Collaborative perception is a promising approach to enhance the robustness and accuracy of 3D object detection for autonomous driving. However, typical methods relying on early fusion of raw data demand prohibitive communication bandwidth, while those using intermediate fusion of compressed features often suffer from significant information loss and detection performance degradation. In this regard, this paper proposes a communication-efficient collaborative perception framework based on a knowledge distillation paradigm. The key idea is to leverage a powerful early-fusion model as a teacher to guide the training of a communication-efficient intermediate-fusion student model. Specifically, the teacher model incorporates a spatial-channel attention module to generate enriched and discriminative feature representations. The student model employs a lightweight yet effective feature compression scheme, combining learnable quantization with entropy coding, to efficiently reduce the volume of transmitted data. Crucially, we propose an adaptive knowledge distillation loss function that selectively transfers knowledge related to potential objects, thereby suppressing background noise and aligning the student’s focus with critical detection regions. Extensive experiments on the OPV2V, V2X-Sim 2.0, and DAIR-V2X datasets demonstrate that our method achieves competitive detection accuracy. Additional on-road tests in campus environments validate the system’s practicality and robustness.

Cooperative perception (CP) is promising to enhance environmental awareness by sharing complementary information among connected vehicles. However, the challenges of resource constraints and edge heterogeneity in vehicular networks often lead to synchronization bottlenecks and degrade CP accuracy under strict delay requirements. This paper presents RT-Cooper, a real-time CP framework that jointly orchestrates computation and communication to optimize perception accuracy with delay guarantees in heterogeneous vehicular networks. We first introduce adaptive model pruning and bandwidth allocation into the CP pipeline to enable fine-grained resource coordination. Then, we formulate the Real-time Cooperative Perception (RCP) problem, which captures coupled computation-communication delays and system-wide synchronization bottlenecks, aiming to maximize perception accuracy under hard deadlines and resource constraints. To handle the closed-box and non-convex nature of the RCP objective, we construct an interpretable feature quality model derived from pruning analysis as a surrogate for accuracy. On this basis, we develop the Two-stage Alternating Resource Optimization (TARO) algorithm, which alternates between closed-form feasibility enforcement and heterogeneity-aware feature quality refinement with provable convergence. Extensive experiments on open-source benchmark and a real-world prototype validate the effectiveness of RT-Cooper in improving perception accuracy while achieving real-time guarantees.

Indoor location-based services (LBS) have witnessed rapid growth in applications such as user tracking, healthcare monitoring, and smart facility management, driving the critical need for efficient and pervasive indoor localization. Traditional WiFi fingerprinting methods face significant challenges: multi-site localization (MSL) relies on densely deployed static WiFi sites, incurring high infrastructure costs and conflicting with the Integrated Sensing and Communication (ISAC) paradigm; single-site localization (SSL) requires complex hardware; and single mobile site localization (SMSL) suffers from poor real-time performance due to long traversal paths. To address these limitations, this paper proposes a Multi-Agent Deep Reinforcement Learning-based Collaborative Indoor Localization (MADRL-CIL) framework. MADRL-CIL leverages multiple collaborative mobile sites to dynamically acquire Received Signal Strength (RSS) fingerprints. By modeling each mobile site as an agent, the framework formulates the path selection and fingerprint acquisition task as a Multi-Agent Deep Reinforcement Learning (MADRL) problem under a Centralized Training with Decentralized Execution (CTDE) paradigm, facilitating effective collaboration among multiple mobile sites to optimize localization accuracy while minimizing localization time. Additionally, a Multi-Site Fingerprint Matching (MS-FM) model is specifically designed to process collaboratively collected RSS fingerprints, enabling fine-grained localization accuracy. Experimental evaluations in a real-world indoor environment demonstrate that MADRL-CIL achieves localization accuracy comparable to dense multi-static site deployments while provides good real-time performance.

