hand gestures impose tensile and compressive deformation on the glove, which dynamically modulates the resistance of the composite conductive network. Accordingly, gesture signals can be acquired by measuring the potential difference across different electrode pairs. As shown in Fig. 2g, the resistance response of the glove was tested at finger bending angles of 0°, 45°, 90° and 135°. During finger bending, the circumferentially wrapped sensing layer undergoes asymmetric deformation: the outer dorsal side bears tensile strain, while the inner side contacting the finger is subjected to compressive strain. The overall resistance decreases stepwise with increasing bending angle, dominated by compressive densification of the conductive network: narrowed interlayer spacing of nanosheets substantially increases effective percolation pathways and reduces resistance. On the other hand, the tensile resistance rise is mitigated by the synergistic percolation-microcrack mechanism elaborated above: large graphene sheets bridge across microcracks and MXene clusters sustain secondary tunneling pathways, jointly suppressing abrupt resistance increase. The resistance reduction from compression outweighs the resistance increase from tension, leading to an overall downward trend. No obvious baseline drift is observed during the holding of a specific gesture. These results preliminarily validate the feasibility of the sensing glove for applications in gesture recognition and HMI.

Structural design and performance evaluation of miniaturized impedance measurement module

To precisely capture subtle impedance variations of the MXene/graphene composite network induced by hand gestures, overcome the size limitations of conventional measuring devices and meet the requirements of wearable applications, a miniaturized impedance measurement module based on a collaborative MCU-CPLD architecture was designed (Fig. 3a–i). The module features an overall physical dimension of only 25.4 mm × 25.4 mm, and integrates four core functional units: main control, excitation generation, signal acquisition and power management (Fig. 3a–ii).

Fig. 3. Structural design and performance evaluation of a miniaturised impedance measurement module.

a Hardware implementation. (i) Physical diagram of the impedance measurement module. (ii) System block diagram of the impedance measurement module. (iii) Schematic diagram of the adjustable gain differential amplifier circuit. b Stability analysis of the excitation current under different load resistors. c Evaluation of the SNR of the standard resistance ring and sensing glove under different gain multiples. d Evaluation of impedance measurement accuracy

In terms of operating principles, the CPLD delivers high-determinism low-level timing control, while the ESP32-C3 MCU performs upper-layer data processing and Wi-Fi transmission. Regarding the signal flow, square waves generated by the CPLD are converted into sinusoidal waves via bandpass filters. These sinusoidal signals drive the Digitally controlled Voltage-Controlled Current Source (D-VCCS) to produce high-precision excitation current, which is then fed into the Multiplexer (MUX) array. A total of 40 sets of boundary voltage signals are collected by eight electrodes on the sensing glove using the adjacent excitation and adjacent measurement strategy (four-terminal measurement). The acquired signals are preprocessed by a gain-adjustable differential amplifier circuit and digitized by the Analog-to-Digital Converter (ADC) before being transmitted back to the main MCU via the Serial Peripheral Interface (SPI). Furthermore, the on-board power management unit provides dual-polarity power supplies of +3.3 V and –2 V, enabling long-term and stable operation of the module under wearable scenarios.

The stable and high-fidelity excitation current is a prerequisite for accurate impedance measurement. The module adopts the conventional constant-current excitation and voltage measurement mode. A highly stable alternating excitation current I_ex is injected into the target measurement channel via the digitally controlled D-VCCS. Geometric deformation of the MXene/graphene composite conductive network during gesture movement leads to dynamic variations in local impedance ΔZ. According to the generalized Ohm’s law, the variation of boundary voltage ΔU_boundary detected between measuring electrode pairs is expressed as:

