Design and intelligent control of a low-cost, sEMG-driven, 3D-printed robotic hand prosthesis: A three-tier benchmarking framework with novel robustness metrics

Clinical and societal context

The World Health Organization (WHO) estimates that approximately 40 million people worldwide require prosthetic or orthotic services, yet only 5–15% of those in need have access to appropriate devices, particularly in developing nations [1]. Separately, an estimated three million people worldwide require a prosthetic arm [2]. Upper-limb loss, whether traumatic or congenital, profoundly impacts quality of life, functional independence, and psychosocial well-being [3]. Myoelectric prostheses—devices controlled by electrical signals generated during voluntary muscle contraction—represent the most physiologically intuitive control paradigm for artificial hands [4]. However, state-of-the-art commercial myoelectric systems, such as the Otto Bock Michelangelo Hand (approximately $75 000) and the DEKA LUKE Arm (approximately $100 000), remain financially inaccessible to the vast majority of potential users globally [5, 6].

This economic barrier has catalyzed a growing research interest in low-cost, open-source prosthetic systems that leverage affordable sensors, consumer-grade microcontrollers, and additive manufacturing (3D printing) technologies [7,8]. Surface electromyography (sEMG), which non-invasively measures the superposition of motor unit action potentials at the skin surface, serves as the primary biosignal modality for such systems [9]. The MyoWare Muscle Sensor (Advancer Technologies, USA), priced at approximately $38, has emerged as a popular low-cost sEMG acquisition platform in the maker and research communities due to its integrated instrumentation amplifier, envelope detection circuitry, and Arduino-compatible analog output [10].

Research gap and motivation

Despite the proliferation of studies using low-cost sEMG sensors and machine learning for prosthetic control, several critical gaps persist in the literature.

Gap 1: Over-reliance on synthetic or benchmark datasets. The majority of published studies report classification accuracies exceeding 95 % on carefully curated, noise-free, or synthetically generated datasets [11, 12]. However, these results rarely translate to real-world deployment scenarios where electrode impedance variations, skin moisture changes, muscle fatigue, electrode displacement, and cross-talk introduce substantial signal degradation [13, 14]. Very few studies provide benchmarking on live-acquired, real-world sEMG signals alongside synthetic baselines, making it difficult to assess the practical viability of proposed systems.

Gap 2: Limited multi-classifier comparison on single-channel systems. While multi-channel sEMG systems (8–16 channels) using devices such as the Myo Armband or high-density electrode arrays have been extensively studied [15, 16], single-channel systems using sensors like the MyoWare have received comparatively less rigorous algorithmic attention. Single-channel systems are inherently more challenging due to reduced spatial information, yet they represent the most accessible and affordable option for resource-constrained settings. Comprehensive head-to-head comparisons of classical machine learning versus deep learning classifiers on single-channel, real-world sEMG data remain scarce.

Gap 3: Absence of integrated cost–performance analysis. Most studies focus exclusively on classification accuracy without contextualizing performance within the total system cost. For a prosthetic system to be genuinely accessible, both the algorithmic performance and the economic feasibility must be jointly evaluated and reported [17].

Gap 4: Lack of standardised robustness metrics. The sEMG literature lacks standardised metrics to quantify how much classification performance degrades when transitioning from controlled laboratory conditions to real-world deployment. Without such metrics, meaningful cross-study comparisons remain impossible, and the practical readiness of proposed systems cannot be objectively assessed.

Gap 5: Lack of reproducibility and open-architecture reporting. Many studies omit critical hardware details (baud rates, sampling protocols, electrode placement specifics) that would enable independent replication, hindering the cumulative progress of the field [18].

Novel contributions of this study

This paper addresses the aforementioned gaps through the following original contributions.

• C1 Three-tier dataset benchmarking framework. We introduce a systematic evaluation methodology comparing five classifiers across three progressively challenging data conditions—synthetic, online benchmark, and real-world live—using sampling-rate-appropriate preprocessing and common evaluation metrics. This framework explicitly quantifies the “reality gap” in sEMG classification, providing the research community with realistic performance expectations for low-cost systems.

