NeuraLoop: a high bandwidth closed-loop human–machine interface Authors: Luis Pelaez Murciego, Elias Thomassen Dam, Hans Henrik Dalgaard, Nikola Jorgovanovic, Matija Strbac, Erika G. Spaich, Strahinja Dosen Published: 2025 ABSTRACT BackgroundMyoelectric interfaces have emerged as powerful tools for human–machine interaction (HMI), enabling intuitive control of virtual and physical devices. However, most existing systems are limited by low spatial resolution and unidirectional communication. To address these limitations, we developed NeuraLoop, a wearable, high-bandwidth, bidirectional interface that integrates myoelectric (EMG) signal acquisition and electrotactile stimulation feedback within a single wearable textile-based platform.MethodsNeuraLoop comprises a flexible matrix of 32 EMG recording and 32 electrotactile stimulation pads controlled by a compact electronic unit. We evaluated the system in two experimental tasks involving ten healthy subjects to demonstrate: (1) online classification of four transient thumb micro-gestures (thumb rightwards, leftwards, upwards, and downwards swipe directions), and (2) closed-loop control of a virtual cursor using micro-gesture commands and spatially encoded tactile feedback. A time-division multiplexing (TDM) strategy was implemented to enable simultaneous stimulation and recording.ResultsThe subjects achieved a median success rate of 82% on the first attempt and over 94% within two attempts during online classification with visual feedback. All four micro-gestures were classified with similar accuracy. In the closed-loop control task with tactile feedback, participants navigated a 3 × 4 grid using only electrotactile stimulation, achieving 70% accuracy for exact target hits and 95% when including the hits in the neighboring cells (1 cell distance error).ConclusionsNeuraLoop demonstrates the feasibility of high-bandwidth, bidirectional HMI using a wearable, textile-based interface. The system enables accurate recognition of subtle micro-gestures and effective delivery of spatially encoded tactile feedback. These capabilities open new possibilities for intuitive control in applications such as prosthetics, rehabilitation, and virtual/augmented reality. Future work will explore multimodal feedback encoding and proportional gesture control.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-025-01815-5. Background Myoelectric signals, which are electrical potentials generated by muscle activity, can be recorded non-invasively using surface electrodes placed on the skin [1]. These signals provide a rich source of information about muscle contractions and can be effectively used to recognize gestures performed by a participant [2, 3]. Traditionally, this approach has been widely used in clinical applications, particularly in rehabilitation [4–6] and prosthesis control [7]. In these contexts, myoelectric control enables intuitive and responsive interaction with assistive devices and virtual environments for gamified exercises, which can lead to enhanced quality of life for individuals with motor impairments [8]. More recently, there has been a growing interest in leveraging myoelectric gesture recognition for broader human–machine interaction (HMI) applications. Unlike data gloves, which require direct contact with the hand and can restrict natural movement [9], or vision-based systems, which depend on the line-of-sight and are sensitive to lighting and occlusion [10, 11], myoelectric interfaces offer a less intrusive and more versatile alternative. Other HMI modalities have also been recently explored, such as ultrasound-based gesture recognition, which offers contactless control but often requires bulky hardware and is sensitive to sensor motion [12, 13]. This has led to the development of commercial EMG-based solutions such as the now-discontinued Myo armband and similar wearable devices, which typically integrate between eight and sixteen EMG channels housed in rigid components distributed around the forearm [14]. While these systems demonstrated the feasibility of myoelectric gesture recognition in consumer applications, their reduced number of channels and fixed sensor placement can constrain their ability to capture the full complexity of forearm muscle activity, thereby limiting their potential. In contrast, high-density EMG (HD-EMG) enables the acquisition of muscle electrical signals from a larger number of spatially distributed measurement points [15–17]. It has been shown that a spatially distributed subset of HD-EMG channels can provide similar classification accuracy compared to the standard uniform circumference distribution commonly used in current systems [18]. Traditionally, HD-EMG requires a cumbersome setup due to the large size and form of the electrode array, which usually needs to be attached to the body using adhesive foams and conductive cream to ensure skin–electrode conductivity [19]. However, leveraging screen-printing technology has made it possible to integrate large arrays of EMG sensors directly onto flexible, textile-based substrates [20, 21], removing one of the main barriers to implementing such systems in a wearable solution. This allows for the creation of compact, lightweight, and conformable interfaces that maintain high signal fidelity while improving user comfort and wearability. An important aspect of enhancing the usability and intuitiveness of such systems is the inclusion of feedback. Closing the control loop by providing haptic feedback to the user allows for more natural and effective interaction [22, 23]. Most existing myoelectric bands include no feedback or, at best, integrate a limited number of simple vibration motors delivering basic tactile cues. While useful, these cues are often limited in resolution and expressiveness. To address this limitation, more advanced forms of haptic feedback have been explored (such as pneumatic actuators [24, 25], or squeezing mechanisms [26]), among which electrotactile stimulation stands out as a particularly promising technology [27]. Electrotactile stimulation involves delivering controlled electrical pulses to the skin to evoke tactile sensations. Unlike mechanical actuators, electrotactile systems are compact, lightweight, and can be easily scaled to include a large number of stimulation sites [28]. This scalability enables the creation of high-resolution tactile displays capable of conveying complex spatial and temporal feedback patterns [29]. Moreover, electrotactile stimulation can be finely tuned in terms of intensity, frequency, and waveform, allowing for a wide range of perceptual experiences [30]. Electrical pulses can lead to uncomfortable or even painful sensations, but this can be easily avoided if the intensity is properly calibrated. In the context of myoelectric interfaces, electrotactile feedback offers several advantages. The stimulation pads can be co-located with the EMG electrodes, enabling a compact and integrated design [31]. It also supports dynamic feedback strategies, where spatiotemporally modulated profiles can be used to convey directional