Inferring Informative Grasp Parameters from Trained Tactile Data Models

When choosing parameters for a tactile action, it is useful to be able to choose parameters that will optimize the discriminating power of the tactile data acquired during the action. We propose a way to extract this information about informative grasp parameters from a deep network trained to perform a classification task on the tactile data. We do this by defining a new cost function on the network that quantifies the certainty of the result, and optimizing the input on that cost function. Though our work on this technique is still in progress, we present some early experiments using this approach on the BiGS tactile grasping dataset, and propose experiments for further investigation of the technique.