"Certified Control" safety architecture for autonomous vehicles : applications with LiDAR

Abstract

Certified control is a safety architecture for autonomous vehicles, in which a safety monitor checks actions proposed by the main controller before they may be executed by the actuators. Unlike conventional runtime monitors, the certified control monitor receives an argument for the safety of the proposed action from the controller (rather than receiving data from the vehicle sensors directly). In this architecture, the monitor has the potential to do all of the following to a reasonable degree: intervene when safety is compromised, not intervene when safety is not compromised, and remain simple enough to be verifiable. First, this work describes the certified control architecture in detail, including how it achieves those three desiderata, which we argue are otherwise difficult to achieve simultaneously. Second, we present two novel applications of certified control: an implementation of LiDAR-based obstacle detection, and a LiDAR-augmented implementation of visual lane following. Finally, we evaluate those two systems using simulation and a physical robot car, and demonstrate that they indeed achieve the three desiderata.