MobiCom 2026 Papers

Accurate sensing and localization of surface force are crucial in various applications, including muscle rehabilitation, robotic manipulation of fragile objects, and finger rehabilitation training. However, existing technologies struggle to achieve high-precision and continuous force sensing, due to the crosstalk interference from dense sensor arrays or ambient light interference from optical sensors or reliance on batteries and power-hungry components.

WiFi-based 3D human pose estimation offers a low-cost and privacy-preserving alternative to vision-based systems for smart interaction. However, existing approaches rely on visual 3D poses as supervision and directly regress CSI to a camera-based coordinate system. We find that this practice leads to coordinate overfitting: models memorize deployment-specific WiFi transceiver layouts rather than only learning activity-relevant representations, resulting in severe generalization failures. To address this challenge, we present PerceptAlign, the first geometry-conditioned framework for WiFi-based cross-layout pose estimation. PerceptAlign introduces a lightweight coordinate unification procedure that aligns WiFi and vision measurements in a shared 3D space using only two checkerboards and a few photos. Within this unified space, it encodes calibrated transceiver positions into high-dimensional embeddings and fuses them with CSI features, making the model explicitly aware of device geometry as a conditional variable. This design forces the network to disentangle human motion from deployment layouts, enabling robust and, for the first time, layout-invariant WiFi pose estimation. To support systematic evaluation, we construct the largest cross-domain 3D WiFi pose estimation dataset. Experiments show that PerceptAlign reduces cross-domain error by more than 60% compared to state-of-the-art baselines. These results establish geometry-conditioned learning as a viable path toward practical WiFi sensing.

Underground cellulars are arguably the most sophisticated public network utilities in modern metropolises due to constrained environments and mobility-driven use cases. To date, their QoS has been largely unsatisfactory to users, especially on subway trains where video apps encounter frequent and lasting video stream stalls (VSSes). With the goal of deeply understanding and practically boosting underground cellular QoS, we conduct the first large-scale study of underground cellular networks, spanning 207 subway lines, 82K BSes, and 1.8M Android phones. Our study reveals the unique multi-stage signal attenuation pattern and the defective infrastructure deployment in today's subway cellular environments, which collectively contribute as the major causes of VSSes. To address this, we enable user devices to actively adapt to the environmental changes, and develop a comprehensive radio/link state collection module to effectively inform video apps. Our optimizations on device-side Android cellular management have reduced 39% VSSes in subway networks; our collaborative optimizations with popular video apps have further reduced 23% VSSes.

Intelligent systems powered by large-scale sensor networks are shifting from predefined monitoring to intent-driven operation, revealing a critical Semantic-to-Physical Mapping Gap. While large language models (LLMs) excel at semantic understanding, existing perception-centric pipelines operate retrospectively, overlooking the fundamental decision of what to sense and when. We formalize this proactive decision as Semantic-Spatial Sensor Scheduling (S3) and demonstrate that direct LLM planning is unreliable due to inherent gaps in representation, reasoning, and optimization. To bridge these gaps, we introduce the Spatial Trajectory Graph (STG), a neurosymbolic paradigm governed by a “verify-before-commit” discipline that transforms open-ended planning into a verifiable graph optimization problem. Based on STG, we implement IoT-Brain, a concrete system embodiment, and construct TopoSense-Bench, a campus-scale benchmark with 5,250 natural-language queries across 2,510 cameras. Evaluations show IoT-Brain boosts task success rate by 37.6% over the strongest search-intensive methods while running nearly 2× faster and using 6.6× fewer prompt tokens. In real-world deployment, it approaches the reliability upper bound set by reducing 4.1× network bandwidth, providing a foundational framework for LLMs to interact with the physical world with unprecedented reliability and efficiency.

UAV-assisted inspection is vital for large-scale power transmission networks. Beyond safe trajectory generation, expert-level planning requires jointly optimizing UAV paths and camera configurations, adapting to diverse structures and inspection standards to ensure reliable visual data. We propose A3TP, an automated, accurate, and adaptive UAV task planning system tailored for large-scale power transmission network inspection. A3TP integrates: (1) Expertise-driven task modeling to encode human planning knowledge from historical plans, (2) Geometry-centric planning to generate reference trajectories and shooting parameters for identifying all inspection targets, and (3) Deployment-adaptive trajectory optimization to ensure safety and inspection coverage in complex environments. Operating in a knowledge-driven manner, A3TP achieves up to 55.78% IoU improvement over supervised baselines, over 96% classification accuracy, and trajectory planning results comparable to human experts. In a real-world deployment covering 373 towers across 65.84 km, A3TP reduced planning time by 95.3%, achieved an over 88% expert acceptance rate, and reliably detected field defects. These results demonstrate A3TP's scalability and practical value for maintaining real-world power transmission networks. Code and dataset are available at: https://github.com/fqykinhgt/A3TP.git.

