Tactile Regrasp: Grasp Adjustments via Simulated Tactile Transformations

This paper presents a novel regrasp control policythat makes use of tactile sensing to plan local grasp adjustments.Our approach determines regrasp actions by virtually searchingfor local transformations of tactile measurements that improvethe quality of the grasp.First, we construct a tactile-based grasp quality metricusing a deep convolutional neural network trained on over2800 grasps. The quality of each grasp, a continuous valuebetween 0 and 1, is determined experimentally by measuringits resistance to external perturbations. Second, we simulatethe tactile imprints associated with robot motions relative tothe initial grasp by performing rigid-body transformations ofthe given tactile measurements. The newly generated tactileimprints are evaluated with the learned grasp quality networkand the regrasp action is chosen to maximize the grasp quality.Results show that the grasp quality network can predict theoutcome of grasps with an average accuracy of 85% on knownobjects and 75% on a cross validation set of 12 objects. Theregrasp control policy improves the success rate of grasp actionsby an average relative increase of 70% on a test set of 8 objects.