Shape from touch by a neural net

The authors report the implementation and testing of a neural network algorithm to solve the fine-form discrimination problem. The skin-like sensor design is briefly reviewed, and the fine-form discrimination task is stated as an inverse problem of contact mechanics. The inverse problem is discussed using a singular value decomposition argument, to enlighten the basic regularity assumptions that are needed in the inversion, and the class of shapes that it is possible to recover from the data. The network design is given, along with the information about training and testing sets and learning rate. Simulated experimental results are reported