mmWave networks promise high bandwidth but face significant challenges in maintaining reliable connections for users moving at high speed. Frequent handovers, complex beam alignment, and signal blockage from car bodies lead to service interruptions and degraded performance. We present Wall-Street, a vehicle-mounted smart surface that enhances mmWave connectivity for in-vehicle users. Wall-Street improves mobility management by (1) steering outdoor mmWave signals into the vehicle for shared coverage and providing a single, collective handover for all users; (2) performing neighbor-cell search without interrupting data transfer, ensuring seamless handovers; and (3) connecting users to a new cell before disconnecting from the old cell for reliable cell transitions. We implemented and integrated Wall-Street into the COSMOS testbed. We collected PHY traces with multiple base station nodes and in-vehicle user nodes with a surface-mounted vehicle, driving on a nearby road. Our trace-driven ns-3 simulation demonstrates a throughput improvement of up to 78% and a latency reduction of up to 34% over the standard Standalone handover scheme.1

To enable commercial-compatible Wi-Fi backscatter transmission, recent works either adopt off-tag mechanisms based on standard OFDM Wi-Fi exciters, or generate OFDM Wi-Fi waveforms on-tag using single-tone exciter. However, they result in either very limited communication range or exorbitant power consumption. We present PassiveOFDM, the first μW-level commercial-compatible OFDM Wi-Fi backscatter system that works at tens of meters exciter-to-tag distance. It enables low-power on-tag OFDM signal forming, decoupling tag placement from the exciter's location. We design a spectrum-filling shaping scheme that retains the spectrum structure of the OFDM Wi-Fi without IFFT, and a DAC-free waveform embedding scheme with low-power RF switches. Our field tests confirm that PassiveOFDM achieves reliable backscatter transmission 65m from the exciter or 80m from the receiver. Its power consumption is 2855.1× smaller than the SOTA on-tag system RF-Transformer. Its exciter-to-tag distance and throughput are 54.2× and 29× higher than those of the SOTA off-tag system TiScatter, respectively.

Accurately modeling the radiance field of 5G signals in urban environments is essential for reliable network operation and scalable planning of next-generation wireless systems. Recent advances in neural radiance fields (NeRF2) have shown great success in capturing the spatially continuous behavior of radio propagation. However, existing NeRF2 frameworks remain limited to small-scale scenes, suffering from severe accuracy degradation and prohibitive computational costs when extended to city-scale deployments. In this work, we present Mixture of Radio Experts (MoRE), a mixture-of-experts architecture designed for city-scale 5G radiance field reconstruction. MoRE integrates two complementary components: (i) a Mixture of Cell Experts, where each expert captures localized attenuation effects within distinct geographic regions; and (ii) a Mixture of Radiance Experts, which leverages shared and dynamically routed experts to model diverse propagation phenomena. To evaluate MoRE, we conducted a large-scale measurement campaign comprising 7,073,565 timestamped and geotagged RF samples from 13,602 base stations across 1,900 km2 in a tier-one city. Experimental results show that MoRE delivers up to a 16 dB improvement over standard NeRF2 by scaling from 0.69 million to 300 million parameters, while activating only 0.28% of the parameters (about 0.85 million) per inference.

GB-scale large apps like on-device LLMs and rich media editors are becoming the next-generation trend, but their heavy memory and I/O demands, especially during multitasking, cause devices to reclaim or kill processes, turning warm apps into cold launches. The challenge lies not in storing them, but in fast, accurate launching. For users, 1s is the usability cliff, yet our measurements show 86.6% of GB-scale cold launches exceed it. Also, Android Vitals flags only ≥5s as slow, exposing a large satisfaction gap. Existing optimizations are designed in isolation and conflict. For example, preloading reduces I/O stalls but consumes scarce memory and is undone by reclamation, while reclamation and killing free memory but sacrifice background survivability, leading to repeated cold relaunches. Our key insight is that, although multitasking makes runtime behavior complex, each app's file access pattern remains predictable. The challenge lies in exploiting this predictability, i.e., preloading without exhausting memory, reclaiming without undoing gains, and killing selectively to preserve background survivability. We introduce AppFlow, a prediction-based system-wide scheduler that integrates a Selective File Preloader, an Adaptive Memory Reclaimer, and a Context-Aware Process Killer. Implemented across the Android framework and Linux kernel without app changes, AppFlow cuts GB-scale cold-launch latency by 66.5% (e.g., 2s→690ms) and sustains 95% of launches within 1s over a 100-day test, significantly improving responsiveness and multitasking experience.

