From Modular Pipelines to AI-first: Software Architectures for Automated Driving

Topic

The objective of this thesis is to conduct a systematic survey of existing software architecture paradigms for automated driving and a comparative analysis of selected, publicly documented open-source and industrial implementations. Based on this analysis, architectural trends, research gaps, and development potentials in the tension field between system performance and assurance/safety shall be identified.

Background und Motivation

Software architecture is a key success factor for automated driving systems. Classical AD stacks predominantly follow a modular pipeline structure (Sense → Plan → Act). In recent years, however, alternative architectural paradigms have gained increasing relevance, including:

• End-to-End approaches, in which large parts of perception, decision-making, and control are integrated into data-driven, largely monolithic models,
• Hybrid architectures, which combine learning-based end-to-end components with explicitly modular system components,
• Agentic approaches, in which the overall system is modeled as a network of autonomous software entities.

While these paradigms promise performance improvements, they pose significant challenges to established concepts of functional safety, verification, and system assurance. In particular, due to their “black-box” nature, strongly data-driven approaches increasingly shift assurance efforts from model-internal proofs toward architecture-driven measures such as supervisor concepts, runtime monitoring, degradation strategies, and system-level assurance.

Aufgaben

• Systematic identification and classification of current software architecture paradigms for automated driving based on a literature review following established systematic review methodologies (e.g., defined search strings, databases, and inclusion/exclusion criteria)

• Development of a consistent taxonomy that enables a transparent comparison of end-to-end, hybrid, and agent-based architectures

• Comparative analysis of selected, publicly documented AD stacks with respect to architectural approach, modularity, sensor setup, maturity level, and existing assurance and monitoring concepts

• Identification of research gaps and architectural trends, particularly in the tension field between performance and the assurability of data-driven systems

• Scientific writing of the thesis in accordance with academic standards (German or English)

Qualifikationen

• Enrollment in Mechanical Engineering, Industrial Engineering, Business Informatics, or a related field

• Interest in software architectures, automated driving, and current developments in robotics

• Basic knowledge of automated driving systems, AI/ML, or vehicle/robotics architectures is an advantage

• Analytical and structured working style and the ability to present complex technical topics in a clear and comprehensible manner

Das bieten wir

• Work on a current and scientifically relevant research topic in the field of automated driving

• Close technical supervision and methodological support during the preparation of a systematic survey

• In-depth insights into modern AD architectures and assurance strategies

• Opportunity to build solid competencies in the field of automated driving

• Option to further exploit the results of the thesis (e.g., conference paper, journal survey, technical presentation), depending on quality and interest

Beginn: By arrangement

Interested? Please contact anian.scheibel∂kit.edu with your current transcript of records and CV.