SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?
In brief
- SpatialScout evaluates VLM agents' ability to build 3D maps via autonomous exploration.
- The benchmark includes 80 indoor scenes and 3,140 objects to locate.
- The top tested model reaches only 41.71% average overlap score.
Researchers introduce SpatialScout, a benchmark to assess whether vision-language model (VLM) agents can build accurate 3D maps through autonomous exploration. According to the announcement, SpatialScout includes 80 indoor scenes with 3,140 target objects to locate. Agents must control a camera to explore each scene and estimate object centers within a shared coordinate frame, under limited interaction budgets.
Results reveal major gaps: the top tested model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with only 3.25% of objects exceeding an intersection-over-union (IoU) threshold of 0.9. SpatialScout provides a quantitative framework for developing agents capable of generating reliable spatial representations through active perception.
This summary gives you the essentials. The full article, its figures and quotes are on Hugging Face Daily Papers.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error