Claude Opus 5.5 precisely maps 3.25% of SpatialScout objects

Claude Opus 5.5, run as a VLM agent that can move a camera around a room, correctly placed 3.25% of target objects in a shared 3D frame when the bar was set to intersection-over-union above 0.9. At the looser IoU 0.5 threshold, 47.23% of targets passed. Its mean overlap across the task was 41.71%. That was the best score in the paper.

The benchmark is called SpatialScout. In a new arXiv preprint, the authors hand a VLM 80 indoor scenes, 3,140 target objects, an object inventory, RGB observations, camera calibration and camera poses, and let it steer the camera to find and localize everything inside a bounded interaction budget.

"Experiments with frontier VLMs reveal substantial limitations," the paper reports. The authors close the abstract bluntly: "These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps."

The abstract names Claude Opus 5.5 as the strongest evaluated model but publishes no per-model figures for the rest of the field, and does not break down which stage of the pipeline (viewpoint selection, cross-view association, geometric estimation, or map revision) accounts for the collapse from 47.23% at IoU 0.5 to 3.25% at IoU 0.9.