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

Plastic pollution in water bodies threatens and disrupts aquatic life, requiring effective cleanup solutions. This paper proposes a strategy for plastic cleanup using a fleet of autonomous surface vehicles in a multitask scenario, with a focus on both exploration and cleaning tasks. The mission is decoupled into two phases: an exploration phase for locating trash and a cleaning phase for collection. A Multitask Deep Q-Network with two heads estimates Q-values for each task, and all ASVs share the same policy through an egocentric state formulation to enhance scalability. A multiobjective learning approach is applied, resulting in distinct policies that balance the duration of the exploration and cleaning phases, leading to the construction of a Pareto front, which provides a visual representation of trade-offs between task priorities. The framework adapts to various environmental conditions, demonstrated in both the larger Malaga Port and the smaller Alamillo Lake. The study also highlights the importance of a dedicated exploration phase for larger areas, while minimal exploration is sufficient for smaller spaces. Compared to the decomposition weighting sum strategy, the approach consistently produces superior Pareto-optimal policies, ensuring broader and more effective exploration of the objective space.

Multiconfiguration quadruped robots offer enhanced adaptability and mobility by adopting diverse locomotion morphologies inspired by different biological counterparts. However, their control frameworks remain insufficiently explored. In this paper, a control framework is proposed that integrates a nonlinear constraint-based whole-body planner with a hierarchical whole-body controller. This framework enables configuration switching and multiconfiguration motion on complex terrains. The motion planner formulates an optimal control problem that incorporates nonlinear constraints, such as reference poses, system dynamics, and foothold positions, to compute optimal states and control inputs. Depending on the terrain, the most suitable configuration of the robot is selected, such as mammal-like or reptile-like, to effectively traverse complex environments and obstacles. The hierarchical whole-body controller accurately tracks prioritized motion tasks of the torso and limbs through QP-based optimization. The proposed framework is validated through real-world experiments, demonstrating robust and versatile performance across challenging scenarios, including slopes, irregular brick terrain, overhanging obstacles, and dual-platform bridges.

The urinary bladder is considered a highly complex organ, capable not only of storing urine but also of sensing intra-vesical volume and dynamically expanding and contracting. Consequently, fully replicating its functions following radical cystectomy remains a significant technological challenge. Hereinafter, an implantable robotic bladder is presented that can change shape and expand its internal volume up to 400 mL, based on the amount of urine collected from kidneys, and monitor the volume in real-time. It can apply on-demand mechanical compression to assist urination, by means of an origami-designed enclosure, coupled to miniaturized mechatronic components. In vitro characterization in a human phantom is demonstrated, and volume monitoring is validated following a realistic filling routine. The tests demonstrate successful expansions for collecting urine, with an average volume reconstruction error of 8.4 ± 6.1 mL, and then 99% of the volume is voided in less than 2 min. The work paves the way for developing active robotic solutions and reproducing bladder functions in patients with cancer and organ removal or impairment.

Self-adaptive, easy-to-control, and low-cost gripper devices are indispensable in manufacturing and agriculture. However, the existing soft grippers cannot provide high response speed and firm grasping. Inspired by the natural, active entanglement behaviors of animals and plants, a rapid, cilia-like soft gripper design is proposed for grasping various objects via envelopes formed by the self-entanglement of multiple hollowed silicone tubes. The basic entanglement unit comprises a hollow, soft silicone tube with an actuation wire inside, which leads to compression and entanglement by fixing the front of the tube and drawing the actuation wire. Using multiple entanglement units enables sufficient mechanical interlocking between deformed tubes and grasped objects, avoiding the reliance on contact force control. Experimental results demonstrate that the developed soft gripper, with a cost lower than one dollar, can complete adaptive grasping within 1 s. The grasping success rate can reach 100% in grasping common irregular-shaped daily objects within the effective grasping range of the entanglement units. The design paves the way for harnessing the potential of embodied intelligence in soft robots, enabling fast and universal grasping.

Deterministic pseudorandom number generators used in generative artificial intelligence (GAI) models produce predictable patterns vulnerable to exploitation by attackers. Conventional defenses against the vulnerabilities often come with significant energy and latency overhead. Herein, hardware-generated true random bits from spin-transfer torque magnetic tunnel junctions (STT-MTJs) are embedded to address the challenges. A highly parallel, field-programmable gate array-assisted prototype computing system delivers megabit-per-second true random numbers, passing NIST randomness tests after in situ operations with minimal overhead. Integrating the hardware random bits into a generative adversarial network trained on CIFAR-10 reduces insecure outputs by up to 18.6 times compared to the low-quality random number generators (RNG) baseline. With nanosecond switching speed, high energy efficiency, and established scalability, the STT-MTJ-based system holds the potential to scale beyond 10⁶ parallel cells, achieving gigabit-per-second throughput suitable for large language model sampling. This advancement highlights spintronic RNGs as practical security components for next-generation GAI systems.

