UrbanGround: a 3D Hong Kong sandbox exposes MLLM agents' poor spatial agency

zhangzhuosheng · x · 2026-09-02

Researchers built UrbanGround, the first city-scale sandbox replicating Hong Kong from 3D geospatial data, testing MLLM agents in first-person closed-loop navigation. Agents stay on pedestrian networks 93.0–96.4% of the time but comply with road closures only 10.0–46.7%, and collision rates reach 76.3–90.0% with moving pedestrians. Models show atomic visual/spatial skills but fail to sustain orientation over long routes. Paper and code are public.

Related event: UrbanGround: Real-Scale 3D Hong Kong Sandbox for Humans and Agents(2 posts)→

Original post →

More from Embodied

Embodied channel →