What's this about?
A benchmark testing whether vision-language-model agents can take real actions on complete 3D scenes, not just describe them or manipulate a single object. Existing 3D benchmarks score text answers or one-object operations, leaving whole-scene action untested (4 upvotes).
What exactly is "acting on a 3D scene"?
It means doing something in a three-dimensional space — moving objects, navigating, rearranging — based on what the agent sees, rather than answering a question about the scene. Describing a room ("the mug is on the table") is easy; reliably picking up the mug and putting it in the sink is the hard, action part.
Why does it matter?
Robotics and embodied agents ultimately need action, not narration. A benchmark that scores whole-scene action pushes VLM agents toward the physical competence they'll need — and exposes how far current models are from it.