Reproduce “NavGSim: High-Fidelity Gaussian Splatting Simulator for Large-Scale Navigation” 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
Jiahang Liu, Yuanxing Duan, Jiazhao Zhang, Minghan Li, Shaoan Wang, Zhizheng Zhang, He Wang
Peking University, Galbot, and BAAI researchers present NavGSim, a simulator leveraging hierarchical 3D Gaussian Splatting (H3DGS) for high-fidelity, large-scale robot navigation environments. It enables a Vision-Language-Action (VLA) model to achieve an average real-world success rate of 69.2% for navigation tasks, outperforming standard 3DGS-trained models by an 81% higher real-world success rate.
Key insights
- H3DGS significantly improves rendering quality (e.g., PSNR 25.396 vs. 23.221 for standard 3DGS) and enhances sim-to-real transferability of navigation policies.
- A multi-slice Gaussian projection technique effectively allows for direct collision detection and extraction of navigable areas from continuous 3D Gaussian fields.
- Integrating pre-training and Dataset Aggregation (DAgger) during training accelerates learning and leads to higher success rates for VLA models in complex navigation scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.