Sources
UrbanGround builds a physically constrained replica of Hong Kong to test whether MLLM agents can act after they move
Submitted 2026-08-27 with 71 upvotes on Hugging Face Daily Papers, UrbanGround is the first sandbox built from territory-wide 3D geospatial data where an agent enters a real-scale city in first person with an interactive map. The three research questions escalate from whether an agent can ground a local scene well enough to answer spatial questions after active observation, up through whether that grounding survives movement. The distinction it draws, between interpreting a street view and retaining useful local evidence once you start navigating, is the same gap that breaks GUI and browser agents on long-horizon tasks.
↳ Follow the thread