Multimodal Material Classification for Robots using Spectroscopy and High Resolution Texture Imaging

摘要: Material recognition can help inform robots about how to properly interact with and manipulate real-world objects. In this paper, we present a multimodal sensing technique, leveraging near-infrared spectroscopy close-range high resolution texture imaging, that enables estimate the materials of household We release dataset images spectral measurements collected from mobile manipulator interacted 144 house-hold then neural network architecture learns compact representation images. When generalizing material classification new objects, show robot recognize greater performance as compared prior state-of-the-art approaches. Finally, combine local robot’s head-mounted camera achieve accurate over scene objects on table.