In-Memory Computation Using CMOS-Integrated Resistive RAM for Robotic Navigation

Abstract

Neuromorphic and in-memory computing (IMC) enabled by Resistive Random Access Memory (ReRAM) holds the potential to dramatically improve the energy efficiency of computation. ReRAM-based IMC facilitates an efficient hardware solution for neural networks reliant on vector matrix multiplication (VMM) and associated applications. To demonstrate these capabilities, we utilized fully CMOS-integrated ReRAM arrays to perform IMC operations for a robotic line following navigation task. In this work, we developed a custom microcontroller-based interface for receiving sensor inputs, performing VMM operations on custom fabricated and packaged ReRAM arrays, and using VMM outputs to guide the robotic demonstrator. This work demonstrates a comprehensive analysis of the impact of ReRAM resistance stochasticity on navigational accuracy, the impact of microcontroller / board design on ReRAM performance, and how careful selection of ReRAM resistance states can mitigate operational errors.