Reproduce “SONG: A Photorealistic 3D Gaussian Simulation Platform for Benchmarking Social 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
Weiqi Huang, Dianyi Yang, Jiaxin Li, Shuangyi Dong, Hao Xu, Zan Wang, Wei Liang
SONG introduces a photorealistic 3D Gaussian Splatting simulation platform and benchmark, SONG-Bench, for vision-based social navigation, leveraging semantics-driven human dynamics. Evaluations on SONG-Bench revealed that existing methods perform poorly in high-fidelity social environments, while fine-tuning on SONG's data improved real-world robot navigation.
Key insights
- Current state-of-the-art vision-based social navigation policies perform poorly on high-fidelity, dynamic human environments, underscoring that the problem is far from solved.
- Existing models demonstrate a foundational deficit in basic safety (collision avoidance and recovery) before progressing to nuanced social compliance.
- The distribution and realism of training data, particularly real-world trajectories, are more crucial for robust policy performance and real-world transfer than the sheer scale of the model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.