The rapid proliferation of smart devices and latency-sensitive applications has intensified demands on edge computing systems, exposing limitations in traditional cloud-based service provisioning. While container-based microservices offer flexibility and scalability, edge environments face three core challenges: dynamically fluctuating service request, imprecise caching control strategy, and concurrent service request bursts. To tackle these issues, we propose TwinOrch, a two-stage orchestration framework that couples predictive planning with real-time refinement. In the twin stage, TwinOrch predicts service request and generates initial dispatch, pre-caching, and scheduling strategies. In the real-time stage, it adaptively refines orchestration decisions and formulating scheduling as a constrained submodular maximization problem solved via greedy local search. This design achieves precise cache control, coordinated resource allocation, and low-latency service delivery under dynamic workloads. Extensive experiments on real-world traces demonstrate that TwinOrch improves service utility by 21.87% and reduces startup latency by 36.11% compared to baseline, validating the effectiveness of combining predictive intelligence with lightweight, scalable online optimization in edge service provisioning.

AI is making mobile applications increasingly cooler, but also introduces serious privacy risks due to the extensive user data collection. Federated learning (FL), as a privacy-preserving machine learning paradigm, enables mobile devices to collaboratively learn a shared prediction model while keeping all training data on devices. However, a key obstacle towards practical cross-device FL training is the huge energy consumption, especially for lightweight mobile devices. Prior literature mostly optimizes the convergence speed and network communication cost. In this work, we first perform the experimental analysis of improving FL performance through low-precision training with energy-friendly Digital Signal Processor (DSP) on mobile devices. Then, we demonstrate that directly integrating the state-of-the-art INT8 (8-bit integer) training algorithm and classic FL protocols will significantly degrade the model accuracy. Finally, we propose a novel FL protocol, namely Q-FedUpdate, incorporates two critical techniques: error-compensated aggregation and pipelined batch quantization. The former can ensure the tiny model updates be accumulated and take effects, and the latter can improve the DSP cache hit rate to reduce the context switching. Extensive experiments show that, Q-FedUpdate can effectively reduce the on-device energy consumption by 21 & times;, and accelerate the FL convergence by 6.1 & times; with only 2% accuracy loss.

Today's mobile applications oftentimes need to detect user-defined complex patterns (e.g., the mysterious "phantom traffic jam") over data streams to support decision making. It is achieved by continuously creating candidate instances that have partially matched a pattern, and meanwhile aggregating common instances (across patterns) for efficiency enhancement. Existing aggregation approaches are taken in a straightforward or intuitive manner, incurring an exponential solution space and thus having to be executed offline. This paper explores how to significantly accelerate aggregation so as to make pattern detection online executable, even suited to the emerging serverless runtime that involves complicated state synchronizations among distributed cloud functions. By comprehensively investigating a wide variety of mobile data streams, we note the existence of a latent hierarchical cluster structure among complex patterns (in terms of their instance similarities), which can be utilized to quickly aggregate common instances without going through the exponential solution space. To extract the latent information, we devise a content-aware structural entropy minimization algorithm to properly determine intra-cluster patterns, together with a lightweight differential compensation mechanism to maintain those inter-cluster "residual" relations among patterns. Evaluations on real-world vehicle and sensor network data streams illustrate that the resulting approach, dubbed PatternInsight, saves the aggregation time by 10 & times; to 50 & times; and reduces the instance size by 40%.

The integration of embodied intelligence into physical environments marks a new frontier in the evolution of intelligent systems.While the Internet of Things(IoT)connects devices and Artificial Intelligence of Things(AIoT)embeds intelligence into them,we argue that a further conceptual leap is required-one that enables the composition of intelligence itself through real-world embodiment,interaction,and evolution.We introduce the paradigm of Embodied Intelligence of Things(EIoT)as a foundational framework for distributed,physically grounded intelligent systems.EIoT systems are structured across three essential dimensions:Enacted,where devices are transformed into embodied agents;Engaged,where agents interact opportunistically based on physical and contextual constraints;and Evolutionary,where intelligence adapts and self-organizes through continuous experience.We further outline a developmental trajectory for EIoT based on the openness of perception and decision spaces,providing a conceptual map from tightly constrained agents to fully autonomous and adaptive systems.This work aims to establish EIoT as a core architectural and theoretical direction for future embodied intelligent systems.

