UAV-SwarmAgent: Scalable Multi-UAV Reconnaissance and Task Allocation via Code Generation and Vision-Language Reasoning

What it claims

UAV-SwarmAgent is a late-breaking report (arXiv, October 2026) from Skoltech and The Hong Kong Polytechnic University. It turns one operator sentence into routes for one to five UAVs, then returns incident locations with camera frames. The finding most useful for planning is the scaling law. Flight time fits T(N) = 332 + 174/N s, so 34% of single-UAV flight time parallelises (95% CI 0.29 to 0.40). The rest is the commute from base to the area and back. The second UAV delivers 50% of the attainable saving and five UAVs deliver 80%. Total mission time, including agent overhead, shows no significant gain beyond two UAVs.

How it works

The system is built on the open-source smolagents library and has two tiers. The Airspace Management Agent (AMA) runs at the ground station in a ReAct loop whose actions are short Python programs run in a sandbox. Route allocation, flight simulation and delegation are deliberately one fixed tool, dispatch_swarm, not model-written code. In an earlier prototype one model silently rescaled waypoint coordinates and shifted every route.

flowchart LR
  S[Operator sentence] --> AMA[AMA: writes Python]
  AMA --> G[VLM grounds waypoints on archival map]
  G --> D[dispatch_swarm fixed tool]
  D --> U1[UAV agent 1]
  D --> UN[UAV agent N]
  U1 --> V{Verifier p > tau?}
  UN --> V
  V -- yes --> R[Report frame + coords]
  V -- no --> X[Discard]
  R --> AMA

1. Grounding. The map is letterboxed to a 1000 × 1000 grid. Objects such as buildings get one centre point. Areas such as forest or water get an even spread of points. At most 15 points are kept. Answers with equally spaced points on one axis are discarded as fabricated.
2. Allocation. The method minimises the longest route rather than total distance. Waypoints are taken farthest-first, and each goes to the UAV whose closed route stays shortest. Each route is then ordered nearest-neighbour.
3. Verification. Valen-Preview-0923, a 2B-parameter VLM, scores a 128 × 128 px frame (about 170 m) every 16 m. It returns P("yes") in one forward pass without generating text. The per-scenario threshold τ is 0.55 to 0.74.

Evidence

The setup is 255 simulated missions over four synthetic incident types: structure fire (30 maps), wildfire, vehicle collision and oil spill (5 maps each). Maps are about 1.4 × 1.4 km, derived from the UAV-VLA benchmark plus coastal web imagery. With N = 3, success was 41/45 for gemini-3.8-flash, 44/45 for gpt-6.1-sol and 42/45 for qwen3.8-max. Median agent times were 152 s, 180 s and 135 s. The authors say the between-model differences on 5-map scenarios are within chance.

Flight time fell from 505 s (1 UAV) to 363 s (5 UAVs). Total time fell from 657 s to 533 s, because agent time grows 6.7 s per UAV from a 140 s intercept. Full strip coverage would take 2432 s and 719 s at those sizes. A power law N^-0.20 fits about equally well (AIC difference 0.8).

Missions returned a median of 37 reported frames. Median precision ranged from 0.57 (wildfire) to 0.95 (oil spill). Of 11 failures, 5 occurred at planning, where no waypoint was placed near the incident and at most 12 of 15 allowed waypoints were used. In the other 6, the drone flew over the incident but the verifier missed it.

How to use it

The page names models and hardware but gives no code link, weights or licence. The verifier ran on a laptop RTX 4050 (6 GB), and the authors say it could run on an embedded Jetson. Planning models are called through an OpenAI-compatible gateway.

Limits

This is simulation only: synthetic composited incidents and a kinematic flight model with no wind, batteries or obstacles. Verifier questions and thresholds were tuned on the benchmark maps with no held-out set. The collision and oil-spill questions were refined and rerun for Gemini only. Each configuration ran once despite LLM variance, and operator sentences were fixed. No human participants were involved, so every operator claim is a hypothesis.