Semi-Supervised Semantic Segmentation (SSS) has emerged as an effective paradigm to reduce the reliance on dense pixel-wise annotations. However, existing pseudo-labeling approaches often suffer from instability and semantic inconsistency, particularly in complex consumer electronic environments where visual inputs are noisy and heterogeneous. To overcome these limitations, we propose a unified framework termed Reliability-Guided Consistency Learning (RGCL), which systematically enhances probabilistic stability and semantic reliability through temporally and spatially consistent supervision. At the core of RGCL lies the Variance-Guided Prior Augmentation (VGPA), which quantifies temporal prediction variance across training checkpoints to identify and weight reliable unlabeled samples, thereby constructing a stability-aware prior. Building upon this, the Dual-Stream Pseudo-Label Refinement (DSPL) module distills high-confidence pseudo-labels from the variance-filtered subset and jointly optimizes them with labeled data in a dual-stream manner, promoting boundary accuracy and structural coherence. To further regularize the predictive manifold, the Interpolation Consistency Regularization (ICR) enforces geometric smoothness by coupling interpolation in the input domain with consistency in the predictive space, ensuring local continuity and mitigating spurious confidence transitions. By jointly addressing temporal uncertainty, semantic reliability, and manifold smoothness, RGCL establishes a reliability-driven learning paradigm for semi-supervised segmentation. Extensive experiments on Cityscapes, Pascal VOC, and ADE20K benchmarks demonstrate that RGCL consistently achieves state-of-the-art performance under various label ratios, confirming its robustness and effectiveness for deployment in consumer electronic applications.

In the collaborative operations of low-altitude Unmanned Aerial Vehicle (UAV) swarms, the reliability of the perception-communication-control (PCC) closed-loop system is critical. However, existing simulation platforms often fail to balance networking fidelity with real-time computational efficiency, relying either on oversimplified distance-based models or computationally expensive discrete-event simulators. To address this challenge, this paper presents a Lightweight Communication Emulation Engine (LCEE), physically consistent network behavior modeling module integrated within a holistic swarm simulation framework. We propose a strict four-stage pipeline—comprising state acquisition, physical layer evaluation, Media Access Control (MAC) contention, and asynchronous delivery—to decouple physical motion from network performance dynamically. Specifically, the proposed model incorporates a low-altitude Rician fading channel and a bandwidth-adaptive contention mechanism, enabling precise link quality assessment for heterogeneous payloads ranging from telemetry to high-bandwidth video. Extensive simulations demonstrate that LCEE provides actionable connectivity metrics with negligible latency, ensuring synchronization with the real-time control loop while adhering to theoretical physical baselines.

Retrieval-augmented generation (RAG) substantially extends the knowledge boundary of large language models. However, it still faces two major challenges when handling complex reasoning tasks: low context utilization and frequent hallucinations. To address these issues, we propose Self-Correcting RAG, a unified framework that reformulates retrieval and generation as constrained optimization and path planning. On the input side, we move beyond traditional greedy retrieval and, for the first time, formalize context selection as a multi-dimensional multiple-choice knapsack problem (MMKP), thereby maximizing information density and removing redundancy under a strict token budget. On the output side, we introduce a natural language inference (NLI)-guided Monte Carlo Tree Search (MCTS) mechanism, which leverages test-time compute to dynamically explore reasoning trajectories and validate the faithfulness of generated answers. Experiments on six multi-hop question answering and fact-checking datasets demonstrate that our method significantly improves reasoning accuracy on complex queries while effectively reducing hallucinations, outperforming strong existing baselines.Our code is available at https://github.com/xjiacs/Self-Correcting-RAG .

Autonomous vehicles rely on cooperative perception to extend environmental awareness and enhance safety. However, contemporary methods typically select cooperative vehicles based solely on distance or link quality, without accounting for occlusions or redundant coverage, thereby incurring higher costs. In this paper, we propose an edge-assisted low-overhead cooperative perception system that achieves a predetermined coverage with a minimum number of vehicles. Specifically, we first introduce a lightweight LiDAR-based occlusion detection method to generate Perception Metadata (PMD), which accurately describes each vehicle’s coverage in terms of locations, orientations, and fields of view. On this basis, we formulate the vehicle selection problem as a Min-Selection problem and solve it using the proposed Q-learning-based Adaptive Vehicle Selection (QAVS) algorithm. The QAVS algorithm is able to select the vehicles with the minimum perception coverage redundancy depending on PMD. Finally, based on the KITTI dataset and the simulation dataset, our LiDAR-based occlusion detection method achieves a competitive recall rate compared to deep-learning methods, while significantly reducing latency. Meanwhile, QAVS selects fewer vehicles than greedy or random selection methods for equal coverage targets.

The increasing sophistication of cyberattacks, particularly multi-stage advanced persistent threats (APTs) and zero-day exploits, poses significant challenges for intrusion detection systems (IDS) in Industrial Internet of Things (IIoT) environments. While machine learning (ML) and deep learning (DL)-based IDS approaches enhance detection of known threats, their effectiveness is frequently limited under evolving attack patterns and distribution shifts.

