Testing certified control for LIDAR and vision perception via physical testing and simulation

Abstract

Certified control is a proposed safety architecture for autonomous vehicles. It consists of a runtime monitor that requires evidence from the vehicle's main controller to prove that the controller's desired action is safe. The monitor passes the command to the vehicle's actuators if the action is deemed safe, and intervenes otherwise. In contrast to conventional runtime monitors that process perception data directly in order to evaluate the main controller's decision, certified control places the burden of finding sufficient safety evidence on the main controller. Thus, a runtime monitor in the certified control architecture need only evaluate the proposed action against the relevant subset of the total perception data, reducing overall complexity of the monitor and opens the door to the application of software checking techniques like formal verification. This work describes deployment, testing, and evaluation of certified control on a physical platform and in simulation, using runtime monitors for LIDAR and vision data developed by fellow researchers. These tests successfully demonstrate certified control in simple scenarios, and lay the groundwork for further research into more complex scenarios via simulation.