Abstract

Recognizing an object's material can inform a robot on how hard it may graspthe object during manipulation, or if the object may be safely heated up. Toestimate an object's material during manipulation, many prior works haveexplored the use of haptic sensing. In this paper, we explore a technique forrobots to estimate the materials of objects using spectroscopy. We demonstratethat spectrometers provide several benefits for material recognition, includingfast sensing times and accurate measurements with low noise. Furthermore,spectrometers do not require direct contact with an object. To illustrate this,we collected a dataset of spectral measurements from two commercially availablespectrometers during which a robotic platform interacted with 50 distinctobjects, and we show that a residual neural network can accurately analyzethese measurements. Due to the low variance in consecutive spectralmeasurements, our model achieved a material classification accuracy of 97.7%when given only one spectral sample per object. Similar to prior works withhaptic sensors, we found that generalizing material recognition to new objectsposed a greater challenge, for which we achieved an accuracy of 81.4% vialeave-one-object-out cross-validation. From this work, we find thatspectroscopy poses a promising approach for further research in materialclassification during robotic manipulation.

 

Introduction (beta)

None

 

Conclusion (beta)

None