Accurate and efficient perception of positive and negative obstacles is critical for the safe autonomous navigation of lunar rovers in extreme unstructured lunar environments. Existing multi-modal fusion methods often suffer from semantic misalignment between modalities, limiting the full utilization of complementary information from RGB and depth data. To address this issue, we propose LuOS, an efficient two-stage multi-modal feature fusion framework for lunar obstacle segmentation. The framework first employs a Contrast-Driven Fusion Module to achieve precise cross-modal semantic alignment by leveraging pre-trained RGB features to guide depth feature learning via contrastive learning. Subsequently, a Cross-Modal Interactive Fusion module fuses the aligned features to fully leverage their complementary information. Extensive experiments on the lunar-specific LunarSeg dataset and real-world road NPO dataset demonstrate that LuOS achieves state-of-the-art performance. Notably, LuOS maintains high real-time inference capability, essential for robotic decision-making, while exhibiting strong cross-domain generalization from lunar to terrestrial environments, validating its effectiveness and practicality for adaptive robot perception.

Generating task-individual independent customized assistive control is a big challenge of wearable powered exoskeletons, which usually relies on accurate and robust motion intention perceptions (MIP) of the wearer's limb. Traditional physics-based models and deep learning models, as two commonly used MIP methods, suffer from poor generalization and strong data dependence, respectively. To overcome these limitations, in this study, we propose a physics-informed neural network (PINN) model named PILaN, which integrates the Lagrange dynamics with a deep learning model, to realize lower limb motion intention perception-based (LLMIP-based) assistive control on a hip-knee powered exoskeleton (HKPE). The proposed PILaN trained by a self-collected small scale dataset is capable of conducting 2-degree of freedom (DoF) lower limb dynamics estimation (LLDE) with the current hip and knee joint states, and using LLDE results to obtain the next joint state according to the Lagrangian dynamics. The assistive control of applied HKPE depends on the estimated joint states from the PILaN. Eight participants are invited to perform designed motion sequences consisting of one or multiple actions with the assistance of the HKPE. Experimental results demonstrate that our proposed PILaN can successfully provide accurate and robust LLMIP for the HKPE assistive control under task-individual independent conditions.

In scenarios where collaborators and noncollaborators coexist, it is important to enable covert information exchange among collaborators while evading detection. Based on the isometric muscle contraction, a surface electromyography (sEMG)-based gesture-free hand intention recognition system has been developed to transmit messages securely. However, nonstationary sEMG signals, electrode displacement, and individual differences restrict the performance of single-day, cross-day, and cross-subject experiments. To overcome the difficulty, a discrete wavelet transform-based sEMG data alignment (DWT-DA) method is proposed to reduce the distribution difference between the source and target domains. The DWT-DA consists of three stages: time-frequency-domain signal decomposition (TFD-SD), subject subsignal recombination (SR), and time-frequency-domain signal reconstruction (TFD-SR). The TFD-SD decomposes the original signal into several subsignals focused on a specific frequency band or time-frequency region. The SR recombines the decomposed subsignals to promote complementary enhancement between the features of each subsignal. The TFD-SR reconstructs the processed subsignal into a time-domain signal to achieve data augmentation and alignment. Experiments are conducted on a self-collected gesture-free hand intention recognition dataset, demonstrating excellent performance of the method on single-day, cross-day, and cross-subject evaluations. The code is available at: https://github.com/ylfzero/EMG_DWT_DA

The prevailing paradigm of perceptive humanoid locomotion relies heavily on active depth sensors. However, this depth-centric approach fundamentally discards the rich semantic and dense appearance cues of the visual world, severing low-level control from the high-level reasoning essential for general embodied intelligence. While monocular RGB offers a ubiquitous, information-dense alternative, end-to-end reinforcement learning from raw 2D pixels suffers from extreme sample inefficiency and catastrophic sim-to-real collapse due to the inherent loss of geometric scale. To break this deadlock, we propose GeoLoco, a purely RGB-driven locomotion framework that conceptualizes monocular images as high-dimensional 3D latent representations by harnessing the powerful geometric priors of a frozen, scale-aware Visual Foundation Model (VFM). Rather than naive feature concatenation, we design a proprioceptive-query multi-head cross-attention mechanism that dynamically attends to task-critical topological features conditioned on the robot's real-time gait phase. Crucially, to prevent the policy from overfitting to superficial textures, we introduce a dual-head auxiliary learning scheme. This explicit regularization forces the high-dimensional latent space to strictly align with the physical terrain geometry, ensuring robust zero-shot sim-to-real transfer. Trained exclusively in simulation, GeoLoco achieves robust zero-shot transfer to the Unitree G1 humanoid and successfully negotiates challenging terrains.

