6G networks promise AI-native RAN architectures that autonomously coordinate applications and infrastructure. Yet a critical gap remains: applications know their constraints (battery-critical mission, safety deadlines) but not network state (cell load, coverage); the RAN knows wireless conditions but not application needs. Static policies ignore context, autonomous ML lacks explainability, manual coordination does not scale. We introduce semantic negotiation: applications and RAN expose constraints via MCP servers, an LLM reasons about tradeoffs asynchronously, and human operators supervise critical decisions. The LLM queries live wireless state, proposes changes, and validates outcomes via query-feedback loops. Operators see why decisions were made: tool rationale, constraint validation, tradeoff reasoning, not just what changed. Our OpenAirInterface 5G proof-of-concept demonstrates explainable coordination for proactive planning in private enterprise networks where human oversight is essential and regulatory compliance requires full auditability.
Dysarthria detection from speech enables scalable neurological health monitoring but raises serious privacy concerns due to leakage of spoken content. We explore whether dysarthria can be detected and classified without intelligible speech using privacy-aware signal processing. Using MFCC features and lightweight classifiers, we show that dysarthria-relevant cues persist under strong privacy filtering, enabling a favorable privacy-utility trade-off.
This paper proposes MetaPower, a design of a magnetic resonance-based wireless power transfer system that integrates capacitive electrodes and metasurface, capable of locating devices and focusing energy towards them for smart charging. Our design builds on the fact that magnetic resonators generate strong localized magnetic fields with lensing effects. We leverage this principle to control the energy distribution by applying different equivalent permeability and phase profiles on our metasurface arrays on each cell. We show how our surface offers improved features and safety guarantees over conventional wireless charging methods, enabling programmability in the transferred power distribution based on the target device's location. Consider, for example, a metasurface-enabled table that automatically detects a tagged device and then actively focuses energy in the near-field based on the object location. We present a detailed design of MetaPower's target localization and near-field magnetic focusing through HFSS and MATLAB simulations. We also present preliminary results from a proof-of-concept implementation of the localization electrodes, and discuss the system-level surface implementation for future work.
Neural Architecture Search (NAS) automatically searches a neural architectural space for onek that achieves high accuracy on a target task, typically image analysis. While current NAS frameworks excel at maximizing accuracy, they have largely overlooked inference latency as a design constraint. In this paper, a novel NAS framework is developed that effectively exploits kernel tuning using the TVM auto-tuning ML compiler to synthesize Latency-Constrained DNNs (NASLCD). We demonstrate the superiority of NAS-LCD over state-of-the-art NAS approaches, including Once-For-All (OFA) and proxy-based ZEN-NAS.
Volumetric video streaming depends on the precise coordination of multi-view video streams, imposing stringent requirements on synchronization and throughput. However, developing robust streaming algorithms is hindered by the high cost of physical capture rigs and the difficulty of reproducing specific network anomalies. We present the Volumetric Stream Simulator (VSS), an end-to-end virtualized testbed that characterizes the impact of mobile network conditions on volumetric reconstruction.
Accurately estimating energy expenditure (EE) is crucial for understanding exercise efficiency, managing fitness goals, and monitoring health conditions. Existing wearable systems either rely primarily on heart rate and motion sensors, resulting in unsatisfactory accuracy, or require bulky setups such as thermal cameras to improve performance by integrating more physiological information, which limits real-world applicability. We present EarCalo, an earable-based system that leverages in-ear audio sensing to estimate EE during running. The system extracts airflow-induced acoustic variations within the ear canal and employs a deep neural network to translate these subtle in-ear sound dynamics into EE estimates. The key insight is that in-ear acoustic signals can capture multiple physiological factors such as breathing and cardiovascular activity, while also reflecting motion-related cues like running intensity. These rich acoustic cues are closely related to EE and provide a unified sensing modality for estimation. We evaluated EarCalo on 21 participants running at varying speeds using a mixed-user setting. EarCalo achieved a mean absolute error (MAE) of 0.67 kcal/min, a mean absolute percentage error (MAPE) of 11.98%, and a Pearson correlation of 0.945, which achieves accuracy close to established physiological standards. This work represents an early step toward practical and personalized earable-based EE estimation in everyday settings.