cues, pressure and force levels in prosthetics [31], or even symbolic information like letters or icons for general communication and human–computer interaction [29, 32]. These capabilities make electrotactile stimulation an attractive candidate for enhancing bidirectional human–machine interfaces, where both input (gesture recognition) and output (haptic feedback) are seamlessly integrated. However, electrotactile stimulation introduces interferences in EMG recording, but this problem can be solved using blanking [33] and/or time division multiplexing [34]. In this paper, we present NeuraLoop, a novel system for bidirectional human–machine interfacing that utilizes a matrix of flexible pads capable of both recording EMG signals and delivering electrotactile stimulation. This dual-functionality design simplifies the hardware architecture and improves the spatial alignment between input (control) and feedback channels. We demonstrate the effectiveness of our system by using it to classify a set of hand micro-gestures performed by healthy participants, while simultaneously providing real-time electrotactile feedback. Some preliminary test results on micro-gesture recognition (without simultaneous haptic feedback) have been reported in a recent conference paper [35]. The present study describes novel technology in detail, reports on comprehensive testing of system capabilities, and highlights the potential of this approach for developing more immersive and responsive wearable neural interfaces. Methods NeuraLoop system The NeuraLoop system represents a step forward in the development of intuitive, high-bandwidth, bidirectional human–machine interfaces. It is designed to enable closed-loop interaction between users and both physical and virtual environments by decoding user motion intent through muscle activity while also delivering real-time, context-aware haptic feedback to the user, thereby closing the control loop. The system is shown in Fig. 1 and comprises a flexible dry electrode laminated onto a stretchable textile band and connected to compact electronics for both recording and stimulation. The electrode is wrapped around the forearm, and the electronics are then attached to the top of the band. The matrix electrode comprises 32 pads for EMG recording and 32 pads for electrotactile stimulation, with a circular shape of 7.5 mm in diameter. All 64 pads are distributed in 6 rows separated by 8.5 mm, following a trapezoidal configuration to better accommodate the placement around the forearm. Pads belonging to the same row were used either for stimulation or recording, alternating the functionality on each row (i.e., first row recording, second row stimulation, third row recording, and so on). The electrode is attached to the NeuraLoop device using custom-made pogo pin connectors. NeuraLoop can thus simultaneously capture high-fidelity muscle activity and deliver rich haptic feedback, enabling high-bandwidth closed-loop human–machine interfacing. The device integrates a WiFi interface that allows full control by a host system (e.g., a laptop). The host can configure recording and stimulation parameters, modify the stimulation parameters online, and start and stop the stimulation, recording, and streaming of the data to the host. The stimulation frequency is common to all channels, but the intensity and pulse width can be independently adjusted, allowing for the generation of dynamic spatially and temporally varied stimulation patterns. High-density EMG recording allows the system to interpret complex and subtle hand and finger gestures and translate them into control commands for a wide range of applications while also providing tactile cues that intuitively convey status information from the application back to the user, as shown in Fig. 2. Fig. 1The NeuraLoop system comprises high-density recording and stimulation electrodes and compact electronics. a Schematic design of the electrode’s pad distribution. Black color indicates the EMG recording interface; red indicates the stimulation interface. The long strips at the top of the electrode serve as the reference pads for recording (including the driven-right-leg pad) and stimulation. The bottom panel shows details of the pogo-pin connectors designed to interface the device with the recording and stimulation contacts in the flexible sleeve. b Frontal view of the open sleeve showing the distribution of the recording and stimulation pads. c User wearing the NeuraLoop system on the forearm The NeuraLoop system comprises high-density recording and stimulation electrodes and compact electronics. a Schematic design of the electrode’s pad distribution. Black color indicates the EMG recording interface; red indicates the stimulation interface. The long strips at the top of the electrode serve as the reference pads for recording (including the driven-right-leg pad) and stimulation. The bottom panel shows details of the pogo-pin connectors designed to interface the device with the recording and stimulation contacts in the flexible sleeve. b Frontal view of the open sleeve showing the distribution of the recording and stimulation pads. c User wearing the NeuraLoop system on the forearm Fig. 2Conceptual diagram of the NeuraLoop system showcasing the closed-loop control to operate a physical or virtual device using hand and finger gestures, while receiving status information from the device encoded as tactile feedback. NeuraLoop integrates high-density EMG recording that allows reliable decoding of gestures and high-density stimulation for providing dynamic and spatially distributed electrotactile stimulation (rich haptic feedback) Conceptual diagram of the NeuraLoop system showcasing the closed-loop control to operate a physical or virtual device using hand and finger gestures, while receiving status information from the device encoded as tactile feedback. NeuraLoop integrates high-density EMG recording that allows reliable decoding of gestures and high-density stimulation for providing dynamic and spatially distributed electrotactile stimulation (rich haptic feedback) The block diagram of the electronic circuit is shown in Fig. 3. The electronic system comprises stimulation and recording units, a power supply module, a WiFi interface for communication with the host, a USB connection for charging and firmware updates, and the main microcontroller board. These modules are described in more detail in the next sections. Microcontroller unit The microcontroller unit (MCU) manages and coordinates all components and system functions. The MCU is a 32-bit ARM microcontroller STM32H743VIT6 from ST Microelectronics, featuring extensive hardware peripherals, high speed, and memory capacity. Communication between the MCU and the electrical stimulation module is established through a serial peripheral interface (SPI) and two compare modules. The SPI is responsible for communicating with the digital-to-analog converter (DAC), allowing control over the current pulse amplitude and enabling the selection of active cathode fields on a 32-channel matrix electrode through high-voltage