SAT-Edge-Agent

English | 中文

中文

SAT-Edge-Agent 是面向星载边缘智能研究的硬件在环（HIL）系统原型。它将浏览器工作台、FastAPI Agent、遥感有向目标检测工具、本地 OpenAI 兼容语言服务和 SSE 事件流组织为端到端任务工作流。

当前公开内容聚焦两类可复现实验：端侧 Agent 的 HIL 编排，以及结构化结果相对于原始图像的下行 payload 成本。

本文不将检测器精度、通用地理配准、飞行验证或校准能耗作为当前贡献。

系统组成

• React + Vite 操作工作台
• 基于 LangChain/LangGraph 组件的 FastAPI Agent 后端
• 项目内部 YOLO-style OBB 检测服务
• 本地 OpenAI 兼容语言模型服务
• start -> tool -> token -> done 的 SSE 可观测事件链

Browser / operator client
  -> React frontend (5173)
     -> FastAPI Agent (<edge-host>:9001)
        -> YOLO-style OBB service (<edge-host>:8003)
        -> Local OpenAI-compatible service (<edge-host>:8080/v1)

所有公开文档使用 <edge-host> 等占位符。私有 IP、主机名、VM 访问方式、内部路径、具体板卡型号和私有模型身份不进入公开材料。

研究制品

HIL orchestration evidence 包含：

• 20 次单图与 20 次双图串行固定工作负载记录；
• 20 次单图 profiler、19 次完整双图 profiler 和 1 次结构化部分失败记录；
• 脱敏 JSON 与标准化 SSE 示例；
• 可重新计算均值、样本标准差、中位数和 nearest-rank P95 的脚本；
• SHA-256 完整性清单和厂商无关运行环境说明。

运行：

python artifacts/hil_orchestration/scripts/summarize_public_metrics.py

论文 LaTeX 源码位于 manuscript，同时提供可直接导入 Overleaf 的 ZIP 包。

20/100 图 payload-size 与链路速率敏感性结果位于 experiments/downlink_payload/results。

数据集

实验样本来自第三方 FAIR1M Kaggle 镜像：

https://www.kaggle.com/datasets/ollypowell/fair1m-satellite-imagery-for-object-detection

本仓库不重新分发图像文件，只公开下载说明、100 文件清单和 SHA-256 校验值。详见 dataset/README.md 与 dataset/DATA_LICENSE.md。

快速启动

前端：

npm install
npm run dev

Agent 后端：

cd yolo_agent-main
uv run uvicorn backend.app:app --host 0.0.0.0 --port 9001

常用环境变量：

• VITE_AGENT_BASE_URL
• VITE_DEMO_PASSWORD（可选；留空时不显示演示登录页）
• NWPU_VHR_API_BASE
• ASSISTANT_LLM_API_BASE

复现边界

• 检测权重是内部训练资产，当前不公开。
• 精确板卡型号与私有语言模型身份保留在内部证据中。
• 读者可使用兼容的 OBB 检测接口与 OpenAI 兼容本地语言接口替换私有组件。

核心文档：

• Reproducibility Guide
• Model and Runtime Card
• Dataset Notes

---

English

SAT-Edge-Agent is a hardware-in-the-loop (HIL) research prototype for onboard edge intelligence. It combines a browser workspace, a FastAPI Agent, an oriented remote-sensing detection tool, a local OpenAI-compatible language service, and an SSE-observable task workflow.

The public research artifacts cover two reproducible evaluations: HIL edge-Agent orchestration and the downlink-payload cost of structured results relative to raw imagery.

The current release does not claim detector accuracy leadership, general georegistration, flight validation, or calibrated energy efficiency.

Architecture

Browser / operator client
  -> React frontend (5173)
     -> FastAPI Agent (<edge-host>:9001)
        -> YOLO-style OBB service (<edge-host>:8003)
        -> Local OpenAI-compatible service (<edge-host>:8080/v1)

Public documentation uses placeholders such as <edge-host>. Private IP addresses, hostnames, VM access details, internal paths, exact board identity, and private model identity are excluded.

Research Artifacts

The HIL orchestration evidence package provides sanitized request-level CSV files, normalized success and partial-failure SSE examples, redacted JSON, a recalculation script, a vendor-agnostic runtime profile, and a SHA-256 manifest.

python artifacts/hil_orchestration/scripts/summarize_public_metrics.py

The LaTeX source is available under manuscript, together with an Overleaf-ready ZIP archive.

The 20-image and 100-image payload-size/link-rate results are available under experiments/downlink_payload/results.

Dataset

The experiments use samples acquired from the third-party FAIR1M Kaggle mirror:

https://www.kaggle.com/datasets/ollypowell/fair1m-satellite-imagery-for-object-detection

Image files are not redistributed. The repository includes acquisition notes, a 100-file manifest, and SHA-256 checksums. See dataset/README.md and dataset/DATA_LICENSE.md.

Getting Started

Frontend:

npm install
npm run dev

Agent backend:

cd yolo_agent-main
uv run uvicorn backend.app:app --host 0.0.0.0 --port 9001

Common overrides:

• VITE_AGENT_BASE_URL
• VITE_DEMO_PASSWORD (optional; leave empty to disable the demo login gate)
• NWPU_VHR_API_BASE
• ASSISTANT_LLM_API_BASE

Reproducibility Boundary

• The detector weight is an internal training asset and is not released.
• The exact board and private language-model identities remain in internal evidence.
• Compatible OBB and OpenAI-style local endpoints may replace the private components.

Core documents:

• Reproducibility Guide
• Model and Runtime Card
• Dataset Notes

License

Project code and evidence-processing scripts are released under the Apache License 2.0. Third-party FAIR1M/Kaggle images, labels, geographic fields, and derived data elements remain governed by the source terms described in dataset/DATA_LICENSE.md.