Engineering of safe autonomous vehicles through seamless integration of system development and system operation

Abstract

Autonomous vehicles will share the road with human drivers within the next couple of years. This will revolutionize road trac and provide a positive benet for road safety, trac density, emissions, and demographic changes. One of the signicant open challenges is the lack of established and cost-ecient veri- cation and validation approaches for assuring the safety of autonomous vehicles. The general public and product liability regulations impose high standards on manufacturers regarding the safe operation of their autonomous vehicles. The vast number of real- world trac situations have to be considered in the verication and validation. Todays conventional engineering methods are not adequate for providing such guarantees for au- tonomous vehicles in a cost-ecient way. One strategy for reducing the costs of quality assurance is transferring a signicant part of the verication and validation from road tests to (system-level) simulations. The vast number and high complexity of real-world situations complicate the exhaustive verication of autonomous vehicles in simulations. It is not clear, how simulations address the vast number of real-world situations with sucient realism and how their results transfer to the real road. Extensive coverage of real-world situations in simulations requires the integration of de- velopment and operation. This thesis presents an engineering approach that integrates the development and operation of autonomous vehicles seamlessly using runtime moni- toring. The runtime monitoring veries if autonomous vehicles satisfy their requirements and operate within safe limits which have been veried in the simulations. Safety of autonomous vehicles is subject to the scope of veried trac situations in simulations. Systematic and comprehensive simulations support the improvement of autonomous vehicles and coverage of trac situations. Results of the runtime monitoring during operation are transferred to the development for the verication of autonomous vehicles and their safe limits in simulations with additional trac situations. The incomplete verication of autonomous vehicles for the vast number of real-world trac situations in simulations requires the validation of simulation results and addi- tional monitoring in the real world. Results from simulations are transferred to the runtime monitoring during operation in the real world for validating the realism of the simulations and maintaining the vehicle safety in critical situations. Vehicle data and real-world situations possess high complexities and, therefore, impact the complexity and eciency of the verication in simulations. The runtime monitoring abstracts from internal data of autonomous vehicles and real-world situations in the evaluation by introducing an abstract semantic representation from natural language requirements. A case study evaluates the engineering approach for an industrial lane change assistant and real-world trac data recorded in road tests on German highways.%%%%Autonome Fahrzeuge werden in den…