WiFi Sensing via Reservoir Computing

Abstract

Practical WiFi sensing must handle clock-asynchronous links, cross-domain variation, and post-deployment updating under the limited compute budget of access point (AP), router, and embedded Internet-of-Things platforms such as ESP-class devices. Reservoir computing (RC) is attractive in this setting because its temporal encoder can remain fixed while only a lightweight readout needs to be optimized and updated. To address these deployment challenges under tight compute budgets, we present ReWiS, a WiFi-sensing-oriented reservoir framework that transforms channel state information (CSI) into structured micro-Doppler streams with common, antenna-specific, and differential motion cues, encodes them with a graph-coupled reservoir, and adapts to a new domain after deployment by freezing the reservoir and fine-tuning only a compact readout with a few labeled target samples. On a large-scale WiFi sensing benchmark, ReWiS achieves 89.2% in-domain macro-F1 and 82.0% mean cross-domain macro-F1 with only a 0.72M trainable readout, improves to 88.5\% after lightweight post-deployment adaptation, and remains competitive with recent deep baselines evaluated under the same protocol, which achieve 87.5%-89.2% mean cross-domain macro-F1, while requiring lower optimization cost and lower CPU latency. These results indicate that ReWiS provides a practical reservoir-based design for deployable WiFi sensing, with further potential for low-power hardware realization.