Classification of Household Materials via Spectroscopy

摘要: Recognizing an object's material can inform a robot on the fragility or appropriate use. To estimate during manipulation, many prior works have explored use of haptic sensing. In this letter, we explore technique for robots to materials objects using spectroscopy. We demonstrate that spectrometers provide several benefits recognition, including fast response times and accurate measurements with low noise. Furthermore, do not require direct contact object. this, collected dataset spectral from two commercially available which robotic platform interacted 50 flat objects, show neural network model accurately analyze these measurements. Due similarity between consecutive measurements, our achieved classification accuracy 94.6% when given only one sample per Similar sensors, found generalizing recognition new posed greater challenge, 79.1% via leave-one-object-out cross validation. Finally, how PR2 leverage everyday in home. From find spectroscopy poses promising approach manipulation.