The integration of satellite communication into mobile devices represents a paradigm shift in connectivity, yet the performance characteristics under motion and environmental occlusion remain poorly understood. We present the Starlink Robot, the first mobile robotic platform equipped with Starlink satellite internet, comprehensive sensor suite including upward-facing camera, LiDAR, and IMU, designed to systematically study satellite communication performance during movement. Our multi-modal dataset captures synchronized communication metrics, motion dynamics, sky visibility, and 3D environmental context across diverse scenarios including steady-state motion, variable speeds, and different occlusion conditions. This platform and dataset enable researchers to develop motion-aware communication protocols, predict connectivity disruptions, and optimize satellite communication for emerging mobile applications from smartphones to autonomous vehicles. In this work, we use LEOViz for real-time data collection and visualization. The project is available at https://starlinkrobot.github.io.

With the rapid development of new smart devices, such as smart home appliances and VR/AR equipment, there is an increasing demand for novel interaction methods. However, many existing interaction methods require external devices, are unintuitive, and demand substantial user learning effort. To fill this gap, we propose ArmPad, a system that leverages the smartwatch's built-in IMU to enablemultidimensional inputon the user's forearm. Methodologically, ArmPad is explicitly designed to address three core research challenges in forearm-based interaction. First, to resolve the inherent feature conflicts between discrete gesture recognition and continuous distance estimation, we propose a multi-task learning framework with a dynamic gating mechanism for cross-task synergy. Second, to tackle the physical limitation of rapid vibration attenuation across the forearm, we introduce a cross-device guidance strategy that incorporates high-fidelity fingertip knowledge during the training phase. Finally, to ensure robust generalization across diverse populations, we develop task-specific data augmentation and a lightweight user registration mechanism to effectively mitigate physiological variances. Experiments on 20 subjects demonstrate that ArmPad achieves an accuracy of 92.51% on nine gestures and a Mean Absolute Error of 1.53$cm$for sliding distance in cross-user settings. Extensive robustness evaluations and case studies further confirm the system's stability and usability under diverse real-world conditions.

Low Earth Orbit (LEO) satellites have emerged as a space-based infrastructure to offer networking services anywhere on Earth. Satellite IoTs enable novel Direct-to-Satellite (DtS) connectivity, allowing IoT devices in remote areas to connect to the Internet via LEO satellites using existing terrestrial technologies like LoRa. This paper presents the first-of-its-kind measurement study on satellite IoTs, investigating the practical characteristics of DtS communications and their suitability for IoT applications. We deployed 27 low-cost ground stations across eight locations worldwide to passively measure the network availability of multiple constellations. Our findings reveal a significant gap between the effective durations of DtS connectivity and their theoretical durations, leading to intermittent connections for satellite IoTs. Additionally, we examine the performance of the Tianqi constellation in supporting real-world IoT traffic (agriculture application). We observed longer delays and higher power consumption in satellite IoTs compared to terrestrial IoTs. Our study identifies the bottlenecks and sheds light on potential optimizations for satellite IoTs.

Daily eating habits shape our long-term health, but most diet apps focus only on calories or macronutrients and overlook deeper issues like chronic inflammation and its effects on cellular aging. Prior medical literature has demonstrated a significant inverse relationship between Dietary Inflammatory Index (DII) and telomere length (TL), a key marker of cellular age. Inspired by this finding, we design ChronoBite, a closed-loop feedback system that pairs real-time inflammation scores with periodic cellular aging insights. In addition to regular calorie tracking, it uses fast-changing DII signals and slow-moving cellular aging markers to guide users toward age-aware eating habits. ChronoBite is a mobile-based prototype that combines food recognition, DII analysis, and cellular aging insights to deliver age-aware dietary feedback. Powered by large language models (LLMs), it offers real-time recommendations while supporting long-term tracking of inflammation patterns and telomere dynamics.

