The paradigm of edge computing is pivotal for deploying deep learning object detectors in time-sensitive applications. Nevertheless, practical efficacy is often impeded by critical impediments: 1) the inherent trade-off between detection precision and model lightweightness; 2) the inflexibility of generalized deployment frameworks for task-specific object detection; and 3) the scarcity of validation in real world operational environments. To address these challenges, we propose the Edge Detection Toolbox (ED-TOOLBOX), which leverages generalizable plug-and-play components to enable edge-site adaptation of object detection models. Specifically, we propose a lightweight Reparameterized Dynamic Convolutional Network (Rep-DConvNet) that employs a weighted multi-shape convolutional branch structure to enhance detection performance. Furthermore, ED-TOOLBOX includes a Sparse Cross-Attention (SC-A) network that adopts a localized-mapping-assisted self-attention mechanism to facilitate a well-crafted Joint Module in adaptively transferring features for further performance improvement. Moreover, we propose an Efficient Head for the classification and location modules to achieve more efficient prediction. Furthermore, we address a critical oversight in industrial safety: conventional helmet detection's neglect of band fastening. To bridge this gap, we construct the Helmet Band Detection Dataset (HBDD) and deploy our ED-TOOLBOX-optimized model on this practical challenge. Extensive experiments validate the efficacy of our components. In surveillance simulations, our model surpasses six state-of-the-art methods, achieving both real-time performance and high accuracy. These results establish our approach as a superior solution for edge object detection.

Autonomous driving is a cornerstone of modern Intelligent Transportation Systems (ITS), relying on Vehicle to-Everything (V2X) communication to achieve collaborative sensing and decision-making. However, the heterogeneity of V2X services (e.g., high-bandwidth V2N vs. low-latency V2V) creates severe contention for limited spectrum resources. Existing static or heuristic allocation methods struggle to adapt to the highly dynamic and stochastic vehicular environment. To address this, we propose a Constraint-aware Knowledge Collaboration framework driven by Deep Reinforcement Learning (DRL). Specifically, we first model the dynamic bandwidth allocation problem using a Hidden Markov Model (HMM) to effectively capture the latent stochasticity of channel conditions and service demands. To mitigate the data inefficiency of varying autonomous vehicles (AVs), we introduce a public knowledge base mechanism that facilitates privacy-preserving knowledge sharing, thereby accelerating policy convergence. Furthermore, we propose an Action Space Constrained Actor-Critic (ASC-AC) algorithm. By redefining collaborative knowledge as gain rules, ASC-AC lever ages global steady-state information to prune the action space, thereby ensuring strict quality-of-service (QoS) compliance while exploring optimal strategies. Simulation results demonstrate that ASC-AC achieves rapid convergence within 100 episodes. Compared to state-of-the-art baselines (e.g., SAC and DDPG), our approach improves throughput by 20%–25%, reduces latency by 48%–63%, and accelerates training convergence by 47%, confirming its superiority in stability and efficiency.

Reinforcement Fine-tuning (RFT) methods such as Group Relative Policy Optimization (GRPO) have demonstrated strong capabilities in aligning Large Language Models with human preferences. However, these approaches often suffer from limited data efficiency, necessitating extensive on-policy rollouts to maintain competitive performance. We propose PSPO (Prompt-Level Prioritization and Experience-Weighted Smoothing for Efficient Policy Optimization), a lightweight yet effective enhancement to GRPO that improves training stability and sample efficiency through two complementary techniques. First, we introduce an experience-weighted reward smoothing mechanism, which uses exponential moving averages to track group-level reward statistics for each prompt. This enables more stable advantage estimation across training steps without storing entire trajectories, allowing the model to capture historical reward trends in a lightweight and memory-efficient manner. Second, we adopt a prompt-level prioritized sampling strategy, which is an online data selection method inspired by prioritized experience replay. It dynamically emphasizes higher-impact prompts based on their relative advantages, thereby improving data efficiency. Experiments on multiple mathematical reasoning benchmarks and models show that PSPO achieves comparable or better accuracy than GRPO, while significantly accelerating convergence, and maintaining low computational and memory overhead.

With the rapid proliferation of intelligent video devices, adaptive and robust bitrate control has become crucial for ensuring stable video transmission in complex networked systems. This paper presents a Lifelong Intelligence Learning (LIL) framework that continuously evolves its decision-making capability through lifelong reinforcement learning. Inspired by human dual-process cognition, LIL integrates fast–slow thinking mechanisms to achieve both rapid adaptation and deep reasoning when facing heterogeneous video tasks. An intelligence gain metric is designed to quantify cognitive gaps for dynamic task adaptation, while an anti-forgetting distillation loss maintains long-term performance stability across sequential learning processes. Extensive experiments in both video-on-demand and live streaming scenarios demonstrate that the proposed method consistently improves bitrate, latency, and overall Quality of Experience (QoE) compared with existing state-of-the-art algorithms. The results verify that LIL serves as a practical and generalizable Artificial Intelligence (AI)-based engineering solution for adaptive video transmission in dynamic network environments.

