Object recognition based on tactile information: A generalized recognition network combining wavelet transform and transformer model for small sample datasets

Abstract

Tactile recognition is a crucial pathway for robots in perception and cognitive processing. While deep learning-based methods have shown excellent performance, training deep neural networks demands a substantial number of manually labeled samples. Unfortunately, current tactile recognition datasets lack the samples needed for robust training. To address this, we introduce a generalized tactile recognition method under low-sample conditions, Wave-Tactile-Transformer. Initially, we expand the tactile data using proposed TacGAN, avoiding traditional processes like rotation and cropping that create redundancy, which yields up to 5.8% accuracy improvement over traditional augmentation techniques. We also propose a Transformer framework integrated with multi-scale wavelet transforms, which is applicable to various tactile data formats. The wavelet transform enhances the model’s ability to discern details in tactile images, while the Transformer network refines the comprehension of feature relationships. This dual approach not only significantly reduces computational costs but also boosts object recognition accuracy. Our approach introduces an innovative framework that harmonizes the processing of tactile data across diverse formats. Cross-format tactile dataset experiments achieved a peak recognition accuracy of 96.7%, outperforming conventional CNN-based methods by up to 2.6%, surpassing previous methods. This generalized tactile recognition network offers innovative solutions for robotic tactile perception and grasp control.