Tactile based Forward Modeling for Contact Location Control

Abstract

In order to grasp and manipulate objects, humans heavily rely on tactile information. Fine manipulation actions specifically rely on information gathered by the receptors locate at the finger-tips. Additionally, humans display some predictive capabilities with respect to their haptic feedback. For manipulation actions, having a robot predict its contact states based on the actions taken, enables the estimation of optimal control actions in order to follow desired state trajectories. We present an approach to control the point of contact between the finger-tip and the surrounding environment. We make use of principal component analysis to reduce the dimensionality of the tactile signals. The resulting lower dimensional representation and the robot motor commands are used as inputs to a Gaussian Process that predicts the point of contact in the next time step. Based on this prediction, a controller estimates the optimal control actions that minimize the error to the desired point of contact. We show that our approach yields promising results in both a simulation environment and real robot experiments.