Search for: All records

The emergence of batteryless sensors offers a sustainable path towards ubiquitous sensing. However, unlike battery-powered devices, batteryless sensors only operate once they have harvested sufficient energy from the environment. This disrupts data availability and timely predictions for time series tasks like human activity recognition (HAR). Existing machine learning (ML) pipelines are ill-equipped for this uncertainty in data acquisition, as they typically assume synchronized and continuous data streams. To address this, we propose BHAR, a system-level optimization framework for Batteryless Human Activity Recognition that learns both the sensor-level data acquisition policy and the downstream activity classification model. Our key insight is that the classification error in batteryless HAR can be decomposed into two components: a passive error governed by the data acquisition policy and an active error dependent on the classifier’s performance. Building on this insight, our BHAR framework explicitly minimizes both errors to handle the sparse and asynchronous nature of batteryless sensor data. We validate BHAR on multiple datasets and architectures and show that it closes the performance gap in F1-score between batteryless and battery-powered HAR by 15–50% compared to baseline approaches.