ggSwarm: Decentralized UAV Swarm Control

Decentralized Formation Control for Autonomous UAV Swarms via Graph Neural Networks and Trajectory Optimization.

Overview

ggSwarm is a scalable, fault-tolerant coordination policy for multicopter swarms. It replaces centralized controllers with decentralized logic where each agent makes autonomous decisions based on local neighborhood data.

Core Features

• Brain (GNN Policy): Graph Attention Networks (GATv2) trained via MARL for permutation invariance.
• Muscles (Trajectory Optimization): Minimum Control (MINCO) trajectory representation for smooth maneuvers.
• Resilience (Consensus Logic): SwarmRaft consensus algorithm for state integrity.
• Showcase (Simulation): NVIDIA Isaac Sim 5.0 & Isaac Lab (OmniDrones).

Directory Structure

• exts/: Custom Omniverse extensions.
• scripts/: Training and evaluation entry points.
• configs/: YAML-based environment and hyperparameter definitions.
• assets/: URDF/USD models and environments.
• research/: Technical documentation and reports.
• weights/: Trained model checkpoints.
• src/ggSwarm/: Core Python package.

Setup

(Details to be added during the planning phase)