Contact Classification for Agriculture Manipulation in Cluttered Canopies

Abstract

In this paper, the authors present a novel way to classify contact
objects using audio signals in a highly cluttered canopy
environment for agriculture manipulation. Rather than solely
relying on visual data to represent the dense canopies as obstacles,
we investigate whether robot can observe latent properties
about safe interactions such as brushing against leaves
using audio signals. We developed a hand-held device to
facilitate the data collection process to distinguish between
three classes: leaf, twig, trunk. Of the time domain, frequency
domain, and cepstrum representations (MFCC), MFCC comparisons
showed the most distinguishable patterns across the
classes. The provided results present a promising direction to
expand this research to leverage deep learning networks to
consistently classify the extracted audio inputs that can lead
to safe and robust agriculture manipulation.