$$\mathrm{\Delta }{U}_{\mathrm{boundary}}=I_{\mathrm{ex}}\mathrm{\Delta }Z$$ 1

The spatial distribution and temporal dynamic fluctuation of boundary voltage ΔU_boundary establish a rigorous linear mapping for the impedance topology ΔZ of the microscopic network. It should be noted that although the system is designated an impedance measurement module for its 25 kHz AC excitation and the device an impedance glove, the MXene/graphene composite sensing layer is intrinsically a pure piezoresistive network. At the operating frequency, parasitic capacitance and inductance of the flexible sensor are minimal, yielding negligible phase shift. The measured complex impedance is thus fully dominated by its real component, which corresponds to pure ohmic resistance, so the acquired boundary voltages ΔU_boundary directly reflect resistive variations of the conductive network. As shown in Fig. 3b, the peak excitation current (I_p) output by the built-in D-VCCS stays stable at approximately 0.8786 mA over a wide dynamic load range of 1 Ω to 1 kΩ. A slight current attenuation complying with physical laws only occurs when the load approaches the high-impedance limit of 10 kΩ. This finding verifies that the system can sustain a constant excitation output under varying loads induced by hand motions in wearable applications.

To address the issue that output signals of flexible sensors are susceptible to environmental noise during dynamic deformation, a gain-adjustable differential amplifier circuit is integrated into the module. Adaptive regulation of the amplifier gain enables high-fidelity acquisition of impedance signals associated with hand gestures (Fig. 3a–iii). Figure 3c illustrates the nonlinear relationship between the SNR and amplifier gain. The SNR values of both standard resistor rings and wearable sensing gloves rise first and then decline with increasing gain, reaching the optimal impedance matching state at a gain of approximately 9. Under this optimal configuration, the peak SNR of the standard resistor ring reaches 70 dB, and the measured SNR of the sensing glove is 61.6 dB. These results strongly demonstrate the excellent noise immunity of the impedance measurement module. Detailed data are presented in Fig. S4 of Supporting Information.

Impedance measurement accuracy was further characterized to evaluate system reliability (Fig. 3d). An average measurement accuracy of 99.786% was achieved, as summarized in Table S2 in Supporting Information, and the contact resistance between stainless steel button electrodes and the composite coating exerts negligible interference on signal acquisition accuracy, as quantitatively verified by the transmission line model method in Fig. S5 of Supporting Information. Comparative analysis between the proposed MCU-CPLD miniaturized impedance module and state-of-the-art data acquisition schemes for flexible sensors is provided in Table S3 in Supporting Information. Traditional commercial instruments suffer from large size and high cost, while conventional portable solutions based on single-chip microcontrollers or wireless SoCs are limited by insufficient SNR and measurement accuracy. In contrast, the developed module achieves a compact footprint of 25.4 mm × 25.4 mm and a low cost of 19 US dollars, while delivering an SNR of 70 dB and a stable impedance measurement accuracy of 99.79%. It exhibits remarkable overall performance. The capability of high-fidelity signal acquisition within a miniature form factor provides solid and reliable data support for subsequent gesture recognition.

Evaluation of the performance and cross-subject generalization capability of the gesture recognition system

To evaluate the interactive feasibility of the gesture recognition system based on the MXene/graphene sensing glove, boundary voltage signals were collected for twelve typical gestures, namely “Hi”, “Protest”, “Thumbs up”, “Victory”, “Up”, “OK”, “Call”, “Question”, “Love”, “Heart”, “Luck”, and “Accept”. Fig. 4a presents the spatial distribution of boundary voltages output by the multi-channel impedance measurement module under different gestures. Owing to the distinct strain responses of the composite conductive network to specific hand postures, the voltage signals of each gesture exhibit highly unique profile variations. The temporal evolution characteristics of multi-channel boundary voltages for gestures during motion transitions are detailed in Fig. S6 of Supporting Information. To further verify the separability of features in the multidimensional space, the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm was adopted for dimensionality reduction and visualization of the high-dimensional boundary voltage data (Fig. 4b). The results show that the twelve gestures form well-defined and mutually independent clusters in the feature space. Meanwhile, the heat map (Fig. 4c) further confirms low correlation and high discriminability among features of different gestures. These findings lay a solid data foundation for high-accuracy classification using subsequent machine learning models.

Fig. 4. The characteristics of different gestures.

a Boundary voltage diagrams, b t-SNE diagrams, and c heat maps of 12 typical gestures collected by the gesture recognition system

Machine learning models deployed on the host computer are employed to train the acquired boundary voltage signals and realize recognition of complex gestures. As illustrated in Fig. 5a, boundary voltage signals of various gestures collected by the miniaturized impedance measurement module serve as the training set for model fitting during the training phase. For gesture recognition, real-time signals are wirelessly transmitted to the host computer and fed into the pre-trained model for the classification of complex gestures.