Cooperative perception enhances autonomous driving safety by enabling vehicles to share complementary sensory information. However, existing methods employ fixed fusion strategies regardless of network conditions and often engage all available vehicles without considering information redundancy, thus leading to excessive communication overhead. To address these challenges, this paper proposes BDGraS, a bandwidth-adaptive dual-relation gravity model for efficient cooperative vehicle selection. Specifically, BDGraS first introduces a bandwidth-aware optimization framework that explicitly considers network constraints and information redundancy, while leveraging a novel dual-relation gravity model that effectively captures both vehicle-to-ego and vehicle-to-vehicle relationships for comprehensive cooperation optimization. Furthermore, an adaptive selection algorithm enables seamlessly transitions between fusion modes and intelligently selects the most valuable subset of cooperative vehicles. Finally, extensive experiments on the OPV2V dataset demonstrate that BDGraS achieves average reductions in communication overhead (25-30%) and latency (up to 25-35%) while maintaining competitive detection accuracy (<= 0.5% drop) across various fusion strategies and bandwidth conditions.

Accurate bird’s-eye-view (BEV) perception is essential for autonomous driving. Existing methods based on projection, lifting, or attention mechanisms face inherent limitations: projection-and attention-based approaches require strong geometric priors or incur high computational cost, while lifting-based methods are sensitive to depth errors, leading to feature misalignment. Such challenges become more severe under conditions of low illumination, heavy occlusion, or long-range perception. To address these issues, we propose DUG-VT, a Depth and Uncertainty Guided View Transformer that explicitly incorporates depth uncertainty into learned feature aggregation. Specifically, the Depth Uncertainty Modeling (DUM) module predicts per-pixel depth distributions, lifts them into 3D space and models spatial uncertainty. The Implicit Geometry-Semantic Manifold (IGSM) module encodes this uncertainty into attention keys for geometry-aware feature transformation, improving alignment between geometric and semantic features. Finally, the Uncertainty-Gated Aggregation (UGA) module adaptively suppresses unreliable features during cross-view fusion, mitigating the effects of ambiguous depth estimates. Results from extensive experiments on the nuScenes benchmark reveal that DUG-VT achieves 39.4

Creating high-fidelity, animatable 3D avatars from a single image remains a formidable challenge. We identified three desirable attributes of avatar generation: 1) the method should be feed-forward, 2) model a 360° full-head, and 3) should be animation-ready. However, current work addresses only two of the three points simultaneously. To address these limitations, we propose OMEGA-Avatar, the first feed-forward framework that simultaneously generates a generalizable, 360°-complete, and animatable 3D Gaussian head from a single image. Starting from a feed-forward and animatable framework, we address the 360° full-head avatar generation problem with two novel components. First, to overcome poor hair modeling in full-head avatar generation, we introduce a semantic-aware mesh deformation module that integrates multi-view normals to optimize a FLAME head with hair while preserving its topology structure. Second, to enable effective feed-forward decoding of full-head features, we propose a multi-view feature splatting module that constructs a shared canonical UV representation from features across multiple views through differentiable bilinear splatting, hierarchical UV mapping, and visibility-aware fusion. This approach preserves both global structural coherence and local high-frequency details across all viewpoints, ensuring 360° consistency without per-instance optimization. Extensive experiments demonstrate that OMEGA-Avatar achieves state-of-the-art performance, significantly outperforming existing baselines in 360° full-head completeness while robustly preserving identity across different viewpoints.

Despite the widespread adoption of Large Language Models (LLMs) in Legal AI, their utility for automated contract revision remains impeded by hallucinated safety and a lack of rigorous behavioral constraints. To address these limitations, we propose the Risk-Constrained Bilevel Stackelberg Framework (RCBSF), which formulates revision as a non-cooperative Stackelberg game. RCBSF establishes a hierarchical Leader Follower structure where a Global Prescriptive Agent (GPA) imposes risk budgets upon a follower system constituted by a Constrained Revision Agent (CRA) and a Local Verification Agent (LVA) to iteratively optimize output. We provide theoretical guarantees that this bilevel formulation converges to an equilibrium yielding strictly superior utility over unguided configurations. Empirical validation on a unified benchmark demonstrates that RCBSF achieves state-of-the-art performance, surpassing iterative baselines with an average Risk Resolution Rate (RRR) of 84.21% while enhancing token efficiency. Our code is available at https://github.com/xjiacs/RCBSF .

Bird’s Eye View (BEV) semantic segmentation is essential for autonomous driving and mobile robotics, yet it still faces significant challenges on accurate segmentation of foreground object and efficient estimating of layout categories obscured by objects. To address these issues, we propose BEV-CAR, a Context-Aware Rasterization method that rasterizes the BEV representation without any coordinate transformations. By optimising each ray and incorporating depth features, BEV-CAR effectively addresses the challenges posed by object occlusions and varying environmental conditions. It ensures robust performance across diverse scenarios, particularly improving the accuracy of foreground object segmentation and layout estimation in occluded areas. And extensive experiments on the nuScenes and Argoverse datasets demonstrate that BEV-CAR achieves state-of-the-art (SOTA) performance. More importantly, the rasterization technique in this paper does not introduce additional computational overhead during the inference process, making it suitable for practical deployment in real-world scenarios. Code and technical appendix are available in supplementary material.

