# ASPEEN — Aerial Semantic PErception ENgine

> A re-promptable aerial-perception SDK that turns live or recorded drone video into an
> operator-steerable, depth-aware **decision stream** — with zero-retrain domain adaptation.

ASPEEN is a product by [Flyvercity](https://flyvercity.com), part of the larger COGNEST effort.

This is the Markdown version of <https://aspeen.flyvercity.com> (the "For Agents" view),
provided for AI agents and crawlers under the [llms.txt](https://llmstxt.org) convention.

---

## What it is

A real-time, re-promptable aerial-perception product that converts a live or recorded drone
RGB feed into an operator-steerable actionable decision stream:

- **Open-vocabulary vision** — detect, segment, and track arbitrary classes from a text or
  exemplar prompt in a single multiplexed pass.
- **Depth-aware objects** — a depth enrichment plugin adds distance-to-object, fused per
  detection for range-to-target downstream.
- **Situational interpretation** — a vision-language model elevates detections into a
  plain-language "what is happening" stream.
- **Plugin architecture** — enrichment capabilities hot-attach to the core loop over one
  shared mask + ROI contract.

The operator re-targets what the system looks for by **prompting** — no retraining, no new labels.

## Core thesis — re-prompt, not retrain

Conventional aerial computer vision is closed-set: every new geography, weather condition, or
target class means a retraining loop with fresh labels before the system is usable. ASPEEN
adapts to a new operating domain by re-prompting an open-vocabulary model at runtime —
delivering usable detection with zero retraining and near-real-time performance on aerial
RGB video.

| Closed-set CV | ASPEEN |
| --- | --- |
| Fixed, pre-trained label set | Open vocabulary, prompt-driven |
| Retrain per domain | Re-target live, no retrain |
| New labels for every new task | Adapts to geography, weather, targets |

## How it works — a tri-directional adaptive loop

Not a linear `source → model → user` chain. Three coupled poles, all adapting live:

1. **Input — dynamic sources:** live feeds can change or switch mid-run.
2. **Compute — adaptive model:** re-prompted at runtime; no retrain, no restart.
3. **Human — adaptive prompts:** the operator re-steers intent live.
4. **Output — decision stream:** continuous, operator-actionable interpretation.

## Markets

Primary frame: **UAS for first responders** — rapid search, situational awareness, and incident
mapping. A dual-use path runs from long-range inspection through automated agriculture to
public-safety operations, with real-time aerial perception throughout.

- **First response** — rapid search, situational awareness, and incident mapping.
- **Inspection** — long-range infrastructure survey.
- **Agriculture** — automated crop and field monitoring.
- **Public safety** — scene assessment and coordinated response.

Target classes: humans, vehicles, buildings, surface type — extensible by prompt, not fixed
by a trained label set.

## Deliverable & posture

- **Form:** an SDK — a library/API you integrate into your own stack, not a turnkey app.
- **Performance:** real-time on commodity end-user GPUs (~RTX 4090 / 32 GB VRAM class).
- **Adaptation:** demonstrated by live re-prompting on the same feed — zero retraining, zero new labels.
- **Edge:** edge deployment is a viable future migration path.

## Demo

A real-time annotation demo (ASPEEN annotating drone video in the `local` real-time pipeline)
is available at:
<https://media.flyvercity.com/aspeen-aerial-perception-demo.mp4>

## Contact

Building aerial autonomy, inspection, or first-responder operations?
Request an SDK evaluation: <mailto:info@flyver.city>

Company: [Flyvercity](https://flyvercity.com)