Fig. 5. Gesture recognition and generalization ability evaluation.

a Flowchart of Model Training and Real-time Gesture Recognition. b Accuracy and Error Analysis of Gesture Recognition for Four Machine Learning Models (ANN, RF, KNN, SVM). c Aggregated Confusion Matrix of Leave-One-Out Cross-Validation (LOOCV) for 15 participants. d Aggregated Confusion Matrix after Incorporating 5% Personal Calibration Data. e Gesture boundary voltage diagrams of 15 participants

During algorithm selection, five-fold cross-validation is applied to systematically evaluate the classification performance of four mainstream algorithms, including ANN, K-Nearest Neighbor (KNN), Support Vector Machine (SVM) and RF (Fig. 5b). Data sampling for the twelve standard gestures was conducted with eight recruited participants. To guarantee the temporal and spatial independence as well as diversity of samples, the experiment conducted 5 rounds of data collection for 12 types of gestures. Fifty repeated samples were recorded for each gesture round, yielding a total dataset of 24,000 voltage signal samples. All samples are divided into a training set and a test set at a ratio of 7:3. Quantitative results reveal that RF achieves the optimal fitting and classification performance when processing multi-channel high-dimensional voltage features. An average recognition accuracy of 99.43 ± 0.18% is attained for the twelve complex gestures (Fig. S7 in Supporting Information), with only minor classification errors observed between a few gestures with highly similar geometric characteristics. Furthermore, quantitative metrics evaluated on the PC platform confirm that the RF algorithm provides the optimal balance between computational speed and resource efficiency compared to other models (Table S4 in Supporting Information). Accordingly, the RF algorithm is ultimately integrated as the core recognition algorithm of the system, and the number of decision trees in the RF is optimized (Fig. S8 in Supporting Information).

To validate the necessity of retaining all 40 dimensions of data, we conducted a SHapley Additive exPlanations (SHAP) analysis (Fig. S9 in Supporting Information). Figure S9a shows that all 40 boundary voltage channels make non-zero contributions to classification decisions, confirming each channel captures unique local spatial strain information. As shown in Fig. S9b, intra-subject recognition accuracy increases monotonically as channels are added in descending order of importance, peaking only when all 40 channels are included. These results demonstrate that the full 40-dimensional dataset provides a holistic, non-redundant spatial mapping of hand impedance topology.

Cross-subject recognition stands as the core challenge for the practical deployment of flexible wearable devices. Inherent differences in physical features and wearing habits among individual users easily lead to drift in the initial impedance baseline at the sensor interface. To evaluate the zero-shot generalization capability of the system, leave-one-out cross-validation (LOOCV) was adopted, with fifteen participants tested individually in sequence. Detailed anthropometric data including gender, hand length, hand width and hand thickness of all participants are summarized in Table S5 of Supporting Information. The aggregated confusion matrix is presented in Fig. 5c, showing an overall accuracy of 91.04 ± 5.48%. Performance fluctuations across different participants were observed due to natural variations in physical characteristics, while the average accuracy remained at 91.04%, which demonstrates the favourable generalizability of the system. Detailed results of the individual LOOCV tests for all participants are provided in Fig. S10 in Supporting Information.

To further mitigate recognition performance degradation caused by individual physical differences and explore the feasibility of practical personalized wearable applications, a few-shot user enrollment mechanism was introduced. Based on leave-one-out cross-validation, 1%, 3%, 5%, 7%, and 9% of each test subject’s gesture samples are separately incorporated into the training set for model fine-tuning, and distinct diminishing marginal improvement in recognition accuracy was observed as the proportion of personalized calibration data increased (Fig. S11 in Supporting Information). Among all tested ratios, 5% personalized calibration data was verified as the optimal trade-off point, lifting the average recognition accuracy to 97.85 ± 2.45% as illustrated in Fig. 5d, with most misclassification entries from the zero-shot test eliminated effectively, verifying that adaptation to individual physical differences can be achieved with only a limited number of user samples.

