Classification of forearm muscle forces for controlling assistive devices

Force myography (FMG)-based wearable interfaces have emerged as promising solutions for non-invasive human–machine interaction (HMI) in assistive and rehabilitation applications. However, achieving reliable real-time gesture recognition using FMG is challenging due to the need for a complex sensor setup and computational algorithms. Moreover, the FMG system suffers from sensor shifting and improper force transmission. Therefore, this work proposes an affordable, multi-channel wearable FMG armband that reliably captures forces at multiple forearm muscle locations during different gestures. A real-time gesture recognition system classifies the FMG data from the armband into discrete instructions for assistive devices. An optimized four-channel sensor configuration for the armband was designed to capture distinct muscle deformation patterns associated with hand gestures. FMG signals were recorded from 10 subjects for their five different hand gestures. The recorded data underwent preprocessing, feature extraction and selection, and then machine-learning-based classification. Five classifiers, Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP), were used to recognize these five gestures. The RF classifier achieved the highest offline classification accuracy of 92% and was subsequently implemented in real time on an ESP32 microcontroller unit (MCU). The system, based on an RF-trained model, depicted an accuracy of 81% in identifying real-time hand gestures. The implemented scheme was further tested to control the movement of an indigenous robotic arm using the predefined gestures. Such a simple, low-cost, and effective scheme based on FMG can help control prosthetic and robotic devices using human intentions.

Gesture recognition has gained significant attention in recent years due to its growing applications in assistive robotics, prosthetic control, rehabilitation systems, and wearable healthcare devices (Zhang et al., 2025; Zhou et al., 2024; Rohr et al., 2025; Zhou and Alici, 2022; Kumar et al., 2021). Such a scheme based on wearable interfaces enables intuitive, natural communication between humans and machines by decoding forearm muscle activity generated during hand movements (Qadir et al., 2021). These systems are particularly important for upper-limb amputees and rehabilitation patients, for whom reliable real-time control of assistive devices is required (Moqadam et al., 2022; Song et al., 2022; Mizrahi and Sintov, 2024; Jiang et al., 2017; Zakia and Menon, 2022; Missiroli et al., 2025; Kahanowich and Sintov, 2021). Therefore, the development of low-cost, wearable, and robust gesture recognition systems has become an active research area in biomedical engineering and assistive technology.

Various sensing modalities have been investigated for wearable gesture recognition, including surface electromyography (sEMG), inertial measurement units (IMU), mechanomyography (MMG), and force myography (FMG) (Zhang et al., 2025; Rohr et al., 2025; Zhou and Alici, 2022; Wang et al., 2022; Belyea et al., 2019; Mizrahi and Sintov, 2024; Jiang et al., 2017; Stefanelli et al., 2026; Missiroli et al., 2025; Zhang et al.,2026). Among these, sEMG has been widely used in prosthetic and rehabilitation applications due to its ability to measure electrical muscle activity. However, sEMG-based systems are highly sensitive to electrode shift, sweat, motion artefacts, and other electrical noise, which affect the robustness and repeatability of real-time operation (Moqadam et al., 2022; Choi et al., 2021; Chen et al., 2023; Oyemakinde et al., 2025). In addition, sEMG systems generally require careful skin preparation and precise electrode placement. IMU and accelerometer-based systems are effective for detecting limb motion but provide limited information regarding muscle contraction patterns. In contrast, FMG measures the mechanical deformation of forearm muscles during contraction using pressure or force sensors (Jiang et al., 2017; Zakia and Menon, 2021). FMG-based systems offer several advantages, including a high signal-to-noise ratio, low computational complexity, reduced sensitivity to skin conditions, simple sensor integration, and good wearability. These advantages make FMG a promising alternative for wearable human–machine interaction systems requiring low-cost, real-time implementation (Zhou et al., 2024; Fujiwara et al., 2021; Zhou and Alici, 2022; Cho et al., 2026; Wang et al., 2022; Jiang et al., 2018; Belyea et al., 2019; Mizrahi and Sintov, 2024; Godiyal and Joshi, 2021).

