Deep GRU-ensembles for active tactile texture recognition with soft, distributed skin sensors in dynamic contact scenarios

Abstract

Tactile feedback is an important sensory information that contributes to our ability to perform various manipulation tasks. Soft sensor skin aims at providing, e.g., anthropomorphic robot hands with the sense of touch to resemble this ability in teleoperation or prosthetic applications. However, the exploration of an objects texture using a robot hand involves various dynamic contact scenarios. Depending on the relative dynamics between the soft sensor skin and the object, the difficulty and the dimensionality of the recognition task changes dynamically during the manipulation of an object. We deployed a deep gated recurrent unit (GRU)-ensemble with unweighted averaging that allows the texture recognition algorithm to dynamically scale with the number of contact points during active tactile texture exploration while maintaining a high accuracy. We experimentally verify the approach by evaluating the prediction performance of the GRU-ensemble on the data that was gathered during the active tactile exploration of four objects of daily living by means of a uSkin sensor module providing up to 16 3-axis force vectors at 100Hz sampling frequency. The accuracy of 100% suggests that deep GRU-ensembles offer a scalable option for reliable texture recognition in active tactile exploration for the implementation into tactile feedback systems.