Interventionalist catheterization actions are essential for assessing tool navigation quality and procedural competence during interventions. Traditional assessment methods are subjective, lack immediate feedback, and limit timely performance improvement. To address these limitations, this study introduces a deep-learning framework designed to systematically analyze catheterization action data, address inherent class imbalances, and enable real-time action recognition. First, the proposed framework leverages advanced generative models to augment minority action classes, thus enhancing data representation and ensuring accurate recognition of catheterization actions. The six generative models utilized in this study undergo rigorous evaluation, achieving high fidelity with average precision and F1-scores exceeding 94% across all models except CTGAN. Second, a convolutional neural network (IAR-Net) tailored to recognize seven distinct catheterization actions is developed. Evaluated using the augmented dataset, IAR-Net achieves an accuracy of 98.9%, surpassing current benchmarks. Comparative analysis with state-of-the-art machine learning and transformer-based models designed for tabular data confirms IAR-Net's performance and robustness in recognizing catheterization actions. Lastly, interpretability methods are incorporated to elucidate the model's decision-making process, improving understanding and increasing the trustworthiness of predictions. These outcomes offer a promising avenue for enhancing trainee assessment and training protocols, thereby accelerating the acceptance and integration of robot-assisted endovascular systems into clinical practice.

This study presents a new method for measuring the effective stiffness of curved actuators. Actuators are loaded into tension, and analytical mechanical equilibrium formulations are used to determine the stress along the actuator. A new mechanical metric, Shape Actuation Modulus (SAM), defines the effective stiffness of the actuator during loading as the ratio of stress change to radius of curvature change. Conductive polylactic-acid shape-memory actuators are produced to benchmark this novel methodology. These actuators display a linear behavior between 25 and 50 mm radius of curvature with SAM of 3.8±0.9 MPa at 50 mm. The interval on which the radius of curvature to stress relationship is linear can be controlled by choosing the radius of curvature of the hinge. For instance, SAM calculation with R² > 0.97 was achieved in ranges of [22.7;79.6] mm and [16.4;51.5]mm for starting radius of curvature of 23.5±0.7 mm and 17.2±0.6 mm, respectively. Hence, the new technique proposed provides guidelines to design actuators. Finally, a comparison of bio-composite actuators made of the same material was conducted. The hygromnemic actuators tested displayed a stiffness more than one order of magnitude larger than the hygromorphic ones for the range of radius of curvature [20;100]mm.

This study investigates the application of Riemannian geometry-based methods for brain decoding using invasive electrophysiological recordings. While Riemannian geometry has been successfully applied in noninvasive settings, its utility for invasive datasets, which are typically smaller and scarcer, remains less explored. Herein, a minimum distance to mean (MDM) classifier is proposed using a Riemannian geometry approach based on covariance matrices extracted from intracortical local field potential (LFP) recordings across various regions during different brain state dynamics. For benchmarking, the performance of the approach is evaluated against convolutional neural networks (CNNs) and Euclidean MDM classifiers. The results indicate that the Riemannian geometry-based classification not only achieves a superior mean F1 macro-averaged score across different channel configurations but also requires up to two orders of magnitude less computational training time. Additionally, the geometric framework reveals distinct spatial contributions of brain regions across varying brain states, suggesting a state-dependent organization that traditional time series-based methods often fail to capture. The findings align with previous studies supporting the efficacy of geometry-based methods and extend their application to invasive brain recordings, highlighting their potential for broader clinical use, such as brain-computer interface applications.

Kirigami, or “jianzhi” in Chinese, is an art in paper-cutting. Using simple tools like scissors, artisans transform paper into intricate designs featuring flowers, animals, or characters (e.g., “囍”). Nowadays, kirigami has emerged as a particularly promising design strategy in engineering. This method involves creating systematic cut patterns on thin, planar sheets, which enables complex mechanical responses by changing dimensions, thereby offering innovative solutions for the development of metamaterials, soft actuators, and robotic systems. The concept of the integration of ancient art and modern science and technology has injected vitality into the development of many disciplines and become the forefront of interdisciplinary research. This review provides a systematic review of recent progress on the design of kirigami and applications in diverse robotic prototypes. The kirigami begins by classifying into two categories from a compliant mechanism perspective, and then it examines the distinctive mechanical properties that altered by cut patterns, followed by reviewing the design of the two types of kirigami. Next, the kirigami-inspired kinematic metamaterials is examined. Finally, applications in soft actuators and robotic systems is demonstrated. The integration of design methods, fabrication techniques, materials research, mechanics modeling, and control systems will further advance this emerging field.

Neuromorphic devices, inspired by the human brain's efficiency and adaptability, hold great potential for artificial intelligence (AI) hardware to overcome the limitations of traditional von Neumann architecture. As a subclass, multimodal and multifunctional neuromorphic devices have recently gained a lot of attention due to their advantages in in-sensor computing and sophisticated behaviors. In this review, recent advances in materials, device structures, and applications in this field are systematically presented. It includes optical, electrical, mechanical, and chemical sensing in multimodal neuromorphic device, which enable in-sensor computing to minimize energy consumption and enhance real-time decision-making. The materials applied in this field such as phase-change, 2D materials, and ferroelectrics are summarized for their roles in achieving synaptic plasticity, nonvolatile memory for multifunctional neuromorphic devices. Structural innovations, including reconfigurable, multi-terminal, and 3D-integrated designs, further optimize parallel processing and multifunctional integration. Besides, application scenarios of multimodal and multifunctional neuromorphic devices and their advantages for improving the efficiency of AI are reviewed. Finally, challenges in material stability and commercialization are discussed, it emphasizes the need for interdisciplinary efforts to bridge the gap. This review provides critical insights and future directions for developing brain-inspired, energy-efficient AI hardware.