Task-oriented communications (ToC), also known as semantic communications, offers a promising paradigm for offloading deep neural network (DNN) inference from edge devices to more powerful remote servers. By transmitting only task-specific information, ToC significantly reduces the communication overhead inherent in conventional offloading systems. However, existing ToC approaches exhibit limited task adaptivity, requiring retraining or separate encoders for each task—hindering their practicality in real-world multi-task settings. In this work, we present TACToC, the first ToC system to achieve task adaptivity with a single encoder, enabled by a novel task-describable encoder structure. TACToC first maps input data into a semantically structured feature space using a self-supervised vision transformer (SSL-ViT), providing a foundation for flexible task-specific feature selection. The encoder then dynamically selects and transmits only important features for the target task, without per-task optimization. This is realized through two key mechanisms: (i) relevant feature estimation, which quantifies each feature's contribution to task inference, and (ii) prompt-based task embedding, which injects task-specific information to guide selective transmission for each task. Extensive experiments demonstrate that TACToC reduces communication cost by up to 1.5-14.5× within the range that allows for a 3%p degradation while supporting diverse vision tasks with seamless task integration and transition. This work marks a significant step toward practical and unified ToC systems for dynamic and multi-task AI services.

Reconfigurable Intelligent Surfaces (RIS) are a key 6G technology for controlling the wireless environment and eliminating signal blockages. However, dominant planar meta-surfaces suffer from high cost, complex control, limited 3D coverage, and high beam alignment overhead. This paper presents OrbWeaver, an RIS that replaces complex metasurfaces with a 3D-printed Luneburg lens and a simple reconfigurable RF switching fabric. It leverages geometric optics for cost-effective, programmable signal redirection across an omnidirectional 3D field of view. A key advantage is its ability to perform in-situ power measurements, enabling a decoupled beam alignment protocol that drastically simplifies multi-node coordination. We validate OrbWeaver with mmWave and sub-6 GHz prototypes, demonstrating its frequency-agnostic design. In real-world 5G and Wi-Fi deployments, OrbWeaver eliminates dead zones, achieving throughput gains up to 430 Mbps and 241 Mbps, respectively. Our results establish OrbWeaver as a practical RIS solution to overcome the deployment barriers of conventional metasurfaces.

Millimeter-wave (mmWave) radar backscatter communication (RBC) is an emerging Integrated Sensing and Communication (ISAC) paradigm that combines high-resolution sensing with low-power connectivity. However, existing systems rely on frequency modulation or coarse frequency-division multiplexing, forcing each tag to occupy the full frequency modulated continuous wave (FMCW) bandwidth or sacrifice data rate for concurrency, which fundamentally limits scalability. This raises a key question: can radar-centric RBC systems support OFDMA-like multi-user access?

Digital twins, virtual simulated replicas of physical scenes, are transforming system design across industries. However, their potential in radio frequency (RF) systems has been limited by the non-differentiable nature of conventional RF simulators. The visibility of propagation paths causes severe discontinuities, and differentiable rendering techniques from computer graphics are no longer applicable due to point-source antennas and dominant specular reflections. In this paper, we present RFDT, a physically based differentiable RF simulation framework that enables gradient-based interaction between virtual and physical worlds. RFDT resolves discontinuities with a physically grounded edge-diffraction transition function, and mitigates non-convexity from Fourierdomain processing through a signal-domain transform surrogate. Our implementation demonstrates RFDT's ability to accurately reconstruct digital twins from real RF measurements. Moreover, RFDT can augment diverse downstream applications, such as test-time adaptation of machine learning-based RF sensing and physically constrained optimization of communication systems. The supplementary materials and code of RFDT are released at rfdigitaltwin.github.io.

The rapid adoption of 5G New Radio (NR), particularly in the millimeter-wave (mmWave) spectrum, imposes stringent demands on baseband processing efficiency. While virtualized Radio Access Networks (vRANs) enable dynamic spectrum sharing, compute resource allocation for baseband processing, especially in multi-cell deployments, remains underexplored. In this paper, we present Nexus, the first system to realize real-time, virtualized multi-cell mmWave baseband processing on a single server with heterogeneous compute resources. For single-cell operation, Nexus employs a random forest (RAF)-based model to predict energy-efficient resource allocations with microsecond-level inference latency. For multi-cell scenarios, Nexus introduces a power-aware scheduler that incorporates a lightweight contention model to adapt allocations under concurrent execution. Through extensive evaluations, we show that Nexus supports up to sixteen concurrent cells under full load, achieving 5.37 Gbps aggregate throughput using sixteen CPU cores and a single physical ACC100 accelerator virtualized into sixteen virtual functions (VFs), while reducing the multi-cell scheduling search space by orders of magnitude.