The advent of Large Language Models (LLMs) has profoundly transformed our lives, revolutionizing interactions with AI and lowering the barrier to AI usage. While LLMs are primarily designed for natural language interaction, the extensive embedded knowledge empowers them to comprehend digital sensor data. This capability enables LLMs to engage with the physical world through IoT sensors and actuators, performing a myriad of AIoT tasks. Consequently, this evolution triggers a paradigm shift in conventional AIoT application development, democratizing its accessibility to all by facilitating the design and development of AIoT applications via natural language. However, some limitations need to be addressed to unlock the full potential of LLMs in AIoT application development. First, existing solutions often require transferring raw sensor data to LLM servers, which raises privacy concerns, incurs high query fees, and is limited by token size. Moreover, the reasoning processes of LLMs are opaque to users, making it difficult to verify the robustness and correctness of inference results. This paper introduces AutoIOT, an LLM-based automated program generator for AIoT applications. AutoIOT enables users to specify their requirements using natural language (input) and automatically synthesizes interpretable programs with documentation (output). AutoIOT automates the iterative optimization to enhance the quality of generated code with minimum user involvement. AutoIOT not only makes the execution of AIoT tasks more explainable but also mitigates privacy concerns and reduces token costs with local execution of synthesized programs. Extensive experiments and user studies demonstrate AutoIOT's remarkable capability in program synthesis for various AIoT tasks. The synthesized programs can match and even outperform some representative baselines.

Ubiquitous acoustic intelligence leverages the pervasiveness of sound and the ubiquity of microphones/speakers to enable intelligent, seamless, and privacy-preserving interaction between humans and mobile devices. However, enabling such intelligence poses multiple challenges: human voice and behavior manifest in acoustics in subtle, complex ways that are difficult to capture robustly; acoustic communication suffers from significant distortion in complex environments and cross-heterogeneous devices; and processing audio data requires careful design to ensure privacy and efficiency. To address these challenges, we developed a series of systems across three pillars: intelligent acoustic sensing, robust acoustic communication, and privacy-aware acoustic computing.

Gait behaviour is a key health metric. Temporal, spatial and kinetic walking gait parameters are valuable in enhancing sport performance and early health diagnostics Full gait assessment requires a gait clinic and existing wearable gait tracking systems typically measure isolated subsets of parameters tailored to specific applications. This is useful when the condition to be monitored is known, but fails to offer a comprehensive view of an individual’s gait traits when their pathology is unknown or changing, or a general assessment is required. To support holistic walking gait tracking, we introduce WalkEar, a novel sensing platform designed to simultaneously track gait parameters using commodity earbuds. WalkEar operates by detecting gait events to derive temporal gait parameters and segment the IMU data. WalkEar then progresses earable gait assessment by, for the first time, estimating kinetic gait parameters and reconstructing the vGRF curve using machine learning. Each parameter is calculated on a step-to-step basis for gait variability and asymmetry. We developed an earbud prototype and collected data from 13 participants using gold standard force plates and instrumented treadmill ground truth. Extensive experiments demonstrate the promising performance of WalkEar, achieving an overall MAPE of 5.1% in estimating gait, 2.0% MAPE on kinetic gait parameters, and an NRMSE of 5.3% for vGRF curve reconstruction.

Code Large Language Models (LLMs) enhance software development efficiency by automatically generating code and documentation based on user requirements. However, code LLMs cannot synthesize specialized programs when tasked with IoT applications that require domain knowledge. While Retrieval-Augmented Generation (RAG) offers a promising solution by fetching relevant domain knowledge, it necessitates powerful cloud LLMs (e.g., GPT-4) to process user requirements and retrieved contents, which raises significant privacy concerns. This approach also suffers from unstable networks and prohibitive LLM query costs. Moreover, it is challenging to ensure the correctness and relevance of the fetched contents. To address these issues, we propose GPIoT, a code generation system for IoT applications by fine-tuning locally deployable Small Language Models (SLMs) on IoT-specialized datasets. SLMs have smaller model sizes, allowing efficient local deployment and execution to mitigate privacy concerns and network uncertainty. Furthermore, by fine-tuning SLMs with our IoT-specialized datasets, the SLMs' ability to synthesize IoT-related programs can be substantially improved. To evaluate GPIoT's capability in synthesizing programs for IoT applications, we develop a benchmark, IoTBench. Extensive experiments and user trials demonstrate the effectiveness of GPIoT in generating IoT-specialized code, outperforming state-of-the-art code LLMs with an average task accuracy increment of 64.7% and significant improvements in user satisfaction.

