Teaching self-driving cars to dream: A deeply integrated, innovative approach for solving the autonomous vehicle validation problem

摘要

Validating autonomous vehicles is a tough problem that, if not solved in a timely manner, might hinder the release of autonomous vehicles. Standard software testing techniques might not be sufficient to validate such complex systems, which is why, in this paper, we present a novel approach to solve this problem. We introduce a new validation space concept that is based on a network of spatiotemporal state lattices to generate the motion of dynamic traffic participants. This space is sparse and fractured in the beginning but becomes more and more integrated and dense as the vehicles experience new situations. Therefore we propose new testing techniques for efficiency as well as an overall workflow that requires a new safety module to be embedded into the automated vehicle's architecture.