switches. The two compare modules, each using a 16-bit timer, drive the current source (H-bridge) to generate biphasic current pulses with precise timing and frequency. To interface the EMG recording module, the MCU uses an isolated SPI channel in conjunction with general-purpose input/output (GPIO) pins. The SPI channel is employed to configure parameters such as the sampling rate and EMG amplification, as well as to retrieve EMG data. GPIOs are used to manage the timing of the sampling sequence. For communication purposes, a UART module connects the MCU to the Wi-Fi module ESP32-S3-WROOM from Espressif Systems. These connections allow the MCU to receive user-specified stimulation commands while simultaneously sending recorded EMG data to a host computer. Fig. 3Block Diagram of the NeuraLoop electronic circuit. The circuit comprises recording and stimulation modules, a power unit, a wireless interface for communication with the host, and a microcontroller unit. The components are explained in text Block Diagram of the NeuraLoop electronic circuit. The circuit comprises recording and stimulation modules, a power unit, a wireless interface for communication with the host, and a microcontroller unit. The components are explained in text Electrical stimulation The electrical stimulation module uses a current-controlled H-bridge configuration to generate biphasic, charge-compensated current pulses. Functioning as a single-channel stimulator, the H-bridge produces a sequence of pulses that are subsequently routed to matrix electrodes through a 32-channel analog switch array in monopolar configuration. Pulse amplitudes are set by a 16-bit DAC, while pulse duration and frequency are managed by the MCU’s compare modules and timers. Both pulse width and amplitude can be individually customized for each electrode pad within specific ranges: 50–1000 µs in 10 µs increments for pulse width, and 0.1–10 mA in 0.1 mA steps for amplitude. The stimulation frequency, which applies globally to all electrode pads, can be varied from 1 to 200 pps in 1 pps increments. Within each stimulation cycle, pulses are sequentially produced and then delivered to the designated active channels with a configurable inter-pulse interval, which is the period between pulses distributed to two channels, ranging from 1 to 25 ms in 1 ms steps. This asynchronous stimulation scheme allows distributing pulses to multiple pads of a matrix electrode (in this case, 32) using a single-channel stimulator. While straightforward to implement, this design imposes certain constraints on parameter combinations. Specifically, all pulses assigned to active channels during a given stimulation cycle must fit within the stimulation period. If pulse widths are uniform, the following condition must be met:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{ch}\bullet\:\left(2{T}_{pulse}+{T}_{TBP}\right)\le\:\frac{1000}{{F}_{stim}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{ch}$$\end{document} denotes the number of active channels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{pulse}$$\end{document} is the pulse duration, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{TBP}$$\end{document} is the time between pulses (both in milliseconds), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{stim}$$\end{document} is the stimulation frequency in pps. For instance, if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{pulse}=0.5\:ms$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{TBP}=1\:ms$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{stim}=100\:pps$$\end{document}, a maximum of five channels can be activated in one stimulation cycle. EMG recording The module for recording EMG signals is built around four ADS1299 integrated circuits (Texas Instruments, Texas, USA), which feature eight differential electrophysiological amplifiers and 24-bit sigma–delta analog-to-digital converters. In NeuraLoop, the EMG unit supports up to 32 monopolar input channels and includes a built-in driven-right-leg (DRL) circuit. To ensure user safety during concurrent EMG recording and electrical stimulation, the EMG module is galvanically isolated from the stimulation circuitry. Power supply The electronic circuit is powered by a single-cell lithium-polymer (Li-Poly) rechargeable battery with a capacity of 1200 mAh. The power management unit generates multiple voltage levels: 3.3 VDC for the microcontroller, + 90 VDC for the current-controlled H-bridge, and isolated supplies for the EMG recording circuitry. It also handles battery charging via the USB interface. The device operates for approximately 3 h on a full charge when both recording and stimulation functions are active simultaneously and continuously. However, most real-life applications will likely include intermittent periods of recording and stimulation, which will prolong the life of the device's operating time. A complete recharge takes approximately 2 h. Firmware The firmware is developed in C using STM32CubeIDE, an integrated development environment provided by STMicroelectronics. The firmware architecture is task-oriented, with three main tasks—communication, stimulation, and EMG recording—executing independently and interacting via hardware resources and interrupt-driven scheduling. The communication task manages bidirectional data transfer over Wi-Fi and USB, sending and receiving textual commands to and from a host computer and dispatching parameter values to the remaining two tasks. The stimulation task handles the generation of electrical pulses in accordance with the parameters provided by the communication task. The EMG recording task configures amplifier settings based on instructions from the communication task. The firmware also ensures that the stimulation and recording parameters and commands sent by the host are valid and comply with the defined constraints (e.g., expression (1)). When the device receives a command, it confirms reception if the command is valid; otherwise, it sends an error indication to the host. Electrode for recording and stimulation Custom-designed flexible, textile-integrated electrode arrays were fabricated using a multilayer screen-printing process, employing Henkel conductive and dielectric inks for biomedical applications. Conductive traces were printed using silver ink (Electrodag PF-410 E&C), followed by a dielectric insulation layer (Electrodag PF-455B) to encapsulate interconnects and define the exposed contact areas. Carbon ink (LOCTITE EDAG 440B E&C) was then printed onto the exposed electrode sites to ensure biocompatibility and a stable skin–electrode contact surface. The multilayer structure was deposited onto ELECROM FLEX 100 MIC, a PET-based substrate from Policrom, which includes a pressure-activated adhesive on its reverse side to facilitate textile integration. Final assembly was achieved through heat lamination onto neoprene, yielding mechanically robust and conformable electrodes suitable for prolonged use as biomedical wearables. Experimental assessment To demonstrate the device’s capabilities, we conducted an experimental assessment