August 14, 2026 · 3 min read
Swarm-Steward: Talking to Drone Swarms in Plain Language
Our Swarm-Steward system lets a single non-expert operator coordinate fleets of drones and ground robots using natural language — no joysticks or programming required. The key idea is a hierarchical LLM-based multi-agent design that separates planning from actuation: specialized agents ground each request in real map features and telemetry, while a deterministic safety gate (geofencing, altitude and separation limits) vets every command before execution, with optional operator preview. To keep reasoning scalable, we use dual retrieval-augmented generation over map features and telemetry state — effective even with 10,000 map features. In end-to-end tests with swarms from 5 to 500 drones, LLM token cost stayed nearly constant, with 92.9% task success across 280 prompts.
A. Jarabo-Peñas, J. Bravo-Arrabal, E.G.A. Rolland, A.L. Christensen