Furthermore, gestures achieving a 100% recognition rate in leave-one-out validation were selected, and an aggregated analysis was conducted on the multi-channel boundary voltage distributions collected from the participants (Fig. 5e). The results reveal that although differences in hand sizes among participants induce baseline drift in absolute voltage amplitude, the spatial topology of multi-channel signals, including peak and valley positions as well as relative variation trends, exhibits highly consistent isomorphic mapping characteristics. This advantage originates from the ultra-wide linear range of 0-90% of the MXene/graphene heterogeneous network, which effectively accommodates initial pre-strain derived from variations in hand dimensions and maintains constant proportional relationships of resistance changes across all channels during deformation. Without data normalization, individual differences are merely reflected in linear scaling of the magnitude of the 40-dimensional voltage vector, while the directional features in the multidimensional space remain unchanged.

As the feature vectors of the twelve gestures are mutually uncorrelated (Fig. 4b), the corresponding data points form distinct radial clusters in the 40-dimensional space. Therefore, the RF algorithm relying on absolute voltage thresholds can efficiently separate these highly distinguishable feature clusters via stepped decision boundaries. Robust cross-subject decoding is realized without resorting to computationally intensive domain adaptation algorithms^24–26.

To intuitively highlight the superior cross-subject generalization performance of the proposed system, a comprehensive comparison with state-of-the-art gesture recognition systems reported in recent literature is presented in Table S6 in Supporting Information. Most existing wearable systems suffer from severe accuracy degradation due to individual physical differences during cross-subject validation, with their accuracy generally dropping to the range of 70%-90%. Although a few systems achieve comparable high accuracy by adopting sophisticated continuous transfer learning, they rely on deep neural networks and intensive GPU computing, which hinders their deployment on wearable devices with limited computational resources. In contrast, the proposed system requires no complex deep learning frameworks. It achieves an accuracy of 91.04% under the zero-shot setting by employing the RF algorithm with extremely low computational cost. The accuracy further increases to 97.85% when merely 5% labeled data are introduced. This work provides a feasible solution for lightweight and rapid user adaptation in cross-subject wearable applications.

Demonstration of intelligent sign language translation and low-latency game control

To verify the application potential of the developed impedance-based gesture recognition system, a demonstration platform consisting of hardware, software and application layers was established, as shown in Fig. 6a. The three layers work collaboratively to realize gesture data acquisition, processing and functional execution. Real-time boundary voltage signals induced by hand gestures are captured by the miniaturized impedance measurement module and wirelessly transmitted to the Flask server via Wi-Fi based on the TCP protocol.

Fig. 6. Demonstration of Real-time sign language translation and HMI Demonstration.

a Overall system architecture diagram covering the hardware layer, software layer and application layer. b Demonstration of continuous action recognition and semantic parsing of the sign language translation system. c Demonstration of low-latency virtual game interaction based on gesture control

The Flask server is responsible for data reception and task scheduling. It forwards the acquired data to the gesture recognition module for feature processing and generates stable recognition results. These results are subsequently delivered to two independent application modules for game control and sign language translation, as detailed in Fig. S12 in Supporting Information.

Real-time gesture recognition and sign language translation implemented by the system are demonstrated in Fig. 6b, Movie S1 and Movie S2 in Supporting Information. To filter out transient mechanical jitter and misclassification inevitably generated during continuous gesture switching, a stabilization verification mechanism based on a sliding window is integrated into the software layer for the sign language translation function. Specifically, the system dynamically caches the latest five frames of classification results. A gesture is regarded as a valid action only when the confidence proportion of the corresponding category within the current window exceeds the threshold of 0.7. The discrete sequence composed of validated stable gestures is then imported into the locally deployed Ollama model. Leveraging the contextual reasoning capability of the large language model, fragmented physical gestures are converted into natural language text that conforms to standard grammatical logic, which is displayed in real time on the Web frontend (Fig. S13a in Supporting Information).

