Ranking-Based Learning for Robust Tactile Roughness Perception under Domain Shift

Tactile roughness perception is commonly formulated as either classification or regression, implicitly assuming that tactile signals can be grounded to a stable absolute scale. In practice, however, tactile measurements are highly sensitive to sensing variability, making this numerical grounding unreliable under limited supervision or domain shift.In this paper, we investigate how the choice of learning objective affects robustness when absolute grounding is unstable. We compare classification, regression, and ranking-based learning on a sandpaper dataset with 5 grit levels and 9,595 samples. Under well-conditioned in-domain settings, all paradigms achieve comparable performance. However, when supervision is scarce (e.g., 5% training data) or when substantial domain shift is introduced through changes in sensor hardware and exploration protocol, their behavior diverges.Under domain shift, regression exhibits a pronounced collapse in ordinal consistency, whereas ranking-based learning preserves relative ordering more effectively, outperforming regression by 0.290 in Spearman rank correlation. These findings suggest that directly optimizing ordinal consistency yields more robust generalization when tactile roughness labels provide reliable ordering but unstable numerical scales.