Kernel operations in resource-constrained edge systems, such as memory swapping and deduplication, use the access frequency (hotness) of memory pages to guide page placement and reclamation. However, these operations suffer from page-hotness skew: a page may contain a mix of highly accessed and infrequently accessed objects, which causes inaccurate page-level classification, wasted DRAM capacity, and expensive I/O. We attribute this skewness to a cross-layer mismatch: the kernel manages memory at page granularity, whereas user-level allocators place objects without considering access hotness.

Drones are widely used in fields such as aerial photography and security inspection, owing to their high mobility and wide field of view. Compared to a single drone, multi-drone systems can capture complementary information from different viewpoints, effectively addressing complex tracking scenarios such as occlusion, viewpoint changes, and target disappearance. However, existing multi-drone tracking methods often suffer from inadequate feature representation and inefficient inter-drone information fusion. To tackle these challenges, we propose FAEMTrack, a novel multi-drone tracking framework that integrates a feature-augmented embedding module (FAEM) with an advanced cross-drone fusion mechanism. FAEM enhances spatial-semantic feature encoding by replacing conventional convolutional layers with a multi-scale depthwise convolution block (MSDB) and spatial attention mechanism (SA), significantly improving the discriminative power of target features across different drone perspectives. Furthermore, to better utilize complementary information from different viewpoints, we introduce a cross-drone fusion strategy based on response confidence and entropy-weighted fusion, dynamically integrating information from multi-drone and promoting adaptive collaboration among them. Extensive experiments on the MDOT benchmark demonstrate that FAEMTrack outperforms existing state-of-the-art single-drone and multi-drone tracking methods in both tracking accuracy and robustness, showcasing its superior performance in complex scenarios. The source code and trained models will be available at https://github.com/wjh-scut/FAEMTrack.

Perception and interaction with articulated objects present a unique challenge for service robots. Although recent research has emphasized understanding articulated shapes and affordance proposals, existing methods only address isolated aspects, failing to develop comprehensive strategies for robotic perception and manipulation of articulated objects. To bridge this gap, we propose GMAP, which systematically integrates the entire process from command to perception and manipulation. Specifically, we first perform precise part-level segmentation of the object and identify the geometric and kinematic parameters of articulated joints. Then, by evaluating point-level affordance proposals, we determine the interaction poses for the robot's end-effector. Finally, the robot's execution trajectory is dynamically computed by combining commands with joint parameters and interaction points. Additionally, a key innovation of GMAP is addressing the scarcity of annotated data. We designed a multi-scale point cloud feature extraction module and introduced pre-training and fine-tuning techniques, significantly enhancing the generalization capability of the perception model. Extensive experiments demonstrate that GMAP achieves state-of-the-art (SOTA) performance in both the perception and manipulation of articulated objects and adapts to real-world scenarios.

Active vision learning aims to develop intelligent systems capable of actively exploring and understanding their surroundings to optimize detection performance. Although current research has begun to explore how reinforcement learning can drive robots to actively perceive their environment, it often overlooks the critical role of integrating historical contextual information. To address this, we propose an innovative approach to active vision learning, designed to enhance target detection performance in unknown environments. Specifically, this method combines reinforcement learning with deep neural network techniques and cleverly designs a sliding window mechanism to integrate observations over multiple steps into the state input. We employ PointNet++ for feature extraction and utilize a Transformer encoder module to process spatial contextual information, thereby constructing a comprehensive representation of the robot’s environment. Additionally, our carefully designed reward mechanism encourages the robot to prioritize the exploration of diverse object categories. With these reward strategies, we achieve significant improvements in detection accuracy and sampling efficiency. We validated our approach on the real-world 3D dataset R3ED, and the results demonstrate that our method outperforms other baseline methods in terms of performance.

Masked Autoencoders (MAE) demonstrated exceptional performance in natural language processing and 2D vision tasks and have now been introduced into point cloud representation learning. We propose Point-UMAE, a novel self-supervised learning method based on a Unet-like structure, designed to enhance the capture of local details and global semantics in point clouds. Point-UMAE employs an asymmetric encoder-decoder architecture with a top-down fine-grained masking strategy to improve multi-scale consistency. The pre-trained model achieves state-of-the-art performance across various downstream tasks. Compared to the baseline Point-BERT, our method achieves a classification performance improvement of 1% and 4.14% on the ModelNet40 and ScanObjectNN datasets, respectively. We also investigate the impact of masking strategies and encoder structures on performance.

Representation and generative learning mutually enhance each other across various domains. However, in point cloud processing, these models are typically trained independently, failing to leverage their synergy. To integrate 3D representation and generative learning deeply, we propose the Point-Mge framework. This framework uses a vector quantized variational autoencoder to reconstruct an occupancy neural field of 3D shapes, learning discrete semantic features of point patches. We introduce sliding masking ratios to smoothly transition from representation to generative learning. Additionally, our method exhibits good generalization in high-capacity models and achieves new state-of-the-art results in multiple down-stream tasks. In shape classification, Point-Mge attained 94.3% (+1.1%) on the ModelNet40 dataset and 92.9% (+5.5%) on the ScanObjectNN dataset. Experimental results also demonstrate that Point-Mge can generate high-quality 3D shapes in both unconditional and conditional settings. Code is available at https://github.com/robhlzeng/Point-Mge.

Talking-face synthesis is the process of generating facial motions synchronized with given audio, serving as an enabling technology for next-generation virtual assistants and video conferencing. However, existing talking-face synthesis systems rely on computation-intensive generative models, posing significant challenges for resource-constrained devices such as smartphones and Personal Digital Assistants (PDAs). This paper presents FastAvatar, an efficient talking-face synthesis system tailored for resource-limited devices. FastAvatar builds on our observations of the unique structure of generative models and the correlation between audio inputs and motion outputs. Specifically, FastAvatar employs a multimodal caching mechanism to avoid redundant computations and a fast adaptation strategy to handle fluctuating workloads, significantly improving end-to-end responsiveness. Our implementation and evaluation on multiple devices demonstrate that FastAvatar reduces latency by 3 times, with slight quality degradation (7.9 % LSE-C and 6.8 % PSNR). To the best of our knowledge, FastAvatar is the first work to support talking-face synthesis on resource-constrained devices.