As shown in Fig. 6c and Movie S3 in Supporting Information, a low-latency instruction mapping framework is established for the game control application. Quantitative measurements reveal that under a local Wi-Fi network, the average user-perceived system latency reaches 105.5 ms (Table S7 in Supporting Information). Real-time gesture recognition outputs are seamlessly converted into directional motion vectors in the virtual space to drive the underlying game engine for state evolution, including coordinate update, collision detection and synchronous calculation of reward mechanisms. All updated system states are finally rendered into real-time graphics on the Web frontend (Fig. S13b in Supporting Information).

Conclusion

An MXene/graphene flexible impedance sensing glove with a wide linear range is developed for lightweight and zero-shot cross-subject gesture recognition. The as-fabricated MXene/graphene composite conductive network exhibits a broad linear tensile range of 0-90%. It adapts to hand size variations among most users and eliminates signal nonlinearity and baseline drift originating from inconsistent initial pre-strain from the material perspective. The self-developed impedance measurement module adopting a collaborative MCU-CPLD architecture features a miniature footprint of 25.4 mm × 25.4 mm and a low cost of 19 US dollars. It achieves a signal-to-noise ratio of 61.6 dB and an average measurement accuracy of 99.79%, alongside a data acquisition rate of 30 frames per second. Its overall performance is superior to that of most conventional portable acquisition systems. Notably, the classification process requires neither complex deep learning models nor large-scale datasets. Outstanding cross-subject recognition performance is realized using the lightweight RF algorithm. The integrated system reaches a zero-shot cross-subject accuracy of 91.04%, which increases to 97.85% after calibration with only 5% few-shot samples. This presents an efficient and practical technical route for rapid user adaptation of wearable devices. Built upon the proposed system, an intelligent sign language translation platform and a low-latency game control system are implemented. The former converts discrete gestures into fluent natural language, and the latter enables real-time mapping from hand movements to motion commands in virtual environments. These applications demonstrate the practical value and promising application potential of the system in barrier-free communication and HMI.

For long-term wearable deployment, future work will systematically investigate the skin biocompatibility of the MXene/graphene composite and optimize the long-term durability of the sensing glove, including washing resistance, sweat resistance and abrasion resistance. This will be carried out as a core direction of our follow-up research to further promote the practical application of the system.

Materials and methods

Fabrication of MXene/graphene composite sensing glove

The composite conductive network precursor solution was prepared via a liquid-phase blending method. First, 10 mL of commercial MXene (Ti₃C₂Tx) aqueous dispersion with a concentration of 5 mg/mL and 50 mg of high-purity graphene powder were precisely measured and co-dispersed in 100 mL of isopropanol solvent. The obtained mixture was subjected to water-bath ultrasonication for 30 min, followed by continuous magnetic stirring at a rotational speed of 800 rpm for 60 min at room temperature. Subsequently, the uniformly dispersed solution was quantitatively coated onto a highly elastic spandex fabric substrate through a spray-coating process. After coating, the fabric was placed in a vacuum drying oven and dried at 60 °C for 3 h to completely volatilize residual solvents, forming a conformally bonded composite conductive network.

For electrode assembly, stainless steel button electrodes were fixed at the proximal joints of each finger of the spandex glove according to human hand anatomical characteristics and the test results of recognition accuracy, forming arrayed impedance measurement contacts (Fig. S3 in Supporting Information).

Morphology characterization and electrical measurements

The microscopic morphology and elemental distribution of the composite conductive network were characterized via scanning electron microscopy (SEM, ZEISS Sigma 300) combined with energy-dispersive X-ray spectroscopy (EDS). In electrical tests, a regulated DC power supply (DC-5030D) was utilized to provide constant DC excitation for the sensor. For tensile performance evaluation, strains ranging from 0-100% were applied to the sensing fabric using a high-precision stepper motor, while normal pressures of 0–10 kPa were loaded for compressive performance testing. A digital oscilloscope (UPO2204) was synchronously employed to acquire real-time electrical signal outputs during all tests.

For the stability test of the D-VCCS in the impedance measurement circuit, a square wave voltage signal with a fixed peak-to-peak value of 3.3 V and a frequency of 25 kHz was adopted as the input to generate a peak current output of 0.8786 mA, with the load resistance varying from 0.1Ω to 10 kΩ.

