Handling Scan-time Parameters in Haptic Surface Classification

摘要

The physical sensations of touch, as measured by the force and vibration felt during contact, strongly depend on the normal force exerted by the end-effector as well as its speed relative to the surface, factors we call “scan-time parameters.” When researchers record surface interactions for machine learning tasks such as haptic surface recognition, they must either (1) precisely control these parameters, (2) record them alongside the rest of the data, or (3) develop postprocessing that makes surface percepts invariant to scan-time parameters. Here we use multi-class support vector machines to compare the second approach (“scan-dependent features” developed in our prior work) with the third approach (“scanfree features” developed by Strese et al.). We first verify our implementation of the scan-free features by testing on Strese et al.'s published dataset of 69 surfaces, achieving 87.4% crossvalidation accuracy. We then compare the two approaches on new surface interaction data gathered from 28 surfaces with our instrument (the Proton Pack), obtaining 57.9% accuracy using scan-dependent features and 93.3% accuracy using scan-free features, demonstrating superiority of the scan-free features for this task. We further calculate the correlation of all features to scan-time parameters, confirming that the scan-free features are relatively invariant to scan-time parameters.