Tactile texture classification on uneven surfaces using a neural network soft voting ensemble

Abstract

With the growing capabilities of intelligent robots in object recognition and manipulation, the ability to sense and interpret physical contact through touch has become a crucial component to enabling effective interaction with the physical world. Although tactile texture classification on flat surfaces has been broadly studied in recent years, uneven surfaces pose additional challenges due to variations in contact geometry and surface normals. To address these challenges, this study introduces a new tactile texture dataset comprising both flat surfaces and several distinct uneven surfaces, and proposes a soft voting-based classification system built on deep neural networks, which combines predictions from multiple temporal window sizes to improve robustness. The dataset is collected using a compliant tactile sensor mounted on the end effector of a UFactory Lite6 robotic arm that combines MARG and barometric data for capturing dynamic contact interactions. The dataset includes six types of uneven surfaces, each including a variety of textures to create diverse and challenging contact conditions. To improve classification robustness and enable multi-scale analysis, the time-series data are segmented using a sliding window approach with varying window sizes. Multiple model architectures are trained on the windowed segments, including 1D Convolutional Neural Networks (1D-CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, hybrid 1D-CNN–BiLSTM models, self-attention-based networks, and hybrid 1D-CNN–self-attention models. Their predictions are combined using a soft voting strategy to enhance overall classification accuracy. Experimental results based on 5-fold cross-validation demonstrate that self-attentionbased models consistently outperform other individual architectures across all window sizes. Moreover, the proposed voting system, which combines predictions from different window sizes, further improves classification performance for all model types by leveraging complementary temporal features. This study demonstrates that combining deep neural networks with a soft voting mechanism across multiple window sizes enables accurate tactile texture classification on various types of uneven surfaces, contributing toward more robust and adaptable robotic perception in complex environments.