Learning to Predict Friction and Classify Contact States by Tactile Sensor

In real-world grasping tasks, unstable grasping will lead to sliding between the object and the robotic nd-effectors. When the target object owns special properties (e.g. glasses, etc.), sliding may cause damage to both the object and the grasping environment, so it is necessary to predict and classify whether sliding occurs at the moment of grasping. Aiming at predicting sliding, this paper proposes a sliding prediction solution for time series tactile data obtained by crude sensor Xela and designs a novel neural network: the force motion tracking network. It predicts the variation trend of the frictional force at the moment of contact, and then concatenate the collected and predicted friction force data as the input of the LSTM network to classify the contact state (whether there is slippage). In this paper, 660 sets of tactile time series data are collected, and we process high dimensional tactile time series data into video data. This data processing method can be applied to other similar tactile sensors. Meanwhile, we also verify the proposed model, the mean square error of our force tracking network is much smaller than ConvLSTM, and the prediction accuracy of our network can reach to 93.5%,which is higher than other methods.