The large-scale expansion of the Internet of Things (IoT), coupled with its integration into application scenarios like smart transportation, remote healthcare, and intelligent agriculture, has led to surging demands for enhanced computational capabilities to process massive real-time data. These demands pose significant challenges to meet the quality of service (QoS) requirements. In order to address these issues, this paper proposes a cloud-edge-end collaborative computing power network (CPN) architecture based on Terahertz (THz) communication and assisted by reconfigurable intelligent surface unmanned aerial vehicle (RIS-UAV). Here, CPN enables flexible resource collaboration for task offloading; THz communication supports large-volume data transmission, while RIS-UAV mitigates THz's limitations of small coverage and blockage sensitivity. To unify control of the computation resource and transmission resource allocation, we introduce the Resource-Allocation Multi-Agent Soft Actor-Critic (RA-MASAC) approach, which is a multi-agent reinforcement learning-based algorithm for joint optimization of resource allocation. Simulation results demonstrate that the proposed method outperforms the existing baselines significantly.

In 6G low-altitude edge intelligent networks, the proliferation of delay-sensitive Internet of Things (IoT) services exacerbates mutual interference among IoT devices, thereby reducing the efficiency of resource allocation. To address this challenge, we propose a blockchain-assisted federated learning (FL)-based four-layer low-altitude framework, where the user equipment (UE) layer generates delay-sensitive tasks, the task UAV (TUAV) layer offers extra computation resources to process UE tasks, the service UAV (SUAV) layer trains the FL local models in a distributed manner, and the base station layer aggregates them to obtain the global model, where the FL models are trained for the joint optimisation including resource allocation, offloading decisions, and caching strategy. Based on this framework, we define a system utility cost that incorporates the total service delay (including the transmission, computation, queueing, and fetching delays) and the total energy consumption, and formulate an optimisation problem to minimise the long-term system utility cost. To solve the formulated problem, we propose an FL and multi-agent proximal policy optimisation (MAPPO) integrated offloading decisions, caching strategy, and resource allocation. Simulation results show that our proposed algorithm converges faster and achieves a lower system utility cost than the benchmarks.

Autonomous navigation of mobile robots in urban environments is crucial for independent travel for the elderly or disabled populations. However, existing robot urban navigation solutions usually rely on the precise coordinates of the target location, and do not comply with traffic rules during navigation. To this end, this paper proposes a demand-oriented robot navigation solution in urban environments based on vector-semantic map (V-S Map), where V-S Map integrates geometrically precise vector layer generated using a Transformer architecture, and task-related semantic layer derived via optical character recognition and point cloud registration. Unlike current mapping approaches, V-S map uniquely incorporates directional road attributes and task-relevant semantic information to address traffic violations in urban navigation while establishing the relevance of tasks to the environment. Furthermore, we develop an LLM-based command parsing strategy that combines optimized prompt engineering with regular expression matching against location lexicons extracted from semantic layer. This strategy eliminates traditional coordinate-input constraints, enhancing human-robot natural language interaction capabilities. Extensive experiments in simulation and real-world environments verify the effectiveness and feasibility of the proposed solution in terms of ambiguous command parsing and robot navigation performance, achieving an 80% overall task success rate with a 63.78% path-violation weighted success rate.

Vision-Language-Action (VLA) policies promise general robotic manipulation, but their robustness against physical-world attacks remains fragile. In particular, we show that physically realizable adversarial patches can reliably induce failures by triggering a mechanism we call policy-critical action-to-vision attention hijacking, where action-conditioned attention is diverted from task-relevant regions to a localized patch. To demonstrate the threat, we propose Attention-Guided Semantic Disruption (AGSD), an Expectation-over-Transformation (EOT) optimized printable patch that jointly (i) concentrates action-to-vision attention on the patch and (ii) disrupts vision-language semantic alignment, yielding strong cross-task and cross-architecture transfer. To mitigate such attacks, we introduce Structure-Aware Robust Fine-Tuning (SARF), a zero-inference-overhead defense that fine-tunes only the visual encoder using feature anchoring, policy-critical attention correction, and language-guided geometric consistency restricted to semantically relevant regions. On LIBERO, SARF reduces OpenVLA's failure rate under AGSD from 100

