Decentralized Drone Swarm Coordination

Experimental validation of decentralized affine transformation for multi-agent aerial systems

First author Published 2026
Crazyflie 2.1 ROS 2 Vicon Crazyswarm2 Python Multi-Agent

Keeping a drone formation safe while it changes shape is hard. When the formation contracts, the drones close in on each other, and each one still has to hold its place. The work each follower does should also stay the same as the swarm grows.

The question: can each follower work out where it should be from the measured positions of its neighbors and a set of fixed weights, while the formation contracts, rotates and deforms?

Decentralized affine transformation with leader-follower architecture

Three leader quadrotors sit at the vertices of a triangle, and the followers stay inside it. Each follower computes its desired position at every control step from the measured positions of its neighbors. The weights it uses are fixed, set once from the initial configuration as barycentric coordinates.

This affine transformation (AT) framework lets the whole formation contract, translate, rotate and deform. Safety in the experiment was shown by measurement, not guaranteed by the paper's theorem, because the flown scaling of 0.5 sits below the theorem's bound.

Decentralized over centralized coordination

A follower works from the measured positions of its neighbors, and the per-follower computation does not grow with the size of the swarm. The experiment used six drones. The paper lists larger teams as future evidence of scalability, and it states the limits of the test setup: a centralized Vicon system and a ground station relaying commands.

WHY → per-follower computation is O(1), whatever the size of the swarm
Fixed barycentric weights from initial configuration

Each follower's communication weights are barycentric coordinates, set once from the initial configuration. They stay fixed for the whole flight. At every control step the follower combines them with the measured positions of its neighbors to get its desired position.

WHY → weights set once from the initial configuration, then left fixed
Crazyflie 2.1 as the experimental platform

The experiment used six Crazyflie 2.1 quadrotors, run with Crazyswarm2 on ROS 2. Commands reach the drones over a 2.4 GHz Crazyradio PA link. Vicon motion capture runs at 100 Hz, and the nominal altitude is 0.75 m.

WHY → Crazyswarm2 on ROS 2 for commands, Vicon motion capture at 100 Hz for position

Key features

A 30-Second Flight
Each test flight lasts 30 seconds and goes through a narrow passage. It starts with takeoff (0-5 s). The formation then contracts with uniform scaling from 1.0 to 0.5 (5-10 s), goes through rigid-body translation and rotation while holding the contracted shape (10-25 s), and ends with non-uniform deformation at scaling 0.6 and 0.9 (25-30 s).
Leader and Follower Controllers
Leaders use PD feedback with a feedforward acceleration term. Followers use PD feedback only, with zero feedforward. Lateral gains are Kp 2.5 and Kd 1.5 for leaders, Kp 3.5 and Kd 2.0 for followers.
50 Hz Control Loop
The controllers run at 50 Hz, and the computation uses 5.23 ms of each 20 ms cycle. Commands go out through the Crazyswarm2 cmdFullState interface.
Separation Measured in Flight
Safety in the experiment was shown by measurement, not guaranteed by the theorem: the flown scaling of 0.5 sits below the theorem's bound. The smallest separation between any two drones over the whole flight was 0.154 m. The safety threshold is 0.13 m, and no pair of drones came inside it.

By the numbers

5.7 cm Leader tracking error, mean RMSE across three valid flights
8.4 cm Follower tracking error, mean RMSE across three valid flights
0.154 m Smallest separation between any two drones (threshold 0.13 m)
6 Crazyflie 2.1 quadrotors
50 Hz Control loop, 5.23 ms of each 20 ms cycle used
AST 2026 Published in Aerospace Science and Technology, vol. 180, art. 113539

The followers were never handed a path. Each one works from the measured positions of its neighbors and a set of fixed weights, through contraction, rigid-body motion and deformation.

Follower error came out higher than leader error: 8.4 cm against 5.7 cm, mean RMSE across three valid flights. That makes sense to me. Followers act on their neighbors' measured positions rather than a planned path. If I were to extend this work, I'd explore adaptive gains that tighten as the formation approaches steady state. What came next was a bigger swarm: I flashed and configured a 10-drone Crazyflie swarm on two Crazyradio 2.0 radios and flew all 10 together, for a paper in preparation.

Mazmanyan, G., and Rastgoftar, H. "Decentralized Affine Transformation for Scalable and Safe Multi-Agent Aerial Coordination." Aerospace Science and Technology, vol. 180, art. 113539, 2026.