Federated Learning (FL) collaboratively trains a shared global model without exposing clients' private data. In practical FL systems, clients (e.g., smartphones and wearables) typically have disparate system resources. Traditional FL, however, adopts a one-size-fits-all solution, where a homogeneous large model is sent to and trained on each client. This method results in an overwhelming workload for less capable clients and starvation for others. To tackle this, we proposeFedConv, a client-friendly FL framework, minimizing the system overhead on resource-constrained clients by providing heterogeneous customized sub-models.FedConvfeatures a novellearning-on-modelparadigm that learns the parameters of heterogeneous sub-models viaconvolutional compression. To aggregate heterogeneous sub-models, we proposetransposed convolutional dilationto convert them back to large models with a unified size while retaining personalized information. The compression and dilation processes, transparent to clients, are tuned on the server using a small public dataset. We further propose ahierarchical and clustering-based local trainingstrategy for enhanced performance. Extensive experiments on six datasets show thatFedConvoutperforms state-of-the-art FL systems in terms of model accuracy (by more than 35% on average), computation and communication overhead (with 33% and 25% reduction, respectively).

Every day, one person loses his life due to drowning in swimming pools, even with professional lifeguards present. Contrary to what the public might assume, drowning swimmers can hardly splash or yell for help. This life-threatening situation calls for a robust SOS channel between the swimmers and the lifeguards. This paper proposes Neusos, a neural-enhanced underwater SOS communication system based on commercial wearable devices and low-cost hydrophones deployed in the swimming pool. Specifically, we repurpose popular wearable devices (e.g., smartwatches) as SOS transmitters, which can send a distress signal when the user is in an emergency. In response, an underwater hydrophone in the swimming pool can detect SOS signals and make alerts immediately to facilitate a timely rescue. The main technical challenge lies in reliably detecting weak SOS signals in non-stationary underwater scenarios. To achieve so, we thoroughly characterize the properties of underwater channels and examine the limitations of the traditional correlation-based signal detection method in underwater communication scenarios. Based on our empirical findings, we developed a robust SOS detection method enhanced with deep learning. By fully embedding signal characteristics into networks, Neusos outperforms traditional signal processing-based underwater SOS detection methods. In particular, our experiments in a real swimming pool show that Neusos can detect SOS signals with a detection rate of 98.2% under various underwater conditions. Given the increasing popularity of smartwatches among swimmers, our system holds immense potential to enhance their safety in swimming pools.

Federated Learning (FL) facilitates collaborative training of a shared global model without exposing clients' private data. In practical FL systems, clients (e.g., edge servers, smartphones, and wearables) typically have disparate system resources. Conventional FL, however, adopts a one-size-fits-all solution, where a homogeneous large global model is transmitted to and trained on each client, resulting in an overwhelming workload for less capable clients and starvation for other clients. To address this issue, we propose FedConv, a client-friendly FL framework, which minimizes the computation and memory burden on resource-constrained clients by providing heterogeneous customized sub-models. FedConv features a novel learning-on-model paradigm that learns the parameters of the heterogeneous sub-models via convolutional compression. Unlike traditional compression methods, the compressed models in FedConv can be directly trained on clients without decompression. To aggregate the heterogeneous sub-models, we propose transposed convolutional dilation to convert them back to large models with a unified size while retaining personalized information from clients. The compression and dilation processes, transparent to clients, are optimized on the server leveraging a small public dataset. Extensive experiments on six datasets demonstrate that FedConv outperforms state-of-the-art FL systems in terms of model accuracy (by more than 35% on average), computation and communication overhead (with 33% and 25% reduction, respectively).

