WorldRover

A Scalable Synthetic Video Data Engine for World Exploration with Rich Annotations

Xiaojie Xu1,2,*   Zhengyuan Lin1,2,*   Runyi Li1,2   Yihao Liu1   Kaipeng Zhang1,†   Yongtao Ge1,†

1 Alaya Lab · 2 The University of Tokyo · * Equal contribution · † Corresponding author

Resources
Paper Data Code
Highlights
  • A scalable UE data engine that ingests UE assets, builds multi-style scenes, plans routes, and renders them headless with streaming post-processing.
  • The same exploration replayed from first-person, third-person and 360° panoramic cameras under different environmental states.
  • Sequences pair RGB with metric depth, optical flow, long-range point tracks, camera trajectories and action signals throughout each exploration.
21.9M
Frames
6,003
Sequences
32
Environments
202.7 h
Duration
18.7 TB
Volume
WorldRover-10M at a glance: diverse viewpoints, scene styles and motion labels.
WorldRover-Engine pipeline, each stage highlighted in turn: asset processing on a Windows editor host, content generation headless on a multi-GPU Linux host.

Multi-View

Three observations of a route with matched timing and geometry.

First-person
Third-person
360° panoramic

Multi-Modal

Colour, depth and motion, one rasterization of the same instant.

RGB video
Metric depth
Optical flow
Long-range point tracks

Multi-Style

Geometry and motion remain fixed while illumination or texture changes.

Day
Golden hour
Night
Snow
Storm
White model · Venice
White model · Train station
White model · Village

Multi-Scene

30+ artist-built UE scenes, interior to city scale.

Paris
Cyberpunk
Mayan
Island city
Art nouveau
Magic school

Multi-Character

70+ animated humanoids, animals and creatures.

Samurai
Archer
Shadow
Wolf
Eagle

BibTeX

@article{worldrover2026,
  title         = {WorldRover: A Scalable Synthetic Video Data Engine
                   for World Exploration with Rich Annotations},
  author        = {Xu, Xiaojie and Lin, Zhengyuan and Li, Runyi and
                   Liu, Yihao and Zhang, Kaipeng and Ge, Yongtao},
  journal       = {arXiv preprint arXiv:2608.15659},
  year          = {2026},
  eprint        = {2608.15659},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2608.15659}
}