Mobile authentication is essential for protecting personal data and payment authorization, yet prevailing approaches face two fundamental limitations: the nonrevocability of physiological biometrics once compromised and the susceptibility of knowledge‐based authentication to observation‐ and trace‐based attacks. This study proposes a next‐generation mobile authentication framework based on dynamic behavioral profiling, in which binary pressure sequences and grip‐position patterns are jointly modeled. The hardware design measures fine spatiotemporal pressure distributions using symmetrically mounted fabric‐based strip sensor arrays along the display bracket, and enhances localized pressure‐transfer efficiency via an electrode‐aligned protrusion film. Closely coupled to this sensing design, the algorithm employs a probability‐based soft‐stacking scheme that combines class‐wise predictive probabilities from base learners to mitigate information loss induced by threshold‐based discretization, increasing mean accuracy from 0.802 to 0.857 compared with fixed‐threshold encoding. The proposed framework achieves an AUC of 0.9867 (EER = 0.049) for 16‐class pressure‐sequence authentication and an AUC of 0.9870 (EER = 0.0595) for 6‐class grip‐scenario classification. These results support the feasibility of using touch‐pressure and grip interactions as re‐enrollable behavioral authentication signals and provide preliminary participant‐level validation of the framework. Further investigation involving larger and more diverse populations and real‐world settings is needed to assess its potential for continuous authentication.
Cellular materials with porous and lightweight structures have attracted attention due to exceptional mechanical tunability and broad applications. Despite advancements in design strategies, traditional simulations remain computationally intensive, motivating developments of efficient optimization methodologies. This study introduces Cellular Material Network (CM‐Net), a physics‐informed machine learning architecture for predicting mechanical properties of cellular materials. By treating basic geometric units as tokens analogous to those in natural language processing, CM‐Net achieves efficient forward prediction and generalization across diverse structures and compositions. Validated against simulations and experiments, CM‐Net accurately predicts nonlinear behaviors, including initial peak compression force, mean compression force, and energy absorption. The model maintains low relative errors for out‐of‐distribution scenarios such as varying wall thicknesses and novel topologies composed of short straight segments. However, extrapolation accuracy decreases for geometries with long curved edges, owing to insufficient representation of deformation mechanisms in training. CM‐Net also predicts the behavior of composite and gradient structures, as well as nonlinear force–displacement curves. We further demonstrate inverse designs of complex irregular cellular materials to meet specific requirements, illustrating CM‐Net's utility as a design assistant, with experimental validation confirming its practical applicability. Its scalability and generalizability make CM‐Net a transformative tool for accelerating lightweight, high‐performance cellular material development.
Bats integrate biosonar sensing with agile flapping flight, making them a promising model for embodied artificial intelligence. Detailed study in complex natural habitats remains difficult, so laboratory paradigms such as mazes of thin wires are used to enable controlled observation. Whether such setups reproduce the acoustic characteristics of natural clutter, including distinct echo signatures associated with different vegetation, remains an open question. Systematic evaluation is computationally demanding because biosonar operates at high frequencies and involves many scatterers, where conventional numerical solvers become prohibitive. A validated framework is presented that combines efficient physical modeling with deep learning to guide wire maze design. An accelerated multiple scattering model for many cylinders provides broadband echo simulations at greatly reduced cost. These simulations train a convolutional classifier that tests acoustic distinguishability among different wire arrangements using echo spectrograms. The classifier reliably separates random wire arrays, indicating that the mazes generate repeatable, structured echo features rather than uninformative noise. The resulting tools enable the design of laboratory wire maze experiments that emulate natural biosonar sensing and maneuvering scenarios while reducing occlusions and supporting detailed recordings of flight and biosonar behavior.
Analog in‐memory computing offers an energy‐efficient solution to resource‐intensive matrix operations in artificial neural networks by performing multiplication via Ohm's law and summation via Kirchhoff's current law. However, two‐dimensional (2D) computing fabrics require unrolling of 2D input into one‐dimensional vectors, demanding an auxiliary memory and diminishing the overall efficiency. Herein, we propose three‐dimensional (3D) architectures that enable single‐step matrix operations without data unrolling. Our proposed hardware interfaces with 2D inputs and outputs directly, allowing data flow seamlessly across the stacked layers. We experimentally validate this concept for matrix‐kernel convolution (MKC) for image processing using both a transistor‐based system on a printed circuit board and a 3D memristive array, and we describe and analyze matrix to matrix multiplication (MMM) as a supported architectural mapping. Our 3D designs are stackable and scalable, enabling parallel multilayer matrix operations in a single step and thereby yielding reduced hardware complexity while delivering over 113‐fold energy savings and orders‐of‐magnitude improvement in operational latency.