The practical application of aerial-aquatic robots is hindered by severe impact loads during water entry. Existing load reduction methods are inefficient for robots requiring rapid, high-frequency aerial-aquatic transitions. Therefore, this study proposes an active learning based joint optimization framework for nose profiles and water-entry strategies, which is fully automated. Specifically, an 8-dimensional parameter space is defined for generating dataset inputs through Latin Hypercube Sampling (LHS). Furthermore, a Deep Kernel Learning (DKL) surrogate model is trained under different water-entry strategies, in order to predict peak impact loads and quantify prediction uncertainty for varying nose profiles. Within each active learning loop, the Non-dominated Sorting Genetic Algorithm III (NSGA-III) guides the selective labelling of samples to expand the dataset, and the DKL model is iteratively retrained until convergence. Compared against state-of-the-art methods, the proposed approach reduces the number of required high-fidelity Computational Fluid Dynamics (CFD) simulations to 65% of that of the comparison methods, while achieving maximum reductions of 66.9% in axial and 72.8% in normal impact loads across the parameter space, respectively. Notably, under the water-entry strategies employed by the hunting behavior of gannets, the approach yields profiles closely matching the skull morphology of the northern gannet. In this case, this work not only delivers an efficient and reliable impact loads reduction solution for aerial-aquatic robots, but also reveals the role of natural selection in minimizing such loads. Note to Practitioners-The purpose of this work is to reduce the impact loads on aerial-aquatic transmedia robots during water entry. Existing approaches generally train a single surrogate model for impact load prediction and utilize optimization algorithms to reduce such loads; however, dataset construction requires tedious manual 3D modeling and time-consuming Computational Fluid Dynamics (CFD) simulations. This study proposes an active learning-based optimization framework that can automatically complete the entire process, from nose profile modeling to surrogate model training, without manual intervention. The framework enhances the prediction accuracy of the surrogate model for inputs conducive to load reduction by selectively adding samples before retraining, while also reducing the number of high-fidelity CFD simulations, thereby saving time and computational costs. The optimized design obtained via this framework effectively reduces the impact loads on the robot, achieving values lower than the minimum identified through extensive manual CFD simulations with uninformed search. Furthermore, the optimized nose profile shows superior load reduction performance compared to the traditional conical nose. Although the optimization was conducted specifically on the Lingyuan aerial-aquatic robot, the proposed automated framework is broadly applicable to the design of any water-entry vehicle with complex geometries for impact load reduction.

The quality of underwater videos depends on absorption and scattering of light, as well as on the performance of transmission devices. In this paper, we leverage the characteristics of underwater imaging and develop a video super-resolution method optimized for underwater scenarios. We first construct a video dataset containing a substantial amount of authentic underwater footage for experimental analysis. Then, we preprocess underwater images based on the principle of minimal color loss, mitigating color distortions caused by lighting effects in underwater videos. Moreover, we incorporate a residual network with spatial attention mechanisms to align features from both reference and adjacent frames, thereby obtaining richer information. Finally, by merging linearly upsampled reference frames with features aligned using spatial attention, we reconstruct high-resolution videos at their original resolution. We evaluate the proposed method both qualitatively and quantitatively on the created underwater dataset. The results demonstrate that our method effectively addresses light attenuation in underwater settings, leading to improved quality in super-resolution reconstruction.

Dexterous robotic manipulation remains a challenge due to the difficulty in capturing and transferring human-like adaptive control and fine force regulation. This paper presents a co-designed data glove system that enables effective skill transfer from human demonstrations to robotic hands through imitation learning. The custom data glove hardware integrates heterogeneous tactile sensing with joint-level inertial motion capture. This system specifically features high-resolution 3D sensors at the fingertips and sparse pressure sensors on the palm, a design that mimics the biological non-uniformity of human tactile sensitivity. Supported by custom electronics, robust mechanical fixation, and a dedicated compression scheme, the system acquires high-fidelity, multi-modal hand-state data in real time.We further introduce a Hierarchical Spatio-Temporal Graph Network (HSTGN) that processes these data via a unified graph representation, incorporating cross-attention mechanisms for fingertip interaction modeling and graph sparsification for computational efficiency. Experimental validation on tasks ranging from precision pinch to whole-hand grasp shows that HSTGN achieves a 91.3% success rate. This work provides a hardware-software integrated framework for capturing and transferring dexterous manipulation skills, closing the gap between human-like adaptability and robotic execution.

