This field report tests open-set 3D scene graphs outdoors, using Terra 3DSG on five outdoor robot datasets. Outliers and multiple modes are common in VLM point embeddings (outlier ratios above 0.1 for about 30% of points), navigation-based object retrieval succeeds about 70% of the time with paths around 66% suboptimal, and region understanding is weak (mean F1 about 0.359). Graphs stay under 600 MB for multi-kilometre trajectories, and new consistency metrics check whether graph properties hold across repeated traversals.
Relevance
Outdoor numbers for 3D scene graphs, which are mostly evaluated indoors. The repeated-traversal consistency metrics are useful for both multi-domain perception and long-term autonomy.
Links
- arXiv
- Subfield: Semantic mapping, Robot learning