Determining Object Properties from Tactile Events During Grasp Failure

Robotic grasp planners are unable to successfully grasp objects 100% of the time. During the failures, inferring a better understanding of the in-hand item could lead to more robust regrasp strategies. This paper explores how relevant object parameters, such as its surface properties and its weight distribution, may be extracted from a time series of tactile feedback generated during a grasp failure. The surface texture of four known objects was classified with an accuracy of 90.22%. Objects whose weight distribution were top heavy, bottom heavy, or evenly distributed were distinguished with an accuracy of 86.9%. In both cases, only a small dataset was required to achieve a relatively high performance.