skip to content

Parallelized Real-Time Drone Swarm Coordination

[in-progress]CUDA, PX4, ROS 2, Gazebo, Python

Project Overview

A project I am currently working on in collaboration with TAI (Turkish Aerospace Industries, a.k.a. TUSAŞ), and soon TÜBİTAK. This is my senior project — currently in the development phase.

Goals

  • Autopilot: PX4 (the open-source flight stack used widely in research drones).
  • Testing environment: Gazebo (the de facto open-source robotics simulator).
  • Parallelization: CUDA for collision avoidance and formation flight algorithms.
  • Topology: Centralized swarm controller — one ground-station node directs the swarm; individual drones execute and report state.

Status

In active development. The architecture is largely settled but most of the CUDA kernels and PX4 integration shims are still under construction. Expect updates as the project progresses through summer 2026.

What I'm Learning

  • ROS 2 and PX4 are both massive ecosystems. The hardest part of swarm work isn't the algorithms — it's getting all the message types, transport layers, and simulation timing to agree.
  • CUDA is overkill for ~10-drone swarms but starts paying off around 50+ drones, where pairwise collision checks scale quadratically. For our target swarm size, the parallelization is more about future-proofing than current necessity.
  • Centralized vs. decentralized swarm coordination is a genuinely interesting tradeoff. Centralized is easier to reason about and easier to debug, but it scales worse and has a single point of failure. We're going centralized for v1 because debugging matters more right now than scaling.