Bernhard Jaeger

selected publications

1. 

   Expert Drivers for Autonomous Driving

   Jaeger, Bernhard

   2021

   Abs Bib PDF

   A popular approach to self-driving is imitation learning, in which a neural network is trained to predict the actions of an expert driver for a given input. For research performed on simulators, these expert drivers are usually computer programs that have privileged access to the simulation. While these approaches are more cost-efficient than using labels from humans, their quality is worse because existing expert driver approaches are not able to safely drive through urban environments. We improve upon prior work and propose SEED, a Simple and Effective Expert Driver that can safely navigate through urban environments even under challenging adversarial traffic scenarios. We show its utility by training an existing imitation learning architecture to imitate SEED. The resulting TransFuser+ method sets a new state-of-the-art on the challenging NEAT validation routes and outperforms the best prior work on the CARLA leaderboard.

   @mastersthesis{jaeger2021expertdrivers,
     bibtex_show = {true},
     title = {Expert Drivers for Autonomous Driving},
     author = {Jaeger, Bernhard},
     year = {2021},
     selected = {true},
     pdf = {master_thesis_bernhard_jaeger.pdf},
     school = {University of Tübingen},
     teaser = {master_thesis_teaser.png}
   }