Intelligent connected vehicles and autonomous service robots are rapidly becoming mainstream consumer electronics that operate in everyday environments and require reliable, low-cost perception capabilities. Among these, monocular Bird’s-Eye-View (BEV) semantic segmentation plays a pivotal role. However, monocular BEV perception remains challenging because of depth ambiguity, occlusion, and inaccurate cross-view feature projection. To address these issues without compromising deployment efficiency, we propose Depth-Centric BEV (DCBEV), a novel framework that achieves high accuracy without introducing inference-time overhead. DCBEV introduces a Geometry-Guided Depth Transformer that fuses multi-scale geometric priors with semantic features through back-projected attention, and a context-aware rasterization mechanism that operates only during training to resolve projection ambiguities in a Polarangle–Cartesian hybrid space, thereby adding no computational overhead at inference time. Extensive experiments on nuScenes and Argoverse demonstrate that DCBEV achieves state-of-the-art performance with mIoU scores of 33.8 and 30.3, respectively, highlighting its practical value for real-world consumer applications such as advanced driver assistance systems, augmented reality navigation, and autonomous mobile robots in complex urban or indoor settings.

Currently, fully supervised methods based on density map estimation are the primary research directions for crowd counting. However, such methods require the location-level annotations of persons in an image, which is time-consuming and laborious. Therefore, a weakly supervised method relying only on count-level annotations is urgently required. As a CNN is not suitable for modelling the global context and interactions between image patches, crowd counting with weakly supervised learning via a CNN generally does not provide good performance. A weakly supervised model using Transformer was sequentially proposed to model the global context and learn the contrast features. However, the Transformer directly partitions crowd images into a series of tokens, which may not be a good choice because each pedestrian is an independent individual, and the number of parameters of the network is very large. Hence, in this study, we propose a weakly supervised crowd counting method using Joint CNN and Transformer Network (JCTNet). JCTNet consists of three parts: a CNN feature extraction module (CFM), a Transformer feature extraction module (TFM), and a counting regression module (CRM). In particular, the CFM extracts crowd semantic information features and then sends their patch partitions to the TFM to model the global context. The CRM is used to predict the number of people. Extensive experiments and visualisations demonstrate that JCTNet can effectively focus on crowded regions and obtain superior performance in weakly supervised counting on five mainstream datasets, i.e., ShanghaiTech Parts A and B, UCF-QNRF, UCF-CC-50, and NWPU-Crowd. The number of parameters of the model can be reduced by approximately 67-73 % compared with that of a pure Transformer network. We also attempted to explain the phenomenon in which a model constrained only by count-level annotations can focus on crowded regions. We believe that our study can promote further research on weakly supervised crowd counting as well as any object counting. The code is available at https://github.com/wfs123456/JCTnet.

To address the inefficiency of intersection from independent control of vehicles and traffic signals in traditional transportation, a cooperative optimization method for connected vehicles and traffic signals via C-V2X (cellular vehicle-to-everything) was proposed. A closed-loop architecture integrating perception, communication, decision-making, and control was designed, leveraging edge intelligence to empower real-time traffic decision-making. A cooperative optimization algorithm, comprising a traffic pressure-based signal phase adjustment strategy and a state machine-based vehicle speed guidance strategy, was developed to achieve the collaborative optimization of signal phases and vehicle speeds. Hardware experiments and traffic flow simulations results demonstrate that the proposed method meets the requirements for hardware deployment and real-time response, effectively reducing vehicle travel time at intersections and improving traffic efficiency.

