Advanced intelligent systems (Weinheim an der Bergstrasse, Germany)最新文献

Quadrupedal Mobile Manipulator

The image shows a quadrupedal mobile manipulator autonomously opening a door using the proposed method, illustrating its ability to perform complex and coordinated tasks in human environments. A novel trajectory generation and control framework that enhances end-effector tracking accuracy while expanding workspace and motion smoothness. This image also means the proposed method helps the quadrupedal mobile robot to open a new future. More details can be found in article 10.1002/aisy.202500242 by Jiawei Chen and co-workers.

Wave-based mechanisms inspired by traveling wave locomotion in animals have shown great potential use in robots to navigate unstructured environments. Herein, the dual actuator wave-like navigator (DAWN), a multisurface robot employing two actuated helical wave generators to produce continuous traveling waves on flexible link tracks enclosed in elastomer skins, is presented. These skins provide mechanical resilience, enhanced friction, and adaptability on uneven terrain. The robot demonstrates steering and controlled locomotion on flat surfaces, inclines, and declines. To characterize the robot, locomotion tests are performed on plywood, PMMA, and sand, achieving average linear speeds of 16.00, 15.76, and 1.63 mm s⁻¹, respectively. A key innovation is cyclic pneumatic actuation of the skins with actuation frequencies of 0.5 and 0.9 Hz, improving locomotion performance on sand to 2.22 and 2.70 mm s⁻¹. DAWN's capability to move on sand, grass, gravel, and wet soil is also demonstrated. Its modular design enables plug-and-play assembly of components including helical wave generators, flexible link tracks, and elastomer skins, allowing for easy maintenance, modification, and replacements. Potential applications include navigation in complex terrains for search and rescue, inspection, and environmental monitoring.

Traditional speech recognition methods rely on software-based feature extraction that introduces latency and high energy costs, making them unsuitable for low-power devices. A proof-of-concept demonstration is provided of a bioinspired tonotopic sensor for speech recognition that mimics the human cochlea, using a spiral-shaped elastic metamaterial. The measured modal response of the structure at different frequencies generates a spatially distributed signal, providing a spatiotemporal map of the input named “tonogram”. The device acts as an in-sensor physical reservoir computing system, working simultaneously as a sensor and as a computing unit, capable of extracting features of spoken words relevant to speech recognition. Results indicate that this can serve as a valid alternative to traditional software-based digital preprocessing, ensuring high accuracy in terms of classification, while reducing computational requirements. This work demonstrates the potential of bioinspired metamaterials for energy-efficient auditory sensing and, beyond speech recognition, for applications such as IoT devices and edge computing artificial intelligence systems.

Under restricted global positioning access, navigating micro aerial vehicles (MAVs) is particularly challenging. Therefore, the ability to autonomously estimate velocity and position based on onboard sensors becomes critical. While vision or radar-based approaches face limitations for MAVs due to payload constraints, poor lighting, or featureless environments, bio-inspired airflow sensing offers a promising alternative. Airflow interaction with MAVs provides continuous motion cues during flight, enabling airflow odometry—a feasible yet accuracy-limited solution. This article presents a novel skin-like airflow odometry system using distributed flexible thermal anemometers embedded in wingtips, allowing real-time motion estimation where conventional methods fail. Computational fluid dynamics is first employed to analyze the feasibility and sensitivity without changing their aerodynamic profile. A transformer-enhanced gated recurrent unit network then extracts high-precision airflow velocity, while model-based multisensor fusion reduces dead-reckoning drift and is validated through experiments. To the authors’ knowledge, this demonstrates the first airflow odometry system with positional accuracy surpassing previous research for MAV applications.

Biological materials heal, learn, and adapt thanks to the collective work of tiny agents acting at their microscale. Extensive research in robotics has tried to duplicate this scheme in a synthetic system, yet in their current centimeter-scale forms, the constituent robots are too large and too few, especially when compared to their biological inspiration. Here, this study shows a new type of high-stiffness, low-density material made entirely from robots of submillimeter dimensions. To bear load, these hundred-micrometer robots directly grow metal onto their bodies and bond together under the control of on-device microelectronics. The resulting aggregates achieve some of the lowest densities of any material and toughness/elastic moduli approaching the fundamental limits for metallic foams. Going beyond static properties, this study shows that robots can be used to actively repair the material microstructure, restoring stiffness and toughness following compressive fatigue. Broadly, these results clear the way for a new breed of programmable materials with bulk properties that can be rationally tuned over several orders of magnitude through the actions of robots too small to see with the naked eye.