Recent studies have demonstrated the effectiveness of FMG and hybrid sensing approaches for wearable gesture recognition systems. Zhang et al. (2025) developed a wearable human–machine interface using collocated EMG and pressure-based FMG sensing for reliable hand gesture recognition. Zhou et al. (2024) proposed a 3D-printed pressure-based FMG armband capable of multipurpose human–machine interaction for assistive applications. Wang et al. (2022) introduced a multimodal wristband integrating PPG, FMG, and accelerometer sensors for robust hand gesture recognition. Dali et al. (2023) developed eco-friendly piezoelectric-based FMG sensors for wearable gesture classification applications. Xiao et al. (2024) investigated optimal FMG channel selection and sensor placement for reducing the number of sensing channels while maintaining high classification accuracy. Rohr et al. (2025) demonstrated that combining FMG and EMG signals improves cross-subject gesture recognition performance due to the complementary information provided by the two sensing modalities. Belyea et al. (2019) compared FMG and EMG for real-time pattern-recognition-based control and reported that FMG offers improved usability and robustness in wearable gesture recognition systems. Jiang et al. (2018) investigated the effect of force exertion variations on FMG-based grasp classification and demonstrated that variations in grasping force significantly influence recognition accuracy. Belyea et al. (2019) compared FMG and EMG for real-time pattern recognition-based control and reported that FMG offers improved robustness, usability, and user comfort in wearable gesture recognition systems. Fang et al. (2025) proposed a hybrid sEMG–FMG sensor fusion framework using cascade fuzzy forest classification for gesture recognition under muscle fatigue conditions, demonstrating improved reliability during prolonged operation. Bamani et al. (2022) developed a wearable FMG-based framework for robust multi-user in-hand object recognition in human–robot collaboration applications and demonstrated reliable performance across different users. Although these studies have shown promising results, many existing systems still rely on numerous sensors, complex sensor fusion frameworks, or computationally expensive processing architectures, thereby limiting their suitability for compact, low-cost, embedded, real-time implementations.

In this work, a low-cost multichannel FMG-based wearable armband is developed for real-time hand gesture classification. Unlike conventional systems that use numerous sensing channels, the proposed system uses an optimized four-channel FMG configuration based on the activation characteristics of the major forearm muscle groups, including the Flexor Carpi, Extensor Digitorum, and Extensor Carpi. The developed framework includes signal preprocessing, feature extraction, feature selection, and machine learning-based gesture classification. Several classifiers, including Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP), were used for recognizing five hand gestures. Furthermore, the trained Random Forest classifier was deployed on an ESP32 microcontroller for embedded real-time implementation. Experimental results demonstrated that the proposed system achieves reliable gesture recognition performance while maintaining low system complexity, good wearability, and suitability for assistive human–machine interaction applications.

Proposed scheme

Figure 1 shows the overall framework of the proposed FMG-based hand gesture recognition system. The primary stage of work involves developing a multi-channel FMG armband to capture muscle forces at different forearm locations. The developed armband was then deployed to acquire FMG data from distinct subjects to capture their hand gestures. The raw recorded FMG data were fed into a machine learning pipeline that included feature preprocessing, extraction, selection, and classification. The hand gestures were classified using multiple machine learning models, including Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbours (KNN), and Multi-Layer Perceptron (MLP). Finally, based on classification accuracy, the best-performing classifier was applied for real-time control.

Armband fabrication

A custom multichannel wearable FMG-based Armband was fabricated to detect forearm muscle deformation during hand gestures. The armband mainly comprises four piezoresistive force-sensing resistors (FSRs), each mounted inside a suitable casing. Direct placement of FSR on the forearm surface may result in inappropriate transmission of muscle contractile force to its sensitive area and may also cause improper bending, leading to ambiguous output. Therefore, custom sensor casing models were designed in Autodesk Fusion 360 and converted into physical structures via 3D printing. Each casing structure shown in Figure 2a consisted of a rigid base, a cylindrical coupler and a top enclosure. The rigid base, with dimensions 24 mm × 24 mm × 1.8 mm, had a circular cavity 18 mm in diameter, 1 mm deep for the precise mounting of the FSR. A cylindrical coupler with a diameter of 15 mm and a height of 3.5 mm was placed directly on the FSR-sensitive area to effectively transmit muscle force. A top enclosure with a circular cutout 16 mm in diameter completes the casing by covering the structure. One end of the coupler was kept slightly convex to receive contractile force from the muscle, while the other end was flat to transmit this force to the sensitive area of the FSR. An armband (in Figure 2c) was formed by arranging four such sensor casings (Figure 2b) on a Velcro tape with plastic strings, allowing the position of each sensor (casing) to be adjusted. The positions of these four sensors can be selected based on their proximity to the distinctive muscle groups of the Flexor Carpi, Extensor Digitorum, and Extensor Carpi to uniquely distinguish hand gestures.