In large-scale Industrial Internet of Things (IIoT) scenarios such as smart ports and open-pit mines, wide-area visual perception is essential for intelligent monitoring, and it may serve as a perception basis for future digital twin applications. Bionic compound-eye cameras provide an ultra-wide field of view and hundred-megapixel resolution, enabling single-node panoramic sensing in Wireless Visual Sensor Networks (WVSNs). However, compound-eye imagery also introduces two major difficulties: heavy transmission overhead and weak perception of tiny targets caused by extreme object-scale variation. To address these issues, this paper proposes an edge-cloud collaborative Internet of Video Things (IoVT) perception framework. In this framework, high-resolution visual streams are processed directly at edge nodes, and only high-value semantic information is transmitted to the cloud. As the core edge inference engine, we develop DPGL-YOLO, a lightweight detector tailored to compound-eye imaging characteristics and edge deployment constraints. Specifically, the proposed method introduces a Dual-Path Attention Fusion Module (DPAM) to enhance cross-scale feature interaction, a Global–Local Collaborative Attention Mechanism (GLAM) to enhance global-local discrimination under cluttered backgrounds, a high-resolution P2 detection head, and an MIF-WiseIoU regression loss to improve localization of dense small objects under weak-feature and cluttered-background conditions. Experiments on two in-house hundred-megapixel compound-eye datasets collected from real port and mining environments show that DPGL-YOLO achieves 86.84% mAP@0.5 and 60.77% mAP@0.5:0.95 on the port dataset, and 72.30% mAP@0.5 and 30.68% mAP@0.5:0.95 on the mining dataset. The proposed method outperforms representative detectors while maintaining edge-oriented inference efficiency. The proposed framework provides an edge-oriented perception architecture that may help reduce transmission burden in industrial IoVT systems.

Diffusion strategies have advanced visual motor control by progressively denoising high-dimensional action sequences, providing a promising method for robot manipulation. However, as task complexity increases, the success rate of existing baseline models decreases considerably. Analysis indicates that current diffusion strategies are confronted with two limitations. First, these strategies only rely on short-term observations as conditions. Second, the training objective remains limited to a single denoising loss, which leads to error accumulation and causes grasping deviations. To address these limitations, this paper proposes Foresight-Conditioned Diffusion (ForeDiffusion), by injecting the predicted future view representation into the diffusion process. As a result, the policy is guided to be forward-looking, enabling it to correct trajectory deviations. Following this design, ForeDiffusion employs a dual loss mechanism, combining the traditional denoising loss and the consistency loss of future observations, to achieve the unified optimization. Extensive evaluation on the Adroit suite and the MetaWorld benchmark demonstrates that ForeDiffusion achieves an average success rate of 80% for the overall task, significantly outperforming the existing mainstream diffusion methods by approximately 20% in high difficulty tasks, while maintaining more stable performance across the entire tasks.

In this paper, we investigate the power control and spectrum allocation challenge in vehicle-to-everything (V2X) networks with uncertain channels, where two types of services coexist, i.e., the large-capacity services supported by vehicle-to-infrastructure (V2I) links and the ultra-reliable services supported by vehicle-to-vehicle (V2V) links. Our goal is to maximize the transmission rate of V2I whilst ensuring the V2V outage probability constraint. To deal with the channel uncertainty, we first mine the channel correlation by mapping historical channel samples into high-dimensional space via a designed deep neural network (DNN). Then, a support vector clustering (SVC) based uncertainty set is derived as a union of convex subsets according to the output of piecewise activation DNN. Afterwards, a power control algorithm is proposed for the transformed non-convex problem. Specifically, a bisection-based method is developed to tackle the fractional signal-to-interference-plus-noise ratio (SINR) expression, and an exploration-comparison-approximation method is proposed to relax the union of convex subsets. Additionally, we design a set of virtual V2V links to enforce the spectrum allocation problem into a feasible bipartite graph matching problem. Simulation results demonstrate the proposed method overcomes the conservatism of previous uncertainty set construction methods, and outperforms other methods on V2I transmission rate.

