Decentralized Drone Swarm Coordination

The Problem

Coordinating a swarm of drones through tight, obstacle-rich environments is hard. Centralized approaches require every agent to know the full plan - which doesn't scale and creates a single point of failure. If the coordinator goes down, the entire swarm goes down.

The question: can follower drones navigate safely through constrained spaces using only local communication with their neighbors - without any knowledge of the leader trajectories?

Approach

Decentralized affine transformation with leader-follower architecture

The core idea is elegant: place three leader quadcopters at the vertices of a triangle, with follower quadcopters inside. Leaders execute known trajectories. Followers compute their desired positions as weighted sums of their neighbors' actual positions, using fixed barycentric weights calculated once from the initial configuration. No global planner, no trajectory broadcasting.

This affine transformation (AT) framework lets the entire formation contract, translate, rotate, and deform - all while maintaining collision-free guarantees through mathematical constraints on the transformation parameters.

Control system block diagram showing the decentralized affine transformation architecture with leader and follower control loops — Control architecture: leaders track reference trajectories via PD feedback, while followers compute desired positions from neighbor states

Decentralized over centralized coordination

A centralized planner must broadcast full trajectories to every agent, creating bandwidth bottlenecks and a single point of failure. The decentralized AT approach requires each follower to communicate only with its immediate neighbors - the same algorithm works whether you have 3 followers or 300.

WHY → scales to large swarms, no single point of failure, minimal communication

Fixed barycentric weights from initial configuration

Rather than continuously recomputing coordination parameters, each follower calculates its barycentric weights once at startup based on its initial position relative to neighbors. These weights remain fixed throughout the mission - drastically simplifying the onboard computation and eliminating a class of runtime errors.

WHY → compute-once simplicity, no runtime weight updates, proven convergence guarantees

Crazyflie 2.1 as the experimental platform

At 27g, the Crazyflie is small enough to fly six simultaneously in a standard motion capture volume. The open-source Crazyswarm2 framework provides the ROS 2 integration needed for synchronized multi-agent control, and the 2.4 GHz Crazyradio PA handles reliable communication with all agents from a single base station.

WHY → lightweight, open-source, proven multi-agent framework with ROS 2 support

Implementation

Key features

Initial pre-AT formation showing three leader quadcopters at triangle vertices with three followers inside — Initial formation: leaders cf1, cf2, cf6 at triangle vertices; followers cf3, cf4, cf5 positioned inside at 0.75m altitude

Multi-Phase AT Protocol

A 30-second experimental sequence exercising three distinct transformation modes: pure contraction (shrinking the formation by 50%), rigid body motion (translation + rotation as a unit), and precise deformation (asymmetric scaling). Each phase validates a different capability of the decentralized framework.

Dual-Loop PD Control

Leaders and followers run independent PD controllers with feedforward terms at 50 Hz. Leaders use conservative gains (Kp=2.5, Kd=1.5 lateral) for smooth reference tracking, while followers use higher gains (Kp=3.5, Kd=2.0) for tighter formation keeping - compensating for the additional uncertainty of computing positions from neighbor states.

Collision Avoidance Guarantees

Safety is enforced mathematically, not reactively. The minimum eigenvalue constraint on the affine transformation ensures inter-agent distances never drop below the safety threshold. With a safety margin of 0.065m and minimum eigenvalue of 0.37, zero collisions occurred across all experimental phases.

Formation during Phase 1 showing contracted configuration after 50% scaling

Formation during Phase 3 showing asymmetric deformation with different X and Y scaling

Results

By the numbers

Tracking error time series showing leader errors of 4-9cm and follower errors of 8-11cm throughout the 30-second experiment

2D XY trajectories comparing desired paths against actual paths for all six agents

4–9cm Leader Tracking Error (mean)

8–11cm Follower Tracking Error (mean)

0 Collisions

6 Synchronized Agents

50Hz Control Loop Rate

arXiv Published Research Paper

Reflection

The most satisfying result wasn't the tracking accuracy - it was watching followers converge to their correct positions without ever being told where to go. They simply asked their neighbors "where are you?" and computed the answer from fixed weights. That's the power of the affine transformation framework: complex global behavior emerges from simple local rules.

The follower tracking errors (8-11cm) are roughly double the leader errors (4-9cm), which makes intuitive sense - followers accumulate uncertainty from estimating positions through neighbor communication rather than tracking a known reference. If I were to extend this work, I'd explore adaptive gains that tighten as the formation approaches steady state, and scale the experiment to larger swarms where the decentralized advantage becomes even more pronounced.

Publication

Read the paper on arXiv View source code Watch on YouTube