Mobile broadband has long been central to increasing access to Internet connectivity [5] and enabling smarter and more connected communities [6]. However, hyperlocal knowledge about where mobile broadband is available and how well it performs is a persistent challenge [7]. The radio propagation models that are used to create mobile broadband coverage maps often overstate coverage [1, 2]. Under the data generated by these models, there exist underserved communities unable to access or effectively advocate for adequate broadband services. Within this critical gap between official datasets and the everyday experience of communities, the United States' Federal Communications Commission (FCC) launched the Mobile Availability Challenge (MAC)-a novel process that allows citizens to dispute the reported availability of mobile broadband service using on-the-ground measurements. This process presented an unprecedented opportunity for civic engagement to inform Internet policy and drive broadband deployment efforts. However, the tools used to collect these measurements have not been designed to center users and collective civic action, which poses a critical threat to meaningful adoption and sustainable impact. As part of our ongoing work to address digital inequalities and empower civic action around mobile broadband, we sought to design an open-source, mobile broadband measurement app that centers users and communities. CellWatch has been carefully co-designed with 69 community stakeholders representing municipal broadband efforts, tribal networks, and nonprofit organizations with the goal of incorporating a more community-centric perspective on mobile broadband measurement and mapping. CellWatch was designed through iterative user workshops using the FCC Speed Test app [3] as an initial model for compliance with the FCC MAC process. Our team worked with the Measurement Lab (M-Lab) team to use the Measurement Swiss Army Knife to integrate multi-stream TCP upload and download speed tests and latency measurements using distributed M-Lab servers as measurement endpoints. In contrast to the FCC Speed Test app, we focused on providing transparency into elements of cognizable challenges. Data collected from the Mobile App is submitted to the FCC Challenge Servers via official API. The CellWatch app is the first third-party application to be officially approved by the FCC for use in the broadband data collection mobile challenge [4]. A copy of the measurements are also stored in the CellWatch backend, implemented as a Supabase data repository with a REST API. The web-based researcher dashboard provides a large map view and comprehensive filtering and download capabilities for visualizing and analyzing measurement data. When a researcher clicks a measurement point on the map, they are able to see the download, upload and latency results collected at that location. The dashboard supports researchers in two ways. First, it allows researchers to explore data that has been collected to quickly make sense of spatial and temporal density of collected measurements. This can inform how data might be utilized for analysis or trace-driven modeling. Second, the dashboard also allows researchers to export data in various formats, including CSV and JSON. This supports researchers in conducting their own offline analysis. Link to demo: https://youtu.be/JUKG5pwoX3s
We present an interactive demo of cloud-assisted real-time semantic segmentation on the edge under strict per-frame inference deadlines. The edge model runs at a fixed cadence on a live camera stream, regardless of network conditions. In parallel, a higher-capacity cloud model processes earlier frames and returns delayed feature representations to the edge device. Our system uses cloud-side predictive temporal alignment to associate delayed cloud-computed features with the latest on-device frame features, then fuses the cloud-computed features into the edge-computed features to improve segmentation quality without violating the inference schedule. The demo lets attendees vary network conditions (round-trip delay, jitter, and disconnections) and compare side-by-side real-time segmentation overlays from an edge-only baseline and a cloud-assisted edge system.
