<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Research |</title><link>https://gspacone.github.io/projects/</link><atom:link href="https://gspacone.github.io/projects/index.xml" rel="self" type="application/rss+xml"/><description>Research</description><generator>HugoBlox Kit (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Sat, 31 Oct 2026 00:00:00 +0000</lastBuildDate><image><url>https://gspacone.github.io/media/icon.svg</url><title>Research</title><link>https://gspacone.github.io/projects/</link></image><item><title>Wearable hand &amp; arm tracking</title><link>https://gspacone.github.io/projects/hand-tracking/</link><pubDate>Thu, 01 Jan 2026 00:00:00 +0000</pubDate><guid>https://gspacone.github.io/projects/hand-tracking/</guid><description>&lt;p&gt;Continuous hand–wrist tracking is key to enabling natural human–machine interaction, yet challenges remain to ensure robustness across users and sensor repositioning. This research combines established modalities (sEMG, IMU) with A-mode ultrasound to capture complementary muscle information for robust, low-power kinematic tracking and gesture recognition, towards reliable real-time interaction across multiple days.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Wearable A-mode US for kinematic tracking&lt;/strong&gt; — designed a four-transducer US armband interfaced with the WULPUS acquisition system; integrated motion capture and an instrumented glove for ground truth; led the data collections; developed XGBoost and CNN regression models, INT8-quantized and deployed on nRF52. &lt;em&gt;3-DoF: 5 subjects over 9 sessions, inter-session RMSE 11.1°. 23-DoF: 11k-parameter CNN, 29.1 ms inference on the nRF52 MCU, &amp;lt; 35 mW for sensing, inference and BLE result transmission.&lt;/em&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Concurrent EMG + US fusion&lt;/strong&gt; — designed a flexible EMG PCB and wearable armband interfacing with the BioGAP acquisition system; implemented firmware-level synchronisation for concurrent EMG and US acquisition; designed and led data collections; developed a CNN-based EMG–US late-fusion architecture. &lt;em&gt;23-DoF: inter-session RMSE 10.6° vs. 12.0° (EMG only) and 13.1° (US only).&lt;/em&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;US + IMU for real-time VR&lt;/strong&gt; — led multi-day data collections, developed late-fusion CNNs with personalised recalibration, and built the real-time pipeline into a Unity VR environment. &lt;em&gt;5 subjects over 5 days: 80% / 77% inter-session accuracy; 88–96% online VR success after 5 min of recalibration.&lt;/em&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Now:&lt;/strong&gt; distributed EMG + US + IMU fusion for VR and robotic control.&lt;/p&gt;</description></item><item><title>Wearable silent-speech interfaces</title><link>https://gspacone.github.io/projects/silentwear/</link><pubDate>Sun, 01 Mar 2026 00:00:00 +0000</pubDate><guid>https://gspacone.github.io/projects/silentwear/</guid><description>&lt;p&gt;Silent-speech interfaces decode speech-related intent from physiological signals, enabling private and accessible communication, but existing approaches rely on obtrusive sensing and struggle to stay robust across sessions. This research explores unobtrusive, wearable interfaces with on-device processing.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Contributions&lt;/strong&gt; — conceptualised SilentWear, a fully wearable, dry-electrode textile sEMG neckband built around the BioGAP platform; conceived and ran multi-day, multi-subject studies with vocalised and silent speech; developed SpeechNet, a compact 15k-parameter CNN for edge deployment, with a personalised fine-tuning strategy for inter-session robustness.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Results&lt;/strong&gt; — 8-command recognition across 4 subjects: 84.8% / 77.5% accuracy (vocalised / silent, aggregated evaluation); fine-tuning recovers &amp;gt; 10 percentage points of accuracy lost to repositioning within &amp;lt; 10 min of personalisation; 2.47 ms inference at the edge (GAP9).&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Now:&lt;/strong&gt; EMG, EEG and audio fusion for continuous speech recognition.&lt;/p&gt;</description></item><item><title>Foundation models for the edge</title><link>https://gspacone.github.io/projects/tinymyo/</link><pubDate>Sun, 01 Feb 2026 00:00:00 +0000</pubDate><guid>https://gspacone.github.io/projects/tinymyo/</guid><description>&lt;p&gt;Foundation models can learn transferable representations from physiological signals, reducing the need for task-specific models and labelled data. However, variability across acquisition systems, subjects and tasks challenges generalisation, and computational cost limits wearable deployment. This research investigates compact foundation models that bridge transferable representation learning and edge inference.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Contributions&lt;/strong&gt; — conceived TinyMyo, the first EMG foundation model at the edge, and defined its research direction: dataset unification, the pre-training framework and multi-task downstream evaluation; supervised its development and experimental validation within a Master&amp;rsquo;s thesis.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Results&lt;/strong&gt; — a 3.6M-parameter transformer backbone evaluated on gesture recognition (NinaPro DB5, UCI-EMG, EPN-612, Generic Neuromotor Interface by Meta), kinematic regression (NinaPro DB8) and silent speech (Gaddy). Deployed on GAP9, where a 1.9M-parameter variant achieves 89 ms inference on 1-s windows.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Now:&lt;/strong&gt; foundation models for A-mode ultrasound at the edge.&lt;/p&gt;</description></item><item><title>Wireless body-area network</title><link>https://gspacone.github.io/projects/body-area-network/</link><pubDate>Wed, 01 Apr 2026 00:00:00 +0000</pubDate><guid>https://gspacone.github.io/projects/body-area-network/</guid><description>&lt;p&gt;I am developing a distributed body-area network built from compact, multimodal sensing nodes with EMG, ultrasound and IMU sensing. A centralised computing &amp;ldquo;brain&amp;rdquo; coordinates the nodes, so the same wearable network can run several applications at once.&lt;/p&gt;</description></item></channel></rss>