Design of Agent Selection Strategies for Generating Adversarial Traffic Scenarios to Test Automated Driving Functions

Context

With the advancement of automated driving technologies, the demand for reliable verification and validation methods is also increasing. One promising approach involves the use of adversarial traffic scenarios—intentionally challenging situations designed to systematically uncover weaknesses in perception, decision-making, and control of automated vehicles. A central question in this context is how traffic participants (agents) can be selected and positioned within a scene in such a way that safety-critical yet plausible scenarios emerge. To ensure meaningful test results, realistic behavior patterns and a consistent scenario logic must be taken into account when generating these scenarios.

Aims

• Development of a method for selecting traffic participants suitable for generating adversarial scenarios in simulation
• Definition of appropriate criteria for agent selection (e.g., spatial proximity, behavior patterns)
• Integration of the selection method into a reinforcement learning training environment
• Generation of challenging traffic scenarios based on the selected agents
• Comparison and evaluation of the developed method

Requirements

• Programming skills in Python, preferably with experience in ML libraries (e.g., PyTorch)