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)→
More from Embodied
- Musk envisions everyone owning a personal C-3PO: one person directing thousands of Optimus robots — elonmusk · 2026-09-23
- Muse previewed on Oakley Meta Vanguard glasses ahead of Meta Connect, app hits #1 in ~7 days — ChrisUniverse · 2026-09-23
- NVIDIA's OpenH-RF dataset hits Hugging Face trending under CC-BY-4.0 — nvidia · 2026-09-23
- Tesla Optimus Gen 3 mockups leaked via Android app APK assets — ChrisGPT · 2026-09-23
- Robotics research dinner weighs synthetic data gaps, onboard GPUs and world models — DynamicWebPaige · 2026-09-23
- Critic calls Optimus walking video fake AI-generated, sparking drama — TechPreacher · 2026-09-23