Robust Neuromorphic Method for Tactile Recognition of Material Surfaces

Abstract

Tactile recognition enables humanoid robots to interact naturally and intelligently. Neuromorphic models, known for their robustness, efficiency, and low energy consumption, offer significant potential. Existing approaches primarily focus on normal force, neglecting other signals. Moreover, the imprecise anti-noise characteristics of tactile-oriented neuromorphic models hinder their optimal performance. To address this, we propose a neuromorphic approach that fuses tactile signals for material recognition, based on mechanoreceptor physiological responses and finger-surface interaction patterns. The method incorporates four models, each handling normal force, tangential force, velocity, and acceleration. Robustness is systematically evaluated at three levels: (1) analyzing the mean difference in spike firing rates for each model; (2) comparing noise resistance through the changing rate of membership degree; and (3) assessing the impact of signal-to-noise ratio on recognition accuracy. The results show that the method's resilience to blue, white, and pink noise decreases in this order. Compared to feature extraction methods, the neuromorphic approach demonstrates superior robustness. This research provides valuable insights for guiding multimodal tactile fusion recognition.