GARUDA: A Bounded, Human-in-the-Loop Decision Model for On-Device Drone Surveillance
Seerror Technologies, Bengaluru, India · 22 October 2026 · Version 2.0
Official record · DOI 10.5281/zenodo.23572020
Abstract
GARUDA is an on-device AI surveillance drone in development. The principle is: Garuda sees, Garuda thinks, the soldier decides. Two detectors find objects, one for day and one for night. Plain code fuses and tracks them. A small fine-tuned language model, the GARUDA Brain, chooses exactly one of 10 non-kinetic actions and writes a short alert. A plain-code Safety Gate approves every action. PX4 flies the aircraft, and the soldier keeps manual override. No weapon, engagement, or targeting action exists.
What was tested
The Brain is a QLoRA fine-tune of Qwen3-1.7B. Training labels come from a 26-rule standard operating procedure, and every decision row is synthetic. Six training runs used the same settings. On the sealed fresh_test_v2, which was frozen before training, the Brain scores 95.7%, misses 0 of 213 urgent alerts, and still produces 2 unsafe rows. The Safety Gate blocks those 2 rows. The full pipeline scores 95.9% with 0 unsafe rows.
An on-device Q6_K file, 1.42 GB, makes the same decision as the trained model on 150 of 150 hard_test rows in llama.cpp, with 0 missed urgent alerts and 0 unsafe rows. A 4-bit file still loses decisions. The night detector, a YOLOv8n model, reaches 87.3% mAP50 on held-out public images. It has not been run on real drone footage.
| Score | Unsafe rows | Missed urgent alerts | |
|---|---|---|---|
| Brain alone | 95.7% | 2 | 0 of 213 |
| Full pipeline | 95.9% | 0 | 0 |
Status
In development. All decision data is synthetic. The sealed test was written by the same author with the same rules. Nothing has been flight-tested. Nothing has been reviewed, approved, or validated by any army. The action space must never be extended to weapon release, targeting, or engagement.
The report, LaTeX source, training notebooks, and result summaries are on Zenodo. The synthetic datasets and the plain-code pipeline are not public. They are available on request for research review at jay@seerror.com. No trained model weights are included.
Cite
Tiwari, J. (2026). GARUDA: A Bounded, Human-in-the-Loop Decision Model for On-Device Drone Surveillance (Version 2.0). Zenodo. https://doi.org/10.5281/zenodo.23572020