in which the system was used for both real-time micro-gesture recognition for control purposes and simultaneous delivery of electrotactile stimulation for haptic feedback. Ten healthy subjects (three females and seven males, mean age of 28.5 ± 4.1 years) participated in the experiment. Prior to the experimental session, all participants received oral and written information about the study and provided written informed consent. The study was conducted in accordance with the Declaration of Helsinki and approved by the North Denmark Region Committee on Health Research Ethics (N-20220001). The experimental session consisted of an offline data collection phase followed by two tasks to showcase the system in different scenarios: (1) micro-gesture control only and (2) micro-gesture control with simultaneous feedback. The subjects sat comfortably in front of a computer screen with the NeuraLoop system worn on the forearm of their dominant arm. A custom graphical user interface (GUI) was developed to facilitate data collection and online control assessment. Myoelectric control We selected the challenging task of recognizing hand micro-gestures, which are brief, subtle, rapid finger movements. HD-EMG has the advantage of providing excellent performance in complex gesture-recognition tasks without the need for accurate electrode positioning [18]. Specifically, we focused on four thumb swipe micro-gestures: upward, downward, leftward, and rightward, chosen for their relevance to everyday interactions with smartphones and wearable devices. In conventional gesture recognition, the subjects are asked to perform a gesture (e.g., close the hand) and maintain it, so that classification is performed using the steady stream of data characterizing the stationary contraction. This approach cannot be directly translated to the recognition of micro-gestures due to their transient nature. To address this, we introduced an additional step in the classification pipeline called activation detection. After EMG signals were collected and segmented into overlapping windows, they were analyzed to detect whether muscle activity was present. If activity was detected, 250 ms of subsequent data was transmitted to the rest of the pipeline and classified. The classifier was therefore mostly dormant and activated to classify transient data only when muscle activity was detected. A refractory period of 750 ms from the onset of detected activation was introduced before the next activation could be detected. This was done to avoid consecutive undesired activations produced by moving the thumb back to the initial position after executing the micro-gesture. The task began with a short calibration phase to manually determine the activation detection threshold for each subject. This was followed by a training data collection phase. In the conventional gesture classification, this is implemented by showing the gesture and instructing the subjects to perform it while maintaining a predefined level of muscle activation by tracking the muscle activation profile shown on the screen [18]. However, this approach cannot be used to collect the micro-gestures as they are transient movements. Instead, the subjects were prompted with a GUI showing the threshold level and the muscle activation profile that produced a peak during the movement execution. During this phase, subjects were instructed to perform each thumb swipe 25 times at a comfortable speed. Each time the activation profile crossed the threshold, it counted as one valid repetition. After data collection, all signal segments where activity was detected were extracted, concatenated, and labeled with the corresponding movement class (micro-gesture). Four time-domain features, including variance, slope-sign changes, zero-crossing, and waveform length, were computed and used to train a support vector machine (SVM) classifier. The selected features are known as the time-domain Hudgins feature set [36], which has been often used for gesture classification. Online control with visual feedback The online control task (without concurrent haptic feedback) consisted of controlling a ball displayed on the screen using four thumb swipe micro-gestures. NeuraLoop was activated to start collecting EMG data and streaming it to the host computer (laptop), which implemented the online classification pipeline. Fig. 4Representation of the GUI used during the online control task. In each trial, the participant was prompted with the centered ball and one of the four possible targets around it (solid circle). After a micro-gesture was detected, the ball moved toward the direction of the predicted class. The four hand illustrations depict the movements (micro-gestures) corresponding to each thumb swipe Representation of the GUI used during the online control task. In each trial, the participant was prompted with the centered ball and one of the four possible targets around it (solid circle). After a micro-gesture was detected, the ball moved toward the direction of the predicted class. The four hand illustrations depict the movements (micro-gestures) corresponding to each thumb swipe At the start of each trial, a red ball appeared at the center of the screen along with a target location indicated by a solid circle (Fig. 4). Participants were instructed to move the ball toward the target using the appropriate thumb swipe. A trial was considered successful if the target was reached within five attempts. This was done to provide some tolerance for misclassifications, similar to online classification of conventional gestures [37]. In the latter case, the incoming data windows are processed and classified continuously, and the participant can maintain the gesture until enough correct classifications are performed for the successful task completion. Micro-gestures are, however, transient events and therefore holding the gesture was replaced by a discrete number of attempts. Five attempts were selected based on pilot tests as a tolerance that would allow exploring how difficult the task is (e.g., the expected average number of attempts < 5). After each trial, a new target position was displayed. The four possible target directions were shown in random order, with each direction appearing 25 times, resulting in 100 consecutive trials. This task was conducted to establish baseline performance for micro-gesture classification. Online control with simultaneous haptic feedback This experiment was conducted to evaluate simultaneous recording and stimulation, a unique feature of NeuraLoop technology. A time-division multiplexing (TDM) strategy was used to avoid stimulation artifacts corrupting the EMG [34]. In this approach, EMG recording and stimulation are performed in dedicated alternating windows. Specifically, the TDM cycle comprised three sequential windows: a stimulation window (0.5 s), a pause window (0.8 s), and a recording window (1.0 s). During the pause phase, neither stimulation nor recording occurred; this was introduced as a “washout” period, ensuring that all residual artifacts from the stimulation phase dissipated