Reproduce “OneOcc: Semantic Occupancy Prediction for Legged Robots with a Single Panoramic Camera” 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

Hao Shi, Ze Wang, Shangwei Guo, Mengfei Duan, Song Wang, Teng Chen, Kailun Yang, Lin Wang and 1 more

OneOcc, a vision-only semantic occupancy prediction framework, was developed for legged robots using a single panoramic camera. This framework effectively mitigates gait-induced body jitter and panoramic distortions, achieving a 20.56 mIoU on the real-world QuadOcc dataset and demonstrating robust generalization on synthetic environments.

Key insights

  • Compensating for gait-induced motion artifacts at the 2D feature sampling stage significantly improves 3D reconstruction accuracy and is more computationally efficient than correcting errors in quantized 3D data.
  • Fusing information from both raw annular and equirectangular panoramic projections, alongside a hybrid Cartesian and cylindrical-polar 3D grid, effectively mitigates distortions and preserves 360-degree continuity.
  • A dynamic, anisotropic 3D decoder utilizing a Mixture-of-Experts approach guided by gradient energy can better handle the diverse scale and feature anisotropies inherent in panoramic scenes, reducing over-smoothing at object boundaries.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.