Yuekun Yang

Real-time sensing and processing of a large amount of tactile information is essential for intelligent robotics and wearable technology. However, physical separation between sensors and processors in the traditional tactile sensing scheme makes these functionalities inaccessible, posing a major roadblock to the rapid advance of skinomorphic electronics. Here, we propose a massively parallel in-sensor skinomorphic computing scheme and demonstrate its promising applications in intelligent tactile perception. This scheme allows for achieving parallel sensing and processing of tactile information directly within sensor. We implement this proposed scheme by fabricating a 32×32 flexible capacitive pressure sensors array with excellent uniformity and endurance, and by cascading the sensors array with a memristive crossbar array. We experimentally demonstrate that the broken pressure patterns of the letter ‘NJU’ loaded on the sensors array can be sensed and restored in parallel, which is inaccessible with previously reported tactile technologies. Moreover, by networking the pressure sensors array with two memristive crossbar arrays, we show that textural features of the loaded complex pressure patterns can be directly extracted in a parallel manner and the tactile information can thus be compressed. Our work opens up an avenue for developing intelligent skins capable of real-time and high-throughput tactile perception.

Mechanical computing represents a highly promising paradigm for environment-adaptive information processing. However, existing implementations are generally constrained by limited architectural scalability, and their modes of application in practical scenarios remain insufficiently defined. Here, we develop a light-programmable mechanical computing system that not only performs scalable logic operations but also enables environment-adaptive optical camouflage. The system is based on a polyaniline composite film (PCF) that integrates light-responsive expansion–contraction elements with a flexible conductive layer. Light illumination dynamically modulates the conductive pathways, giving rise to optically controlled single-pole single-throw (SPST) and single-pole double-throw (SPDT) relays that reconfigure signal transmission routes. Interconnecting these relays enables the construction of basic logic gates and 2-bit full-adder circuits, establishing a scalable paradigm for light-programmable mechanical computation. Moreover, we implement an adaptive camouflage function that senses environmental textures and generates matching optical patterns, demonstrating potential for intelligent skin applications capable of environmental interaction. This work establishes a light-programmable, pathway-reconfigurable mechanical computing framework, expanding possibilities for autonomous and adaptive intelligent systems.

The rapid growth of artificial intelligence (AI) is increasingly constrained by fundamental hardware bottlenecks in computation throughput and energy efficiency. Bioinspired computing (BIC) offers a promising alternative by emulating the intrinsic advantages of biological systems, such as parallelism, adaptability, and robustness. Progress in BIC hardware demands interdisciplinary convergence that bridges materials science and device physics with neuroscience, computer science, mathematics, and information science. Therefore, the development of this cross-disciplinary field urgently requires a comprehensive roadmap that analyzes systematically and in-depth the frontier issues and the latest progress. In this roadmap, we categorize the critical challenges into three components: hardware foundations, architectures, and prototype realizations. We highlight how biological features inspire the design of BIC hardware through device physics and discuss their performance metrics and engineering challenges. We then describe how diverse signaling rules and structural organizations in BIC architectures support specific computational prototypes, including electronic and photonic BIC chips, and present a technological roadmap that outlines opportunities to expand the functional scope of BIC hardware through coordinated advances in devices, architectures, and system demonstrations. This ongoing convergence of interdisciplinary knowledge can help accelerate the shift toward high-efficiency AI hardware.

Conventional machine vision systems acquire visual data frame by frame and transmit all the data to the backend for processing, which induces redundant data movement and a substantial processing load, especially for real-time detection of moving objects. In this work, we fabricate an MoS2/h-BN/PdSe2 heterostructure featuring infrared-visible dual-band rapid and relaxation responses to reduce redundancy and improve efficiency for motion detection. Building on these properties, we realize an in-sensor two-layer motion-prediction network at the device level by exploiting the tunable rapid infrared photoresponse and the relaxation visible photoresponse. The first layer based on rapid infrared photoresponse performs target classification and imposes a coarse constraint on possible motion directions. The second layer based on the relaxation visible photoresponse constitutes a physical reservoir that maps optical inputs to photocurrents carrying temporal information for motion-direction prediction. By fusing the device’s dual-band optoelectronic features with algorithmic design, we achieve accurate motion detection and prediction directly in-sensor with reduced data transmission and computational load.

Memristor-based computing-in-memory (CIM) architectures have emerged as a promising solution for enhancing the computational efficiency of deep neural networks (DNNs). However, memristors are inherently vulnerable to non-idealities, including manufacturing-induced variations and operational fluctuations. These non-idealities cause deviations in programmed weights, compromising the computational accuracy of neural networks. Although software-based approaches exist to mitigate these issues, they typically target specific error types or require additional components and specialized architectures. These requirements significantly limit the scalability and broader adaptation of CIM systems. In this study, we propose a multi-teacher robust distillation training framework to address the discrepancies between analytically derived training information and the imprecision inherent in analog devices. This framework offers a generalized solution that is independent of network architecture and task type, making it applicable to a wide range of tasks, including image classification, object segmentation, and image denoising. Moreover, it can be seamlessly integrated into models exhibiting various non-ideal behaviors. The proposed method significantly mitigates accuracy degradation in memristor-based CIM systems, paving the way for more reliable and scalable deployment in real-world applications. Experimentally, we created a one-transistor-one-memristor (1T1R) chip to verify the classification and denoising tasks. The results demonstrated that, in a statistical distribution with a standard deviation of 0.5 in weight variations, the accuracy rate was 33.7% higher than that of nominal networks on classification with CIFAR-10. Compared to other variation-aware algorithms, it also achieved the best performance and generalization ability.

Conventional computer systems based on the Von Neumann architecture rely on silicon transistors with binary states for information representation and processing. However, exploiting emerging materials’ intrinsic physical properties and dynamic behaviors offers a promising pathway for developing next-generation brain-inspired neuromorphic hardware. Here, we introduce a stable and controllable photoelectricity-induced halide-ion segregation effect in epitaxially grown mixed-halide perovskite CsPbBr_1.5I_1.5 microwire networks on mica, as confirmed by various in-situ measurements. The dynamic segregation and recovery processes show the reconfigurable, self-powered photoresponse, enabling non-volatile light information storage and precise modulation of optoelectronic properties. Furthermore, our microwire array successfully addressed a typical graphical neural network problem and an image restoration task without external circuits, underscoring the potential of in-material dynamics to achieve highly parallel and energy-efficient physical computing in the post-Moore era.

Noncooperative target detection in real-world scenarios relies on large-scale deep neural networks after image capture. However, directly implementing this detection pipeline under conventional optoelectronic sensors and computing units leads to physical bottlenecks in latency and energy consumption. Here, inspired by the biological visual attention mechanism and leveraging fabricated two-dimensional optoelectronic van der Waals heterostructure devices, we present a highly efficient neuromorphic in-sensor target detection pipeline. The inherent physical process of infrared self-driven visible photoresponse in heterostructures is used to simplify the originally complex processing in artificial intelligence (AI) optical detection algorithms. Specifically, the high-cost target localization and image fusion process can be directly implemented in the sensing unit. This manipulation decreases redundant information at the sensor level, reducing the burden on data transmission and backend computation. The results show that our bioinspired pipeline achieves a mean average precision (mAP) of 95.85% in detecting real-world scenes, even in extreme environments. Meanwhile, significant reductions in computing load (31.65%), latency (95.66%), and energy consumption (21.25%) can be attained compared with previous research. Our work provides a scalable material-for-AI solution for real-time, highly efficient target detection applications in real-world settings.