FlyMirage A Fully Automated Generation Pipeline for Diverse and Scalable UAV Flight Data via Generative World Model

Jinhan Li#,1,2, Xijie Huang#,1,2, Weiqi Gai2, Zhaoqi Wang2, Qiyi He2, Yijin Wang1,2, Mo Zhu1,2, Zhiyang Liu2, Fei Gao1,2, Xin Zhou2, Yuze Wu*1,2, Ziming Ding2

1Zhejiang University   2Differential Robotics * Corresponding author # Equal contribution

Paper (Arxiv) Dataset (Huggingface)
A collection of panoramic indoor and outdoor scenes from FlyMirage
1,100 Scenes
110K+ Trajectories
100% Automated

Pipeline

Dataset Generation Pipeline

FlyMirage dataset generation pipeline, stage 1

The pipeline begins by selecting from a wide range of indoor and outdoor scene types. It creates a detailed description and reference image, then transforms them into a realistic, explorable 3D world.

Stage 1 (World Generation) Outcome

Scenes

Public & Institutional
Commercial & Leisure
Workplace & Office
Scene category distribution Distribution of 1,100 scenes across six environment categories. 1,100 Scenes
  • Public & Institutional 23.6%
  • Commercial & Leisure 23.0%
  • Workplace & Office 13.3%
  • Industrial & Utility 10.6%
  • Residential 20.1%
  • Transportation 9.4%
Industrial & Utility
Residential
Transportation

Stage 2 (Scene Annotation) Outcome

Object Bounding Box

Stage 3 (Navigation & Collection) Outcome

Trajectories

Trajectory Length Distribution

Trajectory length distribution A bar chart showing trajectory counts in two-meter length bins. The largest groups are two to four meters with 48,815 trajectories and four to six meters with 36,115 trajectories. 0 0–2 48,815 2–4 36,115 4–6 18,464 6–8 9,661 8–10 1,004 10–12 85 12–14 47 14–16 84 16+ Length (m) # Trajectories