Semi-supervised approaches for medical image segmentation are frequently mired by the concomitant challenges of noisy pseudo-labels and ill-defined boundaries. We address these entangled issues with DSHENet, a novel architecture that synergizes dual-stream hybrid encoding and progressive consistency co-training. Specifically, we design a Dual-Stream Hybrid Encoder (DSHE) that jointly captures global semantic context and high-frequency structural details within a shared feature space via delayed global-feature injection and multi-scale context fusion. We further introduce Bilateral Attention Decoding (BLAD), where the main decoder focuses on region consistency while the auxiliary decoder enhances boundary awareness through morphology-inspired structural modeling. To improve pseudo-label reliability, we develop a Progressive Consistency Co-Training (PCCT) strategy that transitions from supervised initialization to dynamic cross pseudo-supervision with confidence-aware masking. Experiments on ACDC, Kvasir-SEG, and ISIC2018 show that DSHENet consistently improves both overlap and boundary metrics over strong baselines under limited annotation. For example, on ACDC with 10

Lung cancer, a global leading cause of incidence and mortality, demands accurate detection. Current research primarily targets early nodule detection and distinguishing benign from malignant tumors, with limited focus on lung cancer classification. Identifying various lung cancer types is vital for tailored treatments, while challenges persist in localizing small lesions and early tumors. To tackle these challenges, we propose the framework integrated the Channel Equally-Divided (CED) and the Coordinate Attention Feature Pyramid Network (CAFPN). CAFPN, an innovative feature pyramid structure, integrates the Semantic Information Enhanced (SIE) module and Coordinate Attention. The SIE module filters redundant semantic information, enriches texture features through Coordinate Attention, emphasizes shallow semantic details, and amplifies trustworthy specifics for enhanced deep semantic information. Additionally, the CED module is devised to proficiently extract local contextual information across channels for more precise feature representations. The superior performance of our proposed method was empirically validated through comparative experiments on two mainstream datasets with Competition Performance Metric (CPM) scores and Mean Average Precision (mAP) values reaching 0.942 and 99.18%, respectively, outperforming current state-of-the-art approaches.

Long-horizon robotic manipulation remains a critical challenge in robotics. Hierarchical reinforcement learning offers a promising solution, but often suffers from an imbalance dilemma: simplifying skill learning increases the complexity of planning, thereby expanding the solution space and computational burden of planning. To tackle this challenge, we propose a Hierarchical Reinforcement Learning framework with Dynamic Kolmogorov-Arnold Network (DyKAN) based Actor Critic, named HIKER. Firstly, HIKER innovates with a dual-chain design that decomposes the complex task into two intersecting sub-chains, reducing the optimization conflict across subtasks and alleviating the burden on the planning model. Secondly, we develop DyKAN, a scalable neural network for both actor and critic in the skill model of HIKER. DyKAN adaptively adjusts grids and basis functions while preserving learned knowledge, enabling efficient learning of complex manipulation skills. Furthermore, to optimize DyKAN's performance, we design a per-layer update module that uses Dynamic Tanh (DyT) and low-rank decomposition to ensure stable, low-cost updates during training. Finally, experiments on long-horizon robotic manipulation tasks demonstrate that HIKER significantly improves efficiency and robustness, yielding higher-quality skill models and achieving a 10.9% increase in task success rate under the high noise condition. Further insights are available on the website: https://sites.google.com/view/hikerdykan.

Purpose Robust replanning capabilities are essential for autonomous robots to navigate dynamic and complex environments. This paper aims to propose the sampling-guided dual-tree (SGDT*), designed to simultaneously improve replanning efficiency and promote the generation of satisfactory paths. Design/methodology/approach SGDT* is composed of configuration space exploration and adaptive replanning. The configuration space is comprehensively and uniformly covered by the configuration space exploration module through the expansion of a main search tree, providing an initial feasible path and all required nodes. Environmental changes are addressed by the adaptive replanning module, where a sampling set is strategically extracted from existing nodes to grow the sub-search tree and rapidly identify new feasible paths. Additionally, feasible paths are used to further contract the sampling set, increasing the probability of selecting valuable nodes and facilitating efficient, high-quality replanning. Findings The properties of SGDT*, including probabilistic completeness, shortest path and runtime are theoretically analysed. The effectiveness and real-time performance of SGDT* are rigorously validated through comparative simulations and real-world mobile robot experiments in various challenging scenarios. Originality/value SGDT* supports stable, real-time updates of collision-free paths in dynamic environments, achieving both efficient replanning and high-quality path generation. SGDT* can be seamlessly integrated into robotic navigation architectures to enhance the reliability and adaptability of autonomous systems that require frequent replanning in complex environments.