Large language models (LLMs) have emerged as a transformative development in artificial intelligence (AI). To fully utilize their potential in mobile edge networks, Docker containers, serving as a lightweight, portable, and isolated framework, enable the efficient deployment of LLM-driven applications. In practice, containerized AI services are deployed using layer-structured container images, and their deployment efficiency is enhanced by the loading and scheduling of image layers at runtime. However, the deployment of containerized AI services for mobile user requests encounters challenges such as erratic mobile requests, delayed image loading, and redundancy in AI services deployment. In this paper, we present MetaPipe, an innovative incremental deployment framework for containerized AI services in mobile edge networks. MetaPipe seeks to optimize startup times through a dynamic workflow that incorporates proactive layer prefetching and reinforcement layer re-scheduling. Before the arrival of mobile requests, proactive layer pre-fetching enables real-time network response and decreases service deployment time through a tensor completion algorithm with Bayesian pruning, which generates layer caching and pre-scheduling strategies. After the arrival of mobile requests, reinforcement layer re-scheduling mitigates pre-fetching inaccuracies by dynamically adjusting layer scheduling strategies using a greedy sequential insertion approximation algorithm to supplement layers missed during pre-fetching. Extensive experiments on real-world datasets show that MetaPipe significantly outperforms conventional methods, achieving a 96.16% reduction in initialization startup time and 85.58% reduction in cold startup time. These results demonstrate its effectiveness in enhancing the performance of AI services deployment within mobile edge networks.

LLMs often struggle with memory-constrained deployment on consumer-grade hardware due to their massive parameter sizes. While existing solutions such as model compression and offloading improve deployment feasibility, they often suffer from substantial accuracy degradation or severe throughput bottlenecks. Recent error compensation methods recover accuracy through auxiliary LoRA-style branches, and we observe that these branches are inherently amenable to offloading: they require substantial parameter storage but access only a small subset of compensation parameters during each inference step. Motivated by this opportunity, we propose HCInfer, a heterogeneous inference system that offloads residual compensation to the CPU while executing the compressed backbone on the GPU, and further introduces an asynchronous compensation pipeline and sensitivity-aware dynamic rank allocation to hide compensation overhead and maximize accuracy recovery. Experimental results show that HCInfer achieves a maximum accuracy improvement of 5.2% on downstream tasks compared to compression model and sustaining a maximum speedup of 10.4x compared to full-precision model.

As drone-based applications proliferate, paramount contactless sensing of airborne drones from the ground becomes indispensable. This work demonstrates concentrating on propeller rotational speed will substantially improve drone sensing performance and proposes an event-camera-based solution, . features two components: Count Every Rotation achieves accurate, real-time propeller speed estimation by mitigating ultra-high sensitivity of event cameras to environmental noise. Every Rotation Counts leverages these speeds to infer both internal and external drone dynamics. Extensive evaluations in real-world drone delivery scenarios show that achieves a sensing latency of 3ms and a rotational speed estimation error of merely 0.23%. Additionally, infers drone flight commands with 96.5% precision and improves drone tracking accuracy by over 22% when combined with other sensing modalities. Demo: https://eventpro25.github.io/EventPro/.

Proactive streaming video understanding requires Video-LLMs to decide when to respond as a video unfolds, a task where existing methods often fall short due to their implicit, query-agnostic modeling of visual evidence. We introduce Response-G1, a novel framework that establishes explicit, structured alignment between the accumulated video evidence and the query's expected response conditions via scene graphs. The framework operates in three fine-tuning-free stages: (1) online query-guided scene graph generation from streaming clips; (2) memory-based retrieval of the most semantically relevant historical scene graphs; and (3) retrieval-augmented trigger prompting for per-frame "silence/response" decisions.By grounding both evidence and conditions in a shared graph representation, Response-G1 achieves more interpretable and accurate response timing decisions. Experimental results on established benchmarks demonstrate the superiority of our method in both proactive and reactive tasks, validating the advantage of explicit scene graph modeling and retrieval in streaming video understanding.

Autoregressive and diffusion models represent two complementary generative paradigms. Autoregressive models excel at sequential planning and constraint composition, yet struggle with tasks that require explicit spatial or physical grounding. Diffusion models, in contrast, capture rich spatial structure through high-dimensional generation, but lack the stepwise logical control needed to satisfy complex, multi-stage constraints or to reliably identify and correct errors. We introduce Collaborative Thoughts, a unified collaborative framework that enables autoregressive and diffusion models to reason and generate jointly through a closed-loop interaction. In Collaborative Thoughts, autoregressive models perform structured planning and constraint management, diffusion models instantiate these constraints as intermediate visual thoughts, and a vision-based critic module evaluates whether the visual thoughts satisfy the intended structural and physical requirements. This feedback is then used to iteratively refine subsequent planning and generation steps, mitigating error propagation across modalities. Importantly, Collaborative Thoughts uses the same collaborative loop regardless of whether the task is autoregressive question answering or diffusion-based visual generation. Through representative examples, we illustrate how Collaborative Thoughts can improve the reliability of spatial reasoning and the controllability of generation.

