UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City
UrbanGround tests whether MLLM agents can keep acting coherently once street-view perception becomes city-scale movement.
The sandbox is a physically constrained 3D replica of Hong Kong built from territory-wide geospatial data. Agents navigate it from a first-person view with an interactive map. The paper finds current systems can handle visual recognition and short-range spatial reasoning, but orientation and pedestrian-aware movement remain unreliable. The main failure is extended exploration: local abilities do not compound into sustained goal-directed behavior, and errors pile up without effective correction. HF Daily Papers' note
The sandbox is a physically constrained 3D replica of Hong Kong built from territory-wide geospatial data. Agents navigate it from a first-person view with an interactive map. The paper finds current systems can handle visual recognition and short-range spatial reasoning, but orientation and pedestrian-aware movement remain unreliable. The main failure is extended exploration: local abilities do not compound into sustained goal-directed behavior, and errors pile up without effective correction. HF Daily Papers' note
score 5