Voice assistants are widely integrated into a variety of mobile devices, enabling users to easily complete daily tasks and even critical operations like online transactions with voice commands. Thus, once attackers replay a secretly-recorded voice command by loudspeakers to compromise users' voice assistants, this operation will cause serious consequences, such as information leakage and property loss. Unfortunately, most voice liveness detection approaches against replay attacks mainly rely on detecting lip motions or subtle physiological features in speech, which are limited within a very short range. In this paper, we propose VoShield to check whether a voice command is from a genuine user or a loudspeaker imposter. VoShield measures sound field dynamics, a feature that changes fast as the human mouths dynamically open and close. In contrast, it would remain rather stable for loudspeakers due to the fixed size. This feature enables VoShield to largely extend the working distance and remain resilient to user locations. Besides, sound field dynamics are extracted from the difference between multiple microphone channels, making this feature robust to voice volume. To evaluate VoShield, we conducted comprehensive experiments with various settings in different working scenarios. The results show that VoShield can achieve a detection accuracy of 98.2% and an Equal Error Rate of 2.0%, which serves as a promising complement to current voice authentication systems for smart mobile devices.

Integrated Sensing and Communication (ISAC) has emerged as a promising technology for next-generation mobile networks. Towards ISAC, we developmmRipplethat empowers commodity mmWave radars with communication capabilities through smartphone vibrations. InmmRipple, a smartphone (transmitter) sends messages by modulating smartphone vibrations, while a mmWave radar (receiver) receives the messages by detecting and decoding the smartphone vibrations. By doing so, a smartphone user can not only be passively sensed by a mmWave radar, but also actively send messages to the radar without any hardware modifications. Although promising, the data rate ofmmRippleis limited by Morse-style communication. To address this, we presentmmRipple+, which leverages the Pulse Width and Amplitude Modulation (PWAM) technique and suppresses inter-symbol interference to enable faster communication. We prototypemmRippleandmmRipple+on commodity mmWave radars and different types of smartphones. Experimental results show thatmmRippleachieves an average vibration pattern recognition accuracy of 98.60% within a$ 2$m communication range, and 97.74% within$ 3$m. The maximum communication range extends to$ 5$m. Meanwhile,mmRipple+achieves a bit rate of 100 bps with a BER of less than 3%, improving the data rate by 4× overmmRippewith the same symbol duration. This work pioneers smartphone-to-COTS mmWave radar communication via vibrations, unlocking diverse applications.

The binaural microphone, which refers to a pair of microphones with artificial human-shaped ears, is widely used in hearing aids and spatial audio recording to improve sound quality. It is crucial for such devices to find the voice direction in many applications such as binaural sound enhancement. However, sound localization with two microphones remains challenging, especially in multi-source scenarios. Most previous work utilized microphone arrays to deal with the multi-source localization problem. Extra microphones yet have space constraints for deployment in many scenarios (e.g., hearing aids). Inspired by the fact that humans have evolved to locate multiple sound sources with only two ears, we propose DeepEar, a binaural microphone-based sound localization system. To this end, we design a multisector-based neural network to locate multiple sound sources simultaneously, where each sector is a discretized region of the space for different angle of arrivals. DeepEar fuses explicit hand-crafted features and implicit latent sound representatives to facilitate sound localization. More importantly, the trained DeepEar model can adapt to new environments with a minimum amount of extra training data. The experiment results show that DeepEar substantially outperforms the state-of-the-art binaural deep learning approach by a large margin in terms of sound detection accuracy and azimuth estimation error.

