Crazyswarm: A Large Nano-Quadcopter Swarm

3 min read

Crazyswarm flies 49 Crazyflie 2.0 nano-quadcopters indoors in dense formation from one base station. A Vicon system supplies raw marker point clouds, and a custom ICP tracker follows vehicles that all carry the same marker pattern. The base station broadcasts compressed poses over three shared radios, and each vehicle runs its own EKF, trajectory evaluator and position controller at 500 Hz. Most of the computation sits onboard, so a dropped packet costs a little accuracy instead of a crash.

Findings

  • Each vehicle weighs 33 g with markers and LED board (27 g bare), measures 92 mm motor to motor, and flies about six minutes.
  • Poses are packed as 24-bit fixed-point positions plus a 32-bit compressed quaternion, so two updates fit in one 32-byte packet and one USB request carries poses for up to four vehicles.
  • End-to-end latency grows roughly linearly with swarm size and reaches 26 ms for 49 vehicles, of which the motion-capture system contributes up to 14 ms.
  • Mean tracking error stays below 2 cm per vehicle with the full swarm flying.
  • Hover error rises from 0.58 cm at a 100 Hz position update rate to 2.18 cm at 2 Hz, which is the authors’ evidence that onboard estimation rides through communication loss.
  • The tracker caps ICP at five iterations to run at 75 Hz with 49 robots, and rejects alignments that imply physically implausible linear or angular velocities.
  • Swarm commands are idempotent and repeated; sending them 5 times every 3 ms was enough to reach all vehicles in a test of 10,000 packets to six Crazyflies.

Limits

  • The authors call the motion-capture system a single point of failure, and the tracker assumes no prolonged occlusions.
  • Vehicle identity comes from a configuration file of initial positions, so a vehicle placed in the wrong spot is mislabelled.
  • Onboard computation makes in-flight debugging hard: telemetry bandwidth and onboard log storage are small (authors admit).
  • Their own future work list is detecting controller failures, better recovery when object tracking fails, and planning that accounts for downwash.
  • The paper does not discuss authentication of the broadcast pose packets.

Relevance

Crazyswarm is the default testbed for indoor aerial swarm experiments. Every pose comes from one central tracker, so a tracker fault hits all 49 vehicles at once, and the motion-capture room is a single point of failure. See also giernacki-2017-crazyflie for the vehicle.