Pebbl system turns wearable data labeling into hands-free trigger-action routines

Pebbl, developed by Zeyu Wang, Yuntao Wang, and colleagues at Tsinghua University and University of Toronto, tackles a core problem in wearable Human Activity Recognition (HAR): collecting richly labeled activity data in real-world settings. Traditional approaches rely on retrospective self-reporting, which is labor-intensive, imprecise, and rarely scales. Pebbl instead uses opportunistic crowdsensing—letting users schedule trigger-action routines on their smartphone. When a trigger fires (the prototype uses four common audio cues like a microwave beep or door knock), the system sends a just-in-time reminder asking the user to confirm or describe the action in open-vocabulary natural language.

Each confirmed action produces a clean sensor window with explicit start/end boundaries and a user-authored label—no manual segmentation or post-hoc annotation needed. The researchers evaluated Pebbl through an expert workshop (N=6), a within-subject in-lab study (N=21), and an 8-person pilot deployment. Results showed high reliability under controlled conditions: 97.30% recall and 97.15% precision, with participants preferring Pebbl over comparison workflows on perceived burden and confidence. The pilot also validated the full interaction and sensing pipeline in free-living settings, though it exposed practical constraints like false triggers and context dependence. Pebbl, accepted to IMWUT 2026, points toward a low-burden, distributable future for user-contributed wearable activity data.