Sensor calibration

To evaluate the working range, sensitivity, and saturation features of these developed sensing units, calibration was performed. A conditioning circuit consisting of a voltage divider followed by a buffer circuit (Figure 2b) was incorporated into each sensing unit to convert a change in the FSR’s resistance into a voltage signal. Calibration was performed by applying a variable load from 0.1 kg to 6 kg to the sensing unit and recording the corresponding voltage outputs. A calibration curve obtained in Figure 3 shows the variation of voltage with applied force (in kgf). A nonlinear relationship was observed between force and voltage, reflecting the piezoresistive nature of the FSR. To select the working range of each sensing unit, the armband was tested on human subjects performing various gestures and a maximum voluntary contraction (MVC). For MVC, the forearm muscle exerted a maximum force of 2 kg. f on the sensing unit, producing a maximum output of 2.8 V. This procedure revealed that 0–2.8 V was the useful range of the sensing unit. In this range, the sensor can be effectively utilized to measure forearm muscle deformation during distinct gestures without saturating.

Data acquisition protocol

The data acquisition included 10 healthy participants: six males and four females. The FMG armband developed in this work was worn on the forearm, with the four sensors placed over the Flexor Carpi, Extensor Digitorum, and Extensor Carpi muscles. These muscle groups in the forearm (shown in Figure 4) are responsible for wrist and finger movement, i.e., hand gestures (Zhang et al., 2025; Rohr et al., 2025). The sensors were positioned to ensure continuous skin (in the vicinity of the muscle) contact with the sensing units during gesture execution. Participants were directed to perform five hand gestures: fist, fist flexion, open hand, wrist extension, and wrist flexion, shown in Figure 5. The selection of these gestures was based on the generation of distinct forearm muscle deformation patterns and on human-machine interface applications. Hand gestures were primarily selected based on the activation behaviour of the targeted forearm muscles. In preliminary tests, selected gestures were found to yield distinguishable FMG signals with a small range of overlap among the four channels. These gestures also ensure that the major contractions can be measured using a limited number of channels, thereby reducing hardware costs. In data acquisition, all the gestures were to be performed by the user in sequence under static conditions. Each gesture was performed for approximately 5 s with a 2 s rest period between each consecutive performance of the same gesture. An average rest period of 5s was maintained between gestures to prevent muscle fatigue and allow acquisition of a steady signal. Each gesture was repeated 10 times to capture variations across trials. The acquisition was performed using a DAQ device (USB-4716) at a sampling frequency of 1 kHz, with all sensing channels supplied from a regulated 5 V supply. All the data were recorded for an offline analysis.

Data preprocessing

The preprocessing techniques were applied to the FMG signals from the four-channel wearable armband to remove noise and fluctuations and maintain consistency in the recorded data. FMG signals have lower noise and fluctuation than other biological signals (Zhou et al., 2024; Xiao et al., 2024). Hence, a Savitzky-Golay low-pass smoothing filter was applied to the FMG signals to remove random noise and motion artefacts, thereby preserving key information about muscle deformation patterns while reducing signal variation. Preprocessing also helped in standardization and uniformity across all the recorded data files obtained from different subjects and gesture trials. After the FMG signals were filtered, they were segmented using a sliding window with a 250 m window length and a 50 m step size. Such window parameters provide the time-domain information required to extract features while maintaining low latency, suitable for real-time gesture recognition. Standard scalar normalization was then applied to the segmented signals to eliminate intersubject variation and make the feature distributions across channels more comparable.

Feature extraction and selection

Feature extraction is crucial in FMG-based gesture recognition because the reliability and performance of the classification strongly depend on the quality of the extracted features. Because differences in forearm structure, muscle activation patterns, sensor placement, and muscle deformation among individuals affect the FMG signal, the selection of appropriate features should capture the muscle activity pattern associated with a specific hand gesture. Since FMG signals consist primarily of low-frequency muscle deformation information with relatively smooth waveforms, using time-domain features is considered computationally efficient and desirable for wearable systems, where real-time operation is important.

From each FMG segmented signal for each of the channels, a set of ten time-domain features described in Table 1 was computed. These features include the mean, median, variance, standard deviation, minimum value, maximum value, Integral absolute value (IAV), mean absolute value (MAV), slope sign change (SSC) and root mean square (RMS) which were chosen as they provide a good indication of the variations in the amplitude and energy of the signal as well as of the shape of the waveform and of the behavior of the muscle contraction for each of the hand gestures. For four different armband channels, a total of 40 features were extracted for each segmented window.