The emergence of Sixth Generation (6G) mobile communication technologies, along with the burgeoning proliferation of Internet of Things (IoT) devices, has significantly increased the requirements for computing resources. This growing demand places considerable pressure on the computing capacity. To address diverse and intensive application requirements, the integration of satellite and terrestrial networks is a key focus in the 6G vision. Within this context, the combination of a geostationary orbit (GEO) satellite and a low Earth orbit (LEO) satellite constellation is a crucial facilitator of IoT applications, addresses their computational demands while offering global coverage, low latency, seamless scalability, and strong reliability. However, several challenges arise in such networks: 1) LEO satellites are battery-powered and resource-constrained; 2) the number of LEO computing nodes is large and distributed; and 3) the network environment is highly dynamic and complex. To address these issues, a collaborative architecture for an LEO satellite constellation and a GEO satellite (CALG) is proposed. Based on the CALG, an energy-balancing strategy is designed to identify the optimal orbital plane for task offloading as well as an efficient routing path among satellites. The task offloading and routing selection are modeled by a Markov Decision Process, with a Proximal Policy Optimization (PPO) algorithm from deep reinforcement learning is applied to optimize decision-making in a dynamic environment. Experimental results demonstrate that the energy-balancing strategy based on PPO meets system performance requirements and outperforms baseline strategies.

Band selection (BS) is a key method in hyperspectral image (HSI) classification that helps to reduce the computational burden and improve the class separability. However, with the emerging of uncrewed aerial vehicle (UAV)-borne HSI datasets, their attributes, such as high spatial and spectral resolution as well as large-scale samples pose serious challenges to the existing BS methods, making them inefficient. In addition, the efficient utilization of the prior knowledge from the data collected by fixed UAV-borne sensors in different regions is often easily overlooked. In view of these issues, this article proposes a neural network-assisted particle swarm optimization (PSO) algorithm for cross-domain BS of UAV-borne HSIs. First, a knowledge learning strategy is designed for the source domain, which applies a neural network model to learn the useful prior knowledge in labeled source domain data. Then, a network-assisted PSO algorithm is introduced to search for the optimal subset of bands in the target domain under the guidance of the valid prior knowledge captured from the source domain by the network model. Moreover, a similarity-based grouping strategy is designed to group similar bands and then select bands from each group with the aims of reducing the redundant information in the subset of bands. Finally, experimental results on three common UAV-borne HSI datasets show that our proposed method can efficiently handle UAV-borne HSI data with large samples, as it is able to find a subset of bands with higher quality compared to several state-of-the-art BS methods.

Augmented reality (AR)-enabled Metaverse is a promising technique to provide immersive service experience for mobile users. However, the limited network resources and unpredictable wireless propagation environments are key design bottlenecks of AR-enabled Metaverse systems. Therefore, this paper presents a resource management framework for simultaneously transmitting and reflecting RIS (STAR-RIS)-assisted AR-enabled Metaverse, where the STAR-RIS is configured to improve the communication efficiency between AR users and the Metaverse server located at the base station (BS). Moreover, we formulate a service latency minimization problem via jointly optimizing the computation resource allocation of the BS, coefficient matrix of the STAR-RIS, central processing unit (CPU) frequency and transmit power of the AR users. To tackle the non-convex problem, we utilize an approximate method to transform it to a tractable form, and decouple the multi-dimensional variables via the alternating optimization method. Particularly, the optimal coefficient matrix is obtained by a penalty function-based method with proved convergence, the CPU frequencies of AR users are derived as the closed-form solution, and the transmit power of AR users and computation resource allocation of the BS are obtained by the Lagrange duality method and convex optimization theory. Finally, simulation results demonstrates that the proposed method achieves remarkable latency reduction than several benchmark methods.

This paper presents a novel vehicle-to-infrastructure (V2I) cooperative perception framework to address inherent limitations of bird's eye view (BEV) systems in autonomous driving. Using sensor constraints that include occlusion and restricted field of view, our framework integrates sensor systems mounted on the infrastructure with perception of the ego vehicle through three innovations. First, a geometry-aware feature alignment module that resolving cross-view discrepancies via projective geometry transformations. Second, attention-optimized fusion architecture (AFFENet) with dual-attention mechanisms for channel recalibration and spatial-contextual enhancement. Third, a multiscale dynamic aggregation protocol enabling context-aware fusion of heterogeneous sensor data. Evaluated on DAIR-V2X dataset, the framework achieves an improvement of 14.03% mAP over baselines from BEV while reducing unobserved areas coverage by 30.46% with real-time efficiency. This work establishes a new paradigm for cooperative perception systems, providing theoretical foundations and practical implementations for multiperspective sensor fusion in intelligent transportation ecosystems. The proposed methodologies address critical challenges in geometric alignment in cross-view and adaptive feature fusion, ultimately advancing robust autonomous driving systems through infrastructure-vehicle perception synergy.