• C2 Novel robustness and cost–performance metrics. We propose two new evaluation metrics: the Robustness Degradation Index (RDI), which quantifies the fraction of above-chance accuracy lost when transitioning from synthetic to real-world conditions; and the Cost–Performance Index (CPI), which jointly evaluates classification agreement and economic cost. These indices complement conventional classification metrics.

• C3 Demonstration of ensemble learning performance on real-world single-channel sEMG. RF significantly outperformed SVM, KNN and LSTM and achieved a higher mean accuracy than the 1D CNN. The RF–1D CNN rank difference of 1.2 did not exceed the Nemenyi critical difference of 1.929 at α = 0.05. These findings are restricted to the five classifiers, datasets and ten able-bodied participants evaluated here.

• C4 Complete cost–performance characterisation of an ultra-low-cost system. We present a fully functional, 3D-printed myoelectric hand prosthesis system with component costs of approximately USD 122–232, a transparent component-level cost breakdown, and end-to-end performance evaluation from signal acquisition to mechanical actuation.

• C5 Reproducible, open-architecture system design. All hardware specifications, software parameters, electrode placement protocols, and data collection procedures are documented in sufficient detail to enable independent replication.

Paper organization

The remainder of this paper is organized as follows. Section Related work reviews related work. Section Materials and methods describes the materials and methods, including hardware design, data acquisition protocol, signal processing, feature extraction, classifier architectures, and the proposed novel metrics. Section Results presents experimental results. Section Discussion provides discussion and analysis. Section Conclusion and future work concludes with limitations and future directions.

Low-cost myoelectric prostheses

The convergence of affordable sEMG sensors, open-source microcontrollers, and desktop 3D printing has enabled a new generation of low-cost prosthetic research. Gusman et al. [19] developed an EMG-controlled prosthetic hand using the MyoWare sensor and Arduino Uno, achieving two-class binary open/close control using simple threshold processing rather than machine learning classification.

The e-NABLE community and projects such as the In-Moov robotic hand [20] have demonstrated the feasibility of 3D-printed prosthetic hands at costs below $500, but these systems typically employ simple on/off EMG control without multiclass pattern recognition capabilities. Tavakoli et al. [21] developed a low-cost myoelectric prosthesis using two EMG channels and an SVM classifier, achieving 85 % accuracy on controlled laboratory data; however, real-world performance was not evaluated and the system cost was not explicitly reported.

sEMG classification approaches

The sEMG classification literature can be broadly categorised into classical machine learning and deep learning approaches. Hudgins et al. [22] established the foundational time-domain feature set—mean absolute value, waveform length, zero crossings, and slope sign changes—for myoelectric pattern recognition. Subsequent work demonstrated the effectiveness of SVM [23], LDA [24], KNN [25], and Random Forest [26] classifiers on these features, with Phinyomark et al. [27] identifying root mean square, mean absolute value, and waveform length as the most robust features across varying conditions.

On the deep learning side, Atzori et al. [28] applied convolutional neural networks (CNNs) to the Ninapro database, demonstrating that discriminative features could be learned directly from raw sEMG signals. Côté-Allard et al. [15] achieved state-of-the-art results using transfer learning with CNNs for inter-subject classification, while Simão et al. [29] applied LSTM networks to exploit temporal dependencies in muscle activation patterns. Importantly, these studies predominantly used multi-channel, research-grade sEMG systems costing $5 000–$50 000, making direct comparison with single-channel, low-cost configurations problematic.

The synthetic-to-real performance gap

A critical but under-addressed issue in sEMG research is the dramatic performance degradation when classifiers trained on clean or synthetic data are deployed on real-world signals. Scheme and Englehart [13] documented accuracy drops of 20–40 percentage points between laboratory and practical conditions, with Campbell et al. [14] identifying electrode shift, muscle fatigue, limb position variation, and inter-session variability as primary contributors. Despite these well-known challenges, few studies explicitly quantify this gap by evaluating identical classifiers across synthetic, benchmark, and live data conditions within a single study, as we do here.