Human-in-the-loop reinforcement learning (HIL-RL) enables robots to learn contact-rich manipulation from limited real-world interaction, but deployment exposes three coupled limitations: static visual reward models fail under scene changes; independently sampled actions cause temporally inconsistent motion; and vision-based policies remain sensitive to appearance shifts. We present EvoHIL, a unified framework that adapts the reward model, action generator, and visual do main within a staged human-in-the-loop learning process. First, self-evolving reward (SER) adapts the success classifier from human-confirmed positives and provisional weak negatives. Second, Action Flow Stabilization (AFS) generates temporally coherent action chunks through flow matching, grounding policy updates in executed action prefixes and demonstrated behavior. Third, retention-aware offline fine-tuning replays relit interaction data while anchoring the AFS actor-critic to prior behavior, adapting the visual domain without additional robot interaction. Across six manipulation tasks on Franka FR3 and SO-101 arms under a controlled lighting shift, EvoHIL improves task success, agreement with human-confirmation labels, motion smoothness, and completion time relative to human-in-the-loop and imitation baselines.Project page: https://anonymous4366.github.io/EvoHIL/

Human-in-the-loop reinforcement learning systems achieve near-perfect success on the workstation where they are trained, but collapse when the same robot is moved to a workstation a few meters away due to shifts in the visual input distribution caused by new lamp positions and window light. Re-collecting demonstrations and re-running HIL on every workstation is incompatible with deployment, and naively fine-tuning on shifted-light data triggers catastrophic forgetting of the source workstation. To close this cross-domain gap, we present RoHIL, an offline fine-tuning framework that uses no extra real-robot interaction. RoHIL combines (i) a world-model-based image relighter that re-synthesises the visual stream of source-workstation trajectories under multiple virtual HDRI environments, leaving actions and rewards real; (ii) Illumination-Retention Replay (IRR), a data-level anti-forgetting mechanism that interleaves relit adaptation transitions with original-light retention transitions to preserve source-workstation Bellman coverage; and (iii) an anchored Bellman-actor regulariser that constrains representation and policy drift from the original source-workstation policy. Across four real-robot manipulation tasks under significant cross-workstation illumination variations, RoHIL substantially improves shifted-light performance where standard HIL-RL collapses, while preserving source-workstation performance, eliminating the need to re-collect data and retrain for every new workstation and environment. Project page: https://anonymous4365.github.io/RoHIL/

Cooperative Perception (CP) has been a promising paradigm to enhance single-vehicle awareness by enabling perception sharing among connected vehicles. However, existing studies often overlook the impact of constrained and heterogeneous edge resources, leading to synchronization bottlenecks and limited deployment efficiency. To address these challenges, this paper proposes EI-Cooper, an Edge Intelligence (EI)-enhanced cooperative framework for adaptive and efficient CP in heterogeneous vehicular networks. The novelty of EI-Cooper is fourfold. First, we leverage key EI techniques including selective cooperation, model pruning and bandwidth allocation to jointly coordinate the perception, computation, communication within the CP pipeline. To the best of our knowledge, EI-Cooper represents the first attempt to extend CP with EI capabilities. Secondly, we formulate a Synchronization-Efficient Cooperative Perception (SECP) problem, which jointly determines edge selection, pruning ratios and bandwidths to balance end-to-end synchronization efficiency and perception accuracy. Thirdly, to tackle the closed-box nature and computational NP-hardness of SECP, we decompose it into two interpretable subproblems, respectively capturing macro-level spatial completeness and micro-level semantic retention. Finally, we develop a Two-Stage Hierarchical Optimization (TSHO) algorithm, where the first stage maximizes coverage via submodular node selection with a $(1-1/e)$ approximation, and the second stage performs alternating optimization of pruning and bandwidth allocation under convergence guarantees. Extensive experiments on public datasets and a real-world prototype demonstrate the superiority of EI-Cooper.

Precise and physically consistent indoor localization is a fundamental requirement for next-generation mobile computing applications. However, it remains challenging due to the inherent asynchrony of heterogeneous sensor streams and the black-box behavior of deep learning models, which can yield kinematically implausible trajectories. In this paper, we propose GraphKANLoc, a novel framework that synergizes physics-informed learning with Kolmogorov-Arnold Networks (KANs) within a latency-aware graph architecture. Unlike traditional methods that rely on implicit data alignment, we construct a heterogeneous temporal graph that explicitly encodes the causal delay between high-frequency inertial dynamics and high-latency Wi-Fi measurements via directed cross-modal edges, effectively transforming sensor latency from a noise source into a structural feature for predictive learning. Furthermore, we use KANs as the trajectory regression head. We theoretically analyze and empirically validate that the inherent $C^{2}$ continuity of KAN spline-based activation functions enables more stable optimization of high-order kinematic constraints than standard Multi-Layer Perceptrons (MLPs). To promote physical realism, we introduce a composite kinematic regularization objective that includes jerk minimization and speed-heading consistency. This objective constrains the solution space to a physically feasible manifold during end-to-end training. Extensive experiments on the public dataset and multiple real-world scenarios demonstrate that GraphKANLoc not only achieves state-of-the-art localization accuracy but also reduces trajectory jitter by over 60%, producing smooth, navigation-ready paths without post-processing filters.