Towards Scalable Collision Avoidance in Dense Airspaces with Deep Multi-Agent Reinforcement Learning

Thibault Roux, Filipo Studzinski Perotto, Gauthier Picard

Keywords: Collision Avoidance, Unmanned Aerial Vehicles, Deep Reinforcement Learning, Aeronautics, Multi-Agent Reinforcement Learning

TL;DR: This paper is a prospective analysis that highlights the potential benefits of applying coordination techniques, in particular MARL to the ACAS problem.

Abstract:

Increasing airspace congestion requires the development of robust collision avoidance systems to mitigate the risk of near mid-air col-
lisions between aircraft. The Airborne Collision Avoidance System-X (ACAS-X) is a next-generation solution that provides both better
conflict resolution maneuvers and fewer unnecessary actions compared to the conventional equipment (TCAS-II) currently used in
most commercial and general aviation aircraft. ACAS-X is achieved through dynamic programming for one-to-one aircraft encounters.
However, this solution still faces significant limitations, in particular the restriction to single intruder scenarios and the reliance
on discretized state and action spaces. In this paper, we show that the naive application of ACAS-X to multi-threat scenarios leads
to suboptimal and even catastrophic results. To address these issues, we formalize the multi-agent aircraft collision problem and
argue for the adoption of deep multi-agent reinforcement learning (MARL) techniques, which have the potential to compute optimal
maneuvers in complex multi-aircraft scenarios. Finally, we identify key challenges and open research questions for the multi-agent
aircraft collision avoidance problem.

Type Of Paper: Work-in-progress paper (max page 6)

Anonymous Submission: Anonymized submission.

Submission Number: 12