Leveraging Time-Frequency Features for Contact Classification and Regression with a Piezoelectric Tactile Skin for Robotic Fingertips

Sensory integration within closed-loop control systems is paramount in robotic grasping to replicate human hand capabilities and improve dexterity. Object manipulation may be decomposed into distinct action phases where accurate sensory feedback, such as contact forces and slippage detection, is crucial for tasks like grip strength regulation and grasp stabilization. This study explores the use of piezoelectric tactile sensors to perform classification and regression of contact force variations and the occurrence of slippage events. Experiments were conducted to characterize a Piezoelectric Tactile Skin (PTS) in terms of its response to different applied force values and sliding motions. Specifically, the extraction of various features (raw tactile signals, Short-Time Fourier Transform, and Discrete Wavelet Transform marginals) from PTS signals was investigated in conjunction with machine learning algorithms, i.e., Support Vector Machine (SVM) and Neural Network (NN). Properly designed acquisition tasks have been performed using both the PTS worn on a human fingertip and on the fingertip of an anthropomorphic robotic hand. In both experimental scenarios, a force sensor was exploited to sensorize a surface on which the actual normal forces applied by the fingertips were accurately measured and used to train the machine learning algorithms. Preliminary results demonstrate that wavelet-based features achieve superior performance in both contact classification and regression, highlighting the importance of multiresolution time-frequency analysis in interpreting complex piezoelectric raw tactile signals.