before the EMG recording resumed. Fig. 5a Selection of pads for stimulation and recording during closed-loop control. Black dots represent the active recording electrodes (all recording electrodes were used), light blue dots show the stimulation pads, and blue dots with a dark blue circumference indicate the stimulation pads that were used in the online control task. Each of the used stimulation pads was associated with a corresponding grid cell in the control task (right panel), in the order from left to right and bottom to top. b Grid shown to the subjects in the conditions with visual feedback. The green square indicated the initial position, while the red square represented the target position of a trial with a distance of n = 4 steps. The task for the subject was to move the virtual cursor from the initial to the target position using thumb movements. A successfully recognized micro-gesture moved the cursor for one step in the direction of the movement, as indicated next to the grid. In the condition with haptic feedback, the current position of the cursor was indicated by activating the corresponding pad of the electrode. The trajectory generated by a representative subject in this condition is indicated by the blue arrows. The subject used six movements (t1 to t6) to reach the final position correctly a Selection of pads for stimulation and recording during closed-loop control. Black dots represent the active recording electrodes (all recording electrodes were used), light blue dots show the stimulation pads, and blue dots with a dark blue circumference indicate the stimulation pads that were used in the online control task. Each of the used stimulation pads was associated with a corresponding grid cell in the control task (right panel), in the order from left to right and bottom to top. b Grid shown to the subjects in the conditions with visual feedback. The green square indicated the initial position, while the red square represented the target position of a trial with a distance of n = 4 steps. The task for the subject was to move the virtual cursor from the initial to the target position using thumb movements. A successfully recognized micro-gesture moved the cursor for one step in the direction of the movement, as indicated next to the grid. In the condition with haptic feedback, the current position of the cursor was indicated by activating the corresponding pad of the electrode. The trajectory generated by a representative subject in this condition is indicated by the blue arrows. The subject used six movements (t1 to t6) to reach the final position correctly A calibration phase was conducted first to adjust the stimulation parameters. The process began by estimating the sensation threshold for each pad individually, by gradually increasing the pulse amplitude in steps of 0.1 mA until the subject indicated they perceived the stimulation (the so-called the method of limits). Once the sensation thresholds were established, the obtained amplitude values were multiplied by 1.5 to produce a comfortable yet clearly localized and perceivable sensation. Next, all pads were activated sequentially, and the subject was asked to confirm that each stimulation point felt equally intense and remained within the comfort range. If discrepancies were reported, individual pad amplitudes were adjusted to equalize the sensations across pads. The total time for electrode calibration was approximately 15 min. Following calibration, the subject was presented with a 3 × 4 grid (Fig. 5b). Each cell of the grid was associated with a corresponding pad on the electrode in a spatially congruent manner. A brief familiarization phase was performed, where the experimenter moved the cursor through the grid while the haptic feedback was provided. As the cursor switched to a new cell, the corresponding pad at the electrode was activated. This is known as spatially encoded tactile feedback, where the elicited tactile sensations moved across the subject’s forearm congruently with the movement of the cursor across the grid cells, thereby providing an intuitive representation of the current position. The familiarization phase allowed the subject to associate each cell with its corresponding tactile sensation on the forearm. The subject then practiced navigating the grid with TDM enabled, using thumb movements while receiving electrotactile feedback. Finally, the online task consisted of navigating the grid from a starting cell to a target cell, relying only on the tactile feedback, with no visual cues provided during movement. At the beginning of each trial, the grid was displayed with a green square indicating the starting position and a red square marking the target, as illustrated in Fig. 5b. After a few seconds for the subject to memorize the positions, the visual prompt was removed, and they were asked to use thumb movements to move the green square toward the target while the current position of the cursor was conveyed using only tactile feedback. Once the subject believed the target had been reached, the trial ended, and a new prompt with a different origin-target combination was shown. Start and target positions were randomly generated for each trial, with distances ranging from one to five cells. The experiment included six repetitions for distances one to three, and three repetitions for distances four and five, totaling 24 trials. In the first half of the experiment, the final position was shown to the subject at the end of each trial to provide reinforced learning. Once all trials were completed, the experiment was repeated without disclosing the final position (validation phase). Data analysis The outcome measures were the success rate for both tasks, and path efficiency and recall for the closed-loop control task. The success rate for the micro-gesture recognition task was calculated as the percentage of successful trials per attempt. The success rate for the closed-loop control task was calculated as the percentage of successful trials with two accuracy levels: (1) the subject reached the target (within-0 error) and (2) the subject reached a neighboring cell around the target (within-1 error). Overall performance was summarized using a confusion matrix for the first attempt across all subjects. Path efficiency was computed as the ratio between the length of the optimal trajectory and the length of the trajectory generated by the participant (number of steps performed inside the grid) across successful trials [38]. The optimal trajectory was defined as the shortest distance in steps required to reach the target cell from the initial position (cell), moving along horizontal/vertical directions (diagonal movements were not allowed). Activation detection recall was calculated as the percentage of the micro-gestures detected relative to the total number of micro-gestures performed. Undetected micro-gestures were counted manually by visually inspecting the participant’s hand during the online control task. The Shapiro–Wilk test was used to assess the normality of the data distributions. As the data did not meet the