Large-scale deep learning models are increasingly constrained by their immense energy consumption, which limits their scalability and applicability for edge intelligence. In-memory computing (IMC) offers a promising solution by addressing the von Neumann bottleneck inherent in traditional deep learning accelerators, significantly reducing energy consumption. However, the analog nature of IMC introduces hardware nonidealities that degrade model performance and reliability. This article presents a novel approach to directly train physical models of IMC, formulated as ordinary differential equation (ODE)-based physical neural networks (PNNs). To enable the training of large-scale networks, a technique called differentiable spike-time discretization is proposed, which reduces the computational cost of ODE-based PNNs by up to 20 times in speed and 100 times in memory. Such large-scale networks enhance learning performance by exploiting hardware nonidealities on the CIFAR-10 dataset. The proposed bottom-up methodology is validated through post-layout SPICE simulations on the IMC circuit with nonideal characteristics using the sky130 process. The proposed PNN approach reduces the discrepancy between model behavior and circuit dynamics by at least an order of magnitude. This work paves the way for leveraging nonideal physical devices, such as nonvolatile resistive memories, for energy-efficient deep learning applications.

Robotic needle steering plays a critical role in improving the precision and safety of percutaneous interventions across various clinical applications. However, manual needle steering remains challenged by operator-dependent variability, physiological tremor, and limited adaptability to dynamic tissue deformation. To address these limitations, this review examines recent advances in robotic needle steering, structured around three core components: 1) sensing for closed-loop needle steering, 2) modeling of soft tissue deformation and needle deflection, and 3) trajectory planning and closed-loop control strategies. Furthermore, emerging trends are discussed in artificial intelligence-driven autonomy and advanced biocompatible materials, highlighting their potential to enhance steering accuracy and real-time adaptability in future robot-assisted percutaneous procedures.

For the first time, a fully complementary metal–oxide–semiconductor process-compatible novel three-transistor (3-T) embedded NOR flash is demonstrated on a 28 nm fully depleted silicon-on-insulator platform. By introducing the band-band hot hole (BBHH) and channel hot electron (CHE) injection enabled by NMOS/PMOS pair coupling transistors, the proposed memory achieves record-fast 28-ns long-term potentiation (LTP) and depression (LTD), offering high-speed and highly reliable synaptic behavior for online training in neuromorphic hardware applications. The proposed 3-T embedded NOR flash demonstrates highly reliable operations: only 1.36% memory window degradation and 4.7% subthreshold swing degradation after 10⁷ program/erase (P/E) cycles. Moreover, its near-zero linearity LTP/LTD operations deliver a high classification accuracy of 92% on the Modified National Institute of Standards and Technology (MNIST) datasets. Finally, the proposed device shows an exceptionally short training latency of 2.68 s for 1M MNIST images, which confirms its suitability for high-performance training accelerators.

Insulin pumps typically use piston-based mechanisms with bulky transmission components to convert rotary motion into the piston's forward motion. These mechanical transmission systems and insulin reservoirs occupy more than one-third of pumps’ volume, significantly limiting miniaturization and making pumps cumbersome for daily use. Herein, a compact, magnetically actuated insulin pump is developed that is less than one-quarter the size of piston-based pumps. Instead of bulky mechanical components, the pump uses a magnetic soft actuator to directly compress the insulin chamber, controlled by a precisely tuned electromagnetic field. This innovative design eliminates the need for large transmission systems, enabling a notably smaller form factor. In addition, the fine-tunable magnetic actuation enables a 0.01 μL delivery resolution, significantly surpassing the 0.25 μL resolution of piston-based pumps. This high-resolution mechanism facilitates further miniaturization by allowing the use of high-concentration insulins, thereby reducing the reservoir size. By varying the magnetic field's waveform, amplitude, and duration, the pump's performance can be further enhanced. The reported magnetic insulin pump exhibits superior repeatability and accuracy across single-pulse, basal, and bolus modes compared to commercial insulin pumps. This miniaturized, high-resolution magnetic insulin pump is anticipated to substantially benefit people with diabetes by improving portability, precision, and cost efficiency.

Neural network (NN)-based transistor compact modeling is a transformative solution for accelerating device modeling and SPICE circuit simulations. However, conventional NN architectures primarily function as black-box problem solvers. This lack of interpretability significantly limits their capacity to extract and convey meaningful insights into learned data patterns, posing a major barrier to their broader adoption in critical modeling tasks. This work introduces Kolmogorov–Arnold network (KAN) for the transistor—a groundbreaking NN architecture that seamlessly integrates interpretability with high precision in physics-based function modeling. The performance of KAN and Fourier KAN (FKAN) is systematically evaluated for fin field-effect transistor (FinFET) compact modeling, benchmarking them against the golden industry-standard compact model and the widely used multi-layer perceptron architecture. The results reveal that KAN and FKAN consistently achieve superior prediction accuracy for critical figures of merit, including gate current (I_D), drain charge (Q_D), and source charge (Q_S). Furthermore, the unique ability of KAN to derive symbolic formulas from learned data patterns is demonstrated and improved—a capability that not only enhances interpretability but also facilitates in-depth transistor analysis and optimization. Additionally, we identify challenges inherent to KAN and FKAN architectures, which limit their ability to achieve a general-purpose SPICE simulation suitability.