Positioning of the present study

Table 1 summarises the positioning of the present study relative to key prior works.

Taken together, the reviewed studies commonly rely on multi-channel or research-grade acquisition, report benchmark or subject-dependent results, and provide limited comparisons of classical and deep learning methods under shared hardware constraints. Reproducible release of hardware, firmware and analysis code is also inconsistent. These limitations motivate the present single-channel feasibility study across synthetic, public benchmark and live multi-subject recordings, with subject-independent evaluation of the recorded datasets.

System architecture overview

The proposed system comprises four functional subsystems: (i) sEMG signal acquisition, (ii) signal preprocessing and feature extraction, (iii) gesture classification, and (iv) mechanical actuation of the 3D-printed hand. The overall system architecture is depicted in Fig. 1.

Figure 1:

Hardware components and cost analysis

Power distribution and load management

The Nano and MyoWare are supplied from a 5 V USB source. The Nano 5V pin supplies MyoWare +V_s, and RAW connects to A0. The five servo signal inputs connect to D3–D7. A separate 2S lithium-ion pack (7.4 V nominal, 2000 mAh) feeds an LM2596 buck module set to 5.0 V, with a nominal 3 A rating, supplying the servo positive terminals directly (Fig. 2). VIN is unused, and servo current does not pass through the Nano onboard regulator. The positive rails remain separate, with return paths meeting at a single star-ground point. A 1000 μF low-ESR electrolytic and a parallel 100 nF ceramic capacitor are placed across the servo rail at the load. This routing reduces shared-return interference between the actuation and sensing circuits.

Figure 2:

A firmware load-management strategy is proposed to limit sustained servo loading. On grasp detection, the servos move progressively to calibrated closure positions. After 500 ms, the controller enters a holding state and slightly relaxes the commanded positions by a calibrated amount to reduce tendon loading while retaining the object. Endpoint limits prevent overtravel. Control pulses are disabled only where passive retention has been verified, and are restored before any subsequent movement. Because the SG90 pulse width commands position rather than motor current, this strategy limits loading indirectly through position control.

Data acquisition protocol

Electrode placement

The forearm was cleaned with 70% isopropyl alcohol to reduce skin impedance. Three Ag/AgCl disposable gel electrodes were applied as follows (see Fig. 3). The MID (positive) electrode was placed on the belly of the flexor digitorum superficialis (FDS) muscle, approximately one-third of the forearm length distal to the medial epicondyle. The END (negative) electrode was placed 2 cm distal to the MID electrode along the longitudinal axis of the FDS. The REF (reference/ground) electrode was placed on the olecranon process of the ulna—a bony elbow prominence with minimal muscular activity—to provide a stable electrical reference.

Figure 3:

Isolated index extension (G2) is produced mainly by the posterior-forearm extensor indicis, anatomically distant from the anterior FDS site. This channel captures its activity largely through crosstalk, so lower discriminability of G2 is expected with this placement (Section Gesture-specific analysis).

Dataset descriptions

Three distinct datasets were used for evaluation, creating a progressively challenging three-tier benchmarking framework.

Signal preprocessing

Preprocessing used Python 3.9, NumPy 1.21, SciPy 1.7 and scikit-learn 1.0. Live RAW data were acquired at Nano A0 at 1 kHz; synthetic windows use the same sample rate. Each recording was bandpass filtered at 20–450 Hz using a fourth-order Butterworth band-pass design, implemented as second-order sections and applied zero-phase (forward–backward), giving an effective eighth-order magnitude response, followed by a second-order 50 Hz IIR notch (Q = 30), non-overlapping 250 ms segmentation (250 samples), and zero-mean window centering. Anti-aliasing relied on the band-limiting characteristics of the MyoWare analog front end, without an additional dedicated analog filter.