assumption of normality, the nonparametric Friedman test was applied with post hoc analysis using the Wilcoxon signed-rank test and Bonferroni correction. Results Fig. 6Segment of the raw EMG signal recorded from a representative subject while performing the four thumb micro-gestures indicated on top of each column. Each micro-gesture produces characteristic patterns of muscle activity across channels Segment of the raw EMG signal recorded from a representative subject while performing the four thumb micro-gestures indicated on top of each column. Each micro-gesture produces characteristic patterns of muscle activity across channels Fig. 7Summary of results for the online micro-gesture recognition. The boxplots show the distribution of a success rates across subjects for each number of attempts, b overall success rates for each micro-gesture and c confusion matrix showing the correct classification rates (diagonal) and errors (off-diagonal) for the first classification attempt. (**, p < 0.01) Summary of results for the online micro-gesture recognition. The boxplots show the distribution of a success rates across subjects for each number of attempts, b overall success rates for each micro-gesture and c confusion matrix showing the correct classification rates (diagonal) and errors (off-diagonal) for the first classification attempt. (**, p < 0.01) An example of EMG signals recorded during the execution of micro-gestures is shown in Fig. 6. All four micro-gestures elicited characteristic patterns of muscle activity across the electrode channels. The total recall (mean ± SD) of the activation detection module was 97.1 ± 0.03%. The success rate for online micro-gesture recognition is shown in Fig. 7. The subjects achieved high performance with an average success rate (median (IQR)) of 82 (12) % on the first attempt, which increased significantly withmultiple attempts, namely, to 94 (5) % within two attempts and above 98% for more than two attempts. Across micro-gestures, success rates on the first attempt were similar, with no significant differences between classes. The lowest median success rate was achieved for the thumb-up movement (80 (24) %). Moreover, as illustrated by the confusion matrix (Fig. 7c), the thumb up and down movements were almost never confused with each other, whereas other error combinations were occasionally observed (e.g., thumb right mistaken for thumb left, thumb up with thumb right, etc.). These errors are not surprising considering the nature of micro-gestures (small transient movements) leading to a degree of similarity in EMG patterns (Fig. 6). Fig. 8Summary of results for the closed-loop control task using tactile feedback. The boxplot shows the distribution of success rates for the reinforced learning (RL, light blue) and validation (yellow) phases. Two success rates are provided: one for reaching the target (within-0) and one for reaching a cell next to the target (within-1) Summary of results for the closed-loop control task using tactile feedback. The boxplot shows the distribution of success rates for the reinforced learning (RL, light blue) and validation (yellow) phases. Two success rates are provided: one for reaching the target (within-0) and one for reaching a cell next to the target (within-1) Results of the closed-loop control task are shown in Fig. 8. The success rates in the reinforced learning and the validation phases were not significantly different (Wilcoxon signed-rank test). Subjects completed the validation phase with a success rate of 70 (23.3)% for landing exactly on the target (within-0 error) and 95 (4.6)% when finishing the trial in a cell next to the target (within-1 error). An example trajectory traced through the grid during the online control task with electrotactile feedback only is depicted in Fig. 5b, showing that the subject generated a movement in the wrong direction (t2) but then detected it through the feedback and corrected, reaching the target in several steps. The results (median (IQR)) of the path efficiency were 1 (0.6), 1 (0.25), 1 (0.6), 0.66 (0.5), and 1 (0.4) for distances from one to five, respectively. Overall, the participants were efficient (median of 1) but often took a few more steps than necessary (IQR) to complete the task compared to the optimal path. This suggests that participants compensated for misclassified micro-gestures by detecting steps in the wrong direction using the tactile feedback. Discussion In this work, we introduce a novel system for bidirectional human–machine interfacing that integrates both high-density EMG detection and electrotactile stimulation within a single wearable platform. Different systems have been presented that provide wearable solutions for high-density EMG pattern recognition in recent years. In [16], a grid of rigid sensors with integrated electronics was developed for hand gesture recognition, and in [39], the authors presented a textile-based sleeve to measure EMG activity from a large surface of the forearm in a clinical setup. However, these interfaces are unidirectional, limited to recording functionality. A compact system for simultaneous electrotactile stimulation and EMG recording was recently developed by our group and presented in [31], but with a reduced number of channels, separate recording and stimulation electrodes, and limited spatial resolution for stimulation. Alternatively, the NeuraLoop system features a dense array of flexible pads with mixed functionality (recording or stimulation) distributed around the forearm on the same electrode. NeuraLoop is envisioned as a general-purpose interface applicable to various use cases, including virtual and augmented reality, robotics, and rehabilitation. It could also be adapted for integration into wearable systems for applications in prosthetic and exoskeleton control. Its wearable, textile-based form and use of dry electrodes ensure easy setup, comfort, and usability in a wide range of scenarios, while its high-resolution sensing and stimulation capabilities make it suitable for both research and general-purpose applications. The neoprene sleeve was designed as a “one-size-fits-all” solution, featuring an adjustable hook-and-loop strap to accommodate a wide range of forearm sizes. However, this design presented limitations, particularly for smaller forearms. Since the printed TPU material lacks stretchability, a forearm that is too small may cause some electrodes to overlap, requiring additional post-processing to remove such channels. Conversely, if the forearm is too large, the distance between edge electrodes increases, which could compromise muscle coverage. We are currently working on developing stretchable electrodes to better adapt to varying forearm sizes. A simpler and more immediate solution could be to offer the sleeve in three different sizes to ensure a more tailored fit. Additionally, the recording and stimulation pads could be integrated into the socket liner for application in prosthetics using screen printing or lamination, while the electronics could be housed inside the socket. An immediate advantage