Julian F. Schumann – TU Delft
For humans to trust autonomous vehicles, they need to have confidence in the vehicles’ ability to reliably resolve space-sharing conflicts with other traffic participants in a safe manner – such as in the case of crossing or merging paths. Planning safe and efficient interactions for autonomous vehicles requires accurate and trustworthy prediction models for behavior of human participants in such situations; these models need to be tested and validated comprehensively. This proposal aims to contribute to addressing this challenge through a research visit of Julian Schumann (PhD student in TU Delft, TAILOR lab) to Institute for Transportation Studies (University of Leeds, non-TAILOR lab). We plan on firstly creating a dataset on which such testing and comparisons between different models can be performed by using a state-of-the-art driving simulator. It will focus on highly safety critical, but rare traffic situations, the inclusion of which is an important requirement for estimating the robustness of the prediction models. Secondly, we will integrate this dataset into a recently developed framework for benchmarking human behavior prediction models to facilitate the fast and easy testing of those models. Lastly, with the help of the updated framework, we plan to test a recently developed model, which, integrating psychological theories on a large-scale, promises a very high robustness. Overall, this project will provide tools for a better and more comprehensive evaluation of a prediction model, and show the impact of including psychological research into such prediction models.
Keywords: autonomous vehicles, human behavior prediction, benchmarking, machine learning, cognitive modeling, robust AI
Scientific Area: Human Robot Interaction