With the rapid proliferation of automotive and edge artificial intelligence applications, hardware‐level security has become a critical concern, exposing intrinsic vulnerabilities of conventional deterministic logic systems. To address these limitations, we propose a cryptographic framework based on probabilistic bits (p‐Bits) implemented with transistor‐free Ti/SiO x /Ti threshold switching (TS) devices. In a 4 × 4 cross‐point TS array, self‐oscillation emerges as output voltage spikes whose probability follows a sigmoidal dependence on input amplitude, enabling experimentally tunable randomness. The intrinsic stochastic oscillation of the TS device under voltage pulses provides controllable entropy that forms the foundation for probabilistic logic operations. Based on this functionality, OR, AND, and XOR gates are constructed using experimentally extracted p‐Bit models in MATLAB and Python. These gate configurations demonstrate invertible logic where fixed outputs constrain plausible input states, which is a drawback in conventional complementary metal–oxide semiconductor logic. Leveraging this invertibility, we design a p‐Bit full‐adder unit (pFA) capable of encryption and decryption of grayscale images within the same hardware module. A key‐update mechanism is introduced to overcome limitations associated with static key‐based encryption, improving security resilience. The pFA achieves reliable image recovery with high fidelity, establishing a versatile platform for secure and energy‐efficient hardware cryptographic accelerators.

As new tactile technologies continually aim to improve the perceptive capabilities of humanoid systems, multiplexed monitoring methods—particularly those using electrical impedance tomography (EIT) hardware—are rapidly gaining popularity in soft robotics. Notably, their potential for system‐level adaptability is promising: by multiplexing between different electrodes, the amount and type of information being generated can be altered in real time, balancing accuracy and speed requirements across new environments and tasks. To solidify this approach, we introduce an approach —Economy‐of‐Touch —which poses tactile sensing as an information‐budgeting problem, enabling operating points to be identified across different tasks. We first propose a three‐part method of quantifying task complexity, grounded in image‐based similarity measures (normalized cross‐correlation, Chamfer distance, and edge‐based convex hulls) which capture both visual similarity and spatial distribution of the tactile stimuli. Validating our approach with thousands of real‐world EIT experiments, we then derive a task‐agnostic mapping between this complexity and the required amount of information: the task's operating point. Three tactile tasks (localization, classification, and pattern recognition) are evaluated with real‐world experiments, across which we demonstrate operating point MAEs as low as 0.659 channels for previously unseen data.

In this work, Voxel‐SLAM (simultaneous localization and mapping) is introduced: a complete, accurate, and versatile LiDAR (light detection and ranging) ‐inertial SLAM system consisting of five modules: initialization, odometry, local mapping (LM), loop closure (LC), and global mapping (GM), all employing the same map representation, an adaptive voxel map. The Voxel‐SLAM effectively utilizes short‐term, mid‐term, long‐term, and multimap data associations to achieve real‐time state estimation and high‐precision mapping. The odometry leverages short‐term data association and estimates current state with minimal latency. The LM, utilizing the mid‐term data association, designs an efficient LiDAR‐inertial bundle adjustment (BA) to refine the local map and states within a sliding window, which can run on an onboard computer in real time. The LC, exploiting the long‐term and multisession data association, can detect loops and support multisession mapping (up to five sessions in the experiments). To further capitalize these two data associations, the GM introduces an efficient global BA method and can even run on an onboard computer. Moreover, Voxel‐SLAM designs a robust initialization module to make the system start normally even under aggressive initial motion. In the presence of severe scene degeneracy or tracking loss, the system can automatically restart and relocalize to the previous tracking‐loss session when revisiting.

The exponential growth of genomic data has created a pressing need for methods capable of interpreting complex biological information, especially in fields like paleogenomics and environmental metagenomics. Ancient DNA (aDNA) analysis faces challenges such as degradation and contamination, while soil metagenomic DNA (soDNA) analysis is hindered by microbial diversity and incomplete reference databases. To address these limitations, this study proposes the polymerase chain reaction (PCR).Perturbation Theory and Machine Learning (PTML) methodology, which integrates machine learning with Perturbation Theory to analyze genetic sequences without the need for alignment. Two models are developed: the first classifies bacterial aDNA sequences extracted from Miocene amber; the second predicts tree health using microbial gene abundance in forest soils. Both rely on entropy‐based descriptors (θ k ) and structural differences (Δθ k ) between query and reference sequences, which serve as perturbation operators for supervised learning algorithms. This approach allows the detection of meaningful patterns even without complete genomic references. The aDNA model achieves 99.65% sensitivity and 99.81% specificity, while the soDNA model reaches 98.85% sensitivity and 92.56% specificity. These results confirm the robustness and applicability of PCR.PTML in diverse genomic contexts, presenting it as a valuable tool for ancient DNA classification and environmental metagenomics analysis.

Accurate and quantitative classification of cell death pathways is fundamental for elucidating disease mechanisms and assessing therapeutic efficacy, as dysregulated cell death underlies a wide range of pathological conditions including cancer and therapy resistance. However, conventional imaging methods such as fluorescence and bright‐field microscopy, or 2D phase imaging, often suffer from phototoxicity, labeling artifacts, or limited morphological contrast. Herein, a real‐time, label‐free platform for classifying cell death phenotypes—apoptosis, necroptosis, and necrosis—is introduced by combining 3D holotomography with deep learning. A convolutional neural network, trained on refractive index (RI)‐based features from HeLa cells, achieves high classification accuracy (99.3%) under varying cell densities. Notably, the model detected early RI changes during necroptosis several hours before conventional fluorescence‐based markers, validated by population‐level comparison with flow cytometry. The framework's adaptability is demonstrated by successfully fine‐tuning the model for A549 lung cancer cells. Collectively, these findings demonstrate the potential of HT‐based AI as a universal, high‐resolution, and label‐free platform for quantitative cell death profiling and translational drug‐response analysis.