The growth of mobile devices, edge computing, and the Internet of Things (IoT) has heightened concerns about the effectiveness of traditional perimeter-based security architectures in safeguarding systems and critical infrastructure. Following numerous security breaches, a significant shift toward adopting the Zero Trust Architecture (ZTA) paradigm is underway. Unfortunately, adapting cloud-based ZTA implementations to systems that include mobile devices and remote networks that experience intermittent connectivity complicates ZTA's requirement for continuous, per-request authentication and authorization. When connectivity is disrupted-whether intentional or accidental-trust verification becomes a single point of failure, leaving edge networks vulnerable. Therefore, research is needed to develop an availability-preserving ZTA that uses local sources of trust to mask cloud outages and support responsive, real-time services. In this paper, we provide a constructive critique of the ZTA tenets and examine six application scenarios that highlight complications with cloud-based ZTA designs. To address the identified complications, we propose the Wandering Area Network Enhanced Security (WANES) framework, an adaptive trust model that enforces ZTA paradigms while preserving operational availability requirements. We also conducted a preliminary evaluation of existing open-source network monitoring solutions. Initial results indicate inadequate performance, highlighting the need for research into locally deployed ZTA monitoring and trust components.
Nearly 90% of Earth's surface remains disconnected from wireless networks, leaving vast oceans, forests, and polar regions invisible to continuous sensing. Existing "Space-IoT" offerings provide global links with proprietary satellite constellations, and paid subscriptions, restricting open and affordable deployments. We present Gaia, the first demonstration of open, continent-scale IoT connectivity that operates entirely on publicly accessible infrastructure. Gaia transmits data by "bouncing" packets off existing amateur-radio satellites, achieving direct ground-to-ground links over 1200 km without store-and-forward relays or proprietary gateways. To enable this, Gaia is designed to overcome extreme path loss and Doppler drift at the physical layer and interference sensing and coexistence challenges at the MAC layer using commodity SDRs. Beyond demonstrating feasibility, Gaia makes the case that if spectrum access analogous to ham radio were allocated for open IoT communication, it could spark a Wi-Fi-like moment for global sensing-democratizing planet-scale connectivity.
Although LEO systems such as Starlink provide low-latency access through high-speed, low-altitude orbits, they also experience frequent handovers and highly variable link conditions that undermine transport-layer performance. In this work, we present a multi-granularity prediction framework that learns to predict future satellite connections and link quality directly from network telemetry and satellite geometry. At the coarse scale, our model predicts upcoming satellite handovers and connection windows. At the fine scale, it predicts per-second performance within each window. Through real-system measurements on Starlink, we demonstrate that LEO link dynamics exhibit strong temporal and spatial correlations that can be learned and predicted accurately. Our results highlight prediction as a key enabler for proactive, satellite-aware optimization across transport, routing, and application layers in future LEO Internet systems.
For English as a Second Language (ESL) learners, unknown words significantly hinder reading fluency. We present EyeLingo, a novel method that detects unknown words in real-time by integrating user gaze trajectories with pre-trained language models (PLMs). Unlike prior gaze-only methods whose performance is highly sensitive to eye-tracking accuracy, EyeLingo leverages the linguistic context captured by RoBERTa to compensate for noisy gaze data. Our system achieves 97.6% accuracy (F1-score 71.1%) with professional eye trackers and maintains 97.3% accuracy (F1-score 65.1%) using commodity webcams.
Autonomous vehicles rely heavily on vision-based sensors that struggle in poor visibility in harsh weather and beyond line-of-sight. While mmWave radars work reliably in these conditions, they suffer from extremely sparse point clouds due to mirror-like reflection where most signals bounce away rather than returning to the radar. We present EyeDAR, a low-power mmWave tag deployed as roadside infrastructure that captures these lost reflections. The tag extracts arrival directions and sends these data back to the radar, providing additional point clouds to enhance radar perception. Like the human eye that uses a lens to map light angles onto different photoreceptors, EyeDAR uses a Luneburg lens to optically map arrival angles to different antennas, replacing O (N-3) direction-finding algorithms with O (N) detection. Combined with backscatter communication, the system operates at low power without power-hungry RF components. Our early prototype achieves 5.4 degrees effective angular resolution with >15 dB passive gain at $7 fabrication cost. We experimentally demonstrate direction-of-arrival estimation error of -0.2 degrees +/- 1.8 degrees with commercial 24 GHz radar. CCS Concepts center dot Hardware -> Wireless devices; center dot Networks -> Cyber-physical networks.
