Fakultät für InformatikLehrstuhl 22Software EngineeringProf. Dr. Alexander Pretschner

Boltzmannstraße 385748 Garching bei München

Tel: +49 (89) 289 - 17885https://www22.in.tum.de

Testing Automated Driving Systems: Fitness Function Deriva-tion for Search-Based Scenario GenerationMaster’s Thesis

Supervisors: Prof. Dr. Alexander Pretschner, Florian HauerEmail: {alexander.pretschner, florian.hauer}@tum.de

Phone: +49 (89) 289 - 17885Starting date: immediately

ContextDriver assistance systems exist for over three decades now with increasing functionality and theoverall goal of highly autonomous driving seems to be not out of reach anymore [1]. However,the systems are getting increasingly complex as they are not only passive, but active systemsinterfering with the driver. Thus, for advanced driver assistance systems (ADAS) extensivetesting needs to be performed, before they can be deployed for series production [2][3].

For an autonomous highway pilot, it is estimated that approximately 6.62 billion kilometers oftest driving on highways are necessary [4]. Considering this and other complexity and feasibilityissues, simulation is arguably the most practical and effective way of testing software systemsused for autonomous driving [5].

A lot of such simulation tools exist, e.g. CarMaker by IPG Automotive [6]. However, withinthese tools, test scenarios are created in a manual and very ad hoc manner. Parameters areadjusted by “trial & error”.

To improve this process, specification-based automated parameter search could be done.In [7], scenario parameters are automatically found for the test of an automated parking system.Similarly, test scenarios are generated in [8] for a braking assistant and in [5] for a pedestriandetection system. In the domain of unmanned aerial vehicles, test scenarios are generated forcollision avoidance systems [9].

Figure 1: Screen shot from CarMaker: Automated emergency braking system detects apedestrian and brakes.

Fakultät für InformatikLehrstuhl 22Software EngineeringProf. Dr. Alexander Pretschner

Boltzmannstraße 385748 Garching bei München

Tel: +49 (89) 289 - 17885https://www22.in.tum.de

GoalThe main goal of this thesis is the development of a framework for test scenario generation.User-provided system behavior specifications are used to derive fitness functions for search-based techniques to find suitable parameters for the test scenarios. This includes several partialgoals:

1. Since test scenarios should be generated for testing as specific system behavior, thisbehavior has to be specified in a way to serve as input for the generation mechanism.For this, a suitable way of description has to be chosen.

2. The generation mechanism will contain search-based techniques, which need so calledfitness functions to optimize the scenario parameters. Such fitness functions have to auto-matically be derived from the specified system behavior, possibly by using transformationtemplates.

3. The search-based techniques have to run automatically after derivation of the fitnessfunctions. For this and the first steps, a framework has to be developed.

Working Plan1. Understand search-based techniques and how they are applied in this context2. Get familiar with CarMaker and understand how simulation based testing is done3. Implement a toolchain for search-based scenario generation4. Develop a description technique for system behavior5. Identify templates like described above6. Round out your framework for automation7. Evaluate your framework by applying it to a lane-keeping system

Deliverables• The framework’s source code and modules• A demo of the framework, including instructions on how to run the demo• Technical report with comprehensive documentation of the implementation, i.e. design

decision, architecture description, API description and usage instructions• Final thesis report written in conformance with TUM guidelines

References[1] Bengler, Three Decades of Driver Assistance Systems, 2014[2] Huang, Autonomous Vehicles Testing Methods Review, 2016[3] Stellet, Testing of advanced driver assistance towards automated driving: A survey and

taxonomy on existing approaches and open questions, 2015[4] Wachenfeld, Freigabe des autonomen Fahrens, 2015[5] Abdessalem, Testing Advaned Driver Assistance Systems using Multi-objective Search

and Neural Networks, 2016[6] IPG Automotive, CarMaker, online at https://ipg-automotive.com/products-services/

simulation-software/carmaker/, retrieved 23th February 2017[7] Bühler, Evolutionary functional testing of an automated parking system, 2003[8] Bühler, Evolutionary functional testing, 2008[9] Zou, Testing method for multi-uav conflict resolution using agend-based simulation and

multi-objective search, 2016