Urbex Reframes City-Scale 3D Grounding as Active VLM Search

A new arxiv preprint proposes handling city-scale 3D grounding as a search problem rather than a ranking one. The paper is titled "Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes."

The setup it argues against is familiar: precompute a pile of instance proposals, then rank them against a referring expression. The authors write that this "becomes inefficient in large outdoor scenes." Their alternative represents a city as an interactive multi-view environment and lets a vision-language agent move through it: "Urbex represents a city-scale 3D scene as an interactive multi-view environment, allowing a VLM agent to query landmarks, zoom into local regions, render oblique views, and commit a grounded bounding box."

The policy is trained with reinforcement learning on CityRefer, "using smooth localization rewards and lightweight shaping rewards for efficient exploration." The paper claims Urbex "achieves the best Hit@1 and IoU on CityRefer, supports candidate-free grounding without ground-truth boxes, and is substantially faster than the strongest city-scale ranking baseline." It also reports zero-shot results on STPLS3D-Refer that "further indicate improved cross-scene robustness."

The abstract publishes no per-benchmark numbers, no latency figures, and no backbone name. It is one of 475 agent-themed items we have tracked in the last 90 days.