Analog in‐memory computing (AIMC) emerges as a promising solution to overcome the energy and latency bottlenecks in von Neumann architectures for artificial intelligence workloads. Computation is performed directly within arrays of nonvolatile memories. Ferroelectric and spintronic devices, relying on the control of polarization and magnetization, offer intrinsically energy‐efficient and low‐latency building blocks for AIMC. This review highlights how these two physical hardware families share a unified computational framework: both utilize the hysteretic dynamics of an order parameter to provide nonvolatile, multistate memory and nonlinear switching. To date, most AIMC implementations have been conceived for mature memristive technologies such as RRAM. Because these devices are current‐driven, unlike ferroic devices whose operation is field‐driven, their associated implementation strategies cannot be directly ported, motivating the need for a new codesign space that this review seeks to establish. We review static, array‐based vector–matrix multiplications for machine learning, and dynamical computing paradigms like reservoir computing and spiking neural networks. While device constraints such as quantized value precision, asymmetric weight updates, low‐frequency noise, and device nonidealities are often seen as limitations, we show how they can instead be leveraged for training. Learning methods include hardware‐aware and physics‐aware training, paving the way toward holistic, brain‐inspired neuromorphic computing.
Over the last decade, soft robots have expanded the boundaries of robotic horizons, showing strong potential in mimicking the true nature of living organisms as well as addressing real‐world problems that conventional rigid‐bodied robots have not achieved. At the heart of these successes are innovative and creative solutions in designing soft gripper bodies and seamlessly integrating actuated systems with hierarchical sensing. Among the possible scenarios that soft robots could aim at, versatile and dexterous grasping and manipulation have been intensively studied, leading researchers, engineers, and scientists to explore multidisciplinary solutions, from material science, mechanical and electrical engineering, and computer science. Herein, achieving both desired mechanical performances (i.e. highly compliant yet strength), and versatile functionality in grasping remain dreaming tasks that many of roboticists desire. In parallel, given that multimodal grasping via morphological computation—where actuation, sensing, and control are well distributed along the soft body—has been conceived as emerging research topics, this comprehensive study aims to provide a comparative analysis of design principles, actuated systems, and sensing strategies, highlighting deep insights and exploring feasible solutions to break through bottlenecks and remaining open challenges in creating totally new intelligent soft machines that can actively interact with environment.
Restoring tactile feedback that can be delivered through selective neural interfaces is essential for prosthetic embodiment. This study presents a biomimetic 3D tactile sensor system designed to convert skin-like mechanical interactions into functionally selective neural stimulation patterns at the fascicle level. The system integrates a polydimethylsiloxane (PDMS)-based artificial skin layer with tunable mechanical compliance, embedded slow-adapting (SA) and fast-adapting (FA) sensor channels, and a neuromorphic encoding framework that transforms analog sensor outputs into receptor-potential-like waveforms and action potential spike trains. MWCNT-PDMS piezoresistive composites provide sustained pressure-sensitive SA outputs, whereas BaTiO3-PDMS triboelectric composites provide transient dynamic-event-sensitive FA outputs. The size, footprint, and embedding depth of the sensors were selected to generate distinct receptive-field-inspired response profiles within the 3D skin matrix. In vivo validation using rat sciatic nerve stimulation showed that encoded pressing and tapping patterns elicited distinguishable electromyographic (EMG) responses in the tibialis anterior (TA) and gastrocnemius (GC) muscles. These results support the feasibility of a functionally selective biomimetic tactile front-end for future closed-loop neuroprosthetic feedback, while further device-level durability, frequency-response characterization, and real-time integration remain necessary for clinical translation.
As intelligent systems increasingly rely on probabilistic inference and large-scale optimization, deterministic hardware faces intrinsic limitations in efficiently exploring complex solution spaces. Here, we present a physics-grounded probabilistic bit (p-bit) that serves as a hardware-native primitive for energy-efficient intelligent inference and optimization. The proposed p-bit exploits intrinsic stochastic electron capture in a multitrap ensemble at the Si-SiNx interface, converting nanoscale defect dynamics into a controllable probabilistic output fully compatible with standard complementary metal-oxide-semiconductor technology. A width-programmed gate-pulse scheme enables robust and continuous probability modulation by controlling trap occupancy through pulse duration rather than voltage amplitude, improving scalability and tolerance to interconnect nonidealities. We develop a physics-based analytical framework that quantitatively links multitrap capture kinetics to macroscopic drain-current statistics and implement it as a SPICE (Simulation Program with Integrated Circuit Emphasis)-compatible compact model, enabling direct cosimulation with conventional digital circuits. The resulting p-bit exhibits a Boltzmann-consistent sigmoid activation, supporting experimentally validated invertible logic and bidirectional probabilistic inference. Using experimentally calibrated characteristics, networks of these p-bits solve a 50-variable, 218-clause 3-satisfiability benchmark via controlled stochastic energy minimization, demonstrating system-level relevance for intelligent optimization. This work establishes a scalable pathway toward hardware-efficient probabilistic computing architectures for intelligent systems.