This paper presents the design and implementation of mmRipple, which empowers commodity mmWave radars with the communication capability through smartphone vibrations. In mmRipple, a smartphone (transmitter) sends messages by modulating smartphone vibrations, while a mmWave radar (receiver) receives the messages by detecting and decoding the smartphone vibrations with mmWave signals. By doing so, a smartphone user can not only be passively sensed by a mmWave radar, but also actively send messages to the radar using her smartphone without any hardware modifications to either the smartphone or the mmWave radar. mmRipple addresses a series of unique technical challenges, including vibration signal generation, tiny vibration sensing, multiple object separation, and movement interference mitigation. We implement and evaluate mmRipple using commodity mmWave radars and smartphones in different practical conditions. Experimental results show that mmRipple achieves an average vibration pattern recognition accuracy of 98.60% within a 2m communication range, and 97.74% within 3m on 11 different types of smartphones. The communication range can be further extended up to 5m with an accuracy of 91.67% with line-of-sight path. To our best knowledge, mmRipple is the first work that allows smartphones to send data to COTS mmWave radars via smartphone vibrations and will enable many new applications such as vibration-based near field communication and pedestrian-to-sensing-infrastructure communication.

Drowning incidents can occur in swimming pools even with professional lifeguards present. This is because drowning swimmers often face difficulties in calling for help due to choking, making it challenging for lifeguards to recognize them and provide a timely rescue. To address this problem, this paper presents AquaHelper, an underwater SOS system that can transmit and detect acoustic SOS signals in swimming pools. Specifically, a wearable device (e.g., a smartwatch) serves as an underwater SOS transmitter, with which a swimmer can call for help in emergency situations. Multiple underwater acoustic receivers are deployed to detect SOS signals and promptly alert lifeguards. The main challenge lies in the low transmission power of lightweight wearable devices, which poses difficulties in detecting weak SOS signals, particularly in low-SNR underwater scenarios. To achieve reliable underwater SOS detection, AquaHelper develops novel techniques (e.g., incorporating high-order harmonics, multi-scale window aggregation, and coherent combining of multiple receivers) to fully leverage the spectral, temporal, and spatial diversity of underwater acoustic signals. We also describe lessons learned and our solutions to address practical challenges involved in underwater SOS transmission and detection. Our experiments demonstrate the effectiveness of AquaHelper in detecting SOS signals in typical swimming pool environments.

Binaural microphones, referring to two microphones with artificial human-shaped ears, are pervasively used in humanoid robots and hearing aids improving sound quality. In many applications, it is crucial for such robots to interact with humans by finding the voice direction. However, sound source localization with binaural microphones remains challenging, especially in multi-source scenarios. Prior works utilize microphone arrays to deal with the multi-source localization problem. Extra arrays yet incur higher deployment costs and take up more space. However, human brains have evolved to locate multiple sound sources with only two ears. Inspired by this fact, we propose DeepEar, a binaural microphone-based localization system that can locate multiple sounds. To this end, we develop a neural network to mimic the acoustic signal processing pipeline of the human auditory system. Different from hand-crafted features used in prior works, DeepEar can automatically extract useful features for localization. More importantly, the trained neural networks can be extended and adapted to new environments with a minimum amount of extra training data. Experiment results show that DeepEar can substantially outperform the state-of-the-art deep learning approach, with a sound detection accuracy of 93.3% and an azimuth estimation error of 7.4 degrees in multisource scenarios.

Millimeter wave (mmWave) radar has demonstrated excellent performance in object tracking and micro-displacement detection. Besides the powerful sensing function, this work brings the communication function, allowing daily devices to communicate with mmWave radars through vibrations. In this work, we present VibBeat, in which a daily device (e.g., smartphone and smartwatch) sends messages by modulating vibrations, while a mmWave radar receives the messages by detecting and decoding the vibrations with reflected mmWave signals. By doing so, the device (user) can not only be passively sensed by a mmWave radar, but also actively send messages to the radar for a personalized response. We implement our system using a COTS mmWave radar and smartphones without any hardware modification. Experimental results show that VibBeat supports multiple object communication and achieves a communication range of up to 5m.