It is noted that adding more features may improve signal representation; however, less informative or redundant features may increase computational complexity and degrade classification accuracy. Therefore, feature selection was performed subsequently to identify the most discriminative and informative features for gesture recognition.

First, the ANOVA F-test was applied to assess the statistical significance of the extracted features and to examine the discriminative power of individual features across classes. The ANOVA F-test compares interclass variance with intraclass variance. Afterwards, RFE (Recursive Feature Elimination) was applied to the set of selected features, iteratively removing the weakest features to find the most relevant feature subset. Five feature subset sizes (15, 20, 25, and 30 features) were tested, and the 20-feature set selected using RFE achieved the highest accuracy while reducing feature and classifier complexity. This feature subset was used to classify gestures using machine learning classifiers.

The extracted feature values were normalized using the StandardScaler method. Each feature was transformed using the Z-score transformation, where x denotes the original feature value, µ denotes the mean, and σ denotes the standard deviation used by the scaler. This normalization was applied to ensure that the extracted features were represented on a comparable scale before the machine-learning classification stage. The same fitted scaler was retained and used during subsequent prediction so that the real-time input features were processed consistently with the offline-trained model. The classification performance was subsequently evaluated using K-fold cross-validation, with the same preprocessing procedure applied consistently within the cross-validation framework. For real-time implementation, the fitted scaler from the trained RF model was retained and applied to incoming feature vectors before classification.

Classification

The features selected after the feature selection process were used as input for the ML-based gesture classification. This paper implements several supervised ML algorithms, including Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP), for hand gesture recognition. These classifiers are appropriate for dealing with the nonlinear and multidimensional feature space of FMG signals.

Offline classification

Offline classification has been performed on the created dataset using 10-fold cross-validation on a machine learning classifier so that the generalization capability and robustness of the model could be observed. In 10-fold cross-validation, the dataset was partitioned into ten subsets. Each time, nine subsets were used for training, and the remaining subset was used for testing; this was repeated until all subsets had been tested. Then evaluate the performance of classifiers based on accuracy, precision, recall and F1-score. This offline classification component could assess the power of the extracted FMG features and compare the classification performance of various machine learning classifiers. In offline classification, the Random Forest achieves the best accuracy.

Real-time classification

Selection of the best classifier was performed by referring to the result of offline classification. The trained Random Forest model was deployed into an ESP32 microcontroller for real-time hand gesture recognition. After receiving the FMG signal, the real-time hand gesture was processed using the same feature extraction method as in the offline phase, and the feature vectors were used in the trained classifier. By detecting gesture data in real-time, LED blinking was controlled to realize a human-machine interface system using FMG signals.

Robotic arm control

The proposed real-time classification system was further implemented to control the operation of a 4-degree-of-freedom (DoF) indigenous robotic arm as a proof of concept. The five discrete outputs of the classifier were used to control the position of servomotors coupled to a robotic arm. The five different hand gestures: fist, fist flexion, wrist flexion, wrist extension, and open hand were used to perform a pick and place task for the robotic hand.

Armband output

The recorded armband output for a subject performing five gestures is shown in the plot in Figure 6. All these patterns obtained with individual sensors were highly distinguishable for distinct gestures.

Offline classification

The classification of hand gestures was performed by five different machine learning models: Random Forest (RF), Support Vector Machine (SVM), XGBoost, Multi-Layer Perceptron (MLP), and K-Nearest Neighbours (KNN). The normalized extracted feature vectors are input to all classifiers to discriminate between hand gestures using the pattern from the FMG signals.

Random Forest is used for classification and builds multiple decision trees, aggregating the predictions by majority vote. Hyperparameters were tuned using RandomizedSearchCV for the number of estimators, maximum tree depth, and class weight balancing. The accuracy score achieved by Random Forest was 91.81%, the best among all classifiers. It also shows good resistance towards variations and noise present in the features.

The SVM classifier performs classification by finding an optimal hyperplane that maximizes the margin between gesture classes in the feature space. Due to the nonlinearity of the FMG signals, a radial basis function kernel was employed to enhance the class separability. Optimized values for the kernel type, regularization parameter, and gamma were selected to improve the performance of the SVM classifier for multiclass gesture recognition, achieving an accuracy of 89.32%.

