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

✨ Quick Summary

OneOcc is a vision-only panoramic semantic occupancy framework for legged/humanoid robots. It handles gait-induced body jitter and 360° continuity, achieves state-of-the-art results, and remains lightweight and deployable.

---

🔬 Method at a Glance

• Dual-Projection Fusion (DP-ER): Uses both annular panorama and equirectangular projections to preserve 360° continuity and improve grid alignment.
• Bi-Grid Voxelization (BGV): Joint reasoning in Cartesian and cylindrical/polar voxel spaces to reduce discretization bias and sharpen boundaries.
• Hierarchical AMoE-3D: Dynamic multi-scale fusion for long-range and occlusion reasoning with a lightweight decoder.
• Gait Displacement Compensation (GDC): Plug-and-play feature-level motion correction without extra sensors to mitigate gait jitter.

---

🧪 Benchmarks & Key Results

• QuadOcc (real quadruped, first-person 360°):
  24K frames (stride-5 training), 6 classes, 64×64×8 grid (0.4 m/voxel); day/dusk/night; heterogeneous sequence split.
  OneOcc mIoU = 20.56, surpassing LiDAR LMSCNet (18.44) and the strongest vision baseline MonoScene (19.19).

• Human360Occ (H3O) (CARLA human-ego 360°, simulated gait):
  16 maps, diverse weather/lighting; provides RGB, depth, occupancy (two resolutions), and poses; report within-city and cross-city.
  OneOcc mIoU = 37.29 (within-city, +3.83) and 32.23 (cross-city, +8.08) versus the best vision baseline.

Lighting robustness on QuadOcc: OneOcc leads the best vision baseline in day (21.15 vs. 18.58) and dusk (19.86 vs. 15.14); at night, it obtains 13.50 mIoU (vs. 14.20) with higher precision.

---

🧱 Repository Status & Roadmap

• Sync preprint abstract/method/results (this README)
• Release training/inference code (oneocc/)
• Release checkpoints (assets/checkpoints/)
• Data preparation guides (QuadOcc / H3O)

Code and datasets will be made public upon publication.

---

📝 Citation

If you find this work useful, please cite:

@misc{shi2025oneoccsemanticoccupancyprediction,
  title         = {OneOcc: Semantic Occupancy Prediction for Legged Robots with a Single Panoramic Camera},
  author        = {Hao Shi and Ze Wang and Shangwei Guo and Mengfei Duan and Song Wang and Teng Chen and Kailun Yang and Lin Wang and Kaiwei Wang},
  year          = {2025},
  eprint        = {2511.03571},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2511.03571}
}