DDA: Deep Drone Acrobatics

Abstract

Performing agile maneuvers such as power loops, barrel rolls, and their combinations with quadrotors is extremely challenging. While expert human pilots can execute these at impressive speeds and reliability, autonomous systems have been far from matching them. We present a system that learns acrobatic maneuvers using deep reinforcement learning trained exclusively in simulation and deployed zero-shot on a physical quadrotor. Key to our approach are a highly accurate physics simulator, comprehensive domain randomization, and a network architecture that effectively encodes maneuver dynamics. We demonstrate power loops, barrel rolls, and combinations thereof on a physical quadrotor, achieving performance comparable to or exceeding expert human pilots. The work was nominated for Best Paper at RSS 2020.

Resources

Video

Citation

@inproceedings{kaufmann2020dda,
  title     = {{DDA}: Deep Drone Acrobatics},
  author    = {Kaufmann, Elia and Loquercio, Antonio and Ranftl, Rene and M{\"{u}}ller, Matthias and Koltun, Vladlen and Scaramuzza, Davide},
  booktitle = {Robotics: Science and Systems (RSS)},
  year      = {2020}
}