A Robot Execution Failures Classification System Based on Analog IC Neural Networks

In this work, an alternative architecture of a pure analog artificial neural network is introduced. This includes circuits operating in the subthreshold region such as the Gaussian bell-shaped curve circuit, sigmoid function circuit, analog multiplier, and current comparator. The overall implementation consumes $1.95\mu W$ and achieves an average accuracy of 94.25%. An additional advantage is its ability to handle a large number of features. It is designed and verified in a TSMC 90nm CMOS process. Furthermore, post-layout results are compared both with software and with the literature. Based on these characteristics, it emerges as a promising candidate for robot execution failures systems.