DB5 remained at its native 200 Hz rate without interpolation. The same zero-phase fourth-order Butterworth implementation was applied with a 20–90 Hz passband, below the 100 Hz Nyquist limit, followed by the same 50 Hz notch. Non-overlapping 250 ms windows contain 50 samples. Frequency features were limited to the corresponding processing band and are not directly comparable with full-band live features.

Feature extraction and standardization

The classifier input comprises MAV, RMS, WL, ZC, SSC, MNF, MDF and PKF, computed from signed zero-mean windows. Let U denote the signal unit (ADC counts for uncalibrated live data). The amplitude features are

These discrete measures have units U. Crossing features are dimensionless counts:

The amplitude threshold is ɛ = 3σ_baseline (in ADC counts for the live data), where σ_baseline is estimated from filtered resting-baseline samples of the training fold and applied unchanged to held-out data. Within each fold, σ_baseline was computed per training subject from the final 1.5 s of each filtered rest period, and the median across training subjects was used. Across LOSO training folds, σ_baseline was 2.1 ± 0.3 ADC counts for the live data (ɛ ≈ 6.3 counts) and 0.9 ± 0.2 a.u. for DB5.

Integrated EMG (IEMG), simple square integral (SSI) and variance (VAR) were not used, because for zero-mean windows they are deterministic functions of MAV and RMS and add no information:

A one-sided Welch PSD is estimated from demeaned segments with a Hann taper. For 250-sample windows, the segment length is 128 samples with 50% overlap; for the 50-sample DB5 windows, 32-sample segments with 50% overlap (16 samples) and a 64-point FFT were used (Δf = 3.125 Hz). Let P_j be the PSD at f_j within the dataset's processing band J, with uniform bin spacing Δf. Then

The feature vector is f ∈ ℝ⁸. For feature k in window w, z_w,k = (f_w,k − μ_k,train)/s_k,train, using only training-fold statistics and applying the fitted transform unchanged to held-out data. All inner-fold fitting is likewise restricted to the training partition. Deep models use the signed preprocessed waveform windows (250 live samples; 50 DB5 samples), with any numerical scaling fitted on training data only.

Classifier architectures

Five classifiers spanning classical machine learning and deep learning were compared. SVM, KNN and RF provide established handcrafted-feature baselines; the 1D CNN and LSTM learn temporal waveform structure. RF was chosen as the representative ensemble method, and Transformer architectures were not considered given their data requirements relative to the dataset size.

Evaluation methodology

Proposed novel performance metrics

To enable standardised comparison of low-cost sEMG prosthetic systems, we introduce two novel evaluation metrics.

Signal characteristics

Figure 4 shows representative raw and preprocessed sEMG signals for the three gestures from Subject S1. Distinct amplitude and temporal pattern differences are visually observable across all three classes. G1 (Open Fist) exhibited sustained, moderate-amplitude activation with a relatively stable envelope, whereas G2 (Single Finger) showed lower amplitude and more localised activation with higher variance. G3 (Repetitive Open/Close) produced a distinctive oscillatory pattern characterised by alternating high-amplitude bursts and low-amplitude rest periods, making it visually the most separable of the three gestures.

Figure 4:

Feature distributions

Figure 5 presents box plots of RMS, MAV and WL, the three most informative features by permutation importance (Section Feature importance analysis), across the three gesture classes for the real-world dataset. While inter-class separation is evident, substantial overlap exists—particularly between G1 and G2—confirming the challenging nature of single-channel classification.

Figure 5:

At the cohort level, subject-level RMS (mean ± SD across N = 10 subjects, uncalibrated ADC counts) was 41.6 ± 12.3 for G1, 27.9 ± 9.8 for G2 and 48.2 ± 14.1 for G3. A Friedman test showed a significant gesture effect (χ²(2) = 15.2, p < 0.001). Holm-corrected Wilcoxon signed-rank tests showed lower RMS for G2 than for G1 (p = 0.006) and G3 (p = 0.006), whereas G1 and G3 did not differ significantly (p = 0.19). Thus, although window-level distributions overlap, the reduced activation during index extension is consistent across subjects.

