Zilo — Smart Ring Gesture Recognition & Musical Interaction

AdventureX 2026 · Zilo Ring Track•2026• Team Project
IMU
Gesture Recognition
BLE
DTW / HMM
Real-time Interaction

Wearable Sensing · Temporal Recognition · Human–Computer Interaction

Project Overview

Zilo is a team-built real-time gesture interaction system using a smart ring with a six-axis inertial measurement unit. The ring streams acceleration and angular-velocity data over Bluetooth Low Energy; the system detects gesture boundaries, processes the time series, recognises short gestures and maps the result to musical performance controls. The project received Third Place in the Zilo Ring Track.

Sense

A smart ring samples three-axis acceleration and three-axis angular velocity at approximately 25 Hz.

Stream

BLE packet handling, reassembly and CRC checks turn device output into a continuous usable signal stream.

Recognise

Motion-energy segmentation isolates gestures before temporal recognition with DTW and Left–Right Gaussian HMM baselines.

Respond

Recognised gestures are mapped to configurable musical techniques and immediate audio feedback.

My Contribution

  • Contributed to the design and integration of an end-to-end wearable interaction prototype built around a six-axis smart-ring IMU.
  • Helped connect real-time BLE data acquisition, motion segmentation, preprocessing, temporal recognition and gesture-to-action mapping into one usable system.
  • Supported system testing across short wrist and finger gestures, including confidence handling, gesture expiry and one-shot triggering.
  • Worked with the team to translate recognised gestures into musical performance controls and real-time audio feedback.

The recognition system was a team outcome. This page describes the integrated project and deliberately does not claim every DTW or HMM implementation as individual work.

From Prototype to Research

The prototype provides a practical baseline for a more rigorous study of personalised and robust wearable gesture recognition.

  • How much does recognition degrade across users, sessions and changes in ring orientation?
  • Can one, three or five personal demonstrations provide effective few-shot calibration?
  • How should the system reject unknown motion instead of forcing every input into a known class?
  • How do DTW, HMM and compact learned models compare under realistic wearable variation?