Tactual Recognition of Soft Objects From Deformation Cues

Recognition of soft objects is a critical step to tailor policy for dexterous manipulation. Deformation property is one of the valuable tactile cues for inferring the identities of soft objects, especially those having similar appearance features. In this work, we implemented tactual grasping for discriminating soft objects from the response curves of indentation displacement in relation to grasping force. Unlike some existing methods defining certain local features from the force-displacement observations, we viewed the force-displacement curve as continuous function rather than discrete observations. We proposed to use functional data analysis for classifying the kind of force-displacement curve data. Functional principal component analysis (FPCA) was used to extract multivariate features from force-displacement curves. Different traditional machine learning models were trained for classifying the multivariate features and the best one was selected based on cross-validation. Case study demonstrated that our method could accurately distinguish eight different soft objects. The FPCA based features outperformed those manually defined features. We indicated that the force had crucial influence on the classification accuracy and proposed how to select the proper force for the classification task. Overall, this work provides practical guidelines on tactual recognition of soft objects.