Robot Grasping Stability Prediction Network based on Feature-fusion and Feature-reconstruction of Tactile Information

Abstract

Grasping stability prediction relies on the contact force between robot end-effector and objects, which can be precisely measured by tactile sensor array. To achieve robust robot grasping stability prediction for different shapes of objects, a novel convolutional neural network structure is proposed to fuse features of different receptive fields and adaptively reconstruct the channel-wise features. First, baseline model based on a relatively shallow convolutional network is built up. Second, considering the spatial distribution of tactile sensor array, three parallel convolutional kernels are introduced into the first convolutional layer. Then data augmentation, optimizations by batch normalization and fully convolutional layer application are involved into the network to achieve better performance. Comprehensive experiments are conducted on the corresponding dataset. It achieves the accuracy, mAP and AUC of (96.75%, 99.23% and 99.41%), with an absolute improvement of (2.51%, 1.32% and 1.12%) compared with the baseline of traditional convolutional neural network. The efficiency and robust are also verified with the feature-fusion and feature-reconstruction based network in predicting robot grasping stability.