Found first: a primary source the press has not covered yet.
Researchers at Peking University and Tianjin University have released SpatialScout, a benchmark that tasks VLM agents with actively exploring indoor environments and constructing metric 3D maps under a fixed interaction budget. Across 3,140 target objects, the best model tested achieved a mean overlap score of 41.71%, with only 3.25% of objects localized at high precision.
What the source says
SpatialScout covers 80 diverse indoor scenes containing 3,140 target objects. Each episode allows 100 model calls and 100 acquired images. Agents receive camera intrinsics and poses and use provided tools, including camera control and multi-view triangulation, to estimate object centers in a shared coordinate frame. Seven frontier models were evaluated: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, GPT-6 Sol, Kimi-K3, Qwen3.8-Max, and DeepSeek-V4.1-Flash. Claude Opus 5.5 was the strongest performer, reaching 41.71% mean overlap; 47.23% of targets exceeded IoU 0.5, and 3.25% exceeded IoU 0.9.
Why it matters
The gap between the IoU 0.5 result (47.23% of objects) and IoU 0.9 (3.25%) shows that current VLMs can recover rough object layouts but almost never achieve metric precision. Robotic manipulation and precise navigation require the latter. Spatial intelligence has drawn a year of lab announcements; the 3.25% figure quantifies the actual gap. Code is at github.com/JesseZZZZZ/SpatialScout.