Tactile and kinesthetic perceptions are crucial for human dexterous manipulation, enabling reliable grasping of objects via proprioceptive sensorimotor integration. For robotic hands, even though acquiring such tactile and kinesthetic feedback is feasible, establishing a direct mapping from this sensory feedback to motor actions remains challenging. In this article, we propose a novel glove-mediated tactile–kinematic perception–prediction framework for grasp skill transfer from human intuitive and natural operation to robotic execution based on imitation learning, and its effectiveness is validated through generalized grasping tasks, including those involving deformable objects. First, we integrate a data glove to capture tactile and kinesthetic data at the joint level. The glove is adaptable for both human and robotic hands, allowing data collection from natural human hand demonstrations across different scenarios. It ensures consistency in the raw data format, enabling evaluation of grasping for both human and robotic hands. Second, we establish a unified representation of multimodal inputs based on graph structures with polar coordinates. We explicitly integrate the morphological differences into the designed representation, enhancing the compatibility across different demonstrators and robotic hands. Furthermore, we introduce the tactile–kinesthetic spatio-temporal graph networks, which leverage multidimensional subgraph convolutions and attention-based long short-term memory (LSTM) layers to extract spatio-temporal features from graph inputs to predict node-based states for each hand joint. These predictions are then mapped to final commands through a force-position hybrid mapping. Comparative experiments and ablation studies demonstrate that our approach surpasses other methods in grasp success rate, finger coordination, contact force management, and both grasp and computational efficiency, achieving results most akin to human grasping.

Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in https://github.com/DongXu-Zhang/IGASA.

The pursuit of humanoid athletic sprints is hindered by a scarcity of humanoid-viable kinematic reference data and the inability of existing frameworks to maintain stability during sprints. To overcome these limitations, we introduce SPRINT, a novel framework driven by efficient, frequency-adaptive spectral priors. By characterizing the fundamental periodicity of human locomotion in the frequency domain using a reference library of five discrete motion sequences, these priors generate kinematically feasible joint trajectories across a broad velocity spectrum, successfully extrapolating to speeds that exceed the reference distribution. Guided by these pretrained priors, the SPRINT policy achieves zero-shot sim-to-real transfer in field experiments on the Unitree G1 platform, reaching a peak sprinting velocity of 6 m/s and demonstrating seamless gait transitions while preserving biomimetic naturalness. Ultimately, this work establishes frequency-adaptive spectral priors as a highly data-efficient foundation for humanoid athletic sprints. The project page is available at https://anonymous.4open.science/w/SPRINT-138A/.

With the growing demand for high-precision spatial perception in consumer electronics, 6D pose estimation technology has become a core enabler for scenarios such as augmented reality and smart home systems. Accurate 6D pose estimation requires effective utilization of depth information and image data. However, most existing pose estimation methods for heterogeneous information fusion merely concatenate the data without considering the implicit correlations between different data distributions. This may increase the algorithm’s sensitivity to outliers and noise in complex environments with dynamic scenes or significant lighting changes. Additionally, redundant information inherent in cross-modality signal representations necessitates rational elimination. These limitations lead to algorithm instability in practical applications. To address these challenges, we propose a novel self-supervised learning method named Cross-modality Masked Autoencoder (CMAE). CMAE effectively integrates information from different sensor modalities by combining the strengths of mask autoencoders. At the same time, it employs a hierarchical cross-modality similarity extraction module to remove redundant information. Experiments show that our algorithm achieves an average accuracy of 90.1% on the Linemod-Occlude dataset and 94.8% on the YCB-Video dataset. These properties make CMAE particularly suitable for consumer electronics applications, where robust and efficient 6D pose estimation is critical.

Mobile manipulators (MMs) need to determine feasible navigation positions before manipulation tasks. Real-world environments, with varying obstacles and objects, pose significant challenges for computing optimal navigation positions due to their variability. In this work, a novel method named graph reinforcement learning (GRL)-based reachability map (GRAM) is proposed. First, GRAM uses a graph attention network (GAT) to capture the spatial relationships between objects. Then, it leverages the Q-value from the pretrained critic network to generate the reachability map. The reachability map is integrated into navigation policies for long-horizon tasks, effectively solving the skill transition problems. Extensive simulation and real-world experiments were conducted on the Fetch mobile robot platform. The results demonstrate the superiority of GRAM, with simulation results showing an average 16.3% performance improvement over the baseline in four flexible environments. In long-horizon tasks, GRAM's overall task success rate (SR) improved by 4.2%.