Recently, wearable devices have become increasingly popular in our lives because of their neat features and stylish appearance. However, their tiny sizes bring about new challenges to human-device interaction such as texts input. Although some novel methods have been put forward, they possess different defects and are not applicable to deal with the problem. As a result, we propose an acoustic-based texts-entry system, i.e., EchoWrite, by which texts can be entered with a finger writing in the air without wearing any additional device. More importantly, different from many previous works, EchoWrite runs in a training-free style which reduces the training overhead and improves system scalability. We implement EchoWrite with commercial devices and conduct comprehensive experiments to evaluate its texts-entry performance. Experimental results show that EchoWrite enables users to enter texts at a speed of 7.5 WPM without practice, and 16.6 WPM after about 30-minute practice. This speed is better than touch screen-based method on smartwatches, and comparable with previous related works. Moreover, EchoWrite provides favorable user experience of entering texts.

Recently, wearable devices have become increasingly popular in our lives because of their neat features and stylish appearance. However, their tiny sizes bring about new challenges to human-device interaction such as texts input. Although some novel methods have been put forward, they possess different defects and are not applicable to deal with the problem. As a result, we propose an acoustic-based texts-entry system, i.e., EchoWrite, by which texts can be entered with a finger writing in the air without wearing any additional device. More importantly, different from many previous works, EchoWrite runs in a training-free style which reduces the training overhead and improves system scalability. We implement EchoWrite with commercial devices and conduct comprehensive experiments to evaluate its texts-entry performance. Experimental results show that EchoWrite enables users to enter texts at a speed of 7.5 WPM without practice, and 16.6 WPM after about 30- minute practice. This speed is better than touch screen-based method on smartwatches, and comparable with previous related works.

Recently, wearable devices have become increasingly popular in our lives because of their neat features and stylish appearance. However, due to the tiny size, it is inconvenient for users to interact with a device using conventional methods, especially for text entry. Although some methods have been proposed to handle this problem, they have different limitations and are not applicable to many existing mobile devices. As a result, we take the first step to propose a digits-entry system, i.e., AcouDigits, in which digits can be entered in the air using a finger without taking help from any additional hardware. We implement AcouDigits on two commercial devices and conduct experiments to evaluate its performance in recognizing ten basic digits. Experimental results show that AcouDigits can achieve average accuracies of 91.7% and 87.4% in recognizing basic digits and 26 English alphabets, respectively.

EMG is becoming an emerging interface for human-computer interface and has been applied to gesture recognition in previous work. However, those existing EMG-based interfaces can only recognize gestures at a coarse-grained level such as hand and arm gestures, which constraints their usage in applications involving fine-grained activities such as text entry via keystrokes. As a result, in this paper, we attempt to push the limit of existing EMG-based interfaces and propose the first wearable text-entry system, named ArmIn, with EMG signals. ArmIn is designed to recognize keystroke gestures with the help of a finger on printed and physical keyboards. We implement ArmIn using commodity EMG sensors and custom hardware board, and conduct experiments to evaluate its performance. By carefully designing the data processing scheme, ArmIn can recognize keystrokes on both kinds of keyboard, with 89.5% and 87.5% accuracy respectively, when it is worn on a user's left arm.

Recently, tiny smart devices have become increasingly popular in our lives because of their neat features and stylish appearance. However, their small form factors, especially screens, make it inconvenient for users to enter texts with conventional methods such as soft keyboards, which need a fairly large screen. To address this problem, we propose a novel texts-input system, called EchoType, with which users can enter texts with a finger writing in the air. EchoType makes use of acoustic sensors (i.e., microphone and speaker) to sense finger gestures and infer texts based on mapping relation between gestures and basic letters. We take a step to enable users to input texts with acoustic signals. Compared with existing approaches, EchoType enjoys merits of low hardware requirements and high scalability to different mobile devices.