Object classification based on piezoelectric actuator-sensor pair on robot hand using neural network

We propose a piezoelectric actuator-sensor pair applicable to object classification, which comprises two piezoelectric films on a polyethylene terephthalate substrate layer. One piezoelectric film is used as an actuator, and another as a sensor. The actuator-sensor pair is installed on a robot hand, and a neural network classifier is used to conduct object classification. Sensor data for learning process are acquired when the actuator is oscillating, while the robot hand grasps objects. Specifically, a sinusoidal input voltage with frequency sweep is supplied to the actuator, and the sensor outputs the signal transferred from the actuator simultaneously. The obtained data undergoes a series of preprocessing procedures to be used as data input for learning. Several neural network classifier models are trained with the preprocessed dataset, and the most suitable model is selected as our classifier. Our classifier successfully predicted the object data in the test set. Furthermore, we develop a real-time recognition system and demonstrate the feasibility of the actuator-sensor pair.