In open‐world robotic manipulation tasks, language‐guided model‐free grasping has garnered increasing attention. However, existing approaches often overlook the geometric structure of target objects, which limits the effectiveness of subsequent tasks such as manipulation and placement. To address this limitation, a novel method called Language‐Guided Grasping via Primitive Fitting is proposed. This approach integrates language instructions with multimodal perception to enhance the semantic interpretability and downstream usability of the grasp through structured geometric modeling. Specifically, the user‐specified object using 2D images and depth data via multimodal understanding is first localized. Then, primitive fitting on the object's point cloud using basic geometric shapes (e.g., cuboids, ellipsoids, truncated cones) to extract approximate size and structural features is performed. Based on the geometric information, a grasp pose generation strategy guided by semantic geometry is defined, and modules for grasp feasibility filtering and task‐oriented optimization to select the optimal grasp pose are introduced. This method is validated in real‐world complex environments and achieved grasp success rates of 95% in structured and 90% in cluttered scenes. Geometric fitting enhances post‐grasp predictability and semantic consistency, enabling better generalization and planning.

Continuum robots enable safe and adaptive interaction with complex, unstructured, or constrained environments through continuous deformation, making them particularly suitable for medical and industrial applications. Accurate contact force sensing is essential to ensure safe and effective physical interaction in such scenarios. Although various embedded force sensors have been developed, sensor‐free approaches offer advantages in miniaturization, cost‐effectiveness, and biocompatibility. This review provides a comprehensive overview of sensor‐free contact force estimation methods for continuum robots, with an emphasis on algorithmic principles rather than specific continuum robot designs or applications. First, contact forces reported in the literature are systematically classified according to their distribution, components, and dynamics. Next, existing force estimation methods are divided into three categories: actuation‐based, deflection‐based, and environment‐based. For each category, the underlying algorithmic principles are discussed, representative challenges are highlighted, and their typical application scenarios are outlined. Finally, emerging trends and potential directions for future research are outlined.

Artificial intelligence (AI) is increasingly shaping modern cardiology by enhancing clinical interpretation through data‐driven insights, surpassing traditional subjective assessments. This review explores AI's impact on the cardiac system, emphasizing the development of an intelligent end‐to‐end platform for prediction, diagnosis, treatment, and rehabilitation. AI enables a unified ecosystem for diagnosing and treating cardiovascular (CV) diseases, both before and after symptom onset, ensuring seamless workflow integration. Key technological advances such as deep learning, federated learning, natural language processing, and multimodal data convergence form the backbone of this collaborative CV ecosystem. AI has the potential to revolutionize CV risk evaluation, personalize treatments, and enable real‐time monitoring. However, challenges remain, including improving algorithm robustness, model reliability, and safeguarding patient privacy. The review also discusses the future role of generative models, edge AI, and federated learning to improve scalability while maintaining privacy. Ultimately, AI aims to shift cardiology toward a more data‐driven, personalized, and efficient system, enhancing both patient experience and care affordability.

Replicating the flexible and efficient locomotion of biological snakes remains a significant challenge in robotics. Conventional snake robots, often built from serially linked rigid joints, require complex control strategies to simultaneously manage body undulation and propulsive ground forces. Based on the previous theoretical study, this article presents the first physical realization of a new principle that simplifies locomotion control by decoupling these two tasks. The core idea is that by using a flexible continuum body, different gaits can be generated by superimposing simple globally‐applied tensions (for vertical bending and axial twisting) onto a basic planar undulation. These global tensions, combined with the robot's compliance and weight, passively shape the required ground contact patterns, eliminating the need for active force control at individual points. The design and implementation of a tendon‐driven continuum snake robot that embodies this principle is presented. The robot uses globally routed tendons, actuated by centrally‐located motors, to create uniform bending and twisting. Through experiments, it is demonstrated that the robot can produce three distinct gaits—forward, backward, and sidewinding—on flat ground simply by changing the global actuation mode, and demonstrate its capability for intuitive steering.

Traditionally, technology computer‐aided design (TCAD) has been employed to analyze the effects of line edge roughness (LER). However, TCAD‐based analysis is computationally prohibitive, particularly for circuit‐level simulations like SRAM, which require extensive computational time. To efficiently analyze the impact of LER, this study proposes a hybrid model combining machine learning and an analytical model, offering high computational efficiency without compromising accuracy. Specifically, this study employs a convolutional neural network (CNN) that directly takes the surface roughness of fin‐shaped field‐effect transistors as input. This allows us to account for spatial characteristics, such as the correlation between two neighboring transistors, which is essential for predicting circuit characteristics. In this work, it first introduces the data generation process, the training process, and the CNN model with the optimized architecture. Next, it introduces a hybrid model that combines both a CNN and an analytical model to predict the voltage transfer characteristics curves and static noise margins for SRAM read and write operations. The proposed hybrid model demonstrates rapid learning and significant time savings in neural network preprocessing while maintaining high accuracy.