Skip to content Skip to navigation

Game of Drones

Game of Drones

Make drones race faster than everyone else

Game of Drones is a NeuRIPS 2019 competition with the goal to push the boundary of building competitive autonomous systems through head-to-head drone races.

The challenge is to race a quadrotor drone in simulation through a series of gates against an opponent.
There will be two drones on the same track at the same time that are allowed to block each other.
The goal is to finish the race quicker than the opponent.

The competition comes in three tiers of different difficulty and level of autonomy.

  • Tier 1: Planning only.
    Given: Ground-truth positions of your and your opponent's drone.
    Challenge: Finish the track quicker than your opponent, without crashing.
  • Tier 2: Perception only.
    Given: Noisy sensor input (RGB from front and downward cameras, IMU)
    Challenge: Finish the track as quick as possible, without crashing.
  • Tier 3: Full autonomy, combined challenge of Tier 1 and 2.
    Given: Noisy sensor input (RGB from front and downward cameras, IMU)
    Challenge: Finish the track quicker than your opponent, without crashing.

What's special about this competition?

Previous drone race challenges, such as the IROS autonomous drone race challenge or the AlphaPilot challenge only need you to race against the clock.  In contrast, here, you'll have to race against an opponent on the same track. This means you can block your opponent if you are ahead and you have to overtake him when you are behind creating a whole range of interaction your policy has to handle.

To broaden the audience and bring people to drones from all kind of disciplines, all racing will be done in simulation. However, we still require you to solve the challenge of state estimation under noise from simulated cameras and an inertial meausurement unit (IMU = accelerometer + gyroscope).

Why should I participate/care about the game of drones?

Besides being fun, this competition provides a fair comparison for the current state of the art.  And if your current focus is either more on the planning or perception side, you'll most probably learn a thing or two. The audience is expected to be large -- NeuRIPS is the largest AI conference after all, so you can expect to get some attention if your algorithm scores high. There will also be some monetary incentives -- a total of more than 10'000 $ in prices.


Prizes for qualification rounds: To encourage a broad audience to participate, and to reach people outside of the research community of NeurIPS that will be present during the live competition, we offer the following prizes to the top three teams at the end of the qualification round.

Qualification PlacementPrize
1st1'000 $
2nd500 $
3rd250 $


Prize for overall winners: The prizes (in USD) for the top three policies as determined
during the live competition are as follows.

PlacementTier ITier IITier III
1st1'000 $2'000 $3'000 $

500 $

1'000 $1'500 $
3rd250 $500 $750 $


All prizes are in USD.

Who is behind this?

The organizing team behind the game of drones are the developers of AirSim from Microsoft Research and robotics researchers from Stanford who share the vision of a simulator realistic enough to provide a training and testing environment for autonomous mobile robots such as drones or self-driving cars.

Many people are helping to make this happen, but the core team consist of

  • Mac Schwager, Assistant Professor at Stanford University and Director of the Multi-Robot Systems Lab,
  • Ashish Kapoor, Principal Research Manager at Microsoft Research where he founded and leads the Aerial Informatics and Robotics group, and
  • Tim Taubner, a Master student from ETH Zurich currently visiting MSL.

If you have questions or suggestions with regard to this competition please write Tim a mail at:

Getting started



  • End of April, 2019: Track design, drone, and sensor parameters.
    We will publish the drone parameters (dynamics) and the sensor parameters here.
    Also, an executable with one
  • End of May, 2019: Release of the baseline algorithm.
    The Python source code of the baseline algorithm (for Tier I) will be released and serves as a starting point.
    The race module for the simulator will be released to enable participants to evaluate there algorithms against the baseline locally.
  • End of July, 2019: Release of reference algorithm.
    We'll release our submission as an opponent in the online interface to benchmark against. Note that the code will not be made available.
  • July -- 21st November, 2019: Qualification Round.
    All submissions made in the timeframe above will be considered for qualification.
  • End of August, 2019: Freeze of reference algorithm.
    Last possible change to the reference MSL submission, to allow participants enough
    time to adjust their algorithms.
  • 15th October, 2019: Detailed qualification rules.
    The detailed qualification rules will be released, specifying what exactly has to be passed to be qualified for the placement rounds.
  • 21st November, 2019: End of qualification.
    Announcement of qualified teams for the live challenge, invitation for poster presen-
  • December, 2019: Live Competition at NeurIPS 2019.
    All present qualified teams participate in the live competition.


The competition will be carried out in the photo-realistic simulator AirSIM developed by Microsoft Research.
We will release executables containing the environment end of April.

Drone parameters

Tbd. Each drone will have a spherical marker of known size on top, to facilitate target tracking.

Sensor parameters

There will be three sensors on the drone: Front- and downward facing camera and an IMU.

Track environments

There will be three track environments:

  1. Easy: Blank room.
  2. Medium: Green soccer field.
  3. Hard: Indoor building environment.


Detailed rules will be published here.