Neural networks for simulating the deformation of soft materials

For robot manipulators, it is not an easy task to handle an object made of soft materials, such as rubber, biological tissue and foods since these materials change their form when a manipulator applies a force in due process of handling or cutting. Deformations of this kind sometimes involve nonlinearities which make it difficult to predict and compensate the error for the manipulator control. In this paper, we introduce a neural-network-aided deformation simulator consisting of: 1) mutually connected BP-nets arranged in 2D to simulate a global deformation of the material; 2) a BP-net for simulating intensive deformations locally produced by the manipulating action; and 3) a deformation detecting network to produce the training data from an image sequence from a TV camera. We show preliminary experimental results on the deformation of a urethan block with a cutting blade pressed on.