of an integrated matrix electrode is that precise identification of electrode location and related socket customization for each user could be avoided. However, the size of the electronics would likely need to be further reduced, which could be achieved by decreasing the number of recording and/or stimulation channels. The experimental assessment showed that the system can be used for closed-loop, high-bandwidth human–machine interfacing that includes online movement intent recognition to generate commands and simultaneous delivery of electrotactile stimulation for haptic feedback. The results demonstrated that NeuraLoop enabled the recognition of subtle and transient finger movements (micro-gestures), such as thumb swipes, with a high success rate. Additionally, we demonstrated that spatially distributed electrotactile stimulation across the matrix of pads can be used as an effective haptic feedback channel during a closed-loop online control task, leading to a high success rate in reaching the target (or close to it). In this task, stimulation was used to intuitively convey the position of a virtual cursor moving through a two-dimensional grid using spatial encoding. Unlike the recognition of full-hand gestures (e.g., hand opening, pinch, lateral and palmar grasps), which has been the focus of research for decades [40], the recognition of micro-gestures is a relatively novel problem and crucial for applying myoelectric control in more general settings. Transient segments of EMG signals have been shown to contain relevant information for decoding movement intention of the upper limbs [41] or predicting force contraction [42] before movement completion. This approach can be used to provide a faster system response in combination with standard steady-state EMG classification. However, the rapid nature of micro-gestures implies that only transient EMG is available for predicting movement. To capture and recognize the micro-gestures, we introduced an activation detection block in the classification pipeline. Once the movement was detected, a single 250 ms data window was used for classification. The window size was selected as a trade-off to ensure the entire transient EMG was captured while risking inclusion of rest EMG data if the movement was performed too quickly. This could be improved by implementing an adaptive window size. In this case, the exact portion of EMG corresponding to muscle activation would be segmented by waiting for the movement onset and detecting its end. The delay in this case would depend on movement duration, which might not be significant compared to the fixed-window approach, particularly in faster thumb swipes (e.g., less than 150 ms). A systematic investigation of the optimal window size will be conducted in future work. The overall performance confirms the feasibility of using the system for recognizing subtle finger movements, which are particularly challenging to detect and classify due to their transient nature. The results demonstrated that the participants were able to complete the online control task using four thumb movements, achieving high success rates. Even if the first attempt failed, the micro-gesture was successfully recognized in most cases after the second attempt (>94 (5)%). These results are in line with other studies investigating micro-gestures. In [17], the authors achieved an average accuracy of 95% for offline recognition of 14 micro-gestures. However, the authors did not report an online assessment of their methods. Most recently, in [43], the authors proposed using a snap gesture to toggle hand-gesture myoelectric control on and off, which significantly reduced false activations during daily activities. Overall, few studies have investigated online control using discrete micro-gestures. Detecting and classifying thumb movements from forearm EMG signals can be particularly challenging because a significant contribution to the movement is produced by intrinsic hand muscles. As demonstrated in [44], combining EMG recordings from both intrinsic and extrinsic muscles significantly improves the classification performance for various grasping gestures. Furthermore, as shown in [45], EMG signals recorded directly from hand muscles can be used to decode a wide range of gestures, including thumb flexion and adduction. Although most musculature responsible for thumb movements is contained within the hand, the results of the present study show that forearm EMG activity provided sufficient discriminative information to detect and classify these specific micro-gestures. This is likely due to the activity of extrinsic muscles that indirectly support the micro-gesture by stabilizing the fingers and facilitating movement execution. In this case, high-density EMG can capture such small variations by leveraging its wider spatial resolution, thereby improving decoding of challenging motor patterns. Although not significant, the downward swipe achieved a somewhat higher and less variable success rate compared to the other three thumb movements. This result is probably due to the significant presence of EMG activity from the extrinsic thumb muscle flexor pollicis longus, which is actively involved in the execution of the movement producing a distinguishable pattern, as shown in Fig. 6. Conversely, the micro-gestures corresponding to right, left and upward directions (involving abduction, adduction, and extension movements, respectively) predominantly engage intrinsic muscles, including the abductor pollicis, abductor pollicis brevis, adductor pollicis, opponens pollicis and flexor pollicis brevis, which contribute more substantially to these specific motor tasks. Extensor pollicis longus is also an extrinsic muscle controlling thumb extension but might have been more difficult to detect. Future directions in micro-gesture classification include detecting not only the type of gesture but also its intensity for proportional control, enabling more natural interactions. Expanding the range of detectable micro-gestures by including, for instance, pinches, rotations, or compound finger movements, could enhance control in applications like augmented and virtual reality and human–computer interfaces. These subtle gestures, though driven largely by intrinsic hand muscles, can still be decoded from forearm EMG signals, especially with high-density arrays that capture fine spatial muscle activity variations. As sensor technology and embedded machine learning algorithms advance, EMG-based systems can become more adaptive and personalized, offering intuitive control solutions for both general users and individuals with motor impairments. One potential avenue for exploration is relocating the NeuraLoop sensors and actuators from the proximal forearm to the wrist. This adjustment may offer practical benefits, as it aligns with the widespread consumer adoption of wrist-worn wearable devices, thereby enhancing user familiarity and comfort. Moreover, recent studies have demonstrated that wrist-based EMG signals can effectively predict individual finger flexions and extensions [46, 47] and