Bluetooth Low Energy (BLE) is a prevalent technology in various applications due to its low power consumption and wide device compatibility. Despite its numerous advantages, the encryption methods of BLE often expose devices to potential attacks. To fortify security, we investigate the application of Physical-layer Key Generation (PKG), a promising technology that enables devices to generate a shared secret key from their shared physical environment. We propose a distinctive approach that capitalizes on the inherent characteristics of BLE to facilitate efficient PKG. We harness the constant tone extension within BLE protocols to extract comprehensive physical layer information and introduce an innovative method that employs Legendre polynomial quantization for PKG. This method facilitates the exchange of secret keys with a high key matching rate and a high key generation rate. The efficacy of our approach is validated through extensive experiments on a software-defined radio platform, underscoring its potential to enhance security in the rapidly expanding field of BLE applications.

The growing demand for deploying large language models (LLMs) in applications requiring low latency and strong privacy has brought increasing attention to edge inference. While distributing computation across edge devices helps alleviate resource bottlenecks, the sequential and layer-wise nature of autoregressive decoding still limits real-time deployment, leading to device underutilization, increased latency, and workload imbalance. Existing methods rely on heavyweight drafts and static policies, which cannot adapt to device heterogeneity and bandwidth variation, leading to limited parallelism and elevated latency. To bridge this gap, we propose EdgeSpec, a hierarchical speculative decoding framework at edges. Instead of relying on last-layer speculation, EdgeSpec enables intermediate transformer layers to emit draft tokens, which are validated and used to predict the next token in parallel across devices via attention masking. The system combines offline entropy profiling with an online Kalman filter to adaptively select confident trigger layers based on runtime load and input complexity. To further improve efficiency, EdgeSpec jointly optimizes model partitioning, entropy thresholds, and candidate sizes through a two-stage strategy that integrates offline planning with online adaptation. Experiments on LLaMA2 showed that EdgeSpec reduced inference latency by up to 17.1× while preserving output quality comparable to standard decoding.

Real-time 6-DoF drone pose tracking enables precise flight control and accurate drone landing. With the widespread availability of urban 3D maps, the Visual Positioning Service (VPS) has been adapted for drone landing to overcome GPS unreliability, serving as a global localization system that yields absolute global coordinates by continuously aligning visual features with the prior 3D map. However, deploying conventional vision-based VPS on highly dynamic drones faces bottlenecks in both pose estimation accuracy and efficiency. In this work, we pioneer EV-Pose, an event camera-enhanced VPS designed to deliver drift-free, absolute global coordinates at ultra-high frequencies for drones. EV-Pose addresses the 2D-3D modality gap via a novel Spatio-Temporal Feature-instructed Pose Estimation module, which extracts a Temporal Distance Field (TDF) to enable continuous, differentiable matching with a prior 3D point map for pose estimation. To fully exploit this formulation, we propose a Motion-aware Hierarchical Fusion and Optimization scheme. This architecture utilizes onboard IMU motion information for fine-grained early-stage event filtering and seamlessly optimizes pose estimation in a later-stage factor graph. Evaluation shows that EV-Pose achieves a rotation accuracy of $1.76^{\circ}$ and a translation accuracy of 7.5$mm$ with a latency of 10.08$ms$, outperforming baselines by $\gt $40% and enabling accurate drone landings. Demo: https://ev-pose.github.io/.

For precise, efficient, and safe drone landings, ground platforms should real-time, accurately locate descending drones and guide them to designated spots. While mmWave sensing combined with cameras improves localization accuracy, lower sampling frequency of traditional frame cameras compared to mmWave radar creates bottlenecks in system throughput. In this work, we upgrade traditional frame camera with event camera, a novel sensor that harmonizes in sampling frequency with mmWave radar within ground platform setup, and introduce mmE-Loc, a high-precision, low-latency ground localization system designed for precise drone landings. To fully exploit the temporal consistency and spatial complementarity between these two modalities, we propose two innovative modules: (i) the Consistency-instructed Collaborative Tracking module, which further leverages the drone's physical knowledge of periodic micro-motions and structure for accurate measurements extraction, and (ii) the Graph-informed Adaptive Joint Optimization module, which integrates drone motion information for efficient sensor fusion and drone localization. Extensive experiments (30+ hours) demonstrate that mmE-Loc attains 0.083$m$ localization accuracy and 5.12$ms$ end-to-end latency, outperforming four state-of-the-art methods by over 48% and 62%, respectively.

