Research · 13 min read
What 1,000 Drones Need to Coordinate: UWB Is Non-Negotiable Above 100
May 19, 2025
Scale a drone swarm from 32 to 1,000 units and something breaks. Not gradually — it breaks hard. Camera-only coordination falls 15.8 percentage points below the omniscient baseline at 1,000 drones and produces 8× more collisions: 3,620 versus 452 in the same mission. That is not a tuning problem. It is an architectural one. UWB ranging is non-negotiable above roughly 100 drones.
Four Sensing Configurations
- Camera-only: onboard vision for neighbor detection and obstacle avoidance.
- UWB-only: ultra-wideband radio ranging — precise distances, independent of lighting or occlusion.
- Hybrid (camera + UWB): both modalities combined.
- Omniscient baseline: perfect ground-truth position knowledge — the ceiling.
At 32 drones the differences are negligible. Camera-only achieves 57.5% coverage, nearly matching the 57.6% omniscient ceiling. This is the regime where most drone swarm demos live. The problem is not visible yet.
By 200 drones, camera-only has opened a 19.2 percentage-point gap (40.0% vs. 59.2%, p<0.001). UWB and Hybrid stay within 1–2 pp of omniscient at every scale. The separation is structural: cameras cannot maintain reliable neighbor awareness when mutual occlusion becomes the norm. UWB ranges through everything.
The Urban Penalty
Urban environments impose a 14–32 percentage-point coverage penalty versus open forest across all swarm sizes and sensing configurations. Two effects compound: visual range is effectively halved (30 m → 15 m in city clutter), and 22% of the map is physically occupied by buildings — coverage saturates at ~28% regardless of mission time.
At 1,000 drones, city environments produce 8–11× more collisions than forest under identical sensing. Camera-only in a city at scale is not just suboptimal — it is operationally unacceptable.
Hybrid Squad Coordination: Centralized Performance at 3.3× Less Overhead
In 16-drone, 8-target experiments:
- Centralized: 95.0±6.1% detection, 3.15 s reaction, 192 msg/s.
- Decentralized: 85.0±5.0% detection, 0.44 s reaction.
- Hybrid squad-based: 95.0±6.1% detection (matching centralized), 1.08 s reaction, 58 msg/s — 3.3× less communication overhead.
For a 1,000-drone fleet: centralized at 192 msg/s per node generates ~192,000 messages/s fleet-wide; hybrid brings that to ~58,000 msg/s. At 200 bytes per message, that is the difference between 38 Mbps and 11 Mbps of coordination traffic before any payload data.
Resilience
GPS denial: with UWB present, coverage drops only 1.3 pp in city. Fleets carrying only GPS lose this fallback entirely.
Communications denial: any architecture that relies on shared state collapses to camera-only performance. UWB hardware provides no benefit when the ranging data cannot be shared.
75% drone attrition: forest coverage holds at 87.7%; city drops to 63.8%. Urban operations require 30–40% more drones to maintain equivalent resilience under attrition.
VLMs on Edge Hardware
SmolVLM-2B on Orin Nano (210 ms latency, 5.06 GB VRAM) provides a 25 pp detection advantage under total comms denial in forest (62.5% vs. 37.5%, p<0.05). In city environments, this advantage disappears: building occlusion prevents the scene context VLM reasoning depends on. Spend those resources on UWB hardware and better mesh radio for urban operations.
Requirements Summary
100-drone fleet: UWB on every airframe (~$20–50/unit), hybrid squad organization (squads of 8–16), mesh radio for 58+ msg/s per drone, GPS primary / UWB fallback.
1,000-drone fleet: UWB without exception, tiered coordination hierarchy (squad → sector → global), 30–40% reserve capacity for urban attrition, comms-denial firmware on every unit.
Full methodology and per-configuration data: Scaling the Separation Principle.
Technical paper
Scaling the Separation Principle: Sensing Requirements for 1000-Drone Swarms in Urban and Natural Environments
Technical report
TL;DR
- Controlled swarm simulations up to 1,000 agents comparing sensing stacks and coordination architectures.
- Camera-only swarms lose 15.8 percentage points of coverage and see 8× the collision rate at 1,000 drones.
- UWB ranging becomes non-negotiable above ~100 drones.
| Max swarm size | 1,000 |
|---|---|
| Coverage drop (camera-only) | 15.8 pp |
| Collision-rate increase | 8× |
| UWB threshold | ~100 drones |
Abstract
Controlled swarm simulations up to 1,000 agents comparing sensing stacks and coordination architectures, with a focus on when ultra-wideband ranging becomes necessary as fleet scale and environment difficulty increase.
Most drone autonomy research is done on a single drone, maybe a few. The interesting engineering questions — and the practical deployment risks — change qualitatively when you scale to hundreds or thousands. This paper measures those changes in controlled simulation, with a specific focus on sensing stack requirements.
The core question
Can a camera-only swarm coordinate effectively at scale, or does it require inter-drone ranging (UWB or equivalent) above some fleet size? The practical stakes are real: UWB adds hardware cost, power consumption, and regulatory considerations. If camera-only works at 1,000 drones, the simpler sensing stack is the right answer. If it fails at 100, the ranging hardware is non-negotiable.
What we measured
We ran controlled simulations with fleet sizes from 10 to 1,000 drones across urban and natural environments, comparing four sensing configurations:
- Camera only
- Camera + UWB ranging
- Camera + GPS (reference case)
- Full sensing stack
Primary metrics: mission coverage (fraction of target area reached), collision rate (inter-drone and obstacle), and coordination latency (time to reach consensus on conflict resolution).
Results
Below roughly 100 drones, camera-only and ranging-equipped swarms perform similarly. The gap opens above that threshold and widens steeply through 1,000.
At 1,000 drones, camera-only swarms show a 15.8 percentage-point drop in coverage compared to ranging-equipped swarms, and an 8× increase in the collision rate. The mechanism is predictable: without inter-drone ranging, each drone's estimate of its neighbors' positions degrades as swarm density increases. Relative position uncertainty compounds across the planning graph, and conflicts that would have been resolved cleanly at small scale result in physical collisions at large scale.
The practical recommendation from this data: for swarms above ~100 drones, treat UWB ranging as infrastructure, not an option. Below that threshold, camera-only coordination is sufficient for most mission profiles.
What simulation can and cannot tell us here
These results are from controlled kinematic simulation. Real swarms add radio congestion, asymmetric packet loss, and physical occlusion effects that the simulation does not model. We expect the qualitative findings — that ranging becomes necessary above a threshold — to hold at real scale, but the threshold number (100 drones) should be read as an order-of-magnitude estimate, not a precise specification.
The companion blog post covers the implications in more operational depth: What 1,000 drones need to coordinate: UWB is non-negotiable above 100.