PhD Proposal: Physical Intelligence from Constrained Observations: Rethinking the Wireless-to-Information Pipeline

Wireless signals are among the most pervasive yet under-utilized sensing resources in the modern world. From cellular base stations and Wi-Fi access points to radar platforms and pervasive connected devices, the electromagnetic environment is continuously filled with signals that not only enable communication, but also encode valuable information about location, motion, occupancy, spectrum usage, and physical interactions with the environment. These ambient transmissions form an invisible layer of infrastructure around everyday life and create new opportunities for sensing and artificial intelligence at the edge — from city-scale localization and spectrum awareness to security monitoring and environmental perception. Yet realizing this potential has remained elusive, because conventional wireless receivers remain too power-hungry and complex for the small, low-cost, energy-constrained devices that could benefit from them most. This dissertation explores this gap, advancing the thesis that the traditional wireless-to-information pipeline is often over-provisioned for wireless systems: instead of relying on exact signal reconstruction, useful information can often be extracted directly from constrained observations through a joint understanding of propagation, reception, and inference.
The contributions of this dissertation span three key areas: enabling pervasive low-power localization, advancing precise positioning through learned inference on compressed signals, and rethinking wideband spectrum monitoring for security-critical environments. First, it introduces LiTEfoot, an ultra-low-power localization system that leverages ambient cellular signals and passive non-linear spectrum folding to enable wide-area positioning, allowing a coin-cell-sized device to simultaneously sense wide cellular spectrum and estimate its own location for years of unattended operation. Second, it presents DeepSync, a deep learning framework for precise localization from compressed cellular spectrum, demonstrating that sub-sample timing information can be recovered even from severely degraded observations, enabling localization accuracy within a few meters under conditions where traditional estimation techniques fail. Third, it develops SpecSentry, a micro-power architecture for wideband spectrum awareness that couples compressed RF observations with learning-based inference, showing how resource-constrained devices can intelligently monitor large swaths of spectrum and detect ephemeral transmissions in security-critical settings. Together, these systems show that effective wireless sensing need not depend on reconstructing the full signal, but rather on preserving and exploiting the right structure for the task.
Across these contributions, the central idea is that wireless sensing should be designed around information requirements rather than receiver conventions. Rather than treating incomplete or transformed measurements as liabilities, this dissertation shows how constrained observations can serve as a first-class substrate for task-driven sensing when paired with the right physical abstractions and inference mechanisms. The result is a new perspective on wireless sensing that favours deployability, persistence, and efficiency without giving up task-relevant precision.