Riccardo Frenner

Abstract

In the pursuit of customizability and evolvability of vehicle functions, manufacturers shift towards software-defined vehicles to enable flexible customization and over-the-air updates. This results in multiple variants and versions of a vehicle model. While shifting to software-defined vehicles (SDVs) adds value and flexibility for customers, manufacturers struggle with homologating new and updated functionality because existing testing processes do not scale for high-frequency release cycles that limit available testing resources. Overcoming this challenge by using a coherent test process designed for testing continuously evolving variant-rich systems will be one of the key enablers. This paper presents an innovative end-to-end pipeline for efficient and comprehensive testing of variant-rich vehicle functionality tailored to an application in continuous development. Our transferable test pipeline employs sample-based variant selection, a software-in-the-loop environment for executing selected variants, and scenario-based testing using mutation-based scenario fuzzing. A central test controller is responsible for managing the process. We evaluate our test pipeline with regard to its fault detection capability, using the YOLO object detection algorithm as a variant-rich test object in the CARLA simulator. Our results show that the testing process outperforms random variant sampling and scenario mutation in detecting faults.

Abstract

Software plays an increasingly important role in modern technical systems. For the broad applicability and to appeal to a wide range of customers, software often provides an enormous degree of configurability. In addition, software is under continuous development, further increasing complexity and introducing additional configuration options. Software testers face significant challenges in this context. While product-wise testing of all possible software variants is no longer feasible, alternative testing approaches that are both efficient and effective in detecting potential software faults are required. Here, t-wise interaction sampling is established as a systematic testing approach to ensure coverage of the configuration space. However, current sampling approaches suffer from severe scalability issues, often requiring hours or even days of computation when applied to large-scale Software Product Lines (SPLs). In this paper, we present a novel technique for pairwise sampling and are the first to address the enormous computational effort required to sample large-scale SPLs with the parallel computing capabilities offered by consumer graphics processing units (GPUs). We build upon techniques used in existing sampling approaches, such as SAT solvers and local search strategies, extending these traditionally sequential techniques for parallel execution. Furthermore, we present a novel, highly efficient sample optimization strategy based on Quadratic Integer Programming (QIP) as an advancement. We evaluate our sampling approach SAMParallel on a diverse set of 20 medium- and large-scale configurable software systems from the existing literature. We show that SAMParallel can generate competitive pairwise covering samples with reliably smaller sizes compared to current sampling approaches, while considerably reducing the time investment previously required. Beyond the practical implications of our contribution regarding faster sampling, the presented technique offers various hyper-parameters to further balance sampling times and sample sizes. As a result, software testers can adapt the sampling process to their specific needs.

Abstract

The necessity of a reconfigurable software-defined vehicle platform that is no longer sold as a one-off, but can be continuously adapted to dynamic customer requirements during its lifetime, is becoming increasingly clear. Political and economic institutions have become aware of these needs and initiated actions in the context of contested markets with declining sales figures. However, according to our research, existing vehicle architecture concepts and frameworks do not cover the upgradeability, dynamic reconfigurability at runtime and continuous adaptability aspects of vehicle sensors to the necessary extent. Vehicle sensor technology constitutes the fundamental basis for safe decisions of autonomous driving functions, we therefore consider the importance of this topic to be significant. Building upon the existing service-oriented E/E architecture framework ROSMARIN, we want to contribute a concept extension, allowing the seamless integration of intelligent sensors in vehicles. In addition to requirements for intelligent sensor technology in flexible architectures such as Plug-and-Play of sensor hardware, upgradeability through virtualization of sensor software and runtime resilience, our concept furthermore integrates a novel generalized approach for dynamic sensor calibration and fusion. We prove our concept on a physical demonstrator setup with ultrasonic, LiDAR and camera sensors and demonstrate the upgradeability and reconfigurability of sensor hardware and software. In addition, we show the dynamic online fusion of camera and LiDAR data with minimal effort, while also proving the resilience and automatic adaptability of our approach in case of unforeseeable events like sensor failures.

Abstract

The progress of digitalization and Internet of Things enables more and more complex, networked and powerful Cyber-physical Systems (CPS) operating in uncertain environments. This complexity and uncertainty, however, makes it unfeasible to model every aspect in advance. This causes the models to leave their scope and reach their capability limits. Specifically, in safe maintenance planning for highly-automated trucks, this fact causes waste of valuable resource, since maintenance models are often more rule-of-thumb (e.g. operation hours) than precise. In order to counteract this issue, we propose extending the digital twin concept by artificial intelligence such that the models become dynamic and adaptive. Having described the general approach and its architecture, we showcase and evaluate the approach in a highly-automated truck scenario.