Overview

Aerial VIO — Landing Page

Status: v1 — opinionated draft. Spec / latency claims marked UNVERIFIED. Depth tier: 🌬️ maintainer anchor (written 1.5–2× deeper than other embodiment axes). TL;DR: Aerial VIO is the strictest test of any state estimator — 200 Hz state rate, sub-10 ms end-to-end latency, metric scale with no GNSS fallback, and IMU robustness against propeller-induced aliasing. Three stacks have actually shipped against this bar: VINS-Fusion (HKUST), OpenVINS (UDel / Skydio-adjacent), and the learned outlier DROID-SLAM (Princeton). Everything else is benchmark theater.

---

Why aerial is harsher than ground / manipulation

A tabletop SfM pipeline can take 100 ms and nobody dies. A drone at 15 m/s in a wind gust integrates 1.5 m of position error in that same 100 ms — past the recoverable envelope of a cascaded attitude controller. The non-negotiables:

Requirement	Why	Typical bar
State rate ≥ 100 Hz	Cascaded attitude controller bandwidth	200 Hz on tuned stacks UNVERIFIED
End-to-end latency ≤ 10 ms	Camera → estimate → controller → motor	5–15 ms VINS-Fusion / OpenVINS UNVERIFIED
Metric scale, no fallback	Position controller integrates in meters; throttle compensates gravity in m/s²	<2% scale error post-init
IMU saturation / aliasing	Props excite IMU at 100–400 Hz	Mechanical isolators + 1 kHz IMU + bandpass

This is why aerial state estimation gets its own folder rather than being absorbed into a generic SLAM section. The constraints that matter here — vibration, latency budget, controller coupling — never show up in a manipulation benchmark.

The three production stacks

1. VINS-Fusion (dissection) — Qin et al. T-RO 2018, the open-source baseline. Tightly coupled, optimization-based, sliding-window factor graph. The reference everyone forks.
2. OpenVINS (dissection) — Geneva et al. ICRA 2020, MSCKF formulation. Lower CPU than VINS-Fusion at comparable accuracy. Lineage adjacent to Skydio-class commercial autonomy.
3. DROID-SLAM (dissection) — Teed & Deng NeurIPS 2021, learned dense bundle adjustment. Best accuracy on hard scenes; loses on Jetson-class real-time. The bridge to the VGGT-vs-VIO question.

Each gets its own dissection because the failure modes are non-overlapping — pick by your CPU / GPU budget, your vibration profile, and how much loop-closure / relocalization tolerance you need.

教学基础 (Tutorial layer · 配合上述三栈使用)

• ekf_from_scratch_dissection.md ⚡ NEW — 15-state EKF + 21-state augmented EKF 从零手写教程（取材 HKUST ELEC5660 L8-L9 + proj3, BSD 3-Clause）。与上述三栈互补 — 它们是「拆库」，本文是「从零写自己的 estimator」。读完你能：(a) 上 Jacobian / 协方差传播自己 onboard 的 sensor 组合；(b) 看懂 OpenVINS / VINS-Fusion 内部 trade-off 来自哪。

跨 zone 代碼層對比

• crossing/slam-vio-migration/orb_slam3_vs_vins_fusion_code_comparison.md ⚡ NEW — 從工程哲學到 main entry / pipeline / 前端 / 狀態 / 優化器 / IMU / 初始化 / 回環 8 維對照。讀完你能秒判：要改前端進哪 / 加 factor 進哪 / multi-session 為什麼選 ORB-SLAM3 / GPS fusion 為什麼選 VINS-Fusion。

2026-05 默认推 OpenVINS（VINS-Fusion #3 global-optimization 自 2019 long-open、VINS-Mono 已 stale 无 ROS 2 port）。OpenVINS 是 aerial VIO 区唯一仍正常维护的官方 repo（2025-11 仍 push，issue 回复活跃）。详见 github_failure_atlas.md。

Cross-references

• Cross-embodiment angle: crossing/slam-vio-migration/vggt_vs_drone_vio.md — why VGGT does not replace VIO at aerial latency budgets.
• Sensor stack pairing: crossing/sensor-stack-matrix/ — IMU + camera + (optional) event-camera trade-offs.
• Feed-forward 3D context: foundations/feed-forward-3d/ — what VGGT-class models can and can’t contribute.
• Event-camera complement: ../event-camera/ — when classical VIO breaks (high speed, low light), event sensors are the hedge.

Boundary

This folder dissects per-paper / per-stack mechanics for aerial VIO. Cross-embodiment comparisons (manipulation ↔ ground ↔ aerial ↔ marine) live in crossing/slam-vio-migration/. Sensor physics (IMU noise models, rolling-shutter geometry) lives in foundations/sensor-physics/. Deployment-side topics (calibration, time-sync, mechanical isolation) live in deployment/.