Flexible Piezoelectric Tactile Sensing System for Intelligent Object Recognition

By providing direct information about the physical properties of objects (e.g., texture and shape), tactile sensing often outperforms vision-based solutions in enabling safe, efficient grasping to enhance robot-object interaction. Numerous approaches to object recognition have been proposed; however, reliable recognition during the early stage of grasping under strict real-time constraints remains a significant challenge. In this study, we propose a flexible, distributed piezoelectric tactile sensing system designed for early-stage and real-time object recognition in robotic grasping. The system was experimentally validated using six representative daily-life objects. Tactile signals acquired during the robot grasping of objects were leveraged to develop and compare three recognition approaches: (i) feature- based learning using Support Vector Machines (SVM), (ii) raw signal learning using 1D Convolutional Neural Networks (1D-CNN) and Convolutional Recurrent Neural Networks (C- RNN), and (iii) a spatio-temporal heatmap representation processed with a 2D Convolutional Neural Network (2D-CNN). Moreover, a comprehensive quantitative analysis was conducted to evaluate the influence of sensor count and temporal window size on recognition performance and real-time feasibility. Results demonstrated that increasing the number of sensing elements substantially enhances the recognition performance of all approaches. For the best-performing approach, specifically the spatio-temporal heatmap representation combined with the 2D-CNN, the F1-score improved from 0.9796 (16 sensors) to 0.9846 (24 sensors), reaching 0.9903 when all 32 sensors were included. The optimal configuration (32 sensors, 100 ms temporal window, corresponding to 50 tactile samples) was deployed on an embedded edge platform, achieving an end-to-end recognition latency of ≈ 101 ms from contact onset, with an energy consumption of 63.8 μJ per inference. These results underscore the effectiveness of the proposed distributed flexible tactile sensing system for fast, embedded, and early-stage object recognition, highlighting its strong potential for real-time robotic manipulation applications.