Extreme Gradient Boosting (XGBoost)is an ensemble boosting algorithm that builds a sequence of weak decision tree learners in which each decision tree corrects the classification error of the past tree. With the help of the gradient boosting optimization and regularization method to prevent over-fitting, XGBoost can enhance the classification accuracy. Using hyperparameters such as the learning rate, maximum tree depth, and number of estimators, the accuracy achieved here was 89.36%.

Multi-Layer Perceptron classifier is an artificial neural network-based classifier, which has an input layer, one or more hidden layers, and an output layer to model the nonlinear function between features and classes. The MLP classifier we utilized in this research had fully connected hidden layers and was activated with a nonlinear activation function, and the learning algorithm was backpropagation to update weights. The hidden layer, learning rate, and activation function were chosen empirically to improve the classification performance. The MLP achieved an accuracy of 90.10%.

The K-Nearest Neighbours classifier is a supervised, distance-based algorithm that assigns class labels to new data samples based on the majority class among the K nearest feature vectors in the multidimensional feature space. Euclidean distance has been used as a distance measure. The number of nearest neighbours, K, is selected by trial and error until the classifier’s performance becomes relatively stable. The KNN algorithm has a simple implementation and good classification performance, but its complexity grows rapidly with dataset size. The KNN accuracy was 90.00%.

All classifiers used in the hand gesture recognition system were evaluated using standard performance metrics, including accuracy, precision, recall, and F1-score. A comparison of classifiers (Table 2) was conducted to identify the most appropriate classifier for recognizing hand gestures by FMG and for real-time use. The confusion matrices for all classifiers are presented in Figure 7.

Real-time classification and control

The Random Forest (RF) classifier was chosen for real-time implementation because it achieved the highest classification accuracy during offline evaluation among all the trained ML models. Its outstanding classification performance resulted from its ensemble of weak learners, which is less prone to overfitting and well-suited to the multi-dimensional FMG feature space. After training, the RF model was converted into Arduino-compatible C/C++ code for a microcontroller, and the trained parameter values and decision tree structures were exported as a header file (.h).

The exported header file was then deployed to the ESP32 embedded code (.ino format) using the Arduino IDE. FMG signals from a four-channel wearable armband were fed to the ESP32 microcontroller in real time during online implementation. The real-time feature extraction stage was implemented using the same signal-processing procedure as in the offline implementation to ensure consistency between training and deployment. The extracted real-time feature vectors are then sent to the trained RF classifier within the ESP32 microcontroller.

Different control signals for controlling the LED were produced by the ESP32 based on the predicted gesture class. Each hand gesture was mapped to a unique LED activation, allowing real-time blinking for the predicted gesture class. For real-time gesture recognition, an evaluation of the designed system was performed on five different subjects, with two males and three females, to assess the real-time behaviour and resilience of the proposed system under real-world scenarios. Figure 8 shows real-time control of LED using fist, fist flexion, open hand, wrist flexion and wrist extensor gestures of a subject. The real-time evaluation consisted of 250 trials across five subjects, five gestures, and ten trials per gesture per subject. A trial was considered successful when the intended gesture was correctly recognized and the corresponding LED was activated within 3 s of gesture initiation. Whereas A trial was considered unsuccessful when the intended gesture was not correctly recognized within 3 s, and an LED corresponding to another gesture was activated. The system successfully recognized 200 out of the 250 trials, resulting in a real-time trial completion/success rate of 80%. A confusion matrix for the RF classifier to perform real-time control of the LED is shown in Figure 9. This successfully demonstrated that low-cost FMG-based gesture recognition is feasible for assistive devices.

Gesture-controlled operation of a robotic arm

Five predefined hand gestures were used to enable controlled operation of the robotic arm. Figure 10 shows a sequence of pick-and-place operations by the robotic arm using these gestures. (a) Initially, the robotic arm was held in a static position to locate the target object. (b) The Fist gesture performed-robotic arm moves downward to grab the target object. (c) The Fist flexion gesture performed-the robotic arm comes to the initial position after grabbing. (d) The wrist flexion- The robotic arm rotates. (e) The wrist extensor- The robotic arm moves downward to release the object. (f) The open hand- The robotic arm returns to its initial position.

The proposed 4-channel wearable FMG-based gesture recognition system showed excellent offline and real-time classification performance for human-machine interaction applications. The time-domain FMG features extracted from a 4-channel wearable armband effectively characterized the forearm muscle deformation pattern corresponding to different hand gestures and yielded high classification performance by using typical machine learning classifiers. Random Forest classifier showed the highest offline classification accuracy 91.81% among all the evaluated classifiers, proving its capability to model non-linear and high-dimensional FMG feature spaces. Using an ANOVA- and RFE-based feature selection approach, the feature dimensionality was significantly reduced while maintaining comparable classification performance, making the model efficient to embed in the target system. Compared to computationally intensive frameworks such as complex deep neural network architectures or multimodal sensor-based gesture recognition, the current system employs only four FMG channels, making the hardware simple and the device portable enough to wear on the forearm.