Classification results

Accuracy degradation across dataset conditions

Table 8 summarises accuracy across all three dataset conditions. Figure 6 presents the corresponding grouped bar chart.

Figure 6:

Key finding: The Random Forest exhibited the smallest accuracy degradation (27.5 pp) when transitioning from synthetic to real-world data, while the LSTM suffered the largest degradation (40.5 pp), indicating greater robustness of RF to the distribution shift between controlled and real-world conditions in this study.

Novel metric evaluation

Statistical significance analysis

Per-subject variability on real-world data

Table 11 presents per-subject Random Forest accuracy on the real-world live data.

Substantial inter-subject variability (range: 55.1–78.3%) was observed, consistent with the known challenge of inter-subject generalisation in sEMG classification [13]. Subjects S5 and S9 showed notably lower accuracies, possibly attributable to anatomical variations such as smaller forearm circumference or a thicker subcutaneous fat layer, or to inconsistent gesture execution during data collection.

Feature importance analysis

Held-out permutation importance was computed for the RF classifier as the mean accuracy decrease over the 10 LOSO folds, with 20 permutations per feature (Table 12). WL ranked highest, followed by RMS and MAV. Because MAV and RMS are strongly correlated, their individual values understate their joint contribution: permuting them together reduced accuracy by 11.2 pp. Time-domain features (WL, RMS, MAV, SSC, ZC) contributed more than the spectral descriptors, with PKF contributing least.

Computational performance

All classifiers execute on the host PC, while the Nano handles ADC sampling, serial communication and servo-control pulses (Table 13). The RF inference time of 1.2 ms per window is well within the control budget, but its 850 KB model exceeds the ATmega328P's 32 KB flash and 2 KB SRAM. Standalone deployment therefore requires a 32-bit MCU together with model compression.

Prosthetic hand actuation results

Functional demonstrations used the RF classifier on the tethered PC, with commands sent to the Nano, and included a ball-grasp trial and return to the open posture (Fig. 9). The average end-to-end latency was 285 ms, timed from completion of an analysis window to the servo reaching 90% of its commanded travel. Its components were serial transfer (43.6 ms), preprocessing and feature extraction (3.1 ms), RF inference (1.2 ms), command return and the next PWM frame (22.0 ms), and servo travel (215 ms), totalling 284.9 ms. The measured transfer time agrees with the calculated wire time of a 250-sample binary packet (two bytes per sample, 8N1 framing) at 115 200 baud, 500 × 10/115200 = 43.4 ms excluding headers. The 250 ms window-buffering interval precedes this timing reference and is reported separately.

Figure 7:

Figure 8:

Figure 9:

Forces were measured with a 5 kg bar load cell and HX711 amplifier sampled at 80 Hz, calibrated with reference masses of 50–1000 g, with five repetitions per condition. Slipping was defined as visible object displacement exceeding 5 mm during a 10 s hold. These benchtop measurements gave a peak individual-fingertip normal force of 1.72 N and a total multi-finger grasp force of approximately 7.24 N (7.22–7.26 N across repetitions). The maximum reliably held payload before slipping was approximately 200 g, corresponding to a 1.96 N gravitational load. These values support light-object grasping and remain below the 15–20 N range often cited for daily grasping tasks.

Why Random Forest performs better under the evaluated conditions

RF achieved higher live accuracy than the CNN and LSTM. It was significantly better than LSTM, while its 8.0 pp mean advantage over the CNN was not statistically significant (rank difference 1.2 < 1.929). Deep networks have outperformed feature-based methods on larger multi-channel datasets [15, 28]; the present findings apply to the small-sample, single-channel regime studied here.

