Bachelor Thesis

Thesis Topic

Systematic analysis and classification of end-to-end architectures for automated driving. The thesis focuses on how current end-to-end approaches differ in terms of their internal structure, latent representations, interfaces, and generated outputs, despite their generally end-to-end differentiable nature.

Based on a structured literature review, a transparent and comprehensible taxonomy is to be developed and applied to selected publicly documented end-to-end architectures.

Background and Motivation

Conventional automated-driving architectures consist of separate modules for perception, environment modelling, planning, and vehicle control. End-to-end approaches, by contrast, aim to connect sensor data and driving decisions within a system that can be trained jointly from end to end.

However, the term “end-to-end” does not describe a single, uniform architecture. Current approaches differ, among other aspects, in:

• the sensor data they process,
• the intermediate representations they learn,
• whether they use explicit interfaces or exclusively latent interfaces,
• how perception, prediction, and planning are coupled,
• the outputs generated by the model,
• and the extent to which individual components are trained jointly.

To date, there is no unified and readily understandable taxonomy that enables different end-to-end architectures to be described and compared systematically.

Objective of the Thesis

The objective of this bachelor’s thesis is to develop a taxonomy for end-to-end architectures in automated driving. The taxonomy should capture key architectural characteristics and enable a consistent classification of current approaches.

The developed taxonomy will subsequently be applied to a limited selection of publicly documented end-to-end architectures.

Tasks

• Conduct a structured literature review of end-to-end architectures for automated driving
• Delineate the term “end-to-end” from modular and hybrid architectures
• Identify relevant characteristics for describing and distinguishing end-to-end approaches
• Investigate different latent and explicit representations, including:
  • bird’s-eye-view representations
  • object and vector representations
  • occupancy and grid representations
  • trajectories and waypoints
  • fully implicit latent feature spaces
• Develop a consistent taxonomy
• Apply the taxonomy to selected publicly documented architectures
• Compare the approaches with regard to architecture, information flow, differentiability, interpretability, and safety assurance
• Derive key architectural trends and open research questions
• Prepare the bachelor’s thesis as a scientific document in either German or English

Qualifications

• Degree programme in mechanical engineering, mechatronics, computer science, industrial engineering and management, or a comparable field
• Interest in autonomous driving, artificial intelligence, and software architectures
• Basic knowledge of artificial intelligence, machine learning, or automated-driving systems is an advantage
• Willingness to become familiar with scientific publications and technical architectures
• Analytical and structured working approach
• Good command of English, as a large proportion of the relevant literature is available in English

Programming skills are helpful but not required. The main focus of the thesis is the systematic analysis and classification of existing approaches.

What We Offer

• The opportunity to work on a current and clearly defined research topic in the field of AI-based autonomous driving
• Close technical and methodological supervision
• An introduction to structured literature-review methods
• Insights into current end-to-end architectures and AI-based driving systems
• The opportunity to build in-depth knowledge of autonomous-driving architectures, latent representations, and the safety assurance of data-driven systems
• The possibility of further using the results, for example in a scientific publication or a subsequent research project, depending on the quality of the work and your interests

Starting date: By arrangement

Interested? Please contact anian scheibel ∂does-not-exist.kit edu and include your current transcript of records and CV.

Keywords: Artificial Intelligence, Autonomous Driving, Automated Driving, Machine Learning, Deep Learning, End-to-End Learning, End-to-End Driving, Software Architectures, Computer Vision, Robotics, Embodied AI, Physical AI, Intelligent Vehicles, Autonomous Systems, Safety Assurance