Predict Tactile Grasp Outcomes Based on Attention and Low-Rank Fusion Network

Abstract

Tactile measurement endows robots with the ability to interact with the environment, which is crucial for accurately predicting grasp outcomes. However, this field has some areas for improvement, particularly regarding the key feature extraction and efficient fusion of multiple tactile modality features. To address these issues, we propose an advanced measurement technique that uses a tactile attention (TacAtt) module and a tactile low-rank tensor fusion (TLRTF) module to enhance the measurement and evaluation capabilities of multiple tactile sensors. By integrating the TacAtt module into the convolutional neural network (CNN), our model enhances the feature extraction capabilities for multiple tactile signals, focusing more on the contact area of the object, which provides highly targeted tactile input features for subsequent feature fusion. Moreover, the TLRTF module successfully addresses the challenges of insufficient fusion and redundant information in traditional concatenation methods when integrating features from multiple tactile sensors. The combination of the two proposed modules forms a strong system for tactile feature extraction and fusion. Our model achieves an accuracy of 78.31% on the public tactile dataset, which represents a significant improvement of 3.16% over the baseline model, thus validating the effectiveness and superiority of our model.