The purpose of service robots is to work and handle a wide range of objects in a human environment. Therefore, they must have the ability to grasp unknown objects with the appropriate force, by dynamically identifying their stiffness, mass, and surface-interaction properties. To ensure the successful grasping of unknown objects, it is significant to estimate the object response properties. In this paper, the cognitive functions to efficiently perform such operations with only tactile motion data and supervised learning using limited data, have been demonstrated. The study uses pinch-grasping performed by a 2-fingered robot experimental setup with a pre-designed motion sequence, over 7 different objects having a wide range of properties to capture the motion data. The proposed approach avails pre-processed motion data to train neural networks for accurately predicting object response properties for 3 unseen test objects. We have done a comparative study among three neural network architectures CNN, LSTM, and CNN-LSTM and we found that novel CNN-LSTM outperforms the other two models in terms of accuracy. The CNN-LSTM regression model achieves 0.98 and 0.97 of R2 value for the prediction of stiffness and mass, respectively and the CNN classification model achieves 94% of accuracy for the surface interaction classification.
|出版ステータス||Published - 2022 3月|
ASJC Scopus subject areas