Emerging network paradigms and applications increasingly rely on federated learning (FL) to enable collaborative intelligence while preserving privacy. However, the sustainability of such collaborative environments hinges on a fair and stable payoff allocation mechanism. Focusing on coalition stability, this paper introduces a payoff allocation framework based on the least core (LC) concept. Unlike traditional methods, the LC prioritizes the cohesion of the federation by minimizing the maximum dissatisfaction among all potential subgroups, ensuring that no participant has an incentive to break away. To adapt this game-theoretic concept to practical, large-scale networks, we propose a streamlined implementation with a stack-based pruning algorithm, effectively balancing computational efficiency with allocation precision. Case studies in federated intrusion detection demonstrate that our mechanism correctly identifies pivotal contributors and strategic alliances. The results confirm that the practical LC framework promotes stable collaboration and fosters a sustainable FL ecosystem.

Volumetric video offers a fantastic six degrees of freedom (6DoF) viewing experience, allowing users to perceive various views from different viewpoints. However, the direct streaming of volumetric videos incurs extremely high bandwidth costs due to sophisticated volumetric content. Current transcode streaming methods, while reducing bandwidth requirements by 2D views of volumetric video, fail to match the user’s actual view precisely. In this work, we propose an Image-based Rendering (IBR)-assisted streaming method that synthesizes volumetric video views with limited computational resources, aiming to maximize user quality of experience (QoE). To achieve this goal, we first model the IBR view synthesis quality and IBR computational cost to capture the tradeoff between its benefit and overhead. Based on these models, we further formulate the IBR-assisted video streaming problem and design a heuristic algorithm to solve it efficiently. Finally, extensive evaluations demonstrate the superiority of our proposed approach compared to the baselines in adaptive volumetric video streaming.

With the development of HTTP-based video streaming, Adaptive Bitrate (ABR) algorithms have become crucial for optimizing video quality. These algorithms dynamically select the bitrate of video chunks based on factors such as network throughput and playback buffer occupancy. However, the volatility of network throughput, conflicting Quality of Experience (QoE) objectives, and cascading effects in decision-making pose significant challenges for ABR algorithms to accurately determine bitrate selections, leading to substantial revenue losses for content providers. This paper proposes a bandwidth prediction-based ABR algorithm for video streaming, termed the BPA algorithm, which consists of two components: a Bandwidth Prediction Model (BPM) and a Bitrate Selection Model (BSM). The BPM leverages a Bidirectional Long Short-Term Memory (BiLSTM) network for bandwidth prediction, while the BSM adopts an Actor-Critic reinforcement learning framework. A reward function based on bandwidth prediction accuracy is proposed, and an end-to-end joint optimization loss function is designed to train the model for optimal video bitrate selection. Under various network conditions, the BPA algorithm outperforms existing baseline algorithms, achieving an improvement of nearly 31.9% compared to traditional heuristic methods and a 9% enhancement over other deep reinforcement learning-based approaches. The BPA algorithm demonstrates excellent performance in terms of bitrate smoothness and QoE.

Train Autonomous Circumambulate System (TACS) represents the next generation of train control systems. Characterized by autonomous travel path planning and train operation adjustments, TACS has heightened demands for real-time communication. Given the high mobility of the system and the surge in train density, handoffs in TACS occur more frequently. An inappropriate handoff decision can lead to extended handoff delays, significantly impacting TACS performance. Current learning-based handoff decision algorithms rely heavily on extensive trial-and-error data from environmental interactions, and it is a challenging feat in urban rail environments. In this paper, we introduce a rolling prediction-decision framework designed to enhance the probability of successful handoffs. We propose STADNet-CP (Spatial-Temporal Attention Deep Network for Channel Prediction), a Deep Learning (DL) based network model enhanced with spatial-temporal attention mechanism to extract the spatial-temporal dependencies in the channel variations and improve the accuracy of channel states prediction. We then formulate a handoff decision optimization problem and adopt the Model Predictive Control (MPC) algorithm to enhance the accuracy and reliability of handoffs utilizing the channel state predictions. Comprehensive experimental results demonstrate that our proposed handoff decision method significantly improves the overall handoff performance.

The integrated geostationary Earth orbit (GEO)-multibeam and low Earth orbit (LEO)-uncrewed aerial vehicle (UAV) network has emerged as a promising paradigm to enhance the coverage and capacity of terrestrial networks. However, due to the high dynamics of the network, handover becomes a critical yet challenging problem. In this paper, we model the handover problem as a decentralized Markov decision process (DEC-MDP), aiming to maximize throughput while minimizing energy consumption and handover cost. A multi-agent dueling double deep Q-network (MAD3QN) algorithm is then designed to solve it, where the centralized training with decentralized execution (CTDE) method is utilized to share intelligent information, and an offline training approach is adopted to support practical deployment. Simulation experiments with real-world satellite deployments demonstrate that MAD3QN effectively converges and significantly outperforms the existing baselines.