Each LOSO training fold contained only 81 live recordings, which limits the data available for learning waveform representations, whereas bootstrap aggregation and engineered features are comparatively data-efficient. The live-recording active-to-rest power ratio, 10 log₁₀(P_active/P_rest) with powers defined as mean squared filtered amplitudes during contraction and the pre-contraction rest period (excluding its initial settling transient), was computed per trial, averaged over each subject's nine trials, and had a cohort median of 10.1 dB (IQR 8.4–11.6 dB). At this modest signal quality, noise-robust summary features such as WL, RMS and MAV, which permutation analysis identified as the most informative, are likely to generalize across subjects better than learned waveform filters. Without the spatial information of multi-channel systems, the temporal modelling of the CNN and LSTM offered no advantage over these features.

The reality gap: quantitative analysis

Synthetic-to-live accuracy drops ranged from 27.5 to 40.5 pp, with RDI values of 43.7–67.4%. The synthetic data are an idealized reference: interpolation and stationary Gaussian noise do not reproduce fatigue, motion artifacts, effort variation or inter-subject anatomy. High synthetic accuracies (above 92%) therefore substantially overestimate live performance, underlining the need for subject-independent validation on recorded data.

The three tiers also differ in acquisition hardware, native bandwidth and approximate gesture mapping. DB5 uses 200 Hz Myo armbands, whereas live data use the 1 kHz MyoWare RAW output, so the tier-to-tier differences reflect several factors rather than noise or sensor quality alone. The RDI is most informative for comparing classifiers evaluated under a common protocol, as done here.

Performance contextualisation

The 68.7 % accuracy achieved by Random Forest on real-world, single-channel data for three gesture classes must be contextualised appropriately. Against the three-class chance level of 33.3 %, the achieved accuracy is 2.06× chance level, with a Cohen's κ of 0.531, indicating “moderate” agreement [34]. Direct comparison with the literature is challenging due to differing experimental conditions; however, Tavakoli et al. [21] achieved 85 % with two channels and controlled conditions, and Phinyomark et al. [37] reported approximately 65–75% accuracy for 3-class problems with single-channel sEMG under realistic conditions—consistent with our result. From a functional perspective, for a three-state prosthetic controller (open, close, pinch), 68.7 % accuracy with appropriate rejection thresholds and majority voting over consecutive windows can yield usable control in practice [38].

These results are a proof of concept from ten able-bodied volunteers, whose activation patterns may differ from residual-limb musculature. A larger cohort and trans-radial amputee testing are planned (Section Future work).

Gesture-specific analysis

The RF confusion matrix (Table 6) reveals that gesture discriminability is asymmetric across the three classes. G3 (Open/Close) was most accurately classified (73.5 % by RF) due to its distinctive oscillatory pattern, which generates time-varying features (WL, ZC, SSC) that differ markedly from sustained contractions. G2 (Single Finger) was most poorly classified (59.4 % by RF), reflecting the inherent challenge of detecting isolated single-finger movements on a single-channel sensor placed over the FDS—a muscle primarily responsible for collective finger flexion. The isolated index extension activates the extensor indicis on the posterior forearm, which is anatomically distant from the sensor placement, resulting in attenuated and ambiguous signals on the single FDS channel. The asymmetric confusion pattern—G2 is confused with both G1 and G3, while G1 and G3 are less often confused with each other—suggests that a hierarchical classification approach, first distinguishing dynamic (G3) from static gestures (G1, G2) and then discriminating G1 from G2, could improve overall accuracy in future work.

Practical implications for low-cost prosthetics

The prototype demonstrates a low-cost research architecture with separate sensing, host-PC classification and tendon actuation. Its modular design allows each subsystem to be upgraded independently. A second sensor on the extensor compartment could improve discrimination of index extension and is proposed for future evaluation.

SG90 plastic gears may be damaged by sustained stall or unexpected tendon resistance. Metal-gear alternatives such as the MG90S are a candidate durability upgrade, subject to torque, mounting, current and endurance testing. The measured grip force and 200 g payload restrict the current prototype to light objects; higher-torque actuators, lower-friction tendon routing and cycling-durability testing are planned.

