Object Attribute Recognition Based on the Flexible Tactile Sensors by the DWT-TCN Model

Through tactile sensing, robots can obtain attribute information such as stiffness, diameter size, and material of the object to achieve precise positioning and perform various operation tasks. To simplify the complexity of robotic object grasping and sensing, while improving sensing accuracy and avoiding damage to objects, we fabricated piezoresistive pressure sensor and piezoresistive strain sensor based on multi-walled carbon nanotubes/porous flexible sponges (MWCNTs/PDMS) conductive composites. The fabricated sensors were installed on the actuator of a robotic arm to sense stiffness and diameter size attributes of objects. A combination of the Discrete Wavelet Transform (DWT) and Temporal Convolutional Network (TCN) methods was utilized to improve the accuracy of object recognition. Initially, the DWT method was employed to extract time-frequency features from temporal data collected by the sensors for 10 different types of stiffness and diameter sizes, and the TCN model was constructed for the prediction of feature sequences. An average recognition accuracy of up to 98.57% was achieved by the constructed TCN model in the classification of the objects. The results demonstrate that the piezoresistive sensors developed in this study effectively sense the multi-attribute features of different objects in terms of stiffness and diameter size. Additionally, the constructed TCN model proves highly effective in discriminating objects with varying stiffness and diameter sizes.