Convolutional neural network (CNN)-based computer-generated holography (CGH) enables high-speed, high-quality holographic display, yet its performance depends critically on the statistical characteristics of training datasets, and systematic frameworks for evaluating the performance boundaries of CNN-based hologram encoders remain lacking. From a dataset design perspective, this study proposes GM-4K, a 4K RGB-D dataset constructed through scene-customized geometric modeling. Using random geometric primitives, procedural textures, and parameterized spatial sampling, GM-4K generates scenes with controllable low-, mid-, high-, and wide-frequency intensity distributions, while also allowing flexible depth-region sampling. Numerical and optical experiments using a U-Net++ encoder demonstrate that the intensity spectral distribution of training data significantly influences the reconstruction quality of multi-depth phase-only holograms and that models trained on wide- or mid-frequency datasets exhibit better overall generalization in complex 3D scenes. Based on this observation, a spectral test framework comprising low-, mid-, and high-frequency scenes is developed to evaluate hologram encoding models under different frequency conditions. The proposed dataset construction paradigm and spectral test framework provide controllable data support for multi-depth hologram generation and offer practical guidance for data design and model evaluation in learning-based CGH.
Stretchable fabric sensors are a promising approach for smart wearable devices owing to their simple fabrication and low cost. However, current practical applications are limited by a lack of seamless integration among the sensor, the embedded platform, and the intelligent processing algorithm. To address this issue, this study proposes a co-design approach that integrates a stretchable fabric sensor, a resource-constrained embedded platform, and a lightweight machine-learning (tinyML) model. The sensor is fabricated from a graphite-based conductive ink-embedded stretch fabric, demonstrating stable performance with a sensitivity coefficient GF approximate to 218 and mechanical durability of up to 3000 working cycles, while maintaining a simple, low-cost fabrication process. Signals from multiple sensor channels are processed directly on the embedded device using a tinyML model with a compact memory footprint, making it suitable for systems with extremely limited resources (<1 MB of flash memory). When implemented in a wireless glove controller for mobile phone-based human interface device protocol, the ultra-lightweight convolutional neural network/random forest (tinyCNN/RF) model achieved up to 99.7/97.4% accuracy in on-device action classification. Experimental results show that the proposed co-design method optimizes system performance, reduces the impact of component-level noise, and broadens the application potential of edge AI-integrated smart wearable devices.
Conventional medical microcatheters, characterized by their stiff, continuous hollow tube design, encounter inevitable mechanical constraints when delivering therapeutic agents through tortuous, highly curved, and narrowing microvascular networks. To overcome this constraint, we introduce an acoustic microbubble-powered segmented delivery system. This microdevice utilizes oscillating microbubbles confined within discrete 3D-printed microchannels, interconnected via rotary microhinges. Under ultrasonic excitation, these series-connected bubble units generate directional acoustic streaming, enabling the targeted delivery of diverse microcargos. Tilted bubble arrays under ultrasound produce asymmetric vortices, breaking flow symmetry and establishing net streamlines that drive unidirectional movement across open channels. Comprehensive experimental and theoretical analyses reveal frequency-dependent flow regimes, with a resonant peak at 102 kHz achieving delivery velocities up to 630 nL/min at 40 V-pp. Geometric optimization reveals a linear correlation among delivery efficiency, acoustic voltage, and bubble quantity and identifies an optimal microchannel length that enhances performance by minimizing vortex interference. Critically, the system enables segmented acoustic delivery around sharp bends (>90 degrees), maintaining hydrodynamic continuity through modular units connected at variable interunit angles. A dual-bubble configuration achieves 100% delivery success across diverse materials and high-curvature conditions. This approach decouples delivery efficacy from catheter stiffness, providing a mechanically flexible and reconfigurable solution for complex luminal geometries.
Soft robotic arms (SRAs) featuring high flexibility and adaptability exhibit great opportunities in unstructured environments, making them an ideal manifestation of embodied intelligence. However, the lack of stable proprioception and integrated structural design limits their real-world applications. In this article, we propose a lightweight, modular shape memory alloy (SMA) driven SRA, integrated with a multi-Hall-magnet sensing system for accurate shape proprioception, while remaining untethered. Each segment is approximately 35 g in weight, allowing for easy assembly in series and seamless integration into complex systems. An omnidirectional shape proprioception method with self-correction is presented, and it maintains high perception accuracy even under pronounced nonconstant-curvature deformations. The results demonstrate a mean tip position error of approximately 1.39 mm in a single segment and a mean shape error of approximately 0.81% in the multisegment SRA. Experiments, including unstructured environment detection, pipeline navigation, and outdoor cooperation with a quadrotor drone, validate the robust performance of the sensing and actuation systems in real-world conditions. The results demonstrate that the precise shape perception and modular integration enable SRAs to achieve strong application performance, further facilitating the practical deployment in real-world scenarios.