Comparison with state of the art

Tables 14 and 15 compare hardware and evaluation scope. The present study combines a low component cost with live-recording LOSO evaluation, but accuracies cannot be ranked directly across different gestures, cohorts, modalities, acquisition conditions and validation methods. NR denotes information not reported in the cited work.

Limitations

First, ten subjects provide an initial feasibility cohort but limit statistical power and generalization. Second, all participants were able-bodied; amputee residual musculature, activation amplitudes and motor control may differ [39]. Third, only three gestures were evaluated using one FDS channel. Fourth, each subject was recorded in one session without repositioning electrodes, so electrode-shift and inter-session robustness were not evaluated [14]. Fifth, no dedicated analog anti-aliasing filter was used, so some aliased content cannot be excluded, and DB5's 200 Hz acquisition limits its spectral band and gesture equivalence. Sixth, Gaussian synthetic data cannot capture the full variation of recorded sEMG. Seventh, grip force and payload are limited to light objects, and long-term cycling durability and the proposed firmware load-management strategy have not yet been validated experimentally. Finally, all classifiers run on a tethered PC; standalone operation requires a suitable 32-bit MCU, model compression and on-device latency verification. Clinical evaluation will be required before the system can be considered for patient use.

Conclusions

This study evaluated a low-cost, single-channel MyoWare/Arduino hand prototype with host-PC classification and tendon-driven PLA fingers in ten able-bodied volunteers. Among the five tested classifiers, RF achieved the highest live LOSO accuracy (68.7%), significantly exceeding SVM, KNN and LSTM; its 8.0 pp mean advantage over the CNN was not statistically significant under the Nemenyi test. RF is therefore a suitable option under the tested conditions, though not necessarily optimal for all sEMG applications.

Synthetic-to-live degradation ranged from 27.5 to 40.5 pp (RDI 43.7–67.4%), with RF the most robust classifier. Held-out permutation importance identified WL, RMS and MAV as the most informative features. The CPI of the prototype is 13.77 using a dimensionless cost ratio and a USD 232 system cost. Above-chance three-class discrimination and light-object actuation were demonstrated, while amputee performance, long-term endurance and standalone operation remain to be established.

Future work

Building upon the current findings, the following directions are proposed.

1. Dual-channel acquisition: Adding a second MyoWare sensor on the extensor compartment to improve G2 classification through antagonist muscle pair analysis.

2. Transfer learning and domain adaptation: Applying domain adaptation techniques (e.g., maximum mean discrepancy minimisation) to bridge the synthetic-to-real gap.

3. Adaptive learning: Implementing online incremental learning in the Random Forest to adapt to individual users' EMG patterns over time.

4. Expanded gesture vocabulary: Increasing to 5–8 gesture classes while maintaining acceptable accuracy.

5. Embedded deployment: Porting the RF model to a 32-bit MCU (e.g. ESP32 or ARM Cortex-M4) using tree pruning and TinyML quantization for fully standalone operation.

6. Clinical validation: Evaluating the system with a larger cohort including trans-radial amputee users in collaboration with rehabilitation clinicians.

7. Sensor upgrade path: Evaluating the MyoWare 2.0 sensor and other emerging low-cost sEMG platforms (e.g., OpenBCI Cyton) to quantify the performance improvement achievable with incremental hardware investment.

8. Multi-session and longitudinal studies: Assessing inter-session robustness and the long-term learning curve of users adapting to the myoelectric controller over weeks of daily use.

9. Rejection thresholds and majority voting: Implementing confidence-based gesture rejection and sliding-window majority voting to improve functional reliability in continuous real-time use.

10. User experience evaluation: Conducting standardised functional assessments (e.g., Southampton Hand Assessment Procedure, Box and Blocks Test) with both healthy subjects and amputee users to complement algorithmic metrics with clinically validated outcome measures.