Wireless connectivity between aerial and underwater networks has inherently been a major challenge, as no single carrier operates well across the boundary. Traditional principles, such as radio in air and acoustics underwater, are fundamentally limited at the cross-media interface because of a stark mismatch in electromagnetic and acoustic properties, such as impedance, between the two different propagation media. To bridge this gap, we propose AquaTera, a new communication architecture that transforms the interface itself into a natural transducer to establish a direct air-to-water link. Namely, AquaTera strategically leverages over-the-air sub-terahertz wireless signals, with unique quasi-optical properties and strong absorption in water, to generate localized, rapid pressure waves. We develop the AquaTera principles of electromagnetic-to-acoustic coupling and design techniques for modulating aerial sub-terahertz signals into information-bearing underwater acoustic signals. We present the AquaTera framework and multiphysics models spanning three distinct, coupled domains, along with detailed preliminary results, and we discuss exciting capabilities of AquaTera to overcome the inherent limitations of existing technologies.
Mridangam and Tabla are key percussion instruments in Indian classical music. This poster presents our broad goal of achieving a fully autonomous percussion tutoring system that uses multi sensor fusion (camera, audio, motion and haptic - force or pressure). In this manuscript we present our proposed design and preliminary explorations for mridangam and discuss how it can translate to tabla analysis.
Integrated Sensing and Communication (ISAC) enables 6G networks to perform environmental sensing using communication infrastructure. We propose O-RAN extensions for monostatic sensing: (1) sensing dApps at the O-DU for IQ processing; (2) E2SM-SENS, a service model for sensing telemetry. Prototype evaluation demonstrates closed-loop latencies compatible with vehicular perception and UAV tracking use cases.
Frequency-selective fading in LTE and 5G can cause severe Signal-to-Noise Ratio (SNR) variation across subbands, leading to higher Block Error Rate (BLER) and conservative Modulation and Coding Scheme (MCS) selection. While subband-level Channel Quality Indicator (CQI) enables finer link adaptation, current systems react to fading rather than proactively mitigating it. We implement a Subband Power-Equalization at the transmitter based on subband-CQI feedback in the srsRAN LTE stack, jointly pre-equalizing data and reference signals to flatten the channel response. Our design addresses quantized/delayed feedback and reference signal consistency, operating within existing protocol limits. The Software Defined Radio (SDR) testbed evaluation shows BLER reductions and throughput gains up to 25%, with the largest improvements in the worst 5-10% of channel realizations. These results demonstrate the practicality and effectiveness of proactive, PHY-layer frequency-selective compensation in real-time networks. CCS CONCEPTS center dot Networks -> Network resources allocation; Data path algorithms; Network experimentation; Network simulations.
Emerging edge-assisted applications demand not only low latency but consistent performance and fairness across multiple users. In this work, we compare the two principal wireless networks, Wi-Fi and Private 5G (P5G), under these multi-user workloads and show that while Wi-Fi performs well for single-user scenarios, its contention-based design struggles under load, leading to jitter and fairness issues. In contrast, P5G's centrally scheduled architecture sustains predictable latency and smoother Quality of Experience (QoE) at scale. We further demonstrate that P5G's programmable control plane enables fine-grained tuning to improve user QoE. Finally, we outline a vision for application-aware wireless, where application intent and QoE feedback guide scheduling and resource allocation across the wireless access network. Our results motivate a radio-agnostic architecture that places QoE, not traditional Quality of Service (QoS), at the center of control and leverages programmability and AI-driven intelligence to deliver high-quality multi-user experiences. CCS Concepts center dot Networks -> Programmable networks; Wireless access points, base stations and infrastructure; Network measurement.
