Power-Efficient and Accurate Texture Sensing Using Spiking Readouts for High-Density e-Skins

Abstract

Fine-grain tactile sensing has recently gained much attention in robotics applications where the manipulation of potentially fragile objects must be provided. This has led to the emergence of electronic skin (e-skin) sensors, usually implemented with conventional frame-based acquisition chains. In addition, prosthetics applications require e-skins with human-level, sub-millimeter spatial resolution. This paper proposes to study two types of spike-based e-skin readout circuits, based on conventional and neuromorphic level crossing architectures. Compared to prior frame-based, coarse spatial resolution readout schemes, a sub-millimeter spiking e-skin scheme is modeled and compared to its frame-based counterpart in terms of power consumption and texture classification accuracy, using a Spiking Neural Network. Our analysis shows that the sparsity-inducing spike-based solutions achieve one order of magnitude lower power consumption while reaching a higher classification accuracy (87.92%) compared to the frame-based readout (74.58%).