A NEUROMORPHIC APPROACH TO TACTILE PERCEPTION

Abstract

This dissertation aims to improve the scalability and responsiveness of tactile sensor arrays. Inspired by biology, we promote the utility of temporal structure in tactile signals to encode stimuli and drive learning processes. We challenge the need for accurate pressure readout, proposing instead a strategy of rapidly extracting change events from large numbers of sensing elements, resulting in sparse spatiotemporal representations analogous to spike patterns from mechanoreceptors. The feasibility of our concept is demonstrated with the realization of a large tactile sensor array capable of an unprecedented 5200 frames-per-second readout of 4096 sensor elements. We demonstrate marked improvements to contact discrimination accuracy using temporally accurate tactile signals while using up to 20 times less communication bandwidth. We also present a learning algorithm for recognizing spatiotemporal patterns through convex optimization. We conclude that neuromorphic techniques offer considerable advantages in the development of efficient tactile sensing systems for robotics and prosthetic applications.