In weakly supervised learning-based medical image segmentation, the incompleteness of labels and the complexity of medical image structures often lead to under-segmentation and over-segmentation issues in practical predictions. Therefore, a weakly supervised brain image segmentation guided by texture knowledge (TKG-Net) is proposed in this paper. Its core idea is based on the interaction between the perceptual system and the rational system in the cognitive dual-path model to locate the fuzzy boundaries of the target. First, a simple and effective perceptual system, called the Stepped Class Activation Map Fusion Algorithm, is designed, which utilizes the ability of deep class activation map (CAM) to capture the localization characteristics of the target position and the ability of shallow CAM to highlight the texture characteristics of the target details, and maximally removes background noise to emphasize foreground targets. Second, a novel rational system, called the Texture Knowledge Extraction Module, is introduced into the network, through the collaborative action of fuzzy K-Means and Prewitt operator to efficiently extract the texture features of the target. Subsequently, these captured target texture structural features are effectively integrated into the texture consistency loss function. Finally, a novel Cross-guided module is constructed, which can fully exploit the correlation between texture information and semantic information, achieving more effective fusion of input vectors and thus achieving more precise segmentation of the target region. Our method is evaluated on the BraTS2019 and INSTANCE2022 datasets. Experimental results demonstrate that the proposed TKG-Net exhibits excellent segmentation performance compared to other state-of-the-art methods.

Soft-bodied organisms exhibit prominent morphological adaptability, dynamically reconfiguring shape and stiffness to achieve versatile behaviors. Inspired by these systems, soft robots with diverse morphologies have emerged, yet a unified control framework that rapidly adapts across configurations remains elusive. Here, we introduce a generalizable control system that enables rapid cross-configuration adaptation via reinforcement learning in a shared linear Koopman embedding space. By encoding robot dynamics into this embedding space, our method decouples control policies from specific morphologies, allowing real-time, model-free policy adaptation without retraining from scratch. We validate our system across 33 distinct robot configurations. Our system achieves a 75 × reduction in transfer samples across configurations, while sustaining robust performance under high-speed motion, heavy payloads, and multiactuator faults, and achieving real-world skills previously unattainable in soft robotics. This work establishes an adaptable control framework for diverse soft robot configurations and may offer insights for generalizable control in complex physical systems.

In this article, an admittance control strategy for aerial manipulators is proposed to achieve contact force tracking in unknown dynamic environments. First, considering the impact of variations in unknown environments on force tracking performance, an adaptive variable stiffness feature is incorporated into an advanced admittance model. The stiffness coefficient is dynamically adjusted using position and force feedback to generate the desired reference trajectory. Second, to address the issue of reference trajectory tracking under disturbances, a pose controller composed of a disturbance observer and barrier Lyapunov function is utilized to achieve stable tracking performance. In the absence of prior knowledge of disturbances, the state variables converge to a constrained range within a finite time, without introducing excessively high control gains. Finally, the stability of the proposed strategy is rigorously analyzed via Lyapunov tools. Both simulations and real-world experimental investigations are conducted to demonstrate the feasibility of the control strategy, highlighting its robust performance in maintaining a stable contact force during interaction with unknown dynamic environments.

A significant bottleneck in humanoid policy learning is the acquisition of large-scale, diverse datasets, as collecting reliable real-world data remains both difficult and cost-prohibitive. To address this limitation, we introduce HumanoidExo, a novel system that transfers human motion to whole-body humanoid data. HumanoidExo offers a high-efficiency solution that minimizes the embodiment gap between the human demonstrator and the robot, thereby tackling the scarcity of whole-body humanoid data. By facilitating the collection of more voluminous and diverse datasets, our approach significantly enhances the performance of humanoid robots in dynamic, real-world scenarios. We evaluated our method across three challenging real-world tasks: table-top manipulation, manipulation integrated with stand-squat motions, and whole-body manipulation. Our results empirically demonstrate that HumanoidExo is a crucial addition to real-robot data, as it enables the humanoid policy to generalize to novel environments, learn complex whole-body control from only five real-robot demonstrations, and even acquire new skills (i.e., walking) solely from HumanoidExo data.