several hand gestures [48]. However, further investigation is required to evaluate the feasibility of detecting micro-gestures with this approach. In addition to micro-gesture recognition, we conducted preliminary tests of the system's electrotactile stimulation capabilities. The calibration process requires adjusting the intensity manually for each pad. Therefore, the total calibration time depends on the spatial resolution of feedback (number of pads) required by the envisioned application (e.g., 12 pads were used in the present study). This is a known challenge when using high-density electrotactile stimulation, but methods are being developed to address it [49]. The results of the closed-loop control task demonstrated that the NeuraLoop system could effectively convey continuous information about cursor position along two degrees of freedom using congruently spatially distributed tactile feedback. The subjects were able to determine the position inside the grid without visual information with reasonable accuracy. The outcome measures of two error levels (within-0 and within-1) were introduced because, in practical applications, some error can usually be tolerated – in such scenarios, it could be enough to navigate the system in the ballpark of the target (e.g., when controlling the wrist angle of a prosthesis, a certain error can be tolerated since it can be compensated by body movements). The initial results are promising, suggesting that the system can deliver meaningful and distinguishable tactile cues across multiple locations on the forearm. Although we did not systematically collect any qualitative feedback from participants regarding system use, participants easily understood how to employ micro-gestures to control the ball during the online task. Similarly, the grid task required little or no additional explanation. This was because the micro-gestures were intuitively related to the commands (moving the ball/cursor). The most challenging aspect appeared to be synchronizing thumb movements with the pace of TDM windows. These results highlight the potential of the NeuraLoop system for real-world applications, particularly in assistive technologies such as prosthetic devices. By using spatial encoding similar to that in the present study, the system could intuitively convey positional information about a prosthetic device, such as wrist orientation, enabling users to perceive limb posture without visual contact [50]. For instance, the activation of the pads in the transversal direction to the forearm could be used to convey the wrist angle, while the proximal-to-distal pad activation could be used to transmit the hand aperture. Although this study focused on spatial haptic cues, the system also supports intensity and frequency modulation, which have yet to be explored. These additional encoding strategies could be used independently or combined with spatial encoding to transmit richer, multidimensional information such as force, velocity, or proximity simultaneously with positional information. For instance, stimulation intensity can be used to convey force feedback from the prosthetic hand to facilitate grasping compliant objects [51]. This multimodal feedback approach could significantly enhance closed-loop control, offering users a more natural and intuitive interaction with the environment. One limitation of the system was associated with implementing TDM, which required precisely matching the timing of movement execution with the recording window. Such timing precision is not an issue for the classification of conventional hand gestures, where the muscle contraction can be naturally held for longer periods. However, in the case of transient, almost ballistic movements such as thumb swipes, the short duration of the EMG burst requires producing the movement entirely inside the recording window, which can be challenging. Additionally, stimulation (tactile feedback) could be perceived after a slight delay (< 1 s) relative to the movement detection, which could negatively affect the control dexterity. The present study showed that time-division multiplexing is a simple yet effective strategy to manage simultaneous stimulation and recording. Nevertheless, implementing a blanking strategy could help eliminate delays, thereby improving system responsiveness. As an alternative to TDM, blanking has been investigated to block the electrical pulses in synchrony with the stimulation frequency [33]. This approach removes the exact segments of the signal when stimulation occurs, allowing continuous-like stimulation without perceiving interruptions. We anticipated that a linear classifier would be sufficient for this task, given the high dimensionality of our feature space (i.e., the large number of channels and extracted features). Indeed, the SVM provided reliable accuracy and computational efficiency in pilot tests. While we acknowledge that other linear and nonlinear machine learning models may also offer strong performance, a comprehensive comparison of classifiers was beyond the scope of the present study, which primarily aimed to demonstrate the feasibility of our novel technology. Future work should include a systematic evaluation of alternative classifiers and the required number of channels to determine the optimal approach for micro-gesture classification. Overall, the results of this study demonstrate the potential of high-density, bidirectional interfaces for advancing human–machine interaction. Systems like NeuraLoop could enable more intuitive, responsive, and accessible control of digital and robotic devices. It is worth noting that this study was exploratory and that several parameters, such as data window size and classifier type, were not exhaustively optimized. Conclusions This study introduced NeuraLoop, a high-bandwidth, bidirectional interface that integrates EMG recording and electrotactile stimulation within a textile-based platform. The system enabled accurate recognition of thumb micro-gestures and effective delivery of spatially encoded tactile feedback, providing intuitive closed-loop control. These results highlight NeuraLoop’s potential for applications in prosthetics, rehabilitation, and immersive virtual environments. Future developments will focus on improving energy efficiency to extend battery life, allowing users to perform daily-life applications that require prolonged, continuous, and simultaneous recording and stimulation, without interrupting device usage. While the system showed promising results, certain limitations remain. The use of time-division multiplexing requires precise timing of gesture execution, which can be challenging for fast, transient movements such as thumb swipes. Additionally, the current implementation focused on a limited set of micro-gestures and spatial feedback. Future work will explore proportional and multimodal feedback encoding, expand the gesture set, and investigate pattern recognition methods to optimize classification of transient EMG signals associated with micro-gestures. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.