A Deep Learning Method Based on Triplet Network Using Self-Attention for Tactile Grasp Outcomes Prediction

Abstract

Recent research work has demonstrated that pregrasp tactile information can be used to effectively predict whether a grasp will be successful or not. However, most of the existing grasp prediction models do not perform satisfactorily with a small available dataset. In this paper, we propose a deep network framework based on triplet network with self-attention mechanisms for grasp outcomes prediction. By forming the samples into contrasting triplets, our method can generate more sample units and discover potential connections between samples by contrasting with the triplet loss. In addition, the inclusion of the self-attention mechanisms helps capture the internal correlation of features, further improving the performance of the network. We also validate that the self-attention module works better as a non-linear projection head for contrast learning than the multilayer perceptron module. Experimental results on the publicly available dataset show that the proposed framework is effective.