The Neverwhere Visual Parkour Benchmark Suite

Scenes for the Neverwhere simulator: real-world hurdles, gaps, ramps and stairs reconstructed as 3D Gaussian Splats with aligned collision meshes for MuJoCo. Paper: https://arxiv.org/abs/2609.16443 · Project page: https://ziyc.github.io/neverwhere-bench/

Layout

scenes/<scene>.zip          one self-contained zip per scene
checkpoints/lucidsim/       hurdle.pt, stairs.pt: the LucidSim example policy
scenes.txt                  scene list with task type
manifest.csv                size and sha256 of every zip

Each scene zip contains

3dgs/model.pt                        gsplat checkpoint used by the simulator
3dgs/model.splat                     same scene for web viewers
geometry/collision_mesh.obj          MuJoCo collision geometry
geometry/visual_mesh_simplified.obj
geometry/collision_tf.json
colmap/sparse/                       camera poses and sparse points
openmvs/                             dense reconstruction
polycam/ or raw_data/                the original capture
*.xml                                scene XML as labeled; the simulator ships its own copies in neverwhere/tasks/

Download

python download.py --scenes hurdle_stata_v1 building_31_stairs_v1
python download.py --all
python download.py --checkpoints

or hf_hub_download("ziyc/neverwhere", "scenes/hurdle_stata_v1.zip", repo_type="dataset").

Citation

@inproceedings{chen2026neverwhere,
  title     = {The Neverwhere Visual Parkour Benchmark Suite},
  author    = {Chen, Ziyu and Bao, Henghui and Chang, Haoran and Yu, Alan and Choi, Ran and McClennen, Kai and Huh, Gio
               and Yang, Kevin and Qiu, Ri-Zhao and Ravan, Yajvan and Leonard, John J. and Wang, Xiaolong and Isola, Phillip
               and Yang, Ge and Wang, Yue},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
}