Concept and Implementation of an Optimization-based Safety Verification Strategy for a Trajectory Following Controller

This paper presents a new approach for safety verification of self-driving systems. A statistical approach to verification is often prohibitive, so a recent trend has been to consider synthetically generated scenarios based on predefined parameters. Instead of covering a large fraction of the parameter space, however, this paper proposes an approach that searches the parameter space systematically by means of an optimization procedure. The main goal is to find worst-case scenarios, also known as corner cases, as quickly as possible (‘pessimizer’). This may lead to a significant speed up of the safety verification process, and it may help with the identification of appropriate safety goals during the development process. To this end, a finite-horizon optimization problem is formulated in which a safety-critical performance measure is minimized. The optimization problem is strongly non-convex and high-dimensional and thus difficult to solve, as it may possess multiple local minima. A tailored evolutionary algorithm is described that iterates towards these local minima, which represent the desired corner cases. The working of the algorithm and the effectiveness of the pessimizer approach are demonstrated in a simulation study for a trajectory following controller. The underlying idea, however, generalizes to many control applications and other functions for safety-critical systems.