Video analytics services utilizing multiple deep neural network models (DNNs) are increasingly being adopted in various edge intelligence applications. To avoid the accuracy reduction caused by data drift, existing works leverage continual learning that directly retrains DNN models in multi-model applications. According to our preliminary experiments, we observe that existing continual learning frameworks neglect the accuracy correlation among multiple DNN models and the opportunity to harvest idle resources during model retraining, leading to considerable resource inefficiency. In this paper, we propose FaaSLearner, an algorithm-system co-designed continual learning framework for multi-model video analytics that leverages the agile and fine-grained resource management of serverless computing. Specifically, FaaSLearner proposes the correlation-aware retraining planning to analyze the multi-model correlation, and then selectively retrain DNN models with the greatest in accuracy gain. In addition, FaaSLearner proposes the resource-efficient retraining scheduling to accurately trigger the continual learning, and then harvest the keep-alive periods for retraining tasks without interfering with normal inference. We evaluate FaaSLearner with four common edge video analytics applications with the Azure public dataset. Extensive experiments show that, FaaSLearner improves the average application accuracy up to 31.9% over representative baselines, and harvests more than 2.51× idle memory resources for retraining.

Emerging heavy-load mobile apps like UHD video and AR/VR access diverse high-throughput hardware devices, e.g., video codecs and cameras. However, today's mobile emulators exhibit poor performance when emulating these devices. We pinpoint the major reason to be the discrepancy between the guest's (system-on-chip) and host's (PC or cloud server) memory architectures for these devices, which makes the shared virtual memory (SVM) architecture of mobile emulators highly inefficient. To address this, we introduce vSoC, the first virtual mobile SoC featuring a unified SVM framework that enables efficient and secure data sharing among virtual devices, as well as an intelligent prefetch engine that effectively eliminates the vast majority of coherence maintenance overhead. While vSoC addresses runtime performance issues, app developers face a severe debugging challenge due to the inability of traditional tools to capture complete system states. Therefore, we devise an adaptive VM (virtual machine) snapshot-based approach that dynamically selects the optimal resource loading strategy to make vSoC debug-friendly. Compared to state-of-the-art emulators, vSoC brings 1.8-9.0 & times; frame rates, 35%-62% lower motion-to-photon latency, and 6.5-14.4 & times; bug reproduction rates for heavy-load apps. It is applicable to a variety of scenarios like end-user high-performance emulation and cloud/web-based rendering.

The history of information infrastructure can be read as a history of connectivity. Transmission Control Protocol/Internet Protocol (TCP/IP) made heterogeneous hosts reachable and the World Wide Web made documents linkable. A new connectivity problem is now emerging: The object to be connected is a capability-bearing agent that can perceive, reason, invoke tools, or act in the physical world. However, these capabilities remain locked inside platforms, organizations, and runtimes, while existing frameworks, interface protocols, and agent-interconnection efforts still leave open the question of what stable public abstraction can connect agents, humans, and infrastructure as a network architecture problem across organizational and physical domains. This vision and position paper argues for agent network: A reference architecture for capability discovery, connection, and coordination among heterogeneous, autonomous, and potentially self-interested agents. We do not present a complete protocol specification or a performance evaluation. Its central position is that open agent networks need a dual narrow-waist architecture. The Agent Locator Protocol (ALP), surfaced through agent://, provides the connectivity waist for making capability-bearing participants globally locatable, reachable, and minimally deliverable across substrates and organizations. The Task Semantic Intermediate Representation (TSIR) protocol provides the semantic waist for circulating task intent and success criteria as signable and reusable objects. Around two waists, we propose the Agent Shared Cognition Protocol (ASCP) as a candidate coordination runtime for adjustable group cognition, with blackboard-like workspaces and sedimentation as reference mechanisms for shared working state and experience reuse. Together, these abstractions define Networked Intent Realization (NIR): A research agenda for preserving high-level intent, routing it to capable participants, and translating it into executable, traceable, and accountable collaboration across open agent networks.

A heterogeneous micro aerial vehicles (MAV) swarm consists of resource-intensive but expensive advanced MAVs (AMAVs) and resource-limited but cost-effective basic MAVs (BMAVs), offering opportunities in diverse fields. Accurate and real-time localization is crucial for MAV swarms, but current practices lack a low-cost, high-precision, and real-time solution, especially for lightweight BMAVs. We find an opportunity to accomplish the task by transforming AMAVs into mobile localization infrastructures for BMAVs. However, translating this insight into a practical system is challenging due to issues in estimating locations with diverse and unknown localization errors of BMAVs, and allocating resources of AMAVs considering interconnected influential factors. This work introduces TransformLoc, a new framework that transforms AMAVs into mobile localization infrastructures, specifically designed for low-cost and resource-constrained BMAVs. We design an error-aware joint location estimation model to perform intermittent joint estimation for BMAVs and introduce a similarity-instructed adaptive grouping-scheduling strategy to allocate resources of AMAVs dynamically. TransformLoc achieves a collaborative, adaptive, and cost-effective localization system suitable for large-scale heterogeneous MAV swarms. We implement and validate TransformLoc on industrial drones. Results show it outperforms all baselines by up to 68% in localization performance, improving navigation success rates by 60%. Extensive robustness and ablation experiments further highlight superiority of its design.