A standard resistor ring model was adopted as the test object for signal-to-noise ratio (SNR) measurement, which was composed of eight high-precision metal film resistors (220 Ω, 1/4 W, ±1% tolerance) connected in series. The SNR was calculated according to the following formula:

$$\mathrm{SNR}\left(\mathrm{dB}\right)=10{\log }_{10}\left(\frac{{\sum }_{i=1}^N{v}_i^2}{{\sum }_{i=1}^N{\left(v_i-\overline{v}\right)}^2}\right)$$ 2

where v_i denotes the boundary voltage acquired in the i-th measurement, and N represents the number of sampling frames, which was set to 100 in this system test. Data acquisition was implemented through eight electrodes in the four-terminal measurement mode. During system operation, the control unit sequentially injects excitation current into adjacent electrode pairs and synchronously measures the potential difference between the remaining adjacent electrodes. With the adopted scanning strategy, a total of 40 sets of boundary voltage data (8 × 5) can be obtained for each complete scanning frame. To eliminate accidental errors and ensure the statistical significance of calculation results, 100 consecutive frames of stable data were collected for SNR analysis.

The Impedance Measurement Accuracy (IMA) was tested based on the four-terminal measurement mode, and the relevant calculation formulas are defined as follows:

$$IME=\frac{∣Z_s-Z_i∣}{Z_i}\times 100\%$$ 3

$$IMA=100\%-IME$$ 4

where Z_s is the actual impedance value measured by the impedance measurement module, and Z_i represents the theoretical impedance value. Specifically, an excitation current with an amplitude of 0.8786 mA was injected into the test load, and the potential difference across the load was measured synchronously. To prevent the amplified differential voltage signal from exceeding the effective range, the circuit gain was reasonably set to 5, and the actual impedance of the test load was calculated by the system according to Ohm’s law. Before all tests, baseline calibration was performed using a high-precision 220 Ω±1% metal film standard resistor to guarantee data reliability. Furthermore, two 500 Ω protective resistors were connected in series within the current excitation circuit to avoid output distortion and oscillation of the D-VCCS under abrupt load impedance changes, effectively ensuring the output stability of the constant current source.

Standardized process for collecting gesture data

To minimize body motion drift and gesture morphological deviations during data acquisition and ensure high intra-class consistency and cross-subject comparability of the dataset. All recruited participants followed a unified, stepwise procedure throughout the entire acquisition process.

Before formal sampling, anthropometric parameters of each participant’s dominant hand, including hand length, hand width and hand thickness, were first measured and recorded to characterize inter-subject differences in hand dimensions. All participants then completed a 10-min familiarization session, during which they repeatedly practiced each of the 12 target gestures under the guidance of on-screen visual cues. This session eliminated the learning effect in early trials and ensured all participants fully mastered the required movement rhythm and postural standards before formal recording.

During formal acquisition, participants maintained a fixed, comfortable seated posture with their forearms resting flat on a stable horizontal surface and upper arms naturally perpendicular to the ground. The wrist joint was kept in a neutral position to avoid compensatory wrist movements. This physical constraint effectively eliminated signal drift caused by unconscious shoulder elevation or involuntary arm movements.

The execution speed, static hold duration and relaxation phase of each gesture were strictly synchronized via visual cues displayed on a screen. Each individual trial followed a standardized temporal sequence: 2 s of resting baseline recording, 3 s of controlled gesture execution, 2 s of static posture hold, 3 s of controlled return to the relaxed position, and 3 s of inter-trial rest. The presentation order of the 12 gesture types was randomized for each sampling round to eliminate potential order effects. This uniform temporal control ensured consistent gesture morphology and temporal characteristics across all trials and participants.

The entire data collection was divided into 5 independent sampling rounds, with 50 repeated trials acquired for each gesture category per round. A mandatory 5-min rest interval was enforced between consecutive sampling rounds, during which participants kept their hands completely relaxed. This measure prevented muscle fatigue, a primary cause of progressive gesture morphological deviation and baseline drift during extended recording sessions. Real-time waveform monitoring was conducted throughout the acquisition process, and trials with obvious baseline drift or abnormal morphology were marked and re-collected on site to guarantee data quality.

Conflict of interest

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41378-026-01446-3.