Friday, September 18, 2026

Development of an Expert Driving Algorithm for End-to-End Learning in Simulation

Topic and Goal of the Thesis

Traditional automated driving stacks split the driving task into separate components, which makes them prone to error propagation and severe failure cases. End-to-end (E2E) driving models instead learn to map sensor data directly to driving actions by imitating demonstrations, which in the real world usually come from human drivers. In simulation, an expert algorithm takes this role and can use privileged information, such as the exact states of all road users.

The objective of this thesis is to develop an expert driving algorithm that generates high-quality demonstrations that E2E-models learn from. The focus is on a rule-based approach, complemented and compared with a reinforcement learning-based method.

Working Points

• Literature review on expert driver algorithms used for training end-to-end driving models
• Development of a rule-based expert driver that handles complex urban traffic scenarios in simulation
• Implementation of a reinforcement learning-based expert and comparison with the rule-based approach
• Evaluation of driving performance on established closed-loop metrics and benchmarks

Requirements

• Reliability, commitment and enjoyment of working independently
• Experience in programming (Ideally Python or C++)
• Experience with the following is beneficial: CARLA, ROS 2, Docker and Machine Learning,

Note: Please attach brief resume and grade summary.