Learning to identify container contents through tactile vibration signatures

We examine using a simple contact sensor coupled with standard machine learning algorithms to classify and count objects shaken in a container. The contact sensor measures the resulting vibrations, and these signatures are used to learn a classifier that maps vibration signatures to known object categories. A linear support vector machine trained on labeled vibration signatures achieves a mean binary classification accuracy of 99% over 66 pairs of objects and a mean multi-class classification accuracy of 94% over 12 classes. It is also shown that useful tasks such as approximate counting of objects over the range 1 to 10 is possible. We see potential applications of these ideas in service robots engaged in cleanup and inventory control in labs, workshops, stores, warehouses and homes.