Table 3 compares the proposed system with recently reported wearable gesture recognition approaches. Several previous studies achieved high-classification accuracy (92%–96%) using large sensor arrays, hybrid sensing modalities, or computationally intensive frameworks (Zhang et al., 2025; Fujiwara et al., 2021; Moqadam et al., 2022; Rohr et al., 2025). Some systems with only a single sensing modality achieved gesture recognition accuracy exceeding 94% (Zhou et al., 2024; Xiao et al., 2024). As the number of sensor channels or sensing modalities increases, so does computational complexity and the overall cost of the recognition system. Software complexity prevents the direct implementation of a trained machine learning algorithm on a low-cost, limited-storage microcontroller unit. Certain schemes using FMG and FMG-EMG achieved offline accuracy of 90%–92%; however, they were infeasible for real-time implementation (Rohr et al., 2025; Dali et al., 2023). The proposed work with only a four-channel FMG has demonstrated decent accuracy in both offline and real-time settings for recognizing hand gestures. The deployment of trained ML models on an embedded microcontroller unit (ESP32) greatly reduced computational complexity, making the system feasible for affordable wearable-assistant and rehabilitation applications.

To implement the real-time recognition system, the trained Random Forest model was converted into Arduino code and uploaded to the ESP32 microcontroller. Real-time preprocessing, feature extraction, and gesture prediction were performed sequentially using the real-time input signals captured by the wearable armband. Predicted gesture output was used to control the LEDs, demonstrating the system for human-machine interaction. When tested on five subjects, the developed system achieved an overall real-time classification accuracy of 80%. Despite a decrease in accuracy compared to offline classification results, which is usually a normal case considering the influence of inter-subject differences, variation of sensor placement on different persons, and noise and disturbances occurring during actual operation, the stable operation demonstrates the feasibility of employing a low-cost wearable FMG-based gesture recognition system for human-machine interaction and rehabilitation.

The FMG armband deployed in this work utilizes piezoresistive based force-sensing resistors (FSRs) to measure the mechanical changes, i.e., muscle contractile force during contraction. The FSR can provide force detection starting from 10 g to several kg. The use of resonant force sensors based on ionic polymer metal composites can further extend the FMG armband’s lower force-sensing range (up to 0.1 g) to detect lower-level contractions (Bonomo et al., 2007).

This paper presents a cost-effective, four-channel, wearable FMG-based hand-gesture recognition system for real-time human-machine interaction. The developed FMG armband was used to measure forearm muscle deformation associated with various hand gestures, employing a few sensing channels to reduce hardware complexity and improve wearability. Multiple ML classifiers, including RF, SVM, KNN, XGBoost, and MLP, for gesture recognition have been implemented. RF classifier obtained the best offline classification accuracy of 91.81%. The obtained RF classifier was then embedded on the microcontroller (ESP32) to achieve real-time hand gesture recognition, with an accuracy of 80% when tested across different subjects. The real-time system demonstrated robust gesture-control functionality by activating different LEDs, thereby demonstrating the feasibility of implementing a cost-effective embedded FMG-based wearable interface for assistive or rehabilitation purposes.

During the implementation of the proposed system, several practical challenges were encountered, including variability in sensor placement across subjects, maintaining consistent contact pressure between the FMG sensors and the skin surface, fluctuations in real-time muscle-contraction signals, and limitations in the embedded microcontroller’s computational capabilities. Each sensor position in the armband was adjustable with plastic strings, which helped address variability in sensor placement across subjects. Stable, isometric muscle contractions were maintained during experiments, ensuring that contact pressure remained constant, and no fluctuations were observed in real-time muscle-contraction signals. The deployment algorithm was further compressed due to the microcontroller’s limited storage.

In addition, intersubject variations in forearm muscle activation patterns affected the overall real-time classification performance. Despite these difficulties, the built system exhibited stable and reasonable performance in practical applications. In future work, improvements in cross-subject generalization, an increase in the number of gestures, and integration with robotic manipulators, prosthetic devices, and other assistive devices for actual human-machine interaction can be implemented.