Multi-Scale Voting System for Robotic Tactile Texture Recognition on Uneven Surfaces

Tactile perception is vital for enabling robots to interpret and interact with complex physical environments. We propose a novel classification framework for recognizing surface textures on uneven terrains using time-series tactile data. Un-like prior approaches that rely on single-window classification and assume flat surfaces, our method introduces a multi-scale, voting-based ensemble that integrates predictions from multiple temporal resolutions (128, 256, and 512 samples). Each segment is independently processed by deep neural networks trained to capture both short- and long-range temporal features. Final decisions are generated via an optimized weighted voting strategy, which assigns importance to each classifier based on validation performance. Our system is trained and evaluated on a new dynamic tactile dataset collected using an arm robot interacting with 12 real-world materials under uneven conditions. The proposed approach achieves a classification accuracy of 94.22 %, outperforming the best single-window baseline (92.11 %). This work contributes a new challenging dataset, establishes a strong baseline for texture recognition on non-planar surfaces, and demonstrates the effectiveness of multi-scale temporal modeling with ensemble voting in tactile perception.