Swarm-Steward SAR Perception-to-Response Artifacts: AerialPerson-YOLO26l, Dataset Splits, and Field Evaluation Data

Supplementary data and models for the paper "One Operator, Many Robots: Natural-Language Command of UAV–UGV Teams for Search and Rescue." This record accompanies a real-world search-and-rescue (SAR) extension of Swarm-Steward, a voice-driven multi-robot coordination system that links multi-UAV aerial search, live aerial victim detection and geolocation, operator confirmation, and UGV deployment. It contains the trained aerial person-detection models, the definition of the AerialPerson-Mix training set, and curated data from twelve outdoor SAR search experiments, organized into three components: A — Detector. The deployed AerialPerson-YOLO26l person detector (PyTorch), the full YOLO26 n/s/m/l/x model-size sweep, a confidence-threshold selection study, and supporting reports, figures, and analysis scripts. On the balanced validation split the deployed model reaches mAP50 = 0.832 and mAP50–95 = 0.501. B — AerialPerson-Mix dataset definition. A balanced, single-class (person) aerial detection set merging ~1,000 images each from SARD, HERIDAL, LaDD, and NOMAD (3,985 images, 80/20 train/val split). Source images are not redistributed; we provide converted single-class YOLO labels, the dataset configuration, and per-source manifests listing exactly which original images were used, so the set can be reconstructed from the original providers. C — Search experiments. Per-run manifests, per-drone telemetry summaries, gimbal/altitude precision tables, and candidate records for the twelve outdoor experiments, covering four victim placements (A–D), one/two/four-UAV teams, different gimbal angles and search speeds, over a ~199 m × 238 m unstructured area (47,362 m², 4.74 ha) flown with DJI Mini 4 Pro UAVs. Licensing: data, labels, manifests, tables, reports, figures, and scripts under CC-BY-4.0; model weights under AGPL-3.0 (consistent with the upstream Ultralytics YOLO framework). The AerialPerson-Mix definition does not relicense any third-party imagery; SARD, HERIDAL, LaDD, and NOMAD each retain their own license and terms of use. If you use this data or these models, please cite the accompanying paper and this record.