End-to-end Driving via Conditional Imitation Learning

Abstract

Standard imitation learning from expert demonstrations is limited in scenarios where goal specification is necessary, such as at road intersections. We propose Conditional Imitation Learning, which branches the policy network based on a high-level command input — for example 'turn left', 'turn right', or 'go straight'. This allows a single network to handle diverse goal-conditioned behaviors without goal ambiguity. We train on data collected from an expert in the CARLA simulator and evaluate across several experimental protocols, including previously unseen weather conditions and a new town layout. Our approach outperforms prior methods on all protocols and sets a new state of the art for imitation learning in autonomous driving.

Resources

Citation

@inproceedings{codevilla2018cil,
  title     = {End-to-end Driving via Conditional Imitation Learning},
  author    = {Codevilla, Felipe and M{\"{u}}ller, Matthias and Dosovitskiy, Alexey and L{\'{o}}pez, Antonio and Koltun, Vladlen},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2018}
}