Successful manipulation of unknown objects requires an understanding of their physical properties. Infrared thermography has the potential to provide real-time, contactless material characterization for unknown objects. In this paper, we propose an approach that utilizes active thermography and custom multi-channel neural networks to perform classification between samples and regression towards the density property. With the help of an off-the-shelf technology to estimate the volume of the object, the proposed approach is capable of estimating the weight of the unknown object. We show the efficacy of the infrared thermography approach to a set of ten commonly used materials to achieve a 99.1% R2 -fit for predicted versus actual density values. The system can be used with tele-operated or autonomous robots to optimize grasping techniques for unknown objects without touching them.
|Original language||English (US)|
|State||Published - 2019|