Since tactile sensing provides rich and delicate sensations, touch-based object recognition has attracted public attention and has been extensively developed for robots. However, robotic grasping recognition in real-life scenarios is highly challenging due to the complexity of real-life objects in shapes, sizes, and other details, as well as the uncertainty of real grabs in orientations and locations. Here, we propose a novel robotic tactile sensing method, utilizing the spatiotemporal sensing of multimodal tactile sensors acquired during hand grasping to simultaneously perceive multi-attributes of the grasped object, including thermal conductivity, thermal diffusivity, surface roughness, contact pressure, and temperature. Multimodal perception of thermal attributes (thermal conductivity, diffusivity, and temperature) and mechanical attributes (roughness and contact pressure) greatly enhance the robotic ability to recognize objects. To further overcome the complexity and uncertainty in real-life grasping recognition, inspired by human logical reasoning “from easy to hard” in solving puzzles, we propose a novel cascade classifier using multilayered long short-term memory neural networks to hierarchically identify objects according to their features. With the enhanced multimodal perception ability of tactile sensors and the novel cascade classifier, the robotic grasping recognition achieves a high recognition accuracy of 98.85% in discriminating diverse garbage objects, showing excellent generalizability. The proposed spatiotemporal tactile sensing with logical reasoning strategy overcomes the difficulty of robotic object recognition in complex real-life scenes and facilitates its practical applications in our daily lives.
Artificial neural networks based on artificial synaptic thin-film transistors (ASTFTs) offer promising opportunities for efficient and scalable information processing. However, implementing both spatial and temporal processing within a unified system requires complex hybrid network architectures, placing high demands on the tunability of individual synaptic devices. Here, we report a reconfigurable photosensitive synaptic transistor composed of FAPbI₃ colloidal quantum dots (CQDs) and an InO x channel, integrated with a modulation scheme tailored for hybrid convolutional-sequence neural computing. The device enables mode-specific plasticity through programmable optical and electrical stimuli. For spatial signal processing, it achieves a long-term potentiation (LTP) linearity of 0.062, long-term depression (LTD) linearity of 0.143, and an asymmetry ratio (AR) of 0.185. For temporal learning, the dynamic fading memory (FM) time constant is tunable from 42 to 75 ms, enabling dynamic short-term memory (STM) behavior. Utilizing these synaptic features, we construct a hybrid CNN-GRU neuromorphic system that achieves 94.2% accuracy in real-time gesture recognition within only 7 training epochs, using a custom dataset over 200 times larger than conventional benchmarks.This work demonstrates a unified neuromorphic hardware platform capable of spatiotemporal cognition, offering new strategies for the development of intelligent optoelectronic systems.
危险品化学品经营许可证（带存储）