Orion: Two-Channel Surface EMG Gesture Recognition for AR/VR Input on Commodity Microcontroller Hardware

Hand-gesture input for augmented and virtual reality is dominated by optical tracking, which fails under occlusion and in low light, and by handheld controllers, which occupy the hands. Surface electromyography (sEMG) senses muscle activation directly and is immune to both problems, but published sEMG gesture systems typically rely on dense electrode arrays or research-grade amplifiers that are impractical for consumer hardware. We present Orion, a complete sEMG gesture recognition pipeline built entirely from commodity parts: two MyoWare analog sEMG sensors and an ESP32 microcontroller, sampled at 50 Hz. We describe the acquisition firmware, a 35-dimensional feature extractor combining raw, rectified and envelope channels with rolling statistical, differential, spectral and higher-moment features, and a three-way soft-voting ensemble over a voting committee, a stacked generalizer, and a standalone gradient-boosted tree model. On a 10,935-sample single-subject corpus spanning four gestures (relaxed, fist, index extension, index+middle extension), the system attains 71.3% classification accuracy and 0.709 macro-F1 against a 25% chance baseline, and runs inference in real time over a serial link. We report a negative result that we believe is more useful than the headline accuracy: our second sensor channel, placed over the extensor digitorum, contributed essentially no discriminative signal (mean envelope amplitude 0.4 of 4095 across three of four gestures), and the two finger-extension gestures proved nearly inseparable on the remaining channel (mean envelope 821.8 vs 827.4), which accounts for almost all of the residual error. We analyse why two-channel low-rate acquisition on this class of hardware is sufficient for coarse gesture states but not for individuated finger gestures, and we state the methodological limits of the evaluation, including the temporal-leakage risk in our sample-level train/test partition.