Grasp stability prediction using tactile information

摘要

The prediction of grasp stability is very important in the fine-grained operation of a robotic manipulator, and it is also the research focus in the field of robot perception now. Most of the current methods to study this problem focus on how to identify the stability of the entire grasping process, while our goal is to predict whether the grasping process stable after the dexterous hand has caught the object and before lift it. The accurate prediction of the grasp stability can guide the posture control of the fingers of the dexterous hand and improve the success rate of the grasp. We solve this problem by the method of CNN (Convolution neural network). We introduce a specifically-designed tactile data collection and analysis module into the robot system, such that the robot can automatically detect the object location, grasp objects, determine the stability and repeat. Particularly, we convert tactile data in time series into image data to satisfy the input format of the CNN algorithm. We use different network structures to find the optimal configuration for grasp stability prediction through tactile data. Experimental results show that this model is available in the prediction of grasp stability, the CNN algorithm is also applicable to the identification of data in time series.