265 — Classification of Household Materials via Spectroscopy

Erickson et al (1805.04051)

Read on 12 May 2018
#spectroscopy  #household  #home  #classification  #machine-learning  #robotics  #robot  #neural-network 

The human visual system is incredibly good at accurately predicting the material of a target object simply by looking at it. When we overlay that information with tactile and touch information, the brain is crazy-good at understanding what material it is interacting with.

Robots? Not so much. Though robots are far better at quantifying materials they can touch, they’re not as great when it comes to visual inputs. This work begins to address that problem by adding spectroscopy information to a machine’s understanding of the world around it.

Spectroscopy is an inexpensive and powerful way to identify the composition of a material, and because household materials are rarely of extreme temperature or speed, common shortcomings in spectroscopy technology are a non-issue here.

The authors categorized household materials into five categories: Metals, plastics, woods, papers, and fabrics. Using deep-learning and SVM complimentary approaches, the authors determined that the machine was able to identify the material category with a maximum encountered accuracy of 97.7%, using only spectroscopy inputs.

This means that spectroscopy is a promising and cheap technology that enables the designers of machines to begin to use the context of the environment in order to improve the understanding of the robot’s context.