Batteryless Gesture Recognition Via Learned Sampling

Geffen Cooper, Radu Marculescu

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: wearables, batteryless sensors, deep learning, gesture recognition, energy harvesting

Abstract: Batteryless wearables offer the potential for continuous and maintenance-free operation through ambient energy harvesting. However, the constrained and uncertain nature of harvested energy makes it impossible to \textit{continuously} sense, process, and transmit data. In this batteryless setting, we no longer have access to complete and continuous data streams. Thus, we must adapt machine learning models to this new paradigm where only a limited \textit{subset} of the data can be sampled and processed. In this work, we consider solar-powered gesture recognition as a target application where sensing and processing are constrained to a strict energy budget. We propose a learning-based approach where a sampling policy and gesture classifier are jointly trained via a shared representation. This enables the system to learn which sensor samples are informative while simultaneously optimizing the classifier for sparse inputs. By actively deciding when to sample, we improve gesture recognition accuracy by up to 10\% compared to fixed-rate subsampling across a wide range of energy budgets. This closes the gap to a standard battery-powered approach by up to 40\%.

Track: 3. Signal processing, machine learning, deep learning, and decision-support algorithms for digital and computational health

NominateReviewer: Geffen Cooper, [e-mail]

Submission Number: 76