Learning-to-Fly (L2F): Learning-based mission-aware Collision Avoidance

Overview

In order for fleets of drones managed by different operators to share an airspace, they require a fast, reliable and fair Collision Avoidance (CA) approach to ensure safety. This project aims to develop such methods by combining learning-based decision making and decentralized Model Predictive Control to perform on-the-fly, predictive Collision Avoidance (CA) between two (co-operative) drones. In addition to performing CA in real-time, L2F also ensures that the drones do not violate their original mission requirements. The video below shows a simulation where two pairs of drones are have conflicting pre-planned trajectories, and also shows the resulting safe trajectories obtained via L2F.

Publications

• Mission-aware, Learning-to-Fly: Learning-based Collision Avoidance for Scalable Urban Air Mobility
  Rodionova, Pant, Jang, Abbas, Mangharam
  IEEE International Conference on Intelligent Transportation Systems (ITSC), 2020.

• Learning-to-Fly RL: Reinforcement Learning-based Collision Avoidance for Scalable Urban Air Mobility
  Jang, Pant, Rodionova, Mangharam
  AIAA/IEEE Digital Avionics Systems Conference (DASC), 2020.

Contributors

Rodionova, Pant, Jang, Abbas, Mangharam

Citation

@misc{rodionova2020learningtoflylearningbasedcollisionavoidance,
      title={Learning-to-Fly: Learning-based Collision Avoidance for Scalable Urban Air Mobility}, 
      author={Alëna Rodionova and Yash Vardhan Pant and Kuk Jang and Houssam Abbas and Rahul Mangharam},
      year={2020},
      eprint={2006.13267},
      archivePrefix={arXiv},
      primaryClass={eess.SY},
      url={https://arxiv.org/abs/2006.13267}, 
}