What's this about?
From local perception to spatial agency in a real-scale city — 105 upvotes. The task is acting purposefully at city scale rather than recognising things in the immediate view.
What is "spatial agency"?
The ability to act toward a goal in space you cannot see all of. Perception answers "what is in front of me"; agency requires maintaining a model of where things are relative to each other beyond your view, planning a route through it, and updating when reality disagrees. At the scale of a city that means a persistent map, not a sequence of frames — which is where most embodied systems fall over, because they are excellent at the current view and have no memory of the last kilometre.
Why does this matter to you?
- It is the same context-and-memory problem as long-horizon agents, expressed spatially.
- Real-scale is the claim to scrutinise: city-scale in simulation and city-scale in reality are very different results.
- Part of this week's embodied and world-model cluster, which dominated the paper rankings.