• Abstractkeyboard_arrow_down

  Serientaugliche, hochautomatisierte Fahrfunktionen sind für die Fahrzeugindustrie ein gänzlich neues Feld. Aufgrund des Neuheitsgrades lässt sich ein derartiges System heute weder vollständig spezifizieren, noch hinreichend validieren.
  Die Frage „Wie kann ich sicherstellen, dass ausreichend getestet wurde und alle Situationen im Straßenverkehr ohne Unfälle oder sonstige Beschädigungen bewältigt werden können?“ lässt sich vorerst nicht vollständig beantworten.
  Die Anzahl der notwendigen Testfälle für hochautomatisierte Fahrzeuge ist theoretisch unendlich groß. Da dies in der Praxis nicht möglich ist, müssen andere Strategien gefunden werden. Eine Strategie ist es, einen großen Teil der Tests in virtuellen Umgebungen durchzuführen, da dies effizienter, schneller und günstiger möglich ist. Nichtsdestotrotz kann so nur eine bestimmte Testabdeckung erzielt werden. Ein Restrisiko bleibt bestehen.
  Um dennoch die Sicherheit des Systems gewährleisten zu können, müssen entsprechende Überwachungs- und Diagnosemechanismen installiert werden. Diese dienen nicht nur der Absicherung des Systems, sondern darüber hinaus auch zur Sammlung von Daten aus dem Realbetrieb, die wieder in die Entwicklung zurückgeführt werden können. Durch diese kontinuierliche Feedbackloop können die Systeme über regelmäßige Software-Updates in sicherer Art und Weise verbessert und erweitert werden.
  Dieses Vorgehen entspricht der DevOps Methodik, die aus der Softwareentwicklung bekannt ist. Ein System wird nach einer Spezifikation entwickelt (DEVelopment) und validiert. Im Betrieb (OPerationS) treten dann Probleme auf, die wieder in den Entwicklungszyklus rückgeführt werden. So „lernt“ das System immer besser zu werden.
  Im Vortrag wird der Validierungsprozess und die Notwendigkeit einer engen Interaktion zwischen Entwicklung und Operations (das DevOps-Vorgehen) beschrieben, wie es für im Forschungsprojekt ASCERTAIN-S3 geplant ist.

Vortragsfolien (passwortgeschützt)

• Abstractkeyboard_arrow_down

  One does not need the gift of clairvoyance to predict that in the near future autonomously driving cars will occupy a significant place in our everyday life. In fact, in designated and even public test-drive areas it is reality even now. Autonomous driving comes with the ambition to make road traffic safer, more efficient, more economic, and more comfortable - and thus to “make the world a bit better”. Recent accidents with automatic cars resulting in severe injuries and death, however, spark a debate on the safety validation of autonomous driving in general. The biggest challenge for autonomous driving to become a reality is thus most likely not the actual development of intelligent vehicles themselves but their rigorous validation that would justify the necessary level of confidence.

  It is common sense that classical test approaches are by far not feasible in this emerging area of autonomous driving as these would induce billions of kilometers of real-world driving in each release cycle. To cope with the tremendous complexity of traffic situations that a self-driving vehicle must deal with - without doing any harm to any other traffic participants - a promising approach to safety validation is virtual simulation, i.e. driving the huge amount of test kilometers in a virtual but realistic simulation environment. A particular interest here is in the validation of the behavior of the autonomous car in rather critical traffic scenarios.

  In this presentation, we concentrate on one important aspect in virtual safety validation of autonomous vehicles, namely how rather abstract traffic scenarios can be properly executed in a simulation environment. The proposed workflow starts with the graphical yet formal specification of abstract traffic scenarios to create a scenario catalog for virtual validation. The formal semantics of these abstract scenarios then gives rise to determine concrete scenario instantiations using formal methods. As the criticality of traffic maneuvers also depends on the actual course of the road, we show how concrete road geometries can be easily created in order to simulate the concretized traffic scenarios on these specified roads. To properly execute a concretized traffic scenario in a simulation environment in the sense that the abstract scenario specification is met, we have to cope with the “autonomy” of the ego car under test, i.e. we need to compensate potential deviations of the ego car behavior related to the story of the concretized scenario by means of suitable reactions of the controllable surrounding traffic. For all that, it cannot be guaranteed in general that the simulated scenario actually complies with the abstract scenario specification. In order to detect such cases automatically, we show how scenario observers are derived from abstract traffic scenarios to monitor the simulation and to report coverage of the scenario catalog.

Vortragsfolien (passwortgeschützt)

• Abstractkeyboard_arrow_down

  To assess the safety of automated driving systems (ADS), all potentially critical situations have to be considered. One way to do so is to test the function performance in scenarios which lead to these situations. A scenario is a description of an evolution of traffic situations. It provides snapshots capturing important intermediate states and descriptions specifying what happens between these states. Semantically, a scenario can be mapped to a set of time series: All concrete instantiations of the dynamic behavior which is specified in the scenario. In this contribution, abstract scenario are used for test specifications, while their concrete instantiations define test cases.
  The number of relevant concrete scenarios is very large, even if a particular ADS is supposed to operate in a restricted traffic environment. Therefore, testing must rely heavily on a virtual, largely automatized exploration of scenario spaces. For that, the abstract scenarios have to be given in language with a formal notation, to enable the dynamic generation of concrete test cases covering the behavior space.
  The presentation will delineate a general approach to safety assessment of ADS by virtual testing. It discusses in particular the nature and building blocks of a formal scenario language and the construction of test specifications. This approach is based on developments of the ongoing research project PEGASUS, where such a method is being defined and instantiated.

Vortragsfolien (passwortgeschützt)