Millimeter-wave Human Pose Estimation (mmWave HPE) promises privacy but suffers from poor generalization under distribution shifts. We demonstrate that brute-force data scaling is ineffective for out-of-distribution (OOD) robustness; efficiency and coverage are the true bottlenecks. To address this, we introduce WiCompass, a coverage-aware data-collection framework. WiCompass leverages large-scale motion-capture corpora to build a universal pose space “oracle” that quantifies dataset redundancy and identifies underrepresented motions. Guided by this oracle, WiCompass employs a closed-loop policy to prioritize collecting informative missing samples. Experiments show that WiCompass consistently improves OOD accuracy at matched budgets and exhibits superior scaling behavior compared to conventional collection strategies. By shifting focus from brute-force scaling to coverage-aware data acquisition, this work offers a practical path toward robust mmWave sensing.

Wireless sensing, traditionally relying on signal processing (SP) techniques, has recently shifted toward data-driven deep learning (DL) to achieve performance breakthroughs. However, existing deep wireless sensing models are typically end-to-end and task-specific, lacking reusability and interpretability. We propose RF-LEGO, a modular co-design framework that transforms interpretable SP algorithms into trainable, physics-grounded DL modules through deep unrolling. By replacing hand-tuned parameters with learnable ones while preserving core processing structures and mathematical operators, RF-LEGO ensures modularity, cascadability, and structure-aligned interpretability. Specifically, we introduce three deep-unrolled modules for critical RF sensing tasks: frequency transform, spatial angle estimation, and signal detection. Extensive experiments using real-world data for Wi-Fi, millimeter-wave, UWB, and 6G sensing demonstrate that RF-LEGO significantly outperforms existing SP and DL baselines, both standalone and when integrated into multiple downstream tasks. RF-LEGO pioneers a novel SP-DL co-design paradigm for wireless sensing via deep unrolling, shedding light on efficient and interpretable deep wireless sensing solutions. Our code is available at https://github.com/aiot-lab/RF-LEGO.

This paper proposes an efficient attack on millimeter-wave (mmWave) security screening systems by designing a novel reflective metasurface, termed MagicPatch, which is applied to prohibited items hidden under clothing and alters imaging results to resemble those of the human body, significantly reducing detection rates. In wideband mmWave imaging systems, MagicPatch is optimized in two key aspects: (1) tuning group delay parameters to manipulate the Time-of-Flight (i.e., perceived object distance); and (2) adjusting reflectivity to control the intensity of the reflected signals (i.e., material properties). To enable automated MagicPatch design, we propose a joint optimization framework that determines the optimal structural configurations of the MagicPatch for targeted attack scenarios, while supporting a low-cost fabrication process. Comprehensive experiments are conducted on two commercial mmWave security scanners with volunteers of diverse body types. The results demonstrate that the MagicPatch effectively conceals eight types of prohibited items with thicknesses ranging from 1 to 5 cm, achieving an alarm suppression rate of 85.7% to 97.0%, significantly higher than baseline methods at 9.4% to 36.7%.

We propose V-Droid, a mobile GUI task automation agent. Unlike previous mobile agents that utilize Large Language Models (LLMs) as generators to directly generate actions at each step, V-Droid employs LLMs as verifiers to evaluate candidate actions before making final decisions. To realize this novel paradigm, we introduce a comprehensive framework for constructing verifier-driven mobile agents: the discretized action space construction coupled with the prefilling-only workflow to accelerate the verification process, the pair-wise progress preference training to significantly enhance the verifier's decision-making capabilities, and the scalable human-agent joint annotation scheme to efficiently collect the necessary data at scale.

Sensing micro-vibration plays a critical role in diverse industrial applications, including structural health monitoring and machinery diagnostics. However, most existing vibration sensing technologies rely on physical contact with the target object or demand complex installation procedures, making them intrusive and costly. This paper presents PocketVib, the first cost-effective, portable smartphone-based laser speckle vibrometry system for multi-point kHz-frequency microvibration sensing. PocketVib integrates built-in smartphone sensors, i.e., Time-of-Flight (ToF) depth cameras and rolling shutter cameras, to achieve non-contact micron-level vibration sensing. The design features three key innovations: 1) emitter-imager synchronization, which ensures stable speckle pattern acquisition, 2) a dual-lens enhanced imager that stretches speckle patterns to improve their limited coverage on the rolling shutter camera, and 3) a reference-free algorithm that leverages temporal correlations across rolling shutter image rows for precise vibration extraction. We implement PocketVib on commodity smartphones with a custom hardware kit. Extensive experimental results demonstrate that PocketVib enables fast and accurate vibration frequency estimation on mobile devices, achieving frequency measurements of up to 1 kHz with an average error of only 0.6 Hz at just 30 fps, while supporting simultaneous measurements at seven points. The project page is available at https://shangchengjin.github.io/PocketVibWeb.