Reproduce “The Neverwhere Visual Parkour Benchmark Suite” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziyu Chen, Henghui Bao, Haoran Chang, Alan Yu, Ran Choi, Kai McClennen, Gio Huh, Kevin Yang and 7 more
Neverwhere is a real-to-sim benchmark suite for evaluating quadruped visual locomotion in photorealistic, reconstructed environments. It combines 3D Gaussian Splatting for rendering with multi-view stereo meshes for MuJoCo collision geometry, covering more than 60 indoor and outdoor scenes and showing that high performance on individual scenes can conceal large generalization gaps of up to roughly 50 percentage points.
Key insights
- Visual quality and physical usability require separate representations: Gaussian splats provide detailed appearance, whereas reconstructed meshes support collision-aware simulation.
- Scene-specific policies can perform well on their training environments yet transfer poorly; multi-scene training reduces but does not remove this gap, exposing the importance of fixed, diverse evaluation sets.
- Geometry guidance generally improves reconstruction quality: for example, Splatfacto novel-view LPIPS falls from 0.172 to 0.161, while Splatfacto-MCMC reconstruction LPIPS decreases from 0.094 to 0.089.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.